text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
arXiv:1712.04604v2 [cs.NE] 30 Jan 2018
Deep Quaternion Networks
Chase J. Gaudet
Anthony S. Maida
School of Computing & Informatics
University of Lousiana at Lafayette
Lafayette, USA
[email protected]
School of Computing & Informatics
University of Lousiana at Lafayette
Lafayette, USA
[email protected]
Abstract—The field of deep learning has seen significant
advancement in recent years. However, much of the existing work
has been focused on real-valued numbers. Recent work has shown
that a deep learning system using the complex numbers can be
deeper for a fixed parameter budget compared to its real-valued
counterpart. In this work, we explore the benefits of generalizing
one step further into the hyper-complex numbers, quaternions
specifically, and provide the architecture components needed to
build deep quaternion networks. We go over quaternion convolutions, present a quaternion weight initialization scheme, and
present algorithms for quaternion batch-normalization. These
pieces are tested in a classification model by end-to-end training
on the CIFAR-10 and CIFAR-100 data sets and a segmentation
model by end-to-end training on the KITTI Road Segmentation
data set. The quaternion networks show improved convergence
compared to real-valued and complex-valued networks, especially
on the segmentation task.
Index Terms—quaternion, complex, neural networks, deep
learning
I. I NTRODUCTION
There have been many advances in deep neural network
architectures in the past few years. One such improvement
is a normalization technique called batch normalization [1]
that standardizes the activations of layers inside a network
using minibatch statistics. It has been shown to regularize the
network as well as provide faster and more stable training.
Another improvement comes from architectures that add so
called shortcut paths to the network. These shortcut paths
connect later layers to earlier layers typically, which allows for
the stronger gradients to propagate to the earlier layers. This
method can be seen in Highway Networks [2] and Residual
Networks [3]. Other work has been done to find new activation
functions with more desirable properties. One example is
the exponential linear unit (ELU) [4], which attempts to
keep activations standardized. All of the above methods are
combating the vanishing gradient problem [5] that plagues
deep architectures. With solutions to this problem appearing
it is only natural to move to a system that will allow one to
construct deeper architectures with as low a parameter cost as
possible.
Other work in this area has explored the use of complex
and hyper-complex numbers, which are a generalization of
the complex, such as quaternions. Using complex numbers
in recurrent neural networks (RNNs) has been shown to
increase learning speed and provide a more noise robust
memory retrieval mechanism [6]–[8]. The first formulation of
complex batch normalization and complex weight initialization
is presented by [9] where they achieve some state of the art
results on the MusicNet data set. Hyper-complex numbers
are less explored in neural networks, but have seen use in
manual image and signal processing techniques [10]–[12].
Examples of using quaternion values in networks is mostly
limited to architectures that take in quaternion inputs or predict
quaternion outputs, but do not have quaternion weight values
[13], [14]. There are some more recent examples of building
models that use quaternions represented as real-values. In [15]
they used a quaternion multi-layer perceptron (QMLP) for
document understanding and [16] uses a similar approach in
processing multi-dimensional signals.
Building on [9] our contribution in this paper is to formulate
and implement quaternion convolution, batch normalization,
and weight initialization 1 . There arises some difficulty over
complex batch normalization that we had to overcome as their
is no analytic form for our inverse square root matrix.
II. M OTIVATION AND R ELATED W ORK
The ability of quaternions to effectively represent spatial
transformations and analyze multi-dimensional signals makes
them promising for applications in artificial intelligence.
One common use of quaternions is for representing rotation
into a more compact form. PoseNet [14] used a quaternion
as the target output in their model where the goal was to
recover the 6−DOF camera pose from a single RGB image.
The ability to encode rotations may make a quaternion network
more robust to rotational variance.
Quaternion representation has also been used in signal
processing. The amount of information in the phase of an
image has been shown to be sufficient to recover the majority
of information encoded in its magnitude by Oppenheim and
Lin [17]. The phase also encodes information such as shapes,
edges, and orientations. Quaternions can be represented as a
2 x 2 matrix of complex numbers, which gives them a group
of phases potentially holding more information compared to a
single phase.
Bulow and Sommer [12] used the higher complexity representation of quaternions by extending Gabor’s complex signal
to a quaternion one which was then used for texture segmentation. Another use of quaternion filters is shown in [11] where
1 Source
code located at https://github.com/gaudetcj/DeepQuaternionNetworks
they introduce a new class of filter based on convolution with
hyper-complex masks, and present three color edge detecting
filters. These filters rely on a three-space rotation about the
grey line of RGB space and when applied to a color image
produce an almost greyscale image with color edges where the
original image had a sharp change of color. More quaternion
filter use is shown in [18] where they show that it is effective in
the context of segmenting color images into regions of similar
color texture. They state the advantage of using quaternion
arithmetic is that a color can be represented and analyzed
as a single entity, which we will see holds for quaternion
convolution in a convolutional neural network architecture as
well in Section III-C.
A quaternionic extension of a feed forward neural network, for processing multi-dimensional signals, is shown in
[16]. They expect that quaternion neurons operate on multidimensional signals as single entities, rather than real-valued
neurons that deal with each element of signals independently.
A convolutional neural network (CNN) should be able to learn
a powerful set of quaternion filters for more impressive tasks.
Another large motivation is discussed in [9], which is that
complex numbers are more efficient and provide more robust
memory mechanisms compared to the reals [10]–[12]. They
continue that residual networks have a similar architecture to
associative memories since the residual shortcut paths compute
their residual and then sum it into the memory provided by
the identity connection. Again, given that quaternions can be
represented as a complex group, they may provide an even
more efficient and robust memory mechanisms.
III. Q UATERNION N ETWORK C OMPONENTS
This section will include the work done to obtain a working
deep quaternion network. Some of the longer derivations are
given in the Appendix.
A. Quaternion Representation
This representation of quaternions is not unique, but we will
stick to the above in this paper. It is also possible to represent
H as M (2, C) where M (2, C) is a 2 x 2 complex matrix.
With our real-valued representation a quaternion real-valued
2D convolution layer can be expressed as follows. Say that the
layer has N feature maps such that N is divisible by 4. We let
the first N/4 feature maps represent the real components, the
second N/4 represent the i imaginary components, the third
N/4 represent the j imaginary components, and the last N/4
represent the k imaginary components.
B. Quaternion Differentiability
In 1833 Hamilton proposed complex numbers C be defined
as the set R2 of ordered pairs (a, b) of real numbers. He
then began working to see if triplets (a, b, c) could extend
multiplication of complex numbers. In 1843 he discovered a
way to multiply in four dimensions instead of three, but the
multiplication lost commutativity. This construction is now
known as quaternions. Quaternions are composed of four
components, one real part, and three imaginary parts. Typically
denoted as
H = {a + bi + cj + dk : a, b, c, d ∈ R}
Since we will be performing quaternion arithmetic using
reals it is useful to embed H into a real-valued representation.
There exists an injective homomorphism from H to the matrix
ring M (4, R) where M (4, R) is a 4x4 real matrix. The 4 x 4
matrix can be written as
a −b −c −d
1 0 0 0
b a −d c
= a 0 1 0 0
c d
0 0 1 0
a −b
d −c b
a
0 0 0 1
0 −1 0 0
1 0 0 0
+ b
0 0 0 −1
0 0 1 0
0 0 −1 0
0 0
0 1
+ c
1 0
0 0
0 −1 0 0
0 0 0 −1
0 0 −1 0
. (4)
+ d
0 1 0
0
1 0 0
0
(1)
In order for the network to perform backpropagation the cost
function and activation functions used must be differentiable
with respect to the real, i, j, and k components of each
quaternion parameter of the network. As the complex chain
rule is shown in [9], we provide the quaternion chain rule
which is given in the Appendix section VII-A.
C. Quaternion Convolution
where a is the real part, (i, j, k) denotes the three imaginary
axis, and (b, c, d) denotes the three imaginary components.
Quaternions are governed by the following arithmetic:
Convolution in the quaternion domain is done by convolving
a quaternion filter matrix W = A+iB+jC+kD by a quaternion
vector h = w + ix + jy + kz. Performing the convolution by
using the distributive property and grouping terms one gets
i2 = j 2 = k 2 = ijk = −1
(2)
W ∗ h = (A ∗ w − B ∗ x − C ∗ y − D ∗ z)+
which, by enforcing distributivity, leads to the noncommutative
multiplication rules
i(A ∗ x + B ∗ w + C ∗ z − D ∗ y)+
j(A ∗ y − B ∗ z + C ∗ w + D ∗ x)+
ij = k, jk = i, ki = j, ji = −k, kj = −i, ik = −j (3)
k(A ∗ z + B ∗ y − C ∗ x + D ∗ w).
(5)
Using a matrix to represent the components of the convolution
we have:
R(W ∗ h)
A
I (W ∗ h) B
J (W ∗ h) = C
K (W ∗ h)
D
−B
A
D
−C
−C −D
w
x
−D C
∗
A −B y
B
A
z
(6)
An example is shown in Fig. 1 where one can see how quaternion convolution forces a linear depthwise mixture of the
channels. This is similar to a mixture of standard convolution
and depthwise separable convolution from [19]. This reuse of
filters on every layer and combination may help extract texture
information across channels as seen in [18].
sition of V−1 where V is the covariance matrix given by:
Vrr Vri Vrj Vrk
Vir Vii Vij Vik
V =
Vjr Vji Vjj Vjk
Vkr Vki Vkj Vkk
Cov(R{x}, R{x})
Cov(I {x}, R{x})
=
Cov(J {x}, R{x})
Cov(K {x}, R{x})
Cov(R{x}, I {x})
Cov(I {x}, I {x})
Cov(J {x}, I {x})
Cov(K {x}, I {x})
Cov(R{x}, J {x})
Cov(I {x}, J {x})
Cov(J {x}, J {x})
Cov(K {x}, J {x})
Cov(R{x}, K {x})
Cov(I {x}, K {x})
Cov(J {x}, K {x})
Cov(K {x}, K {x})
D. Quaternion Batch-Normalization
Batch-normalization [1] is used by the vast majority of all
deep networks to stabilize and speed up training. It works
by keeping the activations of the network at zero mean and
unit variance. The original formulation of batch-normalization
only works for real-values. Applying batch normalization to
complex or hyper-complex numbers is more difficult, one can
not simply translate and scale them such that their mean is
0 and their variance is 1. This would not give equal variance
in the multiple components of a complex or hyper-complex
number. To overcome this for complex numbers a whitening
approach is used [9], which scales the data by the square root
of their variances along each of the two principle components.
We use the same approach, but must whiten 4D vectors.
However, an issue arises in that there is no nice way to
calculate the inverse square root of a 4 x 4 matrix. It turns out
that the square root is not necessary and we can instead use the
Cholesky decomposition on our covariance matrix. The details
of why this works for whitening given in the Appendix section
VII-B. Now our whitening is accomplished by multiplying the
0-centered data (x − E[x]) by W:
x̃ = W(x − E[x])
(7)
where Cov(·) is the covariance and R{x}, I {x}, J {x}, and
K {x} are the real, i, j, and k components of x respectively.
Real-valued batch normalization also uses two learned
parameters, β and γ. Our shift parameter β must shift a
quaternion value so it is a quaternion value itself with real,
i, j, and k as learnable components. The scaling parameter γ
is a symmetric matrix of size matching V given by:
Vrr Vri Vrj Vrk
Vri Vii Vij Vik
γ=
(9)
Vrj Vij Vjj Vjk
Vrk Vik Vjk Vkk
Because of its symmetry it has only ten learnable parameters.
The variance of the components of input
√ x̃ are variance 1 so
the diagonal of γ is initialized to 1/ 16 in order to obtain a
modulus of 1 for the variance of the normalized value. The off
diagonal terms of γ and all components of β are initialized
to 0. The quaternion batch normalization is defined as:
BN(x̃) = γx̃ + β
(10)
E. Quaternion Weight Initialization
The proper initialization of weights is vital to convergence
of deep networks. In this work we derive our quaternion
weight initialization using the same procedure as Glorot and
Bengio [20] and He et al. [21].
To begin we find the variance of a quaternion weight:
W = |W |e(cosφ1 i+cosφ2 j+cosφ3 k)θ
= R{W } + I {W } + J {W } + K {W }.
where W is one of the matrices from the Cholesky decompo-
(8)
(11)
where |W | is the magnitude, θ and φ are angle arguments,
and cos2 φ1 + cos2 φ2 + cos2 φ3 = 1 [22].
Fig. 1. An illustration of quaternion convolution.
Variance is defined as
Var(W ) = E[|W |2 ] − (E[W ])2 ,
(12)
but since W is symmetric around 0 the term (E[W ])2 is 0.
We do not have a way to calculate Var(W ) = E[|W |2 ] so
we make use of the magnitude of quaternion normal values
|W |, which follows an independent normal distribution with
four degrees of freedom (DOFs). We can then calculate the
expected value of |W |2 to find our variance
Z ∞
E[|W |2 ] =
x2 f (x) dx = 4σ 2
(13)
∞
where f (x) is the four DOF distribution given in the Appendix.
And since Var(W ) = E[|W |2 ], we now have the variance
of W expressed in terms of a single parameter σ:
Var(W ) = 4σ 2 .
(14)
To follow the Glorot and Bengio [20] initialization we have
Var(W ) = 2/(nin +nout ), where nin and nout are the number
of input and output units respectivly.
p Setting this equal to (14)
and solving for σ gives σ = 1/ 2(nin + nout ). To follow
He et al. [21] initialization that is specialized for rectified
linear units (ReLUs) [23], then we have Var(W ) = 2/nin ,
which √
again setting equal to (14) and solving for σ gives
σ = 1/ 2nin .
As shown in (11) the weight has components |W |, θ, and
φ. We can initialize the magnitude |W | using our four DOF
distribution defined with the appropriate σ based on which
initialization scheme we are following. The angle components
are initialized using the uniform distribution between −π and
π where we ensure the constraint on φ.
IV. E XPERIMENTAL R ESULTS
Our experiments covered image classification using both the
CIFAR-10 and CIFAR-100 benchmarks and image segmentation using the KITTI Road Estimation benchmark.
A. Classification
We use the same architecture as the large model in [9],
which is a 110 layer Residual model similar to the one in
[3]. There is one difference between the real-valued network
and the ones used for both the complex and hyper-complex
valued networks. Because the datasets are all real-valued the
network must learn the imaginary or quaternion components.
We use the same technique as [9] where there is an additional
block immediately after the input which will learn the hypercomplex components
BN → ReLU → Conv → BN → ReLU → Conv.
Since all datasets that we work with are real-valued, we
present a way to learn their imaginary components to let the
rest of the network operate in the complex plane. We learn the
initial imaginary component of our input by performing the
operations present within a single real-valued residual block
This means that to maintain the same approximate parameter count the number of convolution kernels for the complex
network was increased. We however did not increase the
number of convolution kernels for the quaternion trials so
any increase in model performance comes from the quaternion
filters and at a lower hardware budget.
The architecture consists of 3 stages of repeating residual
blocks where at the end of each stage the images are downsized by strided convolutions. Each stage also doubles the
previous stage’s number of convolution kernels. The last layers
are a global average pooling layer followed by a single fully
connected layer with a softmax function used to classify the
input as either one of the 10 classes in CIFAR-10 or one of
the 100 classes in CIFAR-100.
We also followed their training procedure of using the
backpropagation algorithm with Stochastic Gradient Descent
with Nesterov momentum [24] set at 0.9. The norm of the
gradients are clipped to 1 and a custom learning rate scheduler
is used. The learning scheduler is the same used in [9] for a
direct comparison in performance. The learning rate is initially
set to 0.01 for the first 10 epochs and then set it to 0.1 from
epoch 10-100 and then cut by a factor of 10 at epochs 120
and 150. Table I presents our results along side the real and
complex valued networks. Our quaternion model outperforms
the real and complex networks on both datasets on a smaller
parameter budget.
Architecture
[3] Real
[9] Complex
Quaternion
CIFAR-10
CIFAR-100
6.37
5.60
27.09
5.44
26.01
TABLE I
C LASSIFICATION ERROR ON CIFAR-10 AND CIFAR-100. N OTE THAT [3]
IS A 110 LAYER RESIDUAL NETWORK , [9] IS 118 LAYER COMPLEX
NETWORK WITH THE SAME DESIGN AS THE PRIOR EXCEPT WITH
ADDITIONAL INITIAL LAYERS TO EXTRACT COMPLEX MAPPINGS .
B. Segmentation
For this experiment we used the same model as the above,
but cut the number of residual blocks out of the model for
memory reasons given that the KITTI data is large color
images. The 110 layer model has 10, 9, and 9 residual blocks
in the 3 stages talked about above, while this model has 2,
1, and 1 and does not perform any strided convolutions. This
gives a total of 38 layers.
The last layer is a 1 × 1 convolution with a sigmoid output
so we are getting a heatmap prediction the same size as the
input. The training procedure is also as above, but the learning
rate is scheduled differently. Here we begin at 0.01 for the first
10 epochs and then set it to 0.1 from epoch 10-50 and then cut
by a factor of 10 at 100 and 150. Table II presents our results
along side the real and complex valued networks where we
used Intersection over Union (IOU) for performance measure.
Quaternion outperformed the other two by a larger margin
compared to the classification tasks.
Architecture
KITTI
Real
0.747
Complex
0.769
Quaternion
0.827
TABLE II
IOU ON KITTI ROAD E STIMATION BENCHMARK .
V. C ONCLUSIONS
We have extended upon work looking into complex valued
networks by exploring quaternion values. We presented the
building blocks required to build and train deep quaternion
networks and used them to test residual architectures on
two common image classification benchmarks. We show that
they have competitive performance by beating both the real
and complex valued networks with less parameters. Future
work will be needed to test quaternion networks for more
segmentation datasets and for audio processing tasks.
VI. ACKNOWLEDGMENT
[15] T. Parcollet, M. Morchid, P.-M. Bousquet, R. Dufour, G. Linarès,
and R. De Mori, “Quaternion neural networks for spoken language
understanding,” in Spoken Language Technology Workshop (SLT), 2016
IEEE, pp. 362–368, IEEE, 2016.
[16] T. Minemoto, T. Isokawa, H. Nishimura, and N. Matsui, “Feed forward
neural network with random quaternionic neurons,” Signal Processing,
vol. 136, pp. 59–68, 2017.
[17] A. V. Oppenheim and J. S. Lim, “The importance of phase in signals,”
Proceedings of the IEEE, vol. 69, no. 5, pp. 529–541, 1981.
[18] L. Shi and B. Funt, “Quaternion color texture segmentation,” Computer
Vision and image understanding, vol. 107, no. 1, pp. 88–96, 2007.
[19] F. Chollet, “Xception: Deep learning with depthwise separable convolutions,” arXiv preprint arXiv:1610.02357, 2016.
[20] X. Glorot and Y. Bengio, “Understanding the difficulty of training deep
feedforward neural networks,” in Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, pp. 249–256,
2010.
[21] K. He, X. Zhang, S. Ren, and J. Sun, “Delving deep into rectifiers:
Surpassing human-level performance on imagenet classification,” in
Proceedings of the IEEE international conference on computer vision,
pp. 1026–1034, 2015.
[22] R. Piziak and D. Turner, “The polar form of a quaternion,” 2002.
We would like to thank James Dent of the University
of Louisiana at Lafayette Physics Department for helpful
discussions. We also thank Fugro for research time on this
project.
[23] V. Nair and G. E. Hinton, “Rectified linear units improve restricted boltzmann machines,” in Proceedings of the 27th international conference on
machine learning (ICML-10), pp. 807–814, 2010.
R EFERENCES
[25] A. Kessy, A. Lewin, and K. Strimmer, “Optimal whitening and decorrelation,” The American Statistician, no. just-accepted, 2017.
[1] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep
network training by reducing internal covariate shift,” in International
Conference on Machine Learning, pp. 448–456, 2015.
[2] R. K. Srivastava, K. Greff, and J. Schmidhuber, “Training very deep networks,” in Advances in neural information processing systems, pp. 2377–
2385, 2015.
[3] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” in Proceedings of the IEEE conference on computer vision
and pattern recognition, pp. 770–778, 2016.
[4] D.-A. Clevert, T. Unterthiner, and S. Hochreiter, “Fast and accurate
deep network learning by exponential linear units (elus),” arXiv preprint
arXiv:1511.07289, 2015.
[5] S. Hochreiter, “Untersuchungen zu dynamischen neuronalen netzen,”
Diploma, Technische Universität München, vol. 91, 1991.
[6] M. Arjovsky, A. Shah, and Y. Bengio, “Unitary evolution recurrent
neural networks,” in International Conference on Machine Learning,
pp. 1120–1128, 2016.
[7] I. Danihelka, G. Wayne, B. Uria, N. Kalchbrenner, and A. Graves, “Associative long short-term memory,” arXiv preprint arXiv:1602.03032,
2016.
[8] S. Wisdom, T. Powers, J. Hershey, J. Le Roux, and L. Atlas, “Fullcapacity unitary recurrent neural networks,” in Advances in Neural
Information Processing Systems, pp. 4880–4888, 2016.
[9] C. Trabelsi, O. Bilaniuk, D. Serdyuk, S. Subramanian, J. F. Santos,
S. Mehri, N. Rostamzadeh, Y. Bengio, and C. J. Pal, “Deep complex
networks,” arXiv preprint arXiv:1705.09792, 2017.
[10] T. Bülow, Hypercomplex spectral signal representations for the processing and analysis of images. Universität Kiel. Institut für Informatik und
Praktische Mathematik, 1999.
[11] S. J. Sangwine and T. A. Ell, “Colour image filters based on hypercomplex convolution,” IEE Proceedings-Vision, Image and Signal
Processing, vol. 147, no. 2, pp. 89–93, 2000.
[12] T. Bulow and G. Sommer, “Hypercomplex signals-a novel extension of
the analytic signal to the multidimensional case,” IEEE Transactions on
signal processing, vol. 49, no. 11, pp. 2844–2852, 2001.
[13] A. Rishiyur, “Neural networks with complex and quaternion inputs,”
arXiv preprint cs/0607090, 2006.
[14] A. Kendall, M. Grimes, and R. Cipolla, “Posenet: A convolutional
network for real-time 6-dof camera relocalization,” in Proceedings of
the IEEE international conference on computer vision, pp. 2938–2946,
2015.
[24] Y. Nesterov, “A method of solving a convex programming problem with
convergence rate o (1/k2),”
VII. A PPENDIX
A. The Generalized Quaternion Chain Rule for a Real-Valued
Function
We start by specifying the Jacobian. Let L be a real valued
loss function and q be a quaternion variable such that q =
a + i b + j c + k d where a, b, c, d ∈ R then,
∂L
∂L
∂L
∂L
∂L
=
+i
+j
+k
(15)
∂q
∂a
∂b
∂c
∂d
∂L
∂L
∂L
∂L
=
+i
+j
+k
∂R(q)
∂I(q)
∂J(q)
∂K(q)
= R(∇L (q)) + I(∇L (q)) + J(∇L (q)) + K(∇L (q))
∇L (q) =
Now let g = m + i n + j o + k p be another quaternion variable
where q can be expressed in terms of g and m, n, o, p ∈ R we
where W is an n x n ‘whitening’ matrix. Since cov(Z) = I it
follows that:
then have,
∂L
∂L
∂L
∂L
∂L
=
+i
+j
+k
(16)
∂g
∂m
∂n
∂o
∂p
∂L ∂a
∂L ∂b
∂L ∂c
∂L ∂d
=
+
+
+
∂a ∂m
∂b ∂m
∂c ∂m
∂d ∂m
∂L ∂a ∂L ∂b
∂L ∂c
∂L ∂d
+i
+
+
+
∂a ∂n
∂b ∂n
∂c ∂n
∂d ∂n
∂L ∂a ∂L ∂b ∂L ∂c ∂L ∂d
+j
+
+
+
∂a ∂o
∂b ∂o
∂c ∂o
∂d ∂o
∂L ∂a ∂L ∂b ∂L ∂c ∂L ∂d
+k
+
+
+
∂a ∂p
∂b ∂p
∂c ∂p
∂d ∂p
∂L ∂a
∂a
∂a
∂a
=
+i
+j
+k
∂a ∂m
∂n
∂o
∂p
∂L ∂b
∂b
∂b
∂b
+
+i
+j
+k
∂b ∂m
∂n
∂o
∂p
∂c
∂c
∂c
∂L ∂c
+i
+j
+k
+
∂c ∂m
∂n
∂o
∂p
∂L ∂d
∂d
∂d
∂d
+
+i
+j
+k
∂d ∂m
∂n
∂o
∂p
∂L
∂a
∂a
∂a
∂a
=
+i
+j
+k
∂R(q) ∂m
∂n
∂o
∂p
∂b
∂b
∂b
∂L
∂b
+i
+j
+k
+
∂I(q) ∂m
∂n
∂o
∂p
∂c
∂c
∂c
∂L
∂c
+i
+j
+k
+
∂J(q) ∂m
∂n
∂o
∂p
∂d
∂d
∂d
∂d
∂L
+i
+j
+k
+
∂K(q) ∂m
∂n
∂o
∂p
∂a
∂a
∂a
∂a
= R(∇L (q))
+i
+j
+k
∂m
∂n
∂o
∂p
∂b
∂b
∂b
∂b
+i
+j
+k
+ I(∇L (q))
∂m
∂n
∂o
∂p
∂c
∂c
∂c
∂c
+ J(∇L (q))
+i
+j
+k
∂m
∂n
∂o
∂p
∂d
∂d
∂d
∂d
+ K(∇L (q))
+i
+j
+k
∂m
∂n
∂o
∂p
∇L (q) =
B. Whitening a Matrix
E[ZZT ] = I
E[W(X − µ)(W(X − µ))T ] = I
E[W(X − µ)(X − µ)T WT ] = I
WΣWT = I
WΣWT W = W
WT W = Σ−1
(18)
From (18) it is clear that the Cholesky decomposition provides
a suitable (but not unique) method of finding W.
C. Cholesky Decomposition
Cholesky decomposition is an efficient way to implement
LU decomposition for symmetric matrices, which allows us
to find the square root. Consider AX = b, A = [aij ]n×n , and
aij = aji , then the Cholesky decomposition of A is given by
A = LL0 where
l11 0 . . . 0
..
l21 l22 . . .
.
(19)
L= .
.
.
.
.
.
.
.
0
.
ln1 ln2 ... lnn
Let lki be the k th row and ith column entry of L, then
k<i
0,
q
Pi−1 2
lki =
aii − j=1 lkj ,
k=i
1 (a − Pi−1 l l ), i < k
ki
j=1 ij kj
lii
D. 4 DOF Independent Normal Distribution
Consider the four-dimensional vector Y = (S, T, U, V )
which has components that are normally distributed, centered
at zero, and independent. Then S, T , U , and V all have density
functions
2
Hx
Let X be an n x n matrix and cov(X) = Σ is the
symmetric covariance matrix of the same size. Whitening a
matrix linearly decorrelates the input dimensions, meaning that
whitening transforms X into Z such that cov(Z) = I where I
is the identity matrix [25]. The matrix Z can be written as:
Z = W(X − µ)
(17)
2
e−x /(2σ )
fS (x; σ) = fT (x; σ) = fU (x; σ) = fV (x; σ) = √
.
2πσ 2
(20)
Let
X
be
the
length
of
Y,
which
means
X
=
√
S 2 + T 2 + U 2 + V 2 . Then X has the cumulative distribution function
ZZZZ
FX (x; σ) =
fS (µ; σ)fT (µ; σ)fU (µ; σ)fV (µ; σ) dA,
where Hx is the four-dimensional sphere
n
o
p
Hx = (s, t, u, v) :
s2 + t2 + u2 + v 2 < x .
(21)
(22)
We then can write the integral in polar representation
Z πZ πZ 2π
Z x
2
1
3 −r
2σ 2 sin(θ)sin(φ)cos(ψ) drdθdφdψ
FX (x; σ) =
r
e
4π 2 σ 4 0 0 0 0
Z x
2
2
1
=
r3 e−r /(2σ ) dr.
(23)
4
2σ 0
The probability density function of X is the derivative of its
cumulative distribution function so we use the funamental
theorem of calculus on (23) to finally arrive at
fX (x; σ) =
d
1 3 −x2 /(2σ2 )
x e
.
FX (x; σ) =
dx
2σ 4
(24)
| 1cs.CV
|
Fast Compressed Self-Indexes with Deterministic
Linear-Time Construction∗
J. Ian Munro1 , Gonzalo Navarro2 , and Yakov Nekrich1
1
Cheriton School of Computer Science, University of Waterloo.
[email protected], [email protected].
CeBiB — Center of Biotechnology and Bioengineering, Department of
Computer Science, University of Chile. [email protected].
2
arXiv:1707.01743v3 [cs.DS] 1 Sep 2017
Abstract
We introduce a compressed suffix array representation that, on a text T of length n over an
alphabet of size σ, can be built in O(n) deterministic time, within O(n log σ) bits of working
space, and counts the number of occurrences of any pattern P in T in time O(|P | + log logw σ) on
a RAM machine of w = Ω(log n)-bit words. This new index outperforms all the other compressed
indexes that can be built in linear deterministic time, and some others. The only faster indexes
can be built in linear time only in expectation, or require Θ(n log n) bits. We also show that, by
using O(n log σ) bits, we can build in linear time an index that counts in time O(|P |/ logσ n +
log n(log log n)2 ), which is RAM-optimal for w = Θ(log n) and sufficiently long patterns.
1998 ACM Subject Classification E.1 Data Structures; E.4 Coding and Information Theory
Keywords and phrases Succinct data structures; Self-indexes; Suffix arrays; Deterministic construction
Digital Object Identifier 10.4230/LIPIcs...
1
Introduction
The string indexing problem consists in preprocessing a string T so that, later, we can
efficiently find occurrences of patterns P in T . The most popular solutions to this problem
are suffix trees [35] and suffix arrays [24]. Both can be built in O(n) deterministic time
on a text T of length n over an alphabet of size σ, and the best variants can count the
number of times a string P appears in T in time O(|P |), and even in time O(|P |/ logσ n)
in the word-RAM model if P is given packed into |P |/ logσ n words [31]. Once counted,
each occurrence can be located in O(1) time. Those optimal times, however, come with two
important drawbacks:
The variants with this counting time cannot be built in O(n) worst-case time.
The data structures use Θ(n log n) bits of space.
The reason of the first drawback is that some form of perfect hashing is always used to
ensure constant time per pattern symbol (or pack of symbols). The classical suffix trees
and arrays with linear-time deterministic construction offer O(|P | log σ) or O(|P | + log n)
counting time, respectively. More recently, those times have been reduced to O(|P | + log σ)
[10] and even to O(|P | + log log σ) [15]. Simultaneously with our work, a suffix tree variant
was introduced by Bille et al. [7], which can be built in linear deterministic time and counts
in time O(|P |/ logσ n + log |P | + log log σ). All those indexes, however, still suffer from the
∗
Funded with Basal Funds FB0001, Conicyt, Chile.
© J. Ian Munro, Gonzalo Navarro, Yakov Nekrich;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
second drawback, that is, they use Θ(n log n) bits of space. This makes them impractical in
most applications that handle large text collections.
Research on the second drawback dates back to almost two decades [30], and has led
to indexes using nHk (T ) + o(n(Hk (T ) + 1)) bits, where Hk (T ) ≤ log σ is the k-th order
entropy of T [25], for any k ≤ α logσ n − 1 and any constant 0 < α < 1. That is, the
indexes use asymptotically the same space of the compressed text, and can reproduce the
text and search it; thus they are called self-indexes. The fastest compressed self-indexes
that can be built in linear deterministic time are able to count in time O(|P | log log σ) [1] or
O(|P |(1 + logw σ)) [6]. There exist other compressed self-indexes that obtain times O(|P |) [5]
or O(|P |/ logσ n + logσ n) for any constant > 0 [19], but both rely on perfect hashing and
are not built in linear deterministic time. All those compressed self-indexes use O(n logb n )
further bits to locate the position of each occurrence found in time O(b), and to extract any
substring S of T in time O(|S| + b).
In this paper we introduce the first compressed self-index that can be built in O(n)
deterministic time (moreover, using O(n log σ) bits of space [28]) and with counting time
O(|P | + log logw σ), where w = Ω(log n) is the size in bits of the computer word. More
precisely, we prove the following result.
I Theorem 1. On a RAM machine of w = Ω(log n) bits, we can construct an index for
a text T of length n over an alphabet of size σ = O(n/ log n) in O(n) deterministic time
using O(n log σ) bits of working space. This index occupies nHk (T ) + o(n log σ) + O(n logb n )
bits of space for a parameter b and any k ≤ α logσ n − 1, for any constant 0 < α < 1. The
occurrences of a pattern string P can be counted in O(|P | + log logw σ) time, and then each
such occurrence can be located in O(b) time. An arbitrary substring S of T can be extracted
in time O(|S| + b).
We obtain our results with a combination of the compressed suffix tree T of T and the
Burrows-Wheeler transform B of the reversed text T . We manage to simulate the suffix tree
traversal for P , simultaneously on T and on B. With a combination of storing deterministic
dictionaries and precomputed rank values for sampled nodes of T , and a constant-time
method to compute an extension of partial rank queries that considers small ranges in B,
we manage to ensure that all the suffix tree steps, except one, require constant time. The
remaining one is solved with general rank queries in time O(log logw σ). As a byproduct, we
show that the compressed sequence representations that obtain those rank times [6] can also
be built in linear deterministic time.
Compared with previous work, other indexes may be faster at counting, but either they
are not built in linear deterministic time [5, 19, 31] or they are not compressed [31, 7]. Our
index outperforms all the previous compressed [13, 1, 6], as well as some uncompressed [15],
indexes that can be built deterministically.
As an application of our tools, we also show that an index using O(n log σ) bits of
space can be built in linear deterministic time, so that it can count in time O(|P |/ logσ n +
log n(log log n)2 ), which is RAM-optimal for w = Θ(log n) and sufficiently long patterns.
Current indexes obtaining similar counting time require O(n log σ) construction time [19] or
higher [31], or O(n log n) bits of space [31, 7].
2
Related Work
Let T be a string of length n over an alphabet of size σ that is indexed to support searches
for patterns P . It is generally assumed that σ = o(n), a reasonable convention we will follow.
© J. Ian Munro, Gonzalo Navarro, Yakov Nekrich;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
Compressed
Compact
Uncompressed
Deterministic
|P | log log σ [1]
|P |(1 + logw σ) [6]
|P |+log logw σ (ours)
|P |/ log n + log n [19]
(constant σ)
|P |/ logσ n+log n(log log n)2
(ours)
|P | + log log σ [15]
|P |/ logσ n+log |P |+log log σ [7]
Randomized
|P |(1 + log logw σ) [6]
|P | [5]
|P |/ logσ n + logσ n [19, 31]
Table 1 Our results in context. The x axis refers to the space used by the indexes (compressed
meaning nHk (T ) + o(n log σ) bits, compact meaning O(n log σ) bits, and uncompressed meaning
Θ(n log n) bits), and the y axis refers to the linear-time construction. In the cells we show the
counting time for a pattern P . We only list the dominant alternatives, graying out those outperformed
by our new results.
Searches typically require to count the number of times P appears in T , and then locate the
positions of T where P occurs. The vast majority of the indexes for this task are suffix tree
[35] or suffix array [24] variants.
The suffix tree can be built in linear deterministic time [35, 26, 34], even on arbitrarily
large integer alphabets [11]. The suffix array can be easily derived from the suffix tree in
linear time, but it can also be built independently in linear deterministic time [23, 22, 21].
In their basic forms, these structures allow counting the number of occurrences of a pattern
P in T in time O(|P | log σ) (suffix tree) or O(|P | + log n) (suffix array). Once counted, the
occurrences can be located in constant time each.
Cole et al. [10] introduced the suffix trays, a simple twist on suffix trees that reduces their
counting time to O(|P | + log σ). Fischer and Gawrychowski [15] introduced the wexponential
search trees, which yield suffix trees with counting time O(|P | + log log σ) and support
dynamism.
All these structures can be built in linear deterministic time, but require Θ(n log n) bits
of space, which challenges their practicality when handling large text collections.
Faster counting is possible if we resort to perfect hashing and give away the linear
deterministic construction time. In the classical suffix tree, we can easily achieve O(|P |) time
by hashing the children of suffix tree nodes, and this is optimal in general. In the RAM model
with word size Θ(log n), and if the consecutive symbols of P come packed into |P |/ logσ n
words, the optimal time is instead O(|P |/ logσ n). This optimal time was recently reached by
Navarro and Nekrich [31] (note that their time is not optimal if w = ω(log n)), with a simple
application of weak-prefix search, already hinted in the original article [2]. However, even
the randomized construction time of the weak-prefix search structure is O(n log n), for any
constant > 0. By replacing the weak-prefix search with the solution of Grossi and Vitter
[19] for the last nodes of the search, and using a randomized construction of their perfect
hash functions, the index of Navarro and Nekrich [31] can be built in linear randomized time
and count in time O(|P |/ logσ n + logσ n). Only recently, simultaneously with our work, a
deterministic linear-time construction algorithm was finally obtained for an index obtaining
O(|P |/ logσ n + log |P | + log log σ) counting time [7].
Still, these structures are not compressed. Compressed suffix trees and arrays appeared in
© J. Ian Munro, Gonzalo Navarro, Yakov Nekrich;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
the year 2000 [30]. To date, they take the space of the compressed text and replace it, in the
sense that they can extract any desired substring of T ; they are thus called self-indexes. The
space occupied is measured in terms of the k-th order empirical entropy of T , Hk (T ) ≤ log σ
[25], which is a lower bound on the space reached by any statistical compressor that encodes
each symbol considering only the k previous ones. Self-indexes may occupy as little as
nHk (T ) + o(n(Hk (T ) + 1)) bits, for any k ≤ α logσ n − 1, for any constant 0 < α < 1.
The fastest self-indexes with linear-time deterministic construction are those of Barbay
et al. [1], which counts in time O(|P | log log σ), and Belazzougui and Navarro [6, Thm. 7],
which counts in time O(|P |(1 + logw σ)). The latter requires O(n(1 + logw σ)) construction
time, but if log σ = O(log w), its counting time is O(|P |) and its construction time is O(n).
If we admit randomized linear-time constructions, then Belazzougui and Navarro [6,
Thm. 10] reach O(|P |(1 + log logw σ)) counting time. At the expense of O(n) further bits,
in another work [5] they reach O(|P |) counting time. Using O(n log σ) bits, and if P comes
in packed form, Grossi and Vitter [19] can count in time O(|P |/ logσ n + logσ n), for any
constant > 0, however their construction requires O(n log σ) time.
Table 1 puts those results and our contribution in context. Our new self-index, with
O(|P | + log logw σ) counting time, linear-time deterministic construction, and nHk (T ) +
o(n log σ) bits of space, dominates all the compressed indexes with linear-time deterministic
construction [1, 6], as well as some uncompressed ones [15] (to be fair, we do not cover the
case log σ = O(log w), as in this case the previous work [6, Thm. 7] already obtains our result).
Our self-index also dominates a previous one with linear-time randomized construction [6,
Thm. 10], which we incidentally show can also be built deterministically. The only aspect
in which some of those dominated indexes may outperform ours is in that they may use
o(n(Hk (T ) + 1)) [6, Thm. 10] or o(n) [6, Thm. 7] bits of redundancy, instead of our o(n log σ)
bits. We also derive a compact index (i.e., using O(n log σ) bits) that is built in linear
deterministic time and counts in time O(|P |/ logσ n + log n(log log n)2 ), which is the only
one in this category unless we consider constant σ for Grossi and Vitter [19].
3
Preliminaries
We denote by T [i..] the suffix of T [0, n − 1] starting at position i and by T [i..j] the substring
that begins with T [i] and ends with T [j], T [i..] = T [i]T [i + 1] . . . T [n − 1] and T [i..j] =
T [i]T [i + 1] . . . T [j − 1]T [j]. We assume that the text T ends with a special symbol $ that
lexicographically precedes all other symbols in T . The alphabet size is σ and symbols
are integers in [0..σ − 1] (so $ corresponds to 0). In this paper, as in the previous work
on this topic, we use the word RAM model of computation. A machine word consists of
w = Ω(log n) bits and we can execute standard bit and arithmetic operations in constant
time. We assume for simplicity that the alphabet size σ = O(n/ log n) (otherwise the text is
almost incompressible anyway [16]). We also assume log σ = ω(log w), since otherwise our
goal is already reached in previous work [6, Thm. 7].
3.1
Rank and Select Queries
We define three basic queries on sequences. Let B[0..n − 1] be a sequence of symbols over
alphabet [0..σ − 1]. The rank query, ranka (i, B), counts how many times a occurs among the
first i + 1 symbols in B, ranka (i, B) = |{ j ≤ i, B[j] = a}|. The select query, selecta (i, B),
finds the position in B where a occurs for the i-th time, selecta (i, B) = j iff B[j] = a and
ranka (j, B) = i. The third query is access(i, B), which returns simply B[i].
© J. Ian Munro, Gonzalo Navarro, Yakov Nekrich;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
We can answer access queries in O(1) time and select queries in any ω(1) time, or vice
versa, and rank queries in time O(log logw σ), which is optimal [6]. These structures use
n log σ + o(n log σ) bits, and we will use variants that require only compressed space. In this
paper, we will show that those structures can be built in linear deterministic time.
An important special case of rank queries is the partial rank query, rankB[i] (i, B), which
asks how many times B[i] occurrs in B[0..i]. Unlike general rank queries, partial rank queries
can be answered in O(1) time [6]. Such a structure can be built in O(n) deterministic time
and requires O(n log log σ) bits of working and final space [28, Thm. A.4.1].
For this paper, we define a generalization of partial rank queries called interval rank
queries, ranka (i, j, B) = hranka (i − 1, B), ranka (j, B)i, from where in particular we can
deduce the number of times a occurs in B[i..j]. If a does not occur in B[i..j], however, this
query just returns null (this is why it can be regarded as a generalized partial rank query).
In the special case where the alphabet size is small, log σ = O(log w), we can represent B
so that rank, select, and access queries are answered in O(1) time [6, Thm. 7], but we are
not focusing on this case in this paper, as the problem has already been solved for this case.
3.2
Suffix Array and Suffix Tree
The suffix tree [35] for a string T [0..n − 1] is a compacted digital tree on the suffixes of T ,
where the leaves point to the starting positions of the suffixes. We call Xu the string leading
to suffix tree node u. The suffix array [24] is an array SA[0..n − 1] such that SA[i] = j if and
only if T [j..] is the (i + 1)-th lexicographically smallest suffix of T . All the occurrences of a
substring P in T correspond to suffixes of T that start with P . These suffixes descend from
a single suffix tree node, called the locus of P , and also occupy a contiguous interval in the
suffix array SA. Note that the locus of P is the node u closest to the root for which P is a
prefix of Xu . If P has no locus node, then it does not occur in T .
3.3
Compressed Suffix Array and Tree
A compressed suffix array (CSA) is a compact data structure that provides the same
functionality as the suffix array. The main component of a CSA is the one that allows
determining, given a pattern P , the suffix array range SA[i..j] of the prefixes starting with
P . Counting is then solved as j − i + 1. For locating any cell SA[k], and for extracting any
substring S from T , most CSAs make use of a sampled array SAMb , which contains the
values of SA[i] such that SA[i] mod b = 0 or SA[i] = n − 1. Here b is a tradeoff parameter:
CSAs require O(n logb n ) further bits and can locate in time proportional to b and extract S
in time proportional to b + |S|. We refer to a survey [30] for a more detailed description.
A compressed suffix tree [33] is formed by a compressed suffix array and other components
that add up to O(n) bits. These include in particular a representation of the tree topology
that supports constant-time computation of the preorder of a node, its number of children,
its j-th child, its number of descendant leaves, and lowest common ancestors, among others
[32]. Computing node preorders is useful to associate satellite information to the nodes.
Both the compressed suffix array and tree can be built in O(n) deterministic time using
O(n log σ) bits of space [28].
3.4
Burrows-Wheeler Transform and FM-index
The Burrows-Wheeler Transform (BWT) [8] of a string T [0..n−1] is another string B[0..n−1]
obtained by sorting all possible rotations of T and writing the last symbol of every rotation
© J. Ian Munro, Gonzalo Navarro, Yakov Nekrich;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
(in sorted order). The BWT is related to the suffix array by the identity B[i] = T [(SA[i] − 1)
mod n]. Hence, we can build the BWT by sorting the suffixes and writing the symbols that
precede the suffixes in lexicographical order.
The FM-index [12, 13] is a CSA that builds on the BWT. It consists of the following
three main components:
The BWT B of T .
The array Acc[0..σ − 1] where Acc[i] holds the total number of symbols a < i in T (or
equivalently, the total number of symbols a < i in B).
The sampled array SAMb .
The interval of a pattern string P [0..m − 1] in the suffix array SA can be computed on
the BWT B. The interval is computed backwards: for i = m − 1, m − 2, . . ., we identify
the interval of P [i..m − 1] in B. The interval is initially the whole B[0..n − 1]. Suppose
that we know the interval B[i1 ..j1 ] that corresponds to P [i + 1..m − 1]. Then the interval
B[i2 ..j2 ] that corresponds to P [i..m − 1] is computed as i2 = Acc[a] + rankc (i1 − 1, B) and
j2 = Acc[a] + rankc (j1 , B) − 1, where a = P [i]. Thus the interval of P is found by answering
2m rank queries. Any sequence representation offering rank and access queries can then be
applied on B to obtain an FM-index.
An important procedure on the FM-index is the computation of the function LF , defined
as follows: if SA[j] = i + 1, then SA[LF (j)] = i. LF can be computed with access and
partial rank queries on B, LF (j) = rankB[j] (i, B) + Acc[B[j]] − 1, and thus constant-time
computation of LF is possible. Using SAMb and O(b) applications of LF , we can locate
any cell SA[r]. A similar procedure allows extracting any substring S of T with O(b + |S|)
applications of LF .
4
Small Interval Rank Queries
We start by showing how a compressed data structure that supports select queries can be
extended to support a new kind of queries that we dub small interval rank queries. An
interval query ranka (i, j, B) is a small interval rank query if j − i ≤ log2 σ. Our compressed
index relies on the following result.
I Lemma 2. Suppose that we are given a data structure that supports access queries on a
sequence C[0..m − 1], on alphabet [0..σ − 1], in time t. Then, using O(m log log σ) additional
bits, we can support small interval rank queries on C in O(t) time.
Proof. We split C into groups Gi of log2 σ consecutive symbols, Gi = C[i log2 σ..(i +
1) log2 σ − 1]. Let Ai denote the sequence of the distinct symbols that occur in Gi . Storing
Ai directly would need log σ bits per symbol. Instead, we encode each element of Ai as its
first position in Gi , which needs only O(log log σ) bits. With this encoded sequence, since we
have O(t)-time access to C, we have access to any element of Ai in time O(t). In addition,
we store a succinct SB-tree [18] on the elements of Ai . This structure uses O(p log log u) bits
to index p elements in [1..u], and supports predecessor (and membership) queries in time
O(log p/ log log u) plus one access to Ai . Since u = σ and p ≤ log2 σ, the query time is O(t)
and the space usage is bounded by O(m log log σ) bits.
For each a ∈ Ai we also keep the increasing list Ia,i of all the positions where a occurs
in Gi . Positions are stored as differences with the left border of Gi : if C[j] = a, we store
the difference j − i log2 σ. Hence elements of Ia,i can also be stored in O(log log σ) bits per
symbol, adding up to O(m log log σ) bits. We also build an SB-tree on top of each Ia,i to
provide for predecessor searches.
© J. Ian Munro, Gonzalo Navarro, Yakov Nekrich;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
Using the SB-trees on Ai and Ia,i , we can answer small interval rank queries ranka (x, y, C).
Consider a group Gi = C[i log2 σ..(i + 1) log2 σ − 1], an index k such that i log2 σ ≤ k ≤
(i + 1) log2 σ, and a symbol a. We can find the largest i log2 σ ≤ r ≤ k such that C[r] = a,
or determine it does not exist: First we look for the symbol a in Ai ; if a ∈ Ai , we find the
predecessor of k − i log2 σ in Ia,i .
Now consider an interval C[x..y] of size at most log2 σ. It intersects at most two groups,
Gi and Gi−1 . We find the rightmost occurrence of symbol a in C[x..y] as follows. First
we look for the rightmost occurrence y 0 ≤ y of a in Gi ; if a does not occur in C[i log2 σ..y],
we look for the rightmost occurrence y 0 ≤ i log2 σ − 1 of a in Gi−1 . If this is ≥ x, we find
the leftmost occurrence x0 of a in C[x..y] using a symmetric procedure. When x0 ≤ y 0 are
found, we can compute ranka (x0 , C) and ranka (y 0 , C) in O(1) time by answering partial rank
queries (Section 3.1). These are supported in O(1) time and O(m log log σ) bits. The answer
is then hranka (x0 , C) − 1, ranka (y 0 , C)i, or null if a does not occur in C[x..y].
J
The construction of the small interval rank data structure is dominated by the time
√
needed to build the succinct SB-trees [18]. These are simply B-trees with arity O( log u)
and height O(log p/ log log u), where in each node a Patricia tree for O(log log u)-bit chunks
of the keys are stored. To build the structure in O(log p/ log log u) time per key, we only
need to build those Patricia trees in linear time. Given that the total number of bits of all
√
the keys to insert in a Patricia tree is O( log u log log u), we do not even need to build the
Patricia tree. Instead, a universal precomputed table may answer any Patricia tree search
for any possible
set of keys and any possible pattern, in constant time. The size of the table
√
√
is O(2O( log u log log u) log u) = o(u) bits (the authors [18] actually use a similar table to
answer queries). For our values of p and u, the construction requires O(mt) time and the
universal table is of o(σ) bits.
5
Compressed Index
We classify the nodes of the suffix tree T of T into heavy, light, and special, as in previous
work [31, 28]. Let d = log σ. A node u of T is heavy if it has at least d leaf descendants and
light otherwise. We say that a heavy node u is special if it has at least two heavy children.
For every special node u, we construct a deterministic dictionary [20] Du that contains
the labels of all the heavy children of u: If the jth child of u, uj , is heavy and the first
symbol on the edge from to u to uj is aj , then we store the key aj in Du with j as satellite
data. If a heavy node u has only one heavy child uj and d or more light children, then we
also store the data structure Du (containing only that heavy child of u). If, instead, a heavy
node has one heavy child and less than d light children, we just keep the index of the heavy
child using O(log d) = O(log log σ) bits.
The second component of our index is the Burrows-Wheeler Transform B of the reverse
text T . We store a data structure that supports rank, partial rank, select, and access queries
on B. It is sufficient for us to support access and partial rank queries in O(1) time and rank
queries in O(log logw σ) time. We also construct the data structure described in Lemma 2,
which supports small interval rank queries in O(1) time. Finally, we explicitly store the
answers to some rank queries. Let B[lu ..ru ] denote the range of Xu , where Xu is the reverse
of Xu , for a suffix tree node u. For all data structures Du and for every symbol a ∈ Du we
store the values of ranka (lu − 1, B) and ranka (ru , B).
Let us show how to store the selected precomputed answers to rank queries in O(log σ)
bits per query. Following a known scheme [17], we divide the sequence B into chunks of
size σ. For each symbol a, we encode the number dk of times a occurs in each chunk k
© J. Ian Munro, Gonzalo Navarro, Yakov Nekrich;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
in a binary sequence Aa = 01d0 01d1 01d2 . . .. If a symbol B[i] belongs to chunk k = bi/σc,
then ranka (i, B) is select0 (k + 1, Aa ) − k plus the number of times a occurs in B[kσ..i]. The
former value is computed in O(1) time with a structure that uses |Aa | + o(|Aa |) bits [9, 27],
whereas the latter value is in [0, σ] and thus can be stored in Du using just O(log σ) bits.
The total size of all the sequences Aa is O(n) bits.
Therefore, Du needs O(log σ) bits per element. The total number of elements in all the
structures Du is equal to the number of special nodes plus the number of heavy nodes with
one heavy child and at least d light children. Hence all Du contain O(n/d) symbols and use
O((n/d) log σ) = O(n) bits of space. Indexes of heavy children for nodes with only one heavy
child and less than d light children add up to O(n log log σ) bits. The structures for partial
rank and small interval rank queries on B use O(n log log σ) further bits. Since we assume
that σ is ω(1), we can simplify O(n log log σ) = o(n log σ).
The sequence representation that supports access and rank queries on B can be made
to use nHk (T ) + o(n(Hk (T ) + 1)) bits, by exploiting the fact that it is built on a BWT [6,
Thm. 10].1 We note that they use constant-time select queries on B instead of constant-time
access, so they can use select queries to perform LF −1 -steps in constant time. Instead, with
our partial rank queries, we can perform LF -steps in constant time (recall Section 3.4), and
thus have constant-time access instead of constant-time select on B (we actually do not use
query select at all). They avoid this solution because partial rank queries require o(n log σ)
bits, which can be more than o(n(Hk (T ) + 1)), but we are already paying this price.
Apart from this space, array Acc needs O(σ log n) = O(n) bits and SAMb uses O(n logb n ).
The total space usage of our self-index then adds up to nHk (T ) + o(n log σ) + O(n logb n ) bits.
6
Pattern Search
Given a query string P , we will find in time O(|P | + log logw σ) the range of the reversed
string P in B. A backward search for P in B will be replaced by an analogous backward
search for P in B, that is, we will find the range of P [0..i] if the range of P [0..i − 1] is
known. Let [li ..ri ] be the range of P [0..i]. We can compute li and ri from li−1 and ri−1 as
li = Acc[a] + ranka (li−1 − 1, B) and ri = Acc[a] + ranka (ri−1 , B) − 1, for a = P [i]. Using
our auxiliary data structures on B and the additional information stored in the nodes of the
suffix tree T , we can answer the necessary rank queries in constant time (with one exception).
The idea is to traverse the suffix tree T in synchronization with the forward search on B,
until the locus of P is found or we determine that P does not occur in T .
Our procedure starts at the root node of T , with l−1 = 0, r−1 = n − 1, and i = 0. We
compute the ranges B[li ..ri ] that correspond to P [0..i] for i = 0, . . . , |P | − 1. Simultaneously,
we move down in the suffix tree. Let u denote the last visited node of T and let a = P [i].
We denote by ua the next node that we must visit in the suffix tree, i.e., ua is the locus of
P [0..i]. We can compute li and ri in O(1) time if ranka (ri−1 , B) and ranka (li−1 − 1, B) are
known. We will show below that these queries can be answered in constant time because
either (a) the answers to rank queries are explicitly stored in Du or (b) the rank query that
must be answered is a small interval rank query. The only exception is the situation when
we move from a heavy node to a light node in the suffix tree; in this case the rank query
takes O(log logw σ) time. We note that, once we are in a light node, we need not descend in
T anymore; it is sufficient to maintain the interval in B.
For ease of description we distinguish between the following cases.
1
In fact it is nHk (T ), but this is nHk (T ) + O(log n) [12, Thm. A.3].
© J. Ian Munro, Gonzalo Navarro, Yakov Nekrich;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
1. Node u is heavy and a ∈ Du . In this case we identify the heavy child ua of u that is
labeled with a in constant time using the deterministic dictionary. We can also find li
and ri in time O(1) because ranka (li−1 − 1, B) and ranka (ri−1 , B) are stored in Du .
2. Node u is heavy and a 6∈ Du . In this case ua , if it exists, is a light node. We then find it
with two standard rank queries on B, in order to compute li and ri or determine that P
does not occur in T .
3. Node u is heavy but we do not keep a dictionary Du for the node u. In this case u has at
most one heavy child and less than d light children. We have two subcases:
a. If ua is the (only) heavy node, we find this out with a single comparison, as the heavy
node is identified in u. However, the values ranka (li−1 − 1, B) and ranka (ri−1 , B) are
not stored in u. To compute them, we exploit the fact that the number of non-a’s in
B[li−1 ..ri−1 ] is less than d2 , as all the children apart from ua are light and less than d.
Therefore, the first and the last occurrences of a in B[li−1 ..ri−1 ] must be at distance
less than d2 from the extremes li−1 and ri−1 , respectively. Therefore, a small interval
rank query, ranka (li−1 , li−1 + d2 , B), gives us ranka (li−1 − 1, B), since there is for sure
an a in the range. Analogously, ranka (ri−1 − d2 , ri−1 , B) gives us ranka (ri−1 , B).
b. If ua is a light node, we compute li and ri with two standard rank queries on B (or
we might determine that P does not appear in T ).
4. Node u is light. In this case, P [0..i − 1] occurs at most d times in T . Hence P [0..i − 1] also
occurs at most d times in T and ri−1 − li−1 ≤ d. Therefore we can compute ri and li in
O(1) time by answering a small interval rank query, hranka (li−1 − 1, B), ranka (ri−1 , B)i.
If this returns null, then P does not occur in T .
5. We are on an edge of the suffix tree between a node u and some child uj of u. In this
case all the occurrences of P [0..i − 1] in T are followed by the same symbol, c, and all
the occurrences of P [0..i − 1] are preceded by c in T . Therefore B[li−1 ..ri−1 ] contains
only the symbol c. This situation can be verified with access and partial rank queries on
B: B[ri−1 ] = B[li−1 ] = c and rankc (ri−1 , B) − rankc (li−1 , B) = ri−1 − li−1 . In this case,
if a 6= c, then P does not occur in T ; otherwise we obtain the new range with the partial
rank query rankc (ri−1 , B), and rankc (li−1 − 1, B) = rankc (ri−1 , B) − (ri−1 − li−1 + 1).
Note that if u is light we do not need to consider this case; we may directly apply case 4.
Except for the cases 2 and 3b, we can find li and ri in O(1) time. In cases 2 and 3b we
need O(log logw σ) time to answer general rank queries. However, these cases only take place
when the node u is heavy and its child ua is light. Since all descendants of a light node are
light, those cases occur only once along the traversal of P . Hence the total time to find the
range of P in B is O(|P | + log logw σ). Once the range is known, we can count and report
all occurrences of P in the standard way.
7
7.1
Linear-Time Construction
Sequences and Related Structures
Apart from constructing the BWT B of T , which is a component of the final structure, the
linear-time construction of the other components requires that we also build, as intermediate
structures, the BWT B of T , and the compressed suffix trees T and T of T and T , respectively.
All these are built in O(n) deterministic time and using O(n log σ) bits of space [28]. We
also keep, on top of both B and B, O(n log log σ)-bit data structures able to report, for any
interval B[i..j] or B[i..j], all the distinct symbols from this interval, and their frequencies in
© J. Ian Munro, Gonzalo Navarro, Yakov Nekrich;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
the interval. The symbols are retrieved in arbitrary order. These auxiliary data structures
can also be constructed in O(n) time [28, Sec. A.5].
On top of the sequences B and B, we build the representation that supports access in
O(1) and rank in O(log logw σ) time [6]. In their original paper, those structures are built
using perfect hashing, but a deterministic construction is also possible [4, Lem. 11]; we give
the details next.
The key part of the construction is that, within a chunk of σ symbols, we must build a
virtual list Ia of the positions where each symbol a occurs, and provide predecessor search on
those lists in O(log logw σ) time. We divide each list into blocks of log2 σ elements, and create
a succinct SB-tree [18] on the block elements, much as in Section 4. The search time inside a
block is then O(t), where t is the time to access an element in Ia , and the total extra space is
O(n log log σ) bits. If there is more than one block in Ia , then the block minima are inserted
into a predecessor structure [6, App. A] that will find the closest preceding block minimum
in time O(log logw σ) and use O(n log log σ) bits. This structure uses perfect hash functions
called I(P ), which provide constant-time membership queries. Instead, we replace them with
deterministic dictionaries [20]. The only disadvantage of these dictionaries is that they require
O(log σ) construction time per element, and since each element is inserted into O(log logw σ)
structures I(P ), the total construction time per element is O(log σ log logw σ). However,
since we build these structures only on O(n/ log2 σ) block minima, the total construction
time is only O(n).
On the variant of the structure that provides constant-time access, the access to an element
in Ia is provided via a permutation structure [29] which offers access time t with extra space
O((n/t) log σ) bits. Therefore, for any log σ = ω(log w), we can have t = O(log logw σ) with
o(n log σ) bits of space.
7.2
Structures Du
The most complex part of the construction is to fill the data of the Du structures. We visit
all the nodes of T and identify those nodes u for which the data structure Du must be
constructed. This can be easily done in linear time, by using the constant-time computation
of the number of descendant leaves. To determine if we must build Du , we traverse its
children u1 , u2 , . . . and count their descendant leaves to decide if they are heavy or light.
We use a bit vector D to mark the preorders of the nodes u for which Du will be
constructed: If p is the preorder of node u, then it stores a structure Du iff D[p] = 1, in
which case Du is stored in an array at position rank1 (D, p). If, instead, u does not store Du
but it has one heavy child, we store its child rank in another array indexed by rank0 (D, p),
using log log σ bits per cell.
The main difficulty is how to compute the symbols a to be stored in Du , and the ranges
B[lu , ru ], for all the selected nodes u. It is not easy to do this through a preorder traversal of T
because we would need to traverse edges that represent many symbols. Our approach, instead,
is inspired by the navigation of the suffix-link tree using two BWTs given by Belazzougui
et al. [3]. Let Tw denote the tree whose edges correspond to Weiner links between internal
nodes in T . That is, the root of Tw is the same root of T and, if we have internal nodes
u, v ∈ T where Xv = a · Xu for some symbol a, then v descends from u by the symbol a in
Tw . We first show that the nodes of Tw are the internal nodes of T . The inclusion is clear by
definition in one direction; the other is well-known but we prove it for completeness.
I Lemma 3. All internal nodes of the suffix tree T are nodes of Tw .
Proof. We proceed by induction on |Xu |, where the base case holds by definition. Now let a
© J. Ian Munro, Gonzalo Navarro, Yakov Nekrich;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
non-root internal node u of T be labeled by string Xu = aX. This means that there are at
least two different symbols a1 and a2 such that both aXa1 and aXa2 occur in the text T .
Then both Xa1 and Xa2 also occur in T . Hence there is an internal node u0 with Xu0 = X
in T and a Weiner link from u0 to u. Since |Xu0 | = |Xu | − 1, it holds by the inductive
hypothesis that u0 belongs to Tw , and thus u belongs to Tw as a child of u0 .
J
We do not build Tw explicitly, but just traverse its nodes conceptually in depth-first order
and compute the symbols to store in the structures Du and the intervals in B. Let u be
the current node of T in this traversal and u its corresponding locus in T . Assume for now
that u is a nod, too. Let [lu , ru ] be the interval of Xu in B and [lu , ru ] be the interval of the
reverse string Xu in B.2 Our algorithm starts at the root nodes of Tw , T , and T , which
correspond to the empty string, and the intervals in B and B are [lu , ru ] = [lu , ru ] = [0, n − 1].
We will traverse only the heavy nodes, yet in some cases we will have to work on all the
nodes. We ensure that on heavy nodes we work at most O(log σ) time, and at most O(1)
time on arbitrary nodes.
Upon arriving at each node u, we first compute its heavy children. From the topology of
T we identify the interval [li , ri ] for every child ui of u, by counting leaves in the subtrees
of the successive children of u. By reporting all the distinct symbols in B[lu ..ru ] with their
frequencies, we identify the labels of those children. However, the labels are retrieved in
arbitrary order and we cannot afford sorting them all. Yet, since the labels are associated
with their frequencies in B[lu ..ru ], which match their number of leaves in the subtrees of u,
we can discard the labels of the light children, that is, those appearing less than d times in
B[lu ..ru ]. The remaining, heavy, children are then sorted and associated with the successive
heavy children ui of u in T .
If our preliminary pass marked that a Du structure must be built, we construct at this
moment the deterministic dictionary [20] with the labels a of the heavy children of u we have
just identified, and associate them with the satellite data ranka (lu − 1, B) and ranka (ru , B).
This construction takes O(log σ) time per element, but it includes only heavy nodes.
We now find all the Weiner links from u. For every (heavy or light) child ui of u, we
compute the list Li of all the distinct symbols that occur in B[li ..ri ]. We mark those symbols
a in an array V [0..σ − 1] that holds three possible values: not seen, seen, and seen (at least)
twice. If V [a] is not seen, then we mark it as seen; if it is seen, we mark it as seen twice;
otherwise we leave it as seen twice. We collect a list Eu of the symbols that are seen twice
along this process, in arbitrary order. For every symbol a in Eu , there is an explicit Weiner
link from u labeled by a: Let X = Xu ; if a occurred in Li and Lj then both aXai and aXaj
occur in T and there is a suffix tree node that corresponds to the string aX. The total time
to build Eu amortizes to O(n): for each child v of u, we pay O(1) time for each child the
node v has in T ; each node in T contributes once to the cost.
The targets of the Weiner links from u in T correspond to the children of the node u in
T . To find them, we collect all the distinct symbols in B[lu ..ru ] and their frequencies. Again,
we discard the symbols with frequency less than d, as they will lead to light nodes, which
we do not have to traverse. The others are sorted and associated with the successive heavy
children of u. By counting leaves in the successive children, we obtain the intervals B[li0 ..ri0 ]
corresponding to the heavy children u0i of u.
We are now ready to continue the traversal of Tw : for each Weiner link from u by
2
In the rest of the paper we wrote B[lu ..ru ] instead of B[lu ..ru ] for simplicity, but this may cause
confusion in this section.
© J. Ian Munro, Gonzalo Navarro, Yakov Nekrich;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
symbol a leading to a heavy node, which turns out to be the i-th child of u, we know
that its node in T is u0i (computed from u using the tree topology) and its interval is
B[li0 ..ri0 ]. To compute the corresponding interval on B, we use the backward step operation,
B[x, y] = B[Acc[a] + ranka (lu − 1, B), Acc[a] + ranka (ru , B) − 1]. This requires O(log logw σ)
time, but applies only to heavy nodes. Finally, the corresponding node in T is obtained in
constant time as the lowest common ancestor of the x-th and the y-th leaves of T .
In the description above we assumed for simplicity that u is a node in T . In the general
case u can be located on an edge of T . This situation arises when all occurrences of Xu
in the reverse text T are followed by the same symbol a. In this case there is at most one
Weiner link from u; the interval in B does not change as we follow that link.
A recursive traversal of Tw might require O(nσ log n) bits for the stack, because we store
several integers associated to heavy children during the computation of each node u. We
can limit the stack height by determining the largest subtree among the Weiner links of
u, traversing all the others recursively, and then moving to that largest Weiner link target
without recursion [3, Lem. 1]. Since only the largest subtree of a Weiner link target can
contain more than half of the nodes of the subtree of u, the stack is guaranteed to be of
height only O(log n). The space usage is thus O(σ log2 n) = O(n log σ).
As promised, we have spent at most O(log σ) time on heavy nodes, which are O(n/d) =
O(n/ log σ) in total, thus these costs add up to O(n). All other costs that apply to arbitrary
nodes are O(1). The structures for partial rank queries (and the succinct SB-trees) can also
be built in linear deterministic time, as shown in Section 4. Therefore our index can be
constructed in O(n) time.
8
A Compact Index
As an application of our techniques, we show that it is possible to obtain O(|P |/ logσ n+log2 n)
search time, and even O(|P |/ logσ n + log n(log log n)2 ), with an index that uses O(n log σ)
bits and is built in linear deterministic time.
We store B in compressed form and a sample of the heavy nodes of T . Following previous
work [19, 31], we start from the root and store a deterministic dictionary [20] with all the
highest suffix tree nodes v representing strings of depth ≥ ` = logσ n. The key associated
with each node is a log(n)-bit integer formed with the first ` symbols of the strings Pv . The
satellite data are the length |Pv |, a position where Pv occurs in T , and the range B[lv ..rv ] of
v. From each of those nodes v, we repeat the process with the first ` symbols that follow
after Pv , and so on. The difference is that no light node will be inserted in those dictionaries.
Let us charge the O(log n) bits of space to the children nodes, which are all heavy. If we
count only the special nodes, which are O(n/d), this amounts to O((n log n)/d) total bits
and construction time. Recall that d is the maximum subtree size of light nodes. This time
will use d = Θ(log n) to have linear construction time and bit space, and thus will not take
advantage of small rank interval queries.
There are, however, heavy nodes that are not special. These form possibly long chains
between special nodes, and these will also induce chains of sampled nodes. While building
the dictionaries for those nodes is trivial because they have only one sampled child, the total
space may add up to O(n log σ) bits, if there are Θ(n) heavy nodes and the sampling chooses
one out of ` in the chains. To avoid this, we increase the sampling step in those chains,
enlarging it to `0 = log n. This makes the extra space spent in sampling heavy non-special
nodes to be O(n) bits as well.
In addition, we store the text T with a data structure that uses nHk (T ) + o(n log σ) for
© J. Ian Munro, Gonzalo Navarro, Yakov Nekrich;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
any k = o(logσ n), and allows us extract O(logσ n) consecutive symbols in constant time [14].
The search for P starts at the root, where its first ` symbols are used to find directly the
right descendant node v in the dictionary stored at the root. If |Pv | > `, we directly compare
the other |Pv | − ` symbols of P with the text, from the stored position where v appears, by
chunks of ` symbols. Then we continue the search from v (ignore the chains for now).
When the next ` symbols of P are not found in the dictionary of the current node v,
or there are less than ` remaining symbols in P , we continue using backward search on B
from the interval B[lv ..rv ] stored at v, and do not use the suffix tree topology anymore. If
there are less than ` remaining symbols in P , we just proceed symbolwise and complete the
search in O(` log logw σ) time. Before this point, we still proceed by chunks of ` symbols,
except when (conceptually) traversing the explicit light suffix tree nodes, where we perform
individual backward steps. Because there are only O(d) suffix tree nodes in the remaining
subtree, the number of backward steps to traverse such light nodes is O(d) and can be
performed in time O(d log logw σ). All the symbols between consecutive light nodes must
be traversed in chunks of `. Let v be our current node and u its desired child, and let
k + 1 be the number of symbols labeling the edge between them. After the first backward
step from v to u leads us from the interval B[lv ..rv ] to B[l..r], we must perform k further
backward steps, where each intermediate interval is formed by just one symbol. To traverse
them fast, we first compute tl = SA[l] and tr = SA[r]. The desired symbols are then
T [tl − 1] = T [tr − 1], . . . , T [tl − k] = T [tr − k]. We thus compare the suffixes T [n − tl ..] and
T [n − tr ..] with what remains of P , by chunks of ` symbols, until finding the first difference;
k is then the number of coincident symbols seen. If the two suffixes coincide up to k + 1
but they differ from P , then P is not in T . Otherwise, we can move (conceptually) to
node u and consume the k coincident characters from P . We can compute the new interval
[lu ..ru ] = [SA −1 [tl − k], SA −1 [tr − k]]. Since SA and its inverse are computed in O(b) time
with arrays similar to SAMb , we spend O(b) time to cross each of the O(d) edges in the final
part of the search.
Let us now regard the case where we reach a sampled node v that starts a sampled chain.
In this case the sampling step grows to `0 . We still compare P with a suffix of T where its
heavy sampled child u appears, in chunks of ` symbols. If they coincide, we continue the
search from u. Otherwise, we resume the search from B[lv ..rv ] using backward search. We
might have to process `0 symbols of P , in time O(`0 log logw σ), before reaching a light node.
Then, we proceed as explained.
Overall, the total space is 2nHk (T ) + o(n log σ) + O(n) + O((n log n)/b) bits, and the
construction time is O(n). To have o(n log σ) + O(n) bits of redundancy in total, we may
choose b = Θ(log n) or any b = ω(logσ n). The search time is O(|P |/ logσ n + `0 log logw σ +
d log logw σ + db). We can, for example, choose b = Θ(log n), to obtain the following result.
I Theorem 4. On a RAM machine of w = Ω(log n) bits, we can construct an index for
a text T of length n over an alphabet of size σ = O(n/ log n) in O(n) deterministic time
using O(n log σ) bits of working space. This index occupies 2nHk (T ) + o(n log σ) + O(n) bits
of space for any k = o(logσ n). The occurrences of a pattern string P can be counted in
O(|P |/ logσ n + log2 n) time, and then each such occurrence can be located in O(log n) time.
An arbitrary substring S of T can be extracted in time O(|S|/ logσ n).
8.1
Faster and Larger
By storing the BWT B of T and the O(n) additional bits to support a compressed suffix
tree T [33], we can reduce the O(log2 n) extra time to O(log n(log log n)2 ).
© J. Ian Munro, Gonzalo Navarro, Yakov Nekrich;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
The idea is to speed up the traversal on the light nodes, as follows. For each light node v,
we store the leaf in T where the heavy path starting at v ends. The heavy path chooses at
each node the subtree with the most leaves, thus any traversal towards a leaf has to switch to
another heavy path only O(log d) times. At each light node v, we go to the leaf u of its heavy
path, obtain its position in T using the sampled array SAMb of B, and compare the rest
of P with the corresponding part of the suffix, by chunks of ` symbols. Once we determine
the number k of symbols that coincide with P in the path from v to u, we perform a binary
search for the highest ancestor v 0 of u where |Pv0 − Pv | ≥ k. If |Pv0 − Pv | > k, then P does
not appear in T (unless P ends at the k-th character compared, in which case the locus of P
is v 0 ). Otherwise, we continue the search from v 0 . Each binary search step requires O(b) time
to determine |Px | [33], so it takes time O(b log d). Since we switch to another heavy path
(now the one starting at v 0 ) O(log d) times, the total time is O(b log2 d) = O(log n(log log n)2 )
instead of O(bd) = O(log2 n).
To store the leaf u corresponding to each light node v, we record the difference between
the preorder numbers of u and v, which requires O(log d) bits. The node u is easily found
in constant time from this information [33]. We have the problem, however, that we spend
O(log d) = O(log log n) bits per light node, which adds up to O(n log log n) bits. To reduce
this to O(n), we choose a second sampling step e = O(log log n), and do not store this
information on nodes with less than e leaves, which are called light-light. Those light
nodes with e leaves or more are called light-heavy, and those with at least two light-heavy
children are called light-special. There are O(n/e) light-special nodes. We store heavy path
information only for light-special nodes or for light-heavy nodes that are children of heavy
nodes; both are O(n/e) in total. A light-heavy node v that is not light-special has at most
one light-heavy child u, and the heavy path that passes through v must continue towards u.
Therefore, if it turns out that the search must continue from v after the binary search on the
heavy path, then the search must continue towards the light-light children of v, therefore no
heavy-path information is needed at node v.
Once we reach the first light-light node v, we proceed as we did for Theorem 4 on light
nodes, in total time O(eb) = O(log n log log n). We need, however, the interval B[lv , rv ]
before we can start the search from v. The interval B[lv , rv ] is indeed known by counting
leaves in T , and we can compute lv = SA −1 [n − 1 − SA[lv ]] and rv = SA −1 [n − 1 − SA[rv ]],
in time O(b).
I Theorem 5. On a RAM machine of w = Ω(log n) bits, we can construct an index for
a text T of length n over an alphabet of size σ = O(n/ log n) in O(n) deterministic time
using O(n log σ) bits of working space. This index occupies 3nHk (T ) + o(n log σ) + O(n)
bits of space for any k = o(logσ n). The occurrences of a pattern string P can be counted
in O(|P |/ logσ n + log n(log log n)2 ) time, and then each such occurrence can be located in
O(log n) time. An arbitrary substring S of T can be extracted in time O(|S|/ logσ n).
9
Conclusions
We have shown how to build, in O(n) deterministic time and using O(n log σ) bits of working
space, a compressed self-index for a text T of length n over an alphabet of size σ that searches
for patterns P in time O(|P | + log logw σ), on a w-bit word RAM machine. This improves
upon previous compressed self-indexes requiring O(|P | log log σ) [1] or O(|P |(1 + logw σ))
[6] time, on previous uncompressed indexes requiring O(|P | + log log σ) time [15] (but that
supports dynamism), and on previous compressed self-indexes requiring O(|P |(1 + log logw σ))
time and randomized construction (which we now showed how to build in linear deterministic
© J. Ian Munro, Gonzalo Navarro, Yakov Nekrich;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
time) [6]. The only indexes offering better search time require randomized construction
[5, 19, 31] or Θ(n log n) bits of space [31, 7].
As an application, we showed that using O(n log σ) bits of space, we can build in O(n)
deterministic time an index that searches in time O(|P |/ logσ n + log n(log log n)2 ).
It is not clear if O(|P |) time, or even O(|P |/ logσ n), query time can be achieved with a
linear deterministic construction time, even if we allow O(n log n) bits of space for the index
(this was recently approached, but some additive polylog factors remain [7]). This is the
most interesting open problem for future research.
References
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
J. Barbay, F. Claude, T. Gagie, G. Navarro, and Y. Nekrich. Efficient fully-compressed
sequence representations. Algorithmica, 69(1):232–268, 2014.
D. Belazzougui, P. Boldi, R. Pagh, and S. Vigna. Fast prefix search in little space, with
applications. In Proc. 18th Annual European Symposium on Algorithms (ESA), LNCS 6346,
pages 427–438, 2010.
D. Belazzougui, F. Cunial, J. Kärkkäinen, and V. Mäkinen. Versatile succinct representations of the bidirectional Burrows-Wheeler transform. In Proc. 21st Annual European
Symposium on Algorithms (ESA), pages 133–144, 2013.
D. Belazzougui, F. Cunial, J. Kärkkäinen, and V. Mäkinen. Linear-time string indexing
and analysis in small space. CoRR, abs/1609.06378, 2016.
D. Belazzougui and G. Navarro. Alphabet-independent compressed text indexing. ACM
Transactions on Algorithms, 10(4):article 23, 2014.
D. Belazzougui and G. Navarro. Optimal lower and upper bounds for representing sequences.
ACM Transactions on Algorithms, 11(4):article 31, 2015.
P. Bille, I. L. Gørtz, and F. R. Skjoldjensen. Deterministic indexing for packed strings.
In Proc. 28th Annual Symposium on Combinatorial Pattern Matching (CPM), LIPIcs 78,
page article 6, 2017.
M. Burrows and D. Wheeler. A block sorting lossless data compression algorithm. Technical
Report 124, Digital Equipment Corporation, 1994.
D. R. Clark. Compact PAT Trees. PhD thesis, University of Waterloo, Canada, 1996.
R. Cole, T. Kopelowitz, and M. Lewenstein. Suffix trays and suffix trists: Structures for
faster text indexing. Algorithmica, 72(2):450–466, 2015.
M. Farach. Optimal suffix tree construction with large alphabets. In Proc. 38th Annual
Symposium on Foundations of Computer Science (FOCS), pages 137–143, 1997.
P. Ferragina and G. Manzini. Indexing compressed text. Journal of the ACM, 52(4):552–
581, 2005.
P. Ferragina, G. Manzini, V. Mäkinen, and G. Navarro. Compressed representations of
sequences and full-text indexes. ACM Transactions on Algorithms, 3(2):article 20, 2007.
P. Ferragina and R. Venturini. A simple storage scheme for strings achieving entropy
bounds. Theoretical Computer Science, 371(1):115–121, 2007.
J. Fischer and P. Gawrychowski. Alphabet-dependent string searching with wexponential
search trees. In Proc. 26th Annual Symposium on Combinatorial Pattern Matching (CPM),
LNCS 9133, pages 160–171, 2015.
T. Gagie. Large alphabets and incompressibility. Information Processing Letters, 99(6):246–
251, 2006.
A. Golynski, J. I. Munro, and S. S. Rao. Rank/select operations on large alphabets: a tool
for text indexing. In Proc. 17th Annual ACM-SIAM Symposium on Discrete Algorithms,
(SODA), pages 368–373, 2006.
© J. Ian Munro, Gonzalo Navarro, Yakov Nekrich;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
R. Grossi, A. Orlandi, R. Raman, and S. S. Rao. More haste, less waste: Lowering the
redundancy in fully indexable dictionaries. In Proc. 26th International Symposium on
Theoretical Aspects of Computer Science (STACS), pages 517–528, 2009.
R. Grossi and J. S. Vitter. Compressed suffix arrays and suffix trees with applications to
text indexing and string matching. SIAM Journal on Computing, 35(2):378–407, 2005.
T. Hagerup, P. Bro Miltersen, and R. Pagh. Deterministic dictionaries. Journal of Algorithms, 41(1):69 – 85, 2001.
J. Kärkkäinen, P. Sanders, and S. Burkhardt. Linear work suffix array construction. Journal
of the ACM, 53(6):918–936, 2006.
D. K. Kim, J. S. Sim, H. Park, and K. Park. Constructing suffix arrays in linear time.
Journal of Discrete Algorithms, 3(2-4):126–142, 2005.
P. Ko and S. Aluru. Space efficient linear time construction of suffix arrays. Journal of
Discrete Algorithms, 3(2-4):143–156, 2005.
U. Manber and G. Myers. Suffix arrays: a new method for on-line string searches. SIAM
Journal on Computing, 22(5):935–948, 1993.
G. Manzini. An analysis of the Burrows-Wheeler transform. Journal of the ACM, 48(3):407–
430, 2001.
E. M. McCreight. A space-economical suffix tree construction algorithm. Journal of the
ACM, 23(2):262–272, 1976.
J. I. Munro. Tables. In Proc. 16th Conference on Foundations of Software Technology and
Theoretical Computer Science (FSTTCS), LNCS 1180, pages 37–42, 1996.
J. I. Munro, G. Navarro, and Y. Nekrich. Space-efficient construction of compressed indexes
in deterministic linear time. In Proc. 28th Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA), pages 408–424, 2017.
J. I. Munro, R. Raman, V. Raman, and S. S. Rao. Succinct representations of permutations
and functions. Theoretical Computer Science, 438:74–88, 2012.
G. Navarro and V. Mäkinen. Compressed full-text indexes. ACM Computing Surveys,
39(1):article 2, 2007.
G. Navarro and Y. Nekrich. Time-optimal top-k document retrieval. SIAM Journal on
Computing, 46(1):89–113, 2017.
G. Navarro and K. Sadakane. Fully-functional static and dynamic succinct trees. ACM
Transactions on Algorithms, 10(3):article 16, 2014.
K. Sadakane. Compressed suffix trees with full functionality. Theory of Computing Systems,
41(4):589–607, 2007.
E. Ukkonen. On-line construction of suffix trees. Algorithmica, 14(3):249–260, 1995.
P. Weiner. Linear pattern matching algorithms. In Proc. 14th Annual Symposium on
Switching and Automata Theory (FOCS), pages 1–11, 1973.
© J. Ian Munro, Gonzalo Navarro, Yakov Nekrich;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
| 8cs.DS
|
A Framework to Synergize Partial Order
Reduction with State Interpolation
Duc-Hiep Chu and Joxan Jaffar
arXiv:1408.0957v1 [cs.LO] 5 Aug 2014
National University of Singapore
hiepcd,[email protected]
Abstract. We address the problem of reasoning about interleavings in
safety verification of concurrent programs. In the literature, there are two
prominent techniques for pruning the search space. First, there are wellinvestigated trace-based methods, collectively known as “Partial Order
Reduction (POR)”, which operate by weakening the concept of a trace by
abstracting the total order of its transitions into a partial order. Second,
there is state-based interpolation where a collection of formulas can be
generalized by taking into account the property to be verified. Our main
contribution is a framework that synergistically combines POR with state
interpolation so that the sum is more than its parts.
1
Introduction
We consider the state explosion problem in safety verification of concurrent programs. This is caused by the interleavings of transitions from different processes.
In explicit-state model checking, a general approach to counter this explosion is
Partial Order Reduction (POR) (e.g., [20, 18, 9]). This exploits the equivalence
of interleavings of “independent” transitions: two transitions are independent if
their consecutive occurrences in a trace can be swapped without changing the
final state. In other words, POR-related methods prune away redundant process
interleavings in a sense that, for each Mazurkiewicz [16]1 trace equivalence class
of interleavings, if a representative has been checked, the remaining ones are
regarded as redundant.
On the other hand, symbolic execution [15] is another method for program
reasoning which recently has made increasing impact on software engineering
research [4]. The main challenge for symbolic execution is the exponential number of symbolic paths. The works [17, 12] tackle successfully this fundamental
problem by eliminating from the concrete model, on-the-fly, those facts which
are irrelevant or too-specific for proving the unreachability of the error nodes.
This learning phase consists of computing state-based interpolants in a similar
spirit to that of conflict clause learning in SAT solvers.
Now symbolic execution with state interpolation (SI) has been shown to be
effective for verifying sequential programs. In SI [17, 12], a node at program
1
We remark that the concept of POR goes beyond the preservation of Mazurkiewicz
traces, e.g. [20]. However, from a practical perspective, it is safe to consider such
form of pruning as a representative example of POR.
point ` in the reachability tree can be pruned, if its context is subsumed by the
interpolant computed earlier for the same program point `. Therefore, even in
the best case scenario, the number of states explored by an SI method must still
be at least the number of all distinct program points2 . However, in the setting of
concurrent programs, exploring each distinct global program point3 once might
already be considered prohibitive. In short, symbolic execution with SI alone is
not efficient enough for the verification of concurrent programs.
Recent work (e.g., [25]) has shown the usefulness of going stateful in implementing a POR method. It directly follows that SI can help to yield even better
performance. In order to implement an efficient stateful algorithm, we are required to come up with an abstraction for each (concrete or symbolic) state.
Unsurprisingly, SI often offers us good abstractions.
The above suggests that POR and SI can be very much complementary to
each other. In this paper, we propose a general framework employing symbolic
execution in the exploration of the state space, while both POR and SI are exploited for pruning. SI and POR are combined synergistically as the concept of
interpolation. Interpolation is essentially a form of learning where the completed
search of a safe subtree is then formulated as a recipe, ideally a succinct formula,
for future pruning. The key distinction of our interpolation framework is that
each recipe discovered by a node is forced to be conveyed back to its ancestors,
which gives rise to pruning of larger subtrees.
In summary, we address the challenge: “combining classic POR methods with
symbolic technique has proven to be difficult” [14], especially in the context
of software verification. More specifically, we propose an algorithm schema to
combine synergistically POR with state interpolation so that the sum is more
than its parts. However, we first need to formalize POR wrt. a symbolic search
framework with abstraction in such a way that: (1) POR can be property driven
and (2) POR, or more precisely, the concept of persistent set, can be applicable
for a set of states (rather than an individual state). While the main contribution
is a theoretical framework (for space reason, we present the proofs of the two
theorems in the Appendix), we also indicate a potential for the development of
advanced implementations.
2
Related Work
Partial Order Reduction (POR) is a well-investigated technique in model checking
of concurrent systems. Some notable early works are [20, 18, 9]. Later refinements
of POR, Dynamic [8] and Cartesian [11] POR (DPOR and CPOR respectively) improve traditional POR techniques by detecting collisions on-the-fly. Recently, [1]
2
3
Whereas POR-related methods do not suffer from this. Here we assume that the
input concurrent program has already been preprocessed (e.g., by static slicing to
remove irrelevant transitions, or by static block encodings) to reduce the size of the
transition system for each process.
The number of global points is the product of the numbers of local program points
in all processes.
has proposed the novel of concept of source sets, optimizing the implementation
for DPOR. These methods, in general, often achieve better reduction than traditional techniques, due to the more accurate detection of independent transitions.
Traditional POR techniques [20, 18, 9] distinguish between liveness and safety
properties. POR has also been extended for symbolic model checking [2]: a symbolic state can represent a number of concrete states. These methods, however,
are not applicable to safety verification of modern concurrent programs (written
in mainstream APIs such as POSIX). One important weakness of traditional
POR is that it is not sensitive wrt. different target safety properties. In contrast, recent works have shown that property-aware reduction can be achieved
by symbolic methods using a general-purpose SAT/SMT solver [24, 14, 22, 5]. Verification is often encoded as a formula which is satisfiable iff there exists an
interleaving execution of the programs that violates the property. Reductions
happen inside the SAT solver through the addition of learned clauses derived by
conflict analysis [19]. This type of reduction is somewhat similar to what we call
state interpolation.
An important related work is [14], which is the first to consider enhancing
POR with property driven pruning, via the use of an SMT solver. Subsequently,
there was a follow-up work [22]. In [14], they began with an SMT encoding of
the underlying transition system, and then they enhance this encoding with a
concept of “monotonicity”. The effect of this is that traces can be grouped into
equivalence classes, and in each class, all traces which are not monotonic will be
considered as unsatisfiable by the SMT solver. The idea of course is that such
traces are in fact redundant. This work has demonstrated some promising results
as most concurrency bugs in real applications have been found to be shallow.
We note that [14] incidentally enjoyed some (weak) form of SI pruning, due to
the similarity between conflict clause learning and state interpolation. However,
there the synergy between POR and SMT is unclear. We later demonstrate in
Sec. 7 that such synergy in [14] is indeed relatively poor.
There is a fundamental problem with scalability in [14], as mentioned in the
follow-up work [22]: “It will not scale to the entire concurrent program” if we
encode the whole search space as a single formula and submit it to an SMT solver.
Let us first compare [14] with our work. Essentially, the difference is twofold.
First, in this paper, the theory for partial order reduction is property driven. In
contrast, the monotonicity reduction of [14] is not. In other words, though property driven pruning is observed in [14], it is contributed mainly by the conflict
clauses learned, not from the monotonicity relation. We specifically exemplify the
power of property driven POR in the later sections. Second, the encoding in [14]
is processed by a black-box SMT solver. Thus important algorithmic refinements
are not possible. Some examples:
• There are different options in implementing SI. Specifically in this paper, we
employ “precondition” computations. Using a black-box solver, one has to rely
on its fixed interpolation methods.
• Our approach is lazy in a sense that our solver is only required to consider
one symbolic path at a time; in [14] it is not the case. This matters most when
the program is unsafe and finding counter-examples is relatively easy (there are
many traces which violate the safety property).
• In having a symbolic execution framework, one can direct the search process.
This is useful since the order in which state interpolants are generated does give
rise to different reductions. Of course, such manipulation of the search process
is hard, if not impossible, when using a black-box solver.
In order to remedy the scalability issue of [14], the work [22] adapted it to
the setting of program testing. In particular, [22] proposed a concurrent trace
program (CTP) framework which employs both concrete execution and symbolic
solving to strike a balance between efficiency and scalability of an SMT-based
method. However, when the input program is safe, i.e., absence of bugs, [22] in
general suffers from the same scalability issue as in [14].
We remark that, the new direction of [22], in avoiding the blow-up of the
SMT solver, was in fact preceded by the work on under-approximation widening
(UW) [10]. As with CTP, UW models a subset, which will be incrementally
enlarged, of all the possible interleavings as an SMT formula and submits it to
an SMT solver. In UW the scheduling decisions are also encoded as constraints,
so that the unsatisfiable core returned by the solver can then be used to further
the search in probably a useful direction. This is the major contribution of UW.
However, an important point is that this furthering of the search is a repeated
call to the solver, this time with a weaker formula; which means that the problem
at hand is now larger, having more traces to consider. On this repeated call, the
work done for the original call is thus duplicated.
At first glance, it seems attractive and simple to encode the problem compactly as a set of constraints and delegate the search process to a general-purpose
SMT solver. However, there are some fundamental disadvantages, and these arise
mainly because it is hard to exploit the semantics of the program to direct the
search inside the solver. This is in fact evidenced in the works mentioned above.
We believe, however, the foremost disadvantage of using a general-purpose
solver lies in the encoding of process interleavings. For instance, even when a
concurrent program has only one feasible execution trace, the encoding formula
being fed to the solver is still of enormous size and can easily choke up the solver.
More importantly, different from safety verification of sequential programs, the
encoding of interleavings (e.g., [14] uses the variable sel to model which process is selected for executing) often hampers the normal derivations of succinct
conflict clauses by means of resolution in modern SMT solvers. We empirically
demonstrate the inefficiency of such approach in Sec. 7.
Another important related work is [21], developed independently4 by Wachter
et al. [21] follows a similar direction as in this paper: combining POR with a
standard state interpolation algorithm, which is often referred to as the IMPACT
4
Our work has been publicly available since 2012 in forms of a draft paper and a
Ph.D. thesis. For anonymity reason, we omit the details here.
algorithm [17]. Nevertheless, it is important to note that the framework presented
in this paper subsumes [21]. While this paper proposes the novel concept of
Property Driven POR before combining it with the state interpolation algorithm,
[21] exploits directly the concept of “monotonicity” as in [14], thus their POR
part does not give rise to property driven pruning.
3
Background
We consider a concurrent system composed of a finite number of threads or
processes performing atomic operations on shared variables. Let Pi (1 ≤ i ≤ n)
be a process with the set transi of transitions. For simplicity, assume that transi
contains no cycles.
We also assume all processes have disjoint sets of transitions. Let T =
∪ni=1 transi be the set of all transitions. Let Vi be the set of local variables
of process Pi , and Vshared the set of shared variables of the given concurrent
program. Let pci ∈ Vi be a special variable representing the process program
counter, and the tuple hpc1 , pc2 · · · , pcn i represent the global program point.
Let SymStates be the set of all global symbolic states of the given program
where s0 ∈ SymStates is the initial state. A state s ∈ SymStates comprises two
parts: its global program point `, also denoted by pc(s), which is a tuple of local
program counters, and its symbolic constraints JsK over the program variables.
In other words, we denote a state s by hpc(s), JsKi.
We consider the transitions of states induced by the program. Following [9],
we only pay attention to visible transitions. A (visible) transition t{i} pertains
to some process Pi . It transfers process Pi from control location `1 to `2 . In
general, the application of t{i} is guarded by some condition cond (cond might
be just true). At some state s ∈ SymStates, when the ith component of pc(s),
namely pc(s)[i], equals `1 , we say that t{i} is schedulable 5 at s. And when s
satisfies the guard cond, denoted by s |= cond, we say that t{i} is enabled at
s. For each state s, let Schedulable(s) and Enabled(s) denote the set of transitions which respectively are schedulable at s and enabled at s. A state s, where
Schedulable(s) = ∅, is called a terminal state.
t
Let s → s0 denote transition step from s to s0 via transition t. This step
is possible only if t is schedulable at s. We assume that the effect of applying
an enabled transition t on a state s to arrive at state s0 is well-understood.
In our symbolic execution framework, executing a schedulable but not enabled
transition results in an infeasible state. A state s is called infeasible if JsK is
unsatisfiable. For technical reasons needed below, we shall allow schedulable
transitions emanating from an infeasible state; it follows that the destination
state must also be infeasible.
For a sequence of transitions w (i.e., w ∈ T ∗ ), Rng(w) denotes the set of
transitions that appear in w. Also let T` denote the set of all transitions which
are schedulable somewhere after global program point `. We note here that the
5
This concept is not standard in traditional POR, we need it here since we are dealing
with symbolic search.
schedulability of a transition at some state s only depends on the program point
component of s, namely pc(s). It does not depend on the constraint component
of s, namely JsK. Given t1 , t2 ∈ T we say t1 can de-schedule t2 iff there exists a
state s such that both t1 , t2 are schedulable at s but t2 is not schedulable after
the execution of t1 from s.
t ···t
Following the above, s1 1=⇒m sm+1 denotes a sequence of state transitions,
tm
t2
t1
sm+1
··· →
s2 →
and we say that sm+1 is reachable from s1 . We call s1 →
a feasible derivation from state s1 , iff ∀ 1 ≤ i ≤ m • ti is enabled at si .
As mentioned earlier, an infeasible derivation results in an infeasible state (an
infeasible state is still aware of its global program point). An infeasible state
satisfies any safety property.
We define a complete execution trace, or simply trace, ρ as a sequence of
ρ
transitions such that it is a derivation from s0 and s0 =⇒ sf and sf is a
terminal state. A trace is infeasible if it is an infeasible derivation from s0 . If a
trace is infeasible, then at some point, it takes a transition which is schedulable
but is not enabled. From thereon, the subsequent states are infeasible states.
We say a given concurrent program is safe wrt. a safety property ψ if ∀s ∈
SymStates • if s is reachable from the initial state s0 then s |= ψ. A trace ρ is
safe wrt. ψ, denoted as ρ |= ψ, if all its states satisfy ψ.
Partial Order Reduction (POR) vs. State-based Interpolation (SI)
We assume the readers are familiar with the traditional concept of POR. Regarding state-based interpolation, we follow the approach of [13, 17]. Here our
symbolic execution is depicted as a tree rooted at the initial state s0 and for each
state si therein, the descendants are just the states obtainable by extending si
with a feasible transition.
Definition 1 (Safe Root). Given a transition system and an initial state s0 ,
let s be a feasible state
a reachable from s0 . We say s is a safe root wrt. a safety
property ψ, denoted ψ (s), iff all states s0 reachable from s are safe wrt. ψ.
Definition 2 (State Coverage). Given a transition system and an initial state
s0 and si and sj are two symbolic states such that (1) si and sj are reachable
from s0 and (2) si and sj share the same program a
point `, we say a
si covers sj
wrt. a safety property ψ, denoted by si ψ sj , iff ψ (si ) implies ψ (sj ).
The impact of state coverage relation is that if (1) si covers sj , and (2) the
subtree rooted at si has been traversed and proved to be safe, then the traversal
of subtree rooted at sj can be avoided. In other words, we gain performance by
pruning the subtree at sj . Obviously, if si naturally subsumes sj , i.e., sj |= si ,
then state coverage is trivially achieved. In practice, however, this scenario does
not happen often enough.
Definition 3 (Sound Interpolant). Given a transition system and an initial
state s0 , given a safety property ψ and program point `, we say a formula Ψ is a
sound interpolant for `, denoted by SI(`, ψ), if for all states s ≡ h`, ·i reachable
from s0 , s |= Ψ implies that s is a safe root.
What we want now is to generate a formula Ψ (called interpolant), which
still preserves the safety of all states reachable from si , but is weaker (more
general) than the original formula associated to the state si . In other words,
we should have si |= SI(`, ψ). We assume that this condition is always ensured
by any implementation of state-based interpolation. The main purpose of using
Ψ rather than the original formula associated to the symbolic state si is to
increase the likelihood of subsumption. That is, the likelihood of having sj |= Ψ
is expected to be much higher than the likelihood of having sj |= si .
In fact, the perfect interpolant should be the weakest precondition [7] computed for program point ` wrt. the transition system and the safety property ψ. We denote this weakest precondition as wp(`, ψ). Any subsequent state
sj ≡ h`, ·i which has sj stronger than this weakest precondition can be pruned.
However, in general, the weakest precondition is too computationally demanding. An interpolant for the state si is indeed a formula which approximates
the weakest precondition at program point ` wrt. the transition system, i.e.,
Ψ ≡ SI(`, ψ) ≡ Intp(si , wp(`, ψ)). A good interpolant is one which closely approximates the weakest precondition and can be computed efficiently.
The symbolic execution of a program can be augmented by annotating each
program point with its corresponding interpolants such that the interpolants
represent the sufficient conditions to preserve the unreachability of any unsafe
state. Then, the basic notion of pruning with state interpolant can be defined as
follows.
Definition 4 (Pruning with Interpolant). Given a symbolic state s ≡ h`, ·i
such that ` is annotated with some interpolant Ψ , we say that s is pruned by the
interpolant Ψ if s implies Ψ (i.e., s |= Ψ ).
Now let us discuss the the effectiveness of POR and SI in pruning the search
space with an example. For simplicity, we purposely make the example concrete,
i.e., states are indeed concrete states.
1 (Closely coupled processes): See Fig. 1. Program points are shown in
angle brackets. Fig. 1(a) shows the control flow graphs of two processes. Process
1 increments x twice whereas process 2 doubles x twice. The transitions associated with such actions and the safety property are depicted in the figure. POR
requires a full search tree while Fig. 1(b) shows the search space explored by SI.
Interpolants are in curly brackets. Bold circles denote pruned/subsumed states.
{1}
Let us first attempt this example using POR. It is clear that t1 is dependent
{2}
{1}
{2}
{2}
{2}
with both t1 and t2 . Also t2 is dependent with both t1 and t2 . Indeed,
each of all the 6 execution traces in the search tree ends at a different concrete
state. As classic POR methods use the concept of trace equivalence for pruning,
no interleaving is avoided: those methods will enumerate the full search tree of
19 states (for space reasons, we omit it here).
Revisit the example using SI, where we use the weakest preconditions [7] as
the state interpolants: the interpolant for a state is computed as the weakest
precondition to ensure that the state itself as well as all of its descendants are
safe (see Fig. 1(b)). We in fact achieve the best case scenario with it: whenever we
EXAMPLE
{x ≤ 0} x=0
h0, 0i
{1}
Shared variables: x
Initially: x = 0
{1}
Process 2
Process 1
h0i
h0i
{1}
t1
{2}
:x=x∗2
{2}
:x=x∗2
t1
:x++
h1i
h1i
{1}
t2
:x++
t2
h2i
h2i
Safety:
{2}
t1
t1
{x ≤ 1}
{x ≤ 2} x=0
x=1
h0, 1i
h1, 0i
ψ≡x≤8
(a) Two Closely Coupled Processes
{2}
{1}
{2}
t2
t1
t1
t2
{x ≤ 6}
{x ≤ 2} {x ≤ 3} x=2 x=1
h0, 2i
h1, 1i
h2, 0i
h1, 1i
x=0
{2} x=2
{1}
{2}
{1}
t1
t2
t2
t2
{x ≤ 4}
x=3
h1, 2i
x=4
h1, 2i
h2, 1i
x=4h2, 1i
x=1
{x ≤ 7}
{2}
{1}
t2
t2
x=8
h2, 2i
{x ≤ 8}
h2, 2i
x=5
(b) Search Space by SI
Fig. 1: Application of SI on 2 Closely Coupled Processes
come to a program point which has been examined before, subsumption happens.
The number of non-subsumed states is still of order O(k 2 ) (where k = 3 in this
particular example), assuming that we generalize the number of local program
points for each process to O(k). Fig. 1(b) shows 9 non-subsumed states and 4
subsumed states.
In summary, the above example shows that SI might outperform POR when
the component processes are closely coupled. However, one can easily devise an
example where the component processes do not interfere with each other at all.
Under such condition POR will require only one trace to prove safety, while SI is
still (lower) bounded by the total number of global program points. In this paper,
we contribute by proposing a framework to combine POR and SI synergistically.
4
Property Driven POR (PDPOR)
“Combining classic POR methods with symbolic algorithms has been proven to
be difficult” [14]. One fundamental reason is that the concepts of (Mazurkiewicz)
equivalence and transition independence, which drive most practical POR implementations, rely on the equivalence of two concrete states. However, in symbolic
traversal, we rarely encounter two equivalent symbolic states.
We now make the following definition which is crucial for the concept of
pruning and will be used throughout this paper.
Definition 5 (Trace Coverage). Let ρ1 , ρ2 be two traces of a concurrent program. We say ρ1 covers ρ2 wrt. a safety property ψ, denoted as ρ1 wψ ρ2 , iff
ρ1 |= ψ → ρ2 |= ψ.
Instead of using the concept of trace equivalence, from now on, we only make
use of the concept of trace coverage. The concept of trace coverage is definitely
weaker than the concept of Mazurkiewicz equivalence. In fact, if ρ1 and ρ2 are
(Mazurkiewicz) equivalent then ∀ψ • ρ1 wψ ρ2 ∧ ρ2 wψ ρ1 . Now we will define
a new and weaker concept which therefore generalizes the concept of transition
independence.
Definition 6 (Semi-Commutative After A State). For a given concurrent
θ
program, a safety property ψ, and a derivation s0 =⇒ s, for all t1 , t2 ∈ T
which cannot de-schedule each other, we say t1 semi-commutes with t2 after
state s wrt. wψ , denoted by hs, t1 ↑ t2 , ψi, iff for all w1 , w2 ∈ T ∗ such that
θw1 t1 t2 w2 and θw1 t2 t1 w2 are execution traces of the program, then we have
θw1 t1 t2 w2 wψ θw1 t2 t1 w2 .
From the definition, Rng(θ), Rng(w1 ), and Rng(w2 ) are pairwise disjoint.
Importantly, if s is at program point `, we have Rng(w1 )∪Rng(w2 ) ⊆ T` \{t1 , t2 }.
We observe that wrt. some ψ, if all important events, those have to do with
the safety of the system, have already happened in the prefix θ, the “semicommutative” relation is trivially satisfied. On the other hand, the remaining
transitions might still interfere with each other (but not the safety), and do not
satisfy the independent relation.
The concept of “semi-commutative” is obviously weaker than the concept of
independence. If t1 and t2 are independent, it follows that ∀ψ ∀s•hs, t1 ↑ t2 , ψi∧
hs, t2 ↑ t1 , ψi. Also note that, in contrast to the relation of transition independence, the “semi-commutative” relation is not symmetric.
We now introduce a new definition for persistent set.
Definition 7 (Persistent Set Of A State). A set T ⊆ T of transitions
schedulable in a state s ∈ SymStates is persistent in s wrt. a property ψ iff,
tm−1
tm
t1
t2
sm including only transitions
for all derivations s →
s1 →
s2 . . . → sm−1 →
ti ∈ T and ti 6∈ T, 1 ≤ i ≤ m, each transition in T semi-commutes with ti after
s wrt. wψ .
For each state, computing a
Safety property ψ and current state s
persistent set from the “semi- h1i T := ∅
commutative” relation is similar h2i Add an enabled transition t into T
to computing the classical per- h3i foreach remaining schedulable transition ti
if ¬(∀ tpj ∈ T • hs, tpj ↑ ti , ψi)
sistent set under the transition h4i
Add ti into T
independence relation. The algo- h5i
rithms for this task can be easily Fig. 2: Computing a Persistent Set of a State
adapted from the algorithms presented in [9]. For convenience, we show one of
such possibilities in Fig. 2.
As in traditional algorithms, the quality (i.e. the size) of the returned persistent set is highly dependent on the first transitioned t to be added and the
order in which the remaining transitions ti are considered. This is, however, not
the topic of the current paper.
With the new definition of persistent set, we now can proceed with the normal selective search as described in classic POR techniques. In the algorithm
presented in Fig. 3, we perform depth first search (DFS), and only accommodate safety verification (invariant property ψ).
Theorem 1. The selective search algorithm in Fig. 3 is sound6 .
In preparing for POR and SI to
work together, we now further modify the concept of persistent set so
that it applies for a set of states
sharing the same program point. We
remark that the previous definitions
apply only for a specific state. The
key intuition is to attach a precondition φ to the program point
of interest, indicating when semicommutativity happens.
t
u
Safety property ψ and initial state s0
h1i Initially : Explore(s0 )
function Explore(s)
h2i if s 6|= ψ Report Error and TERMINATE
h3i T := Persistent Set(s)
h4i foreach enabled transition t in T do
t
h5i
s −→ s0
/* Execute t */
h6i
Explore(s0 )
end function
Fig. 3: New Selective Search Algorithm
Definition 8 (Semi-Commutative After A Program Point). For a given
concurrent program, a safety property ψ, and t1 , t2 ∈ T , we say t1 semi-commutes
with t2 after program point ` wrt. wψ and φ, denoted as h`, φ, t1 ↑ t2 , ψi, iff
for all state s ≡ h`, ·i reachable from the initial state s0 , if s |= φ then t1 semicommutes with t2 after state s wrt. wψ .
Definition 9 (Persistent Set Of A Program Point). A set T ⊆ T of transitions schedulable at program point ` is persistent at ` under the interpolant Ψ
wrt. a property ψ iff, for all state s ≡ h`, ·i reachable from the initial state s0 ,
t
t
tm−1
t
1
2
m
s1 →
if s |= Ψ then for all derivations s →
s2 . . . → sm−1 →
sm including only
transitions ti ∈ T and ti 6∈ T, 1 ≤ i ≤ m, each transition in T semi-commutes
with ti after state s wrt. wψ .
Assume
V that T = {tp1 , tp2 , · · · tpk }. The interpolant Ψ can now be computed
as Ψ = φji for 1 ≤ j ≤ k, 1 ≤ i ≤ m such that h`, φji , tpj ↑ ti , ψi.
For each program point, it is possible to have different persistent sets associated with different interpolants. In general, a state which satisfies a stronger
interpolant will have a smaller persistent set, therefore, it enjoys more pruning.
5
Synergy of PDPOR and SI
We now show our combined framework. We assume for each program point,
a persistent set and its associated interpolant are computed statically, i.e., by
separate analyses. In other words, when we are at a program point, we can right
away make use of the information about its persistent set.
The algorithm is in Fig. 4. The function Explore has input s and assumes the
safety property at hand is ψ. It naturally performs a depth first search of the
state space.
6
Proof outline is in the Appendix.
Two Base Cases: The function
Explore handles two base cases. Assume safety property ψ and initial state s0
h1i Initially : Explore(s0 )
One is when the current state
function Explore(s)
is subsumed by some computed
Let s be h`, ·i
(and memoed) interpolant Ψ . No
h2i if (memoed(s, Ψ )) return Ψ
further exploration is needed, and
h3i if (s 6|= ψ) Report Error and TERMINATE
Ψ is returned as the interpolant
h4i Ψ := ψ
h5i hT, Ψ trace i := Persistent Set(`)
(line 2). The second base case is
h6i if (s |= Ψ trace )
when the current state is found to
h7i
Ts := T
be unsafe (line 3).
h8i
Ψ := Ψ ∧ Ψ trace
Combining Interpolants: We
h9i else Ts := Schedulable(s)
make use of the (static) persistent h10i foreach t in (Ts \ Enabled(s)) do
set T computed for the current h11i
Ψ := Ψ ∧ pre(t, false)
program point. We comment fur- h12i foreach t in (Ts ∩ Enabled(s)) do
t
ther on this in the next section.
/* Execute t */
h13i
s −→ s0
0
The set of transitions to be h14i
Ψ := Explore(s0 )
0
considered is denoted by Ts. h15i
Ψ := Ψ ∧ pre(t, Ψ )
When the current state implies h16i memo and return (Ψ )
the interpolant Ψ trace associated end function
with T , we need to consider only Fig. 4: A Framework for POR and SI (DFS)
those transitions in T . Otherwise,
we need to consider all the schedulable transitions. Note that when the persistent set T is employed, the interpolant Ψ trace must contribute to the combined
interpolant of the current state (line 8). Disabled transitions at the current state
will strengthen the interpolant as in line 11. Finally, we recursively follow those
transitions which are enabled at the current state. The interpolant of each child
state contributes to the interpolant of the current state as in line 15. In our
framework, interpolants are propagated back using the precondition operation
pre, where pre(t, φ) denotes a safe approximation of the weakest precondition
wrt. the transition t and the postcondition φ [7].
Theorem 2. The algorithm in Fig. 4 is sound7 .
6
t
u
Implementation of PDPOR
We now elaborate on the remaining task: how to estimate the semi-commutative
relation, thus deriving the (static) persistent set at a program point. Similar to
the formalism of traditional POR, our formalism is of paramount importance
for the semantic use as well as to construct the formal proof of correctness
(see the Appendix). In practice, however, we have to come up with sufficient
conditions to efficiently implement the concepts. In this paper, we estimate the
semi-commutative relation in two steps:
1. We first employ any traditional POR method and first estimate the “semicommutative” relation as the traditional independence relation (then the
7
Proof outline is in the Appendix.
corresponding condition φ is just true). This is possible because the proposed concepts are strictly weaker than the corresponding concepts used in
traditional POR methods.
2. We then identify and exploit a number of patterns under which we can statically derive and prove the semi-commutative relation between transitions.
In fact, these simple patterns suffice to deal with a number of important
real-life applications.
In the rest of this section, we outline three common classes of problems, from
which the semi-commutative relation between transitions can be easily identified
and proved, i.e., our step 2 becomes applicable.
Resource Usage of Concurrent Programs: Programs make use of limited
resource (such as time, memory, bandwidth). Validation of resource usage in
sequential setting is already a hard problem. It is obviously more challenging in
the setting of concurrent programs due to process interleavings.
Here we model this class of problems by using a resource variable r. Initially,
r is zero. Each process can increment or decrement variable r by some concrete
value (e.g., memory allocation or deallocation respectively). A process can also
double the value r (e.g., the whole memory is duplicated). However, the resource
variable r cannot be used in the guard condition of any transition, i.e., we cannot
model the behavior of a typical garbage collector. The property to be verified is
that, “at all times, r is (upper-) bounded by some constant”.
Proposition 1. Let r be a resource variable of a concurrent program, and assume the safety property at hand is ψ ≡ r ≤ C, where C is a constant. For all
transitions (assignment operations only) t1 : r = r+c1 , t2 : r = r∗2, t3 : r = r−c2
where c1 , c2 > 0, we have for all program points `:
h`, true, t1 ↑ t2 , ψi ∧ h`, true, t1 ↑ t3 , ψi ∧ h`, true, t2 ↑ t3 , ψi
t
u
Informally, other than common mathematical facts such as additions can
commute and so do multiplications and subtractions, we also deduce that additions can semi-commute with both multiplications and subtractions while multiplications can semi-commute with subtractions. This Proposition can be proved
by using basic laws of algebra.
2 : Let us refer back to the example of two closely coupled processes
introduced in Sec. 3, but now under the assumption that x is the resource variable
of interest. Using the semi-commutative relation derived from Proposition 1, we
need to explore only one complete trace to prove this safety.
We recall that, in contrast, POR (and DPOR)-only methods will enumerate the
full execution tree which contains 19 states and 6 complete execution traces. Any
technique which employs only the notion of Mazurkiewicz trace equivalence for
pruning will have to consider all 6 complete traces (due to 6 different terminal
states). SI alone can reduce the search space in this example, and requires to
explore only 9 states and 4 subsumed states (as in Sec. 3).
Detection of Race Conditions: [23] proposed a property driven pruning algorithm to detect race conditions in multithreaded programs. This work has
achieved more reduction in comparison with DPOR. The key observation is that,
EXAMPLE
at a certain location (program point) `, if their conservative “lockset analysis”
shows that a search subspace is race-free, the subspace can be pruned away. As we
know, DPOR relies solely on the independence relation to prune redundant interleavings (if t1 , t2 are independent, there is no need to flip their execution order).
In [23], however, even when t1 , t2 are dependent, we may skip the corresponding
search space if flipping the order of t1 , t2 does not affect the reachability of any
race condition. In other words, [23] is indeed a (conservative) realization of our
PDPOR, specifically targeted for detection of race conditions. Their mechanism
to capture the such scenarios is by introducing a trace-based lockset analysis.
Ensuring Optimistic Concurrency: In the implementations of many concurrent protocols, optimistic concurrency, i.e., at least one process commits, is
usually desirable. This can be modeled by introducing a flag variable which will
be set when some process commits. The flag variable once set can not be unset. It is then easy to see that for all program point ` and transitions t1 , t2 , we
have h`, flag = 1, t1 ↑ t2 , ψi. Though simple, this observation will bring us more
reduction compared to traditional POR methods.
7
Experiments
This section conveys two key messages. First, when trace-based and state-based
methods are not effective individually, our combined framework still offers significant reduction. Second, property driven POR can be very effective, and applicable not only to academic programs, but also to programs used as benchmarks
in the state-of-the-art.
We use a 3.2 GHz Intel processor and 2GB memory running Linux. Timeout
is set at 10 minutes. In the tables, cells with ‘-’ indicate timeout. We compare the
performance of Partial Order Reduction alone (POR), State Interpolation alone
(SI), the synergy of Partial Order Reduction and State Interpolation (POR+SI),
i.e., the semi-commutative relation is estimated using only step 1 presented in
Sec. 6, and when applicable, the synergy of Property Driven Partial Order Reduction and State Interpolation (PDPOR+SI), i.e., the semi-commutative relation
is estimated using both steps presented in Sec. 6. For the POR component, we
use the implementation from [3].
Table 1 starts with parameterized versions of the producer/consumer example
because its basic structure is extremely common. There are 2 ∗ N producers and
1 consumer. Each producer will do its own non-interfered computation first,
modeled by a transition which does not interfere with other processes. Then
these producers will modify the shared variable x as follows: each of the first N
producers increments x, while the other N producers double the value of x. On
the other hand, the consumer consumes the value of x. The safety property is
that the consumed value is no more than N ∗ 2N .
Table 1 clearly demonstrates the synergy benefits of POR and SI. POR+SI
significantly outperforms both POR and SI. Note that this example can easily
be translated to the resource usage problem, where our PDPOR requires only a
single trace (and less than 0.01 second) in order to prove safety.
We next use the parameTable 1: Synergy of POR and SI
terized version of the dining
POR
SI
POR+SI
philosophers. We chose this Problem States T(s) States T(s) States T(s)
for two reasons. First, this is
p/c-2
449 0.03
514 0.17
85 0.03
p/c-3 18745 2.73
6562 2.43
455 0.19
a classic example often used
p/c-4 986418 586.00 76546 37.53
2313 1.07
in concurrent algorithm dep/c-5
−
−
−
− 11275 5.76
sign to illustrate synchrop/c-6
−
−
−
− 53261 34.50
nization issues and techp/c-7
−
−
−
− 245775 315.42
niques for resolving them. din-2a
22 0.01
21 0.01
21 0.01
Second, previous work [14] din-3a
646 0.05
153 0.03
125 0.02
647 0.09
has used this to demon- din-4a 155037 19.48 1001 0.17
−
−
6113 1.01
4313 0.54
strate benefits from combin- din-5a
din-6a
−
− 35713 22.54 24201 4.16
ing POR and SMT.
din-7a
−
− 202369 215.63 133161 59.69
The first safety property
bak-2
48 0.03
38 0.03
31 0.02
used in [14], “it is not that
bak-3
1003 1.85
264 0.42
227 0.35
all philosophers can eat sibak-4 27582 145.78
1924 5.88
1678 4.95
multaneously”, is somewhat
bak-5
−
− 14235 73.69 12722 63.60
trivial. Therefore, here we
verify a tight property, which is (a): “no more than half the philosophers can
eat simultaneously”. To demonstrate the power of symbolic execution, we verify
this property without knowing the initial configurations of all the forks. Table 1,
again, demonstrates the significant improvements of POR+SI over POR alone and
SI alone. We note that the performance of our POR+SI algorithm is about 3 times
faster than [14].
We additionally considered a second safety property as in [14], namely (b):
“it is possible to reach a state in which all philosophers have eaten at least once”.
Our symbolic execution framework requires only a single trace (and less than
0.01 second) to prove this property in all instances, whereas [14] requires even
more time compared to proving property (a). This illustrates the scalability issue
of [14], which is representative for other techniques employing general-purpose
SMT solver for symbolic pruning.
We also perform experiments on the “Bakery” algorithm. Due to existence
of infinite domain variables, model checking hardly can handle this case. Here
we remark that in symbolic methods, loop handling is often considered as an
orthogonal issue. Programs with statically bounded loops can be easily unrolled
into equivalent loop-free programs. For unbounded loops, either loop invariants
are provided or the employment of some invariant discovery routines, e.g., [12], is
necessary. In order for our algorithm to work here, we make use of the standard
loop invariant for this example.
To further demonstrate the power our synergy framework over [14]8 as well
as the power of our property driven POR, we experiment next on the Sumof-ids program. Here, each process (of N processes) has one unique id and will
8
[14] is not publicly available. Therefore, it is not possible for us to make more comprehensive comparisons.
increment a shared variable sum by this id. We prove that in the end this variable
will be incremented by the sum of all the ids.
See Table 2, where we exTable 2: Comparison with [14]
periment with Z3 [6] (ver[14] w. Z3
POR+SI PDPOR+SI
sion 4.1.2) using the encodT(s) #C #D T(s) States T(s) States T(s)
ings presented in [14]. #C desum-6 1608 1795 0.08
193 0.05
7 0.01
notes the number of conflicts
sum-8 54512 59267 10.88
1025 0.27
9 0.01
−
−
−
5121 1.52
11 0.01
while #D denotes the num- sum-10
−
−
− 24577 8.80
13 0.01
ber of decisions made by Z3. sum-12
sum-14
−
−
−
114689
67.7
15 0.01
We can see that our synergy
framework scale much better than [14] with Z3. Also, this example can also
be translated to resource usage problem, our use of property-driven POR again
requires one single trace to prove safety.
Finally, to benchmark our
Table 3: Experiments on [5]’s Programs
framework with SMT-based meth[5]
SI
PDPOR+SI
ods, we select four safe proProblem
C T(s) States T(s) States T(s)
grams from [5] where the exper- micro 2
17 1095 20201 10.88
201 0.04
imented methods did not perform stack
12 225
529 0.26
529 0.26
29 0.03
29 0.03
well. Those programs are micro 2, circular buffer ∞ 477
10 95 1681 1.13
41 0.01
stack, circular buffer, and state- stateful20
ful20. We note that safe programs allow fairer comparison between different
approaches since to verify them we have to cover the whole search space. Table 3 shows the running time of SI alone and of the combined framework. For
convenience, we also tabulate the best running time reported in [5] and C is
the context switch bound used. We assume no context switch bound, hence the
corresponding value in our framework is ∞.
We can see that even our SI alone significantly outperforms the techniques
in [5]. We believe it is due to the inefficient encoding of process interleavings
(mentioned in Sec. 2) as well as the following reasons. First, our method is lazy,
which means that only a path is considered at a time: [5] itself demonstrates
partially the usefulness of this. Second, but importantly, we are eager in discovering infeasible paths. The program circular buffer, which has only one feasible
complete execution trace, can be efficiently handled by our framework, but not
SMT. This is one important advantage of our symbolic execution framework over
SMT-based methods, as discussed in [17].
It is important to note that, PDPOR significantly improves the performance
of SI wrt. programs micro 2 and stateful20. This further demonstrates the applicability of our proposed framework.
8
Conclusion
We present a concurrent verification framework which synergistically combines
trace-based reduction techniques with the recently established notion of state
interpolant. One key contribution is the new concept of property-driven POR
which serves to reduce more interleavings than previously possible. We believe
that fully automated techniques to compute the semi-commutative relations for
specific, but important, application domains will be interesting future work.
References
1. P. Abdulla, S. Aronis, B. Jonsson, and K. Sagonas. Optimal dynamic partial order
reduction. In POPL, 2014.
2. R. Alur, R. K. Brayton, T. A. Henzinger, S. Qadeer, and S. K. Rajamani. Partialorder reduction in symbolic state space exploration. In CAV, 1997.
3. P. Bokor, J. Kinder, M. Serafini, and N. Suri. Supporting domain-specific state
space reductions through local partial-order reduction. In ASE, 2011.
4. C. Cadar, P. Godefroid, S. Khurshid, C. S. Păsăreanu, K. Sen, N. Tillmann, and
W. Visser. Symbolic execution for software testing in practice: preliminary assessment. In ICSE, 2011.
5. L. Cordeiro and B. Fischer. Verifying multi-threaded software using smt-based
context-bounded model checking. In ICSE, 2011.
6. L. De Moura and N. Bjørner. Z3: an efficient smt solver. In TACAS, 2008.
7. E. W. Dijkstra. Guarded commands, nondeterminacy and formal derivation of
programs. Commun. ACM, 1975.
8. C. Flanagan and P. Godefroid. Dynamic partial-order reduction for model checking
software. In POPL, 2005.
9. P. Godefroid. Partial-Order Methods for the Verification of Concurrent Systems:
An Approach to the State-Explosion Problem. Springer-Verlag New York, Inc.,
1996.
10. O. Grumberg, F. Lerda, O. Strichman, and M. Theobald.
Proof-guided
underapproximation-widening for multi-process systems. In POPL, 2005.
11. G. Gueta, C. Flanagan, E. Yahav, and M. Sagiv. Cartesian partial-order reduction.
In SPIN, 2007.
12. J. Jaffar, J.A. Navas, and A. Santosa. Unbounded Symbolic Execution for Program
Verification. In RV, 2011.
13. J. Jaffar, A. E. Santosa, and R. Voicu. An interpolation method for clp traversal.
In CP, 2009.
14. V. Kahlon, C. Wang, and A. Gupta. Monotonic partial order reduction: An optimal
symbolic partial order reduction technique. In CAV, 2009.
15. J. C. King. Symbolic Execution and Program Testing. Com. ACM, 1976.
16. A. W. Mazurkiewicz. Trace theory. In Advances in Petri Nets, 1986.
17. K. L. McMillan. Lazy annotation for program testing and verification. In CAV,
2010.
18. D. Peled. All from one, one for all: on model checking using representatives. In
CAV, 1993.
19. J. P. M. Silva and K. A. Sakallah. Grasp-a new search algorithm for satisfiability.
In ICCAD, 1996.
20. A. Valmari. Stubborn sets for reduced state space generation. In Applications and
Theory of Petri Nets, 1989.
21. B. Wachter, D. Kroening, and J. Ouaknine. Verifying multi-threaded software with
impact. In FMCAD, 2013.
22. C. Wang, S. Chaudhuri, A. Gupta, and Y. Yang. Symbolic pruning of concurrent
program executions. In ESEC/FSE, 2009.
23. C. Wang, Y. Yang, A. Gupta, and G. Gopalakrishnan. Dynamic model checking
with property driven pruning to detect race conditions. In ATVA, 2008.
24. C. Wang, Z. Yang, V. Kahlon, and A. Gupta. Peephole partial order reduction. In
TACAS, 2008.
25. Y. Yang, X. Chen, G. Gopalakrishnan, and R. M. Kirby. Efficient stateful dynamic
partial order reduction. In SPIN, 2008.
Appendix: Proofs of the Two Theorems
Theorem 1. The selective search algorithm in Fig. 3 is sound.
Proof (Outline). Assume that there exist some traces which violate the property
ψ and are not examined by our selective search. Let denote the set of such traces
tm
t2
t1
sm , ρ ∈ Wviolated , let
s2 · · · →
s1 →
as Wviolated . For each trace ρ = s0 →
f irst(ρ) denote the smallest index i such that ti is not in the persistent set
tm
t2
t1
sm
s2 · · · →
s1 →
of si−1 . Without loss of generality, assume ρmax = s0 →
having the maximum f irst(ρ). Let i = f irst(ρmax ) < m. As the “commuter”
and “commutee” cannot “de-schedule” each other, in the set {ti+1 · · · tm } there
must be a transition which belongs to the persistent set of si−1 (otherwise, the
must exist some transition that belongs to the persistent set of si−1 which is
schedulable at sm . Therefore sm is not a terminal state). Let j be the smallest
index such that tj belongs to the persistent set of si−1 . By definition, wrt. wψ
and after si−1 , tj semi-commutes with ti , ti+1 , · · · tj−1 . Also due to the definition
of the “semi-commutative” relation we deduce that all the following traces (by
making tj repeatedly commute backward):
ρ01 = t1 t2 · · · ti−1 ti ti+1 · · · tj tj−1 tj+1 · · · tm
..
.
ρ0j−i−1 = t1 t2 · · · ti−1 ti tj ti+1 · · · tj−1 tj+1 · · · tm
ρ0j−i = t1 t2 · · · ti−1 tj ti ti+1 · · · tj−1 tj+1 · · · tm
must violate the property ψ too. However, f irst(ρ0j−i ) > f irst(ρmax ). This contradicts the definition of ρmax .
t
u
s0
θ2
θ1
Subsumed
si
t1
t2
t3
sj
t1
si+1
sj+1
B
A
t2
t3
B’
A’
Fig. 5: Inductive Correctness
Theorem 2. The synergy algorithm in Fig. 4 is sound.
Proof (Outline). We use structural induction. Refer to Fig. 5. Assume that from
s0 we reach state si ≡ h`, ·i. W.l.o.g., assume that at si there are three transitions
which are schedulable, namely t1 , t2 , t3 , of which only t1 and t2 are enabled. Also
assume that under the interpolant Ψ 2 , the persistent set of `, and therefore of
si , is just {t1 }. From the algorithm, we will extend si with t1 (line 13) and
attempt to verify the subtree A (line 14). Our induction hypothesis is that we
have finished considering A, and indeed, it is safe under the interpolant Ψ A .
That subtree will contribute Ψ 1 = pre(t1 , Ψ A ) (line 15) to the interpolant of si .
Using the interpolant Ψ 2 , the branch having transition t2 followed by the
subtree B is pruned and that is safe, due to Theorem 1. Also, the disabled
transition t3 contributes Ψ 3 (line 11) to the interpolant of si . Now we need to
prove that Ψ = Ψ 1 ∧ Ψ 2 ∧ Ψ 3 is indeed a sound interpolant for program
point `.
Assume that subsequently in the search, we reach some state sj ≡ h`, ·i. We
will prove that Ψ is a sound interpolant of ` by proving that if sj |= Ψ , then the
pruning of sj is safe.
First, sj |= Ψ implies that sj |= Ψ 3 . Therefore, at sj , t3 is also disabled.
t1
sj+1 (if not the pruning
Second, assume that t1 is enabled at sj and sj −→
0
of t1 followed by A is definitely safe). Similarly, sj |= Ψ implies that sj |= Ψ 1 .
Consequently, sj+1 |= Ψ A and therefore the subtree A0 is safe too. Lastly, sj |= Ψ
implies that sj |= Ψ 2 . Thus the reasons which ensure that the traces ending with
subtree A cover the traces ending with subtree B also hold at sj . That is, the
traces ending with subtree A0 also cover the traces ending with subtree B 0 . t
u
| 6cs.PL
|
Numerical Model For Vibration Damping Resulting
From the First Order Phase Transformations
arXiv:cs/0702172v1 [cs.CE] 28 Feb 2007
L. X. Wang 1 ∗ and Roderick V.N. Melnik
1
MCI, Faculty of Science and Engineering,
University of Southern Denmark,
2
2
Sonderborg, DK-6400, Denmark
Mathematical Modelling and Computational Sciences,
Wilfrid Laurier University,
75 University Avenue West,
Waterloo, ON, Canada, N2L 3C5
Abstract
A numerical model is constructed for modelling macroscale damping effects induced by the first order martensite phase transformations in a shape memory alloy
rod. The model is constructed on the basis of the modified Landau-Ginzburg theory
that couples nonlinear mechanical and thermal fields. The free energy function for
the model is constructed as a double well function at low temperature, such that
the external energy can be absorbed during the phase transformation and converted
into thermal form. The Chebyshev spectral methods are employed together with
backward differentiation for the numerical analysis of the problem. Computational
experiments performed for different vibration energies demonstrate the importance
of taking into account damping effects induced by phase transformations.
Key words: Martensite transformation, thermo-mechanical coupling, vibration
damping, Ginzburg-Landau theory.
1
Introduction
Shape Memory Alloy (SMA) materials are able to directly transduce thermal energy into
mechanical and vice versa. Their unique properties make them very attractive in many
∗
Corresponding Address:Tel.:+45 6550 1686, Fax:+45 6550 1660, E-mail:[email protected]
1
engineering applications, including mechanical and control engineering, biomedicine, communication, robotics and so on [1]. Among these applications, dampers made from
SMA for passive and semi-active vibration damping, are quoted perhaps most frequently
[2]. These devices exhibit fairly complicated nonlinear (hysteretic) behaviour induced by
martensite phase transformations. An appropriate mathematical model involving phase
transformations and thermo-mechanical coupling is essential for a better understanding
of the dynamic behavior of SMA dampers.
The damping effects of SMA cannot be understood without incorporating into the
model the dynamics of the first order martensite transformations induced by mechanical
loadings, and hysteresis as a consequence. SMAs have more than one martensite variants
at low temperature that correspond to the same elastic potential energy of the material
stored via deformation. When the martensite phase transformation is induced mechanically, part of the external energy will be consumed to transform the material from one
variant to another, without increasing the total elastic energy stored in the material, but
the deformation will be oriented in a way favored by the external loadings. The work
done by the external loadings during the phase transformation process cannot be restored
reversely when the loading is removed, since it is converted into thermal form, and will be
dissipated eventually via, for example, thermal dissipation. This dissipation of mechanical energy due to phase transformation can be measured by the hysteretic behaviour of
the material under mechanical loadings [3, 4]. Because of the nonlinear nature of phase
transformations and the nonlinear coupling between the mechanical and thermal fields,
the dynamic behavior of SMA dampers becomes very complex.
Some investigations have been already carried out to understand the damping effects
of SMA at either micro-scale or macro-scale [4, 5, 6]. Among many efforts, it was shown
in Ref. [4, 7] ( and references therein) that the damping effects of SMA are influenced by
the vibration frequency, heating/cooling rate (temperature change rate), and vibration
amplitude . This influence is discussed and analyzed in terms of strain formation, interface
moving, internal friction, and other factors, which are all considered either at micro-scale
or mesoscale.
For most current engineering applications of SMA dampers, damping effects due to
hysteresis is more pronounced at macroscale, and one needs to model the dynamical
behaviour and damping effects of SMAs at macroscale, which demands that the model
should have the capability to capture all the contributions to vibration damping, particularly that due to hysteresis induced by cyclic mechanical loadings [3, 7]. Since vibration
damping can be induced by hysteretic behaviour of the SMAs, interface moving, internal
friction, and thermo-mechanical coupling, the model for damping devices made from SMA
has to be constructed on the basis of dynamics of the materials, which involves the phase
2
transformation and thermo-mechanical coupling. Progressively, the impact induced phase
transformation in SMA wires were investigated in [6, 7] where the constitutive laws were
approximated linearly, and the thermo-mechanical coupling was neglected. Such models
have obvious limitations.
In what follows, we attempt to overcome such limitations by better capturing the
thermo-mechanical coupling, hysteresis, and nonlinear nature of phase transformations
with the Ginzburg-Landau theory, originally discussed in the context of SMAs in [8,
9] and later applied to model the dynamical behavior of SMA rods under dynamical
loadings (e.g., [10, 11, 12] and references therein). In particular, in this paper we employ
a mathematical model based on the modified Ginzburg - Landau theory for the SMA
damper, the dynamics of the SMA damper are described by Navier’s equation with a
non-convex constitutive relation. The damper is connected with a mass block with a
given initial velocity. The movement of the mass block is then simulated as a single
degree of freedom system, subject to the damping force from the SMA rod.
2
Mathematical Modelling
The physical problem investigated here is sketched in Figure (1). There is a mass block
connected to a SMA rod. The SMA rod has a cross section area of β and length L,
and is initially at rest. The block has an initial velocity vm . The purpose of the current
investigation is to show the damping effect of the SMA rod on the vibration of the mass
block.
In the current damping system, the mass block is a lumped sub-system, while the SMA
rod is a distributed sub-system. Here, we construct the model describing their movements
separately and then couple them together. For the mass block, the model can be easily
formulated using the momentum conservation law as:
mẍm + µm ẋm + km (xm − L) = fm ,
(1)
where xm is the position of the block, µm the friction coefficient, km the spring stiffness,
m the mass of the block, and fm is the external force on the mass block which is caused
by the SMA rod discussed below.
The mathematical model for the dynamics of the SMA rod can be constructed on the
basis of the modified Ginzburg - Landau theory, as done in ([8, 9, 10, 12, 14] and references
therein). For the convenience of discussion about damping and energy absorbing here,
the model is constructed using the extended Hamilton’s principle, but using the same
non-convex potential energy.
3
In order to formulate the governing equation for the mechanical field, the Lagrangean
L is introduced for material particles as follows:
ρ
(2)
L = (u̇)2 − F ,
2
where ρ is the density of the material and F is the potential energy function of the
material. The differntiating feature of the Ginzburg - Landau theory is that the potential
energy function is constructed as a non-convex function of the chosen order parameters
and temperature θ. It is a sum of local energy density (Fl ) and non-local energy density
is chosen as the
(Fg ). For the current one-dimensional problem, the strain ε(x, t) = ∂u
∂x
order parameter, and the local free energy function can be constructed as the Landau free
energy function Fl (θ, ε) [8, 9, 10, 11, 16]:
k1 (θ − θ1 ) 2 k2 4 k3 6
ε + ε + ε,
(3)
2
4
6
where k1 , k2 , and k3 are material-specific constants, θ1 is the reference transformation
temperature.
The non-local free energy function is usually constructed in a way similar to the class
of ferroelastic materials as follows [10, 11]:
Fl (θ, ε) =
∂ε
1
(4)
Fg (∇ε) = kg ( )2 ,
2 ∂x
where kg is a material-specific constant. The non-local term above accounts for inhomogeneous strain field. It represents energy contributions from domain walls of different phases
(hence, an analogy with the Ginzburg term in semiconductors). In order to account for
dissipation effects accompanying phase transformations, a Rayleigh dissipation term is
introduced here as follows [13]:
1 ∂ε
(5)
FR = − ν( )2 ,
2 ∂t
where ν is a material-specific constant. The above dissipation term accounts for the
internal friction. At macroscale, it is translated into the viscous effects [18].
By substituting the potential energy function into the Lagrangian function given in
Equation (2), and using the extended Hamilton’s principle, the governing equation for
the dynamics of the mechanical field can be obtained as follows:
3
5 !
∂u
∂ ∂2u
∂4u
∂u
∂u
∂
k1 (θ − θ1 )
+ν
+ k2
−
k
.
(6)
+ k3
ρü =
g
∂x
∂x
∂x
∂x
∂t ∂x2
∂x4
In order to formulate the governing equation for the thermal field, the conservation
law of the internal energy is employed here:
ρ
∂e ∂q
∂ε
∂ε ∂ε
∂2ε
+
−σ −ν
− kg
= 0,
∂t ∂x
∂t
∂t ∂t
∂x∂t
4
(7)
where e is the internal energy, q = −k∂θ/∂x is the (Fourier) heat flux, k is the heat
conductance coefficient of the material, and σ is the stress. In order to capture the
coupling between the mechanical and thermal fields, the internal energy is associated
with the non-convex potential energy mentioned above via the Helmholtz free energy
function as follows [8, 9, 10]:
H(θ, ε) = F − cv θ ln θ
(8)
and the thermodynamical equilibrium condition gives:
e=H−θ
∂H
,
∂θ
σ=
∂H
,
∂ε
(9)
where cv is the specific heat capacitance. By substituting the above relationships into
Equation (7), the governing equation for the thermal field can be formulated as:
∂θ
∂2θ
∂ε
cv
= k 2 + k1 θε + ν
∂t
∂x
∂t
∂ε
∂t
2
.
(10)
It is shown clearly from the above derivation that the governing equations for both
the mechanical and thermal fields are constructed using the same potential function
Fl (θ, ε) + Fg (∇ε), which is taken here as the Ginzburg-Landau free energy function. It
has been shown in Ref.([13, 14, 15]) that the mathematical model such as the one based
Equation (6) and (10) is capable to capture the first order phase transformations in ferroelastic materials, and the thermo-mechanical coupling, by suitably choosing the coefficients
according experiment results [10, 11, 15], as shown in section 3. The thermo-mechanical
coupling is also shown by the term k1 (θ − θ1 )ε, which accounts for the influence of the
2
+ ν ∂ε
account
thermal field on the mechanical one, and the other two terms k1 θε ∂ε
∂t
∂t
for the conversion of mechanical energy into thermal form. Due to strong nonlinearity in
the mechanical field and nonlinear coupling between the mechanical and thermal fields,
numerical simulations based on this model is far from trivial [9, 10].
For the sake of convenience of numerical analysis discussed later on, the model is recast
into a differential-algebraic system in a way similar to [9, 14]:
∂σ
∂4u
∂2u
=
−
k
+ F,
g
∂t2
∂x
∂x4
2
∂2θ
∂ε
∂ε
∂θ
= k 2 + k1 θε + ν
+ G,
cv
∂t
∂x
∂t
∂t
∂ε
σ ∗ = k1 (θ − θ1 )ε − k2 ε3 + k3 ε5 + ν .
∂t
ρ
(11)
where u is the displacement, θ is the temperature, ρ is the density, k1 , k2 , k3 , cv , kg , ν and
k are normalized material-specific constants, σ ∗ is the effective stress incorporating the
5
viscous effects, and F and G are distributed mechanical and thermal loadings, which are
all zero for the current problem.
To couple the mathematical model for the SMA rod and that of the block, boundary
conditions for Equation (11) should be associated with the movement of the mass block,
which can be formulated as follows:
∂θ
= 0, u = 0, at x = 0,
∂x
(12)
∂θ
fm
= 0, σ(L) =
, at x = L.
∂x
β
It is worthy to note that the above boundary conditions are not sufficient for the
system to have a unique solution. Indeed, when the nonlinear constitutive law is used, as
given in the last line of Equation (11), the relationship between σ and ε is no longer a
one to one map. For a given stress value, one might be able to find three possible strain
values which all satisfy the constitutive law. Therefore, an additional boundary condition
on u or ε is necessary to guarantee uniqueness. Here we employ the idea given in Ref.
[10] and set the additional boundary condition as ∂ε/∂x = 0. It is shown in Ref.([15, 18])
that, provided with such boundary conditions, the model has an unique solution and can
be numerically analyzed.
The geometrical coupling can be written as xm = L + u(L), so the velocity of the
mass block can be written as vm = ẋm = u̇(L). By putting the two mathematical models,
Equation (1) and Equation (11), and their coupling conditions together, our mathematical
model can be formulated finally as:
∂σ
∂4u
∂2u
=
−
k
,
g
∂t2
∂x
∂x4
2
∂2θ
∂ε
∂ε
∂θ
= k 2 + k1 θε + ν
cv
,
∂t
∂x
∂t
∂t
∂ε
σ ∗ = k1 (θ − θ1 )ε − k2 ε3 + k3 ε5 + ν ,
∂t
∂ε
∂θ
= 0, u = 0,
= 0, at x = 0,
∂x
∂x
1
∂θ
= 0, σ = (mü(L) + µm u̇(L) + km u(L)),
∂x
β
ρ
(13)
∂ε
= 0,
∂x
at x = L.
The above model incorporates the phase transformation, thermo-mechanical coupling,
viscous effects, and interfacial contributions, it is natural to expect that the model is able
to model the vibration damping effects due to these contributions.
3
Vibration Damping Due to Phase Transformation
As mentioned earlier, damping effects in SMA rod can be attributed to several factors.
The internal friction (viscous effect) contributes to the damping effects when phase trans6
formation is taking place. Since the occurence of thermo-mechanical coupling, the SMA
materials also have damping effects due to the energy conversion between thermal and
mechanical fields. The most pronouncing contribution to the damping effects in the SMA
rod is the mechanical hysteresis due to phase transformation between martensite variants.
Unlike other contributions to the damping effects, the contribution due to hysteresis can
be graphically explained using the non-convex potential energy function and the nonconvex constitutive relation.
In the heart of the Ginzburg-Landau theory for the first order phase transformation
in SMA, lies the free energy function defined as a non-convex function of the chosen order
parameter for characterization of all the phases and variants (e.g., [10, 8, 12, 16]), which
is defined as Equation (3). One example of the Landau free energy function is shown
in Figure (2), for which the material and its physical parameters are given in section 5.
It is shown clearly in the figure that the function has only one local minimum at high
temperature (θ = 280K), which is associated with the only stable phase (austenite) in
this case. At lower temperature (θ = 210K), there are two symmetrical local minima
(marked with small gray rectangles), which are associated with martensite plus (P) and
minus (M), respectively. For vibration damping purpose, it is always beneficial to make
the SMA work at low temperature, so the hysteresis loop will be larger.
The hysteretic behaviour and mechanical energy dissipation of the SMA rod at low
temperature can be explained using its nonlinear constitutive relation as shown in Figure
(3), which is obtained using the thermodynamical equilibrium condition [10]:
σ=
∂F
∂ε
(14)
For a given external loading, it will induce an internal stress and deformation in the
SMA rod, part of the work done by the loading is stored in the rod via its deformation
and can be calculated as:
Z εe
(−σε)dε,
(15)
W =
εs
where εs and εe are starting and ending strain values, respectively.
For pure elastic material, the stored energy can be released without loss when the
external loading is removed, there is no mechanical energy dissipation. For thermomechanical coupling material, the stored energy is also obtainable when the loading is
removed, but part of the input energy will be converted into thermal form and can not be
fully released. If the material is insulated, its temperature will be slightly increased due
to the input of mechanical energy. When viscous effects of the material are taken into
account, the fraction of mechanical energy converted into thermal form will be larger,
which means that dissipation of mechanical energy into thermal form is enhanced by
7
viscous effects. But, compared to those due to mechanical hysteresis, all these energy
dissipation are not remarkable.
The energy dissipation due to hysteresis can be presented schematically by considering
the mechanical energy loss (converted from mechanical to thermal) in one cycle of loading
[7, 12]. Assume that the loading process starts from point A in Figure (3), and the load
is continuously increased to point E. Then, the constitutive relation in this case will
be the curve ABCDE, and the work done by the loading is represented by the area of
ABCDERA. When the load is decreased from E to A, the system will follow another
constitutive curve EDF BA, and the work done by the rod is represented by the area
of EDF BARE. There is a difference between the two areas, which is enclosed by the
curve BCDF B and hatched in the figure. This area represents the mechanical energy
loss due to the hysteretic behaviour. At the same time, due to the thermo-mechanical
coupling, the temperature of the material will be increased, as indicated by the source
2
∂t
+ ν ∂ε
in the energy equation.
terms k1 θε ∂t
∂t
The physics behind this is that, when the external energy is applied to the SMA rod,
part of the energy will be demanded by the material to convert itself from one variant
(P) to another (M), as sketched in Figure (2). Both P and M are stable states, and
have the same potential, so the material will not recover its previous configuration when
the loading is removed. In this sense, the SMAs have plastic-like property. Part of the
input mechanical energy is consumed just for converting the material from one variant to
another (martensite phase transformation), and finally converted into thermal energy via
the thermo-mechanical coupling. If one want to reverse the transformation, extra input
energy is demanded.
4
Numerical Methodology
The constructed model for the vibration damping system given in Equation (13) is a
strongly nonlinear coupled system. Experiences gained from the numerical experiments
on similar problems shows that the development of an efficient algorithm is necessary for
the numerical analysis of the model because standard iteration algorithms have difficulties
to cope with the very strong nonlinearity [10, 12, 14, 16]. For the current problem, an
additional difficulty is that the boundary conditions are system dependent. Indeed, we
observe that the stress on the right end of the rod is a function of its displacement,
velocity, and acceleration. In order to deal with these difficulties, the model is reduced
to a differential-algebraic system and solved by using backward differentiation in time.
The practical implementation is done similarly to [9, 14, 16]. Note, however, that here for
the Chebyshev pseudo-spectral approximation, a set of Chebyshev points {xi } are chosen
8
along the length direction as follows:
πi
xi = L 1 − cos( ) /2,
N
i = 0, 1, . . . , N.
(16)
Using these nodes, ε, v, θ, and σ distributions in the rod can be expressed in terms of
the following linear approximation:
f (x) =
N
X
fi φi (x),
(17)
i=0
where f (x) stands for ε, v, θ, or σ, and fi is the function value at xi . Function φi (x) is
the ith interpolating polynomial which has the following property:
(
1, i = j,
(18)
φi (xj ) =
0, i 6= j.
It is easy to see that the Lagrange interpolants would satisfy the the above requirements. Having obtained f (x) approximately, the derivative ∂f (x)/∂x can be obtained by
taking the derivative of the basis functions φi (x) with respect to x:
N
X ∂φi (x)
∂f
=
fi
.
∂x
∂x
i=1
(19)
All these approximations are formulated in matrix form, for the convenience of actual
programming implementation. For approximation to higher order derivatives, similar
matrix form can be easily obtained. By substituting all the approximation into the DAE
system, it will be recast into a set of nonlinear algebraic equations, which can be solved
by Newton-Raphson iteration, following a similar way as done in Ref. ([9, 10, 14, 16]).
5
Numerical Experiments
Several different numerical experiments have been carried out to demonstrate the damping
effects of hysteresis in the SMA rod induced by the phase transformations. All experiments
reported in this section have been performed on a Au23 Cu30 Zn47 rod, with length of 1 cm.
Most of the physical parameters for this specific material can be found in the literature
(e.g., [8, 14, 15, 16]), which are listed as follows for the sake of convenience:
k1 = 480 g/ms2cmK,
θ1 = 208K,
k2 = 6 × 106 g/ms2cmK,
ρ = 11.1g/cm3 ,
cv = 3.1274g/ms2cmK,
k3 = 4.5 × 108 g/ms2 cmK,
k = 1.9 × 10−2 cmg/ms3 K.
There are two coefficients, kg and ν, are not easy to find its value. Normally, kg is
very small and here we take it as 5 by referring to the value indicated in Ref.([10, 11]).
9
The value of ν is take as 10, with the same unit system. To demonstrate the damping
effects of the SMA rod, we set µm = km = 0 in Equation (13). Then, only the force from
the SMA rod is taken into account.
For all numerical experiments, the initial conditions for the rod are chosen as θ = 210o ,
ε = 0.115, u = v = 0, the simulation time span is [0, 4]ms, and 40 nodes have been used
for the Chebyshev approximation.
In the first experiment reported here, the parameter m/β is chosen as 200 (g/cm2 )
and the initial value of vm is chosen −3 (cm/ms). The numerical results are presented in
Figure (4).
From the upper left plot it can be observed that the velocity of the block is damped
effectively, and gradually the velocity tends to a small value. The entire damping process
can be roughly divided into two stages. The first stage is t ∈ [0, 2.6]ms in which the
damping is more effective since the vibration energy is still large enough to switching the
whole SMA rod between martensite plus and minus at low temperature (martensite phase
transformation). While in the the second stage, the contribution from phase transformation becomes weaker and finally fades because the temperature is raised, the mechanical
energy is also dissipated and not large enough to induce phase transformation. The motion
of the rod finally becomes a thermo-mechanical vibration without phase transformations
at the end of the simulation. The strain distribution, presented in the lower right part of
the plot, shows this clearly. In the lower left plot, we present the average temperature of
the rod, which could be regarded as a measurement of thermal energy. Since insulation
thermal conditions are used and there is no thermal dissipation included in the model,
the temperature is increased rapidly when there is phase transformation induced because
the large change rate of strain, and much slower when there is no transformation. This
observation indicates that the conversion of mechanical energy into thermal form is much
faster when phase transformation takes place. The oscillation of average temperature is
caused by the phase transformation and thermo-mechanical coupling.
To analyze the damping effects further, we have increased the initial vibration energy
to m/β = 500. The numerical results for this case are presented in Figure (5), in a
similar way as those in Figure (4). It is seen from the strain plot that the whole rod is
switched between compressed (ε < 0) and stretched (ε > 0) states in the entire simulation
range, and vm is decreased continuously. This strain switching causes the temperature
oscillation. The dissipation of mechanical energy is faster at the beginning and becomes
slower after a while in the simulation, which is indicated by the velocity plot (consider
that mechanical energy is proportional to the square of velocity). This is due to the fact
that phase transformation only takes place in a short period in which the rod temperature
is still low, which can be roughly read as [0, 1.4]ms form the temperature plot, in which
10
the temperature increases faster. Afterwards, there is no phase transformation since the
rod is at high temperature state and the motion becomes thermo-mechanical vibration
again. This observation leads to same conclusion that phase transformation contribute
most to the damping effects (energy conversion).
To show the damping effects contributed from the viscous effects and thermo-mechanical
coupling, the last numerical example presented here has been dealing with the analysis
of the damping effects when there is no phase transformation induced in the SMA rod
at all. For this purpose, the parameters are chosen as m/β = 20, and initial velocity
vm = −1 such that the input mechanical energy is too small to induce phase transformation, and the simulation time span is also extented to [0, 20]ms to capture the long
time behaviour since the mechanical energy dissiption might be much slower. For the
simulation of damping effects contributed from thermo-mechanical coupling alone, the
parameter is set ν = 0 (no viscous effects), and the numerical results are presented on
the left column of Figure (6), by plotting the evolution of mass block velocty (top) and
average temperature of the SMA rod (bottom). It indicates that the coupling alone does
contribute to damping effects, the mechanical energy stored via the mass block motion
is finally converted into thermal form, and increase the SMA rod temperature slightly.
But, the dissipation of mechanical energy into thermal form is rather slow, there are still
oscillations at the end of the simulation. In order to include the contribution from viscos
effects, now the viscosity parameter is set ν = 20, numerical results are presented on
the right column similarly in the figure. It is shown clearly that when viscous effects
are incorportate (right), the conversion of energy due to thermo-mechanical coupling is
enhanced, the the system is already at rest at t = 15ms. In both case, the oscillation
of velocity and average temperature has a series of peaks, which indicates a complicate
nature of the thermo-mechanical energy conversion.
In all these three numerical experiments, the energy conversion between mechanical
and thermal form due to thermo-mechanical coupling is well captured. It is shown that
when the phase transformation is induced, the vibration damping is remarkabley enhanced
because the energy conversion is enhanced. Because there is no heat loss considered here,
the average temperature of the rod increases continuously. It is clear that if the SMA
damper is controlled at low temperature by external efforts, it will be an effective damper.
6
Conclusions
In this paper, we constructed a mathematical model for vibration damping of a mass
block connected with a SMA rod. We employed the modified Ginzburg - Landau theory
for the mathematical modelling of the SMA rod to capture the phase transformation and
11
thermo-mechanical coupling. The model for the vibration of the block is coupled with
the dynamics of the SMA rod by adjusting the boundary conditions. The model is then
numerically analyzed and the damping characteristics of the SMA rod due to the first
order martensite phase transformation are investigated. It is shown that the vibration
can be effectively damped, if the first order phase transformation is induced.
References
[1] Birman,V., Review of mechanics of shape memory alloys structures, Appl.Mech.Rev.
50 (1997) 629-645.
[2] Song, G., Ma, N., and Li, H. N., Application of shape memory alloys in civil structures,
Engineering Structures 28 (2006) 1266 - 1274.
[3] Juan, J. S., and No, M. L., Damping behaviour during martensite transformation in
shape memory alloys, Journal of Alloys and Compounds 355 (2003) 65 - 71.
[4] Humbeeck, J. V., Damping capacity of thermoelastic martensite shape memory alloys,
Journal of Alloys and Compounds, 355 (2003) 58 -64
[5] Piedboeuf, M. C., Gauvin, R., and Thomas, M., Damping behaviour of shape memory
alloys: Strain amplitude, frequency and temperature effects. Journal of Sound and
Vibration 214 (5) (1998) 885 - 901
[6] Chen, Y.C., Lagoudas, D. C., Impact induced phase transformation in shape memory
alloys, Journal of the Mechanics and Physics of Solids, 48 (2000) 275 - 300.
[7] Masuda, A., and Noori, M. Optimisation of hysteretic characteristics of damping
devices based on pseudoelastic shape memory alloys. Inter. J. Nonlinear Mech. 37
(2002) 1375-1386.
[8] Falk, F., Model free energy, mechanics, and thermomechanics of shape memory alloys.
Acta Metallurgica, 28, (1980) 1773-1780.
[9] Matus, P., Melnik, R., Wang, L., Rybak, I.: Application of fully conservative schemes
in nonlinear thermoelasticity: Modelling shape memory materials. Mathematics and
Computers in Simulation. Mathematics And Computers in Simulation 65 4-5 489-509
[10] Bunber, N., Landau-Ginzburg model for a deformation-driven experiment on shape
memory alloys. Continuum Mech. Thermodyn. 8 (1996) 293-308.
12
[11] Bubner, N., Mackin, G., and Rogers, R. C., Rate depdence of hysteresis in onedimensional phase transitions, Comp. Mater. Sci., 18 (2000) 245 - 254.
[12] Koucek, P, Reynolds, D. R., and Seidman, T. I., Computational modelling of vibration damping in SMA wires, Continuum Mech. Thermodyn. 16 (2004) 495 - 514.
[13] Bales, G.S. and Gooding, R.J., Interfacial dynamics at a 1st-order phase-transition
involving strain dynamic twin formation. Phys. Rev. Lett.,67, (1991) 3412.
[14] Melnik, R., Roberts, A., Thomas, K.: Phase transitions in shape memory alloys
with hyperbolic heat conduction and differential algebraic models. Computational
Mechanics 29 (1) (2002) 16-26.
[15] Niezgodka, M., Sprekels, J.: Convergent numerical approximations of the thermomechanical phase transitions in shape memory alloys. Numerische Mathematik 58(1991)
759-778.
[16] Wang, L.X, and R.V.N.Melnik, Thermomechanical waves in SMA patches under
small mechanical loadings. in Lecture Notes in Computer Science 3039, (2004), 645652.
[17] Chaplygin, M. N., Darinskii, B. M., and Sidorkin, A. S., Nonlinear Waves in Ferroelastics, Ferroelectrics, 307 (2004) 1-6.
[18] Abeyaratne, R;, and Knowles, J. K., On a shock-induced martensitic phase transition,
J. Appl. Phys. 87 (2000), 1123 - 1134.
13
x
Km
m, vm
fm(t)
L
Figure 1: Vibration damping of a mass block connected to a shape memory alloy rod.
200
150
θ = 280 K
150
Landau Free Energy
Landau Free Energy
100
100
50
θ = 245 K
0
θ = 230 K
50
θ = 210 K
0
−50
−50
θ = 210 K
M
P
−100
−0.15
−0.1
−0.05
0
ε
0.05
0.1
0.15
−100
−0.15
−0.1
−0.05
0
ε
0.05
0.1
0.15
Figure 2: Non-convex free energy and its dependence on temperature
4000
E
3000
2000
D
C
σ
1000
0
−1000
B
F
−2000
−3000
−4000
−0.2
R
A
−0.15
−0.1
−0.05
ε
0
0.05
0.1
0.15
Figure 3: hysteretic behaviour and its damping effect caused by the non-convex free
energy
14
0.2
3
0.15
2
Displacement [cm]
Velocity [cm/ms]
0.1
1
0
−1
0.05
0
−0.05
−0.1
−2
−3
0
−0.15
0.5
1
1.5
2
2.5
3
3.5
−0.2
0
4
Time [ms]
0.5
1
1.5
2
2.5
Time [ms]
3
3.5
4
260
250
θ [K]
240
230
220
210
200
0
0.5
1
1.5
2
2.5
Time [ms]
3
3.5
4
Figure 4: Damping effect of a SMA rod involving mechanically induced phase transformations
15
3
0.2
0.15
2
Displacement [cm]
Velocity [cm/ms]
0.1
1
0
−1
0.05
0
−0.05
−0.1
−2
−3
0
−0.15
0.5
1
1.5
2
2.5
Time [ms]
3
3.5
−0.2
0
4
1
2
3
4
5
Time [ms]
320
300
θ [K]
280
260
240
220
200
0
0.5
1
1.5
2
2.5
Time [ms]
3
3.5
4
Figure 5: Damping effect of a SMA rod with large initial vibration energy
16
0.8
0.6
0.6
0.4
0.4
0.2
0.2
Velocity [cm/ms]
Velocity [cm/ms]
ν =20
0.8
0
−0.2
−0.4
0
−0.2
−0.4
−0.6
−0.6
−0.8
−0.8
−1
0
5
10
Time [ms]
15
−1
0
20
214
214
212
212
θ [K]
θ [K]
216
210
208
206
206
10
Time [ms]
15
20
15
20
210
208
5
10
Time [ms]
ν=20
ν =0
216
204
0
5
15
20
204
0
5
10
Time [ms]
Figure 6: Effects of internal friction on damping performance of shape memory alloy rods
(no phase transformations)
17
0.125
Displacement [cm]
0.12
0.115
0.11
0.105
0.1
0
0.5
1
1.5
2
2.5
Time [ms]
3
3.5
4
215
214
213
θ [K]
212
211
210
209
208
207
206
0
0.5
1
1.5
2
2.5
Time [ms]
3
3.5
4
0.8
0.6
Velocity [cm/ms]
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
0
0.5
1
1.5
2
2.5
Time [ms]
3
3.5
4
| 5cs.CE
|
The Likelihood Ratio Test in High-Dimensional Logistic
Regression Is Asymptotically a Rescaled Chi-Square
arXiv:1706.01191v1 [math.ST] 5 Jun 2017
Pragya Sur∗
Yuxin Chen†
Emmanuel J. Candès∗‡
June 2017
Abstract
Logistic regression is used thousands of times a day to fit data, predict future outcomes, and assess the
statistical significance of explanatory variables. When used for the purpose of statistical inference, logistic
models produce p-values for the regression coefficients by using an approximation to the distribution of
the likelihood-ratio test. Indeed, Wilks’ theorem asserts that whenever we have a fixed number p of
variables, twice the log-likelihood ratio (LLR) 2Λ is distributed as a χ2k variable in the limit of large
sample sizes n; here, χ2k is a chi-square with k degrees of freedom and k the number of variables being
tested. In this paper, we prove that when p is not negligible compared to n, Wilks’ theorem does not
hold and that the chi-square approximation is grossly incorrect; in fact, this approximation produces
p-values that are far too small (under the null hypothesis).
Assume that n and p grow large in such a way that p/n → κ for some constant κ < 1/2. We prove that
d
for a class of logistic models, the LLR converges to a rescaled chi-square, namely, 2Λ → α(κ)χ2k , where
the scaling factor α(κ) is greater than one as soon as the dimensionality ratio κ is positive. Hence, the
LLR is larger than classically assumed. For instance, when κ = 0.3, α(κ) ≈ 1.5. In general, we show how
to compute the scaling factor by solving a nonlinear system of two equations with two unknowns. Our
mathematical arguments are involved and use techniques from approximate message passing theory, from
non-asymptotic random matrix theory and from convex geometry. We also complement our mathematical
study by showing that the new limiting distribution is accurate for finite sample sizes.
Finally, all the results from this paper extend to some other regression models such as the probit
regression model.
Keywords. Logistic regression, likelihood-ratio tests, Wilks’ theorem, high-dimensionality, goodness of
fit, approximate message passing, concentration inequalities, convex geometry, leave-one-out analysis
1
Introduction
Logistic regression is by far the most widely used tool for relating a binary response to a family of explanatory
variables. This model is used to infer the importance of variables and nearly all standard statistical softwares
have inbuilt packages for obtaining p-values for assessing the significance of their coefficients. For instance,
one can use the snippet of R code below to fit a logistic regression model from a vector y of binary responses
and a matrix X of covariates:
fitted <- glm(y ~ X+0, family = ‘binomial’)
pvals <- summary(fitted)$coefficients[,4]
∗ Department
of Statistics, Stanford University, Stanford, CA 94305, U.S.A.
of Electrical Engineering, Princeton University, Princeton, NJ 08544, U.S.A.
‡ Department of Mathematics, Stanford University, Stanford, CA 94305, U.S.A.
† Department
1
The vector pvals stores p-values for testing whether a variable belongs to a model or not, and it is well known
that the underlying calculations used to produce these p-values can also be used to construct confidence
intervals for the regression coefficients. Since logistic models are used hundreds of times every day for
inference purposes, it is important to know whether these calculations—e.g. these p-values—are accurate
and can be trusted.
1.1
Binary regression
Imagine we have n samples of the form (yi , Xi ), where yi ∈ {0, 1} and Xi ∈ Rp . In a generalized linear
model, one postulates the existence of a link function µ(·) relating the conditional mean of the response
variable to the linear predictor Xi> β,
E[yi |Xi ] = µ(Xi> β),
(1)
where β = [β1 , β2 , . . . , βp ]> ∈ Rp is an unknown vector of parameters. We focus here on the two most
commonly used binary regression models, namely, the logistic and the probit models for which
(
et /(1 + et ) in the logistic model,
µ(t) :=
(2)
Φ(t)
in the probit model;
here, Φ is the cumulative distribution function (CDF) of a standard normal random variable. In both cases,
the Symmetry Condition
µ(t) + µ(−t) = 1
(3)
holds, which says that the two types yi = 0 and yi = 1 are treated in a symmetric fashion. Assuming that
the observations are independent, the negative log-likelihood function is given by [1, Section 4.1.2]
n
X
µi
+ log (1 − µi ) ,
µi := µ(Xi> β).
yi log
` (β) := −
1
−
µ
i
i=1
Invoking the symmetry condition, a little algebra reveals an equivalent expression
Xn
` (β) :=
ρ − ỹi Xi> β ,
i=1
where
(
1
if yi = 1,
ỹi :=
−1 if yi = 0,
(
and
ρ(t) :=
log (1 + et ) in the logistic case,
− log Φ (−t) in the probit case.
(4)
(5)
Throughout we refer to this function ρ as the effective link.
1.2
The likelihood-ratio test and Wilks’ phenomenon
Researchers often wish to determine which covariates are of importance, or more precisely, to test whether
the jth variable belongs to the model or not: formally, we wish to test the hypothesis
Hj :
βj = 0 versus
βj 6= 0.
(6)
Arguably, one of the most commonly deployed techniques for testing Hj is the likelihood-ratio test (LRT),
which is based on the log-likelihood ratio (LLR) statistic
Λj := ` β̂(−j) − ` β̂ .
(7)
Here, β̂ and β̂(−j) denote respectively the maximum likelihood estimates (MLEs) under the full model and
the reduced model on dropping the jth predictor; that is,
β̂ = arg minp `(β)
β∈R
β̂(−j) = arg
and
2
min
β∈Rp ,βj =0
`(β).
Inference based on such log-likelihood ratio statistics has been studied extensively in prior literature [12, 37,
52]. Arguably, one of the most celebrated results in the large-sample regime is the Wilks’ theorem.
To describe the Wilk’s phenomenon, imagine we have a sequence of observations (yi , Xi ) where yi ∈ {0, 1},
Xi ∈ Rp with p fixed. Since we are interested in the limit of large samples, we may want to assume that the
covariates are i.i.d. drawn from some population with non-degenerate covariance matrix so that the problem
is fully p-dimensional. As before, we assume a conditional logistic model for the response. In this setting,
Wilks’ theorem [52] calculates the asymptotic distribution of Λj (n) when n grows to infinity:
(Wilks’ phenomenon) Under suitable regularity conditions which, for instance, guarantee that the
MLE exists and is unique,1 the LLR statistic for testing Hj : βj = 0 vs. βj 6= 0 has asymptotic
distribution under the null given by
d
2Λj (n) → χ21 ,
as n → ∞.
(8)
This fixed-p large-n asymptotic result, which is a consequence of asymptotic normality properties of the
MLE [50, Theorem 5.14], applies to a much broader class of testing problems in parametric models; for
instance, it applies to the probit model as well. We refer the readers to [34, Chapter 12] and [50, Chapter
16] for a thorough exposition and details on the regularity conditions under which Wilks’ theorem holds.
Finally, there is a well-known extension which states that if we were to drop k variables from the model,
then the LLR would converge to a chi-square distribution with k degrees of freedom under the hypothesis
that the reduced model is correct.
1.3
Inadequacy of Wilks’ theorem in high dimensions
The chi-square approximation to the distribution of the LLR statistic is used in standard statistical softwares
to provide p-values for the single or multiple coefficient likelihood ratio tests. Here, we perform a simple
experiment on synthetic data to study the accuracy of the chi-square approximation when p and n are
both decently large. Specifically, we set β = 0 and test β1 = 0 vs. β1 6= 0 using the LRT in a setting
i.i.d.
where p = 1200. In each trial, n = 4000 observations are produced with yi ∼ Bernoulli(1/2), and X :=
[X1 , · · · , Xn ]> ∈ Rn×p is obtained by generating a random matrix composed of i.i.d. N (0, 1) entries. We
fit a logistic regression of y on X using R, and extract the p-values for each coefficient. Figure 1 plots the
pooled histogram that aggregates 4.8 × 105 p-values in total (400 trials with 1200 p-values obtained in each
trial).
If the χ21 approximation were true, then we would expect to observe uniformly distributed p-values. The
histrogram from Fig. 1 is, however, far from uniform. This is an indication of the inadequacy of Wilks’
theorem when p and n are both large. The same issue was also reported in [11], where the authors observed
that this discrepancy is highly problematic since the distribution is skewed towards smaller values. Hence,
such p-values cannot be trusted to construct level-α tests and the problem is increasingly severe when we
turn attention to smaller p-values as in large-scale multiple testing applications.
1.4
The Bartlett correction?
A natural question that arises immediately is whether the observed discrepancy could be an outcome of
a finite-sample effect. It has been repeatedly observed that the chi-square approximation does not yield
accurate results with finite sample size. One correction to the LRT that is widely used in finite samples is
the Bartlett correction, which dates back to Bartlett [5] and has been extensively studied over the past few
decades (e.g. [8, 10, 13, 15, 33]). In the context of testing for a single coefficient in the logistic model, this
correction can be described as follows [38]: compute the expectation of the LLR statistic up to terms of
order 1/n2 ; that is, compute a parameter α such that
1
α
,
E[2Λj ] = 1 + + O
n
n2
1 Such
conditions would also typically imply asymptotic normality of the MLE.
3
12500
30000
15000
10000
7500
Counts
Counts
Counts
20000
10000
10000
5000
5000
2500
0
0
0.00
0.25
0.50
P−Values
0.75
1.00
0
0.00
(a)
0.25
0.50
P−Values
0.75
1.00
(b)
0.00
0.25
0.50
P−Values
0.75
1.00
(c)
Figure 1: Histogram of p-values for logistic regression under i.i.d. Gaussian design, when β = 0, n = 4000,
p = 1200, and κ = 0.3: (a) classically computed p-values; (b) Bartlett-corrected p-values; (c) adjusted
p-values.
which suggests a corrected LLR statistic
2Λj
1 + αnn
(9)
with αn being an estimator of α. With a proper choice of αn , one can ensure
2Λj
1
E
=1+O
1 + αnn
n2
in the classical setting where p is fixed and n diverges. In expectation, this corrected statistic is closer to a
χ21 distribution than the original LLR for finite samples. Notably, the correction factor may in general be a
function of the unknown β and, in that case, must be estimated from the null model via maximum likelihood
estimation.
In the context of GLMs, Cordeiro [13] derived a general formula for the Bartlett corrected LLR statistic,
see [14, 17] for a detailed survey. In the case where there is no signal (β = 0), one can compute αn for the
logistic regression model following [13] and [38], which yields
αn =
n
2
Tr Dp2 − Tr Dp−1
.
2
(10)
−1 >
>
Here, Dp is the diagonal part of X(X > X)−1 X > and Dp−1 is that of X(−j) X(−j)
X(−j)
X(−j) in
which X(−j) is the design matrix X with the jth column removed. Comparing the adjusted LLRs to a χ21
distribution yields adjusted p-values. In the setting of Fig. 1(a), the histogram of Bartlett corrected p-values
is shown in Fig. 1(b). As we see, these p-values are still far from uniform.
If the mismatch is not due to finite sample-size effects, what is the distribution of the LLR in high
dimensions? Our main contribution is to provide a very precise answer to this question; below, we derive the
high-dimensional asymptotic distribution of the log-likelihood ratios, i.e. in situations where the dimension
p is not necessarily negligible compared to the sample size n.
4
2
Main results
2.1
Modelling assumptions
In this paper, we focus on the high-dimensional regime where the sample size is not much larger than the
number of parameters to be estimated—a setting which has attracted a flurry of activity in recent years.
In particular, we assume that the number p(n) of covariates grows proportionally with the number n of
observations; that is,
p(n)
= κ,
(11)
lim
n→∞
n
where κ > 0 is a fixed constant independent of n and p(n). In fact, we shall also assume κ < 1/2 for both
the logistic and the probit models, as the MLE is otherwise at ∞; see Section 2.2.
To formalize the notion of high-dimensional asymptotics when both n and p(n) diverge, we consider a
sequence of instances {X(n), y(n)}n≥0 such that for any n,
• X(n) ∈ Rn×p(n) has i.i.d. rows Xi (n) ∼ N (0, Σ), where Σ ∈ Rp(n)×p(n) is positive definite;
ind.
• yi (n) | X(n) ∼ yi (n) | Xi (n) ∼ Bernoulli µ(Xi (n)> β(n)) , where µ satisfies the Symmetry Condition;
• we further assume β(n) = 0. From the Symmetry Condition it follows that µ(0) = 1/2, which directly
implies that y(n) is a vector with i.i.d Bernoulli(1/2) entries.
The MLE is denoted by β̂(n) and there are p(n) LLR statistics Λj (n) (1 ≤ j ≤ p(n)), one for each of the
p(n) regression coefficients. In the sequel, the dependency on n shall be suppressed whenever it is clear from
the context.
2.2
When does the MLE exist?
Even though we are operating in the regime where n > p, the existence of the MLE cannot be guaranteed
for all p and n. Interestingly, the norm of the MLE undergoes a sharp phase transition in the sense that
kβ̂k = ∞ if κ > 1/2
and
kβ̂k < ∞ if κ < 1/2.
Here, we develop some understanding about this phenomenon. Given that ρ(t) ≥ ρ(−∞) = 0 for both
the logistic and probit models, each summand in (4) is minimized if ỹi Xi> β = ∞, which occurs when
sign(Xi> β) = sign(ỹi ) and kβk = ∞. As a result, if there exists a nontrivial ray β such that
Xi> β > 0
if ỹi = 1
and
Xi> β < 0
if ỹi = −1
(12)
for any 1 ≤ i ≤ n, then pushing kβk to infinity leads to an optimizer of (4). In other words, the solution to
(4) becomes unbounded (the MLE is at ∞) whenever there is a hyperplane perfectly separating the two sets
of samples {i | ỹi = 1} and {i | ỹi = −1}.
Under the assumptions from Section 2.1, ỹi is independent of X and the distribution of X is symmetric.
Hence, to calculate the chance that there exists a separating hyperplane, we can assume ỹi = 1 (1 ≤ i ≤ n)
without loss of generality. In this case, the event (12) becomes
{Xβ | β ∈ Rp } ∩ Rn++ 6= ∅,
(13)
where Rn++ is the positive orthant. Write X = ZΣ1/2 so that Z is an n × p matrix with i.i.d. standard
Gaussian entries, and θ = Σ1/2 β. Then the event (13) is equivalent to
{Zθ | θ ∈ Rp } ∩ Rn++ 6= ∅.
5
(14)
Now the probability that (14) occurs is the same as that
{Zθ | θ ∈ Rp } ∩ Rn+ 6= {0}
(15)
occurs, where Rn+ denotes the non-negative orthant. From the approximate kinematic formula [3, Theorem
I] in the literature on convex geometry, the event (15) happens with high probability if and only if the total
statistical dimension of the two closed convex cones exceeds the ambient dimension, i.e.
δ ({Zθ | θ ∈ Rp }) + δ Rn+ > n + o(n).
(16)
Here, the statistical dimension of a closed convex cone K is defined as
δ(K) := Eg∼N (0,I) kΠK (g) k2
(17)
with ΠK (g) := arg minz∈K kg − zk the Euclidean projection. Recognizing that [3, Proposition 2.4]
δ ({Zθ | θ ∈ Rp }) = p and δ(Rn+ ) = n/2,
we reduce the condition (16) to
p + n/2 > n + o(n)
or
p/n > 1/2 + o(1),
thus indicating that kβ̂k = ∞ when κ = lim p/n > 1/2.
This argument only reveals that kβ̂k = ∞ in the regime where κ > 1/2. If κ = p/n < 1/2, then
kΣ1/2 β̂k = O(1) with high probability, a fact we shall prove in Section 5. In light of these observations we
work with the additional condition
κ < 1/2.
(18)
2.3
The high-dimensional limiting distribution of the LLR
In contrast to the classical Wilks’ result, our findings reveal that the LLR statistic follows a rescaled chisquare distribution with a rescaling factor that can be explicitly pinned down through the solution to a
system of equations.
2.3.1
A system of equations
We start by setting up the crucial system of equations. Before proceeding, we first recall the proximal
operator
1
(19)
proxbρ (z) := arg min bρ(x) + (x − z)2
x∈R
2
defined for any b > 0 and convex function ρ(·). As in [18], we introduce the operator
Ψ(z; b) := bρ0 (proxbρ (z)),
(20)
which is simply the proximal operator of the conjugate (bρ)∗ of bρ.2 To see this, we note that Ψ satisfies the
relation [18, Proposition 6.4]
Ψ(z; b) + proxbρ (z) = z.
(21)
The claim that Ψ(·; b) = prox(bρ)∗ (·) then follows from the Moreau decomposition
proxf (z) + proxf ∗ (z) = z,
2 The
∀z,
conjugate f ∗ of a function f is defined as f ∗ (x) = supu∈dom(f ) {hu, xi − f (u)}.
6
(22)
25.0
●
●
20.0
2.00
●
15.0
●
Rescaling Constant
Rescaling Constant
●
10.0
●
1.75
●
●
●
●
●
●
1.50
●
●
●
●
●
1.25
1.00
●●
●●
●●
●●●
0.00
0.05
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
5.0
●
●
●
●
1.0
0.10
0.15
0.20
κ
0.25
0.30
0.35
0.40
●
2.5
●
●●● ● ●
●●●●
●●●
●●●
●●●
●●
●●●
●●●
●●
●●
●●
●●
●
●
●●
●
●
●
●
●
●
0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45
κ
Figure 2: Rescaling constant τ∗2 /b∗ as a function of κ for the logistic model. Note the logarithmic scale in
the right panel. The curves for the probit model are nearly identical.
which holds for a closed convex function f [39, Section 2.5]. Interested readers are referred to [18, Appendix
1] for more properties of proxbρ and Ψ.
We are now in position to present the system of equations that plays a crucial role in determining the
distribution of the LLR statistic in high dimensions:
i
1 h
2
E (Ψ (τ Z; b)) ,
κ
κ = E Ψ0 (τ Z; b) ,
τ2 =
(23)
(24)
where Z ∼ N (0, 1), and Ψ0 (·, ·) denotes differentiation with respect to the first variable. The fact that this
system of equations would admit a unique solution in R2+ is not obvious a priori. We shall establish this for
the logistic and the probit models later in Section 6.
2.3.2
Main result
Theorem 1. Consider a logistic or probit regression model under the assumptions from Section 2.1. If
κ ∈ (0, 1/2), then for any 1 ≤ j ≤ p, the log-likelihood ratio statistic Λj as defined in (7) obeys
d
2Λj →
τ∗2 2
χ ,
b∗ 1
as n → ∞,
(25)
where (τ∗ , b∗ ) ∈ R2+ is the unique solution to the system of equations (23) and (24). Furthermore, the LLR
statistic obtained by dropping k variables for any fixed k converges to (τ∗2 /b∗ )χ2k . Finally, these results extend
to all binary regression models with links obeying the assumptions listed in Section 2.3.3.
Hence, the limiting distribution is a rescaled chi-square with a rescaling factor τ∗2 /b∗ that only depends
on the aspect ratio κ. Fig. 2 illustrates the dependence of the rescaling factor on the limiting aspect ratio κ
for logistic regression. The figures for the probit model are similar as the rescaling constants actually differ
by very small values.
To study the quality of approximation for finite samples, we repeat the same numerical experiments as
before but now obtain the p-values by comparing the LLR statistic with the rescaled chi-square suggested by
7
Theorem 1. For a particular run of the experiment (n = 4000, p = 1200, κ = 0.3), we compute the adjusted
LLR statistic (2b∗ /τ∗2 )Λj for each coefficient and obtain the p-values based on the χ21 distribution. The
pooled histogram that aggregates 4.8 × 105 p-values in total is shown in Fig. 1(c).
As we clearly see, the p-values are much closer to a uniform distribution now. One can compute the
chi-square goodness of fit statistic to test the closeness of the above distribution to uniformity. To this end,
we divide the interval [0, 1] into 20 equally spaced bins of width 0.05 each. For each bin we compute the
observed number of times a p-value falls in the bin out of the 4.8 × 105 values. Then a chi-square goodness of
fit statistic is computed, noting that the expected frequency is 24000 for each bin. The chi-square statistic
in this case is 16.049, which gives a p-value of 0.654 in comparison with a χ219 variable. The same test when
performed with the Bartlett corrected p-values (Fig. 1(b)) yields a chi-square statistic 5599 with a p-value
of 0. 3 Thus, our correction gives the desired uniformity in the p-values when the true signal β = 0.
Practitioners would be concerned about the validity of p-values when they are small—again, think about
multiple testing applications. In order to study whether our correction yields valid results for small p-values,
we compute the proportion of times the p-values (in all the three cases) lie below 5%, 1%, 0.5%, 0.1% out
of the 4.8 × 105 times. The results are summarized in Table 1. This further illustrates the deviation from
uniformity for the classical and Bartlett corrected p-values, whereas the “adjusted” p-values obtained invoking
Theorem 1 are still valid.
P{p-value ≤ 5%}
P{p-value ≤ 1%}
P{p-value ≤ 0.5%}
P{p-value ≤ 0.1%}
P{p-value ≤ 0.05%}
P{p-value ≤ 0.01%}
Classical
11.1044%(0.0668%)
3.6383%(0.038%)
2.2477%(0.0292%)
0.7519%(0.0155%)
0.4669%(0.0112%)
0.1575%(0.0064%)
Bartlett-corrected
6.9592%(0.0534%)
1.6975%(0.0261%)
0.9242%(0.0178%)
0.2306%(0.0078%)
0.124%(0.0056%)
0.0342%(0.0027%)
Adjusted
5.0110%(0.0453%)
0.9944%(0.0186%)
0.4952%(0.0116%)
0.1008%(0.0051%)
0.0542%(0.0036%)
0.0104%(0.0014%)
Table 1: Estimates of p-value probabilities with estimated Monte Carlo standard errors in parentheses under
i.i.d. Gaussian design.
2.3.3
Extensions
As noted in Section 1.1, the Symmetry Condition (3) allows to express the negative log-likelihood in the
form (4), which makes use of the effective link ρ(·). Theorem 1 applies to any ρ(·) obeying the following
properties:
1. ρ is non-negative, has up to three derivatives, and obeys ρ(t) ≥ t.
2. ρ0 may be unbounded but it should grow sufficiently slowly, in particular, we assume |ρ0 (t)| = O(|t|)
and ρ0 (proxcρ (Z)) is a sub-Gaussian random variable for any constant c > 0 and any Z ∼ N (0, σ 2 ) for
some finite σ > 0.
3. ρ00 (t) > 0 for any t which implies that ρ is convex, and supt ρ00 (t) < ∞.
4. supt |ρ000 (t)| < ∞.
5. Given any τ > 0 ,the equation (24) has a unique solution in b.
6. The map V(τ 2 ) as defined in (57) has a fixed point.
3 Note that the p-values obtained at each trial are not exactly independent. However, they are exchangeable, and weakly
dependent (see the proof of Corollary 1 for a formal justification of this fact). Therefore, we expect the goodness of fit test to
be an approximately valid procedure in this setting.
8
It can be checked that the effective links for both the logistic and the probit models (5) obey all of the above.
The last two conditions are assumed to ensure existence of a unique solution to the system of equations (23)
and (24) as will be seen in Section 6; we shall justify these two conditions for the logistic and the probit
models in Section 6.1.
2.4
Reduction to independent covariates
In order to derive the asymptotic distribution of the LLR statistics, it in fact suffices to consider the special
case Σ = Ip .
Lemma 1. Let Λj (X) be the LLR statistic based on the design matrix X, where the rows of X are
i.i.d. N (0, Σ) and Λj (Z) that where the rows are i.i.d. N (0, Ip ). Then
d
Λj (X) = Λj (Z).
Proof: Recall from (4) that the LLR statistic for testing the jth coefficient can be expressed as
Λj (X) = min
β
n
X
ρ(−ỹi e>
i Xβ)
i=1
n
X
− min
β:βj =0
ρ(−ỹi e>
i Xβ).
i=1
Write Z 0 = XΣ−1/2 so that the rows of Z 0 are i.i.d. N (0, Ip ) and set θ 0 = Σ1/2 β. With this reparameter0
p
ization, we observe that the constraint βj = 0 is equivalent to a>
j θ = 0 for some non-zero vector aj ∈ R .
This gives
n
n
X
X
> 0 0
0 0
ρ(−ỹ
e
Z
θ
)
−
min
ρ(−ỹi e>
Λj (X) = min
i i
i Z θ ).
0
θ
0
θ 0 :a>
j θ =0
i=1
i=1
p
Now let Q be an orthogonal matrix mapping aj ∈ R into the vector kaj kej ∈ Rp , i.e. Qaj = kaj kej .
0
Additionally, set Z = Z 0 Q (the rows of Z are still i.i.d. N (0, Ip )) and θ = Qθ 0 . Since a>
j θ = 0 occurs if
and only if θj = 0, we obtain
Λj (X) = min
θ
n
X
i=1
ρ(−ỹi e>
i Zθ) − min
θ:θj =0
n
X
ρ(−ỹi e>
i Zθ) = Λj (Z),
i=1
which proves the lemma.
In the remainder of the paper we, therefore, assume Σ = Ip .
2.5
Proof architecture
This section presents the main steps for proving Theorem 1. We will only prove the theorem for {Λj }, the
LLR statistic obtained by dropping a single variable. The analysis for the LLR statistic obtained on dropping
k variables (for some fixed k) follows very similar steps and is hence omitted for the sake of conciceness. As
discussed before, we are free to work with any configuration of the yi ’s. For the two steps below, we will
adopt two different configurations for convenience of presentation.
2.5.1
Step 1: characterizing the asymptotic distributions of β̂j
Without loss of generality, we assume here that yi = 1 (and hence ỹi = 1) for all 1 ≤ i ≤ n and, therefore,
the MLE problem reduces to
Xn
minimizeβ∈Rp
ρ(−Xi> β).
i=1
We would first like to characterize the marginal distribution of β̂, which is crucial in understanding the LLR
statistic. To this end, our analysis follows by a reduction to the setup of [18–21], with certain modifications
9
that are called for due to the specific choices of ρ(·) we deal with here. Specifically, consider the linear model
y = Xβ + w,
(26)
and prior work [18–21] investigating the associated M-estimator
Xn
minimizeβ∈Rp
ρ(yi − Xi> β).
i=1
(27)
Our problem reduces to (27) on setting y = w = 0 in (27). When ρ(·) satisfies certain assumptions
(e.g. strong convexity), the asymptotic distribution of kβ̂k has been studied in a series of works [19–21] using
a leave-one-out analysis and independently in [18] using approximate message passing (AMP) machinery.
An outline of their main results is described in Section 2.7. However, the function ρ(·) in our cases has
vanishing curvature and, therefore, lacks the essential strong convexity assumption that was utilized in both
the aforementioned lines of work. To circumvent this issue, we propose to invoke the AMP machinery as
in [18], in conjunction with the following critical additional ingredients:
• (Norm Bound Condition) We utilize results from the conic geometry literature (e.g. [3]) to establish
that
kβ̂k = O(1)
with high probability as long as κ < 1/2. This will be elaborated in Theorem 4.
• (Likelihood Curvature Condition) We establish some regularity conditions on the Hessian of the loglikelihood function, generalizing the strong convexity condition, which will be detailed in Lemma 4.
• (Uniqueness of the Solution to (23) and (24)) We establish that for both the logistic and the probit
case, the system of equations (23) and (24) admits a unique solution.
We emphasize that these elements are not straightforward, require significant effort and a number of novel
ideas, which form our primary technical contributions for this step.
These ingredients enable the use of the AMP machinery even in the absence of strong convexity on ρ(·),
finally leading to the following theorem:
Theorem 2. Under the conditions of Theorem 1,
lim kβ̂k2 =a.s. τ∗2 .
n→∞
(28)
This theorem immediately implies that the marginal distribution of β̂j is normal.
Corollary 1. Under the conditions of Theorem 1, for every 1 ≤ j ≤ p, it holds that
√
d
pβ̂j → N (0, τ∗2 ),
as n → ∞.
(29)
Proof: From the rotational invariance of our i.i.d. Gaussian design, it can be easily verified that β̂/kβ̂k
is uniformly distributed on the unit sphere Sp−1 and is independent of kβ̂k. Therefore, β̂j has the same
√
distribution as kβ̂kZj /kZk, where Z = (Z1 , . . . , Zp ) ∼ N (0, Ip ) independent of kβ̂k. Since pkβ̂k/kZk
√
converges in probability to τ∗ , we have, by Slutsky’s theorem, that pβ̂j converges to N (0, τ∗2 ) in distribution.
2.5.2
Step 2: connecting Λj with β̂j
Now that we have derived the asymptotic distribution of β̂j , the next step involves a reduction of the LLR
statistic to a function of the relevant coordinate of the MLE. Before continuing, we note that the distribution
of Λj is the same for all 1 ≤ j ≤ p due to exchangeability. As a result, going forward we will only analyze
Λ1 without loss of generality. In addition, we introduce the following convenient notations and assumptions:
10
• the design matrix on dropping the first column is written as X̃ and the MLE in the corresponding
reduced model as β̃;
• write X = [X1 , · · · , Xn ]> ∈ Rn×p and X̃ = [X̃1 , · · · , X̃n ]> ∈ Rn×(p−1) ;
• without loss of generality, assume that ỹi = −1 for all i in this subsection, and hence the MLEs under
the full and the reduced models reduce to
Xn
β̂ = arg minβ∈Rp
`(β) :=
ρ(Xi> β),
(30)
i=1
Xn
˜
β̃ = arg minβ∈Rp−1 `(β)
:=
ρ(X̃i> β).
(31)
i=1
With the above notations in place, the LLR statistic for testing β1 = 0 vs. β1 6= 0 can be expressed as
n n
o
X
˜ β̃) − `(β̂) =
Λ1 := `(
ρ(X̃i> β̃) − ρ(Xi> β̂) .
(32)
i=1
To analyze Λ1 , we invoke Taylor expansion to reach
Λ1 =
n
1X
2
ρ00 Xi> β̂ X̃i> β̃ − Xi> β̂
ρ0 Xi> β̂ X̃i> β̃ − Xi> β̂ +
2 i=1
i=1
{z
}
|
n
X
:=Qlin
n
+
3
1 X 000
ρ (γi ) X̃i> β̃ − Xi> β̂ ,
6 i=1
(33)
where γi lies between X̃i> β̃ and Xi> β̂. A key observation is that the linear term Qlin in the above equation
vanishes. To see this, note that the first-order optimality conditions for the MLE β̂ is given by
Xn
ρ0 (Xi> β̂)Xi = 0.
(34)
i=1
0
Replacing X̃i> β̃ with Xi>
in Qlin and using the optimality condition, we obtain
β̃
Xn
> 0
0
>
Qlin =
ρ Xi β̂ Xi
− β̂ = 0.
i=1
β̃
Consequently, Λ1 simplifies to the following form
n
n
3
2 1 X
1 X 00 > >
ρ (Xi β̂) X̃i β̃ − Xi> β̂ +
ρ000 (γi ) X̃i> β̃ − Xi> β̂ .
Λ1 =
2 i=1
6 i=1
(35)
Thus, computing the asymptotic distribution of Λ1 boils down to analyzing Xi> β̂ − X̃i> β̃. Our argument is
inspired by the leave-one-predictor-out approach developed in [19, 20].
We re-emphasize that our setting is not covered by that of [19,20], due to the violation of strong convexity
and some other technical assumptions. We sidestep this issue by utilizing the Norm Bound Condition and
the Likelihood Curvature Condition. In the end, our analysis establishes the equivalence of Λ1 and β̂1 up to
some explicit multiplicative factors modulo negligible error terms. This is summarized as follows.
Theorem 3. Under the assumptions of Theorem 1,
p
P
2Λ1 − β̂12 → 0,
b∗
as n → ∞.
(36)
Theorem 3 reveals a simple yet surprising connection between the LLR statistic Λ1 and the MLE β̂. As
we shall see in the proof of the theorem, the quadratic term in (35) is 12 bp∗ β̂12 + o(1), while the remaining
third-order term of (35) is vanishingly small. Finally, putting Corollary 1 and Theorem 3 together directly
establishes Theorem 1.
11
2.6
Comparisons with the classical regime
We pause to shed some light on the interpretation of the correction factor τ∗2 /b∗ in Theorem 1 and understand
the differences from classical results. Classical theory (e.g. [30,31]) asserts that when p is fixed and n diverges,
the MLE for a fixed design X is asymptotically normal, namely,
√
d
n(β̂ − β) → N (0, Iβ−1 ),
where
Iβ =
1 >
X Dβ X
n
with Dβ :=
ρ00 X1> β
(37)
..
.
ρ
00
Xn> β
(38)
is the normalized Fisher information at the true value β. In particular, under the global null and i.i.d. Gaussian design, this converges to
(
1
I, for the logistic model
EX [Iβ ] = 42
π I, for the probit model
as n tends to infinity [50, Example 5.40].
The behavior in high dimensions is different. In particular, Corollary 1 states that under the global null,
we have
√
d
p(β̂j − βj ) → N (0, τ∗2 ).
(39)
Comparing the variances in the logistic model, we have that
(
√
4κ, in classical large-sample theory;
lim Var
pβ̂j =
n→∞
τ∗2 , in high dimensions.
Fig. 3 illustrates the behavior of the ratio τ∗2 /κ as a function of κ. Two observations are immediate:
• First, in Fig. 3(a) we have τ∗2 ≥ 4κ for all κ ≥ 0. This indicates an inflation in variance or an “extra
Gaussian noise” component that appears in high dimensions, as discussed in [18]. The variance of the
“extra Gaussian noise” component increases as κ grows.
• Second, as κ → 0, we have τ∗2 /4κ → 1 in the logistic model, which indicates that classical theory
becomes accurate in this case. In other words, our theory recovers the classical prediction in the
regime where p = o(n).
Further, for the testing problem considered here, the LLR statistic in the classical setup can be expressed,
through Taylor expansion, as
2Λ1 = n(β̂ − β̃)> Iβ (β̂ − β̃) + oP (1),
(40)
where β̃ is defined in (31). In the high-dimensional setting, we will also establish a quadratic approximation
of the form
1
2Λ1 = n(β̂ − β̃)> G(β̂ − β̃) + oP (1),
G = X > Dβ̂ X.
n
In Theorem 7, we shall see that b∗ is the limit of n1 Tr(G−1 ), the Stieltjes transform of the empirical spectral
distribution of G evaluated at 0. Thus, this quantity in some sense captures the spread in the eigenvalues
of G one would expect to happen in high dimensions.
12
10
●
●
30
●
●
●
●
7.5
●
●
●
●
20
●
τ2 κ
τ2 κ
●
●
●
●
●
●
●
5
●
●
●
●
●
●
●
●
10
4
●●
●●
●●
●●
●●
●●
●
●
●●●
●●● ●
0.00
0.05
0.10
0.15
●
●
●
●
●
●
●
●
2.5
●
●●● ●
1.57
0.20
κ
0.25
0.30
0.35
0.40
●●
●●
●●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40
κ
(a) logistic regression
(b) probit regression
Figure 3: Ratio of asymptotic variance and dimensionality factor κ as a function of κ.
2.7
Prior art
Wilks’ type of phenomenon in the presence of a diverging dimension p has received much attention in the past.
For instance, Portnoy [43] investigated simple hypotheses in regular exponential families, and established
the asymptotic chi-square approximation for the LLR test statistic as long as p3/2 /n → 0. This phenomenon
was later extended in [46] to accommodate the MLE with a quadratic penalization, and in [53] to account
for parametric models underlying several random graph models. Going beyond parametric inference, Fan
et al. [22, 24] explored extensions to infinite-dimensional non-parametric inference problems, for which the
MLE might not even exist or might be difficult to derive. While the classical Wilks’ phenomenon fails to
hold in such settings, Fan et al. [22, 24] proposed a generalization of the likelihood ratio statistics based
on suitable non-parametric estimators and characterized the asymptotic distributions. Such results have
further motivated Boucheron and Massart [9] to investigate the non-asymptotic Wilks’ phenomenon or,
more precisely, the concentration behavior of the difference between the excess empirical risk and the true
risk, from a statistical learning theory perspective. The Wilks’ phenomenon for penalized empirical likelihood
has also been established [48]. However, the precise asymptotic behavior of the LLR statistic in the regime
that permits p to grow proportional to n is still beyond reach.
On the other hand, as demonstrated in Section 2.5.1, the MLE here under the global null can be viewed as
an M-estimator for a linear regression problem. Questions regarding the behavior of robust linear regression
estimators in high dimensions—where p is allowed to grow with n—–were raised in Huber [30], and have
been extensively studied in subsequent works, e.g. [36, 40–42]. When it comes to logistic regression, the
behavior of the MLE was studied for a diverging number of parameters by [28], which characterized the
squared estimation error of the MLE if (p log p)/n → 0. In addition, the asymptotic normality properties of
the MLE and the penalized MLE for logistic regression have been established by [35] and [23], respectively.
A very recent paper by Fan et al. [25] studied the logistic model under the global null β = 0, and investigated
the classical asymptotic normality as given in (37). It was discovered in [25] that the convergence property
(37) breaks down even in terms of the marginal distribution, namely,
√
Iβ
nβ̂i
d
−1/2 9 N (0, 1) ,
i,i
13
Iβ =
1 >
X X,
4n
12500
30000
15000
10000
7500
Counts
Counts
Counts
20000
10000
10000
5000
5000
2500
0
0
0.00
0.25
0.50
P−Values
0.75
1.00
(a)
0
0.00
0.25
0.50
P−Values
(b)
0.75
1.00
0.00
0.25
0.50
P−Values
0.75
1.00
(c)
Figure 4: Histogram of p-values for logistic regression under i.i.d. Bernoulli design, when β = 0, n = 4000,
p = 1200, and κ = 0.3: (a) classically computed p-values; (b) Bartlett corrected p-values; (c) adjusted
p-values.
as soon as p grows at a rate exceeding n2/3 . This result, however, does not imply the asymptotic distribution
of the likelihood-ratio statistic in this regime. In fact, our theorem implies that LLR statistic 2Λj goes to
χ21 (and hence Wilks phenomenon remains valid) when κ = p/n → 0.
The line of work that is most relevant to the present paper was initially started by El Karoui et al. [21].
Focusing on the regime where p is comparable to n, the authors uncovered, via a non-rigorous argument,
that the asymptotic `2 error of the MLE could be characterized by a system of nonlinear equations. This
seminal result was later made rigorous independently by Donoho et al. [18] under i.i.d. Gaussian design and
by El Karoui [19, 20] under more general i.i.d. random design as well as certain assumptions on the error
distribution. Both approaches rely on strong convexity on the function ρ(·) that defines the M-estimator,
which does not hold in the models considered herein.
2.8
Notations
We adopt the standard notation f (n) = O (g(n)) or f (n) . g(n) which means that there exists a constant
c > 0 such that |f (n)| ≤ c|g(n)|. Likewise, f (n) = Ω (g(n)) or f (n) & g(n) means that there exists a
constant c > 0 such that |f (n)| ≥ c |g(n)|, f (n) g(n) means that there exist constants c1 , c2 > 0 such that
(n)
c1 |g(n)| ≤ |f (n)| ≤ c2 |g(n)|, and f (n) = o(g(n)) means that limn→∞ fg(n)
= 0. Any mention of C, Ci , c, ci
for i ∈ N refers to some positive universal constants whose value may change from line to line. For a square
symmetric matrix M , the minimum eigenvalue is denoted by λmin (M ). Logarithms are base e.
3
3.1
Numerics
Non-Gaussian covariates
In this section we first study the sensitivity of our result to the Gaussianity assumption on the design matrix.
To this end, we consider a high dimensional binary regression set up with a Bernoulli design matrix. We
i.i.d.
simulate n = 4000 i.i.d. observations (yi , Xi ) with yi ∼ Bernoulli(1/2), and Xi generated independent of
yi , such that each entry takes on values in {1, −1} w.p. 1/2. At each trial, we fit a logistic regression model
to the data and obtain the classical, Bartlett corrected and adjusted p-values (using the rescaling factor
τ∗2 /b∗ ). Figure 4 plots the histograms for the pooled p-values, obtained across 400 trials.
14
It is instructive to compare the histograms to that obtained in the Gaussian case (Figure 1). The classical
and Bartlett corrected p-values exhibit similar deviations from uniformity as in the Gaussian design case,
whereas our adjusted p-values continue to have an approximate uniform distribution. We test for deviations
from uniformity using a formal chi-squared goodness of fit test as in Section 2.3.2. For the Bartlett corrected
p-values, the chi-squared statistic turns out to be 5885, with a p-value 0. For the adjusted p-values,the
chi-squared statistic is 24.1024, with a p-value 0.1922.4
Once again, the Bartlett correction fails to provide valid p-values whereas the adjusted p-values are
consistent with a uniform distribution. These findings indicate that the distribution of the LLR statistic
under the i.i.d. Bernoulli design is in agreement to the rescaled χ21 derived under the Gaussian design in
Theorem 1, suggesting that the distribution is not too sensitive to the Gaussianity assumption. Estimates
of p-value probabilities for our method are provided in Table 2.
P{p-value ≤ 5%}
P{p-value ≤ 1%}
P{p-value ≤ 0.5%}
P{p-value ≤ 0.1%}
P{p-value ≤ 0.05%}
P{p-value ≤ 0.01%}
Adjusted
5.0222%(0.0412%)
1.0048%(0.0174%)
0.5123%(0.0119%)
0.1108%(0.005%)
0.0521%(0.0033%)
0.0102%(0.0015%)
Table 2: Estimates of p-value probabilities with estimated Monte Carlo standard errors in parentheses under
i.i.d. Bernoulli design.
3.2
Quality of approximations for finite sample sizes
In the rest of this section, we report some numerical experiments which study the applicability of our theory
in finite sample setups.
Validity of tail approximation The first experiment explores the efficacy of our correction for extremely small p-values. This is particularly important in the context of multiple comparisons, where practitioners care about the validity of exceedingly small p-values. To this end, the empirical cumulative distribution of the adjusted p-values is estimated under a standard Gaussian design with n = 4000, p = 1200 and
4.8 × 105 p-values. The range [0.1/p, 12/p] is divided into points which are equi-spaced with a distance of
1/p between any two consecutive points. The estimated empirical CDF at each of these points is represented
in blue in Figure 5. The estimated CDF is in near-perfect agreement with the diagonal, suggesting that the
adjusted p-values computed using the rescaled chi-square distribution are remarkably close to a uniform, even
when we zoom in at very small resolutions as would be the case when applying Bonferroni-style corrections.
Moderate sample sizes The final experiment studies the accuracy of our asymptotic result for moderately large samples. This is especially relevant for applications where the sample sizes are not too large. We
repeat our numerical experiments with n = 200, p = 60 for i.i.d. Gaussian design, and 4.8 × 105 p-values.
The empirical CDF for these p-values are estimated and Figure 6 shows that the adjusted p-values are nearly
uniformly distributed even for moderate sample sizes such as n = 200.
4
Preliminaries
This section gathers a few preliminary results that will be useful throughout the paper. We start by collecting
some facts regarding i.i.d. Gaussian random matrices.
4 Recall
our earlier footnote about the use of a χ2 test.
15
0.0100
●
●●
●
●
●●
●●
●
●●
●●
●
●●
●●
●
●●
●
●●
●●
●
●
●●
●●
●
●●
●●
●
●●
●●
●●
●●
●
●
●
●●
●●
●●
●●
●
●
●●
●
●
●●
●●
●
●
●●
●●
●●
●●
●
●●
●●
●●
●●
●
●
●●
●●
●
●●
●●
●
●
Empirical cdf
0.0075
0.0050
0.0025
0.0000
0.000
0.002
0.004
0.006
0.008
0.010
t
Figure 5: Empirical CDF of adjusted pvalues for logistic regression when β = 0, n = 4000, p = 1200. Here,
the blue points represent the empirical CDF (t vs. the fraction of p-values below t), and the red line is the
diagonal.
1.00
Empirical cdf
0.75
0.50
0.25
0.00
●●
●●
●●
●●
●
●
●●
●●
●●
●●
●
●
●●
●●
●●
●●
●
●
●●
●●
●●
●●
●
●
●●
●●
●●
●●
●
●
●●
●●
●●
●●
●
●
●●
●●
●●
●●
●
●
●●
●●
●●
●●
●
●
●●
●●
●●
●●
●
●●
●●
●●
●●
●
●
●●
0.00
0.25
0.50
0.75
1.00
t
Figure 6: Empirical CDF of adjusted pvalues for logistic regression when β = 0, n = 200, p = 60. Here,
the blue points represent the empirical CDF (t vs. the fraction of p-values below t), and the red line is the
diagonal.
16
Lemma 2. Let X = [X1 , X2 , . . . Xn ]> be an n × p matrix with i.i.d. standard Gaussian entries. Then
P kX > Xk ≤ 9n ≥ 1 − 2 exp(−n/2);
(41)
√
√
P sup1≤i≤n kXi k ≤ 2 p ≥ 1 − 2n exp(−( p − 1)2 /2).
Proof: This is a straighforward application of [51, Corollary 5.35] and the union bound.
(42)
Lemma 3. Suppose X is an n × p matrix
with
pentries i.i.d N (0, 1), then there exists a constant 0 such
√
that whenever 0 ≤ ≤ 0 and 0 ≤ t ≤ 1 − − p/n,
!
r
2
√
1X
p
>
λmin
1−−
Xi Xi
− t , ∀S ⊆ [n] with |S| = (1 − )n
(43)
≥
n
n
i∈S
2
with probability exceeding 1 − 2 exp − (1−)t
−
H
()
n . Here, H() = − log − (1 − ) log(1 − ).
2
Proof: See Appendix A.1.
The above facts are useful in establishing an eigenvalue lower bound on the Hessian of the log-likelihood
function. Specifically, recall that
Xn
∇2 `(β) =
ρ00 Xi> β Xi Xi> ,
(44)
i=1
and the result is this:
Lemma 4 (Likelihood Curvature Condition). Suppose that p/n < 1 and that ρ00 (·) ≥ 0. Then there exists a
constant 0 such that whenever 0 ≤ ≤ 0 , with probability at least 1 − 2 exp (−nH ()) − 2 exp (−n/2), the
matrix inequality
!2
!
r
r
√
p
H()
1 2
00
∇ `(β)
inf
−2
I
(45)
ρ (z)
1−−
3kβk
n
n
1−
z:|z|≤ √
holds simultaneously for all β ∈ Rp .
Proof: See Appendix A.2.
The message of Lemma 4 is this: take > 0 to be a sufficiently small constant. Then
1 2
∇ `(β) ω(kβk) I
n
for some non-increasing and positive function ω(·) independent of n. This is a generalization of the strong
convexity condition.
5
5.1
When is the MLE bounded?
Phase transition
In Section 2.2, we argued that the MLE is at infinity if we have less than two observations per dimension or
κ > 1/2. In fact, a stronger version of the phase transition phenemonon occurs in the sense that
kβ̂k = O(1)
as soon as κ < 1/2. This is formalized in the following theorem.
17
Theorem 4 (Norm Bound Condition). Fix any small constant > 0, and let β̂ be the MLE for a model
with effective link satisfying the conditions from Section 2.3.3.
(i) If p/n ≥ 1/2 + , then
kβ̂k = ∞
with probability exceeding 1 − 4 exp −2 n/8 .
(ii) There exist universal constants c1 , c2 , C2 > 0 such that if p/n < 1/2 − c1 3/4 , then5
kβ̂k <
4 log 2
2
with probability at least 1 − C2 exp(−c2 2 n).
These conclusions clearly continue to hold if β̂ is replaced by β̃ (the MLE under the restricted model obtained
on dropping the first predictor).
The rest of this section is devoted to proving this theorem. As we will see later, the fact that kβ̂k = O(1)
is crucial for utilizing the AMP machinery in the absence of strong convexity.
5.2
Proof of Theorem 4
As in Section 2.5.1, we assume ỹi ≡ 1 throughout this section, and hence the MLE reduces to
Xn
minimizeβ∈Rp `0 (β) :=
ρ(−Xi> β).
i=1
5.2.1
(46)
Proof of Part (i)
Invoking [3, Theorem I] yields that if
δ ({Xβ | β ∈ Rp }) + δ Rn+ ≥ (1 + ) n,
or equivalently, if p/n ≥ 1/2 + , then
P {Xβ | β ∈ Rp } ∩ Rn+ 6= {0} ≥ 1 − 4 exp −2 n/8 .
As is seen in Section 2.2, kβ̂k = ∞ when {Xβ | β ∈ Rp } ∩ Rn+ 6= {0}, establishing Part (i) of Theorem 4.
5.2.2
Proof of Part (ii)
We now turn to the regime in which p/n ≤ 1/2 − O(3/4 ), where 0 < < 1 is any fixed constant. Begin by
observing that the least singular value of X obeys
√
(47)
σmin (X) ≥ n/4
2
with probability at least 1 − 2 exp − 21 34 − √12 n (this follows from Lemma 3 using = 0). Then for any
β ∈ Rp obeying
Xn
`0 (β) =
ρ −Xj> β ≤ n log 2 = `0 (0)
(48)
j=1
and
5 When
kβk ≥
4 log 2
,
2
Xi ∼ N (0, Σ) for a general Σ 0, one has kΣ1/2 β̂k . 1/2 with high probability.
18
(49)
we must have
n
X
max −Xj> β, 0 =
(a)
−Xj> β ≤
X
j: Xj> β<0
j=1
X
(b)
ρ −Xj> β ≤ n log 2;
j: Xj> β<0
(a) follows since t ≤ ρ(t) and (b) is a consequence of (48). Continuing, (47) and (49) give
√ kXβk
√
n log 2 ≤ 4 n
log 2 ≤ 2 nkXβk.
kβk
This implies the following proposition: if the solution β̂—which necessarily satisfies `0 (β̂) ≤ `0 (0)—has
2
norm exceeding kβ̂k ≥ 4 log
2 , then X β̂ must fall within the cone
n
o
Xn
√
max {−uj , 0} ≤ 2 nkuk .
A := u ∈ Rn
(50)
j=1
2
Therefore, if one wishes to rule out the possibility of having kβ̂k ≥ 4 log
2 , it suffices to show that with high
probability,
{Xβ | β ∈ Rp } ∩ A = {0} .
(51)
This is the content of the remaining proof.
We would like to utilize tools from conic geometry [3] to analyze the probability of the event (51). Note,
however, that A is not convex, while the theory developed in [3] applies only to convex cones. To bypass the
non-convexity issue, we proceed in the following three steps:
1. Generate a set of N = exp 22 p closed convex cones {Bi | 1 ≤ i ≤ N } such that it forms a cover of
A with probability exceeding 1 − exp −Ω(2 p) .
√ 3
√
2. Show that if p < 21 − 2 2 4 − 2H(2 ) n and if n is sufficiently large, then
(
1
P {{Xβ | β ∈ R } ∩ Bi =
6 {0}} ≤ 4 exp −
8
p
√ 3
√
p
1
− 2 2 4 − 10H(2 ) −
2
n
2 )
n
for each 1 ≤ i ≤ N .
3. Invoke the union bound to reach
P {{Xβ | β ∈ Rp } ∩ A 6= {0}}
≤
P {{Bi | 1 ≤ i ≤ N } does not form a cover of A}
+
N
X
P {{Xβ | β ∈ Rp } ∩ Bi 6= {0}}
i=1
≤ exp −Ω(2 p) ,
where we have used the fact that
2 )
√
√
3
1
1
p
P {{Xβ | β ∈ Rp } ∩ Bi 6= {0}} ≤ 4N exp −
− 2 2 4 − 10H(2 ) −
n
8
2
n
i=1
(
! )
2
√ 3
√
1 1
p
2
< 4 exp −
− 2 2 4 − 10H(2 ) −
− 2 n
8 2
n
< 4 exp −2 n .
2
√ 3
√
Here, the last inequality holds if 21 − 2 2 4 − 10H(2 ) − np
> 242 , or equivalently, np < 12 −
√ 3
√
√
2 2 4 − 10H(2 ) − 24.
(
N
X
19
√
√ 3
√
Taken collectively, these steps establish the following claim: if np < 12 − 2 2 4 − 10H(2 ) − 24, then
4 log 2
P kβ̂k >
< exp −Ω(2 n) ,
2
thus establishing Part (ii) of Theorem 4. We defer the complete details of the preceding steps to Appendix
D.
Asymptotic `2 error of the MLE
6
This section aims to establish Theorem 2, which characterizes precisely the asymptotic squared error of the
MLE β̂ under the global null β = 0. As described in Section 2.5.1, it suffices to assume that β̂ is the solution
to the following problem
Xn
minimizeβ∈Rp
ρ(−Xi> β).
(52)
i=1
In what follows, we derive the asymptotic convergence of kβ̂k under the assumptions from our main theorem.
Theorem 5. Under the assumptions of Theorem 1, the solution β̂ to (52) obeys
lim kβ̂k2 =a.s. τ∗2 .
n→∞
(53)
Theorem 5 is derived by invoking the AMP machinery [6, 7, 32]. The high-level idea is the following:
in order to study β̂, one introduces an iterative algorithm (called AMP) where a sequence of iterates β̂ t is
formed at each time t. The algorithm is constructed so that the iterates asymptotically converge to the MLE
in the sense that
lim lim kβ̂ t − β̂k2 =a.s. 0.
(54)
t→∞ n→∞
On the other hand, the asymptotic behavior (asymptotic in n) of β̂ t for each t can be described accurately
by a scalar sequence {τt }—called state evolution (SE)—following certain update equations [6]. This, in turn,
provides a characterization of the `2 loss of β̂.
Further, in order to prove Theorem 2, one still needs to justify
(a) the existence of a solution to the system of equations (23) and (24),
(b) and the existence of a fixed point for the iterative map governing the SE sequence updates.
We will elaborate on these steps in the rest of this section.
6.1
State evolution
We begin with the SE sequence {τt } introduced in [18]. Starting from some initial point τ0 , we produce two
sequences {bt } and {τt } following a two-step procedure.
• For t = 0, 1, . . .:
– Set bt to be the solution in b to
κ = E Ψ0 (τt Z; b) ;
(55)
– Set τt+1 to be
2
τt+1
=
1 2
E (Ψ (τt Z; bt )) .
κ
20
(56)
10.0
7.5
7.5
ν(τ2)
ν(τ2)
10.0
5.0
2.5
0.0
0.0
5.0
2.5
2.5
5.0
τ2
0.0
7.5
(a) logistic regression
0
1
2
τ2
3
4
(b) probit regression
Figure 7: The variance map for both the logistic and the probit models when κ = 0.3: (blue line) variance
map V(τ 2 ) as a function of τ 2 ; (red line) diagonal.
Suppose that for given any τ > 0, the solution in b to (55) with τt = τ exists and is unique, then one can
denote the solution as b(τ ), which in turn allows one to write the sequence {τt } as
2
τt+1
= V(τt2 )
with the variance map
1 2
E Ψ (τ Z; b(τ )) .
(57)
κ
As a result, if there exists a fixed point τ∗ obeying V(τ∗2 ) = τ∗2 and if we start with τ0 = τ∗ , then by induction,
V(τ 2 ) =
τt ≡ τ∗
and bt ≡ b∗ := b(τ∗ ),
t = 0, 1, . . .
Notably, (τ∗ , b∗ ) solves the system of equations (23) and (24). We shall work with this choice of initial
condition throughout our proof.
The preceding arguments hold under two conditions: (i) the solution to (24) exists and is unique for
any τt > 0; (ii) the variance map (57) admits a fixed point. To verify these two conditions, we make two
observations.
• Condition (i) holds if one can show that the function
G(b) := E [Ψ0 (τ Z; b)] ,
b>0
(58)
is strictly monotone for any given τ > 0, and that limb→0 G(b) < κ < limb→∞ G(b).
• Since V(·) is a continuous function, Condition (ii) becomes self-evident once we show that V(0) > 0
and that there exists τ > 0 obeying V(τ 2 ) < τ 2 . The behavior of the variance map is illustrated in
Figure 7 for the logistic and probit regression when κ = 0.3. One can in fact observe that the fixed
point is unique. For other values of κ, the variance map shows the same behavior.
In fact, the aforementioned properties can be proved for a certain class of effective links, as summarized
in the following lemmas. In particular, they can be shown for the logistic and the probit models.
21
Lemma 5. Suppose the effective link ρ satisfies the following two properties:
(a) ρ0 is log-concave.
(b) For any fixed τ > 0 and any fixed z, bρ00 (proxbρ (τ z)) → ∞ when b → ∞.
Then for any τ > 0, the function G(b) defined in (58) is an increasing function in b (b > 0), and the equation
G(b) = κ
has a unique positive solution.
Proof: See Appendix B.
Lemma 6. Suppose that 0 < κ < 1/2 and that ρ = log(1 + et ) or ρ = − log Φ(−t). Then
(i) V(0) > 0;
(ii) V(τ 2 ) < τ 2 for some sufficiently large τ 2 .
Proof: See Appendix C and the supplemental material [47].
Remark 1. A byproduct of the proof is that the following relations hold for any constant 0 < κ < 1/2:
• In the logistic case,
2
limτ →∞ V (τ2 )
τ
lim
τ →∞
b(τ )
τ
=
x2 P{Z>x}+E[Z 2 1{0<Z<x} ]
P{0<Z<x}
;
x=Φ−1 (κ+0.5)
= Φ−1 (κ + 0.5).
• In the probit case,
lim b(τ ) =
τ →∞
2κ
1 − 2κ
and
V(τ 2 )
= 2κ.
τ →∞ τ 2
lim
(59)
Remark 2. Lemma 6 is proved for the two special effective link functions, the logistic and the probit cases.
However, the proof sheds light on general conditions on the effective link that suffice for the lemma to hold.
Such general sufficient conditions are also discussed in the supplemental material [47].
6.2
AMP recursion
In this section, we construct the AMP trajectory tracked by two sequences {β̂ t (n) ∈ Rp } and {η t (n) ∈ Rn }
for t ≥ 0. Going forward we suppress the dependence on n to simplify presentation. Picking β̂ 0 such that
lim kβ̂ 0 k2 = τ02 = τ∗2
n→∞
and taking η −1 = 0 and b−1 = 0, the AMP path is obtained via Algorithm 1, which is adapted from the
algorithm in [18, Section 2.2].
22
Algorithm 1 Approximate message passing.
For t = 0, 1, · · · :
1. Set
η t = X β̂ t + Ψ η t−1 ; bt−1 ;
(60)
κ = E [Ψ0 (τt Z; b)] ,
(61)
2. Let bt be the solution to
where τt is the SE sequence value at that time.
3. Set
1
β̂ t+1 = β̂ t − X > Ψ η t ; bt .
p
(62)
Here, Ψ(·) is applied in an entrywise manner, and Ψ0 (., .) denotes derivative w.r.t the first variable.
As asserted by [18], the SE sequence {τt } introduced in Section 6.1 proves useful as it offers a formal
procedure for predicting operating characteristics of the AMP iterates at any fixed iteration. In particular
it assigns predictions to two types of observables: observables which are functions of the β̂ t sequence and
those which are functions of η t . Repeating identical argument as in [18, Theorem 3.4], we obtain
lim kβ̂ t k2 =a.s. τt2 ≡ τ∗2 ,
n→∞
6.3
t = 0, 1, . . . .
(63)
AMP converges to the MLE
We are now in position to show that the AMP iterates {β̂ t } converge to the MLE in the large n and t limit.
Before continuing, we state below two properties that are satisfied under our assumptions.
• The MLE β̂ obeys
lim kβ̂k < ∞
n→∞
(64)
almost surely.
• And there exists some non-increasing continuous function 0 < ω (·) < 1 independent of n such that
1 2
P
∇ ` (β) ω (kβk) · I, ∀β ≥ 1 − c1 e−c2 n .
(65)
n
In fact, the norm bound (64) follows from Theorem 4 together with Borel-Cantelli, while the likelihood
curvature condition (65) is an immediate consequence of Lemma 4. With this in place, we have:
Theorem 6. Suppose (64) and (65) hold. Let (τ∗ , b∗ ) be a solution to the system (23) and (24), and assume
that limn→∞ kβ̂ 0 k2 = τ∗2 . Then the AMP trajectory as defined in Algorithm 1 obeys
lim lim kβ̂ t − β̂k =a.s. 0.
t→∞ n→∞
Taken collectively, Theorem 6 and Eqn. (63) imply that
lim kβ̂k =a.s. lim lim kβ̂ t k =a.s. τ∗ ,
n→∞
t→∞ n→∞
thus establishing Theorem 5. In addition, an upshot of these theorems is a uniqueness result:
Corollary 2. The solution to the system of equations (23) and (24) is unique.
23
(66)
Proof: When the AMP trajectory β̂ t is started with the initial condition from Theorem 6, limn→∞ kβ̂k2 =a.s.
τ∗2 . This holds for any τ∗ such that (τ∗ , b∗ ) is a solution to (23) and (24). However, since the MLE problem
is strongly convex and hence admits a unique solution β̂, this implies that τ∗ must be unique, which together
with the monotonicity of G(·) (cf. (58)) implies that b∗ is unique as well.
Proof of Theorem 6: To begin with, repeating the arguments in [18, Lemma 6.9] we reach
lim lim kβ̂ t+1 − β̂ t k2 =a.s. 0;
t→∞ n→∞
(67)
1 t+1
kη
− η t k2 =a.s. 0.
(68)
n
To show that the AMP iterates converge to the MLE, we shall analyze the log-likelihood function. Recall
from Taylor’s theorem that
>
E 1
D
β̂ − β̂ t ∇2 ` β̂ t + λ(β̂ − β̂ t ) β̂ − β̂ t
`(β̂) = `(β̂ t ) + ∇`(β̂ t ), β̂ − β̂ t +
2
holds for some 0 < λ < 1. To deal with the quadratic term, we would like to control the Hessian of the
likelihood at a point between β̂ and β̂ t . Invoking the likelihood curvature condition (65), one has
D
E 1
n
o
`(β̂ t ) ≥ `(β̂) ≥ `(β̂ t ) + ∇`(β̂ t ), β̂ − β̂ t + nω max kβ̂k, kβ̂ t k kβ̂ − β̂ t k2
(69)
2
with high probability. Apply Cauchy-Schwarz to yield that with exponentially high probability,
lim lim
t→∞ n→∞
kβ̂ − β̂ t k ≤
1
1
2
2
n
o
∇`(β̂ t ) ≤
∇`(β̂ t ) ,
t
n
n
t
ω kβ̂k ω kβ̂ k
ω max kβ̂k, kβ̂ k
where the last inequality follows since 0 < ω(·) < 1 and ω(·) is non-decreasing.
It remains to control k∇`(β̂ t )k. The identity Ψ(z; b∗ ) = z − proxb∗ ρ (z) and (60) give
proxb∗ ρ η t−1 = X β̂ t + η t−1 − η t .
(70)
0
In addition, substituting Ψ (z; b) = bρ (proxρb (z)) into (62) yields
p t
(β̂ − β̂ t−1 ) = −X > ρ0 proxb∗ ρ (η t−1 ) = −X > ρ0 X β̂ t + η t−1 − η t .
b∗
We are now ready to bound k∇`(β̂ t )k. Recalling that
∇`(β̂ t ) = X > ρ0 (X > β̂ t ) = X > ρ0 X β̂ t + η t−1 − η t + X > ρ0 (X > β̂ t ) − ρ0 X β̂ t + η t−1 − η t
and that supz ρ00 (z) < ∞, we have
∇`(β̂ t )
≤
−X > ρ0 X β̂ t + η t−1 − η t + kXk ρ0 X β̂ t + η t−1 − η t − ρ0 (X β̂ t )
p
t
t−1
00
kβ̂ − β̂ k + kXk sup ρ (z) kη t−1 − η t k.
≤
b∗
z
This establishes that with probability at least 1 − c1 e−c2 n ,
2
p
1
kβ̂ − β̂ t k ≤
kβ̂ t − β̂ t−1 k +
sup ρ00 (z) kXkkη t−1 − η t k .
(71)
b∗ n
n
z
ω kβ̂k ω kβ̂ t k
√
Using (42) together with Borel-Cantelli yields limn→∞ kXk/ n < ∞ almost surely. Further, it follows
from (63) that limn→∞ kβ̂ t k is finite almost surely as τ∗ < ∞. These taken together with (64), (67) and (68)
yield
lim lim kβ̂ − β̂ t k =a.s. 0
(72)
t→∞ n→∞
as claimed.
24
7
Likelihood ratio analysis
This section presents the analytical details for Section 2.5.2, which relates the log-likelihood ratio statistic
Λi with β̂i . Recall from (35) that the LLR statistic for testing β1 = 0 vs. β1 6= 0 is given by
n
Λ1 =
>
3
1X
1
X̃ β̃ − X β̂ Dβ̂ X̃ β̃ − X β̂ +
ρ000 (γi ) X̃i> β̃ − Xi> β̂ ,
2
6 i=1
where
ρ00 X1> β̂
(73)
Dβ̂ :=
..
.
ρ00 Xn> β̂
(74)
and γi lies between Xi> β̂ and X̃i> β̃. The asymptotic distribution of Λ1 claimed in Theorem 3 immediately
follows from the result below, whose proof is the subject of the rest of this section.
Theorem 7. Let (τ∗ , b∗ ) be the unique solution to the system of equations (23) and (24), and define
G̃ =
1 >
X̃ Dβ̃ X̃
n
and
α̃ =
1
Tr(G̃−1 ).
n
(75)
Suppose p/n → κ ∈ (0, 1/2) . Then
(a) the log-likelihood ratio statistic obeys
P
2Λ1 − pβ̂12 /α̃ → 0;
(76)
(b) and the scalar α̃ converges,
P
α̃ → b∗ .
7.1
(77)
More notations and preliminaries
Before proceeding, we introduce some notations that will be used throughout. For any matrix X, denote
by Xij and X·j its (i, j)-th entry and jth column, respectively. We denote an analogue r = {ri }1≤i≤n
(resp. r̃ = {r̃i }1≤i≤n ) of residuals in the full (resp. reduced) model by
ri := −ρ0 Xi> β̂
and
r̃i := −ρ0 X̃i> β̃ .
(78)
As in (74), set
Dβ̃ :=
ρ00 X̃1> β̃
..
.
ρ00 X̃n> β̃
and Dβ̂,b̃ :=
ρ00 (γ1∗ )
..
.
ρ
00
(γn∗ ),
,
(79)
where γi∗ is between Xi> β̂ and Xi> b̃, and b̃ is to be defined later in Section 7.2. Further, as in (75), introduce
the Gram matrices
G :=
1 >
X Dβ̂ X
n
and Gβ̂,b̃ =
1 >
X Dβ̂,b̃ X.
n
(80)
Let G̃(i) denote the version of G̃ without the term corresponding to the ith observation, that is,
G̃(i) =
1X
ρ00 (X̃j> β̃)X̃j X̃j> .
j:j6=i
n
25
(81)
Additionally, let β̂[−i] be the MLE when the ith observation is dropped and let G[−i] be the corresponding
Gram matrix,
G[−i] =
1X
ρ00 (Xj> β̂[−i] )Xj Xj> .
j:j6=i
n
(82)
Further, let β̃[−i] be the MLE when the first predictor and ith observation are removed, i.e.
X
β̃[−i] := arg min
ρ(X̃j> β).
β∈Rp−1
j:j6=i
Below G̃[−i] is the corresponding version of G̃,
G̃[−i] =
1X
ρ00 (X̃j> β̃[−i] )X̃j X̃j> .
j:j6=i
n
(83)
For these different versions of G, their least eigenvalues are all bounded away from 0, as asserted by the
following lemma.
Lemma 7. There exist some absolute constants λlb , C, c > 0 such that
P(λmin (G) > λlb ) ≥ 1 − Ce−cn .
Moreover, the same result holds for G̃, Gβ̂,b̃ , G̃(i) , G[−i] and G̃[−i] for all i ∈ [n].
Proof: This result follows directly from Lemma 2, Lemma 4, and Theorem 4.
Throughout the rest of this section, we restrict ourselves (for any given n) to the following event:
An := {λmin (G̃) > λlb } ∩ {λmin (G) > λlb } ∩ {λmin (Gβ̂,b̃ ) > λlb }
∩ {∩ni=1 λmin (G̃(i) ) > λlb } ∩ {∩ni=1 λmin (G̃[−i] ) > λlb } ∩ {∩ni=1 λmin (G[−i] ) > λlb }.
(84)
By Lemma 7, An arises with exponentially high probability, i.e.
P(An ) ≥ 1 − exp(−Ω(n)).
7.2
(85)
A surrogate for the MLE
In view
of (73), the main step in controlling Λ1 consists of characterizing the differences X β̂ − X̃ β̃ or
0
β̂ −
. Since the definition of β̂ is implicit and not amenable to direct analysis, we approximate β̂ by
β̃
a more amenable surrogate b̃, an idea introduced in [19–21]. We collect some properties of the surrogate
which will prove valuable in the subsequent analysis.
To begin with, our surrogate is
0
1
b̃ =
+ b̃1
,
(86)
−G̃−1 w
β̃
where G̃ is defined in (80),
w :=
1 Xn
1
ρ00 (X̃i> β̃)Xi1 X̃i = X̃ > Dβ̃ X·1 ,
i=1
n
n
and b̃1 is a scalar to be specified later. This vector is constructed in the hope that
(
β̂1 ≈ b̃1 ,
β̂ ≈ b̃,
or equivalently,
β̂2:p − β̃ ≈ −b̃1 G̃−1 w,
26
(87)
(88)
where β̂2:p contains the 2nd through pth components of β̂.
Before specifying b̃1 , we shall first shed some insights into the remaining terms in b̃. By definition,
"
#
>
>
>
X
D
X
X
D
X̃
X·1
Dβ̃ X·1 nw>
0
·1
·1
·1
2
>
β̃
β̃
∇ `
= X Dβ̃ X =
.
=
β̃
X̃ > Dβ̃ X·1 X̃ > Dβ̃ X̃
nw
nG̃
Employing the first-order approximation of ∇`(·) gives
0
0
0
∇2 `
β̂ −
≈ ∇`(β̂) − ∇`
.
β̃
β̃
β̃
(89)
0
are
β̂
also very close to each other. Therefore, taking the 2nd through pth components of (89) and approximating
them by zero give
h
i
0
w, G̃ β̂ −
≈ 0.
β̃
Suppose β̂2:p is well approximated by β̃. Then all but the first coordinates of ∇`(β̃) and ∇`
This together with a little algebra yields
β̂2:p − β̃ ≈ −β̂1 G̃−1 w ≈ −b̃1 G̃−1 w,
which coincides with (88). In fact, for all but the 1st entries, b̃ is constructed by moving β̃ one-step in the
direction which takes it closest to β̂.
Next, we come to discussing the scalar b̃1 . Introduce the projection matrix
H := I −
and define b̃1 as
b̃1 :=
1 1/2
1/2
D X̃ G̃−1 X̃ > Dβ̃ ,
n β̃
>
X·1
r̃
,
1/2
1/2
>
X·1 Dβ̃ HDβ̃ X·1
(90)
(91)
where r̃ comes from (78). In fact, the expression b̃1 is obtained through similar (but slightly more complicated) first-order approximation as for b̃2:p , in order to make sure that b1 ≈ β̂1 ; see [21, Pages 14560-14561]
for a detailed description.
We now formally justify that the surrogate b̃ and the MLE β̂ are close to each other.
Theorem 8. The MLE β̂ and the surrote b̃ (86) obey
kβ̂ − b̃k . n−1+o(1) ,
(92)
|b̃1 | . n−1/2+o(1) ,
(93)
sup |Xi> b̃ − X̃i> β̃| . n−1/2+o(1)
(94)
and
1≤i≤n
with probability tending to one as n → ∞.
Proof: See Section 7.4.
The global accuracy (92) immediately leads to a coordinate-wise approximation result between β̂1 and
b̃1 .
27
Corollary 3. With probability tending to one as n → ∞,
√
n|b̃1 − β̂1 | . n−1/2+o(1) .
(95)
Another consequence from Theorem 8 is that the value Xi> β̂ in the full model and its counterpart X̃i> β̃
in the reduced model are uniformly close.
Corollary 4. The values Xi> β̂ and X̃i> β̃ are uniformly close in the sense that
sup Xi> β̂ − X̃i> β̃
. n−1/2+o(1)
(96)
1≤i≤n
holds with probability approaching one as n → ∞.
Proof: Note that
sup Xi> β̂ − X̃i> β̃ ≤
1≤i≤n
sup Xi> (β̂ − b̃) + sup Xi> b̃ − X̃i> β̃ .
1≤i≤n
1≤i≤n
The second term in the right-hans side is upper bounded by n−1/2+o(1) with probability 1 − o(1) according
to Theorem 8. Invoking Lemma 2 and Theorem 8 and applying Cauchy-Schwarz inequality yield that the
first term is O(n−1/2+o(1) ) with probability 1 − o(1). This establishes the claim.
7.3
Analysis of the likelihood-ratio statistic
We are now positioned to use our surrogate b̃ to analyze the likelihood-ratio statistic. In this subsection we
establish Theorem 7(a). The proof for Theorem 7(b) is deferred to Appendix I.
Recall from (35) that
n
2Λ1 = (X̃ β̃ − X β̂)> Dβ̂ (X̃ β̃ − X β̂) +
1 X 000
ρ (γi )(X̃i> β̃ − Xi> β̂)3 .
3 i=1
{z
}
|
:=I3
000
To begin with, Corollary 4 together with the assumption supz ρ (z) < ∞ implies that
I3 . n−1/2+o(1)
with probability 1 − o(1). Hence, I3 converges to zero in probability.
Reorganize the quadratic term as follows:
2
X
(X̃ β̃ − X β̂)> Dβ̂ (X̃ β̃ − X β̂) =
ρ00 (Xi> β̂) Xi> β̂ − X̃i> β̃
i
=
X
=
X
h
i2
ρ00 (Xi> β̂) Xi> (β̂ − b̃) + (Xi> b̃ − X̃i> β̃)
i
ρ00 (Xi> β̂)(Xi> (β̂ − b̃))2 + 2
X
ρ00 (Xi> β̂)Xi> (β̂ − b̃)(Xi> b̃ − X̃i> β̃)
i
i
+
X
ρ
00
(Xi> β̂)
Xi> b̃
2
− X̃i> β̃ .
(97)
i
We control each of the three terms in the right-hand side of (97).
• Since supz ρ00 (z) < ∞, the first term in (97) is bounded by
X
X
ρ00 (Xi> β̂)(Xi> (β̂ − b̃))2 . ||β̃ − b̃||2
Xi Xi>
i
i
with probability 1 − o(1), by an application of Theorem 8 and Lemma 2.
28
. n−1+o(1)
• From the definition of b̃, the second term can be upper bounded by
p
X
X
1
≤ |b̃1 | · kβ̂ − b̃k ·
Xi Xi> · 1 + w> G̃−2 w
2
ρ00 (Xi> β̂)(β̂ − b̃)> Xi Xi> b̃1
−1
i
−G̃ w
i
1
. n− 2 +o(1)
with probability 1 − o(1), where the last line follows from a combination of Theorem 8, Lemma 2 and
the following lemma.
Lemma 8. Let G̃ and w be as defined in (80) and (87), respectively. Then
P w> G̃−2 w . 1 ≥ 1 − exp(−Ω(n)).
(98)
Proof: See Appendix E.
• The third term in (97) can be decomposed as
X
ρ00 (Xi> β̂)(Xi> b̃ − X̃i> β̃))2
i
=
X
=
X
X
ρ00 (Xi> β̂) − ρ00 (X̃i> β̃) (Xi> b̃ − X̃i> β̃))2 +
ρ00 (X̃i> β̃)(Xi> b̃ − X̃i> β̃)2
i
i
ρ000 (γ̃i )(Xi> β̂ − X̃i> β̃) Xi> b̃ − X̃i> β̃
i
2
+
X
2
ρ00 (X̃i> β̃) Xi> b̃ − X̃i> β̃
(99)
i
for some γ̃i between Xi> β̂ and X̃i> β̃. From Theorem 8 and Corollary 4, the first term in (99) is
O(n−1/2+o(1) ) with probability 1 − o(1). Hence, the only remaining term is the second.
In summary, we have
2Λ1 −
X
2
P
→ 0,
ρ00 (X̃i> β̃) Xi> b̃ − X̃i> β̃
(100)
i
|
{z
=v > X > Dβ̃ Xv
}
where v := b̃1
1
according to (86). On simplification, the quadratic form reduces to
−G̃−1 w
>
v > X > Dβ̃ Xv = b̃21 X·1 − X̃ G̃−1 w Dβ̃ X·1 − X̃ G̃−1 w
>
>
= b̃21 X·1
Dβ̃ X·1 − 2X·1
Dβ̃ X̃ G̃−1 w + w> G̃−1 X̃ > Dβ̃ X̃ G̃−1 w
>
= b̃21 X·1
Dβ̃ X·1 − nw> G̃−1 w
!
1/2
> 1/2
2 1
= nb̃1
X·1 Dβ̃ HDβ̃ X·1 ,
|n
{z
}
:=ξ
recalling the definitions (80), (87), and (90). Hence, the log-likelihood ratio 2Λ1 simplifies to nb̃21 ξ + oP (1)
on An .
Finally, rewrite v > X > Dβ̃ Xv as n(b̃21 − β̂12 )ξ + nβ̂12 ξ. To analyze the first term, note that
1
n|b̃21 − β̂12 | = n|b̃1 − β̂1 | · |b̃1 + β̂1 | ≤ n|b̃1 − βˆ1 |2 + 2n|b̃1 | · |b̃1 − β̂1 | . n− 2 +o(1)
29
(101)
with probability 1 − o(1) in view of Theorem 8 and Corollary 3. It remains to analyze ξ. Recognize that X·1
1/2
1/2
is independent of Dβ̃ HDβ̃ . Applying the Hanson-Wright inequality [27, 44] and the Sherman-MorrisonWoodbury formula (e.g. [26]) leads to the following lemma:
Lemma 9. Let α̃ =
1
−1
),
n Tr(G̃
where G̃ =
1
>
n X̃ Dβ̃ X̃.
Then one has
1 > 1/2
p−1
1/2
− α̃ X·1
Dβ̃ HDβ̃ X·1 . n−1/2+o(1)
n
n
(102)
with probability approaching one as n → ∞.
Proof: See Appendix F.
In addition, if one can show that α̃ is bounded away from zero with probability 1 − o(1), then it is seen from
Lemma 9 that
p P
ξ−
→ 0.
(103)
nα̃
To justify the above claim, we observe that since ρ00 is bounded, λmax (G̃) . λmax (X̃ > X̃)/n . 1 with
exponentially high probability (Lemma 2). This yields
α̃ = Tr(G̃−1 )/n & p/n
with probability 1 − o(1). On the other hand, on An one has
α̃ ≤ p/(nλmin (G̃)) . p/n.
Hence, it follows that ξ = Ω(1) with probability 1 − o(1). Putting this together with (101) gives the
approximation
v > X > Dβ̃ Xv = nβ̂12 ξ + o(1).
(104)
Taken collectively (100), (103) and (104) yields the desired result
P
2Λ1 − pβ̂12 /α̃ → 0.
7.4
Proof of Theorem 8
This subsection outlines the main steps for the proof of Theorem 8. To begin with, we shall express the
difference β̂ − b̃ in terms of the gradient of the negative log-likelihood function. Note that ∇`(β̂) = 0, and
hence
Xn
∇`(b̃) = ∇`(b̃) − ∇`(β̂) =
Xi [ρ0 (Xi> b̃) − ρ0 (Xi0 β̂)]
i=1
Xn
=
ρ00 (γi∗ )Xi Xi> (b̃ − β̂),
i=1
where γi∗ is between Xi> β̂ and Xi> b̃. Recalling the notation introduced in (80), this can be rearranged as
b̃ − β̂ =
1 −1
G ∇`(b̃).
n β̂,b̃
Hence, on An , this yields
kβ̂ − b̃k ≤
k∇`(b̃)k
.
λlb n
0
The next step involves expressing ∇`(b̃) in terms of the difference b̃ −
.
β̃
30
(105)
Lemma 10. On the event An (84), the negative log-likelihood evaluated at the surrogate b̃ obeys
n
X
00 ∗
0
00
>
>
∇`(b̃) =
ρ (γi ) − ρ (X̃i β̃) Xi Xi b̃ −
,
β̃
i=1
where γi∗ is some quantity between Xi> b̃ and X̃i> β̃.
Proof: The proof follows exactly the same argument as in the proof of [20, Proposition 3.11], and is thus
omitted.
0
The point of expressing ∇`(b̃) in this way is that the difference b̃ −
is known explicitly from the
β̃
definition of b̃. Invoking Lemma 10 and the definition (86) allows one to further upper bound (105) as
kβ̂ − b̃k .
1
∇`(b̃)
n
. sup ρ00 (γi∗ ) − ρ00 (X̃i> β̃)
i
. sup Xi> b̃ − X̃i> β̃ |ρ000 |∞
i
n
1X
0
Xi Xi> b̃ −
β̃
n i=1
p
1 Xn
Xi Xi> · |b̃1 | 1 + w> G̃−2 w
i=1
n
. |b̃1 | sup Xi> b̃ − X̃i> β̃
(106)
i
with probability at least 1 − exp(−Ω(n)). The last inequality here comes from our assumption that
supz |ρ000 (z)| < ∞ together with Lemmas 2 and 8.
In order to bound (106), we first make use of the definition of b̃ to reach
sup Xi> b̃ − X̃i> β̃ = |b̃1 | sup |Xi1 − X̃i> G̃−1 w|.
i
(107)
i
The following lemma provides an upper bound on supi |Xi1 − X̃i> G̃−1 w|.
Lemma 11. With G̃ and w as defined in (80) and (87),
> −1
o(1)
P sup Xi1 − X̃i G̃ w ≤ n
≥ 1 − o(1).
(108)
1≤i≤n
Proof: See Appendix G.
In view of Lemma 11, the second term in the right-hand side of (107) is bounded above by no(1) with
high probability. Thus, in both the bounds (106) and (107), it only remains to analyze the term b̃1 . To this
end, we control the numerator and the denominator of b̃1 separately.
>
• Recall from the definition (91) that the numerator of b̃1 is given by X·1
r̃ and that r̃ is independent
>
of X·1 . Thus, conditional on X̃, the quantity X·1 r̃ is distributed as a Gaussian with mean zero and
variance
Xn
2
σ2 =
ρ0 (X̃i> β̃) .
i=1
0
Since |ρ (x)| = O(|x|), the variance is bounded by
Xn
σ 2 . β̃ >
X̃i X̃i> β̃ . nkβ̃k2 . n
i=1
(109)
with probability at least 1 − exp(−Ω(n))), a consequence from Theorem 4 and Lemma 2. Therefore,
with probability 1 − o(1), we have
1
>
√ X·1
r̃ . no(1) .
(110)
n
31
• We now move on to the denominator of b̃1 in (91). In the discussion following Lemma 9 we showed
1/2
1
> 1/2
n X·1 Dβ̃ HDβ̃ X·1 = Ω(1) with probability 1 − o(1).
Putting the above bounds together, we conclude
1
P |b̃1 | . n− 2 +o(1) = 1 − o(1).
(111)
Substitution into (106) and (107) yields
kβ̂ − b̃k . n−1+o(1)
and
sup Xi> b̃ − X̃i> β̃ . n−1/2+o(1)
i
with probability 1 − o(1) as claimed.
8
Discussion
In this paper, we derived the high-dimensional asymptotic distribution of the LLR under our modelling
assumptions. In particular, we showed that the LLR is inflated vis a vis the classical Wilks’ approximation
and that this inflation grows as the dimensionality κ increases. This inflation is typical of high-dimensional
problems, and one immediate practical consequence is that it explains why classically computed p-values are
completely off since they tend to be far too small under the null hypothesis. In contrast, we have shown
in our simulations that our new limiting distribution yields reasonably accurate p-values in finite samples.
Having said this, our work raises a few important questions that we have not answered and we conclude this
paper with a couple of them.
• We expect that our results continue to hold when the covariates are not normally distributed, see
Section 3 for some numerical evidence in this direction. To be more precise, we expect the same
limiting distribution to hold when the variables are simply sub-Gaussian. If this were true, then this
would imply that our rescaled chi-square has a form of universal validity.
• The major limitation of our work is arguably the fact that our limiting distribution holds under the
global null; that is, under the assumption that all the regression coefficients vanish. It is unclear to
us how the distribution would change in the case where the coefficients are not all zero. In particular,
would the limiting distribution depend upon the unknown values of these coefficients? Are there
assumptions under which it would not? Suppose for instance that we model the regression coefficients
as i.i.d. samples from the mixture model
(1 − )δ0 + Π? ,
where 0 < < 1 is a mixture parameter, δ0 is a point mass at zero and Π? is a distribution with
vanishing mass at zero. Then what would we need to know about and Π? to compute the asymptotic
distribution of the LLR under the null?
Acknowledgements
E. C. was partially supported by the Office of Naval Research under grant N00014-16-1-2712, and by the
Math + X Award from the Simons Foundation. Y. C. and P. S. are grateful to Andrea Montanari for his help
in understanding AMP and [18]. Y. C. thanks Kaizheng Wang and Cong Ma for helpful discussion about [20],
and P. S. thanks Subhabrata Sen for several helpful discussions regarding this project. E. C. would like to
thank Iain Johnstone for a helpful discussion as well.
32
A
A.1
Proofs for Eigenvalue Bounds
Proof of Lemma 3
p
√
Fix ≥ 0 sufficiently small. For any given S ⊆ [n] obeying |S| = (1 − )n and 0 ≤ t ≤ 1 − − p/n it
follows from [51, Corollary 5.35] that
!
r
2
√
√ 2
1 p
p
1X
√
>
λmin
<
Xi Xi
|S| − p − t n =
1−−
−t
n
n
n
i∈S
2
2
n
. Taking the union bound over all possible
holds with probability at most 2 exp − t 2|S| = 2 exp − (1−)t
2
subsets S of size (1 − )n gives
(
1
λmin
n
P ∃S ⊆ [n] with |S| = (1 − )n s.t.
!
X
Xi Xi>
<
√
r
1−−
i∈S
p
−t
n
2 )
n
(1 − ) t2 n
2 exp −
(1 − )n
2
2
(1 − ) t
n ,
≤ 2 exp nH () −
2
≤
where the last line is a consequence of the inequality
A.2
n
(1−)n
≤ enH() [16, Example 11.1.3].
Proof of Lemma 4
Define
SB (β) := i : |Xi> β| ≤ Bkβk
for any B > 0 and any β. Then
n
X
ρ00 Xi> β Xi Xi>
X
i=1
ρ00 Xi> β Xi Xi>
i∈SB (β)
inf
z:|z|≤Bkβk
ρ00 (z)
X
Xi Xi> .
i∈SB (β)
If one also has |SB (β) | ≥ (1 − )n (for ≥ 0 sufficiently small), then this together with Lemma 3 implies
that
r
2
n
√
1 X 00
p
ρ Xi> β Xi Xi>
inf
ρ00 (z)
1−−
−t I
n i=1
n
z:|z|≤Bkβk
2
with probability at least 1 − 2 exp − (1−)t
− H () n .
2
Thus if we can ensure that with high probability, |SB (β) | ≥ (1 − )n holds simultaneously for all β, then
we are done. From Lemma 2 we see that n1 X > X ≤ 9 with probability exceeding 1 − 2 exp (−n/2). On
this event,
2
kXβk ≤ 9nkβk2 ,
∀β.
(112)
On the other hand, the definition of SB (β) gives
2
kXβk ≥
X
2
Xi> β
i∈S
/ B (β)
|SB (β)|
B 2 kβk2 .
≥ n − |SB (β)| (Bkβk) = n 1 −
n
2
Taken together, (112) and (113) yield
|SB (β)| ≥
1−
9
B2
33
n,
∀β
(113)
with probability at least 1−2 exp(−n/2). Therefore, with probability 1−2 exp(−n/2), S3/√ (β) ≥ (1 − ) n
q
holds simultaneously for all β. Putting the above results together and setting t = 2 H()
1− give
n
X
ρ
00
Xi> β
Xi Xi>
00
ρ (z)
inf
z:|z|≤
i=1
√
r
1−−
3kβk
√
!2
r
p
H()
−2
I
n
1−
simultaneously for all β with probability at least 1 − 2 exp (−nH ()) − 2 exp (−n/2).
B
Proof of Lemma 5
Applying an integration by parts leads to
Z ∞
1
E [Ψ0 (τ Z; b)] =
Ψ0 (τ z; b)φ(z)dz = Ψ(τ z; b)φ(z)
τ
−∞
Z
1 ∞
Ψ(τ z; b)φ0 (z)dz
= −
τ −∞
with φ(z) =
√1
2π
∞
−∞
−
1
τ
Z
∞
Ψ(τ z; b)φ0 (z)dz
−∞
exp(−z 2 /2). This reveals that
0
G (b)
=
=
Z
Z
ρ0 proxbρ (τ z)
1 ∞ ∂Ψ(τ z; b) 0
1 ∞
φ0 (z)dz
−
φ (z)dz = −
τ −∞
∂b
τ −∞ 1 + bρ00 proxbρ (τ z)
!
Z
ρ0 proxbρ (−τ z)
ρ0 proxbρ (τ z)
1 ∞
−
φ0 (z)dz,
τ 0
1 + xρ00 proxbρ (−τ z)
1 + xρ00 proxbρ (τ z)
(114)
where the second identity comes from [18, Proposition 6.4], and the last identity holds since φ0 (z) = −φ0 (−z).
Next, we claim that
(a) The function h (z) :=
ρ0 (z)
1+bρ00 (z)
is increasing in z;
(b) proxbρ (z) is increasing in z.
These two claims imply that
ρ0 proxbρ (−τ z)
1 + bρ00 proxbρ (−τ z)
−
ρ0 proxbρ (τ z)
< 0,
1 + bρ00 proxbρ (τ z)
∀z > 0,
which combined with the fact φ0 (z) < 0 for z > 0 reveals
sign
ρ0 proxbρ (−τ z)
ρ0 proxbρ (τ z)
−
1 + bρ00 proxbρ (−τ z)
1 + bρ00 proxbρ (τ z)
!
!
0
φ (z)
= 1,
∀z > 0.
In other words, the integrand in (114) is positive, which allows one to conclude that G0 (b) > 0.
We then move on to justify (a) and (b). For the first, the derivative of h is given by
h0 (z) =
ρ00 (z) + b(ρ00 (z))2 − bρ0 (z)ρ000 (z)
2
(1 + bρ00 (z))
.
Since ρ0 is log concave, this directly yields (ρ00 )2 −ρ0 ρ000 > 0. As ρ00 > 0 and b ≥ 0, the above implies h0 (z) > 0
∂proxbρ (z)
for all z. The second claim follows from
≥ 1+bkρ1 00 k∞ > 0 (cf. [18, Equation (56)]).
∂z
34
It remains to analyze the behavior of G in the limits when b → 0 and b → ∞. From [18, Proposition 6.4],
G(b) can also be expressed as
1
G(b) = 1 − E
.
1 + bρ00 (proxbρ (τ Z))
Since ρ00 is bounded and the integrand is at most 1, the dominated convergence theorem gives
lim G(b) = 0.
b→0
When b → ∞, bρ00 (proxbρ (τ z)) → ∞ for a fixed z. Again by applying the dominated convergence theorem,
lim G(b) = 1.
b→∞
It follows that limb→0 G(b) < κ < limb→∞ G(b) and, therefore, G(b) = κ has a unique positive solution.
Remark 3. Finally, we show that the logistic and the probit effective links obey the assumptions of Lemma
5. We work with a fixed τ > 0.
• A direct computation shows that ρ0 is log-concave for the logistic model. For the probit, it is well-known
that the reciprocal of the hazard function (also known as Mills’ ratio) is strictly log-convex [4].
• To check the other condition, recall that the proximal mapping operator satisfies
bρ0 (proxbρ (τ z)) + proxbρ (τ z) = τ z.
(115)
For a fixed z, we claim that if b → ∞, proxbρ (τ z) → −∞. To prove this claim, we start by assuming
that this is not true. Then either proxbρ (τ z) is bounded or diverges to ∞. If it is bounded, the LHS
above diverges to ∞ while the RHS is fixed, which is a contradiction. Similarly if proxbρ (τ z) diverges
to ∞, the left-hand side of (115) diverges to ∞ while the right-hand side is fixed, which cannot be
true as well. Further, when b → ∞, we must have proxbρ (τ z) → −∞, bρ0 (proxbρ (τ z)) → ∞, such that
the difference of these two is τ z. Observe that for the logistic, ρ00 (x) = ρ0 (x)(1 − ρ0 (x)) and for the
probit, ρ00 (x) = ρ0 (x)(ρ0 (x) − x) [45]. Hence, combining the asymptotic behavior of proxbρ (τ z) and
bρ0 (proxbρ (τ z)), we obtain that bρ00 (proxbρ (τ z)) diverges to ∞ in both models when b → ∞.
C
C.1
Proof of Lemma 6
Proof of Part (i)
Recall from [18, Proposition 6.4] that
"
#
1
.
κ = E [Ψ (τ Z; b(τ ))] = 1 − E
1 + b(τ )ρ00 proxb(τ )ρ (τ Z)
0
If we denote c := proxbρ (0), then b(0) is given by the following relation:
1−κ=
1
1 + b(0)ρ00 (c)
=⇒
b(0) =
κ
>0
ρ00 (c)(1 − κ)
as ρ00 (c) > 0 for any given c > 0. In addition, since ρ0 (c) > 0, we have
V(0) =
Ψ(0, b(0))2 (a) b(0)2 ρ0 (c)2
=
> 0,
κ
κ
where (a) comes from (20).
35
(116)
C.2
Proof of Part (ii)
We defer the proof of this part to the supplemental materials [47].
D
Proof of Part (ii) of Theorem 4
As discussed in Section 5.2.2, it suffices to (1) construct a set {Bi | 1 ≤ i ≤ N } that forms a cover of the
cone A defined in (50), and (2) upper bound P {{Xβ | β ∈ Rp } ∩ Bi 6= {0}}. In what follows, we elaborate
on these two steps.
• Step 1. Generate N = exp 22 p i.i.d. points z (i) ∼ N (0, p1 Ip ), 1 ≤ i ≤ N , and construct a collection
of convex cones
z (i)
p
u, (i)
Ci := u ∈ R
≥ kuk ,
1 ≤ i ≤ N.
kz k
In words, Ci consists of all directions that have nontrivial positive correlation with z (i) . With high probability, this collection {Ci | 1 ≤ i ≤ N } forms a cover of Rp , a fact which is an immediate consequence
of the following lemma.
Lemma 12. Consider any given constant 0 < < 1, and let N = exp 22 p . Then there exist
some
positive universal constants c5 , C5 > 0 such that with probability exceeding 1 − C5 exp −c5 2 p ,
N
X
1{hx,z(i) i≥kxkkz(i) k} ≥ 1
i=1
holds simultaneously for all x ∈ Rp .
With our family {Ci | 1 ≤ i ≤ N } we can introduce
n
(i)
X
√
z
Bi := Ci ∩ u ∈ Rn |
max {−uj , 0} ≤ n u, (i)
,
kz k
1 ≤ i ≤ N,
(117)
j=1
which in turn forms a cover of the nonconvex cone A defined in (50). DTo justifyE this, note that for
(i)
any u ∈ A, one can find i ∈ {1, · · · , N } obeying u ∈ Ci , or equivalently, u, kzz(i) k ≥ kuk, with high
probability. Combined with the membership to A this gives
n
X
√
√
z (i)
max {−uj , 0} ≤ 2 nkuk ≤ n u, (i)
,
kz k
j=1
indicating that u is contained within some Bi .
• Step 2. We now move on to control P {{Xβ | β ∈ Rp } ∩ Bi 6= {0}}. If the statistical dimensions of
the two cones obey δ (Bi ) < n − δ ({Xβ | β ∈ Rp }) = n − p, then an application of [3, Theorem I] gives
(
2 )
1 n − δ ({Xβ | β ∈ Rp }) − δ (Bi )
p
√
P {{Xβ | β ∈ R } ∩ Bi 6= {0}} ≤ 4 exp −
8
n
(
)
2
(n − p − δ(Bi ))
≤ 4 exp −
.
(118)
8n
It then comes down to upper bounding δ(Bi ), which is the content of the following lemma.
36
Lemma 13. Fix > 0. When n is sufficiently large, the statistical dimension of the convex cone Bi
defined in (117) obeys
√ 3
√
1
4
δ(Bi ) ≤
(119)
+ 2 2 + 10H(2 ) n,
2
where H(x) := −x log x − (1 − x) log(1 − x).
Substitution into (118) gives
P {{Xβ | β ∈ Rp } ∩ Bi 6= {0}}
2
√ 3
√
1
4 − 10H(2
−
2
2
)
n
−
p
2
≤ 4 exp −
8n
(
2 )
√ 3
√
p
1 1
4
− 2 2 − 10H(2 ) −
n .
= 4 exp −
8 2
n
(120)
Finally, we prove Lemmas 12-13 in the next subsections. These are the only remaining parts for the proof
of Theorem 4.
D.1
Proof of Lemma 12
To begin with, it is seen that all kz (i) k concentrates around 1. Specifically, apply [29, Proposition 1] to get
r
2t
t
(i) 2
P kz k > 1 + 2
+
≤ e−t ,
p
p
and set t = 32 p to reach
o
n
o
n
√
2
P kz (i) k2 > 1 + 10 ≤ P kz (i) k2 > 1 + 2 3 + 62 ≤ e−3 p .
Taking the union bound we obtain
n
o
P ∃1 ≤ i ≤ N s.t. kz (i) k2 > 1 + 10
2
≤ N e−3
p
2
= e− p .
(121)
Next, we note that it suffices to prove Lemma 12 for all unit vectors x. The following lemma provides a
bound on z (i) , x for any fixed unit vector x ∈ Rp .
Lemma 14. Consider any fixed unit vector x ∈ Rp and any given constant 0 < < 1, and set N =
exp 22 p . There exist positive universal constants c5 , c6 , C6 > 0 such that
P
)
7 2
7
1{hz(i) ,xi≥ 1 } ≤ exp (1 − o (1)) p
≤ exp −2 exp (1 − o (1)) 2 p
.
2
4
4
i=1
(N
X
(122)
Recognizing that Lemma 12 is a uniform result, we need to extend Lemma 14 to all
x simultaneously,
which we achieve via the standard covering argument. Specifically, one can find a set C := x(j) ∈ Rp | 1 ≤ j ≤ K
p
of unit vectors with cardinality K = 1 + 2p2 to form a cover of the unit ball of resolution p−2 [51, Lemma
5.2]; that is, for any unit vector x ∈ Rp , there exists a x(j) ∈ C such that
kx(j) − xk ≤ p−2 .
37
Apply Lemma 14 and take the union bound to arrive at
N
X
7
1{hz(i) ,x(j) i≥ 1 } ≥ exp (1 − o(1)) 2 p > 1,
1≤j≤K
(123)
2
4
i=1
with probability exceeding 1−K exp −2 exp (1 − o(1)) 47 2 p ≥ 1−exp −2 (1 − o (1)) exp (1 − o(1)) 47 2 p .
This guarantees that for each x(j) , one can find at least one z (i) obeying
D
E 1
z (i) , x(j) ≥ .
2
This result together with (121) yields that with probability exceeding 1 − C exp −c2 p , for some universal
constants C, c > 0.
E
E D
E
D
D
E D
≥
z (i) , x(j) − kz (i) k · kx(j) − xk
z (i) , x ≥ z (i) , x(j) − z (i) , x(j) − x
1
1
1
1
kz (i) k − 2 kz (i) k
− 2 kz (i) k ≥ √ 2
2
p
p
1 + 10
1
kz (i) k
≥
30
holds simultaneously for all unit vectors x ∈ Rp . Since > 0 can be an arbitrary constant, this concludes
the proof.
≥
Proof of Lemma 14: Without loss of generality, it suffices to consider x = e1 = [1, 0, · · · , 0]> . For any
t > 0 and any constant ζ > 0, it comes from [2, Theorem A.1.4] that
(
)
N
1 X
√
√
1{hz(i) ,e1 i<ζ } > (1 + t) Φ (ζ p) ≤ exp −2t2 Φ2 (ζ p) N .
P
N i=1
√
Setting t = 1 − Φ ζ p gives
(
)
N
1 X
√
√
√ 2
√
P
1{hz(i) ,e1 i<ζ } > (2 − Φ (ζ p)) Φ (ζ p) ≤ exp −2 (1 − Φ (ζ p)) Φ2 (ζ p) N .
N i=1
Recall that for any t > 1, one has (t−1 − t−3 )φ(t) ≤ 1 − Φ(t) ≤ t−1 φ(t) which implies that
(1 + o (1)) ζ 2 p
√
1 − Φ (ζ p) = exp −
.
2
Taking ζ = 12 , we arrive at
√
√
(2 − Φ (ζ p)) Φ (ζ p)
=
√ 2
√
(1 − Φ (ζ p)) Φ2 (ζ p)
=
1
1 − exp − (1 + o (1)) ζ 2 p = 1 − exp − (1 + o (1)) 2 p ,
4
1
1
exp − (1 + o (1)) ζ 2 p = exp − (1 + o (1)) 2 p .
4
N
This justifies that
(N
(
)
)
N
X
1 2
1 X
√
√
P
1{hz(i) ,e1 i≥ 1 } ≤ N exp − (1 + o (1)) p
=P
1 (i)
> (2 − Φ (ζ p)) Φ (ζ p)
2
4
N i=1 {hz ,e1 i<ζ }
i=1
1 2
≤ exp −2 exp − (1 + o (1)) p N
4
7
= exp −2 exp (1 − o (1)) 2 p
4
as claimed.
38
D.2
Proof of Lemma 13
First of all, recall from the definition (17) that
h
i
2
2
2
2
δ(Bi ) = E kΠBi (g)k = E kgk − min kg − uk = n − E min kg − uk
u∈Bi
u∈Bi
2
≤ n − E min kg − uk ,
u∈Di
where g ∼ N (0, In ), and Di is a superset of Bi defined by
n
o
Xn
√
Di := u ∈ Rn |
max {−uj , 0} ≤ nkuk .
(124)
j=1
Recall from the triangle inequality that
kg − uk
≥
kuk − kgk > kgk = kg − 0k,
∀u : kuk > 2kgk.
Since 0 ∈ Di , this implies that
arg min kg − uk ≤ 2kgk,
u∈Di
revealing that
2
E min kg − uk = E
u∈Di
2
kg − uk
min
.
u∈Di ,kuk≤2kgk
In what follows, it suffices to look at the set of u’s within Di obeying kuk ≤ 2kgk, which verify
Xn
√
√
max {−uj , 0} ≤ nkuk ≤ 2 nkgk.
(125)
j=1
It is seen that
kg − uk2
≥
X
2
(gi − ui ) =
≥
gi2
+
i:gi <0,ui ≥0
≥
X
i:gi <0, −
i:gi <0, −
X
i:gi <0, ui >−
√
X
+
i:gi <0, −
i:gi <0, ui ≤−
√
n
(gi − ui )
n kgk<ui <0
gi2 − 2ui gi
√
n kgk<ui <0
X
i:gi <0, −
2ui gi .
√
(126)
n kgk<ui <0
1. Regarding the first term of (126), we first recognize that
r
Pn
P
√
max {−ui , 0}
i: ui <0 |ui |
i | ui ≤ −
kgk ≤ p
= i=1 p
≤ 2 n,
n
n kgk
n kgk
where the last inequality follows from the constraint (125). As a consequence,
X
X
X
gi2 ≥
gi2 −
gi2
√
√
i:gi <0
i:gi <0, ui >−
2
(gi − ui )
kgk
2
√
gi2 −
n kgk
n kgk<ui <0
X
+
√
X
gi2 +
i:gi <0,ui ≥0
≥
X
i:gi <0,ui ≥0
X
i:gi <0
X
n kgk
i:ui ≤−
X
≥
i:gi <0
39
gi2 −
n kgk
max
√
S⊆[n]: |S|=2 n
X
i∈S
gi2 .
2. Next, we turn to the second term of (126), which can be bounded by
v
u
u
X
X
u
ui gi ≤ u
u2i
t
√
√
i:gi <0, −
n kgk<ui <0
i:gi <0, −
≤
v
u
u
t
√
i:gi <0, −
!
i:−
≤
n kgk<ui <0
X
n kgk<ui <0
v
ur
u
t
kgk
n
· kgk2
|ui |
i:ui <0
!
X
n kgk<ui <0
!
X
|ui |
√ max
gi2
· kgk2 ≤
|ui |
√
3
2 4 kgk2 ,
i:ui <0
where the last inequality follows from the constraint (125).
Putting the above results together, we have
X
2
kg − uk ≥
gi2 −
i:gi <0
√
S⊆[n]: |S|=2 n
for any u ∈ Di obeying kuk ≤ 2kgk, whence
X
2
≥ E
gi2 −
E min kg − uk
u∈Di
i:gi <0
=
X
max
√ 3
1
− 2 2 4
2
√ 3
gi2 − 2 2 4 kgk2
i∈S
max √
S⊆[n]: |S|=2 n
X
√
3
gi2 − 2 2 4 kgk2
i∈S
#
"
n−E
max √
S⊆[n]: |S|=2 n
X
gi2
.
(127)
i∈S
√
Finally, it follows from [29, Proposition 1] that for any t > 2 n,
(
)
)
(
X
X
p
2
2
P
gi ≥ 5t ≤ P
gi ≥ |S| + 2 |S|t + 2t ≤ e−t ,
i∈S
i∈S
which together with the union bound gives
(
)
X
2
P
max √
gi ≥ 5t ≤
S⊆[n]: |S|=2 n
i∈S
(
X
√
S⊆[n]: |S|=2 n
P
)
X
gi2
≥ 5t
√
≤ exp H 2 n − t .
i∈S
This gives
"
E
#
max √
S⊆[n]: |S|=2 n
X
gi2
Z
=
P
0
i∈S
≤
(
∞
)
max √
S⊆[n]: |S|=2 n
√
5H 2 n +
√
< 10H 2 n,
Z
X
gi2
≥ t dt
i∈S
∞
√
1
exp
H
2
n
−
t
dt
√
5
5H (2 )n
for any given > 0 with the proviso that n is sufficiently large. This combined with (127) yields
√ 3
√
1
2
4
E min kg − uk
≥
− 2 2 − 10H(2 ) n
u∈Di
2
as claimed.
40
(128)
E
Proof of Lemma 8
Throughout, we shall restrict ourselves on the event An as defined in (84), on which G̃ λlb I. Recalling
the definitions of G̃ and w from (80) and (87), we see that
w> G̃−2 w =
≤
−2
1 >
X̃ Dβ̃ X̃
X̃ > Dβ̃ X·1
n
−2
1
1 >
D X̃
X̃ Dβ̃ X̃
X̃ > Dβ̃ .
n β̃
n
1 >
X D X̃
n2 ·1 β̃
>
X·1
n
2
1/2
If we let the singular value decomposition of √1n Dβ̃ X̃ be U ΣV > , then a little algebra gives Σ
and
−2
1 1/2
1 0
1/2
D X̃
X̃ Dβ̃ X̃
X̃ > Dβ̃ = U Σ−2 U > λ−1
lb I.
n β̃
n
(129)
√
λlb I
Substituting this into (129) and using the fact kX·1 k2 . n with high probability (by Lemma 2), we obtain
w> G̃−2 w .
1
kX·1 k2 . 1
nλL
with probability at least 1 − exp(−Ω(n)).
F
Proof of Lemma 9
Throughout this and the subsequent sections, we consider Hn and Kn to be two diverging sequences with
the following properties:
Hn = o (n ) , Kn = o (n ) , n2 exp −c1 Hn2 = o(1), n exp −c2 Kn2 = o(1),
(130)
for any constants ci > 0, i = 1, 2 and any > 0. This lemma is an analogue of [20, Proposition 3.18]. We
modify and adapt the proof ideas to establish the result in our setup. Throughout we shall restrict ourselves
to the event An , on which G̃ λlb I.
Due to independence between X·1 and {Dβ̃ , H}, one can invoke the Hanson-Wright inequality [44,
Theorem 1.1] to yield
P
!
1 > 1/2
1 1/2
1/2
1/2
> t H, Dβ̃
X D HDβ̃ X·1 − Tr Dβ̃ HDβ̃
n ·1 β̃
n
2
t
t
≤ 2 exp −c min
,
K24 D 1/2 HD 1/2 2 K 2 D 1/2 HD 1/2
n
n
F
β̃
β̃
β̃
β̃
2
t
t
,
≤ 2 exp −c min
,
K 4 D 1/2 HD 1/2 2 K 2 D 1/2 HD 1/2
n
n
β̃
β̃
β̃
β̃
√
1/2
1/2
where k.kF denotes the Frobenius norm. Choose t = C 2 Dβ̃ HDβ̃ Hn / n with C > 0 a sufficiently
large constant, and take Hn to be as in (130). Substitution into the above inequality and unconditioning
41
give
P
1 > 1/2
1 1/2
1 2
1/2
1/2
1/2
1/2
√
C Hn kDβ̃ HDβ̃ k
X D HDβ̃ X·1 − Tr Dβ̃ HDβ̃
>
n ·1 β̃
n
n
4 2
√
C Hn C 2 nHn
,
= C exp −cHn2 = o(1),
≤ 2 exp −c min
K4
K2
(131)
for some universal constants C, c > 0.
1/2
1/2
We are left to analyzing Tr Dβ̃ HDβ̃ . Recall from the definition (90) of H that
1/2
1/2
Dβ̃ HDβ̃
= Dβ̃ −
1
D X̃ G̃−1 X̃ > Dβ̃ ,
n β̃
and, hence,
Tr
1/2
1/2
Dβ̃ HDβ̃
=
n
X
ρ
00
(X̃i> β̃)
i=1
ρ00 (X̃i> β̃)2 > −1
X̃i G̃ X̃i
−
n
!
.
(132)
This requires us to analyze G̃−1 carefully. To this end, recall that the matrix G̃(i) defined in (81) obeys
G̃(i) = G̃ −
1 00 >
ρ (X̃ β̃)X̃i X̃i> .
n
Invoking Sherman-Morrison-Woodbury formula (e.g. [26]), we have
G̃
−1
=
G̃−1
(i)
−
ρ00 (X̃i> β̃) −1
G̃(i) X̃i X̃i> G̃−1
(i)
n
ρ00 (X̃i> β̃)
X̃i> G̃−1
(i) X̃i
n
1+
It follows that
X̃i> G̃−1 X̃i
=
X̃i> G̃−1
(i) X̃i
−
.
ρ00 (X̃i> β̃)
2
(Xi> G̃−1
(i) X̃i )
n
1+
ρ00 (X̃i> β̃)
X̃i> G̃−1
(i) X̃i
n
(133)
,
which implies that
X̃i> G̃−1
(i) X̃i
X̃i> G̃−1 X̃i =
1+
ρ00 (X̃i> β̃)
X̃i> G̃−1
(i) X̃i
n
.
(134)
The relations (132) and (134) taken collectively reveal that
n
1X
1 1/2
1/2
=
Tr Dβ̃ HDβ̃
n
n i=1 1 +
ρ00 (X̃i β̃)
ρ00 (X̃i> β̃)
X̃i> G̃−1
(i) X̃i
n
.
(135)
We shall show that the trace above is close to Tr(I − H) up to some factors. For this purpose we analyze
the latter quantity in two different ways. To begin with, observe that
1/2 !
1/2
Tr(I − H) = Tr
Dβ̃ X̃ G̃−1 X̃ > Dβ̃
n
= Tr(G̃G̃−1 ) = p − 1.
(136)
On the other hand, it directly follows from the definition of H and (134) that the ith diagonal entry of H
is given by
1
Hi,i =
.
ρ00 (X̃i> β̃)
> G̃−1 X̃
1+
X̃
i
i
(i)
n
42
Applying this relation, we can compute Tr(I − H) analytically as follows:
Tr(I − H) =
X
i
=
1+
X
ρ00 (X̃i> β̃)
X̃i> G̃−1
(i) X̃i
n
1+
1
n Tr
− ρ00 (X̃i> β̃)α̃
ρ00 (X̃i> β̃)
X̃i> G̃−1
(i) X̃i
n
ρ00 (X̃i> β̃)α̃Hi,i +
X ρ00 (X̃i> β̃)
i
where α̃ :=
(137)
ρ00 (X̃i> β̃)
X̃i> G̃−1
(i) X̃i
n
X ρ00 (X̃i> β̃)α̃ +
i
=
ρ00 (X̃i> β̃)
X̃i> G̃−1
(i) X̃i
n
1+
i
1
> −1
n X̃i G̃(i) X̃i
− α̃
ρ00 (X̃i> β̃)
X̃i> G̃−1
(i) X̃i
n
,
(138)
G̃−1 .
1/2
1/2
Observe that the first quantity in the right-hand side above is simply α̃Tr Dβ̃ HDβ̃
denote
1
ηi = X̃i> G̃−1
(i) X̃i − α̃.
n
. For simplicity,
(139)
Note that G̃(i) 0 on An and that ρ00 > 0. Hence the denominator in the second term in (138) is greater
than 1 for all i. Comparing (136) and (138), we deduce that
p−1
1 X 00 >
1 1/2
1/2
− Tr Dβ̃ HDβ̃
α̃ ≤ sup |ηi | ·
|ρ (X̃i β̃)| . sup |ηi |
(140)
n
n
n i
i
i
on An . It thus suffices to control supi |ηi |. The above bounds together with Lemma (85) and the proposition
below complete the proof.
Proposition 1. Let ηi be as defined in (139). Then there exist universal constants C1 , C2 , C3 > 0 such that
C1 Kn2 Hn
√
≥ 1 − C2 n2 exp −c2 Hn2 − C3 n exp −c3 Kn2
P sup |ηi | ≤
n
i
− exp (−C4 n (1 + o(1))) = 1 − o(1),
where Kn , Hn are diverging sequences as specified in (130)
Proof of Proposition 1: Fix any index i. Recall that β̃[−i] is the MLE when the 1st predictor and ith
observation are removed. Also recall the definition of G̃[−i] in (83). The proof essentially follows three steps.
First, note that X̃i and G̃[−i] are independent. Hence, an application of the Hanson-Wright inequality [44]
yields that
!
2
1 −1
t
1 > −1
t
X̃i G̃[−i] X̃i − Tr G̃[−i] > t G̃[−i] ≤ 2 exp −c min
,
P
K24 G̃−1 2 K 2 G̃−1
n
n
[−i]
n
[−i] F
n
2
t
t
.
≤ 2 exp −c min
,
K 4 G̃−1 2 K 2 G̃−1
n
[−i]
n
[−i]
√
We choose t = C 2 G̃−1
[−i] Hn / n, where C > 0 is a sufficiently large constant. Now marginalizing gives
P
1 > −1
1
H
√n
X̃i G̃[−i] X̃i − Tr G̃−1
> C 2 G̃−1
[−i]
[−i]
n
n
n
43
≤ 2 exp −c min
≤ 2 exp −C 0 Hn2 ,
√
C 4 Hn2 C 2 nHn
,
K4
K2
where C 0 > 0 is a sufficiently large constant. On An , the spectral norm G̃−1
(i) is bounded above by λlb for
all i. Invoking (85) we obtain that there exist universal constants C1 , C2 , C3 > 0 such that
1 > −1
1 −1
Hn
P sup X̃i G̃[−i] X̃i − Tr G̃[−i] > C1 √
≤ C2 n exp −C3 Hn2 .
(141)
n
n
n
i
−1
> −1
> −1
The next step consists of showing that Tr G̃−1
[−i] (resp. X̃i G̃[−i] X̃i ) and Tr G̃(i) (resp. X̃i G̃(i) X̃i ) are
uniformly close across all i. This is established in the following lemma.
Lemma 15. Let G̃(i) and G̃[−i] be defined as in (81) and (83), respectively. Then there exist universal
constants C1 , C2 , C3 , C4 , c2 , c3 > 0 such that
Kn2 Hn
1 > −1
1 > −1
P sup X̃i G̃(i) X̃i − X̃i G̃[−i] X̃i ≤ C1 √
n
n
n
i
= 1 − C2 n2 exp −c2 Hn2 − C3 n exp −c3 Kn2 − exp (−C4 n (1 + o(1))) = 1 − o(1), (142)
1
Kn2 Hn
1
−1
√
−
≤
C
Tr
G̃
P sup Tr G̃−1
1
(i)
[−i]
n
n
n
i
2
2
= 1 − C2 n exp −c2 Hn − C3 n exp −c3 Kn2 − exp (−C4 n (1 + o(1))) = 1 − o(1), (143)
where Kn , Hn are diverging sequences as defined in (130).
This together with (141) yields that
1
Kn2 Hn
1
−1
√
X̃
−
>
C
Tr(
G̃
)
P sup X̃i> G̃−1
i
1
(i)
(i)
n
n
n
i
≤ C2 n2 exp −c2 Hn2 + C3 n exp −c3 Kn2 + exp (−C4 n (1 + o(1))) . (144)
1
−1
The final ingredient is to establish that n1 Tr G̃−1
are uniformly close across i.
(i) and n Tr G̃
Lemma 16. Let G̃ and G̃(i) be as defined in (80) and (81), respectively. Then one has
1
−1
−
Tr
G̃
≤
P Tr G̃−1
≥ 1 − exp (−Ω(n)) .
(i)
λlb
(145)
This completes the proof.
Proof of Lemma 15: For two invertible matrices A and B of the same dimensions, the difference of their
inverses can be written as
A−1 − B −1 = A−1 (B − A)B −1 .
Applying this identity, we have
−1
−1
G̃−1
−
G̃
=
G̃
G̃
−
G̃
G̃−1
[−i]
(i)
(i)
[−i]
(i)
[−i] .
From the definition of these matrices, it follows directly that
1 X 00
ρ X̃j> β̃[−i] − ρ00 X̃j> β̃ X̃j X̃j> .
G̃[−i] − G̃(i) =
n
(146)
j:j6=i
As ρ000 is bounded, by the mean-value theorem, it suffices to control the differences Xj> β̃[−i] − X̃j> β̃
uniformly across all j. This is established in the following lemma, the proof of which is deferred to Appendix
H.
44
Lemma 17. Let β̂ be the full model MLE and β̂[−i] be the MLE when the ith observation is dropped.
Let qi be as described in Lemma 18 and Kn , Hn be as in (130). Then there exist universal constants
C1 , C2 , C3 , C4 , c2 , c3 > 0 such that
!
Kn2 Hn
>
>
P sup Xj β̂[−i] − Xj β̂ ≤ C1 √
n
j6=i
≥ 1 − C2 n exp −c2 Hn2 − C3 exp −c3 Kn2 − exp (−C4 n (1 + o(1))) = 1 − o(1), (147)
Kn2 Hn
P sup |Xi> β̂ − proxqi ρ (Xi> β̂[−i] )| ≤ C1 √
n
i
2
≥ 1 − C2 n exp −c2 Hn − C3 exp −c3 Kn2 − exp (−C4 n (1 + o(1))) = 1 − o(1). (148)
Invoking this lemma, we see that the spectral norm of (146) is bounded above by some constant times
Kn2 Hn X
√
X̃j X̃j> /n
n
j:j6=i
with high probability as specified in (147). From Lemma 2, the spectral norm here is bounded by some
constant with probability at least 1 − c1 exp(−c2 n). These observations together with (85) and the fact that
on An the minimum eigenvalues of G̃(i) and G̃[−i] are bounded by λlb yield that
Kn2 Hn
−1
√
≤
C
−
G̃
P G̃−1
≥ 1 − C2 n exp −c2 Hn2 − C3 exp −c3 Kn2 − exp (−C4 n (1 + o(1))) .
1
(i)
[−i]
n
This is true for any i. Hence, taking the union bound we obtain
Kn2 Hn
−1
√
≤
C
−
G̃
P sup G̃−1
1
(i)
[−i]
n
i
≥ 1 − C2 n2 exp −c2 Hn2 − C3 n exp −c3 Kn2 − exp (−C4 n (1 + o(1))) . (149)
In order to establish the first result, note that
sup
i
1
kX̃i k2
−1
> −1
sup kG̃−1
X̃i> G̃−1
X̃
−
X̃
G̃
X̃
≤
sup
i
i
i
(i) − G̃[−i] k.
(i)
[−i]
n
n
i
i
To obtain the second result, note that
sup
i
p−1
1
1
−1
Tr(G̃−1
) − Tr(G̃−1
) ≤
sup kG̃−1
(i)
[−i]
(i) − G̃[−i] k.
n
n
n
i
Therefore, combining (149) and Lemma 2 gives the desired result.
Proof of Lemma 16: We restrict ourselves to the event An throughout. Recalling (133), one has
−1
Tr(G̃−1
)=
(i) ) − Tr(G̃
ρ00 (X̃i> β̃)
n
1+
X̃i> G̃−2
(i) X̃i
ρ00 (X̃i> β̃)
X̃i> G̃−1
(i) X̃i
n
.
In addition, on An we have
1
1
> −2
> −1
X̃i> G̃−1
X̃
−
X̃
G̃
X̃
=
X̃
G̃
G̃
−
λ
I
G̃−1
i
i
lb
(i)
i
(i)
(i)
(i) X̃i ≥ 0.
λlb
λlb i (i)
Combining these results and recognizing that ρ00 > 0, we get
−1
Tr(G̃−1
) ≤
(i) ) − Tr(G̃
ρ00 (X̃i> β̃)
n
1+
1
> −1
λlb X̃i G̃(i) X̃i
ρ00 (X̃i> β̃)
X̃i> G̃−1
(i) X̃i
n
≤
1
λlb
(150)
as claimed.
45
G
Proof of Lemma 11
Again, we restrict ourselves to the event An on which G̃ λlb I. Note that
X̃i> G̃−1 w =
1 > −1 >
X̃ G̃ X̃ Dβ̃ X·1 .
n i
Note that {G̃, X̃} and X·1 are independent. Conditional on X̃, the left-hand side is Gaussian with mean
zero and variance n12 X̃i> G̃−1 X̃ > Dβ̃2 X̃ G̃−1 X̃i . The variance is bounded above by
1 > −1 > 2
1
X̃ G̃ X̃ Dβ̃ X̃ G̃−1 X̃i ≤ sup ρ00 (X̃i> β̃) · 2 X̃i> G̃−1 X̃ > Dβ̃ X̃ G̃−1 X̃i
n2 i
n
i
1
1
= sup ρ00 (X̃i> β̃) · X̃i> G̃−1 X̃i . kX̃i k2
n i
n
2
σX
:=
(151)
In turn, Lemma 2 asserts that n−1 kX̃i k2 is bounded by a constant with high probability. As a result,
applying Gaussian concentration results [49, Theorem 2.1.12] gives
|X̃i> G̃−1 w| . Hn
with probability exceeding 1 − C exp −cHn2 , where C, c > 0 are universal constants.
In addition, supi |Xi1 | . Hn holds with probability exceeding 1 − C exp −cHn2 . Putting the above
results together, applying the triangle inequality |Xi1 − X̃i> G̃−1 w| ≤ |Xi1 | + |X̃i> G̃−1 w|, and taking the
union bound, we obtain
P( sup |Xi1 − X̃i> G̃−1 w| . Hn ) ≥ 1 − Cn exp −cHn2 = 1 − o(1).
1≤i≤n
H
Proof of Lemma 17
The goal of this section is to prove Lemma 17, which relates the full-model MLE β̂ and the MLE β̂[−i] . To
this end, we establish the key lemma below.
Lemma 18. Suppose β̂[−i] denote the MLE when the ith observation is dropped. Further let G[−i] be as in
(82), and define qi and b̂ as follows:
1 > −1
X G Xi ;
n i [−i]
1
0
>
b̂ = β̂[−i] − G−1
X
ρ
prox
X
β̂
.
i
[−i]
q
ρ
i
i
n [−i]
qi =
(152)
Suppose Kn , Hn are diverging sequences as in (130). Then there exist universal constants C1 , C2 , C3 > 0
such that
Kn2 Hn
P kβ̂ − b̂k ≤ C1
≥ 1 − C2 n exp(−c2 Hn2 ) − C3 exp(−c3 Kn2 ) − exp(−C4 n(1 + o(1)));
(153)
n
P sup
j6=i
Xj> β̂[−i]
−
Xj> b̂
Kn Hn
≤ C1 √
n
!
≥ 1 − C2 n exp −c2 Hn2 − C3 exp −c3 Kn2 − exp (−C4 n (1 + o(1))) . (154)
46
The proof ideas are inspired by the leave-one-observation-out approach of [20]. We however emphasize
once more that the adaptation of these ideas to our setup is not straightforward and crucially hinges on
Theorem 4, Lemma 7 and properties of the effective link function.
Proof of Lemma 18: Invoking techniques similar to that for establishing Lemma 7, it can be shown that
n
1 X 00 ∗
ρ (γi )Xi Xi> λlb I
n i=1
(155)
with probability at least 1 − exp(Ω(n)), where γi∗ is between Xi> b̂ and Xi> β̂. Denote by Bn the event
where (155) holds. Throughout this proof, we work on the event Cn := An ∩ Bn , which has probability
1 − exp (−Ω(n)). As in (105) then,
1
∇`(b̂) .
(156)
kβ̂ − b̂k ≤
nλlb
Next, we simplify (156). To this end, recall the defining relation of the proximal operator
bρ0 (proxbρ (z)) + proxbρ (z) = z,
which together with the definitions of b̂ and qi gives
Xi> b̂ = proxqi ρ Xi> β̂[−i] .
(157)
Now, let `[−i] denote the negative log-likelihood function when the ith observation is dropped, and hence
∇`[−i] β̂[−i] = 0. Expressing ∇`(b̂) as ∇`(b̂) − ∇`[−i] β̂[−i] , applying the mean value theorem, and using
the analysis similar to that in [20, Proposition 3.4], we obtain
i
1
1 X h 00 ∗
∇`(b̂) =
ρ (γj ) − ρ00 (Xj> β̂[−i] ) Xj Xj> b̂ − β̂[−i] ,
(158)
n
n
j:j6=i
where γj∗ is between Xj> b̂ and Xj> β̂[−i] . Combining (156) and (158) leads to the upper bound
kβ̂ − b̂k ≤
1
λlb
1 X
1 −1
Xj Xj> · sup ρ00 (γj∗ ) − ρ00 Xj> β̂[−i] ·
G[−i] Xi · ρ0 proxqi ρ (Xi> β̂[−i] ) . (159)
n
n
j6=i
j:j6=i
We need to control each term in the right-hand side. To start with, the first term is bounded by a
universal constant with probability 1 − exp(−Ω(n)) (Lemma 2). For the second term, since γj∗ is between
Xj> b̂ and Xj> β̂[−i] and kρ000 k∞ < ∞, we get
sup ρ00 (γj∗ ) − ρ00 (Xj> β̂[−i] ) ≤ kρ000 k∞ kXj> b̂ − Xj> β̂[−i] k
(160)
j6=i
1 > −1
Xj G[−i] Xi ρ0 proxqi ρ Xi> β̂[−i]
n
1
0
>
≤ kρ000 k∞ sup Xj> G−1
X
·
ρ
prox
(X
β̂
)
.
i
[−i]
qi ρ
i
[−i]
n j6=i
≤ kρ000 k∞
(161)
(162)
Given that {Xj , G[−i] } and Xi are independent for all j 6= i, conditional on {Xj , G[−i] } one has
> −2
Xj> G−1
X
∼
N
0,
X
G
X
.
i
j
j
[−i]
[−i]
In addition, the variance satisfies
|Xj> G−2
[−i] Xj | ≤
47
kXj k2
.n
λ2lb
(163)
with probability at least 1 − exp(−Ω(n)). Applying standard Gaussian concentration results [49, Theorem
2.1.12], we obtain
1
≥
C
H
≤ C2 exp −c2 Hn2 + exp (−C3 n (1 + o(1))) .
P √ Xj> G−1
X
(164)
1 n
i
[−i]
p
By the union bound
P
1
√ sup Xj> G−1
[−i] Xi ≤ C1 Hn
p j6=i
!
≥ 1 − nC2 exp −c2 Hn2 − exp (−C3 n (1 + o(1))) .
(165)
Consequently,
1
sup ρ00 (γj∗ ) − ρ00 (Xj> β̂[−i] ) . sup kXj> b̂ − Xj> β̂[−i] k . √ Hn ρ0 proxqi ρ (Xi> β̂[−i] ) .
n
j6=i
j6=i
In addition, the third term in the right-hand side of (159) can be upper bounded as well since
1
1
1 q > −2
|Xi G[−i] Xi | . √
kG−1
X
k
=
i
[−i]
n
n
n
(166)
(167)
with high probability.
It remains to bound ρ0 proxqi ρ (Xi> β̂[−i] ) . To do this, we begin by considering ρ0 (proxcρ (Z)) for any
constant c > 0 (rather than a random variable qi ). Recall that for any constant c > 0 and any Z ∼ N (0, σ 2 )
with finite variance, the random variable ρ0 (proxcρ (Z)) is sub-Gaussian. Conditional on β̂[−i] , one has
Xi> β̂[−i] ∼ N 0, kβ̂[−i] k2 . This yields
!#
"
C32 Kn2
0
>
P ρ proxcρ (Xi β̂[−i] ) ≥ C1 Kn ≤ C2 E exp −
kβ̂[−i] k2
≤ C2 exp −C3 Kn2 + C4 exp (−C5 n)
(168)
for some constants C1 , C2 , C3 , C4 , C5 > 0 sincekβ̂[−i] k is bounded with high probability (see Theorem 4).
∂prox
(z)
bρ
≤ 0 by [18, Proposition 6.3]. Hence, in order to move over from the above concenNote that
∂b
tration result established for a fixed constant c to the random variables qi , it suffices to establish a uniform
lower bound for qi with high probability. Observe that for each i,
qi ≥
kXi k2
1
n
G[−i]
≥ C∗
with probability 1 − exp(−Ω(n)), where C ∗ is some universal constant. On this event, one has
ρ0 proxqi ρ Xi> β̂[−i]
≤ ρ0 proxC ∗ ρ Xi> β̂[−i] .
This taken collectively with (168) yields
P ρ0 (proxqi ρ (Xi> β̂[−i] )) ≤ C1 Kn ≥ P ρ0 (proxC ∗ ρ (Xi> β̂[−i] )) ≤ C1 Kn
≥ 1 − C2 exp −C3 Kn2 − C4 exp (−C5 n) .
(169)
(170)
This controls the last term.
To summarize, if {Kn } and {Hn } are diverging sequences satisfying the assumptions in (130), combining
(159) and the bounds for each term in the right-hand side finally gives (153). On the other hand, combining
(165) and (170) yields (154).
48
With the help of Lemma 18 we are ready to prove Lemma 17. Indeed, observe that
Xj> (β̂[−i] − β̂) ≤ Xj> (b̂ − β̂) + Xj> (β̂[−i] − b̂) ,
and hence by combining Lemma 2 and Lemma 18, we establish the first claim (147). The second claim (148)
follows directly from Lemmas 2, 18 and (157).
I
Proof of Theorem 7(b)
This section proves that the random sequence α̃ = Tr G̃−1 /n converges in probability to the constant b∗
defined by the system of equations (23) and (24). To begin with, we claim that α̃ is close to a set of auxiliary
random variables {q̃i } defined below.
Lemma 19. Define q̃i to be
q̃i =
1 > −1
X̃ G̃ X̃i ,
n i [−i]
where G̃[−i] is defined in (83). Then there exist universal constants C1 , C2 , C3 , C4 , c2 , c3 > 0 such that
Kn2 Hn
P sup |q̃i − α̃| ≤ C1 √
n
i
≥ 1 − C2 n2 exp c2 Hn2 − C3 n exp −c3 Kn2 − exp (−C4 n (1 + o(1))) = 1 − o(1),
where Kn , Hn are as in (130).
Proof: This result follows directly from Proposition 1 and equation (142).
A consequence is that proxq̃i ρ Xi> β̂[−i]
becomes close to proxα̃ρ Xi> β̂[−i] .
Lemma 20. Let q̃i and α̃ be as defined earlier. Then one has
Kn3 Hn
P sup proxq̃i ρ Xi> β̂[−i] − proxα̃ρ Xi> β̂[−i] ≤ C1 √
n
i
2
2
≥ 1 − C2 n exp −c2 Hn − C3 n exp −c3 Kn2 − exp (−C4 n (1 + o(1))) = 1 − o(1), (171)
where Kn , Hn are as in (130).
The key idea behind studying proxα̃ρ Xi> β̂[−i] is that it is connected to a random function δn (·) defined
below, which happens to be closely related to the equation (24). In fact,
we will show that δn (α̃) converges
in probability to 0; the proof relies on the connection between proxα̃ρ Xi> β̂[−i] and the auxiliary quantity
proxq̃i ρ Xi> β̂[−i] . The formal results is this:
Proposition 2. For any index i, let β̂[−i] be the MLE obtained on dropping the ith observation. Define
δn (x) to be the random function
n
δn (x) :=
p
1X
1
.
−1+
n
n i=1 1 + xρ00 prox
>
xρ Xi β̂[−i]
P
Then one has δn (α̃) → 0.
49
(172)
Furthermore, the random function δn (x) converges to a deterministic function ∆(x) defined by
1
∆(x) = κ − 1 + EZ
,
1 + xρ00 (proxxρ (τ∗ Z))
(173)
where Z ∼ N (0, 1), and τ∗ is such that (τ∗ , b∗ ) is the unique solution to (23) and (24).
P
Proposition 3. With ∆(x) as in (173), ∆(α̃) → 0.
In fact, one can easily verify that
∆(x) = κ − E Ψ0 (τ∗ Z; x) ,
(174)
and hence by Lemma 5, the solution to ∆(x) = 0 is exactly b∗ . As a result, putting the above claims together,
we show that α̃ converges in probability to b∗ .
It remains to formally prove the preceding lemmas and propositions, which is the goal of the rest of this
section.
Proof of Lemma 20: By [18, Proposition 6.3], one has
∂proxbρ (z)
ρ0 (x)
=−
∂b
1 + bρ00 (x)
,
x=proxbρ (z)
which yields
sup proxq̃i ρ Xi> β̂[−i] − proxα̃ρ Xi> β̂[−i]
i
0
ρ
(x)
= sup
· |q̃i − α̃|
1 + qα̃,i ρ00 (x) x=prox
> β̂
i
X
[−i]
qα̃,i ρ
i
≤ sup ρ0 proxqα̃,i (Xi> β̂[−i] ) · sup |q̃i − α̃|,
i
(175)
i
where qα̃,i is between q̃i and α̃. Here, the last inequality holds since qα̃,i , ρ00 ≥ 0.
In addition, just as in the proof of Lemma 18, one can show that qi is bounded below by some constant
√
C ∗ > 0 with probability 1 − exp(−Ω(n)). Since qα̃,i ≥ min{q̃i , α̃}, on the event supi |q̃i − α̃| ≤ C1 Kn2 Hn / n,
which happens with high probability (Lemma 19), qα̃,i ≥ Cα for some universal constant Cα > 0. Hence, by
an argument similar to that establishing (170), we have
P sup ρ0 proxqα̃,i Xi> β̂[−i]
≥ C1 Kn
i
≤ C2 n2 exp −c2 Hn2 + C3 n exp −c3 Kn2 + exp (−C4 n (1 + o(1))) .
This together with (175) and Lemma 19 concludes the proof.
Proof of Proposition 2: To begin with, recall from (136) and (137) that on An ,
n
ρ00 (X̃i> β̃)
X̃i> G̃−1
(i) X̃i
n
p−1 X
=
n
i=1 1 +
ρ00 (X̃i> β̃)
n
n
X̃i> G̃−1
(i) X̃i
=1−
50
1X
n i=1 1 +
1
ρ00 (X̃i> β̃)
X̃i> G̃−1
(i) X̃i
n
.
(176)
Using the fact that
1
1+x
−
1
1+y
≤ |x − y| for x, y ≥ 0, we obtain
n
1X
n i=1 1 +
n
1
ρ00 (X̃i> β̃)
X̃i> G̃−1
(i) X̃i
n
−
1X
1
00
n i=1 1 + ρ (X̃i> β̃)α̃
n
≤
1 X 00 >
1
1 > −1
00
ρ (X̃i β̃) X̃i> G̃−1
(i) X̃i − α̃ ≤ kρ k∞ sup n X̃i G̃(i) X̃i − α̃
n i=1
n
i
Kn2 Hn
= kρ00 k∞ sup |ηi | ≤ C1 √
,
n
i
with high probability (Proposition 1). This combined with (176) yields
!
n
p−1
Kn2 Hn
1X
1
≥ C1 √
P
−1+
n
n i=1 1 + ρ00 (X̃i> β̃)α̃
n
≤ C2 n2 exp −c2 Hn2 + C3 n exp −c3 Kn2 + exp (−C4 n (1 + o(1))) .
Pn
Pn
1
1
1
The above bound concerns n1 i=1 1+ρ00 (X̃
> β̃)α̃ , and it remains to relate it to n
i=1 1+ρ00 (prox (X̃ > β̃ ))α̃ .
α̃ρ
i
i
To this end, we first get from the uniform boundedness of ρ000 and Lemma 17 that
Kn2 Hn
P sup ρ00 (X̃i> β̃) − ρ00 proxq̃i ρ (X̃i> β̃[−i] ) ≥ C1 √
n
i
≤ C2 n exp(−c2 Hn2 ) + C3 exp(−c3 Kn2 ) + exp(−C4 n(1 + o(1))). (177)
Note that
n
n
1
1
1X
1X
−
>
00
00
n i=1 1 + ρ (X̃i β̃)α̃ n i=1 1 + ρ (proxα̃ρ (X̃i> β̃[−i] ))α̃
≤ |α̃| sup ρ00 (X̃i> β̃) − ρ00 proxα̃ρ (X̃i> β̃[−i] )
i
n
o
≤ |α̃| sup ρ00 X̃i> β̃ − ρ00 proxq̃i ρ (X̃i> β̃[−i] ) + ρ00 proxq̃i ρ (X̃i> β̃[−i] ) − ρ00 proxα̃ρ (X̃i> β̃[−i] )
.
i
By the bound (177), an application of Lemma 20, and the fact that α̃ ≤ p/(nλlb ) (on An ), we obtain
!
n
1X
1
Kn3 Hn
p
−1+
P
≥ C1 √
n
n i=1 1 + ρ00 proxα̃ρ (Xi> β̂[−i] ) α̃
n
≤ C2 n2 exp − c2 Hn2 + C3 n exp − c3 Kn2 + exp − C4 n(1 + o(1)) .
P
This establishes that δn (α̃) → 0.
Proof of Proposition 3: Note that since 0 < α ≤ p/(nλlb ) := B on An , it suffices to show that
P
sup |δn (x) − ∆(x)| → 0.
x∈[0,B]
We do this by following three steps. Below, M > 0 is some sufficiently large constant.
1. First we truncate the random function δn (x) and define
δ̃n (x) =
n
X
p
−1+
n
00 prox
i=1 1 + xρ
51
1
xρ
Xi> β̂[−i] 1{kβ̂[−i] k≤M }
.
P
The first step is to show that supx∈[0,B] δ̃n (x) − δn (x) → 0. We stress that this truncation does not
arise in [20], and we keep track of the truncation throughout the rest of the proof.
P
2. Show that supx∈[0,B] δ̃n (x) − E δ̃n (x) → 0.
P
3. Show that supx∈[0,B] E δ̃n (x) − ∆(x) → 0.
1
1
− 1+z
| ≤ |y−z| for any y, z > 0 and that
To argue about the first step, observe that | 1+y
Proposition 6.3]. Then
∂proxcρ (x)
∂x
≤ 1 [18,
|δn (x) − δ̃n (x)| ≤ |x| · kρ000 k∞ · sup Xi> β̂[−i] − Xi> β̂[−i] 1{kβ̂[−i] k≤M } .
i
For a sufficiently large constant M > 0, P(kβ̂[−i] k ≥ M ) ≤ exp(−Ω(n)) by Theorem 4. Hence, for any > 0,
!
>
P
sup |δn (x) − δ̃n (x)| ≥ ≤ P sup Xi β̂[−i] 1{kβ̂[−i] k≥M } ≥
Bkρ000 k∞
i
x∈[0,B]
≤
n
X
P kβ̂[−i] k ≥ M = o(1),
(178)
i=1
establishing Step 1.
To argue about Step 2, note that for any x and z,
n
δ̃n (x) − δ̃n (z) ≤
1
1X
1
−
n i=1 1 + xρ00 prox
>
1 + zρ00 proxzρ Xi> β̂[−i] 1{kβ̂[−i] k≤M }
xρ Xi β̂[−i] 1{kβ̂[−i] k≤M }
n
1X
xρ00 proxxρ Xi> β̂[−i] 1{kβ̂[−i] k≤M } − zρ00 proxzρ Xi> β̂[−i] 1{kβ̂[−i] k≤M }
n i=1
n
0
ρ
(x)
1 X 00
kρ k∞ |x − z| + |z| · kρ000 k∞
≤
|x − z|
n i=1
1 + z̃ρ00 (x) x=prox X > β̂[−i] 1
z̃ρ
i
{kβ̂[−i] k≤M }
!
n
1X 0
>
00
000
ρ proxz̃ρ Xi β̂[−i] 1{kβ̂[−i] k≤M }
≤ |x − z| kρ k∞ + |z| · kρ k∞
,
n i=1
≤
where z̃ ∈ (x, z). Setting
Yn := kρ00 k∞ + Bkρ000 k∞
1 Xn
ρ0 proxz̃ρ Xi> β̂[−i] 1{kβ̂[−i] k≤M }
,
i=1
n
then for any , η > 0 we have
!
P
sup
x,z∈(0,B],|x−z|≤η
|δ̃n (x) − δ̃n (z)| ≥
η
≤
E[Yn ] ≤ ηC1 (),
≤ P Yn ≥
η
(179)
where C1 () is some function independent of n. The inequality (179) is an analogue of [20, Lemma 3.24].
We remark that the truncation is particularly important here in guaranteeing that E[Yn ] < ∞.
Set
Gn (x) := E δ̃n (x) ,
52
and observe that
i
h
.
Gn (x) − Gn (z) ≤ |x − z| kρ00 k∞ + |z|kρ000 k∞ E ρ0 proxz̃ρ Xi> β̂[−i] 1{kβ̂[−i] k≤M }
A similarly inequality applies to ∆(x) in which Xi> β̂[−i] 1{kβ̂[−i] k≤M } is replaced by τ∗ Z. In either case,
|Gn (x) − Gn (z)| ≤ C2 η
sup
|∆(x) − ∆(z)| ≤ C3 η
sup
and
x,z∈(0,B],|x−z|≤η
(180)
x,z∈(0,B],|x−z|≤η
for any η > 0.
For any 0 > 0, set K = max{C1 (0 ), C2 }. Next, divide [0, B] into finitely many segments
[0, x1 ), [x1 , x2 ), . . . , [xK−1 , xK := B]
such that the length of each segment is η/K for any η > 0. Then for every x ∈ [0, B], there exists l such
that |x − xl | ≤ η/K. As a result, for any x ∈ [0, B],
sup
sup
δ̃n (x) − δ̃n (xl ) + sup δ̃n (xl ) − Gn (xl ) .
x,xl ∈(0,B],|x−xl |≤η/K
1≤l≤k
δ̃n (x) − Gn (x) ≤ η +
x∈(0,B]
Now fix δ > 0, > 0. Applying the above inequality gives
!
P
sup |δ̃n (x) − Gn (x)| ≥ δ
x∈(0,B]
!
δ−η
≤P
sup
|δ̃n (x) − δ̃n (xl )| ≥
2
x,xl ∈(0,B],|x−xl |≤η/K
δ−η
+ P sup |δ̃n (xl ) − Gn (xl )| ≥
2
1≤l≤k
η
δ−η
δ−η
≤ C1
+ P sup |δ̃n (xl ) − Gn (xl )| ≥
.
K
2
2
1≤l≤k
(181)
Choose η < min{/2, δ}, K = max{C1 ( δ−η
2 ), C2 }. Then the first term in the right-hand side is at most /2.
Furthermore, suppose one can establish that for any fixed x,
P
|δ̃n (x) − Gn (x)| → 0.
(182)
Since the second term in the right-hand side is a supremum over finitely many points, there exists an integer
N such that for all n ≥ N , the second term is less than or equal to /2. Hence, for all n ≥ N , right-hand
side is at most , which proves Step 2. Thus, it remains to prove (182) for any fixed x ≥ 0. We do this after
the analysis of Step 3.
We argue about Step 3 in a similar fashion and letting K = max{C2 , C3 }, divide (0, B] into segments of
length η/K. For every x ∈ (0, B] there exists xl such that |x − xl | ≤ η/K. Then
|Gn (x) − ∆(x)| ≤ |Gn (x) − Gn (xl )| + |Gn (xl ) − ∆(xl )| + |∆(xl ) − ∆(x)|
≤ 2η + |Gn (xl ) − ∆(xl )|,
=⇒
sup |Gn (x) − ∆(x)| ≤ 2η + sup |Gn (xl ) − ∆(xl )|.
1≤l≤k
[0,B]
Hence, it suffices to show that for any fixed x, |Gn (x) − ∆(x)| → 0. To this end, observe that for any fixed x,
|Gn (x) − ∆(x)|
p
≤
− κ + EX
n
1 + xρ00 prox
1
xρ
X1> β̂[−1] 1{kβ̂[−1] k≤M }
53
− EZ
"
1 + xρ00
1
proxxρ (τ∗ Z)
#
.
Additionally,
X1> β̂[−1] 1{kβ̂[−1] k≤M } = kβ̂[−1] k1{kβ̂[−1] k≤M } Z̃,
P
where Z̃ ∼ N (0, 1). Since kβ̂[−1] k1{kβ̂[−1] k≤M } → τ∗ , by Slutsky’s theorem, X1> β̂[−1] 1{kβ̂[−1] k≤M } converges
weakly to τ∗ Z̃. Since t 7→ 1/(1 + xρ”(proxxρ (t))) is bounded, one directly gets
P
Gn (x) − ∆(x) → 0
for every x.
Finally, we establish (182). To this end, we will prove instead that for any given x,
L
δn (x) − Gn (x) →2 0.
Define
Mi :=
Xi> β̂[−i] 1{kβ̂[−i] k≤M }
Then δn (x) − Gn (x) =
1
n
Pn
i=1
and
f (Mi ) :=
1 + xρ00
"
#
1
1
−E
.
proxxρ (Mi )
1 + xρ00 proxxρ (Mi )
f (Mi ). Hence, for any x ∈ [0, B],
n
1 X 2
1 X
E f (Mi ) + 2
E f (Mi )f (Mj )
2
n i=1
n
i6=j
2
E f (M1 )
n(n − 1)
=
+
E f (M1 )f (M2 ) .
n
n2
Var[δn (x)] =
The first term in the right-hand side is at most 1/n. Hence, it suffices to show that E f (M1 )f (M2 ) → 0.
Let β̂[−12] be the MLE when the 1st and 2nd observations are dropped. Define
1 X 00 >
1
ρ Xj β̂[−12] Xj Xj> ,
G[−12] :=
q2 := X2> G−1
[−12] X2 ,
n
n
j6=1,2
1
0
>
b̂[−1] := β̂[−12] + G−1
X
−ρ
prox
X
β̂
.
2
[−12]
q2 ρ
2
n [−12]
By an application of Lemma 18,
K 2 Hn
≤ C2 n exp −c2 Hn2 − C3 exp −c3 Kn2 − exp (−C4 n(1 + o(1))) . (183)
P β̂[−1] − b̂[−1] ≥ C1 n
n
Also, by the triangle inequality,
X1> (β̂[−1] − β̂[−12] ) ≤ X1> (β̂[−1] − b̂[−1] ) +
1
0
>
X1> G−1
−ρ
prox
X
β̂
.
X
q2 ρ
2 [−12]
[−12] 2
n
Invoking Lemma 2, (183), and an argument similar to that leading to (170) and (164), we obtain
Kn2 Hn
>
P X1 β̂[−1] − β̂[−12] ≥ C1 √
n
≤ C2 n exp −c2 Hn2 − C3 exp −c3 Kn2 − exp (−C4 n (1 + o(1))) .
The event {kβ̂[−1] k ≤ M } ∩ {kβ̂[−12] k ≤ M } occurs with probability at least 1 − C exp(−cn). Hence, one
obtains
Kn2 Hn
>
P X1 β̂[−1] 1kβ̂[−1] k≤M − β̂[−12] 1kβ̂[−12] k≤M ≤ C1 √
n
≥ C2 n exp −c2 Hn2 − C3 exp −c3 Kn2 − exp (−C4 n (1 + o(1))) . (184)
54
A similar statement continues to hold with X1 replaced by X2 and β̂[−1] replaced by β̂[−2] . Some simple
computation yields that kf 0 k∞ is bounded by some constant times |x|. By the mean value theorem and the
fact that kf k∞ ≤ 1,
f (M1 )f (M2 ) − f X1> β̂[−12] 1{kβ̂[−12] k≤M } f X2> β̂[−12] 1{kβ̂[−12] k≤M }
n
≤ kf k∞ f (M1 ) − f X1> β̂[−12] 1{kβ̂[−12] k≤M }
o
+ f (M2 ) − f X2> β̂[−12] 1{kβ̂[−12] k≤M }
≤ C|x| · X1> β̂[−1] 1{kβ̂[−1] k≤M } − X1> β̂[−12] 1{kβ̂[−12] k≤M }
+ |x| · X2> β̂[−2] 1{kβ̂[−2] k≤M } − X2> β̂[−12] 1{kβ̂[−12] k≤M } .
Consequently,
P
f (M1 )f (M2 ) − f X1> β̂[−12] 1{kβ̂[−12] k≤M } f X2> β̂[−12] 1{kβ̂[−12] k≤M } → 0.
As kf k∞ ≤ 1, this implies convergence in L1 . Thus, it simply suffices to show that
h
i
E f X1> β̂[−12] 1{kβ̂[−12] k≤M } f X2> β̂[−12] 1{kβ̂[−12] k≤M }
→ 0.
Denote the design matrix on dropping the first and second row as X[−12] . Note that conditional on X[−12] ,
X1> β̂[−12] 1{kβ̂[−12] k≤M } and X2> β̂[−12] 1{ kβ̂[−12] k ≤ M } are independent and have distribution
N 0, kβ̂[−12] k2 1{kβ̂[−12] k≤M } .
Using this and by arguments similar to [20, Lemma 3.23], one can show that
"
E e
i tX1> β̂[−12] 1{kβ̂
[−12] k≤M }
+wX2> β̂[−12] 1{kβ̂
#
[−12] k≤M }
itX1> β̂[−12] 1{kβ̂
−E e
[−12] k≤M }
iwX2> β̂[−12] 1{kβ̂
k≤M }
[−12]
E e
→ 0. (185)
On repeated application of the multivariate inversion theorem for obtaining densities from characteristic
functions, we get that
h
i
E f X1> β̂[−12] 1{kβ̂[−12] k≤M } f X2> β̂[−12] 1{kβ̂[−12] k≤M }
h
i h
i
− E f X1> β̂[−12] 1{kβ̂[−12] k≤M } E f X2> β̂[−12] 1{kβ̂[−12] k≤M }
→ 0.
Since f is centered, this completes the proof.
References
[1] Alan Agresti and Maria Kateri. Categorical data analysis. Springer, 2011.
[2] Noga Alon and Joel H Spencer. The probabilistic method (3rd edition). John Wiley & Sons, 2008.
[3] Dennis Amelunxen, Martin Lotz, Michael B McCoy, and Joel A Tropp. Living on the edge: Phase
transitions in convex programs with random data. Information and Inference, page iau005, 2014.
55
[4] Árpád Baricz. Mills’ ratio: monotonicity patterns and functional inequalities. Journal of Mathematical
Analysis and Applications, 340(2):1362–1370, 2008.
[5] Maurice S Bartlett. Properties of sufficiency and statistical tests. Proceedings of the Royal Society of
London. Series A, Mathematical and Physical Sciences, pages 268–282, 1937.
[6] Mohsen Bayati and Andrea Montanari. The dynamics of message passing on dense graphs, with applications to compressed sensing. IEEE Transactions on Information Theory, 57(2):764–785, 2011.
[7] Mohsen Bayati and Andrea Montanari. The LASSO risk for Gaussian matrices. IEEE Transactions on
Information Theory, 58(4):1997–2017, 2012.
[8] Peter J Bickel and JK Ghosh. A decomposition for the likelihood ratio statistic and the bartlett
correction–a bayesian argument. The Annals of Statistics, pages 1070–1090, 1990.
[9] Stéphane Boucheron and Pascal Massart. A high-dimensional Wilks phenomenon. Probability theory
and related fields, 150(3-4):405–433, 2011.
[10] George Box. A general distribution theory for a class of likelihood criteria. Biometrika, 36(3/4):317–346,
1949.
[11] Emmanuel Candès, Yingying Fan, Lucas Janson, and Jinchi Lv. Panning for gold: Model-free knockoffs
for high-dimensional controlled variable selection. arXiv preprint arXiv:1610.02351, 2016.
[12] Herman Chernoff. On the distribution of the likelihood ratio. The Annals of Mathematical Statistics,
pages 573–578, 1954.
[13] Gauss M Cordeiro. Improved likelihood ratio statistics for generalized linear models. Journal of the
Royal Statistical Society. Series B (Methodological), pages 404–413, 1983.
[14] Gauss M Cordeiro and Francisco Cribari-Neto. An introduction to Bartlett correction and bias reduction.
Springer, 2014.
[15] Gauss M Cordeiro, Franciso Cribari-Neto, Elisete CQ Aubin, and Silvia LP Ferrari. Bartlett corrections
for one-parameter exponential family models. Journal of Statistical Computation and Simulation, 53(34):211–231, 1995.
[16] Thomas M Cover and Joy A Thomas. Elements of information theory. John Wiley & Sons, 2012.
[17] Francisco Cribari-Neto and Gauss M Cordeiro. On bartlett and bartlett-type corrections francisco
cribari-neto. Econometric reviews, 15(4):339–367, 1996.
[18] David Donoho and Andrea Montanari. High dimensional robust M-estimation: Asymptotic variance
via approximate message passing. Probability Theory and Related Fields, pages 1–35, 2013.
[19] Noureddine El Karoui. Asymptotic behavior of unregularized and ridge-regularized high-dimensional
robust regression estimators: rigorous results. arXiv preprint arXiv:1311.2445, 2013.
[20] Noureddine El Karoui. On the impact of predictor geometry on the performance on high-dimensional
ridge-regularized generalized robust regression estimators. Probability Theory and Related Fields, pages
1–81, 2017.
[21] Noureddine El Karoui, Derek Bean, Peter J Bickel, Chinghway Lim, and Bin Yu. On robust regression
with high-dimensional predictors. Proceedings of the National Academy of Sciences, 110(36):14557–
14562, 2013.
[22] Jianqing Fan and Jiancheng Jiang. Nonparametric inference with generalized likelihood ratio tests.
Test, 16(3):409–444, 2007.
56
[23] Jianqing Fan and Jinchi Lv. Nonconcave penalized likelihood with NP-dimensionality. IEEE Transactions on Information Theory, 57(8):5467–5484, 2011.
[24] Jianqing Fan, Chunming Zhang, and Jian Zhang. Generalized likelihood ratio statistics and Wilks
phenomenon. Annals of statistics, pages 153–193, 2001.
[25] Yingying Fan, Emre Demirkaya, and Jinchi Lv. Nonuniformity of p-values can occur early in diverging
dimensions. https://arxiv.org/abs/1705.03604, May 2017.
[26] William W Hager. Updating the inverse of a matrix. SIAM review, 31(2):221–239, 1989.
[27] David L. Hanson and Farroll T. Wright. A bound on tail probabilities for quadratic forms in independent
random variables. The Annals of Mathematical Statistics, 42(3):1079–1083, 1971.
[28] Xuming He and Qi-Man Shao. On parameters of increasing dimensions. Journal of Multivariate Analysis,
73(1):120–135, 2000.
[29] Daniel Hsu, Sham Kakade, and Tong Zhang. A tail inequality for quadratic forms of subgaussian random
vectors. Electron. Commun. Probab, 17(52):1–6, 2012.
[30] Peter J Huber. Robust regression: asymptotics, conjectures and Monte Carlo. The Annals of Statistics,
pages 799–821, 1973.
[31] Peter J Huber. Robust statistics. Springer, 2011.
[32] Adel Javanmard and Andrea Montanari. State evolution for general approximate message passing
algorithms, with applications to spatial coupling. Information and Inference, page iat004, 2013.
[33] DN Lawley. A general method for approximating to the distribution of likelihood ratio criteria.
Biometrika, 43(3/4):295–303, 1956.
[34] Erich L Lehmann and Joseph P Romano. Testing statistical hypotheses. Springer Science & Business
Media, 2006.
[35] Hua Liang, Pang Du, et al. Maximum likelihood estimation in logistic regression models with a diverging
number of covariates. Electronic Journal of Statistics, 6:1838–1846, 2012.
[36] Enno Mammen. Asymptotics with increasing dimension for robust regression with applications to the
bootstrap. The Annals of Statistics, pages 382–400, 1989.
[37] Peter McCullagh and James A Nelder. Generalized linear models. Monograph on Statistics and Applied
Probability, 1989.
[38] Lawrence H Moulton, Lisa A Weissfeld, and Roy T St Laurent. Bartlett correction factors in logistic
regression models. Computational statistics & data analysis, 15(1):1–11, 1993.
[39] Neal Parikh and Stephen Boyd.
1(3):127–239, 2014.
Proximal algorithms.
Foundations and Trends in Optimization,
[40] Stephen Portnoy. Asymptotic behavior of M-estimators of p regression parameters when p2 /n is large.
i. consistency. The Annals of Statistics, pages 1298–1309, 1984.
[41] Stephen Portnoy. Asymptotic behavior of M-estimators of p regression parameters when p2 /n is large;
ii. normal approximation. The Annals of Statistics, pages 1403–1417, 1985.
[42] Stephen Portnoy. Asymptotic behavior of the empiric distribution of m-estimated residuals from a
regression model with many parameters. The Annals of Statistics, pages 1152–1170, 1986.
57
[43] Stephen Portnoy et al. Asymptotic behavior of likelihood methods for exponential families when the
number of parameters tends to infinity. The Annals of Statistics, 16(1):356–366, 1988.
[44] Mark Rudelson, Roman Vershynin, et al. Hanson-Wright inequality and sub-gaussian concentration.
Electron. Commun. Probab, 18(82):1–9, 2013.
[45] Michael R Sampford. Some inequalities on mill’s ratio and related functions. The Annals of Mathematical
Statistics, 24(1):130–132, 1953.
[46] Vladimir Spokoiny. Penalized maximum likelihood estimation and effective dimension. arXiv preprint
arXiv:1205.0498, 2012.
[47] Pragya Sur, Yuxin Chen, and Emmanuel Candès. Supplemental materials for “the likelihood ratio
test in high-dimensional logistic regression is asymptotically a rescaled chi-square”. http: // statweb.
stanford. edu/ ~candes/ papers/ supplement_ LRT. pdf , 2017.
[48] Cheng Yong Tang and Chenlei Leng. Penalized high-dimensional empirical likelihood. Biometrika, pages
905–919, 2010.
[49] Terence Tao. Topics in random matrix theory, volume 132. American Mathematical Society Providence,
RI, 2012.
[50] A. W. Van der Vaart. Asymptotic statistics, volume 3. Cambridge university press, 2000.
[51] Roman Vershynin. Introduction to the non-asymptotic analysis of random matrices. Compressed Sensing: Theory and Applications, pages 210 – 268, 2012.
[52] Samuel S Wilks. The large-sample distribution of the likelihood ratio for testing composite hypotheses.
The Annals of Mathematical Statistics, 9(1):60–62, 1938.
[53] Ting Yan, Yuanzhuang Li, Jinfeng Xu, Yaning Yang, and Ji Zhu. High-dimensional Wilks phenomena
in some exponential random graph models. arXiv preprint arXiv:1201.0058, 2012.
58
| 7cs.IT
|
Design of an intermediate representation for query
languages
Internship report
Romain Vernoux, supervised by Kim Nguyễn (VALS team, LRI)
arXiv:1607.04197v1 [cs.PL] 14 Jul 2016
July 15, 2016
Contents
1 Introduction
1
2 Notations and conventions
3
3 Architecture
3
4 QIR calculus
4.1 QIR data model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 QIR operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3 QIR expressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5
5
5
6
5 QIR evaluation
5.1 QIR reduction rules . . . . . . . . .
5.2 A measure for “good” output plans .
5.3 An exhaustive reduction strategy . .
5.4 A heuristic-based reduction strategy
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
7
8
8
10
11
6 Experiments
18
6.1 Example of code factorization: analytics queries . . . . . . . . . . . . . . . . . . . 18
6.2 Example of fragment grouping: dynamic queries . . . . . . . . . . . . . . . . . . . 20
6.3 Example of incompatibility: caching . . . . . . . . . . . . . . . . . . . . . . . . . 21
7 Related work
21
8 Conclusion and future work
23
Appendices
24
A References
24
B Internship timeline
25
C Formal definition of occurences
25
D Proofs
26
E Estimation of the heuristic fuel value
30
F Source code and other resources
30
1
Introduction
This internship is part of a joint collaboration between Laurent Daynes at Oracle Labs, Giuseppe
Castagna and Julien Lopez (also a MPRI student) at PPS and Kim Nguyễn and myself at LRI.
Each of these three groups worked on different aspects of the project and synchronized during
regular meetings. In this report, I will do my best to focus on my contribution and explain only
the necessary of the other parts of the project. I will use “I” to emphasize my work and “we” in
formal definitions and technical sections.
The general context
Data oriented applications, usually written in a high-level, general-purpose programming language
(such as Java, Ruby, or JavaScript) interact with databases through a very coarse interface.
Informally, the text of a query is built on the application side (either via plain string concatenation
or through an abstract notion of statement) and shipped to the database. After its evaluation, the
results are then serialized and sent back to the “application-code” where they are translated in the
application language datatypes. Such roundtrips represent a significant part of the application
running time. Moreover, this programming model prevents one from using richer query constructs,
such as user-defined functions (UDF). UDFs are functions defined by the application developer,
most of the time in a language more expressive than the query language, and used in a query (e.g.,
in a filter condition of a SQL query). While some databases also possess a “server-side” language
for this purpose (e.g., PL/SQL in Oracle Database and PL/pgSQL in PostgreSQL), its integration
with the very-optimized query execution engine is still minimal, resulting in notoriously slow
query evaluations. The alternative is to evaluate UDFs in the application runtime, leading to
additional application-database roundtrips and even poorer performance[8][7]. These problems,
often refered to as language-integrated query issues, are in no way specific to relational databases
and also affect the so-called NoSQL databases, which now all provide some form of declarative
query language.
In this setting, Oracle Labs is developing Truffle[14], a framework for representing dynamic
languages programs and functions as abstract syntax tree (AST) nodes with particular evaluation
rules, and Graal[14], a just-in-time (JIT) compiler written in Java, leveraging Oracle’s experience
from the Java JIT compiler and achieving high performance in evaluating Truffle nodes. A
proof of concept, in the form of a full-blown JavaScript runtime prototype written entirely in
Truffle and open-source projects for Ruby, R and Python runtimes[14], is already available.
In principle, this framework could enable to efficiently evaluate UDFs written in high-level
dynamic languages directly inside a database embedding a Java Virtual Machine (such as Oracle
DB[15], Cassandra[11] or Hive[16]), thus bridging the aforementioned impedence mismatch gap,
considerably decreasing the cost of roundtrips between query and UDFs evaluation, and providing
the right level of abstraction for a deep integration in the database query optimization process.
The research problem
Language-integrated queries and impedence mismatch were already research topics in the 1980s[1],
although they have received increased attention with the popularization of web applications.
Various solutions have already been proposed in the past[4] and constant demand for this matter
has for instance lead Microsoft to include LINQ[12]–a framework integrating Microsoft SQL
Server querying primitives in C#, among other languages–in .NET 3.5, and to continually add
new features in the subsequent releases.
In our opinion, the imperfections of the existing solutions and the new perspectives opened
by the Truffle/Graal project justified a reinvestigation of this (still trending) topic. Furthermore,
taking in consideration the wide adoption of NoSQL databases, special care should be taken to
propose a solution that is not specific to relational databases, and to the extent of our knowledge
1
such work has not been done yet. Thus, the problematic of the QIR project is articulated around
the following questions:
1. How to draw the line between application code that should be evaluated in the application
runtime, in the database native query processor and in the application language runtime
embedded in the database?
2. How to represent queries in a way that is agnostic of the application language (for reusability)
and the target database (to support relational and non-relational ones)?
3. How to rewrite this query representation to present an efficient query to the database?
4. How to translate this query representation to the database native language?
5. How to evaluate application (Truffle) code inside the database?
Although I took part in discussions revolving around all these questions (cf. Section 3), my
internship was centered on questions 2 and 3. Questions 1 and 4 were investigated at PPS and
question 5 at Oracle Labs.
My contribution
A review of existing language-integrated query frameworks (cf. Appendix F) and results from
a previous internship[13] highlighted that existing database query languages (including SQL)
share high-level querying primitives (e.g., filtering, joins, aggregation) that can be represented by
operators, but differ widely regarding the semantics of their expression language.
In order to represent queries in an application language- and database-agnostic manner
(question 2 above), I designed a small calculus, dubbed “QIR” for Query Intermediate Representation, expressive enough to capture querying capabilities offered by mainstream languages.
QIR contains expressions, corresponding to a small extension of the pure lambda-calculus, and
operators to represent usual querying primitives (Section 4).
In the effort to send efficient queries to the database (question 3 above), I abstracted the
idea of “good” query representations in a measure on QIR terms. Then, I designed an evaluation
strategy rewriting QIR query representations into “better” ones (Section 5).
Arguments supporting its validity
As an abstraction layer between application languages and databases, QIR guarantees the
robustness of the entire framework regarding changes in these two tiers. But to provide formal
evidence of the relevance of my solutions, I wrote an optimality proof for my evaluation strategy
on a particular subset of QIR terms, with respect to the aforementioned measure (Theorem 6).
Additionally, I implemented a prototype of the QIR evaluation strategy and ran experiments
showing that (i) the measure captures well the idea of “good” queries, (ii) rewritings have a low
cost while enabling considerable speedup opportunities for the database and (iii) the evaluation
strategy outputs optimal results well outside the scope of the QIR subset captured by the
optimality proof (Section 6).
Summary and future work
The ability to evaluate snippets from the application code inside the database opened new
perspectives on the language-intergrated query problems. My approach allows a complete
separation between application language and target database considerations, and to this extent
is an interesting contribution even outside the scope of our project. Future work includes testing
this concept with more combinations of application languages and databases, and observing the
impact of such an architecture on real-world web applications.
2
2
Notations and conventions
Trees In this document, we will use an infix parenthesis notation for trees. The tree leaves are
represented by their name and a tree rooted at a node N with n children S1 , . . . , Sn is denoted
by N (r1 , . . . , rn ) where ri is the representation of the subtree rooted at node Si . For instance, if
a tree T is composed of a node N which has two children P and Q such that Q has one child R,
we will write T as N (P, Q(R)).
Contexts Contexts are trees in which some subtrees have been removed and replaced by “holes”.
A hole is denoted by []. For instance, a context C can be obtained by removing the subtree
rooted at Q in T and will be denoted by N (P, []). Context holes can be filled with arbitrary
subtrees. If a context C contains one hole, C[S] corresponds to the tree C where the hole has
been replaced by the tree S. For instance, if C = N (P, []) then T = C[Q(R)]. This definition
generalizes to contexts with n holes using the C[S1 , . . . , Sn ] notation. Notice that in this case,
the context holes are ordered using a prefix (depth-first, left-to-right) traversal of the tree, which
corresponds to the left-to-right order while reading the tree notation we use in this document.
For instance, if C = N ([], Q([])) then T = C[P, R].
List comprehensions Similarly to sets and the {x | P } notation for set comprehensions, we
use the [x | P ] notation for list comprehensions. List comprehensions are order-preserving, that
is, a list [f (x) | x ∈ L] respects the order of L if L is a list. In this document, we will use them
to build the list of a node’s children in a tree. For instance, 0([x + 1 | x ∈ [1, 2, 3]]) will stand for
the tree 0(2, 3, 4). A node with an empty children list will be considered to be a leaf.
Proofs For space reasons, some proofs have been moved to the Appendices. When no proof is
given directly below a lemma or a theorem, it can be found in Appendix D.
3
Architecture
Query language
In this Section, I present the
global architecture of the project
and quickly describe the purpose
of the modules I did not work on.
Application language
① Translation to QIR
Oracle
② QIR evaluation
QIR
Truffle
PPS
LRI
Capabilities DSL
③ Translation to native queries
Native queries
Truffle
④ Query evaluation
Since the beginning of my internship, I took part in the design of the global architecture of
the project, whose current state
is illustrated in Figure 1. It
works as follows.
Figure 1: Architecture
1. Queries specified in the application language (possibly augmented with a special syntax for
querying primitives) are detected in the application code. We refer to the result of this
detection as the query boundary : code outside this boundary will be translated to Truffle
nodes and evaluated in the application runtime, whereas code inside this boundary will be
evaluated by the database. Inside the boundary, query parts that can be expressed in QIR
are mapped to QIR constructs, and the remaining is propagated as Truffle nodes. This
translation phase also maps the application language data model to the QIR data model.
For now, we make the assumption that a same construct has identical semantics in the
application language and in QIR (e.g., tuple navigations such as user.name return the
same result in JavaScript and in QIR, and Ruby integers behave like QIR integers), and we
allow a lossy translation when this is not the case. The QIR constructs will be described
in Section 4. The translation phase, designed at PPS, was not the point of my internship
3
and will not be further described. To understand the remaining of this document, you
can safely assume that query boundaries are specified by the application developer with a
quote/antiquote syntax as in [3]. For example, in the following code snippet (written in an
imaginary application language in which queries are specified using a SQL-like syntax),
1 f ( i d ) := i d = 1 ;
2 r e s u l t := null ;
3 <@ %r e s u l t := s e l e c t ∗ from db( " ads " ) as t where %f ( t . i d ) ) @>;
4 f o r t in r e s u l t do . . .
the <@ @> quotes delimit the query boundary, whereas the % antiquote annotate identifiers
bound outside the query boundary. During the application execution, the f and result
assignments are evaluated in the application runtime, the quoted code is translated to a
QIR expression containing a Truffle reference %f and the dependencies of this quoted code
(i.e., the Truffle definition of f) are passed, along with the QIR code, to the next module
in order to be evaluated in the database.
2. QIR constructs are partially evaluated within the framework. The goal of this rewriting
pass is to feed the database with queries that can be easily optimized by its query optimizer.
This module relies on a formal description of the operators that the database can natively
evaluate, in order to progressively “improve” the query representation with local reductions.
Providing this description through a well-defined API is the first role of the “Capabilities
DSL” (domain-specific language), thus maintaining the nice abstraction over the choice of
target database. The partial evaluation of QIR expressions will be described in Section 5.
This Capabilities DSL and description mechanism, designed at PPS, was not the point
of my internship and will not be further described. To understand the remaining of this
document, you can safely assume that it is an API providing information such as This
database supports filters, but not joins and a filter is compatible if the filter condition has
the shape λt. t.attr = x and if x is a constant value, but not if x is a free variable.
3. After the rewriting pass, QIR expressions that can be expressed in the database native
query language are mapped to native query constructs, and the remaining is translated
to Truffle nodes. This translation phase also maps the QIR data model to the database
data model. As in step 1, we make the assumption that a same construct has identical
semantics in QIR and in the database query language, and we allow a lossy translation
when this is not the case. Providing a way to translate compatible QIR operators into
constructs of the database query language through a well-defined API is the second role
of the Capabilities DSL, thus maintaining again the abstraction over the choice of target
database. For the same reason as above, you can simply assume that this API provides a
method mapping natively supported operators (according to the Capabilities DSL) to the
corresponding query string in the database query language.
4. Native queries go through the database optimizer before their evaluation in the database
query processor. Such queries may contain references to Truffle definitions for three reasons:
(i) the query in the application language was refering to an identifier defined outside the
query boundary, (ii) the quoted application code contained features that could not be
translated into QIR, and (iii) the query representation in QIR contained features that
could not be translated into the database query language. In all these cases, the Truffle
runtime embedded in the database is called by the query processor to evaluate the code
corresponding to these references. The result of the query is then returned to the host
language. This backward translation maps database native data values to the QIR values
(using again the Capabilities DSL translation API), then QIR values to the application
language data model. Again, this evaluation mechanism, designed at Oracle Labs, was not
the point of my internship and will not be further described in this document.
This architecture has the advantage of interleaving an abstraction layer, composed of QIR
and Truffle, between the application layer and the database layer. Beyond added cleanliness,
4
this means that no work has to be done on existing database layer modules to support a new
application language, and no work has to be done on existing application layer modules to
support a new database.
4
QIR calculus
For the architecture described in Section 3 to work, QIR constructs have to be simple enough
to be mapped to as many mainstream application language constructs and database language
constructs as possible, but also expressive enough to represent interesting queries.
To design a QIR calculus satisfying these requirements, I first reviewed existing languageintegrated query frameworks (cf. Appendix F), work that I had for the most part already done in
a previous internship on a similar area[13]. I observed that existing database query languages all
share high-level querying primitives (e.g., filtering, joins, aggregation), even though they support
these features to various degrees. Moreover, the corresponding language constructs are almost
always only a thin language layer hiding to various extents an algebra of operators inspired by
the relational algebra. I also noticed the tension between database languages with a declarative
descriptions of queries and application languages in which queries are built imperatively.
Therefore, in order to represent queries in a manner agnostic of the application and database
languages, I designed QIR as a small extension of the pure lambda-calculus, made of (i) expressions, i.e., variables, lambdas, applications, constants, functions (conditionals, arithmetic
operations, etc.) as well as list/tuple constructors and destructors, and (ii) operators, i.e.,
constructors representing the common querying primitives mentioned above. Thus, queries
built imperatively in the application code can easily be translated to expressions manipulating
operators whereas declarative queries correspond more to trees of operators, and one can go from
the first representation to the second by partially evaluating the corresponding QIR term. The
remaining of this section is a formal description of the QIR calculus.
4.1
QIR data model
QIR data objects are called values. The following grammar describes their syntax (for future use
in this document) and hierarchy.
hValuei
::= hNumber i
| hStringi
| hBool i
hNumber i ::= h64-bit IEEE 754 i
hStringi
::= hUTF-16 i
hBool i
::= true
| false
For now, the QIR data model only contains the necessary for a proof of concept. It will have
to be extended with more primitive types for our framework to be used in real-world applications.
4.2
QIR operators
Operators represent computation on database tables. Each operator has children expressions (in
parentheses) and configuration expressions (in subscript). As in the relational algrebra, children
provide the operator input tables whereas configurations describe the computation on table
elements. The following grammar describes their syntax (for future use in this document) and
hierarchy.
hOperator i ::= ScanhExpr i ()
| SelecthExpr i (hExpr i)
| LimithExpr i (hExpr i)
| GrouphExpr i,hExpr i (hExpr i)
| ProjecthExpr i (hExpr i)
| SorthExpr i (hExpr i)
| JoinhExpr i (hExpr i, hExpr i)
5
These operators represent the following database computations:
• Scantable () outputs the list (in a non-deterministic order) of elements in the target database
table denoted by the expression table.
• Selectf ilter (input) outputs the list of elements v in the list corresponding to input such
that f ilter v reduces to true.
• Projectf ormat (input) outputs the list of elements corresponding to f ormat v, with v
ranging in the list corresponding to input.
• Sortcomp (input) outputs the list of elements v in the list corresponding to input ordered
according to comp v ascending.
• Limitlimit (input) outputs the limit first elements of the list corresponding to input.
• Groupeq,agg (input) outputs the list of elements corresponding to agg g for each group g
in the partition of the elements v in the list corresponding to input, according to eq v.
• Joinf ilter (input1 , input2 ) outputs the join of the elements v1 in the list corresponding to
input1 and the elements v2 in the list corresponding to input2 , such that f ilter v1 v2
reduces to true.
For the same reason as above, QIR operators are still limited. Nervertheless, they already
capture a consequent part of the relational algebra.
4.3
QIR expressions
QIR expressions are the core of the QIR calculus. The following grammar describes their syntax
(for future use in this document) and hierarchy.
hExpr i
hLambdai
::=
|
|
|
|
|
|
|
|
|
|
hVariablei
hLambdai
hApplicationi
hConstanti
hValueConsi
hValueDestr i
hValueFuni
hBuiltinFuni
hDataRef i
hTruffleNodei
hOperator i
hTupleConsi ::= ‘tnil’
| ‘tcons’ hStringi hExpr i hExpr i
hValueDestr i ::= hListDestr i
| hTupleDestr i
hListDestr i
hTupleDestr i ::= ‘tdestr’ hExpr i hStringi
hValueFuni
::= ‘λ’ hVariablei ‘.’ hExpr i
hApplicationi ::= hExpr i hExpr i
hConstanti
::= hValuei
::= ‘-’ hExpr i
| hExpr i ‘and’ hExpr i
| ‘if’ hExpr i ‘then’ hExpr i ‘else’
hExpr i
| ...
hBuiltinFuni ::= ‘avg’ hExpr i
| ...
hValueConsi ::= hListConsi
| hTupleConsi
hListConsi
::= ‘destr’ hExpr i hExpr i hExpr i
hDataRef i
::= ‘nil’
| ‘cons’ hExpr i hExpr i
::= ‘db.table_name’
hTruffleNodei ::= ‘truffle<id>’
Lambda-abstractions (resp. applications, constants) represent functions (resp. function
applications, constants) of the application language. The constructor for lists takes an expression
for the head and an expression for the tail. Tables are represented by lists, since the output of
queries might be ordered. Tuples are constructed as a list of mappings: the constructor for tuples
takes a string for the mapping name, an expression for the mapping value and an expression for
the tail of the mapping list. The list destructor has three arguments: the list to destruct, the
term to return when the list is nil and a function with two arguments λh.λt.M to treat the
case when the list has a head and a tail. The tuple destructor has two arguments: the tuple
to destruct and the attribute name of the value to return. Finally, built-in functions represent
6
common database functions (such as aggregation) for easier recognition and translation into
native languages. The formal definition of expression reductions is given in Section 5.
In this document, we will sometimes use the syntactic sugar let x = M in N for (λx. N )M
but the let . . . in constructor is not part of the expression language. Similarly, we will sometimes
use the notation t.attr for tdestr t "attr", not x for if x then false else true and let rec x =
M in N as a shortcut for the application of a fixpoint combinator.
5
QIR evaluation
As discussed in the beginning of Section 4, the purpose of the QIR partial evaluation mechanism
is to transform the QIR representation of a query written in the (possibly imperative) style
of the application language into a QIR representation of the same query that is easier to
translate to the database native language and offers more optimization opportunities to its
optimizer. Today’s databases use a tree of algebra operators (plan) to represent computation.
Optimizations consist in commuting operators in the plan or applying local tree rewritings. The
plan evaluation usually corresponds to (a variation of) a bottom-up evaluation of each operator,
where a particular implementation is chosen for each operator for best performance. Therefore,
for a same computation, such database engines benefit from working on a large, single plan
instead of multiple small plans using the result of each other, since (i) more optimizations become
available, (ii) more statistics can be gathered to pick the best implementation for each operator
and (iii) no intermediate result needs to be materialized and transferred.
Consider the following example of query (written in an imaginary application language in
which queries are specified using a SQL-like syntax).
1 r e s u l t := null ;
2 <@ u s e r s := db( " u s e r s " ) ; ads := db( " ads " ) ;
3 f i l t e r ( x , y ) := ( x . i d = y . u s e r i d ) and ( y . timestamp > 1 2 3 4 ) ;
4 p r e t t i f y ( s ) := . . . ;
5 %r e s u l t = s e l e c t p r e t t i f y ( name ) as u_name , d e s c r as a_descr
6
from u s e r s as u , ads as a where f i l t e r ( u , a ) @>
7 f o r t in r e s u l t do . . .
The quoted code contains a query as well as the definition of two UDFs: filter, which is a
quite simple filter function and prettify, which contains application languages features that
cannot be translated to QIR. A direct translation of the quoted code to QIR would be as follows.
1
2
3
4
5
let users = db.users in let ads = db.ads in
let f ilter = λx. λy. (tdestr x "id") = (tdestr y "userid") and (tdestr y "timestamp" > 1234) in
let prettif y = truffle<0> in
Projectλt. tcons "u_name" (prettif y (tdestr t "name")) (tcons "a_descr" (tdestr t "descr") tnil) (
Joinλu. λa. f ilter u a (Scanusers (), Scanads ()))
Notice that the definition of prettify had to be translated to Truffle nodes (not shown here)
and is represented in the QIR code by a Truffle reference. Assuming that the target database
can only natively evaluate operators (e.g., SQL), the direct translation of the above QIR code to
the database language would lead to a query in which all operators issue calls to UDFs, greatly
impacting performances. Moreover, with no information about filter, the database optimizer
could not decide to push the selection timestamp > 1234 below the join or choose an index-based
implementation for the join. In fact, a better QIR representation for this query is as follows.
1
2
Projectλt. tcons "u_name" (truffle<0> (tdestr t "name")) (tcons "a_descr" (tdestr t "descr") tnil) (
Joinλu. λa. (tdestr u "id") = (tdestr a "userid") and (tdestr a "timestamp">1234) (Scandb.users (), Scandb.ads ()))
This second version can be obtained from the first by a reduction of the QIR code. As we will
see later in this section, applying classical reduction strategies (e.g., call-by-value, call-by-name,
lazy evaluation) until a normal form is reached does not satisfy our requirements, since contracting
7
some redexes might scatter parts of a same query. Instead, we will characterize the shape of
“good” QIR code with a measure on QIR expressions in Section 5.2, then investigate reduction
strategies guided by this measure in Sections 5.3 and 5.4.
5.1
QIR reduction rules
In this section, I formally define the reduction of QIR expressions, then state two useful properties
on the QIR calculus equipped with this reduction.
Definition 1. The capture-avoiding variable substitution in QIR expressions (cf. Section 4.3) is
defined as in the pure lambda-calculus. The substitution of a variable x in e1 by an expression e2
is denoted e1 {e2 /x}.
Definition 2. The reduction rules for QIR expressions consist in the β-reduction rule augmented
with the δ-reduction rules for destructors and ρ-reduction rules for primitive types.
(λx. e1 ) e2
→β
e1 {e2 /x}
destr nil enil econs
destr (cons ehead etail ) enil econs
tdestr (tcons "name1" eval1 etail ) "name1"
tdestr (tcons "name1" eval1 etail ) "name2"
→δ
→δ
→δ
→δ
enil
econs ehead etail
eval1
tdestr etail "name2"
if true then e1 else e2
if false then e1 else e2
true and e1
false and e1
...
→ρ
→ρ
→ρ
→ρ
e1
e2
true
e1
In this document, we will use the notation → for the relation →β ∪ →δ ∪ →ρ and talk about
redexes for β-redexes, δ-redexes and ρ-redexes indifferently. Furthermore, we will denote by →n
the n-step reduction relation and by →∗ the reflexive, transitive closure of →.
Theorem 1. QIR with the reduction rules of Definition 2 satisfies the Church-Rosser property.
As we will see in the following sections, this is an important property, since it allows us to
apply any reduction strategy while preserving the semantics of the input expression. Consequently,
extending QIR with non-confluent constructs (e.g., side-effects) would require significant work.
Theorem 2. QIR with the reduction rules of Definition 2 verifies the Standardization theorem,
that is, if an expression e has a normal form, it can be obtained from e by reducing successively
its leftmost outermost redex.
5.2
A measure for “good” output plans
In this Section, I formally define the measure characterizing “good” QIR expressions.
Definition 3. Supported operators are operators that are supported by the database, according
to the mechanism discussed in Section 3.
Definition 4. Supported expressions are expressions that are supported by the database inside
operator configurations, according to the mechanism discussed in Section 3.
Definition 5. An operator is called compatible if it is supported and has supported expressions
as configuration. Any other operator or expression that is not an operator is called incompatible.
Note that the compatibility of an operator is independent of the compatibility of its children.
8
Definition 6. Let e be an expression and F a
context. We say that F is a fragment if either
e = C[t(t1 , . . . , ti−1 , F [e1 , . . . , en ], ti+1 , . . . , tj )] or
e = F [e1 , . . . , en ], where:
• C is a one-hole context with arbitrary expressions
• t is an incompatible j-ary expression
• F is a n-hole context made only of compatible
operators
• t1 , . . . , ti−1 , ti+1 , . . . , tj and F [e1 , . . . , en ] are
the j children of t
• e1 , . . . , en are incompatible expressions
C
t
F
e1
…
…
en
Figure 2: A fragment F
This definition is illustrated in Figure 2.
Definition 7. Let e be an expression. We define Op (e) as the number of operators in e, Comp (e)
as the number of compatible operators in e and Frag (e) as the number of fragments in e.
Definition 8. Let e be an expression. We define the measure for good output expressions as
the pair M (e) = (Op (e) − Comp (e) , Frag (e)). Moreover, expressions are ordered using the
lexicographical order on this measure.
The intuition about this measure is as follows. We want to consider reduction chains for
which the measure decreases, that is, reductions e →∗ e0 verifying one of the following conditions.
• Op (e0 ) − Comp (e0 ) < Op (e) − Comp (e)
• Op (e0 ) − Comp (e0 ) = Op (e) − Comp (e) and Frag (e0 ) < Frag (e)
When a reduction transforms an incompatible operator into a compatible operator, the measure
Op (e) − Comp (e) decreases. For instance, in Example 1, Op (e) − Comp (e) = 2 − 1 = 1 (since f
is a variable, it is not supported by the database) and Op (e0 ) − Comp (e0 ) = 2 − 2 = 0.
Example 1.
e
→∗
let f = λx. x = 2 in Selectλt.f (t.id) (Scane1 ()) →∗ Selectλt.
e0
t.id=2 (Scane1 ())
When a reduction inlines operators as children of other operators, Op (e) − Comp (e) stays the
same (since no operator is created nor modified) and Frag (e) decreases. For instance, in Examples
2 and 3, Op (e) − Comp (e) = Op (e0 ) − Comp (e0 ) = 2 − 2 = 0, Frag (e) = 2 and Frag (e0 ) = 1.
Example 2.
→∗
→∗ Selectλt.
e0
t.id=2 (Scane1 ())
→∗
∗
t.id=2 (x) in f (Scane1 ()) → Selectλt.
e0
t.id=2 (Scane1 ())
e
let x = Scane1 () in Selectλt.
t.id=2 (x)
Example 3.
e
let f = λx. Selectλt.
In the infinite recursion of Example 4, unfolding one step of recursion does not change the
measure, which hints that the plan is not getting better.
Example 4.
e
let rec f = λx. Selectλt.
f (Scane1 ())
e0
let rec f = λx. Selectλt. t.id=2 (f x) in
→∗ Select
λt. t.id=2 (f (Scane1 ()))
→∗
t.id=2 (f
x) in
This intuition generalizes well to binary operators, with which a reduction can duplicate
operators (e.g., Example 5). If all duplicated operators are compatible, Op (e) − Comp (e) stays
the same (since the duplicated operators are compatible) therefore it is a good reduction if the
9
number of fragments decreases. In this case, the database will get a bigger plan in which it can
perform optimizations (such as caching). Conversely, if a duplicated operator is incompatible,
Op (e) − Comp (e) increases, meaning that the reduction should not happen. In this situation,
this operator should in fact be evaluated once in memory and its result cached to be reused
multiple times during the expression evaluation. For instance, in Example 5, if e1 is supported,
Op (e) − Comp (e) = 3 − 3 = 0, Op (e0 ) − Comp (e0 ) = 5 − 5 = 0, Frag (e) = 2 and Frag (e0 ) = 1,
and if e1 is not supported, Op (e) − Comp (e) = 3 − 2 = 1, Op (e0 ) − Comp (e0 ) = 5 − 3 = 2.
Example 5.
e
let x = Selecte1 (Scane2 ()) in
Joinλt1.λt2. t1.id=t2.name (x, x)
e0
Joinλt1.λt2. t1.id=t2.name (Selecte1 (Scane2 ()),
→∗ Select (Scan ()))
e1
e2
→∗
Lemma 1. The measure M of Definition 8 induces a well-founded order on expressions.
5.3
An exhaustive reduction strategy
Given the measure M described in Section 5.2 and an expression e, the question is now to find a
reduced term e0 such that e →∗ e0 and M (e0 ) is minimal. During my internship, I first designed
and implemented the exhaustive reduction strategy presented formally in this section.
Definition 9. Let e be an expression. We denote by Rdxs (e) the set of redexes in e, and for
r ∈ Rdxs (e), we write e →r e0 to state that e0 can be obtained from e in one reduction step by
contracting r.
Definition 10. Let e be an expression. We recursively define Red0 (e) as the singleton {e} and
00
0
0
0
00
∀n
S > 0, Redn (e) = {e | e ∈ Redn−1 (e) , r ∈ Rdxs (e ) , e →r e }. We write 0 Reds (e) to denote
n∈N Redn (e). Finally we define MinReds (e) as the set argmine0 ∈Reds(e) M (e ).
Notice that for some expressions e (e.g., expressions containing recursions), Reds (e) and
MinReds (e) may be infinite and an algorithm that iteratively constructs the Redn (e) may not
terminate. Conversely, if e is strongly normalizing, Reds (e) and MinReds (e) are finite and such an
algorithm terminates. In general, the parallel exploration of all redexes leads to a combinatorial
explosion, and I will show in Section 6 that my implementation of this algorithm timeouts on
moderately large expressions.
Theorem 3. The reduction strategy of Definition 10 is exhaustive, i.e., ∀e, ∀e0 , e →∗ e0 ⇔ e0 ∈
Reds (e).
Proof. (⇐) If e0 ∈ Reds (e) then ∃n ∈ N, e0 ∈ Redn (e). The proof follows an induction on n.
If n = 0 then e = e0 therefore e →∗ e0 . For n > 0, we know that ∃e00 ∈ Redn−1 (e) , ∃r ∈
Rdxs (e00 ) , e00 →r e0 . Thus, e00 → e0 and by induction hypothesis e →∗ e00 therefore e →∗ e0 .
(⇒) If e →∗ e0 then ∃n ∈ N, e →n e0 . The proof follows an induction on n. If n = 0 then e = e0
and e0 ∈ Red0 (e) ⊆ Reds (e). For n > 0, we know that ∃e00 , e →n−1 e00 → e0 . By induction
hypothesis, e00 ∈ Reds (e) and ∃n0 ∈ N, e00 ∈ Redn0 (e). Therefore e0 ∈ Redn0 +1 (e) ⊆ Reds (e).
Lemma 2. ∀e, MinReds (e) 6= ∅.
Theorem 4. The reduction strategy of Definition 10 is optimal, i.e., ∀e, ∀e0 , e →∗ e0 ⇒ e0 ∈
MinReds (e) ∨ ∃e00 ∈ MinReds (e) , M (e00 ) < M (e0 ).
Proof. Suppose that e →∗ e0 . Using Theorem 3 we know that e0 ∈ Reds (e). Using Lemma
2, we know that MinReds (e) =
6 ∅ and we denote by Mmin the measure M of the elements of
MinReds (e). Then, either M (e0 ) = Mmin and e0 ∈ MinReds (e) by definition, or M (e0 ) > Mmin
and using again Lemma 2 we can find e00 ∈ MinReds (e) , M (e00 ) < M (e0 ).
As discussed above, this reduction strategy is not realistic in terms of complexity for a direct
implementation and may not even terminate in some cases. Nevertheless, Theorem 4 states that
it can be used as a yardstick to discuss the optimality of other reduction strategies.
10
5.4
A heuristic-based reduction strategy
Given the conclusion of Section 5.3, I started working on a more efficient evaluation strategy. In
this section, I formally describe an efficient, always-terminating heuristic, corresponding to a
partial exploration of the possible reductions, then give proofs of its properties.
This heuristic-based reduction strategy supposes the existence of an integer constant Φ
representing the “fuel” that can be consumed by the reduction. It consists of two main passes.
The first pass tries to reduce redexes that could make operators compatible by assuming that
(i) operators with free variables in their configuration have few chances to be compatible and
(ii) reducing redexes inside operator configurations increases the odds of making an operator
compatible. The second pass tries to decrease the number of fragments by reducing redexes
inside the children expressions of the operators. Both passes guarantee that the measure of the
global expression always decreases after a number of reduction steps bounded by Φ.
For readability purposes, I first describe the search space tree explored by the heuristic then
define the result of the algorithm as a particular expression in this search space. The actual
implementation corresponds to a depth-first exploration of the search space, with decision points
and back-tracking.
Definition 11. Let e be an expression. We define its operator contexts OpContexts (e) as the set
of contexts {C[] | e = C[op], op is an operator}.
Definition 12. Let e be an expression and C[] an operator context in OpContexts (e). We
define the free variables of a configuration ConfigFVars (e, C[]) as the list of variables [v | e =
C[opC 0 [v] (. . .)], v is a free variable in C 0 [v]] sorted using a depth-first left-to-right traversal of
C 0 [v].
Definition 13. Let e be an expression and C[] an operator context in OpContexts (e). We define
the redexes of a configuration ConfigRdxs (e, C[]) as the list of redexes [r | e = C[opC 0 [r] (. . .)], r is a
redex] sorted using a depth-first left-to-right traversal of C 0 [v].
Definition 14. Let e be an expression and C[] an operator context in OpContexts (e). We define
the children expressions Children (e, C[]) as the list of expressions [ci | e = C[op... (c1 , . . . , cn )]].
Definition 15. The following three definitions are mutually recursive and therefore presented
together in this document.
Let e be an expression and e0 a subexpression of e. We define HMakeRedex (e, e0 ) as
• e00 , such that e →r e00 , if e0 is already a redex r
• HMakeRedex (e, e00 ) if e0 = (e00 e1 ), e0 = destr e00 e1 e2 , e0 = tdestr e00 s, e0 = if e00 then e1
else e2 , e0 = e00 and (true/false) or e0 = (true/false) and e00 (and similarily for other
ρ-redexes).
• HInlineVar (e, v) if e0 is a variable v
• None otherwise
Let e be an expression and e0 a subexpression of e. We define HContractLam (e, e0 ) as
• HMakeRedex (e, e00 ) where e = C[e00 ], if e = C[e0 e1 ], e = C[e1 e0 ], e = C[destr e0 e1 e2 ], e =
C[destr e1 e0 e2 ], e = C[destr e1 e2 e0 ], e = C[tdestr e0 s], e = C[if e1 then e0 else e2 ]
or e = C[if e1 then e2 else e0 ]
• HContractLam (e, e00 ) where e = C[e00 ], if e = C[λv. e0 ], e = C[cons e0 e1 ], e = C[cons e1 e0 ],
e = C[tcons s e0 e1 ] or e = C[tcons s e1 e0 ]
• None otherwise
Let e be an expression and v a variable in e. We define the inlining HInlineVar (e, v) as
• None if v is free in e
• HContractLam (e, l), where l is the λ binding v in e, otherwise
11
HMakeRedex (e, e0 ) corresponds to (i) contracting e0 if e0 is already a redex and (ii) contracting
a necessary redex in order for e0 to become a redex, otherwise. HContractLam (e, e0 ) corresponds
to (i) contracting the redex containing e0 if e0 is already part of a redex and (ii) contracting a
necessary redex in order for e0 to become part of a redex, otherwise. HInlineVar (e, v) corresponds
to contracting a necessary redex in order to eventually substitute v by its definition.
Consider for instance the expression e given in Example 6. To go towards the inlining of the
variable tl in the configuration of the Scan operator, the heuristic calls HInlineVar (e, tl), which
tries to contract the binding lambda by calling HContractLam (e, λtl. Scanf tl ()). This lambda
cannot be contracted before the lambda above it, and therefore HContractLam (e, λhd. λtl. . . .)
is called recursively. Again, this lambda cannot be contracted before the head destr, but this
requires its first child expression to be reduced to a cons. This is the reason why the heuristic
recursively calls HMakeRedex (e, (λx. x) (cons . . .)), which contracts its argument redex.
Example 6.
e
destr ((λx. x) (cons 1 (cons 2 nil)))
false (λhd. λtl. Scanf tl ())
e0
destr (cons 1 (cons 2 nil))
→
false (λhd. λtl. Scanf tl ())
→
Definition 16. Let e be an expression, C[] an operator context in OpContexts (e), V = [v ∈
ConfigFVars (e, C[]) | HInlineVar (e, v) =
6 None] and R = ConfigRdxs (e, C[]). We define the onestep reduction of the operator configuration HOneRedConfig (e, C[]) as
• HInlineVar (e, v) such that v is the first element of L, if L is non empty
• e0 , such that e →r e0 and r is the first element of R, if R is non empty
• None otherwise
This means that for a given operator, the heuristic first tries to find a free variable that can be
inlined, then if there is no such variable, it reduces the configuration using a leftmost outermost
(call-by-name) reduction strategy, and finally if there is no redex to reduce, it returns None.
In the following definitions, we will describe (search space) trees of expressions using the
notations given in Section 2.
Definition 17. Let e be an expression, C[] an operator context in OpContexts (e), φ an integer
and e0 = HOneRedConfig (e, C[]). We define HRedConfigφ (e, C[]) as
• e() if φ = 0 or if e0 = None
• e0 ([HRedConfigΦ (e0 , op) | op ∈ OpContexts (e0 ) ]) if M (e0 ) < M (e)
• e([HRedConfigφ−1 (e0 , op) | op ∈ OpContexts (e0 ) ]) otherwise
Definition 18. Let e be an expression. The search space HConfigSSΦ (e) after the first pass is
defined as e([HRedConfigΦ (e, op) | op ∈ OpContexts (e) ]).
HConfigSSΦ (e) corresponds to the recursive exploration of the search space tree for reductions
in configurations, with the guarantee that search space subtrees in which the measure does not
decrease have a depth bounded by Φ.
Definition 19. Let e be an expression. The result expression HConfigΦ (e) of the first pass is
defined as follows. First, the rewrite rule x(s1 , . . . , si , x(), si+2 , sn ) → x(s1 , . . . , si , si+2 , sn ) is
applied on HConfigSSΦ (e) as much as possible. Then, the leftmost leaf is chosen.
The rewrite rule corresponds to pruning search space subtrees in HConfigSSΦ (e) that failed to
reach an expression with smaller measure, whereas taking the leftmost leaf is a heuristic decision
corresponding to never backtrack to expressions with larger measure than the current candidate
(for performance reasons).
Definition 20. Let e be an expression, C[] an operator context in OpContexts (e) and L = [c ∈
Children (e, C[]) | HMakeRedex (e, c) =
6 None]. We define the one-step reduction of the operator
children HOneRedChild (e, C[]) as
12
• HMakeRedex (e, c) such that c is the first element of L, if L is non empty
• None otherwise
This means that for a given operator, the heuristic tries to find a child expression that can
be reduced (using a call-by-name reduction strategy) in order to regroup fragments, and if there
is no such child, it returns None.
Definition 21. Let e be an expression, C[] an operator context in OpContexts (e), φ an integer
and e0 = HOneRedChild (e, C[]). We define HRedChildφ (e, C[]) as
• e() if φ = 0 or if e0 = None
• e0 ([HRedChildΦ (e0 , op) | op ∈ OpContexts (e0 ) ]) if M (e0 ) < M (e)
• e([HRedChildφ−1 (e0 , op) | op ∈ OpContexts (e0 ) ]) otherwise
Definition 22. Let e be an expression. The search space HChildrenSSΦ (e) after the second pass
is defined as e([HRedChildΦ (e, op) | op ∈ OpContexts (e) ]).
HChildrenSSΦ (e) corresponds to the recursive exploration of the search space tree for reduction
in children, with the guarantee that search space subtrees in which the measure does not decrease
have a depth bounded by Φ.
Definition 23. Let e be an expression. The result expression HChildrenΦ (e) of the second pass
is defined as follows. First, the rewrite rule x(s1 , . . . , si , x(), si+2 , sn ) → x(s1 , . . . , si , si+2 , sn ) is
applied on HChildrenSSΦ (e) as much as possible. Then, the leftmost leaf is chosen.
Similarly to Definition 19, the rewrite rule corresponds to pruning search space subtrees in
HChildrenSSΦ (e) that failed to reach an expression with smaller measure, whereas taking the
leftmost leaf is a heuristic decision.
Definition 24. Let e be an expression. The result expression HMinRedΦ (e) of the heuristic
reduction is defined as HChildrenΦ (HConfigΦ (e)).
Figure 3 illustrates the construction of the two search spaces (cf. Definitions 18 and 22)
leading to the computation of HMinRedΦ (e). Dots and stars represent expressions considered
by the heuristic as a possible reduction of their parent, but stars correspond to the special case
where the measure of the expression is smaller than the measure of its parent, i.e., when the
heuristic made progress.
We will now continue towards a proof of termination of this reduction strategy (Theorem 5).
Definition 25. Considering an expression with subexpressions as a tree with subtrees, we define
a traversal of an expression as follows, starting from the head expression:
• for a nullary expression, visit this expression.
• for λx. e1 , cons e1 e2 and tcons s e1 e2 , visit this expression then traverse the children
from left to right (e1 then e2 . . . then en ).
• for other n-ary expressions (e.g., (e1 e2 ), destr e1 e2 e3 , etc.), traverse the children from
left to right (e1 then e2 . . . then en ), then visit the expression.
We denote by TraversalPos (e, e0 ) the position of a subexpression e0 in this traversal of e.
TraversalPos (e, .) actually induces a well-founded order on the locations of the subexpressions that are arguments of the recursive calls to HInlineVar (., .), HMakeRedex (., .) and
HContractLam (., .) made by the heuristic. It will be used in inductions in some of the following
proofs.
13
destr 16
… …
(
…
…
λx 1
x 2
HConfigSSФ
HMinRedФ
… …
false 9
) 8
cons 5
2 6
…
HChildrenSSФ
λtl 11
cons 3
1 4
λhd 10
nil 7
Scan 15
(
f 12
Figure 3: HMinRedΦ (e) and the search spaces
HConfigSSΦ (e) and HChildrenSSΦ (HConfigΦ (e))
) 14
tl 13
Figure 4: Example of expression traversal
For example, consider again the expression e from Example 6 and its traversal given in
Figure 4. Remember the chain of recursive calls made by the heuristic (HInlineVar (e, tl), which
calls HContractLam (e, λtl. Scanf tl ()), which calls HContractLam (e, λhd. λtl. . . .), which indirectly calls HMakeRedex (e, (λx. x) (cons . . .)), which finally contracts this redex) and notice
that TraversalPos (e, tl) > TraversalPos (e, λtl. Scanf tl ()) > TraversalPos (e, λhd. λtl. . . .) >
TraversalPos (e, (λx. x) (cons . . .)).
Lemma 3. For all expression e the following properties hold.
• For a variable e0 in e, HInlineVar (e, e0 ) returns None or contracts a redex.
• For a subexpression e0 in e, HMakeRedex (e, e0 ) returns None or contracts a redex.
• For a subexpression e0 in e, HContractLam (e, e0 ) returns None or contracts a redex.
These three properties not only show that the objects used to describe the heuristic are well
defined, but also mean that correct implementations of HInlineVar (., .), HMakeRedex (., .) and
HContractLam (., .) terminate.
Lemma 4. For all expression e, the search space HConfigSSΦ (e) (resp. HChildrenSSΦ (e)) has
bounded size.
Theorem 5. This reduction strategy always terminates, that is, for an input expression e, it
always considers a finite number of expressions e0 such that e →∗ e0 in order to find HMinRedΦ (e).
Proof. Using Lemma 4, this reduction strategy only considers a finite number of expressions
e1 such that e →∗ e1 in order to find HConfigΦ (e), then using Lemma 4 again, it only
considers a finite number of expressions e2 such that HConfigΦ (e) →∗ e2 in order to find
HChildrenΦ (HConfigΦ (e)) = HMinRedΦ (e).
We will now continue towards a proof that under some hypotheses this reduction strategy is
complete, that is, it returns an optimal result for a big enough fuel value (Theorem 6).
In the following proofs, we will need a way to track subexpressions of an expression and
understand how they are duplicated, erased or simply moved across successive reductions of the
main expression. Borrowing an idea used in [2] to define residuals, we annotate all subexpressions
of the original expression with unique identifiers, simply propagate such identifiers along reductions
without altering them, and reason on identifiers in the final expression.
Definition 26. Let e0 be a QIR expression, e a subexpression of e0 and en a QIR expression
such that e0 →∗ en . We define Occurences (en , e) as the set of subexpressions of en with the same
identifiers as e when identifiers are propagated along the reduction chain e0 →∗ en .
For instance, the QIR reduction (λx. λy. f x x) z z → (λy. f z z) z → f z z
can be annotated as follows: (λ1 x. λ2 y. f 3 x4 x5 ) z 6 z 7 → (λ2 y. f 3 z 6 z 6 ) z 7 → f 3 z 6 z 6 .
14
One can see, just by looking at identifiers, that the first z variable produced both z variables of
the final expression: it has two occurences in the final expression. The other one disappeared: it
has no occurence in the final expression.
A more formal definition of Occurences (en , e) is given in Appendix C. The following definitions
use the Occurences (., .) notation but do not depend on the internals of its definition.
Definition 27. Let e be an expression, e0 a subexpression of e and r a redex such that e →r e00
and Occurences (e00 , e0 ) = ∅. Borrowing a concept from [2], we say that r erases e0 if e0 is not part
of the redex r, that is if e0 is a subexpression of an expression e00 such that one of the following
conditions holds.
• r = (λx. e1 ) e00 and x is not free in e1
• r = if true then e1 else e00
00
• r = destr nil e1 e
• r = if false then e00 else e1
• r = destr (cons . . .) e00 e1
• r = true and e00 (and similarly for other
00
• r = tdestr (tcons "s1" e1 e ) "s1"
ρ-redexes)
• r = tdestr (tcons "s1" e00 e1 ) "s2"
Definition 28. Let e be an expression, e0 a subexpression of e and r a redex such that e →r e00
and Occurences (e00 , e0 ) = ∅. We say that r consumes e0 if e0 is part of the redex r, that is if one
of the following conditions holds.
• r = tdestr e0 s
• r = e0
0
• r = if e0 then e1 else e2
• r = (e e1 )
0
• r = e0 and e1 (and similarly for other
• r = (λv. e1 ) e2 and e is the variable v
free in e1
ρ-redexes)
• r = destr e0 e1 e2
Lemma 5. Let e be an expression, e0 a subexpression of e and r a redex in e. If e →r e00 and
Occurences (e00 , e0 ) = ∅, then r either consumes or erases e0 .
Definition 29. Let e be an expression, e0 a subexpression of e and r a redex in e such that
e →r e00 , we say that r duplicates e0 if |Occurences (e00 , e0 ) | > 1.
Definition 30. Let e0 be an expression and e0 a subexpression of e0 . We say that e0 cannot be
erased (resp. consumed, duplicated) if there is no reduction chain e0 →r1 . . . →rn en such that
rn erases (resp. consumes, duplicates) an element of Occurences (en−1 , e0 ).
Lemma 6. Let e be an expression and e0 a subexpression of e. The following properties hold.
• If e0 is a variable v and if HInlineVar (e, v) returns None then e0 cannot be consumed.
• If HMakeRedex (e, e0 ) returns None then e0 cannot be consumed.
• If HContractLam (e, e0 ) returns None then e0 cannot be consumed.
Definition 31. We say that an expression e has the fixed operators property if no operator in e
can be erased or duplicated.
Remember that the definition of compatible operators (Definitions 3, 4 and 5) depends on an
arbitrary external module describing the capabilities of the target database (cf. Section 3). The
following restriction allows us to reason on this module.
Definition 32. We say that a database verifies the stable compatibility property if, given an
expression e, an operator op in {op | C[] ∈ OpContexts (e) , e = C[op]} such that op is compatible
and an expression e0 such that e →∗ e0 , each operator op0 ∈ Occurences (op, e0 ) is also compatible.
This last definition should hold for a realistic database and an accurate description of its
capabilities. Indeed, it basically says that if an operator is compatible, any reduction either does
not affect the operator or helps the database by simplifying its configuration.
Lemma 7. Let e be an expression with fixed operators and r a redex in e. For a database with
stable compatibility, if e →r e0 then M (e0 ) ≤ M (e).
15
Lemma 8. Let e be a weakly-normalizing expression with fixed operators. For a database with
stable compatibility, the normal form of e has minimal measure.
Lemma 9. Let e be a weakly-normalizing expression with fixed operators, emin an expression
in MinReds (e) and e0 an expression such that e →∗ e0 and Op (e0 ) − Comp (e0 ) = Op (emin ) −
Comp (emin ). For a database with stable compatibility, an operator is compatible in emin if and
only if it is compatible in e0 .
Theorem 6. For databases with stable compatibility, the reduction strategy of Definition 24 is
complete on strongly-normalizing expressions with fixed operators. That is, for a database with
stable compatibility, given a strongly-normalizing input expression e with fixed operators, there
exists a fuel value Φ such that HMinRedΦ (e) ∈ MinReds (e).
Proof. Remember from Definition 10 that all expressions in MinReds (e) have same (minimal)
measure. Using Lemma 8, we know that the normal form eN of e is in MinReds (e). Let Mmin be
its measure. Consider now eh = HMinRedΦ (e). Using Theorem 4, we know that eh ∈ MinReds (e)
or M (eh ) > Mmin and we want to prove that the latter cannot happen. Suppose to the contrary
that M (eh ) > Mmin . Using the definition of M (cf. Definition 8), this means that one of the
two following statements holds.
• Op (eh ) − Comp (eh ) is greater than the first component of Mmin
• Op (eh ) − Comp (eh ) is equal to the first component of Mmin and Frag (eh ) is greater than
the second component of Mmin
We will prove that none of these cases can happen.
• Suppose that Op (eh ) − Comp (eh ) is greater than the first component of Mmin . Since
e has the fixed operators property, there is a one-to-one correspondance between the
operators of eN and eh . Therefore, we know that Comp (eh ) < Comp (eN ) and there exists
an operator op in e such that Occurences (eh , op) = {oph }, Occurences (eN , op) = {opN },
opN is compatible and oph is not compatible. Let ch (resp. cN ) be the configuration of oph
(resp. opN ). The question now is to understand how the first pass of the heuristic-based
algorithm (cf. Definition 19) could fail to make op compatible. Remember Lemma 3 telling
that HInlineVar (., .) and HMakeRedex (., .) either contract a redex or return None, and keep
in mind that such reductions maintain a single instance of ch in the reduced forms of eh
(fixed operator hypothesis). Since e is strongly-normalizing, this means that there is a fuel
value Φ allowing the heuristic to make enough calls to HInlineVar (., .) on the free variables
of ch in order to get to an expression e0h = C 0 [opc0h (. . .)] such that (i) there is no free variable
in c0h or (ii) calls to HInlineVar (e0h , .) return None for every free variable of c0h . Continuing
from this point, since e is strongly-normalizing, e0h and c0h are also strongly normalizing.
Thus, Theorem 2 tells that there is a fuel value Φ allowing the heuristic to reduce all
redexes of c0h and reach a normal form c00h and an expression e00h = C 0 [opc00h (. . .)]. Since we
supposed that the heuristic failed to make op compatible, this means that c00h is different
from cN . Using Theorem 1 (confluence), we know there is reduction e00h →∗ eN . Since
the redexes contracted in this chain cannot erase nor duplicate operators (fixed operator
hypothesis), the reduction chain can only affect c00h in the following ways.
– Substitute free variables in c00h . This cannot happen: by hypothesis, either (i) there is
no free variable in c0h and therefore in c00h or (ii) calls to HInlineVar (e0h , .) return None
for every free variable of c0h and using Lemma 6, such a variable cannot be consumed.
– Reduce redexes located inside c00h . This cannot happen since c00h is in normal form.
– Leave c00h untouched. This leads to a contradiction: c00h is equal to cN .
Therefore, there is a fuel value such that the heuristic makes op compatible. Now, taking
the maximum of the required values of Φ to make each operator compatible, there exists a
value of Φ such that Op (eh ) − Comp (eh ) is equal to the first component of Mmin .
• Suppose now that Op (eh ) − Comp (eh ) is equal to the first component of Mmin and Frag (eh )
is greater than the second component of Mmin . Since e has the fixed operators property,
there is a one-to-one correspondance between the operators of eN and eh . Using Lemma
9, we know that there exists an operator op in e such that Occurences (eh , op) = {oph },
16
Occurences (eN , op) = {opN }, opN and oph are both compatible, opN has a compatible
child operator cN and the child expression ch of oph is incompatible (i.e., not a compatible
operator). The question now is to understand how the second pass of the heuristic-based
algorithm (cf. Definition 23) could fail to reduce ch to a compatible operator. Remember
Lemma 3 telling that HMakeRedex (., .) either contracts a redex or returns None, and keep
in mind that such reductions maintain a single instance of oph in the reduced forms of eh
(fixed operator hypothesis). Since e is strongly-normalizing, this means that there is a fuel
value Φ allowing the heuristic to make enough calls to HMakeRedex (., .) on ch in order to
get to an expression e0h = C 0 [op... (. . . , c0h , . . .)] such that calls to HMakeRedex (c0h , .) returns
None. Since we supposed that the heuristic failed to reduce ch to a compatible operator,
this means that the head of c0h is different from the head of cN (which is a compatible
operator). Using Lemma 6, c0h cannot be consumed, and as the child expression of an
operator that cannot be erased, c0h cannot be erased either. According to Lemma 5 this
contradicts the confluence theorem telling that e0h →∗ eN . Therefore, there is a fuel value
such that the heuristic reduces ch to a compatible operator. Now, taking the maximum of
the required values of Φ to reduce the children of all operators, there exists a value of Φ
such that Frag (eh ) is equal to the second component of Mmin .
We also conjecture that the result of Theorem 6 still holds for weakly-normalizing expressions.
Conjecture 1. For databases with stable compatibility, the reduction strategy of Definition 24 is
complete on weakly-normalizing expressions with fixed operators. That is, for a database with
stable compatibility, given a weakly-normalizing input expression e with fixed operators, there
exists a fuel value Φ such that HMinRedΦ (e) ∈ MinReds (e).
See Appendix D for an intuition of why this holds and the current state of the proof. We will
now explain why none of the remaining hypotheses can be removed.
Stable compatiblity Consider a database for which the Scan operator is compatible if
its configuration has more than two free variables. Obviously, it would not have the stable
compatibility property, since inlining the definition of these variables could reduce the number of
free variables in the configuration. Take now expression e from Example 7. Since the heuristic
tries to inline variables before reducing redexes in the configurations, it will never consider
expression e0 , which is the only element of MinReds (e).
Example 7.
(λt. Scan(λx.
x=x)
e
t ()) 1
→∗
→∗ (λt. Scanλx.
e0
t=t ())
1
For the next counterexamples, we will suppose a simplistic database capabilities description,
for which all operators are compatible as long as there is no free variable in their configuration
(such a database would have the stable compatibility property).
Non-normalizing expressions Consider expression e from Example 8. Obviously, it is nonnormalizing because of the Ω = (λx. x x) (λx. x x) in the configuration. Since the heuristic
applies a call-by-name reduction strategy on the configurations once all free variables are inlined,
it will consume all the fuel on Ω and never consider e0 , which is the only element of MinReds (e).
Example 8.
e
Scan((λx.
x x) (λx. x x)) ((λx. 1) y) ()
→∗
→∗ Scan((λx.
x x) (λx. x
e0
x)) 1 ()
Operator erasure Consider expression e from Example 9. Obviously, the Scanx () operator
can be erased. Since the heuristic tries to inline variables in configurations, reduce configurations
then regroup fragments, it will never consider e0 , which is the only element of MinReds (e).
Example 9.
e
if false then Scanx () else Scandb.table ()
→∗
→∗ Scandb.table ()
17
e0
Operator duplication Consider expression e from Example 10. Obviously, the Scanz operator
can be duplicated. The heuristic will try to inline y in the configuration of the two Join operators,
which requires to inline x first. Since this two-step reduction decreases the measure and because
the heuristic chooses the leftmost leaf of the configuration search space, e0 will never be considered
although it is the only element of MinReds (e).
Example 10.
e
(λx. λy. Joinif false then y else true (
Scandb.table (), Joinif false then y else
)) Scanz () false
E
H
SN
P
FO
Figure 5: Situation summary
→∗
e0
(λx. λy. Jointrue (
∗ Scan
(x,
x)
→
db.table (), Jointrue (x, x)
true
)) Scanz () false
Figure 5 sums up the situation. In this Venn diagram, E
stands for the set of all QIR expressions, SN for the set
of strongly-normalizing QIR expressions, F O for the set
of QIR expression with fixed operators, H for the set of
QIR expressions on which the heuristic returns an optimal
result and P for the set of QIR expressions for which I
proved that the heuristic returns such a result. As I will
show in Section 6, H is larger than P : in fact the heuristic
returns optimal results on all our real-world use cases.
In Appendix E, I discuss how the fuel value Φ can be estimated by analyzing the QIR code.
Though, an implementation of a module doing such an estimation is future work, and for now, Φ
has to be set by the developer.
6
Experiments
In this section, I present real-world use cases and show how the QIR evaluation strategies
discussed in Section 5 perform on these examples. In the following experiments, the target
database language is SQL and for conciseness, I never show the application code but rather
quickly describe the applications themselves.
The provided timings have been measured on my personal laptop (Core i5 M450 @2.4GHz,
8Go RAM). Although my setup is not precise enough for fine-grained performance comparison,
it enables an order of magnitude analysis and delivers interesting results.
6.1
Example of code factorization: analytics queries
Consider the following analytics query written in SQL and freely inspired by the TPC-H
benchmark.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
SELECT
l _ r e t u r n f l a g AS r e t u r n _ f l a g , l _ l i n e s t a t u s AS l i n e _ s t a t u s ,
SUM( l _ e x t e n d e d _ p r i c e ) AS sum_base_price ,
SUM( l _ e x t e n d e d _ p r i c e ∗ ( 1 − l _ d i s c o u n t ) ) AS sum_disc_price ,
SUM( l _ e x t e n d e d _ p r i c e ∗ ( 1 − l _ d i s c o u n t ) ∗ ( 1 + l_tax ) ) AS sum_charge ,
SUM( l _ e x t e n d e d _ p r i c e ) ∗ 0 . 7 5 AS sum_real_cost ,
SUM( l _ e x t e n d e d _ p r i c e ) ∗ 0 . 2 5 AS sum_margin ,
AVG( l _ e x t e n d e d _ p r i c e ) AS avg_base_price ,
AVG( l _ e x t e n d e d _ p r i c e ∗ ( 1 − l _ d i s c o u n t ) ) AS avg_disc_price ,
AVG( l _ e x t e n d e d _ p r i c e ∗ ( 1 − l _ d i s c o u n t ) ∗ ( 1 + l_tax ) ) AS avg_charge ,
AVG( l _ e x t e n d e d _ p r i c e ) ∗ 0 . 7 5 AS avg_real_cost ,
AVG( l _ e x t e n d e d _ p r i c e ) ∗ 0 . 2 5 AS avg_margin
FROM db . l i n e i t e m
GROUP BY l _ r e t u r n _ f l a g , l _ l i n e s t a t u s
ORDER BY l _ r e t u r n _ f l a g , l _ l i n e s t a t u s
18
Notice in particular the many common expressions that are used to compose the aggregation
functions of the query. To factorize code and increase the maintainability of the code, the
application developer might want to define these expressions only once and reuse them throughout
the query. For better readability, he might also store subexpressions in variables instead of
writing the query in one block. The direct translation of this kind of code to QIR would lead to
the following query representation (remember that in QIR, as opposed to SQL’s strange syntax,
aggregations functions belong to the configuration of the Group operator).
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
/∗ C o n s t r u c t s t h e l i s t o f p r o j e c t e d a t t r i b u t e s ∗/
let project_list = λtup.
let real_cost = 0.75 in
let margin = 0.25 in
let return_f lag = tdestr tup "l_returnflag" in
let line_status = tdestr tup "l_linestatus" in
let sum_base_price = tdestr tup "sum_base_price" in
let sum_disc_price = tdestr tup "sum_disc_price" in
let sum_charge = tdestr tup "sum_charge" in
let avg_base_price = tdestr tup "avg_base_price" in
let avg_disc_price = tdestr tup "avg_disc_price" in
let avg_charge = tdestr tup "avg_charge" in
tcons "return_flag" return_f lag (
tcons "line_status" line_status (
tcons "sum_base_price" sum_base_price (
tcons "sum_disc_price" sum_disc_price (
tcons "sum_charge" sum_charge (
tcons "sum_real_cost" (sum_base_price ∗ real_cost) (
tcons "sum_margin" (sum_base_price ∗ margin) (
tcons "avg_base_price" avg_base_price (
tcons "avg_disc_price" avg_disc_price (
tcons "avg_charge" avg_charge (
tcons "avg_real_cost" (avg_base_price ∗ real_cost) (
tcons "avg_margin" (avg_base_price ∗ margin) (
tnil)))))))))))) in
/∗ C o n s t r u c t s t h e l i s t o f g r o u p i n g / s o r t i n g a t t r i b u t e s ∗/
let group_sort_attr = λtup.
cons (tdestr tup "l_returnflag") (
cons (tdestr tup "l_linestatus") (
tnil)) in
/∗ C o n s t r u c t s t h e a g g r e g a t e f u n c t i o n s ∗/
let group_agg = λtup.
let extended_price = tdestr tup "l_extended_price" in
let discount = tdestr tup "l_discount" in
let tax = tdestr tup "l_tax" in
let disc_price = extended_price ∗ (1 − discount) in
let charge = disc_price ∗ (1 + tax) in
tcons "sum_base_price" (sum extended_price) (
tcons "sum_disc_price" (sum disc_price) (
tcons "sum_charge" (sum charge) (
tcons "avg_base_price" (avg extended_price) (
tcons "avg_disc_price" (avg disc_price) (
tcons "avg_charge" (avg charge) (
tnil)))))) in
/∗ Main q u e r y ∗/
Projectproject_list (Sortgroup_sort_attr (Groupgroup_sort_attr,group_agg (Scandb.lineitem ())))
On this example, my implementation of the exhaustive reduction strategy does not terminate
in reasonable time due to a combinatorial explosion. Nevertheless, it eventually finds the reduced
expression with minimal measure, i.e the original expression with all definitions and common
19
expressions inlined. On the other hand, my implementation of the heuristic-based strategy quickly
finds the same result for a fuel value Φ ≥ 11 (timings are given in Table 1). Furthermore, the
reduced expression is a direct translation of the SQL query above, which represents an argument
in favor of the measure described in Section 5.2 and used by both reduction strategies.
6.2
Example of fragment grouping: dynamic queries
Consider now a simple website on which users can read small ads from people selling furniture,
cars, etc. and in particular a page to browse through the offers. This page would consist mainly
in a form with (i) a dropdown menu to optionally sort ads by date or price, (ii) a set of checkboxes
to filter by category and (iii) an integer field with a default value to specify the number of
results to display on the page. The corresponding query, fetching the results, would have to be
built dynamically depending on the presence of filters and ordering. The following QIR code
is a possible query representation of this logic, in which we assume that variables is_sorted,
sort_attr, cat_list and limit are provided by the application context and encode the presence
of a sorting attribute, the list of selected categories and the number of results to display.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
/∗ R e c u r s i v e l y c o n s t r u c t s a l i s t o f OR o f t h e s e l e c t e d c a t e g o r i e s ∗/
let make_cat_f ilter = λcat_list. λtup.
let rec aux = λf ilter. λcat_list.
destr cat_list f ilter (λhd. λtl. aux ((tdestr tup "category" = hd) or f ilter) tl) in
let aux2 = λcat_list.
destr cat_list nil (λhd. λtl. aux (tdestr tup "category" = hd) tl) in
aux2 cat_list in
/∗ C o n s t r u c t s t h e o r d e r i n g a t t r i b u t e s ∗/
let make_order = λattr. λtup.
if attr = "price" then cons (tdestr tup "price") nil
else if attr = "date" then cons (tdestr tup "timestamp") nil
else nil in
/∗ C o n s t r u c t s t h e l i s t o f p r o j e c t e d a t t r i b u t e s ∗/
let project_list = λtup.
tcons "title" (tdestr tup "title") (
tcons "description" (tdestr tup "description") (
tnil)) in
/∗ Base t a b l e ∗/
let ads = Scandb.ads () in
/∗ A f t e r c a t e g o r y f i l t e r s ∗/
let ads_f iltered = destr cat_list ads (λhd. λtl. Selectmake_cat_f ilter
/∗ A f t e r ( o p t i o n a l ) o r d e r i n g ∗/
let ads_ordered = if is_sorted then Sortmake_order
cat_list (ads))
sort_attr (ads_f iltered)
in
else ads_f iltered in
/∗ Main q u e r y ∗/
Projectproject_list (Limitlimit (ads_ordered))
On this example, my implementation of the exhaustive reduction strategy does not terminate,
since one can obtain infinitely many distinct reductions of the original expression by unfolding the
recursive call of aux in make_cat_filter. Conversely, my implementation of the heuristic-based
strategy quickly finds the result expression with minimal measure (timings are given in Table 1).
For instance, with cat_list set to nil, is_sorted set to false and limit set to 20 the
result found with Φ ≥ 15 is
Projectλt. tcons "title" (tdestr t "title") (tcons "description" (tdestr t "description") tnil) (
Limit20 (Scandb.ads ()))
20
And with cat_list set to cons "housing" (cons "cars" nil), is_sorted set to true,
sort_attr set to date and limit set to 30 the result found with Φ ≥ 24 is
Projectλt. tcons "title" (tdestr t "title") (tcons "description" (tdestr t "description") tnil) (
Limit30 (Sortλt. cons (tdestr t "timestamp") nil (
Selectλt. tdestr t "category"="cars" or tdestr t "category"="housing" (Scandb.ads ()))))
Again, the result expression corresponds clearly to a SQL query, and this is a second example
of the effectiveness of the measure described in Section 5.2.
6.3
Example of incompatibility: caching
Following on the previous example, consider a page on which an admin can detect if an unexperienced user has published the same ad twice. Assume there is a function unexperienced(user_id)
outside the query boundary, telling if a user is unexperienced. This function would be translated
into a Truffle dependency, and represented in the QIR code by a Truffle reference truffle<0>.
The following QIR code could correspond to the query used by the page.
1
2
3
4
5
6
7
8
9
10
11
12
13
/∗ C o n s t r u c t s t h e l i s t o f p r o j e c t e d a t t r i b u t e s ∗/
let project_list = λtup. tcons "user_id" (tdestr tup "user_id") tnil in
/∗ C o n s t r u c t s t h e j o i n c o n d i t i o n ∗/
let join_cond = λtup1. λtup2.
(tdestr tup1 "title" = tdestr tup2 "title") and
not (tdestr tup1 "ad_id" = tdestr tup2 "ad_id") in
/∗ U n e x p e r i e n c e d u s e r s ∗/
let ads_unex_users = Selectλtup.
truffle<0> (tdestr tup "user_id") (Scandb.ads ())
in
/∗ Main q u e r y ∗/
Projectproject_list (Joinjoin_cond (ads_unex_users, ads_unex_users))
This is one example of situation where reducing a redex (e.g., inlining ads_unex_users) is
not beneficial for the database. In this case, both the implementations of the exhaustive strategy
and of the heuristic-based strategy (for Φ ≥ 1) quickly return the correct answer (timings are
given in Table 1), that is
let ads_f rom_unex_users = Selectλtup. truffle<0> (tdestr tup "user_id") (Scandb.ads ()) in
Projecttcons "user_id" (tdestr tup "user_id") tnil (
Join(tdestr tup1 "title"=tdestr tup2 "title") and not (tdestr tup1 "ad_id"=tdestr tup2 "ad_id") (
ads_f rom_unex_users, ads_f rom_unex_users))
This corresponds to a computation where the Truffle runtime embedded in the database
evaluates once ads_unex_users, stores the result on the database storage then passes the main
query to the database engine. Once again this is an example of the effectiveness of the measure
described in Section 5.2.
Experiment
Red. strategy
Timing
Optimal result
Section 6.1
Exh. Heur.
56.63s 6ms
X
X
Section 6.2 (1)
Exh.
Heur.
7ms
X
Section 6.2 (2)
Exh.
Heur.
7ms
X
Section 6.3
Exh. Heur.
5ms
4ms
X
X
Table 1: Timings of experiments from Sections 6.1, 6.2 and 6.3
7
Related work
As stated in the introduction, traditional data-oriented application frameworks, such as PHP with
the MySQL extension or Java with the JDBC driver, construct queries in the application code
21
using string concatenation or statements, with the aforementioned impedence mismatch issue.
Object-relational mappings (ORM), such as criteria queries in Hibernate and ActiveRecord in
Ruby on Rails, offer a syntactic abstraction to represent querying primitives with constructs of
the application language, and provide features to automatically handle translations from the
application language data model to the database data model an vice versa.
Although they solve part of the problem, these solutions restrict the application code sent to
the database to the provided querying primitives. This results in numerous application-database
roundtrips when a function of the application code (a UDF) is called from inside a query, and
forms a well-known performance bottleneck[8][7]. A recent contribution[5] focuses on analyzing
the application code surrounding querying primitives to translate it into the database language
when possible, and inline it in the query. But this detection is based on a list of handcrafted
patterns in which the choices of application language and database are hardcoded assumptions.
More advanced solutions include the Forward project[6] (on which I was working last year)
and Microsoft LINQ[12]. Forward offers as application language SQL++, a declarative extension of SQL tailored for web programming, and evaluates with an in-memory query processor
language features not supported by the database. Although the similarities between database
and application languages in terms of data models and querying primitives enable a seamless
integration of queries in the application language, this framework fails to push to the database
application code snippets containing unsupported features, and therefore does not fully solves
the UDF performance problem. Compared to this contribution, the QIR framework offers the
choice of the application language, a clean isolation of this application language from the rest of
the framework, formal semantics for operators and expressions, allowing a precise description of
how queries integrated in the application code are detected, translated and optimized, as well as
the possibility to evaluate any application code in the database, regardless of its capabilities.
LINQ integrates querying primitives for Microsoft SQL server in the application languages
part of the .NET framework (e.g., C#, F#, Visual Basic). Similarly to ORMs, it provides
automatic mechanisms to translate from the application language data model to the database
data model an the other way around. Since LINQ has no released specification, we rely on
the formal description and optimality proofs of T-LINQ[3] to compare our contribution. This
approach adds to the application language a small language of expressions similar to QIR’s
(variables, lambdas, applications, list and tuples) as well as a quotation/antiquotation system (to
define the boundary between application code to evaluate in the application runtime and in the
database query processor). Querying primitives correspond to patterns in application code (e.g.,
a Scan corresponds to a for loop and a Select to an if conditional. The translation to SQL
includes a rewriting phase corresponding to successive iterations, on the application code AST,
of the β-reduction and ad-hoc rules on these patterns. Although T-LINQ successfully gathers
application code around the querying primitives and reduces it to produce efficient queries, it can
only rely on the SQL capabilities of the database (i.e., there is no equivalent of the Truffle runtime
inside the database). Thus, this framework only targets a limited number of application language
and a single database, and conflates the three following issues: detection of the query boundary,
optimization of the query representation and translation of QIR into the database language. This
fact is particularly observable in the restrictions enforced on expressions inside quotes (i.e., to be
sent to the database): quoted expressions have to be reductible to expressions supported by the
database. Compared to this contribution, the QIR framework offers a clean separation between
application language, query representation and database language. Moreover, it does not impose
restrictions on application code that can be send to the database. One notices that, although
they start from different architecture assumptions and use different proof ideas, Theorem 6
ended up requiring hypotheses similar to the restrictions necessary to the optimality proof of
T-LINQ[3], hinting that such assumptions are reasonable. Nevertheless, while T-LINQ cannot
handle expressions outside this scope (indeed, no example from Section 6 can be reproduced in
T-LINQ), the QIR heuristic can still produce (good) results on them, thanks to the embedded
Truffle runtime. Looking at T-LINQ examples in [3], one can easily spot the differences between
22
the two frameworks: examples 1-4 and 6-7 can easily be reproduced in QIR and similar examples
have already been given in this document, examples 5 and 8-9 contain nesting and unnesting
operators and use ad-hoc reduction rules to rewrite them into a flat relational algebra (which we
do not want to engage into, as a design choice), and examples 10-13 have to use the application
language for features such as recursion, whereas QIR offers them natively.
8
Conclusion and future work
The ability to evaluate snippets from the application code inside the database opened new
perspectives on the language-intergrated query problems. We introduced Query Intermediate
Representation (QIR), an abstraction layer between application language and database language,
that can be used to represent queries. QIR comes with a rewriting strategy transforming queries
written in a natural way in the application code into queries heavily optimizable by the database.
Such rewritings are guided by a notion of “good” queries abstracted away in a measure on QIR
expressions. Moreover, we produced formal proofs and experiments based on our QIR prototype
showing that the QIR framework causes a minimal overhead while enabling a considerable
speedup of query evaluation.
Nevertheless, as explained in Section 3, we made the assumption that a same construct has
identical semantics in the applicaton language, QIR and the database language (e.g., a tuple
navigation user.name returns the same result in JavaScript, QIR and SQL), and we allow a lossy
translation from a language to another when this is not the case (in fact, exisiting frameworks
have the same limitation). This issue, which extends to discrepancies in the data model (e.g.,
JavaScript integers and Cassandra integers might be different), could be fixed by introducing
more abstraction in the QIR data model. On a different matter, the rewriting heuristic presented
in Section 5 uses a fuel value that has to be set by the developer. This obligation could become a
simple option if we provide a module “guessing” what fuel value to set, using the estimation logic
discussed at the end of Section 5 and described in Appendix E. These ideas are future work. We
also plan to try our framework with more combinations of application languages and databases,
then measure the impact of such an architecture on real-world large-scale web applications.
23
Appendices
A
References
[1] Malcolm P. Atkinson and O. Peter Buneman. Types and persistence in database programming
languages. ACM Comput. Surv., 19(2):105–170, June 1987.
[2] Henk P Barendregt. The Lambda Calculus: Its Syntax and Semantics, revised ed., vol. 103
of Studies in Logic and the Foundations of Mathematics. North-Holland, Amsterdam, 1984.
[3] James Cheney, Sam Lindley, and Philip Wadler. A practical theory of language-integrated
query. SIGPLAN Not., 48(9):403–416, September 2013.
[4] James Cheney, Sam Lindley, and Philip Wadler. Language-integrated query using comprehension syntax: state of the art, open problems, and work in progress. report, University of
Edinburg, 2014.
[5] Alvin Cheung, Samuel Madden, Armando Solar-Lezama, Owen Arden, and Andrew C
Myers. Using program analysis to improve database applications. IEEE Data Eng. Bull.,
37(1):48–59, 2014.
[6] Yupeng Fu, Kian Win Ong, and Yannis Papakonstantinou. Declarative ajax web applications
through SQL++ on a unified application state. In Proceedings of the 14th International
Symposium on Database Programming Languages (DBPL 2013), August 30, 2013, Riva del
Garda, Trento, Italy., 2013.
[7] Yupeng Fu, Kian Win Ong, Yannis Papakonstantinou, and Romain Vernoux. Holistic Data
Access Optimization for Analytics Reports. report, Department of Computer Science, UC
San Diego, 2013.
[8] George Giorgidze, Torsten Grust, Tom Schreiber, and Jeroen Weijers. Haskell boards the
ferry. In Implementation and Application of Functional Languages, pages 1–18. Springer,
2011.
[9] Delia Kesner, Carlos Lombardi, and Alejandro Ríos. A standardisation proof for algebraic
pattern calculi. In Proceedings 5th International Workshop on Higher-Order Rewriting, HOR
2010, Edinburgh, UK, July 14, 2010., pages 58–72, 2010.
[10] Jan Willem Klop, Vincent van Oostrom, and Roel C. de Vrijer. Lambda calculus with
patterns. Theor. Comput. Sci., 398(1-3):16–31, 2008.
[11] Avinash Lakshman and Prashant Malik. Cassandra: a decentralized structured storage
system. Operating Systems Review, 44(2):35–40, 2010.
[12] Erik Meijer, Brian Beckman, and Gavin Bierman. Linq: Reconciling object, relations
and xml in the .net framework. In Proceedings of the 2006 ACM SIGMOD International
Conference on Management of Data, SIGMOD ’06, pages 706–706, New York, NY, USA,
2006. ACM.
[13] Kian Win Ong, Yannis Papakonstantinou, and Romain Vernoux. The SQL++ semistructured data model and query language: A capabilities survey of sql-on-hadoop, nosql
and newsql databases. CoRR, abs/1405.3631, 2014.
[14] Oracle. Graal/truffle overview. http://www.oracle.com/technetwork/oracle-labs/
program-languages/overview/", accessed July 21, 2015.
[15] Oracle Database. https://www.oracle.com/database/index.html.
24
[16] Ashish Thusoo, Joydeep Sen Sarma, Namit Jain, Zheng Shao, Prasad Chakka, Ning Zhang,
Suresh Anthony, Hao Liu, and Raghotham Murthy. Hive - a petabyte scale data warehouse
using Hadoop. In ICDE, pages 996–1005, 2010.
B
Internship timeline
The Gantt diagram of Figure 6 describes the use of my time during this internship. In addition
to the elements presented in this report, I had the opportunity to present my work at Oracle
Labs in July. It was a very interesting experience, a good presentation exercice and a chance to
organize my results before writing this report.
Weeks
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19 20
Introduction to Truffle
Related work
QIR calculus
QIR reduction strategies
Prototype implementation
Proofs
Talk at Oracle Labs
Internship report
Weekly meetings
Figure 6: Internship timeline
C
Formal definition of occurences
Definitions 33 and 34 correspond to a formalization of the concept of occurences presented quickly
in Definition 26.
Definition 33. We define the language QIR* as the language of QIR constructs (cf. Section 4)
in which integer identifiers can be attached to each subexpression and reductions are extended
accordingly. Non-exhaustively, this means that
id
• QIR* operators contain Scane1 (), Scanid
e1 (), Selecte1 (e2 ), Selecte1 (e2 ), etc.
• QIR* expressions contain x, xid , λx. e1 , λid x. e1 , (e1 e2 ), (e1 e2 )id , true, trueid etc.
• xid {e1 /x} = e1 , x{e1 /x} = e1
0
0
• (λid x. e1 ) e2 → e1 {e2 /x}, destrid (consid e1 e2 ) e3 e4 → e4 e1 e2 , trueid andid e1 →
true, etc.
where id, id0 are integer identifiers and the ei are QIR* expressions.
Let e be a QIR expression and e1 a subexpression of e. We denote by ?(e) the QIR* expression
in which a unique identifier is attached to each subexpression of e. There is a one-to-one
correspondance between the subexpressions of e and the subexpressions of ?(e), thus we will also
denote by ?sub (e1 ) the subexpression of ?(e) corresponding to e1 .
Let e0 be a QIR* expression and e01 a subexpression of e0 . We denote by ?−1(e0 ) the QIR expression
in which all identifers are removed from e0 . There is a one-to-one correspondance between the
0
subexpressions of e0 and the subexpressions of ?−1(e0 ), thus we will also denote by ?−1
sub(e1 ) the
subexpression of ?−1(e0 ) corresponding to e01 .
Let e0 be a QIR* expression and e01 a subexpression of e0 . We denote by ExprId (e0 , e01 ) the identifier
of e01 in e0 , if it has one (it is not defined otherwise).
25
Lemma 10. Let e0 be a QIR expression and e00 a QIR* expression such that ?−1(e00 ) = e0 .
Any QIR reduction chain e0 →r1 . . . →rn en can be simulated by the QIR* reduction chain
0
e00 →r10 . . . →rn0 e0n such that, for all 1 ≤ i ≤ n, ?−1(e0i ) = ei and ?−1
sub(ri ) = ri .
Definition 34. Let e0 be a QIR expression, e a subexpression of e0 and en a QIR expression such
that e0 →∗ en . Using Lemma 10 and the fact that ?−1(?(e0 )) = e0 , we know that this QIR reduction chain can be simulated by a QIR* reduction chain e00 →∗ e0n starting from e00 = ?(e0 ). This
0
0
0
0
0
way, we define Occurences (en , e) as the set {?−1
sub(e ) | e is a subexpression of en , ExprId (en , e ) =
ExprId (?(e0 ) , ?sub (e))}.
For instance, the QIR reduction
(λx. λy. f x x) z z → (λy. f z z) z → f z z
can be simulated by the QIR* reduction
(λ1 x. λ2 y. f 3 x4 x5 ) z 6 z 7 → (λ2 y. f 3 z 6 z 6 ) z 7 → f 3 z 6 z 6 .
One can see, just by looking at identifiers, that the first z variable produced both z variables
of the final expression: it has two occurences in the final expression. The other one disappeared:
it has no occurence in the final expression.
D
Proofs
Theorem 1. QIR with the reduction rules of Definition 2 satisfies the Church-Rosser property.
Proof. Encoding (i) integer (resp. boolean, string) values and functions in the lambda-calculus
(Church encoding), (ii) operators, built-in functions, Truffle nodes, Data references, nil, cons,
tnil and tcons with constructors and (iii) destr and tdestr with (linear) pattern-matching,
QIR can be transposed in a lambda-calculus with patterns[10] which satisfies the Church-Rosser
property and for which the β, δ and ρ reduction rules of Definition 2 are compatible.
Theorem 2. QIR with the reduction rules of Definition 2 verifies the Standardization theorem,
that is, if an expression e has a normal form, it can be obtained from e by reducing successively
its leftmost outermost redex.
Proof. With the same encoding as in the proof of Theorem 1, QIR can be transposed in a
lambda-calculus with patterns[9] in which the Standardization theorem holds and for which the
β, δ and ρ reduction rules of Definition 2 are compatible.
Lemma 1. The measure M of Definition 8 induces a well-founded order on expressions.
Proof. The order induced by M is a lexicographical order on the natural order of positive integers,
which is well-founded, and the lexicographical order preserves well-foundness.
Lemma 2. ∀e, MinReds (e) 6= ∅.
Proof. Since e ∈ Red0 (e) ⊆ Reds (e), we know that Reds (e) =
6 ∅. Moreover, the order induced
by M is well-founded (Lemma 1) therefore M has a minimum on Reds (e) and MinReds (e) is
non-empty.
Lemma 3. For all expression e the following properties hold.
• For a variable e0 in e, HInlineVar (e, e0 ) returns None or contracts a redex.
• For a subexpression e0 in e, HMakeRedex (e, e0 ) returns None or contracts a redex.
• For a subexpression e0 in e, HContractLam (e, e0 ) returns None or contracts a redex.
Proof. The proof follows an induction on TraversalPos (e, e0 ).
• If e0 is free in e then HInlineVar (e, e0 ) returns None. Otherwise, let l be the lambda binding
e0 . HInlineVar (e, e0 ) corresponds to HContractLam (e, l), and since TraversalPos (e, l) <
TraversalPos (e, e0 ), the induction hypothesis yields the result.
26
• If e0 is a redex, HMakeRedex (e, e0 ) contracts this redex. If e0 = (e00 e1 ), e0 = destr e00 e1 e2 ,
e0 = tdestr e00 s or e0 = if e00 then e1 else e2 (and similarly for other ρ-redexes),
HMakeRedex (e, e0 ) corresponds to HMakeRedex (e, e00 ), and since TraversalPos (e, e00 ) <
TraversalPos (e, e0 ) and the induction hypothesis yields the result. If e0 is a variable,
HMakeRedex (e, e0 ) corresponds to HInlineVar (e, e0 ) and we proved the result above. Otherwise, HMakeRedex (e, e0 ) returns None.
• We denote by e00 the parent expression of e0 . If e00 = (e0 e1 ), e00 = destr e0 e1 e2 ,
e00 = tdestr e0 s or e00 = if e0 then e1 else e2 (and similarly for other ρ-redexes),
HContractLam (e, e0 ) corresponds to HMakeRedex (e, e00 ) which either contracts e00 and we
are done, or corresponds to HMakeRedex (e, e0 ) and we proved the result above. If e00 =
(e1 e0 ), e00 = destr e1 e0 e2 , e00 = destr e1 e2 e0 , e00 = if e1 then e0 else e2 or e00 =
if e1 then e2 else e0 (and similarly for other ρ-redexes), HContractLam (e, e0 ) corresponds
to HMakeRedex (e, e1 ) and since TraversalPos (e, e1 ) < TraversalPos (e, e0 ), the induction
hypothesis yields the result. If e00 = λx. e0 , e00 = cons e0 e1 , e00 = cons e1 e0 , e00 =
tcons s e0 e1 , e00 = tcons s e1 e0 , HContractLam (e, e0 ) corresponds to HContractLam (e, e00 )
and since TraversalPos (e, e00 ) < TraversalPos (e, e0 ), the induction hypothesis yields the
result. Otherwise, HContractLam (e, e0 ) returns None.
Lemma 4. For all expression e, the search space HConfigSSΦ (e) (resp. HChildrenSSΦ (e)) has
bounded size.
Proof. Each recursive call to HRedConfigφ (e, C[]) (resp. HRedChildφ (e, C[])) is well defined
(corollary of Lemma 3) and is such that the pair (M (e), φ) decreases. Using lemma 1 it is easy to
prove that the lexicographical order induced by this pair is also well-founded, therefore the search
space has bounded depth. Moreover, any expression has a bounded number of operators and
operator contexts therefore each node in the search space has a bounded number of children.
Lemma 10. Let e0 be a QIR expression and e00 a QIR* expression such that ?−1(e00 ) = e0 .
Any QIR reduction chain e0 →r1 . . . →rn en can be simulated by the QIR* reduction chain
0
e00 →r10 . . . →rn0 e0n such that, for all 1 ≤ i ≤ n, ?−1(e0i ) = ei and ?−1
sub(ri ) = ri .
Proof. The proof follows an induction on n. For n = 0, the result is trivial. For n > 0, contracting
0
0
the redex r10 such that ?−1
sub(r1 ) = r1 yields a QIR* expression e1 . It is easy to prove by case
−1
0
analysis on r1 that ? (e1 ) = e1 . By induction hypothesis, the remaining of the reduction chain
e1 →∗ en can be simulated by e01 →∗ e0n , which means that the entire chain e0 →∗ en can be
simulated by e00 →∗ e0n .
Lemma 5. Let e be an expression, e0 a subexpression of e and r a redex in e. If e →r e00 and
Occurences (e00 , e0 ) = ∅, then r either consumes or erases e0 .
Proof. By case analysis on r. All cases are similar therefore we will only write the proof for
r = (λx.e1 ) e2 . By hypothesis, there is a context C[] such that e = C[r]. Then we do a case
analysis on the position of e0 in e. We know that e0 cannot be a subexpression of C, otherwise
Occurences (e00 , e0 ) 6= ∅. For the same reason, we also know that e0 is not a subexpression of
e1 other than the free variable x. If e0 = r, if e0 = λx.e1 or if e0 is the variable x free in
e1 , then e is consumed. Otherwise, e0 is a subexpression of e2 and x is not free in e1 (since
Occurences (e00 , e0 ) = ∅), and in this case e0 is erased.
Lemma 6. Let e be an expression and e0 a subexpression of e. The following properties hold.
• If e0 is a variable v and if HInlineVar (e, v) returns None then e0 cannot be consumed.
• If HMakeRedex (e, e0 ) returns None then e0 cannot be consumed.
• If HContractLam (e, e0 ) returns None then e0 cannot be consumed.
Proof. The proof follows an induction on TraversalPos (e, e0 ).
27
• Suppose that HInlineVar (e, v) returns None. If v is free in e then by definition it cannot be consumed. Otherwise, denoting by l the lambda binding v, this means that
HContractLam (e, l) returns None. Since TraversalPos (e, l) < TraversalPos (e, v), the induction hypothesis tells that l cannot be consumed and (by definition of the consumption of a
variable) neither can e0 .
• Suppose that HMakeRedex (e, e0 ) returns None. If e0 is a variable, this implies that
HInlineVar (e, e0 ) returns None and we proved the result above. If e0 is a redex, this
is absurd (HMakeRedex (e, e0 ) cannot return None). If e0 = (e00 e1 ), e0 = destr e00 e1 e2 ,
e0 = tdestr e00 s or e0 = if e00 then e1 else e2 (and similarly for other ρ-redexes), this means
that HMakeRedex (e, e00 ) returns None and since TraversalPos (e, e00 ) < TraversalPos (e, e0 )
the induction hypothesis tells that e00 cannot be consumed and (by definition of the consumption of the considered expressions for e0 ) neither can e0 . If e0 = λx. e1 , e0 = cons e1 e2 ,
e0 = tcons s e1 e2 , e0 = nil, e0 = tnil or if e0 is a constant (integer, boolean, etc.), by
looking at all the call sites of HMakeRedex (., .) we know that the parent of e0 is such that
e0 is in an ill-formed term (e.g., destr tnil e1 e2 ) and therefore cannot be consumed. The
remaining possible expressions for e0 (e.g., Truffle nodes, Data references, operators, etc.)
cannot be consumed in any expression.
• Suppose that HContractLam (e, e0 ) returns None. We denote by e00 the parent expression
of e0 . If e00 = (e0 e1 ), e00 = destr e0 e1 e2 , e00 = tdestr e0 s or e00 = if e0 then e1 else e2
(and similarly for other ρ-redexes), this means that HMakeRedex (e, e00 ) returns None, which
in turn means that HMakeRedex (e, e0 ) returns None and we proved the result above. If
e00 = (e1 e0 ), e00 = destr e1 e0 e2 , e00 = destr e1 e2 e0 , e00 = if e1 then e0 else e2
or e00 = if e1 then e2 else e0 (and similarly for other ρ-redexes), this means that
HMakeRedex (e, e00 ) returns None, which in turn means that HMakeRedex (e, e1 ) returns
None. Since TraversalPos (e, e1 ) < TraversalPos (e, e0 ), the induction hypothesis tells that e1
cannot be consumed and (by definition of the consumption of the considered expressions
for e00 ) neither can e00 , which implies that e0 cannot be consumed either. If e00 = λx. e0 ,
e00 = cons e0 e1 , e00 = cons e1 e0 , e00 = tcons s e0 e1 , e00 = tcons s e1 e0 , this means that
HContractLam (e, e00 ) returns None and since TraversalPos (e, e00 ) < TraversalPos (e, e0 ), the
induction hypothesis tells that e00 cannot be consumed and (by definition of the consumption
of the considered expressions for e0 ) neither can e0 . Similarly to above, the remaining
expressions to consider for e00 either correspond to ill-formed expressions (that cannot be
consumed) or expressions that can never be consumed.
Lemma 7. Let e be an expression with fixed operators and r a redex in e. For a database with
stable compatibility, if e →r e0 then M (e0 ) ≤ M (e).
Proof. By case analysis on r. All cases are similar therefore we will only write the proof for
r = (λx.e1 ) e2 . By hypothesis, there is a context C[] such that e = C[r]. Since e has fixed
operators, there is a one-to-one correspondance between the operators of e and the operators
of e0 . For each operator op in e, denoting by op0 the corresponding operator in e0 , the stable
compatibility hypothesis tells us that if op is compatible, then op0 is also compatible. Since
no redex can create operators, this implies that Op (e0 ) − Comp (e0 ) ≤ Op (e) − Comp (e). The
only case to treat is when Op (e0 ) − Comp (e0 ) = Op (e) − Comp (e). Looking at the definition
of fragments (cf. Definition 6), we see that there is only three ways to increase the number of
fragments in e.
• Duplicate an existing fragment. This cannot happen under the fixed operator hypothesis,
since a fragment contains at least one operator.
• Create a new fragment by making an incompatible operator compatible. This cannot
happen either. Indeed, if r turns an incompatible operator into a compatible one, using
the stable compatibility hypothesis, we know that all other compatible operators in e are
still compatible in e0 which contradicts Op (e0 ) − Comp (e0 ) = Op (e) − Comp (e).
• Split an existing fragment into at least two fragments. This again, cannot happen. Indeed,
let F be a fragment in e = C[(λx.e1 ) e2 ]. By definition of a fragment, we only have to
consider the following cases:
28
– if F is a subexpression of e2 or C[], it is intact in e0 .
– if F is a subexpression of e1 , either x is not free in F and F is not affected by the
reduction or x is in the configuration of some operators of F and (stable compatibility)
these operators stay compatible after reduction and r cannot split F .
– if r is in the fragment, it is necessarily in a configuration of an operator of the fragment
which (stable compatibility) stays compatible after reduction, therefore r cannot split
F.
Lemma 8. Let e be a weakly-normalizing expression with fixed operators. For a database with
stable compatibility, the normal form of e has minimal measure.
Proof. Since e is weakly-normalizing, it has a normal form eN . Suppose, to the contrary, that
there exists e0 such that e →∗ e0 and M (e0 ) < M (eN ). Using Theorem 1 (confluence) and the
fact that eN is normal, we know there is a finite reduction chain e0 →∗ eN and applying Lemma
7 on each reduction of the chain leads to a contradiction.
Lemma 9. Let e be a weakly-normalizing expression with fixed operators, emin an expression
in MinReds (e) and e0 an expression such that e →∗ e0 and Op (e0 ) − Comp (e0 ) = Op (emin ) −
Comp (emin ). For a database with stable compatibility, an operator is compatible in emin if and
only if it is compatible in e0 .
Proof. Using the fixed operator hypothesis, we know that there is a one-to-one correspondance
between the operators of e and the operators in any reduced form of e. Therefore, Comp (e0 ) =
Comp (emin ).
Suppose, to the contrary, that there exists an operator op in e such that Occurences (e0 , op) = {op0 },
Occurences (emin , op) = {opmin }, op0 compatible and opmin not compatible. Using Theorem 1
(confluence), we know there is an expression e00 such that e0 →∗ e00 and emin →∗ e00 . Using the
stable compatibility hypothesis, op is compatible in e00 and all operators compatible in emin stay
compatible in e00 , which contradicts the minimality of the measure of emin .
Suppose now, to the contrary, that there exists an operator op in e such that Occurences (e0 , op) =
{op0 }, Occurences (emin , op) = {opmin }, op0 not compatible and opmin compatible. The minimality
of emin tells that all operators compatible in e00 are also compatible in emin and the stable
compatibility hypothesis tells that each operator compatible in e0 are still compatible in e00 , which
contradicts the fact that Comp (e0 ) = Comp (emin ).
Conjecture 1. For databases with stable compatibility, the reduction strategy of Definition 24 is
complete on weakly-normalizing expressions with fixed operators. That is, for a database with
stable compatibility, given a weakly-normalizing input expression e with fixed operators, there
exists a fuel value Φ such that HMinRedΦ (e) ∈ MinReds (e).
Looking at the proof of the theorem for strongly-normalizing expressions (Theorem 6), one
notices that the strongly-normalizing hypothesis itself is only used three times: (i) to prove that
iteratively calling HInlineVar (., .) on the free variables of an operator configuration terminates,
(ii) to use Theorem 2 on an operator configuration and (iii) to prove that iteratively calling
HMakeRedex (., .) on an operator child expression terminates. Case (ii) is not an issue, since
Theorem 2 only requires the expression to be weakly-normalizing. Case (iii) can be dealt with
using the fixed operator hypothesis: HMakeRedex (., .) contracts redex in the same order as the
call-by-name reduction strategy and none of these redexes can be erased (because this would
require to erase the parent operator), therefore Theorem 2 yields the result. The only remaining
issue is case (i): HInlineVar (., .) reduces redexes in an order that cannot be compared with the
call-by-name reduction strategy, therefore Theorem 2 cannot be used here. However, the intuition
is that operator configurations in which one can inline variables infinitely have to be inside some
kind of fixpoint, which is forbidden by the fixed operator hypothesis.
29
E
Estimation of the heuristic fuel value
For now, we assume that the fuel value (cf. Section 5.4) for the heuristic is empirically chosen
by the developer. Nevertheless, one can give a simple estimate with the following reasoning. In
usual QIR query representations, query fragments are not passed as function arguments and the
tree of operators is almost already built.
In this case, inlining a variable in a configuration is most of the time a single-step heuristic
reduction (e.g., the inlining of x in let x = . . . in C[opC 0 [x] ]) or a multi-step heuristic reduction
in which each step inlines a variable (e.g., the inlining of y in (λx. λy. C[opC 0 [x,y] ]) e1 e2 ). Thus,
denoting by V the maximum (on all operators) number of variables to inline in a configuration,
we can estimate that after V reduction steps all variables are inlined, and providing more fuel
would not inline more variables.
Furthermore, configuration are usually small pieces of code, that can reach a normal form in a
few reduction steps. The most expensive configurations to reduce are then expressions containing
recursions. Thus, denoting by R the maximum number of recursive calls and R0 the maximum
number of reductions used inside a single recursive call, we can estimate that after RR0 reduction
steps a configuration is in normal form, and providing more fuel would not enable to consider
more reduced expressions of this configuration.
Finally, children of operators are most of the time variables (e.g., let subquery = . . . in
C[op(subquery)]) or conditionals (e.g., C[op(if c then subquery1 else subquery2 )]) and both
can be reduced to an operator in a small number of heuristic reduction steps, denoted by C.
Thus, since QIR operators have at most two children, we can estimate that after 2C reduction
steps all the children of an operator are reduced to an operator, and providing more fuel would
not enable to regroup more fragments.
Therefore, given the two-pass behavior of the heuristic, we can estimate the fuel value required
for a comprehensive exploration of the search space as max(V + RR0 , 2C). Looking back at the
complex examples of Section 6, we can compare the aforementioned estimation (calculated by
hand) to the fuel value Φ effectively required to reach the optimal result.
Example
6.1
6.2 (1)
6.2 (2)
6.3
V
1
2
2
1
R
1
3
3
0
R0
10
8
8
0
C
0
2
2
1
max(V + RR0 , 2C)
11
26
26
1
Φ
11
24
15
1
Although using the Φ estimation would lead to the optimal result in all these cases, computing
Φ requires to estimate the values for V , R, R0 and C, which is far from easy to automate (for R
it corresponds to estimating in how many steps a recursive program terminates). Nevertheless,
the Φ ≈ max(V + RR0 , 2C) formula can be useful to the application developer when setting
Φ, since V , R, R0 and C correspond to characteristics indirectly observable in the application
language, as opposed to the Φ black-box.
F
Source code and other resources
The implementation source code and other resources mentioned in this document are available
for download at http://vernoux.fr/data/internship-mpri/. This folder contains
/
integration_survey.xlsx .... Notes on existing language-integrated query frameworks
qir_reduction. ............................................... Prototype source code
30
README.txt ............................................... Installation instructions
src...................................................................Source code
compat_sql.ml.....................................SQL capabilities description
compat_sql.mli....................................Interface of compat_sql.ml
lexer.mll...........................................Lexing of QIR expressions
parser.mly.........................................Parsing of QIR expressions
qir.ml.............................................QIR constructs and utilities
qir.mli....................................................Interface of qir.ml
qir_reduction.ml ............................ Measure and reduction strategies
qir_reduction.mli.............................Interface of qir_reduction.ml
reduce.ml..................Main program and command-line arguments parsing
utils.ml..............................................................Utilities
test.......................................................Test files and examples
ads_view.txt............................Input of the experiment of Section 6.2
analytics.txt...........................Input of the experiment of Section 6.1
caching.txt ............................. Input of the experiment of Section 6.3
linq_1.txt................................Reproduction of an example from [3]
linq_2.txt................................Reproduction of an example from [3]
linq_xpath.txt...........................Reproduction of an example from [3]
nesting_1.txt.......................................................Test case
testchild_1.txt .................................................... Test case
testchild_2.txt .................................................... Test case
testchild_3.txt .................................................... Test case
testconfig_1.txt ................................................... Test case
testconfig_2.txt ................................................... Test case
testconfig_3.txt ................................................... Test case
testnoinline_1.txt.................................................Test case
testnoinline_2.txt.................................................Test case
testtpch_1.txt......................................................Test case
31
| 6cs.PL
|
Logical Methods in Computer Science
Vol. 5 (2:12) 2009, pp. 1–30
www.lmcs-online.org
Submitted Apr. 2, 2006
Published May 28, 2009
SIMULATING REACHABILITY USING FIRST-ORDER LOGIC WITH
APPLICATIONS TO VERIFICATION OF LINKED DATA STRUCTURES ∗
TAL LEV-AMI a , NEIL IMMERMAN b , THOMAS W. REPS c , MOOLY SAGIV d , SIDDHARTH SRIVASTAVA e ,
AND GRETA YORSH f
a,d,f
b,e
c
School of Computer Science, Tel Aviv University
e-mail address: [email protected], {msagiv,gretay}@post.tau.ac.il
Department of Computer Science, University of Massachusetts, Amherst
e-mail address: {immerman,siddharth}@cs.umass.edu
Computer Science Department, University of Wisconsin, Madison
e-mail address: [email protected]
A BSTRACT. This paper shows how to harness existing theorem provers for first-order logic to automatically verify safety properties of imperative programs that perform dynamic storage allocation
and destructive updating of pointer-valued structure fields. One of the main obstacles is specifying
and proving the (absence) of reachability properties among dynamically allocated cells.
The main technical contributions are methods for simulating reachability in a conservative way
using first-order formulas—the formulas describe a superset of the set of program states that would
be specified if one had a precise way to express reachability. These methods are employed for semiautomatic program verification (i.e., using programmer-supplied loop invariants) on programs such
as mark-and-sweep garbage collection and destructive reversal of a singly linked list. (The mark-andsweep example has been previously reported as being beyond the capabilities of ESC/Java.)
1. I NTRODUCTION
This paper explores how to harness existing theorem provers for first-order logic to prove reachability properties of programs that manipulate dynamically allocated data structures. The approach
that we use involves simulating reachability in a conservative way using first-order formulas—i.e.,
the formulas describe a superset of the set of program states that would be specified if one had an
accurate way to express reachability.
1998 ACM Subject Classification: F.3.1, F.4.1, F.3.2.
Key words and phrases: First Order Logic, Transitive Closure, Approximation, Program Verification, Program
Analysis.
∗
A preliminary version of this paper appeared in Automated Deduction - CADE-20, 20th International Conference on
Automated Deduction, Tallinn, Estonia, July 22-27, 2005.
a
This research was supported by an Adams Fellowship through the Israel Academy of Sciences and Humanities.
b,e
Supported by NSF grants CCF-0514621,0541018,0830174.
c
Supported by ONR under contracts N00014-01-1-{0796,0708}.
f
Partially supported by the Israeli Academy of Science.
l
LOGICAL METHODS
IN COMPUTER SCIENCE
c
DOI:10.2168/LMCS-5 (2:12) 2009
CC
T. Lev-Ami, N. Immerman, T. Reps, M. Sagiv, S. Srivastava, and G. Yorsh
Creative Commons
2
T. LEV-AMI, N. IMMERMAN, T. REPS, M. SAGIV, S. SRIVASTAVA, AND G. YORSH
Automatically establishing safety and liveness properties of sequential and concurrent programs that permit dynamic storage allocation and low-level pointer manipulations is challenging.
Dynamic allocation causes the state space to be infinite; moreover, a program is permitted to mutate
a data structure by destructively updating pointer-valued fields of nodes. These features remain even
if a programming language has good capabilities for data abstraction. Abstract-datatype operations
are implemented using loops, procedure calls, and sequences of low-level pointer manipulations;
consequently, it is hard to prove that a data-structure invariant is reestablished once a sequence of
operations is finished [Hoa75]. In languages such as Java, concurrency poses yet another challenge:
establishing the absence of deadlock requires establishing the absence of any cycle of threads that
are waiting for locks held by other threads.
Reachability is crucial for reasoning about linked data structures. For instance, to establish
that a memory configuration contains no garbage elements, we must show that every element is
reachable from some program variable. Other cases where reachability is a useful notion include
• Specifying acyclicity of data-structure fragments, i.e., from every element reachable from node
n, one cannot reach n
• Specifying the effect of procedure calls when references are passed as arguments: only elements
that are reachable from a formal parameter can be modified
• Specifying the absence of deadlocks
• Specifying safety conditions that allow establishing that a data-structure traversal terminates, e.g.,
there is a path from a node to a sink-node of the data structure.
The verification of such properties presents a challenge. Even simple decidable fragments of firstorder logic become undecidable when reachability is added [GME99, IRR+ 04a]. Moreover, the
utility of monadic second-order logic on trees is rather limited because (i) many programs allow nontree data structures, (ii) expressing the postcondition of a procedure (which is essential for modular
reasoning) usually requires referring to the pre-state that holds before the procedure executes, and
thus cannot, in general, be expressed in monadic second-order logic on trees—even for procedures
that manipulate only singly-linked lists, such as the in-situ list-reversal program shown in Fig. 6,
and (iii) the complexity is prohibitive.
While our work was actually motivated by our experience using abstract interpretation – and,
in particular, the TVLA system [LAS00, SRW02, RSW04] – to establish properties of programs
that manipulate heap-allocated data structures, in this paper, we consider the problem of verifying
data-structure operations, assuming that we have user-supplied loop invariants. This is similar to
the approach taken in systems like ESC/Java [FLL+ 02], and Pale [MS01].
The contributions of the paper can be summarized as follows:
Handling FO(TC) formulas using FO theorem provers. We want to use first-order theorem
provers and we need to discuss the transitive closure of certain binary predicates, f . However,
first-order theorem provers cannot handle transitive closure. We solve this conundrum by adding
a new relation symbol ftc for each such f , together with first-order axioms that assure that ftc is
interpreted correctly. The theoretical details of how this is done are presented in Section 3. The
fact that we are able to handle transitive closure effectively and reasonably automatically is quite
surprising.
As explained in Section 3, the axioms that we add to control the behavior of the added predicates, ftc , must be sound but not necessarily complete. One way to think about this is that we are
simulating a formula, χ, in which transitive closure occurs, with a pure first-order formula χ′ . If our
axioms are not complete then we are allowing χ′ to denote more stores than χ does. The study of
methods that are sound but potentially incomplete is motivated by the fact that abstraction [CC77]
SIMULATING REACHABILITY USING FIRST-ORDER LOGIC
3
can be an aid in the verification of many properties. In terms of logic, abstraction corresponds to
using formulas that describe a superset of the set of program states that can actually arise. A definite
answer about whether a property always holds can sometimes be obtained even when information
has been lost because of abstraction.
If χ′ is proven valid in FO then χ is also valid in FO(TC); however, if we fail to prove that χ′ is
valid, it is still possible that χ is valid: the failure would be due to the incompleteness of the axioms,
or the lack of time or space for the theorem prover to complete the proof.
As we will see in Section 3, it is easy to write a sound axiom, T1 [f ], that is “complete” in the
very limited sense that every finite, acyclic model satisfying T1 [f ] must interpret ftc as the reflexive,
transitive closure of its interpretation of f . However, in practice this is not worth much because, as
is well-known, finiteness is not expressible in first-order logic. Thus, the properties that we want to
prove do not follow from T1 [f ]. We do prove that T1 [f ] is complete for positive transitive-closure
properties (Proposition 3.2). The real difficulty lies in proving properties involving the negation of
ftc , i.e., that a certain f -path does not exist.
Induction axiom scheme. To solve the above problem, we add an induction axiom scheme.
Although in general, there is no complete, recursively-enumerable axiomatization of transitive closure (Proposition 4.1), we have found, on the practical side, that on the examples we have tried, T1
plus induction allows us to automatically prove all of our desired properties. On the theoretical side,
we prove that our axiomatization is complete for word models (Theorem 4.8).
We think of the axioms that we use as aides for the first-order theorem prover that we employ
(S PASS [WGR96]) to prove the properties in question. Rather than giving S PASS many instances of
the induction scheme, our experience is that it finds the proof faster if we give it several axioms that
are simpler to use than induction. As already mentioned, the hard part is to show that certain paths
do not exist.
Coloring axiom schemes. In particular, we use three axiom schemes, having to do with partitioning memory into a small set of colors. We call instances of these schemes “coloring axioms”.
Our coloring axioms are simple, and are easily proved using S PASS (in under ten seconds) from
the induction axioms. For example, the first coloring axiom scheme, NoExit[A, f ], says that if no
f -edges leave color class, A, then no f -paths leave A. It turns out that the NoExit axiom scheme
implies – and thus is equivalent to – the induction scheme. However, we have found in practice
that explicitly adding other coloring axioms (which are consequences of NoExit) enables S PASS to
prove properties that it otherwise fails at.
We first assume that the programmer provides the colors by means of first-order formulas with
transitive closure. Our initial experience indicates that the generated coloring axioms are useful to
S PASS. In particular, it provides the ability to verify programs like the mark phase of a mark-andsweep garbage collector. This example has been previously reported as being beyond the capabilities
of ESC/Java. TVLA also succeeds on this example; however our new approach provides verification
methods that can in some instances be more precise than TVLA.
Prototype implementation. Perhaps most exciting, we have implemented the heuristics for
selecting colors and their corresponding axioms in a prototype using S PASS. We have used this
to automatically choose useful color axioms and then verify a series of small heap-manipulating
programs. We believe that the detailed examples presented here give convincing evidence of the
promise of our methodology. Of course much further study is needed.
Strengthening Nelson’s results. Greg Nelson considered a set of axiom schemes for reasoning
about reachability in function graphs, i.e., graphs in which there is at most one f -edge leaving any
node [Nel83]. He left open the question of whether his axiom schemes were complete for function
4
T. LEV-AMI, N. IMMERMAN, T. REPS, M. SAGIV, S. SRIVASTAVA, AND G. YORSH
graphs. We show that Nelson’s axioms are provable from T1 plus our induction axioms. We also
show that Nelson’s axioms are not complete: in fact, they do not imply NoExit.
Outline. The remainder of the paper is organized as follows: Section 2 explains our notation
and the setting; Section 3 fills in our formal framework, introduces the induction axiom scheme,
and presents the coloring axiom schemes; Section 4 provides more detail about TC-completeness
including a description of Nelson’s axioms, a proof that they are not TC-complete for the functional
case, and a proof that our axiomatization is TC-complete for words; Section 5 presents our heuristics including the details of their successful use on a variety of examples; Section 6 describes the
applicability of our methodology, relating it to the reasoning done in the TVLA system; Section 7
describes some related work; and Section 8 describes some conclusions and future directions.
2. P RELIMINARIES
This section defines the basic notations used in this paper and the setting.
2.1. Notation. Syntax: A relational vocabulary τ = {p1 , p2 , . . . , pk } is a set of relation symbols,
each of fixed arity. We use the letters u, v, and w (possibly with numeric subscript) for first-order
variables. We write first-order formulas over τ with quantifiers ∀ and ∃, logical connectives ∧,
∨, →, ↔, and ¬, where atomic formulas include: equality, pi (v1 , v2 , . . . vai ), and TC[f ](v1 , v2 ),
where pi ∈ τ is of arity ai and f ∈ τ is binary. Here TC[f ](v1 , v2 ) denotes the existence of a finite
path of 0 or more f edges from v1 to v2 . A formula without TC is called a first-order formula.
We use the following precedence of logical operators: ¬ has highest precedence, followed by
∧ and ∨, followed by → and ↔, and ∀ and ∃ have lowest precedence.
Semantics: A model, A, of vocabulary τ , consists of a non-empty universe, |A|, and a relation
pA over the universe interpreting each relation symbol p ∈ τ . We write A |= ϕ to mean that the
formula ϕ is true in the model A. For Σ a set of formulas, we write Σ |= ϕ (Σ semantically implies
ϕ) to mean that all models of Σ satisfy ϕ.
2.2. Setting. We are primarily interested in formulas that arise while proving the correctness of
programs. We assume that the programmer specifies pre and post-conditions for procedures and
loop invariants using first-order formulas with transitive closure on binary relations. The transformer
for a loop body can be produced automatically from the program code.
For instance, to establish the partial correctness with respect to a user-supplied specification of
a program that contains a single loop, we need to establish three properties: First, the loop invariant
must hold at the beginning of the first iteration; i.e., we must show that the loop invariant follows
from the precondition and the code leading to the loop. Second, the loop invariant provided by the
user must be maintained; i.e., we must show that if the loop invariant holds at the beginning of an
iteration and the loop condition also holds, the transformer causes the loop invariant to hold at the
end of the iteration. Finally, the postcondition must follow from the loop invariant and the condition
for exiting the loop.
In general, these formulas are of the form
ψ1 [τ ] ∧ T r[τ, τ ′ ] → ψ2 [τ ′ ]
SIMULATING REACHABILITY USING FIRST-ORDER LOGIC
5
where τ is the vocabulary of the before state, τ ′ is the vocabulary of the after state,1 and T r is
the transformer, which may use both the before and after predicates to describe the meaning of the
module to be executed. If symbol f denotes the value of a predicate before the operation, then f ′
denotes the value of the same predicate after the operation.
An interesting special case is the proof of the maintenance formula of a loop invariant. This
has the form:
LC[τ ] ∧ LI[τ ] ∧ T r[τ, τ ′ ] → LI[τ ′ ]
Here LC is the condition for entering the loop and LI is the loop invariant. LI[τ ′ ] indicates that the
loop invariant remains true after the body of the loop is executed.
The challenge is that the formulas of interest contain transitive closure; thus, the validity of
these formulas cannot be directly proven using a theorem prover for first-order logic.
3. A XIOMATIZATION
OF
T RANSITIVE C LOSURE
The original formula that we want to prove, χ, contains transitive closure, which first-order
theorem provers cannot handle. To address this problem, we replace χ by a new formula, χ′ , where
all appearances of TC[f ] have been replaced by the new binary relation symbol, ftc .
We show in this paper that from χ′ , we can often automatically generate an appropriate firstorder axiom, σ, with the following two properties:
(1) if σ → χ′ is valid in FO, then χ is valid in FO(TC).
(2) A theorem prover successfully proves that σ → χ′ is valid in FO.
We now explain the theory behind this process. A TC model, A, is a model such that if f and
ftc are in the vocabulary of A, then (ftc )A = (f A )⋆ ; i.e., A interprets ftc as the reflexive, transitive
closure of its interpretation of f .
A first-order formula ϕ is TC valid iff it is true in all TC models. We say that an axiomatization,
Σ, is TC sound if every formula that follows from Σ is TC valid. Since first-order reasoning is
sound, Σ is TC sound iff every σ ∈ Σ is TC valid.
We say that Σ is TC complete if for every TC-valid ϕ, Σ |= ϕ. If Σ is TC complete and TC
sound, then for all first-order ϕ,
Σ |= ϕ
⇔
ϕ
is TC valid
Thus a TC-complete set of axioms proves exactly the first-order formulas, χ′ , such that the
corresponding FO(TC) formula, χ, is valid.
All the axioms that we consider are TC valid. There is no recursively enumerable TC-complete
axiom system (Proposition 4.1). However, the axiomatization that we give does allow S PASS to
prove all the desired properties on the examples that we have tried. We do prove that our axiomatization is TC complete for word models (Theorem 4.8).
1In some cases it is useful for the postcondition formula to refer to the original vocabulary as well. This way the
postcondition can summarize some of the behavior of the transformer, e.g., summarize the behavior of an entire procedure.
6
T. LEV-AMI, N. IMMERMAN, T. REPS, M. SAGIV, S. SRIVASTAVA, AND G. YORSH
3.1. Some TC-Sound Axioms. We begin with our first TC axiom scheme. For any binary relation
symbol, f , let,
T1 [f ]
≡
∀u, v . ftc (u, v) ↔ (u = v) ∨ ∃w . f (u, w) ∧ ftc (w, v)
We first observe that T1 [f ] is “complete” in a very limited way for finite, acyclic graphs, i.e.,
T1 [f ] exactly characterizes the meaning of ftc for all finite, acyclic graphs. The reason that we say
this is limited is that it does not give us a complete set of first-order axioms: as is well known, there
is no first-order axiomatization of “finite”.
Proposition 3.1. Any finite and acyclic model of T1 [f ] is a TC model.
Proof. Let A |= T1 [f ] where A is finite and acyclic. Let a0 , b ∈ |A|. Assume that there is an f -path
from a0 to b. Since A |= T1 [f ], it is easy to see that A |= ftc (a0 , b). Conversely, suppose that
A |= ftc (a0 , b). If a0 = b, then there is a path of length 0 from a0 to b. Otherwise, by T1 [f ], there
exists an a1 ∈ |A| such that A |= f (a0 , a1 ) ∧ ftc (a1 , b). Note that a1 6= a0 since A is acyclic. If
a1 = b then there is an f -path of length 1 from a to b. Otherwise there must exist an a2 ∈ |A| such
that A |= f (a1 , a2 ) ∧ ftc (a2 , b) and so on, generating a set {a1 , a2 , . . .}. None of the ai can be
equal to aj , for j < i, by acyclicity. Thus, by finiteness, some ai = b. Hence A is a TC model.
Let T1′ [f ] be the ← direction of T1 [f ]:
T1′ [f ]
≡
∀u, v . ftc (u, v) ← (u = v) ∨ ∃w . f (u, w) ∧ ftc (w, v)
Proposition 3.2. Let ftc occur only positively in ϕ. If ϕ is TC valid, then T1′ [f ] |= ϕ.
Proof. Suppose that T1′ [f ] 6|= ϕ. Let A |= T1′ [f ] ∧ ¬ϕ. Note that ftc occurs only negatively in
¬ϕ. Furthermore, since A |= T1′ [f ], it is easy to show by induction on the length of the path, that
if there is an f -path from a to b in A, then A |= ftc (a, b). Define A′ to be the model formed from
A by interpreting ftc in A′ as (f A )⋆ . Thus A′ is a TC model and it only differs from A by the fact
′
that we have removed zero or more pairs from (ftc )A to form (ftc )A . Because A |= ¬ϕ and ftc
occurs only negatively in ¬ϕ, it follows that A′ |= ¬ϕ, which contradicts the assumption that ϕ is
TC valid.
Proposition 3.2 shows that proving positive facts of the form ftc (u, v) is easy; it is the task of
proving that paths do not exist that is more subtle.
Proposition 3.1 shows that what we are missing, at least in the acyclic case, is that there is no
first-order axiomatization of finiteness. Traditionally, when reasoning about the natural numbers,
this problem is mitigated by adding induction axioms. We next introduce an induction scheme that,
together with T1 , seems to be sufficient to prove any property we need concerning TC.
Notation: In general, we will use F to denote the set of all binary relation symbols, f , such
that TC[f
V] occurs in a formula we are considering. If ϕ[f ] is a formula in which f occurs, let
ϕ[F ] = f ∈F ϕ[f ]. Thus, for example, T1 [F ] is the conjunction of the axiom T1 [f ] for all binary
relation symbols, f , under consideration.
Definition 3.3. For any first-order formulas Z(u), P (u), and binary relation symbol, f , let the
induction principle, IND[Z, P, f ], be the following first-order formula:
(∀w . Z(w) → P (w))
∧ (∀u, v . P (u) ∧ f (u, v) → P (v))
→ ∀u, w . Z(w) ∧ ftc (w, u) → P (u)
In order to explain the meaning of IND and other axioms it is important to remember that we
are trying to write axioms, Σ, that are,
SIMULATING REACHABILITY USING FIRST-ORDER LOGIC
7
• TC valid, i.e., true in all TC models, and
• useful, i.e., all models of Σ are sufficiently like TC models that they satisfy the TC-valid properties we want to prove.
To make the meaning of our axioms intuitively clear, in this section we will say, for example, that
“y is ftc -reachable from x” to mean that ftc (x, y) holds. Later, we will assume that the reader has
the idea and just say “reachable” instead of “ftc -reachable”.
The intuitive meaning of the induction principle is that if every zero point satisfies P , and P
is preserved when following f -edges, then every point ftc -reachable from a zero point satisfies P .
Obviously this principle is TC valid, i.e., it is true for all structures such that ftc = f ⋆ .
As an easy application of the induction principle, consider the following cousin of T1 [f ],
T2 [f ]
≡
∀u, v . ftc (u, v) ↔ (u = v) ∨ ∃w . ftc (u, w) ∧ f (w, v)
The difference between T1 and T2 is that T1 requires that each path represented by ftc starts with
an f edge and T2 requires the path to end with an f edge. It is easy to see that neither of T1 [f ],
T2 [f ] implies the other. However, in the presence of the induction principle they do imply each
other. For example, it is easy to prove T2 [f ] from T1 [f ] using IND[Z, P, f ] where Z(v) ≡ v = u
and P (v) ≡ u = v ∨ ∃w . ftc (u, w) ∧ f (w, v). Here, for each u we use IND[Z, P, f ] to prove by
induction that every v reachable from u satisfies the right-hand side of T2 [f ].
Another useful axiom scheme provable from T1 plus IND is the transitivity of reachability:
Trans[f ] ≡ ∀u, v, w . ftc (u, w) ∧ ftc (w, v) → ftc (u, v)
3.2. Coloring Axioms. We next describe three TC-sound axioms schemes that are not implied by
T1 [F ] ∧ T2 [F ], and are provable from the induction principle. We will see in the sequel that these
coloring axioms are very useful in proving that paths do not exist, permitting us to verify a variety
of algorithms. In Section 5, we will present some heuristics for automatically choosing particular
instances of the coloring axiom schemes that enable us to prove our goal formulas.
The first coloring axiom scheme is the NoExit axiom scheme:
(∀u, v . A(u) ∧ ¬A(v) → ¬f (u, v)) → ∀u, v . A(u) ∧ ¬A(v) → ¬ftc (u, v)
for any first-order formula A(u), and binary relation symbol, f , NoExit[A, f ] says that if no f -edge
leaves color class A, then no point outside of A is ftc -reachable from A.
Observe that although it is very simple, NoExit[A, f ] does not follow from T1 [f ] ∧ T2 [f ].
Let G1 = (V, f, ftc , A) be a model consisting of two disjoint cycles: V = {1, 2, 3, 4}, f =
{h1, 2i, h2, 1i, h3, 4i, h4, 3i}, and A = {1, 2}. Let ftc have all 16 possible pairs. Thus G1 satisfies T1 [f ] ∧ T2 [f ] but violates NoExit[A, f ]. Even for acyclic models, NoExit[A, f ] does not
follow from T1 [f ] ∧ T2 [f ] because there are infinite models in which the implication does not hold
(Proposition 4.7).
NoExit[A, f ] follows easily from the induction principle: if no f -edges leave A, then induction
tells us that everything ftc -reachable from a point in A satisfies A. Similarly, NoExit[A, f ] implies
the induction axiom, IND[Z, A, f ], for any formula Z.
The second coloring axiom scheme is the GoOut axiom: for any first-order formulas A(u), B(u),
and binary relation symbol, f , GoOut[A, B, f ] says that if the only f -edges leaving color class A
are to B, then any ftc -path from a point in A to a point not in A must pass through B.
(∀u, v . A(u) ∧ ¬A(v) ∧ f (u, v) → B(v)) →
∀u, v . A(u) ∧ ¬A(v) ∧ ftc (u, v) → ∃w . B(w) ∧ ftc (u, w) ∧ ftc (w, v)
8
T. LEV-AMI, N. IMMERMAN, T. REPS, M. SAGIV, S. SRIVASTAVA, AND G. YORSH
To see that GoOut[A, B, f ] follows from the induction principle, assume that the only f -edges out
of A enter B. For any fixed u in A, we prove by induction that any point v ftc -reachable from u is
either in A or has a predecessor, b in B, that is ftc -reachable from u.
The third coloring axiom scheme is the NewStart axiom, which is useful in the context of
dynamically changing graphs: for any first-order formula A(u), and binary relation symbols f and
g, think of f as the previous edge relation and g as the current edge relation. NewStart[A, f, g] says
that if there are no new edges between A nodes, then any new path, i.e., gtc but not ftc , from A must
leave A to make its change:
(∀u, v . A(u) ∧ A(v) ∧ g(u, v) → f (u, v)) →
∀u, v . gtc (u, v) ∧ ¬ftc (u, v) → ∃w . ¬A(w) ∧ gtc (u, w) ∧ gtc (w, v)
NewStart[A, f, g] follows from the induction principle by a proof that is similar to the proof of
GoOut[A, B, f ].
3.2.1. Linked Lists. The spirit behind our consideration of the coloring axioms is similar to that
found in a paper of Greg Nelson’s in which he introduced a set of reachability axioms for a functional predicate, f , i.e., there is at most one f edge leaving any point [Nel83]. Nelson asked whether
his axiom schemes are complete for the functional setting. We remark that Nelson’s axiom schemes
are provable from T1 plus our induction principle. However, Nelson’s axiom schemes are not complete: we constructed a functional graph that satisfies Nelson’s axioms but violates NoExit[A, f ]
(Proposition 4.7).
At least one of Nelson’s axiom schemes seems orthogonal to our coloring axioms and may
be useful in certain proofs. Nelson’s fifth axiom scheme states that the points reachable from a
given point are linearly ordered. The soundness of the axiom scheme is due to the fact that f is
functional. We make use of a simplified version of Nelson’s ordering axiom scheme: Let Func[f ] ≡
∀u, v, w . f (u, v) ∧ f (u, w) → v = w; then,
Order[f ] ≡ Func[f ] → ∀u, v, w . ftc (u, v) ∧ ftc (u, w) → ftc (v, w) ∨ ftc (w, v)
3.2.2. Trees. When working with programs manipulating trees, we have a fixed set of selectors Sel
and transitive closure is performed on the down relation, defined as
_
∀v1 , v2 . down(v1 , v2 ) ↔
s(v1 , v2 )
s∈Sel
Trees have no sharing (i.e., the down relation is injective), thus a similar axiom to Order[f ] is used:
∀u, v, w . downtc (v, u) ∧ downtc (w, u) → downtc (v, w) ∨ downtc (w, v)
Another important property of trees is that the subtrees below distinct children of a node are disjoint.
We use the following axioms to capture this, where s1 6= s2 ∈ Sel:
∀v, v1 , v2 , w . ¬(s1 (v, v1 ) ∧ s2 (v, v2 ) ∧ downtc (v1 , w) ∧ downtc (v2 , w))
SIMULATING REACHABILITY USING FIRST-ORDER LOGIC
9
4. O N TC-C OMPLETENESS
In this section we consider the concept of TC-Completeness in detail. The reader anxious to
see how we use our methodology is encouraged to skim or skip this section.
We first show that there is no recursively enumerable TC-complete set of axioms.
Proposition 4.1. Let Γ be an r.e. set of TC-valid first-order sentences. Then Γ is not TC-complete.
Proof. By the proof of Corollary 9, page 11 of [IRR+ 04a], there is a recursive procedure that, given
any Turing machine Mn as input, produces a first-order formula ϕn in a vocabulary τn such that
ϕn is TC-valid iff Turing machine, Mn , on input 0 never halts. The vocabulary τn consists of the
two binary relation symbols, E, Etc , constant symbols, a, d, and some unary relation symbols. It
follows that if Γ were TC-complete, then it would prove all true instances of ϕn and thus the halting
problem would be solvable.
Proposition 4.1 shows that even in the presence of only one binary relation symbol, there is no
r.e. TC-complete axiomatization.
In [Avr03], Avron gives an elegant finite axiomatization of the natural numbers using transitive
closure, a successor relation and the binary function symbol, “+”. Furthermore, he shows that
multiplication is definable in this language. Since the unique TC-model for Avron’s axioms is the
standard natural numbers it follows that:
Corollary 4.2. Let Γ be an arithmetic set of TC-valid first-order sentences over a vocabulary including a binary relation symbol and a binary function symbol (or a ternary relation symbol). Then
Γ is not TC-complete.
In Proposition 3.1 we showed that any finite and acyclic model of T1 [f ] is a TC model. This
can be strengthened to
Proposition 4.3. Any finite model of T1 plus IND is a TC-model.
Proof. Let A be a finite model of T1 plus IND. Let f be a binary relation symbol, and let a, b be
elements of the universe of A. Since A |= T1 , if there is an f path from a to b then A |= ftc (a, b).
Conversely, suppose that there is no f path from a to b. Let Ra be the set of elements of the
universe of A that are reachable from a. Let k = |Ra |. Since A is finite we may use existential
quantification to name exactly all the elements of Ra : x1 , . . . , xk . We can then define the color
class: C(y) ≡ y = x1 ∨ · · · ∨ y = xk . Then we can prove using IND, or equivalently NoExit, that
no vertex outside this color class is reachable from a, i.e., A |= ¬ftc (a, b). Thus, as desired, A is a
TC-model.
4.1. More About TC-Completeness. Even though there is no r.e. set of TC-complete axioms in
general, there are TC-complete axiomatizations for certain interesting cases. Let Σ be a set of
formulas. We say that ψ is TC-valid wrt Σ iff every TC-model of Σ satisfies ψ. Let Γ be TC-sound.
We say that Γ is TC-complete wrt Σ iff Γ ∪ Σ ⊢ ψ for every ψ that is TC-valid wrt Σ. We are
interested in whether T1 plus IND is TC-complete with respect to interesting theories, Σ.
Since TC[s](a, b) asserts the existence of a finite s-path from a to b, we can express that a
structure is finite by writing the formula: Φ ≡ Func[s] ∧ ∃x∀y . stc (x, y). Observe that every TCmodel that satisfies Φ is finite. Thus, if we are in a setting – as is frequent in logic – where we may
add a new binary relation symbol, s, then finiteness is TC-expressible.
10
T. LEV-AMI, N. IMMERMAN, T. REPS, M. SAGIV, S. SRIVASTAVA, AND G. YORSH
Proposition 4.4. Let Σ be a finite set of formulas, and Γ an r.e., TC-complete axiomatization wrt Σ
in a language where finiteness is TC-expressible. Then finite TC-validity for Σ is decidable.
Proof. Let Φ be a formula as above that TC-expresses finiteness. Let ψ be any formula. If ψ is
not finite TC-valid wrt Σ, then we can find a finite TC model of Σ where ψ is false. If ψ is finite
TC-valid, then Γ ∪ Σ ⊢ Φ → ψ, and we can find this out by systematically generating all proofs
from Γ.
From Proposition 4.4 we know that we must restrict our search for cases of TC-completeness
to those where finite TC-validity is decidable. In particular, since the finite theory of two functional
relations is undecidable, e.g., [IRR+ 04a], we know that,
Corollary 4.5. There are no r.e. TC-valid axioms for the functional case even if we restrict to at
most two binary relation symbols.
4.2. Nelson’s Axioms. Our idea of considering transitive-closure axioms is similar in spirit to the
approach that Nelson takes [Nel83]. To prove some program properties, he introduces a set of
reachability axiom schemes for a functional predicate, f . By “functional” we mean that f is a
partial function: Func[f ] ≡ ∀u, v, w . f (u, v) ∧ f (u, w) → v = w.
We remark that Nelson’s axiom schemes are provable from T1 plus our induction principle. At
least two of his schemes may be useful for us to add in our approach. Nelson asked whether his
axioms are complete for the functional setting. It follows from Corollary 4.5 that the answer is no.
We prove below that Nelson’s axioms do not prove NoExit.
f
Nelson’s basic relation symbols are ternary. For example, he writes “u →
v” to mean that there
x
x (u, v), where, for
is an f -path from u to v that follows no edges out of x. We encode this as, ftc
x
each parameter x we add a new relation symbol, f , together with the assertion: ∀u, v . f x (u, v) ↔
f (u, v) ∧ (u 6= x). Nelson also includes a notation for modifying the partial function f . He writes,
(p)
fq for the partial function that agrees with f everywhere except on argument p where it has value
q. Nelson’s eighth axiom scheme asserts a basic consistency property for this notation. In our
(p)
translation we simply assert that fq (u, v) ↔ (u 6= p ∧ f (u, v)) ∨ (u = p ∧ v = q). When we
translate Nelson’s eighth axiom scheme the result is tautological, so we can safely omit it.
Using our translation, Nelson’s axiom schemes are the following.
x (u, v) ↔ (u = v) ∨ ∃z . (f x (u, z) ∧ f x (z, v))
(N1) ftc
tc
x (u, v) ∧ f x (v, w) → f x (u, w)
(N2) ftc
tc
tc
x (u, v) → f (u, v)
(N3) ftc
tc
y
z (u, y) → f z (u, x)
(N4) ftc
(u, x) ∧ ftc
tc
y
x (u, y)
(N5) ftc (u, x) → ftc
(u, x) ∨ ftc
y
z (u, y) → f z (x, y)
(N6) ftc
(u, x) ∧ ftc
tc
x (u, v)
(N7) f (x, u) ∧ ftc (u, v) → ftc
These axiom schemes can be proved using appropriate instances of T1 and the induction principle. Just as we showed in Proposition 3.1 that any finite and acyclic model of T1 [f ] is a TC model,
we have that,
Proposition 4.6. Any finite and functional model of Nelson’s axioms is a TC-model.
SIMULATING REACHABILITY USING FIRST-ORDER LOGIC
11
Proof. Consider any finite and function model, M. We claim that for each f and x ∈ |M|,
x )M = ((f x )M )⋆ . If there is an f x path from u to v, then it follows from repeated uses of
(ftc
x holds.
(N1) that ftc
If there is no f x path from u to v and u is not on an f -cycle, then using (N1) we can follow
x does not hold.
f -edges from u to the end and prove that ftc
x
If there is no f path from u to v and u is on an f -cycle containing x, then using (N1) we can
x (u, v) does not hold.
follow f -edges from u to x to prove that ftc
Finally, if there is no f path from u to v and u is on an f -cycle, suppose for the sake of a
x (u, v) must
contradiction that ftc (u, v) holds. Let x be the predecessor of u on the cycle. By N7, ftc
hold. However, this contradicts the previous paragraph.
Axiom schemes (N5) and (N7) may be useful for us to assert when f is functional. (N5) says
that the points reachable from u are totally ordered in the sense that if x and y are both reachable
from u, then in the path from u either x comes first or y comes first. (N7) says that if there is an
edge from x to u and a path from u to v, then there is a path from u to v that does not go through
x. This implies the useful property that no vertex not on a cycle is reachable from a vertex on the
cycle.
We conclude this section by proving the following,
Proposition 4.7. Nelson’s axioms do not imply NoExit.
0 , f 1 , f 2 , . . . , f ∞ , A) such that V = N ∪ {∞}, the
Proof. Consider the structure G = (V, f, ftc , ftc
tc
tc tc
set of natural numbers plus a point at infinity. Let A = N, i.e., the color class A is interpreted as all
points except ∞. Define f = {hu, u + 1i | u ∈ N}, i.e., there is an edge from every natural number
to its successor, but ∞ is isolated. However, let ftc = {hu, vi | u ≤ v}, i.e., G believes that there is
k = {hu, vi|u ≤ v ∧ (k <
a path from each natural number to infinity. Similarly, for each k ∈ V , ftc
u ∨ v ≤ k)}.
It is easy to check that G satisfies all of Nelson’s axioms.
The problem is that G |= ¬NoExit[A, f ]. It follows that Nelson’s axioms do not entail
NoExit[A, f ]. This is another proof that they are not TC complete.
4.3. TC-Completeness for Words. In this subsection, we prove that T1 plus IND is TC-complete
for words.
For any alphabet, Σ, let the vocabulary of words over Σ be vocab(Σ) = h0, max; s2 , s2tc , Pσ1 :
σ ∈ Σi . The domain of a word model is an ordered set of positions, and the unary relation Pσ (x)
expresses the presence of symbol σ at position x. s is the successor relation over positions, and stc
is its transitive closure. The constants 0 and max represent the first and last positions in the word.
A simple axiomatization of words is AΣw , the conjunction of the following four statements:
(A1) ∀x . (¬s(x, 0) ∧ ¬s(max, x) ∧ (x 6= 0 → ∃y . s(y, x)) ∧ (x 6= max → ∃y . s(x, y)))
(A2) ∀xyz . ((s(x, y) ∧ s(x, z)) ∨ (s(y, x) ∧ s(z, x))) → y = z
(A3) ∀x . stc (0, x) ∧ stc (x, max)
_
^
(A4) ∀x .
(Pσ (x) ∧
¬Pτ (x))
σ∈Σ
τ 6=σ
In particular, observe that a TC-model of AΣw is exactly a Σ word. Let Γ = IND ∪ {T1 }. We
wish to prove the following:
Theorem 4.8. Γ is TC-complete wrt AΣw .
12
T. LEV-AMI, N. IMMERMAN, T. REPS, M. SAGIV, S. SRIVASTAVA, AND G. YORSH
We first note that Γ ∪ {AΣw } implies acyclicity: ∀xy . s(x, y) → ¬stc (y, x). The proof using
induction proceeds as follows: in the base case, there is no loop at 0. Inductively, suppose there is
no loop starting at x, s(x, y) holds, but there is a loop at y, i.e., ∃z . s(y, z) ∧ stc (z, y). Then by T1
and IND we know ∃x′ . stc (z, x′ ) ∧ s(x′ , y), and stc (y, x′ ). (A2) asserts that the in-degree of s is 1,
which means x′ = x and we have a contradiction: stc (y, x).
In order to prove Theorem 4.8, we need to show that if ϕ is true in all TC models of Γ ∪ {AΣw },
i.e., in all words, then Γ ∪ {AΣw } ⊢ ϕ. By the completeness of first-order logic it suffices to show
that Γ ∪ {AΣw } |= ϕ. We prove the contrapositive of this in Lemma 4.10. In order to do so, we first
construct a DFA Dϕ that has some desirable properties.
Lemma 4.9. For any ϕ ∈ L(vocab(Σ)) we can build a DFA Dϕ = (Qϕ , Σ, δϕ , q1 , Fϕ ), satisfying
the following properties:
(1) The states q1 , q2 , . . . qn of Dϕ are first-order definable as formulas q11 , q21 , . . . qn1 , where intuitively qi (x) will mean that Dϕ is in state qi after reading symbols at word positions 0, 1, . . . , x.
(2) The transition function δϕ of Dϕ is captured by the first-order definitions of the states. That is,
for all i ≤ n, Γ ∪ AΣw semantically
implies the following two formulas for every state qi :
_
(a) qi (0) ↔
Pσ (0).
σ∈Σ,δϕ (q1 ,σ)=qi
(b) ∀u, v . s(u, v) → qi (v)
↔
_
(Pσ (v) ∧ qj (u)) .
σ∈Σ,δϕ (qj ,σ)=qi
(3) Γ ∪ {AΣw } |= ϕ ↔ F (max),
where F (u) ≡
_
qi (u).
qi ∈Fϕ
Proof. We prove properties 1, 2, and 3 while constructing Dϕ and the first-order definitions of
its states by induction on the length of ϕ. The reward is that we get a generalized form of the
McNaughton-Papert [MP71] construction that works on non-standard models.
Some subformulas of ϕ may have free variables, e.g., x, y. In the inductive step considering
such subformulas, we expand the vocabulary of the automaton to Σ′ = {x, ǫ} × {y, ǫ} × Σ. We
write Pσ (u) ∧ (x = u) ∧ (y 6= u) to mean that at position u, symbol σ occurs, as does x, but not y.
Note: Since every structure gives a unique value to each variable, x, we are only interested in
strings in which x occurs at exactly one position.
For the following induction, let B be any model of Γ ∪ {AΣw }. For the intermediate stages of
induction where some variables may occur freely, we assume that B interprets these free variables.
We prove that the formulas of properties 2 and 3 must hold in B at each step of the induction.
Base cases: ϕ is either Pσ (x), x = y, s(x, y), or stc (x, y).
ϕ = Pσ (x): The automaton for Pσ (x) and its state definitions are shown in Fig 1.
State predicate
q1 (v)
q2 (v)
q3 (v)
Definition
¬stc (x, v)
stc (x, v) ∧ Pσ (x)
stc (x, v) ∧ ¬Pσ (x)
Table 1: DPσ (x)
Figure 1: DPσ (x)
SIMULATING REACHABILITY USING FIRST-ORDER LOGIC
13
Properties 2 and 3 can be verified as follows:
For property 2b, suppose that B |= s(u, v). We must show that B |= q2 (v) iff one of two rules
leading to state q2 holds. These two rules correspond to the edge from q1 (if x = v), and the self
loop on q2 (if x 6= v). Suppose B |= q2 (v) ∧ (v = x). Expanding the definition of q2 , we get
B |= stc (x, v) ∧ Pσ (x) ∧ (v = x). But this means B |= ¬stc (x, u) since B |= Γ ∪ {AΣw } and
we have acyclicity. Therefore, we have B |= q1 (u) by definition of q1 , and we get the desired
conclusion, B |= q1 (u) ∧ Pσ (v).
The case corresponding to x 6= v is also easy, and relies on the fact that B |= stc (x, v) ∧
s(u, v) ∧ (x 6= v) → stc (x, u). In other words, if q2 (v) holds and x 6= v, then q2 holds at v’s
predecessor too.
This proves one direction of property 2b for state q2 . The other direction for q2 , and the proofs
for other states proceed similarly. The proof for 2a is similar.
For property 3, we need to show that B |= Pσ (x) ↔ q2 (max). This can be verified easily
from the definition of q2 .
ϕ = (x = y) or s(x, y): The automata and their state definitions for ϕ = (x = y) and ϕ = s(x, y)
are shown in Figs 2 and 3. Properties 2 and 3 can be verified easily for these definitions.
State predicate
q1 (v)
q2 (v)
q3 (v)
Definition
¬stc (x, v)
(x = y) ∧ stc (x, v)
(x 6= y) ∧ stc (x, v)
Table 2: Dx=y
Figure 2: Dx=y
State predicate
q1 (v)
q2 (v)
q3 (v)
q4 (v)
Definition
¬stc (x, v)
x=v
s(x, y) ∧ stc (y, v)
stc (x, v) ∧ (x 6= v)∧
¬s(x, y)
Table 3: Ds(x,y)
Figure 3: Ds(x,y)
ϕ = stc (x, y): The automaton for ϕ = stc (x, y), and its state definitions are shown in Fig 4.
We provide a sketch of the proof of property 2b for state q3 . Proofs for other states follow
using similar arguments. Suppose B |= q3 (v) ∧ s(u, v). Expanding the definition of q3 (v), we
get B |= stc (x, y) ∧ stc (y, v) ∧ s(u, v).
There are two possibilities: v 6= y and v = y, corresponding to the loop on state q3 , and the
incoming edges from q2 or q1 . Suppose v = y. Now we have two further cases, x = y and x 6= y.
14
T. LEV-AMI, N. IMMERMAN, T. REPS, M. SAGIV, S. SRIVASTAVA, AND G. YORSH
State predicate Definition
q1 (v)
¬stc (x, v)
q2 (v)
stc (x, v)∧
¬(stc (x, y) ∧ stc (y, v))
q3 (v)
stc (x, y) ∧ stc (y, v)
Table 4: Dstc (x,y)
Figure 4: Dstc (x,y)
If x = y = v, we get B |= ¬stc (x, u), or B |= q1 (u) ∧ s(u, x) ∧ (x = y = v), denoting the
appropriate transition from state q1 .
On the other hand, if B |= (x 6= y), we need to show that q3 was reached via q2 . Expanding the
definition of q3 (v) we have B |= stc (x, y)∧stc (y, v). Since y = v, we get B |= stc (x, u)∧s(u, y).
But by definition of q2 , this means B |= q2 (u). Thus, we have B |= q2 (u) ∧ s(u, v) ∧ v = y, the
appropriate transition rule for moving from state q2 to q3 .
For this direction of property 2b, the only remaining case is y 6= v. In this case, it is easy to
prove that we entered state q3 at y, and looped thereafter using the appropriate transition for the
loop.
For the reverse direction, we need to prove that if a transition rule is applicable at a position
then the corresponding next state must hold at the next position. This is easily verified using the
state-definitions. Property 2 for other states follows by similar arguments. Property 3 can also be
verified easily using the definition of q3 .
Inductive steps: ϕ is either ϕ1 ∧ ϕ2 , or ¬ψ, or ∃x . ψ(x).
ϕ = ϕ1 ∧ϕ2 : Inductively we have Dϕ1 and Dϕ2 with final state definitions qf1 and qf2 respectively.
To construct Dϕ , we perform the product construction: let qi be state definitions of Dϕ1 and qi′
those of Dϕ2 . Then the state definitions of Dϕ are qhi,ji, and we have qhi,ji(u) ≡ qi (u) ∧ qj′ (u).
The accepting states are
_
qhf1 ,f2 i (u).
Fϕ1 ∧ϕ2 (u) ≡
f1 ∈F1 ∧f2 ∈F2
Property 1 holds because we are still in first-order. Property 2 follows because we are just
performing logical transliterations of the standard DFA conjunction operation. Property 3 follows
from the fact that we already have B |= F1 (max) ↔ ϕ1 and B |= F2 (max) ↔ ϕ2 , and from the
definition of Fϕ1 ∧ϕ2 .
ϕ = ¬ψ: In this case, we take the complement of Dψ which is easy because our automata are
deterministic. Let the final state of Dψ be F ′ . Dϕ has the same state definitions as ψ, but its final
state definition is F (u) ≡ ¬F ′ (u). It is easy to see that properties 1, 2 and 3 hold in this case.
ϕ = ∃x . ψ(x):
Inductively we have Dψ = ({q1 , . . . , qn }, Σ × {x, ǫ}, δψ , q1 , Fψ ).
First we transform Dψ to an NFA Nϕ = ({p1 , . . . , pn , p′1 , . . . , p′n }, Σ, δ, p1 , F ), where F =
′
{pi |qi ∈ Fψ } and δ(pi , σ) = {pj , p′k |δψ (qi , σ ∧ ¬x) = qj , δψ (qi , σ ∧ x) = qk }.
Thus Nϕ no longer sees x’s. Instead, it guesses the one place that x might occur, and that is
where the transition from pi to p′i occurs. (See Fig. 5)
Let pi (u) ≡ ∃x . ¬stc (x, u) ∧ qi (u); p′i (u) ≡ ∃x . stc (x, u) ∧ qi (u).
SIMULATING REACHABILITY USING FIRST-ORDER LOGIC
15
Figure 5: N∃x . Pσ (x)
Define Dϕ to be the DFA equivalent to Nϕ using the subset construction. Let S0 = {pi0 , p′j |j ∈
J0 }, S1 = {pi1 , p′j |j ∈ J1 } be two states of Dϕ . (Note that each reachable state of Dϕ has exactly
one element of {p1 , . . . , pn }.)
Observe that in a “run” of Nϕ on B, we can be in state pi at position u iff B |= pi (u) and we
can be in state p′i of u iff B |= p′i (u). Thus, the first-order formula capturing state S0 is
^
^
¬p′j (u)
p′j (u) ∧
S0 (u) ≡ pi0 ∧
j∈J0
j ∈J
/ 0
Conditions 2 and 3 for Dϕ thus follow by these conditions for Dψ , which hold by inductive
assumption.
For example, if δϕ (S0 , σ) = S1 , then δψ (pi0 , σ ∧ ¬x) = pi1 , and j ∈ J1 iff δψ (qi0 , σ ∧ x) = qj
or δψ (qj0 , σ ∧ ¬x) = qj for some j0 ∈ J0 .
Thus, we have inductively constructed the Dϕ and proved that it satisfies properties 1, 2, and 3.
Lemma 4.9 tells us that for any model B of Γ ∪ {AΣw }, B |= ϕ iff B |= Fϕ (max). In other
words, B |= ϕ iff B “believes” that there is a path from the start state to some qf in Fϕ . As a part of
the next lemma, we use induction to prove that this implies that there actually must be a path in Dϕ
from the start state to some qf in Fϕ .
Lemma 4.10. Suppose B |= Γ ∪ {AΣw } ∪ {ϕ}. Then, there exists a word, w0 , such that its
corresponding word model, B0 , satisfies ϕ.
Proof. By Lemma 4.9, we can construct Dϕ , and we have B |= Fϕ (max). So B “believes” that
there is a path to some qf ∈ Fϕ . Suppose there is no such path in Dϕ . Let C denote the disjunction
of all states that are truly reachable from the start state in Dϕ . This situation can be expressed as
follows: ∀u, v . C(u) ∧ s(u, v) → C(v). But this is exactly the premise for the axiom scheme
NoExit, which must hold since B |= Γ. Therefore, we have B |= ∀u, v . C(v) ∧ stc (u, v) → C(v).
This implies some accepting state qf should be in C, because B |= ∀u . stc (u, max) ∧ Fϕ (max),
and we get a contradiction.
Therefore, there has to be a real path from the start state to a final state qf in Dϕ . This implies
that the DFA Dϕ accepts some standard word, w0 . Let B0 be the word model corresponding to w0 .
Thus B0 |= Fϕ (max), and therefore by Lemma 4.9 B0 |= ϕ as desired.
16
T. LEV-AMI, N. IMMERMAN, T. REPS, M. SAGIV, S. SRIVASTAVA, AND G. YORSH
Node reverse(Node x){
[0] Node y = null;
[1] while (x != null){
[2]
Node t = x.next;
[3]
x.next = y;
[4]
y = x;
[5]
x = t;
[6] }
[7] return y;
}
Figure 6: A simple Java-like implementation of the in-place reversal of a singly linked list.
5. H EURISTICS
FOR
U SING
THE
C OLORING A XIOMS
This section presents heuristics for using the coloring axioms. Toward that end, it answers the
following questions:
• How can the coloring axioms be used by a theorem prover to prove χ? (Section 5.2)
• When should a specific instance of a coloring axiom be given to the theorem prover while trying
to prove χ? (Section 5.4)
• What part of the process can be automated? (Section 5.5)
We first present a running example (more examples are described in Section 5.6 and used in later
sections to illustrate the heuristics). We then explain how the coloring axioms are useful, describe
the search space for useful axioms, give an algorithm for exploring this space, and conclude by
discussing a prototype implementation we have developed that proves the example presented and
others.
5.1. Reverse Specification. The heuristics described in Sections 5.2–5.4 are illustrated on problems that arise in the verification of partial correctness of a list reversal procedure. Other examples
proven using this technique can be found in Section 5.6.
The procedure reverse, shown in Fig. 6, performs in-place reversal of a singly linked list, destructively updating the list. The precondition requires that the input list be acyclic and unshared
(i.e., each heap node is pointed to by at most one heap node). For simplicity, we assume that there
is no garbage. The postcondition ensures that the resulting list is acyclic and unshared. Also, it
ensures that the nodes reachable from the formal parameter on entry to reverse are exactly the nodes
reachable from the return value of reverse at the exit. Most importantly, it ensures that each edge in
the original list is reversed in the returned list.
The specification for reverse is shown in Fig. 7. We use unary predicates to represent program
variables and binary predicates to represent data-structure fields. Fig. 7(a) defines some shorthands.
To specify that a unary predicate z can point to a single node at a time and that a binary predicate
f of a node can point to at most one node (i.e., f is a partial function), we use unique[z] and
f unc[f ] . To specify that there are no cycles of f -fields in the graph, we use acyclic[f ]. To specify
that the graph does not contain nodes shared by f -fields, (i.e., nodes with 2 or more incoming f fields), we use unshared[f ]. To specify that all nodes in the graph are reachable from z1 or z2 by
following f -fields, we use total[z1 , z2 , f ]. Another helpful shorthand is rx,f (v) which specifies that
v is reachable from the node pointed to by x using f -edges.
SIMULATING REACHABILITY USING FIRST-ORDER LOGIC
17
The precondition of the reverse procedure is shown in Fig. 7(b). We use the predicates xe
and ne to record the values of the variable x and the next field at the beginning of the procedure.
The precondition requires that the list pointed to by x be acyclic and unshared. It also requires
that unique[z] and f unc[f ] hold for all unary predicates z that represent program variables and all
binary predicates f that represent fields, respectively. For simplicity, we assume that there is no
garbage, i.e., all nodes are reachable from x.
The post-condition is shown in Fig. 7(c). It ensures that the resulting list is acyclic and unshared. Also, it ensures that the nodes reachable from the formal parameter x on entry to the
procedure are exactly the nodes reachable from the return value y at the exit. Most importantly, we
wish to show that each edge in the original list is reversed in the returned list (see Eq. (5.9)).
A loop invariant is given in Fig. 7(d). It describes the state of the program at the beginning of
each loop iteration. Every node is in one of two disjoint lists pointed to by x and y (Eq. (5.10)). The
lists are acyclic and unshared. Every edge in the list pointed to by x is exactly an edge in the original
list (Eq. (5.12)). Every edge in the list pointed to by y is the reverse of an edge in the original list
(Eq. (5.13)). The only original edge going out of y is to x (Eq. (5.14)).
The transformer is given in Fig. 7(e), using the primed predicates n′ , x′ , and y ′ to describe the
values of predicates n, x, and y, respectively, at the end of the iteration.
5.2. Proving Formulas using the Coloring Axioms. All the coloring axioms have the form A ≡
PA → CA , where PA and CA are closed formulas. We call PA the axiom’s premise and CA the
axiom’s conclusion. For an axiom to be useful, the theorem prover will have to prove the premise
(as a subgoal) and then use the conclusion in the proof of the goal formula χ. For each of the
coloring axioms, we now explain when the premise can be proved, how its conclusion can help, and
give an example.
NoExit. The premise PNoExit [C, f ] states that there are no f -edges exiting color class C.
When C is a unary predicate appearing in the program, the premise is sometimes a direct result of
the loop invariant. Another color that will be used heavily throughout this section is reachability
from a unary predicate, i.e., unary reachability, formally defined in Eq. (5.6). Let us examine
two cases. PNoExit [rx,f , f ] is immediate from the definition of rx,f and the transitivity of ftc .
PNoExit [rx,f , f ′ ] actually states that there is no f -path from x to an edge for which f ′ holds but f
does not, i.e., a change in f ′ with respect to f . Thus, we use the absence of f -paths to prove the
absence of f ′ -paths. In many cases, the change is an important part of the loop invariant, and paths
from and to it are part of the specification.
A sketch of the proof by refutation of PNoExit [rx′ ,n , n′ ] that arises in the reverse example is
given in Fig. 8. The numbers in brackets are the stages of the proof.
(1) The negation of the premise expands to:
∃u1 , u2 , u3 . x′ (u1 ) ∧ ntc (u1 , u2 ) ∧ ¬ntc (u1 , u3 ) ∧ n′ (u2 , u3 )
(2) Since u2 is reachable from u1 and u3 is not, by T2 , we have ¬n(u2 , u3 ).
(3) By the definition of n′ in the transformer, the only edge in which n differs from n′ is out of x
(one of the clauses generated from Eq. (5.15) is ∀v1 , v2 . ¬n′ (v1 , v2 )∨ n(v1 , v2 )∨ x(v1 )) . Thus,
x(u2 ) holds.
(4) By the definition of x′ it has an incoming n edge from x. Thus, n(u2 , u1 ) holds.
The list pointed to by x must be acyclic, whereas we have a cycle between u1 and u2 ; i.e., we have
a contradiction. Thus, PNoExit [rx′ ,n , n′ ] must hold.
CNoExit [C, f ] states there are no f paths (ftc edges) exiting C. This is useful because proving
the absence of paths is the difficult part of proving formulas with TC.
18
T. LEV-AMI, N. IMMERMAN, T. REPS, M. SAGIV, S. SRIVASTAVA, AND G. YORSH
def
unique[z] = ∀v1 , v2 .z(v1 ) ∧ z(v2 ) → v1 = v2
def
f unc[f ] = ∀v1 , v2 , v.f (v, v1 ) ∧ f (v, v2 ) → v1 = v2
def
acyclic[f ] = ∀v1 , v2 .¬f (v1 , v2 ) ∨ ¬TC[f ](v2 , v1 )
(a)
def
unshared[f ] = ∀v1 , v2 , v.f (v1 , v) ∧ f (v2 , v) → v1 = v2
def
total[z1 , z2 , f ] = ∀v.∃w.(z1 (w) ∨ z2 (w)) ∧ TC[f ](w, v)
def
rx,f (v) = ∃w . x(w) ∧ TC[f ](w, v)
(5.1)
(5.2)
(5.3)
(5.4)
(5.5)
(5.6)
def
− (v) = ∃w . x(w) ∧ TC[f ](v, w)
rx,←
f
(5.7)
def
(b)
pre = total[xe, xe, ne] ∧ acyclic[ne] ∧ unshared[ne] ∧
unique[xe] ∧ f unc[ne]
(c)
post = total[y, y, n] ∧ acyclic[n] ∧ unshared[n] ∧
∀v1 , v2 .ne(v1 , v2 ) ↔ n(v2 , v1 )
(d)
LI[x, y, n] = total[x, y, n] ∧ ∀v.(¬rx,n (v) ∨ ¬ry,n (v)) ∧
acyclic[n] ∧ unshared[n]
unique[x] ∧ unique[y] ∧ f unc[n] ∧
∀v1 , v2 .(rx,n (v1 ) → (ne(v1 , v2 ) ↔ n(v1 , v2 ))) ∧
∀v1 , v2 .(ry,n (v2 ) ∧ ¬y(v1 ) → (ne(v1 , v2 ) ↔ n(v2 , v1 ))) ∧
∀v1 , v2 , v.y(v1 ) → (x(v2 ) ↔ ne(v1 , v2 ))
def
def
(5.8)
(5.9)
(5.10)
(5.11)
(5.12)
(5.13)
(5.14)
def
(e)
T = ∀v.(y ′ (v) ↔ x(v)) ∧ ∀v.(x′ (v) ↔ ∃w.x(w) ∧ n(w, v)) ∧
∀v1 , v2 .n′ (v1 , v2 ) ↔
((n(v1 , v2 ) ∧ ¬x(v1 )) ∨ (x(v1 ) ∧ y(v2 )))
(5.15)
Figure 7: Example specification of reverse procedure: (a) shorthands, (b) precondition pre, (c)
postcondition post, (d) loop invariant LI[x, y, n], (e) transformer T (effect of the loop
body).
n[4]
x′ [1]
{
ntc [1]
@ABC
@ABC
/ GFED
/ GFED
u2 o
u1 K
KK¬ntc [1] n′ [1] ss
KK
ss
KK%
ysss
@ABC
GFED
u3 i ¬n[2]
x[3]
Figure 8: Proving PNoExit [rx,n , n′ ].
GoOut. The premise PGoOut [A, B, f ] states that all f edges going out of color class A, go to
B. When A and B are unary predicates that appear in the program, again the premise sometimes
holds as a direct result of the loop invariant. An interesting special case is when B is defined as
SIMULATING REACHABILITY USING FIRST-ORDER LOGIC
19
∃w . A(w) ∧ f (w, v). In this case the premise is immediate. Note that in this case the conclusion is provable also from T1 . However, from experience, the axiom is very useful for improving
performance (2 orders of magnitude when proving the acyclic part of reverse’s postcondition).
CGoOut [A, B, f ] states that all paths out of A must pass through B. Thus, under the premise
PGoOut [A, B, f ], if we know that there is a path from A to somewhere outside of A, we know that
there is a path to there from B. In case all nodes in B are reachable from all nodes in A, together
with the transitivity of ftc this means that the nodes reachable from B are exactly the nodes outside
of A that are reachable from A.
For example, CGoOut [y ′ , y, n′ ] allows us to prove that only the original list pointed to by y is
reachable from y ′ (in addition to y ′ itself).
NewStart. The premise PNewStart [C, g, h] states that all g edges between nodes in C are also
h edges. This can mean the iteration has not added edges or has not removed edges according to the
selection of h and g. In some cases, the premise holds as a direct result of the definition of C and
the loop invariant.
CNewStart [C, g, h] means that every g path that is not an h path must pass outside of C. Together with CNoExit [C, g], it proves there are no new paths within C.
For example, in reverse the NewStart scheme can be used as follows. No outgoing edges were
added to nodes reachable from y. There are no n or n′ edges from nodes reachable from y to nodes
not reachable from y. Thus, no paths were added between nodes reachable from y. Since the list
pointed to by y is acyclic before the loop body, we can prove that it is acyclic at the end of the loop
body.
We can see that NewStart allows the theorem prover to reason about paths within a color, and
the other axioms allow the theorem prover to reason about paths between colors. Together, given
enough colors, the theorem prover can often prove all the facts that it needs about paths and thus
prove the formula of interest.
5.3. The Search Space of Possible Axioms. To answer the question of when we should use a
specific instance of a coloring axiom when attempting to prove the target formula, we first define
the search space in which we are looking for such instances. The axioms can be instantiated with the
colors defined by an arbitrary unary formula (one free variable) and one or two binary predicates.
First, we limit ourselves to binary predicates for which TC was used in the target formula. Now,
since it is infeasible to consider all arbitrary unary formulas, we start limiting the set of colors we
consider.
The initial set of colors to consider are unary predicates that occur in the formula we want to
prove. Interestingly enough, these colors are enough to prove that the postcondition of mark and
sweep is implied by the loop invariant, because the only axiom we need is NoExit[marked, f ].
An immediate extension that is very effective is forward and backward reachability from unary
predicates, as defined in Eq. (5.6) and Eq. (5.7), respectively. Instantiating all possible axioms from
the unary predicates appearing in the formula and their unary forward reachability predicates, allows
us to prove reverse. For a list of the axioms needed to prove reverse, see Fig. 9. Other examples are
presented in Section 5.6. Finally, we consider Boolean combinations of the above colors. Though
not used in the examples shown in this paper, this is needed, for example, in the presence of sharing
or when splicing two lists together.
All the colors above are based on the unary predicates that appear in the original formula. To
prove the reverse example, we needed x′ as part of the initial colors. Table 5 gives a heuristic for
finding the initial colors we need in cases when they cannot be deduced from the formula, and how
it applies to reverse.
20
T. LEV-AMI, N. IMMERMAN, T. REPS, M. SAGIV, S. SRIVASTAVA, AND G. YORSH
NoExit[rx′ ,n , n′ ] GoOut[x, x′ , n]
NoExit[rx′ ,n′ , n] GoOut[x, y, n′ ]
NoExit[ry,n , n′ ]
NoExit[ry,n′ , n]
NewStart[rx′ ,n , n, n′ ]
NewStart[rx′ ,n′ , n, n′ ]
NewStart[ry,n , n, n′ ]
NewStart[ry,n′ , n, n′ ]
NewStart[rx′ ,n , n′ , n]
NewStart[rx′ ,n′ , n′ , n]
NewStart[ry,n , n′ , n]
NewStart[ry,n′ , n′ , n]
Figure 9: The instances of coloring axioms used in proving reverse.
Group
Roots[f]
StartChange[f,g]
EndChange[f,g]
(a)
Criteria
All changes are reachable from one of the colors using ftc
All edges for which f and g differ start from a node in these colors
All edges for which f and g differ end at a node in these colors
Group
Roots[n]
Roots[n′ ]
StartChange[n, n′ ]
EndChange[n, n′ ]
(b)
Colors
x(v), y(v)
x′ (v), y ′ (v)
x(v)
y(v), x′ (v)
Table 5: (a) Heuristic for choosing initial colors. (b) Results of applying the heuristic on reverse.
An interesting observation is that the initial colors we need can, in many cases, be deduced
from the program code. As in the previous section, we have a good way for deducing paths between
colors and within colors in which the edges have not changed. The program usually manipulates
fields using pointers, and can traverse an edge only in one direction. Thus, the unary predicates that
represent the program variables (including the temporary variables) are in many cases what we need
as initial colors.
5.4. Exploring the Search Space. When trying to automate the process of choosing colors, the
problem is that the set of possible colors to choose from is doubly-exponential in the number of
initial colors; giving all the axioms directly to the theorem prover is infeasible. In this section, we
define a heuristic algorithm for exploring a limited number of axioms in a directed way. Pseudocode
for this algorithm is shown in Fig. 10. The operator ⊢ is implemented as a call to a theorem prover.
Because the coloring axioms have the form A ≡ PA → CA , the theorem prover must prove PA
or the axiom is of no use. Therefore, the pseudocode works iteratively, trying to prove PA from the
current ψ ∧ Σ, and if successful it adds CA to Σ.
The algorithm tries colors in increasing levels of complexity. BC(i, C) gives all the Boolean
combinations of the predicates in C up to size i. After each iteration we try to prove the goal
formula. Sometimes we need the conclusion of one axiom to prove the premise of another. The
NoExit axioms are particularly useful for proving PNewStart . Therefore, we need a way to order
instantiations so that axioms useful for proving the premises of other axioms are acquired first.
The ordering we chose is based on phases: First, try to instantiate axioms from the axiom scheme
GoOut. Second, try to instantiate axioms from the axiom scheme NoExit. Finally, try to instantiate
axioms from the axiom scheme NewStart. For NewStart[c, f, g] to be useful, we need to be able
SIMULATING REACHABILITY USING FIRST-ORDER LOGIC
21
explore(Init, χ) {
Let χ = ψ → ϕ
Σ := {Trans[f ], Order[f ] | f ∈ F }
Σ := Σ ∪ {T1 [f ], T2 [f ] | f ∈ F }
C := {rc,f (v) | c ∈ Init, f ∈ F }
C := C ∪ Init
i := 1
forever {
C ′ := BC(i, C)
// Phase 1
foreach f ∈ F, cs 6= ce ∈ C ′
if Σ ∧ ψ ⊢ PGoOut [cs , ce , f ]
Σ := Σ ∪ {CGoOut [cs , ce , f ]}
// Phase 2
foreach f ∈ F, c ∈ C ′
if Σ ∧ ψ ⊢ PNoExit [c, f ]
Σ := Σ ∪ {CNoExit [c, f ]}
// Phase 3
foreach CNoExit [c, f ] ∈ Σ, g 6= f ∈ F
if Σ ∧ ψ ⊢ PNewStart [c, f, g]
Σ := Σ ∪ {CNewStart [c, f, g]}
if Σ ∧ ψ ⊢ ϕ
return SUCCESS
i := i + 1
}
}
Figure 10: An iterative algorithm for instantiating the axiom schemes. Each iteration consists of
three phases that augment the axiom set Σ
to show that there are either no incoming f -paths or no outgoing f -paths from c. Thus, we only try
to instantiate such an axiom when either PNoExit [c, f ] or PNoExit [¬c, f ] has been proven.
5.5. Implementation. The algorithm presented here was implemented using a Perl script and
the S PASS theorem prover [WGR96] and used successfully to verify the example programs of Section 5.1 and Section 5.6.
The method described above can be optimized. For instance, if CA has already been added to
the axioms, we do not try to prove PA again. These details are important in practice, but have been
omitted for brevity.
When trying to prove the different premises, S PASS may fail to terminate if the formula that it
is trying to prove is invalid. Thus, we limit the time that S PASS can spend proving each formula. It
is possible that we will fail to acquire useful axioms this way.
5.6. Further Examples. This section shows the code (Fig. 11) and the complete specification of
two additional examples: appending two linked lists, and the mark phase of a simple mark and
sweep garbage collector.
22
T. LEV-AMI, N. IMMERMAN, T. REPS, M. SAGIV, S. SRIVASTAVA, AND G. YORSH
Node append(Node x, Node y) {
[0] Node last = x;
[1] if (last == null)
[2]
return y;
[3] while (last.next != null) {
[4]
last = last.next;
[5] }
[6] last.next = y;
[7] return x;
}
(a)
void mark(NodeSet root, NodeSet marked) {
[0] Node x;
[1] if(!root.isEmpty()){
[2]
NodeSet pending = new NodeSet();
[3]
pending.addAll(root);
[4]
marked.clear();
[5]
while (!pending.isEmpty()) {
[6]
x = pending.selectAndRemove();
[7]
marked.add(x);
[8]
if (x.car != null &&
[9]
!marked.contains(x.car))
[10]
pending.add(x.car);
[11]
if (x.cdr != null &&
[12]
!marked.contains(x.cdr))
[13]
pending.add(x.cdr);
}
}
}
(b)
Figure 11: A simple Java-like implementation of (a) the concatenation procedure for two singlylinked lists; (b) the mark phase of a mark-and-sweep garbage collector.
5.6.1. Specification of append. The specification of append (see Fig. 11(a)) is given in Fig. 12. The
specification includes procedure’s pre-condition, a transformer of the procedure’s body effect, and
the procedure’s post-condition. The pre-condition (Fig. 12(a)) states that the lists pointed to by
x and y are acyclic, unshared and disjoint. It also states there is no garbage. The post condition
(Fig. 12(b)) states that after the procedure’s execution, the list pointed to by x′ is exactly the union
of the lists pointed to by x and y. Also, the list is still acyclic and unshared. The transformer is given
in Fig. 12(c). The result of the loop in the procedure’s body is summarized as a formula defining
the last variable. The only change to n is the addition of an edge between last and y.
The coloring axioms needed to prove append are given in Fig. 13.
SIMULATING REACHABILITY USING FIRST-ORDER LOGIC
23
def
(a)
pre = acyclic[n] ∧ unshared[n] ∧
unique[x] ∧ unique[y] ∧ f unc[n] ∧
(∀v.¬rx,n (v) ∨ ¬ry,n (v)) ∧ ∀v.rx,n (v) ∨ ry,n (v)
(5.16)
def
(b)
post = acyclic[n′ ] ∧ unshared[n′ ] ∧
unique[x′ ] ∧ unique[last] ∧ f unc[n′ ] ∧
(∀v . rx′ ,n′ (v) ↔ (rx,n (v) ∨ ry,n (v))) ∧
∀v1 , v2 . n′ (v1 , v2 ) ↔ n(v1 , v2 ) ∨ (last(v1 ) ∧ y(v2 )) (5.17)
T is the conjunction of the following formulas:
(c)
∀v.x′ (v)
∀v.last(v)
↔
↔
∃v.
′
∀v1 , v2 .n (v1 , v2 ) ↔
x(v)
rx,n (v) ∧ ∀u.¬n(v, u)
last(v)
n(v1 , v2 ) ∨ (last(v1 ) ∧ y(v2 ))
(5.18)
(5.19)
(5.20)
(5.21)
Figure 12: Example specification of append procedure: (a) precondition pre, (b) postcondition
post, (c) transformer T (effect of the procedure body).
NoExit[ry,n , n′ ]
GoOut[last, y, n′ ]
NewStart[rx,n , n, n′ ] NewStart[rx,n , n′ , n]
NewStart[ry,n , n, n′ ] NewStart[ry,n , n′ , n]
Figure 13: The instances of coloring axioms used in proving append.
5.6.2. Specification of the mark phase. Another example proven is the mark phase of a mark-andsweep sequential garbage collector, shown in Fig. 11(b). The example goes beyond the reverse
example in that it manipulates a general graph and not just a linked list. Furthermore, as far as
we know, ESC/Java [FLL+ 02] was not able prove its correctness because it could not show that
unreachable elements were not marked. Note that the axiom needed to prove this property is NoExit,
which we have shown to be beyond the power of Nelson’s axiomatization.
The loop invariant of mark is given in Fig. 14(a). The first disjunct of the formula holds only in
the first iteration, when only the nodes in root are pending and nothing is marked. The second holds
from the second iteration on. Here, the nodes in root are marked or pending (they start as pending,
and the only way to stop being pending is to become marked). No node is both marked and pending
(because the procedure checks if the node is marked before adding it to pending). All nodes that
are marked or pending are reachable from the root set (we start with only the root nodes as pending,
and after that only nodes that are neighbors of pending nodes became pending; furthermore, only
pending nodes may become marked). There are no edges between marked nodes and nodes that are
neither marked nor pending (because when we mark a node we add all its neighbors to pending,
unless they are marked already). Our method succeeded in proving the loop invariant in Fig. 14(a)
using only the positive axioms.
The post-condition of mark is given in Fig. 14(b). To prove it, we had to use the fact that
there are no edges between marked and unmarked nodes (i.e, there are no pending nodes at the end
24
T. LEV-AMI, N. IMMERMAN, T. REPS, M. SAGIV, S. SRIVASTAVA, AND G. YORSH
(a)
(b)
((∀v . root(v) ↔ pending(v)) ∧
(∀v . ¬ marked(v)))
∨
((∀v . root(v) → marked(v) ∨ pending(v)) ∧
(∀v . ¬pending(v) ∨ ¬marked(v)) ∧
(∀v . pending(v) ∨ marked(v) → rroot,f (v)) ∧
(∀v1 , v2 . marked(v1 ) ∧ ¬marked(v2 ) ∧ ¬pending(v2 )
→ ¬f (v1 , v2 )))
∀v . marked(v) ↔ rroot,f (v)
(5.22)
(5.23)
(5.24)
(5.25)
(5.26)
(5.27)
(5.28)
Figure 14: Example specification of mark procedure: (a) The loop invariant of mark, (b) The postcondition of mark.
of the loop). Thus, we instantiate the axiom NoExit[marked, f ], and this is enough to prove the
post-condition.
6. A PPLICABILITY
OF THE
C OLORING A XIOMS
The coloring axioms are applicable to a wide variety of verification problems. To demonstrate
this, we describe the reasoning done by the TVLA system and how it can be simulated using the
coloring axioms. TVLA is based on the theory of abstract interpretation [CC79] and specifically on
canonical abstraction [SRW02]. TVLA has been successfully used to analyze a large verity of small
but intricate heap manipulating programs (see e.g., [LAS00, BLARS07]), including the verification
of several algorithms (see e.g., [LARSW00, LRS06]). Furthermore, the axioms described in this
paper have been used to integrate S PASS as the reasoning engine behind the TVLA system. The
integrated system is used to perform backward analysis on heap manipulating programs as described
in [LASR07].
In [SRW02], logical structures are used to represent the concrete stores of the program, and
FO(TC) is used to specify the concrete transformers. This provides great flexibility in what programming-language constructs the method can handle. For the purpose of this section, we assume that the
vocabulary used is fixed and always contains equality. Furthermore, we assume that the transformer
cannot change the universe of the concrete store. Allocation and deallocation can be easily modeled
by using a designated unary predicate that holds for the allocated heap cells. Similarly, we assume
that the universe of the concrete store is non-empty. Abstract stores are represented as finite 3valued logical structures. We shall explain the meaning of a structure S by describing the formula
γ
b(S) to which it corresponds.
The individuals of a 3-valued logical structure are called abstract nodes. We use an auxiliary
unary predicate for each abstract node to capture the concrete nodes that are mapped to it. For an
abstract structure with universe {node1 , . . . , noden }, let {a1 , . . . an } be the corresponding unary
predicates.
For each k-ary predicate p in the vocabulary, each k-tuple hnode1 , . . . , nodek i in the abstract
structure (called an abstract tuple) can have one of the following truth values {0, 1, 21 } as follows:
SIMULATING REACHABILITY USING FIRST-ORDER LOGIC
25
• The truth value 1 means that the predicate p universally holds for all of the concrete tuples mapped
to this abstract tuple, i.e.,
∀v1 , . . . , vk . a1 (v1 ) ∧ . . . ∧ ak (vk ) → p(v1 , . . . , vk )
(6.1)
• The truth value 0 means that the predicate p universally does not hold, for all of the concrete
tuples mapped to this abstract tuple, i.e.,
∀v1 , . . . , vk . a1 (v1 ) ∧ . . . ∧ ak (vk ) → ¬p(v1 , . . . , vk )
(6.2)
1
2
• The truth value means that we have no information about this abstract tuple, and thus the value
of the predicate p is not restricted.
We use a designated set of unary predicates called abstraction predicates to control the distinctions among concrete nodes that can be made in an abstract element, which also places a bound
on the size of abstract elements. For each abstract node nodei , Ai denotes the set of abstraction
predicates for which nodei has the truth value 1, and Ai denotes the set of abstraction predicates
for which nodei has the truth value 0. Every pair nodei , nodej of different abstract nodes either
we require that the abstract nodes in the structure represent
Ai ∩Aj 6= ∅ or Ai ∩Aj 6= ∅. In addition,
W
all the concrete nodes, i.e., ∀v . i ai (v). Thus, the abstract nodes form a bounded partition of the
concrete nodes. Finally, each node must represent at least one concrete node, i.e., ∃v . ai (v).
The vocabulary may contain additional predicates called derived predicates, which are explicitly defined from other predicates using a formula in FO(TC). These derived predicates help
the precision of the analysis by recording correlations not captured by the universal information.
Some of the unary derived predicates may also be abstraction predicates, and thus can induce finergranularity abstract nodes.
We say that S1 ⊑ S2 if there is a total mapping m between the abstract nodes of S1 and
the abstract nodes of S2 such that S2 represents all of the concrete stores that S1 represents when
considering each abstract node of S2 as a union of the abstract nodes of S1 mapped to it by m.
Formally, b
γ (S1 ) ∧ ψm → γ
b(S2 ) where
^
ψm =
∀v . ai (v) → a′j (v)
nodei ∈ S1
m(nodei ) = node′j
The order is extended to sets using the induced Hoare order (i.e., XS1 ⊑ XS2 if for each element
S1 ∈ XS1 there exists an element S2 ∈ XS2 such that S1 ⊑ S2 ).
In the original TVLA implementation [LAS00] the abstract transformer is computed by a three
step process:
• First, a heuristic is used to perform case splits by refining the partition induced by the abstraction
predicates. This process is called Focus.
• Second, the formulas comprising the concrete transformer are used to conservatively approximate
the effect of the concrete transformer on all the represented memory states. Update formulas are
either handwritten or derived using finite differencing [RSL03].
• Third, a constraint solver called Coerce is used to improve the precision of the abstract element
by taking advantage of the inter-dependencies between the predicates dictated by the defining
formulas of the derived predicates and constraints of the programming language semantics.
Most of the logical reasoning performed by TVLA is first order in nature. The transitive-closure
reasoning is comprised of three parts:
(1) The update formulas for derived predicates based on transitive closure use first-order formulas
to update the transitive-closure relation, as explained in Section 6.1.
26
T. LEV-AMI, N. IMMERMAN, T. REPS, M. SAGIV, S. SRIVASTAVA, AND G. YORSH
(2) The Coerce procedure relates the definition of the edge relation with its transitive closure by
performing Kleene evaluation (see below).
(3) Handwritten axioms are given to Coerce to allow additional transitive-closure reasoning. They
are usually written once and for all per data-structure analyzed by the system.
To compare the transitive-closure reasoning of TVLA and the coloring axioms presented in
this paper, we concentrate on programs that manipulate singly-linked lists and trees, although the
basic argument holds for other data-structures analyzed by TVLA as well. The handwritten axioms
used by TVLA for these cases are all covered by the axioms described in Section 3.2. The issue
of update formulas is covered in detail in Section 6.1. A detailed description of Kleene evaluation
is beyond the scope of this paper and can be found in [SRW02]. Kleene evaluation of transitive
closure is equivalent to applying transitivity to infer the existence of paths, and finding a subset of
the partition that has no outgoing edges to infer the absence of paths. The latter is equivalent to
applying the NoExit axiom on the formula that defines the appropriate partition.
6.1. Precise Update. Maintenance of transitive closure through updates in the underlying relation
is required for the verification of heap-manipulating programs. In general, it is not possible to
update transitive closure for arbitrary change using first-order-logic formulas. Instead, we limit
the discussion to unit changes (i.e., the addition or removal of a single edge). Work in descriptive
dynamic complexity [PI97, Hes03] and database theory [DS95] gives first-order update formulas to
unit changes in several classes of graphs, including functional graphs and acyclic graphs.
We demonstrate the applicability of the proposed axiom schemes by showing how they can be
used to prove the precise update formula for unit changes in several classes of graphs.
6.1.1. Edge addition. We refer to the edge relation before the update by e and the edge relation
after the update by e′ . Adding an edge from s to t can be formulated as
∀v1 , v2 . e′ (v1 , v2 ) ↔ (e(v1 , v2 ) ∨ (s(v1 ) ∧ t(v2 ))).
The precise update formula for this change is
∃vs , vt . s(vs ) ∧ t(vt ) ∧ ∀v1 , v2 . e′tc (v1 , v2 ) ↔ (etc (v1 , v2 ) ∨ (etc (v1 , vs ) ∧ etc (vt , v2 )))
We have used S PASS to prove the validity of this update formula using the color axioms described in this paper. The basic colors needed are rt,e , i.e., forward reachability from the target of
−
the new edge, and rs,←
e , i.e., backward reachability from the source of the new edge. The axioms
instantiated in the proof are given in Table 6(a).
6.1.2. Edge removal. There is no known precise formula for updating the transitive closure of a
general graph. For general acyclic graphs, Dong and Su [DS95] give a precise update formula
that is beyond the scope of this work. For functional graphs, Hesse [Hes03] gives precise update
formulas based on either an auxiliary binary relation, or by using a ternary relation to describe paths
in the graph that pass through each node. Without these additions, it is not possible to give precise
update formulas in the presence of cyclicity.
When limiting the discussion to acyclic graphs in which between any two nodes there is at most
one path (such as acyclic functional graphs and trees) it is possible to give a simple precise update
formula. As before, let s be the source of the edge to be removed and t be the target of the edge.
The formula for removing an edge is
∀v1 , v2 . e′ (v1 , v2 ) ↔ (e(v1 , v2 ) ∧ ¬(s(v1 ) ∧ t(v2 ))).
SIMULATING REACHABILITY USING FIRST-ORDER LOGIC
NewStart[true, e, e′ ]
′
−
NewStart[rt,e ∧ ¬rs,←
e , e , e]
′
−
NewStart[¬rt,e ∧ rs,←
e , e , e]
′
NewStart[¬rt,e , e , e]
′
−
NewStart[¬rs,←
e , e , e]
′
−
NoExit[¬rs,←
e ,e ]
′
NoExit[rt,e , e ]
(a)
NewStart[true, e′ , e]
NewStart[rt,e , e, e′ ]
′
−
NewStart[rs,←
e , e, e ]
NewStart[¬rt,e , e, e′ ]
′
−
NewStart[¬rs,←
e , e, e ]
′
−
NoExit[rs,←
e ,e ]
(b)
27
NewStart[true, e′ , e]
NewStart[rt,e , e, e′ ]
′
−
NewStart[rs,←
e , e, e ]
NewStart[¬rt,e , e, e′ ]
′
−
NewStart[¬rs,←
e , e, e ]
′
NoExit[¬rt,e , e ]
(c)
Table 6: Axioms instantiated for the proof of the precise update formula of: (a) adding an edge to
a general graph, (b) removing an edge from an acyclic functional graph, and (c) removing
an edge from a tree.
The precise update formula for this change is
∃vs , vt . s(vs ) ∧ t(vt ) ∧ ∀v1 , v2 . e′tc (v1 , v2 ) ↔ (etc (v1 , v2 ) ∧ ¬(etc (v1 , vs ) ∧ etc (vt , v2 ))).
We have used S PASS to prove the validity of this update formula for the case of acyclic func−
tional graphs and the case of trees. As in edge addition, rt,e and rs,←
e are used as the basic colors.
The axioms instantiated in the proof are given in Table 6(b) and Table 6(c).
7. R ELATED W ORK
Shape Analysis. This work was motivated by our experience with TVLA [LAS00, SRW02],
which is a generic system for abstract interpretation [CC77]. The TVLA system is more automatic
than the methods described in this paper since it does not rely on user-supplied loop invariants.
However, the techniques presented in the present paper are potentially more precise due to the use
of full first-order reasoning. It can be shown that the NoExit scheme allows us to infer reachability at
least as precisely as evaluation rules for 3-valued logic with Kleene semantics. In the future, we hope
to develop an efficient non-interactive theorem prover that enjoys the benefits of both approaches.
An interesting observation is that the colors needed in our examples to prove the formula are the
same unary predicates used by TVLA to define its abstraction. This similarity may, in the future,
help us find better ways to automatically instantiate the required axioms. In particular, inductive
logic programming has recently been used to learn formulas to use in TVLA abstractions [LRS05],
which holds out the possibility of applying similar methods to further automate the approach of the
present paper.
Decidable Logics. Decidable logics can be employed to define properties of linked data structures: Weak monadic second-order logic has been used in [EMS00, MS01] to define properties of
heap-allocated data structures, and to conduct Hoare-style verification using programmer-supplied
loop invariants in the PALE system [MS01]. A decidable logic called Lr (for “logic of reachability
expressions”) was defined in [BRS99]. Lr is rich enough to express the shape descriptors studied
in [SRW98] and the path matrices introduced in [Hen90]. More recent decidable logics include
Logic of Reachable Patterns [YRS+ 06] and a decision procedure for linked data structures that can
handle singly linked lists [BR06].
The present paper does not develop decision procedures, but instead suggests methods that can
be used in conjunction with existing theorem provers. Thus, the techniques are incomplete and
the theorem provers need not terminate. However, our initial experience is that the extra flexibility
gained by the use of first-order logic with transitive closure is promising. For example, we can prove
28
T. LEV-AMI, N. IMMERMAN, T. REPS, M. SAGIV, S. SRIVASTAVA, AND G. YORSH
the correctness of imperative destructive list-reversal specified in a natural way and the correctness
of mark and sweep garbage collectors, which are beyond the scope of Mona and Lr .
Indeed, in [IRR+ 04b], we have tried to simulate existing data structures using decidable logics
and realized that this can be tricky because the programmer may need to prove a specific simulation
invariant for a given program. Giving an inaccurate simulation invariant causes the simulation to be
unsound. One of the advantages of the technique described in the present paper is that soundness is
guaranteed no matter which axioms are instantiated. Moreover, the simulation requirements are not
necessarily expressible in the decidable logic.
Other First-Order Axiomatizations of Linked Data Structures. The closest approach to
ours that we are aware of was taken by Nelson as we describe in Section 4. This also has some
follow-up work by Leino and Joshi [Lei98]. Our impression from their write-up is that Leino and
Joshi’s work can be pushed forward by using our coloring axioms.
A more recent work by Lahiri and Qadeer [LQ06] uses first-order axiomatization. This work
can be seen as a specialization of ours to the case of (cyclic) singly linked lists.
Dynamic Maintenance of Transitive Closure. Another orthogonal but promising approach to
transitive closure is to maintain reachability relations incrementally as we make unit changes in the
data structure. It is known that in many cases, reachability can be maintained by first-order formulas
[DS95, PI97] and even sometimes by quantifier-free formulas [Hes03]. Furthermore, in these cases,
it is often possible to automatically derive the first-order update formulas using finite differencing
[RSL03].
8. C ONCLUSION
This paper reports on our proposal of a new methodology for using off-the-shelf first-order
theorem provers to reason about reachability in programs. We have explored many of the theoretical
issues as well as presenting examples that, while still preliminary, suggest that this is indeed a viable
approach.
As mentioned earlier, proving the absence of paths is the difficult part of proving formulas
with TC. The promise of our approach is that it is able to handle such formulas effectively and
reasonably automatically, as shown by the fact that it can successfully handle the programs described
in Section 5 and the success of the TVLA system, which uses similar transitive-closure reasoning.
Of course, much further work is needed including the following:
• Exploring other heuristics for identifying color classes.
• Exploring variations of the algorithm given in Fig. 10 for instantiating coloring axioms.
• Exploring the use of additional axiom schemes, such as two of the schemes from [Nel83], which
are likely to be useful when dealing with predicates that are partial functions. Such predicates
arise in programs that manipulate singly-linked or doubly-linked lists—or, more generally, data
structures that are acyclic in one or more “dimensions” [HHN92] (i.e., in which the iterated
application of a given field selector can never return to a previously visited node).
• Additional work should be done on the theoretical power of T1 +IND and related axiomatizations
of transitive closure. We conjecture, for example, that T1 + IND is TC-complete for trees.
Acknowledgements. Thanks to Aharon Abadi and Roman Manevich for interesting suggestions.
Thanks to Viktor Kuncak for useful conversations including his observation and proof of Proposition 4.4.
SIMULATING REACHABILITY USING FIRST-ORDER LOGIC
29
R EFERENCES
[Avr03]
A. Avron. Transitive closure and the mechanization of mathematics. In Thirty Five Years of Automating
Mathematics, pages 149–171. Kluwer Academic Publishers, 2003.
[BLARS07] I. Bogudlov, T. Lev-Ami, T. Reps, and M. Sagiv. Revamping tvla: Making parametric shape analysis
competitive. In CAV, 2007.
[BR06]
J. Bingham and Z. Rakamaric. A logic and decision procedure for predicate abstraction of heapmanipulating programs. In VMCAI, pages 207–221, 2006.
[BRS99]
M. Benedikt, T. Reps, and M. Sagiv. A decidable logic for describing linked data structures. In European
Symp. On Programming, pages 2–19, March 1999.
[CC77]
Patrick Cousot and Radhia Cousot. Abstract interpretation: a unified lattice model for static analysis of programs by construction or approximation of fixpoints. In POPL ’77: Proceedings of the 4th ACM SIGACTSIGPLAN symposium on Principles of programming languages, pages 238–252. ACM Press, 1977.
[CC79]
P. Cousot and R. Cousot. Systematic design of program analysis frameworks. In Symp. on Princ. of Prog.
Lang., pages 269–282, New York, NY, 1979. ACM Press.
[DS95]
G. Dong and J. Su. Incremental and decremental evaluation of transitive closure by first-order queries. Inf.
& Comput., 120:101–106, 1995.
[EMS00]
J. Elgaard, A. Møller, and M.I. Schwartzbach. Compile-time debugging of C programs working on trees.
In European Symp. On Programming, pages 119–134, 2000.
[FLL+ 02]
C. Flanagan, K.R.M. Leino, M. Lillibridge, G. Nelson, J.B. Saxe, and R. Stata. Extended static checking
for java. In SIGPLAN Conf. on Prog. Lang. Design and Impl., 2002.
[GME99]
E. Grädel, M.Otto, and E.Rosen. Undecidability results on two-variable logics. Archive of Math. Logic,
38:313–354, 1999.
[Hen90]
L. Hendren. Parallelizing Programs with Recursive Data Structures. PhD thesis, Cornell Univ., Ithaca, NY,
Jan 1990.
[Hes03]
W. Hesse. Dynamic Computational Complexity. PhD thesis, Department of Computer Science, UMass,
Amherst, July 2003.
[HHN92]
L. Hendren, J. Hummel, and A. Nicolau. Abstractions for recursive pointer data structures: Improving the
analysis and the transformation of imperative programs. In SIGPLAN Conf. on Prog. Lang. Design and
Impl., pages 249–260, New York, NY, June 1992. ACM Press.
[Hoa75]
C.A.R. Hoare. Recursive data structures. Int. J. of Comp. and Inf. Sci., 4(2):105–132, 1975.
[IRR+ 04a] N. Immerman, A. Rabinovich, T. Reps, M. Sagiv, and G. Yorsh. The boundary between decidability and
undecidability of transitive closure logics. In CSL’04, 2004.
[IRR+ 04b] N. Immerman, A. Rabinovich, T. Reps, M. Sagiv, and G. Yorsh. Verification via structure simulation. In
Proc. Computer-Aided Verif., pages 281–294, 2004.
[LARSW00] T. Lev-Ami, T. Reps, M. Sagiv, and R. Wilhelm. Putting static analysis to work for verification: A case
study. In ISSTA 2000: Proc. of the Int. Symp. on Software Testing and Analysis, pages 26–38, 2000.
[LAS00]
T. Lev-Ami and M. Sagiv. TVLA: A system for implementing static analyses. In Static Analysis Symp.,
pages 280–301, 2000.
[LASR07] T. Lev-Ami, M. Sagiv, and T. Reps. Backward analysis for inferring quantified preconditions. Submitted
for publication, 2007.
[Lei98]
R. Leino. Recursive object types in a logic of object-oriented programs. Nordic J. of Computing, 5:330–360,
1998.
[LQ06]
S. K. Lahiri and S. Qadeer. Verifying properties of well-founded linked lists. In POPL, pages 115–126,
2006.
[LRS05]
A. Loginov, T. Reps, and M. Sagiv. Abstraction refinement via inductive learning. In Proc. Computer-Aided
Verif., 2005.
[LRS06]
A. Loginov, T. Reps, and M. Sagiv. Automatic verification of the Deutsch-Schorr-Waite tree-traversal algorithm. In SAS, 2006.
[MP71]
R. McNaughton and S. Papert. Counter-Free Automata. MIT Press, 1971.
[MS01]
A. Møller and M.I. Schwartzbach. The pointer assertion logic engine. In SIGPLAN Conf. on Prog. Lang.
Design and Impl., pages 221–231, 2001.
[Nel83]
G. Nelson. Verifying reachability invariants of linked structures. In Symp. on Princ. of Prog. Lang., pages
38–47, 1983.
[PI97]
S. Patnaik and N. Immerman. Dyn-FO: A parallel, dynamic complexity class. Journal of Computer and
System Sciences, 55(2):199–209, October 1997.
30
[RSL03]
[RSW04]
[SRW98]
[SRW02]
[WGR96]
[YRS+ 06]
T. LEV-AMI, N. IMMERMAN, T. REPS, M. SAGIV, S. SRIVASTAVA, AND G. YORSH
T. Reps, M. Sagiv, and A. Loginov. Finite differencing of logical formulas for static analysis. In European
Symp. On Programming, pages 380–398, 2003.
T. Reps, M. Sagiv, and R. Wilhelm. Static program analysis via 3-valued logic. In CAV, pages 15–30, 2004.
M. Sagiv, T. Reps, and R. Wilhelm. Solving shape-analysis problems in languages with destructive updating. Trans. on Prog. Lang. and Syst., 20(1):1–50, January 1998.
M. Sagiv, T. Reps, and R. Wilhelm. Parametric shape analysis via 3-valued logic. Trans. on Prog. Lang.
and Syst., 2002.
Christoph Weidenbach, Bernd Gaede, and Georg Rock. Spass & flotter version 0.42. In CADE-13: Proceedings of the 13th International Conference on Automated Deduction, pages 141–145. Springer-Verlag,
1996.
G. Yorsh, A. Rabinovich, M. Sagiv, A. Meyer, and A. Bouajjani. A logic of reachable patterns in linked
data-structures. In FOSSACS, 2006.
This work is licensed under the Creative Commons Attribution-NoDerivs License. To view
a copy of this license, visit http://creativecommons.org/licenses/by-nd/2.0/ or
send a letter to Creative Commons, 171 Second St, Suite 300, San Francisco, CA 94105,
USA, or Eisenacher Strasse 2, 10777 Berlin, Germany
| 6cs.PL
|
1
3D Object Dense Reconstruction
from a Single Depth View
arXiv:1802.00411v1 [cs.CV] 1 Feb 2018
Bo Yang, Stefano Rosa, Andrew Markham, Niki Trigoni, Hongkai Wen*
Abstract—In this paper, we propose a novel approach, 3D-RecGAN++, which reconstructs the complete 3D structure of a given object
from a single arbitrary depth view using generative adversarial networks. Unlike existing work which typically requires multiple views of
the same object or class labels to recover the full 3D geometry, the proposed 3D-RecGAN++ only takes the voxel grid representation of a
depth view of the object as input, and is able to generate the complete 3D occupancy grid with a high resolution of 2563 by recovering
the occluded/missing regions. The key idea is to combine the generative capabilities of autoencoders and the conditional Generative
Adversarial Networks (GAN) framework, to infer accurate and fine-grained 3D structures of objects in high-dimensional voxel space.
Extensive experiments on large synthetic datasets and real-world Kinect datasets show that the proposed 3D-RecGAN++ significantly
outperforms the state of the art in single view 3D object reconstruction, and is able to reconstruct unseen types of objects.
Index Terms—3D Reconstruction, Shape Completion, Shape inpainting, Single Depth View, Adversarial Learning, Conditional GAN.
F
1
I NTRODUCTION
T
O reconstruct the complete and precise 3D geometry of
an object is essential for many graphics and robotics
applications, from AR/VR [1] and semantic understanding, to robot grasping [2] and obstacle avoidance. Classic
approaches use the off-the-shelf low-cost depth sensing
devices such as Kinect and RealSense cameras to recover
the 3D shape of an object from captured depth images.
Those approaches typically require multiple depth images
from different viewing angles of an object to estimate the
complete 3D structure [3] [4] [5]. However, in practice it is
not always feasible to scan all surfaces of an object before
reconstruction, which leads to incomplete 3D shapes with
occluded regions and large holes. In addition, acquiring and
processing multiple depth views require more computing
power, which is not ideal in many applications that require
real-time performance.
In this paper, we aim to tackle the problem of estimating
the complete 3D structure of an object using a single depth
view. This is a very challenging task, since the partial
observation of the object (i.e. a depth image from one viewing angle) can be theoretically associated with an infinite
number of possible 3D models. Traditional reconstruction
approaches typically use interpolation techniques such as
plane fitting, Laplacian hole filling [6] [7], or Poisson surface
estimation [8] [9] to infer the underlying 3D structure. However, they can only recover very limited occluded or missing
regions, e.g. small holes or gaps due to quantization artifacts, sensor noise and insufficient geometry information.
Interestingly, humans are surprisingly good at solving
such ambiguity by implicitly leveraging prior knowledge.
•
•
Bo Yang, Stefano Rosa, Andrew Markham and Niki Trigoni are with the
Department of Computer Science, University of Oxford, UK.
E-mail:{bo.yang,stefano.rosa,andrew.markham,niki.trigoni}@cs.ox.ac.uk
Corresponding to Hongkai Wen, who is with the Department of Computer
Science, University of Warwick, UK.
E-mail: [email protected]
For example, given a view of a chair with two rear legs
occluded by front legs, humans are easily able to guess the
most likely shape behind the visible parts. Recent advances
in deep neural networks and data driven approaches show
promising results in dealing with such a task.
In this paper, we aim to acquire the complete and highresolution 3D shape of an object given a single depth view.
By utilizing the high performance of 3D convolutional neural nets and large open datasets of 3D models, our approach
learns a smooth function to map a 2.5D view to a complete
and dense 3D shape. In particular, we train an end-to-end
model which estimates full volumetric occupancy from a
single 2.5D depth view of an object.
While state-of-the-art deep learning approaches [10] [11]
[2] for 3D shape reconstruction from a single depth view
achieve encouraging results, they are limited to very small
resolutions, typically at the scale of 323 voxel grids. As
a result, the learnt 3D structure tends to be coarse and
inaccurate. However, to increase the 3D shape resolution
without sacrificing recovery accuracy is challenging, as even
a slightly higher resolution would exponentially increase
the search space of potential 2.5D to 3D mapping functions,
resulting in difficulties in convergence of neural nets.
Recently, deep generative models achieve impressive
success in modeling complex high-dimensional data distributions, among which Generative Adversarial Networks
(GANs) [12] and Variational Autoencoders (VAEs) [13]
emerge as two powerful frameworks for generative learning, including image and text generation [14] [15], and latent
space learning [16] [17]. In the past few years, a number of
works [18] [19] [20] [21] applied such generative models to
learn latent space to represent 3D object shapes, in order to
solve simple discriminative tasks such as new image generation, object classification, recognition and shape retrieval.
In this paper, we propose 3D-RecGAN++, a novel model
that combines a skip-connected 3D autoencoder with adversarial learning to generate a complete and fine-grained 3D
2
structure conditioned on a single 2.5D view. Particularly, our
model firstly encodes the 2.5D view to a low-dimensional
latent space vector which implicitly represents general 3D
geometric structures, then decodes it back to recover the
most likely full 3D shape. The rough 3D shape is then
fed into a conditional discriminator which is adversarially
trained to distinguish whether the coarse 3D structure is
plausible or not. The autoencoder is able to approximate the
corresponding shape, while the adversarial training tends
to add fine details to the estimated shape. To ensure the
final generated 3D shape corresponds to the input single
partial 2.5D view, adversarial training of our model is based
on a conditional GAN [22] instead of random guessing.
The above novel and efficient network design excels the
competing approaches [2] [11] [23], which either use a single
fully connected layer [2], a low capacity decoder [11], or the
multi-stage and inefficient LSTMs [23] to estimate the full
3D shapes.
Our contributions are as follows:
(1) We propose a novel generative model to reconstruct
the complete and accurate 3D structure using a single arbitrary depth view. Particularly, our model takes a simple
occupancy grid map as input without requiring object class
labels or any annotations, while predicting a compelling
shape within a high resolution of 2563 voxel grid. By
drawing on both autoencoder and GAN, our approach is
end-to-end trainable with high level of generality. To the
best of our knowledge, this is the first work that reconstructs
such high resolution of 3D shapes using a single view.
(2) We exploit conditional GAN during training to refine
the 3D shape estimated by the autoencoder. Our contribution here is that we use the mean value of a latent vector
feature, instead of a single scalar, as the output of the
discriminator to stabilize GAN training.
(3) We conduct extensive experiments for single category
and multi-category object reconstruction, outperforming the
state of the art. Importantly, our approach is also able to
generalize to previously unseen object categories. At last,
our model also performances robustly on real-world dataset
collected by Kinect, after being trained purely on synthetic
datasets.
(4) To the best of our knowledge, there are no good open
datasets which have the ground truth for occluded/missing
parts and holes for each 2.5D view in real world scenarios.
We therefore contribute our real world testing dataset to the
community.
A preliminary version of this work has been published in
ICCV 2017 workshops [24]. Our code and data are available
at: https://github.com/Yang7879/3D-RecGAN-extended
2
R ELATED W ORK
We review different pipelines for 3D reconstruction or shape
completion. Both conventional geometry based techniques
and the state of the art deep learning based approaches are
covered.
(1) 3D Model/Shape Completion. [25] uses plane fitting
to complete small missing regions, while [26] [27] [28]
[29] [30] apply shape symmetry to fill in holes. Although
these methods show good results, relying on predefined
geometric regularities fundamentally limits the structure
space to hand-crafted shapes. Besides, these approaches are
likely to fail when missing or occluded regions are relatively
big. Another similar fitting pipeline is to leverage database
priors. Given a partial shape input, [31] [32] [33] [34] [35]
[36] try to retrieve an identical or most likely 3D model and
align it with the partial scan. However, these approaches
explicitly assume the database contains identical or very
similar shapes, thus being unable to generalize to novel
objects or categories.
(2) Multiple RGB/Depth Images Reconstruction. Traditionally, 3D dense reconstruction requires a collection of
RGB images [37]. Geometric shape is recovered by dense
feature extraction and matching [38], or by directly minimizing reprojection errors [39] from color images. Recently, [40]
[41] [42] [43] [44] [45] leverage deep neural nets to learn the
3D shape from multiple RGB images. However, resolution
of the recovered occupancy shape is usually up to a small
scale of 323 . With the advancement of depth sensors, depth
images are also used to recover the object shape. Classic
approaches usually fuse multiple depth images through
iterative closest point (ICP) algorithms [3] [46] [47], while
recent work [48] learns the 3D shape using deep neural nets
from multiple depth views.
(3) Single RGB Image Reconstruction. Predicting a
complete 3D object model from a single view is a longstanding and extremely challenging task. When reconstructing a specific object category, model templates can be used.
For example, morphable 3D models are exploited for face
recovery [49] [50]. This concept was extended to reconstruct
simple objects in [51]. For general and complex object reconstruction from a single RGB image, recent works [52] [53]
[54] aim to infer 3D shapes using multiple RGB images for
weak supervision. However, the training procedure of [54]
is two stage rather than end-to-end, while [53] uses a simple autoencoder instead of designing sophisticated learning frameworks for shape learning, and [52] still requires
3D shape priors for constraints. Shape prior knowedge is
also required in [55] [56] [57]. To recover high resolution
3D shapes, [58] [59] use Octree representation, while [60]
proposed an inverse discrete cosine transform (IDCT) technique. Lin et al. [61] designed a pseudo-renderer to predict
dense 3D shapes, while [62] sequentially estimates 2.5D
sketches and dense 3D shapes from a single RGB image.
(4) Single Depth View Reconstruction. The task of
reconstruction from a single depth view is to complete
the occluded 3D structures behind the visible parts. 3D
ShapeNets [10] is among the early work using deep neural
nets to estimate 3D shapes from a single depth view. Firman
et al. [63] trained a random decision forest to infer unknown
voxels. Originally designed for shape denoising, VConvDAE [1] can also be used for shape completion. To facilitate
robotic grasping, Varley et al. proposed a neural network
to infer the full 3D shape from a single depth view in [2].
However, all these approaches are only able to generate low
resolution voxel grids which are less than 403 and unlikely
to capture fine geometric details. Recent works [11] [64] [23]
[65] can infer higher resolution 3D shapes. However, 3DEPN [11] relies on a shape database to synthesize higher
resolution shapes after learning a small 323 voxel grid from
a depth view, while SSCNet [64] requires strong voxel-level
annotations for supervised scene completion and semantic
3
true full 3D shape
concat
up-sampling
y
concat
U-net
loss
real
reconstruction
decoder
x encoder
fake
reconstruction
input
2.5D view
conditional
discriminator
Fig. 2: Overview of the network architecture for training.
Fig. 1: t-SNE embeddings of 2.5D partial views and 3D
complete shapes of multiple object categories.
label prediction. Both [23] and [65] are originally designed
for shape inpainting instead of directly reconstructing the
complete 3D structure from a partial depth view. The recent 3D-PRNN [66] predicts simple shape primitives using
RNNs, but the estimated shapes do not have finer geometric
details.
(5) Deep Generative Frameworks. Deep generative
frameworks, such as VAEs [13] and GANs [12], have
achieved impressive success in image super-resolution [67],
image generation [15], text to image synthesis [68], etc.
Recently, [69] [70] [71] [21] applied generative networks for
3D structure generation. However, most of them generate
3D shapes from random noise instead of reconstructing
structures from a specific single image.
3
3.1
3D-R EC GAN++
Overview
Our method aims to estimate a complete and dense 3D
structure of an object, which only takes an arbitrary single
2.5D depth view as input. The output 3D shape is automatically aligned with the corresponding 2.5D partial view.
To achieve this task, each object model is represented in a
high resolution 3D voxel grid. We use the simple occupancy
grid for shape encoding, where 1 represents an occupied cell
and 0 an empty cell. Specifically, the input 2.5D partial view,
denoted as x, is a 643 occupancy grid, while the output 3D
shape, denoted as y , is a high resolution 2563 probabilistic
voxel grid. The input partial shape is directly calculated
from a single depth image given camera parameters. We use
the ground truth dense 3D shape with aligned orientation
as same as the input partial 2.5D depth view to supervise
our network.
To generate ground truth training and evaluation pairs,
we virtually scan 3D objects from ShapeNet [72]. Figure 1
is the t-SNE visualization [73] of partial 2.5D views and the
corresponding full 3D shapes for multiple general chair and
bed models. Each green dot represents the t-SNE embedding
of a 2.5D view, whilst a red dot is the embedding of the corresponding 3D shape. It can be seen that multiple categories
inherently have similar 2.5D to 3D mapping relationships.
Essentially, our neural network is to learn a smooth function,
denoted as f , which maps green dots to red dots as close as
possible in high dimensional space as shown in Equation 1.
The function f is parametrized by convolutional layers in
general.
3
y = f (x)
x ∈ Z 64 , where Z = {0, 1}
(1)
input
2.5D view
x encoder
decoder
up-sampling
y
U-net
Fig. 3: Overview of the network architecture for testing.
After generating training pairs, we feed them into our
network. The first part of our network loosely follows the
idea of an autoencoder with the U-net architecture [74].
The skip-connected autoencoder serves as an initial coarse
generator which is followed by an up-sampling module to
further generate a high resolution 3D shape within a 2563
voxel grid. This whole generator aims to learn a correlation
between partial and complete 3D structures. With the supervision of complete 3D labels, the generator is able to learn
a function f and infer a reasonable 3D shape given a brand
new partial 2.5D view. In the testing phase, however, the
results tend to be grainy and without fine details.
To address this issue, in the training phase, the reconstructed 3D shape from the generator is further fed
into a conditional discriminator to verify its plausibility.
In particular, a partial 2.5D input view is paired with its
corresponding complete 3D shape, which is called the ‘real
reconstruction’, while the partial 2.5D view is paired with
its corresponding output 3D shape from generator, which
is called the ‘fake reconstruction’. The discriminator aims
to discriminate all ‘fake reconstruction’ from ‘real reconstruction’. In the original GAN framework [12], the task
of the discriminator is to simply classify real and fake input, but its Jensen-Shannon divergence-based loss function
is difficult to converge. The recent WGAN [75] leverages
Wasserstein distance with weight clipping as a loss function
to stabilize the training procedure, whilst the extended work
WGAN-GP [76] further improves the training process using
a gradient penalty with respect to its input. In our 3DRecGAN++, we apply WGAN-GP as the loss function of our
conditional discriminator, which guarantees fast and stable
convergence. The overall network architecture for training
is shown in Figure 2, while the testing phase only needs the
well trained generator as shown in Figure 3.
Overall, the main challenge of 3D reconstruction from an
arbitrary single view is to generate new information including filling the missing and occluded regions from unseen
views, while keeping the estimated 3D shape corresponding
to the specific input 2.5D view. In the training phase, our
3D-RecGAN++ firstly leverages a skip-connected autoencoder together with an up-sampling module to generate
a reasonable ‘fake reconstruction’ within a high resolution
occupancy grid, then applies adversarial learning to refine
the ‘fake reconstruction’ to make it as similar to ‘real reconstruction’ by jointly updating parameters of the generator.
4
In the testing phase, given a novel 2.5D view as input, the
jointly trained generator is able recover a full 3D shape with
satisfactory accuracy, while the discriminator is no longer
used.
3.2
Architecture
Figure 4 shows the detailed architecture of our proposed 3DRecGAN++. It consists of two main networks: the generator
as in Figure 4a and the discriminator as in Figure 4b.
The generator consists of a skip-connected autoencoder
and an up-sampling module. Unlike the vanilla GAN generator which generates data from arbitrary latent distributions, our 3D-RecGAN++ generator synthesizes data from
latent distributions of 2.5D views. Particularly, the encoder
has five 3D convolutional layers, each of which has a bank of
4x4x4 filters with strides of 1x1x1, followed by a leaky ReLU
activation function and a max pooling layer with 2x2x2
filters and strides of 2x2x2. The number of output channels
of max pooling layer starts with 64, doubling at each subsequent layer and ends up with 512. The encoder is lastly
followed by two fully-connected layers to embed semantic
information into a latent space. The decoder is composed
of 5 symmetric up-convolutional layers which are followed
by ReLU activations. Skip-connections between encoder and
decoder guarantee propagation of local structures of the
input 2.5D view. The skip-connected autoencoder is followed by the up-sampling module which simply consists of
two layers of up-convolutional layers as detailed in Figure
4a. This simple yet efficient up-sampling module directly
upgrades the output 3D shape to a high resolution of 2563
without requiring complex network design and operations.
It should be noted that without the two fully connected
layers and skip-connections, the vanilla autoencoder would
be unable to learn reasonable complete 3D structures as
the latent space is limited and the local structure is not
preserved. Without the efficient up-sampling module, it is
unable to finally generate high resolution 3D shapes. Although a more complicated and dedicated network design
could also output 2563 shapes, it would be unlikely to
be effectively trained on a single GPU because of the extremely high computation consumption for high resolution
3D shape generation. The loss function and optimization
methods are described in Section 3.3.
The discriminator aims to distinguish whether the estimated 3D shapes are plausible or not. Based on the conditional GAN, the discriminator takes both real reconstruction
pairs and fake reconstruction pairs as input. Particularly, it
consists of six 3D convolutional layers, the first of which
concatenates the generated 3D shape (i.e. a 2563 voxel grid)
and the input 2.5D partial view (i.e. a 643 voxel grid) which
is reshaped as a 256x256x4 tensor. The reshaping process
is done straightforwardly using Tensorflow ‘tf.reshape()’.
Basically, this is to inject the condition information with
a matched tensor dimension, and then leave the network
itself to learn useful features from this condition input. Each
convolutional layer has a bank of 4x4x4 filters with strides of
2x2x2, followed by a ReLU activation function except for the
last layer which is followed by a sigmoid activation function. The number of output channels of the convolutional
layers starts with 8, doubling at each subsequent layer and
ends up with 256.
At the early stage of GAN training, the high dimensional real and fake distributions may not overlap, then
the discriminator can separate them perfectly using a single
scalar output, which is theoretically analyzed in [77]. In our
experiments, the original WGAN-GP always crashes in the
early 3 epochs due to the extremely high dimensionality (i.e.
2563 + 643 dimensions). To stabilize it, we propose to use
mean feature (i.e. mean of a vector feature) for discriminator.
As the mean vector feature captures more information from
the input overall, it is more difficult for the discriminator
to easily distinguish whether the mean feature is from
fake or real input. This enables useful information to backpropagate to the generator. A theoretical study of the mean
feature matching method for GAN is in [78]; mean feature
matching is also applied in [79] to stabilize GAN.
Therefore, our discriminator is to distinguish the distributions of mean feature of fake and real reconstructions,
while the generator is trained to make the two distributions
of mean feature as similar as possible. We apply WGAN-GP
as loss functions for our modified mean feature matching.
3.3
Objectives
The objective function of our 3D-RecGAN++ includes two
main parts: an object reconstruction loss `ae for the generator; the objective function `gan for the conditional GAN.
(1) `ae For the generator, inspired by [80], we use modified binary cross-entropy loss function instead of the standard version. The standard binary cross-entropy weights
both false positive and false negative results equally. However, most of the voxel grid tends to be empty, so the network easily gets a false positive estimation. In this regard,
we impose a higher penalty on false positive results than
on false negatives. Particularly, a weight hyper-parameter α
is assigned to false positives, with (1-α) for false negative
results, as shown in Equation 2.
`ae
N
1 X
=
−αy¯i log(yi )−(1−α)(1− y¯i ) log(1−yi ) (2)
N i=1
where y¯i is the target value {0,1} of a specific ith voxel in
the ground truth voxel grid ȳ , and yi is the corresponding
estimated value (0,1) in the same voxel from the autoencoder output y . We calculate the mean loss over the total N
voxels in the whole voxel grid.
(2) `gan For the discriminator, we leverage the state of
the art WGAN-GP loss functions. Unlike the original GAN
loss function which presents an overall loss for both real
and fake inputs, we separately represent the loss function
`ggan in Equation 3 for generating fake reconstruction pairs
and `dgan in Equation 4 for discriminating fake and real
reconstruction pairs. Detailed definitions and derivation of
the loss functions can be found in [75] [76], but we modify
them for our conditional GAN settings.
`ggan = −E D(y|x)
(3)
`dgan = E D(y|x) − E D(ȳ|x)
2
+λE
∇ŷ D(ŷ|x) 2 − 1
(4)
43 conv
lrelu
2 maxpool
43 conv
lrelu
23 maxpool
3
x
43 conv
lrelu
2 maxpool
3
x
643
43 conv
lrelu
3
323 2 maxpool16
3
43 conv
43 conv 3
lrelu 83
lrelu 4
3
512 channels
2 maxpool
256 channels
23 maxpool
128 channels
43 deconv
relu
43 deconv
relu
43 deconv
43 deconv
3
43 deconv
relu
4
3
16
512
channels
relu
32
dense
256 channels relu
relu
128 channels
32768
64 channels
reshape
43 deconv
relu
43 deconv
relu
3
flatten
64 channels
1 channel
43 conv
43 conv
lrelu
lrelu
3
2 maxpool 23 maxpool
dense
dense
relu
relu
2000
dense
32768
flatten
relu
4 deconv
relu
reshape
43
3
43 deconv
relu
3
deconv8
relu
y
643
16 channels
concat
323
643
64 channels
1 channel
163
128 channels
43
43
concat
2000
83
83
163
512 channels 256 channels
256 channels 512 channels
concat
128 channels
32768
32768
323
64 channels
concat
concat
concat
concat
43 deconv
relu
5
43 deconv
relu
1283
y
2563
8 channels
1 channel
up-sampling module
643
16 channels
1283
2563
8 channels
1 channel
up-sampling module
concat
concat
concat
(a) Generator for 3D shape estimation from a single depth view.
y
2563
1 channel
x
reshape
2563
1 channel
x
643
1 channel
256x256x4
reshape
43 conv
lrelu
y
fake
reconstruction
real
fake
reconstruction
reconstruction
43 conv
lrelu
256x256x260
643
real
reconstruction
256x256x4
1 channel
3
4 conv
lrelu
128x128x130
8 channels
43 conv
lrelu
43 conv
lrelu
43 conv
lrelu
43 conv
lrelu
flatten
loss
4x4x5
8x8x943 conv
43 conv
16x16x17
43 conv
43 conv
32x32x33 lrelu 64 channels
128 channels 256 channels
lrelu
lrelu
lrelu
32
channels
flatten
20480
16 channels
loss
64x64x65
1 channel
1 channel
ground truth
1 channel
43 conv
lrelu
256x256x260
128x128x130
8 channels
64x64x65
16 channels
32x32x33
32 channels
16x16x17 8x8x9
64 channels
4x4x5
128 channels 256 channels
20480
1 channel
3
256
1 channel
ground truth
(b) Discriminator for 3D shape refinement.
3
256of 3D-RecGAN++, showing the two main building blocks. Note that, although these are shown
Fig. 4: Detailed architecture
1 channel
as two separate modules, they are trained end-to-end.
where ŷ = ȳ + (1 − )y, ∼ U [0, 1], x is the input partial
depth view, y is the corresponding output of autoencoder,
ȳ is the corresponding ground truth. λ controls the trade-off
between optimizing the gradient penalty and the original
objective in WGAN.
For the generator in our 3D-RecGAN++ network, there
are two loss functions, `ae and `ggan , to optimize. As we
discussed in Section 3.1, minimizing `ae tends to learn the
overall 3D shapes, whilst minimizing `ggan estimates more
plausible 3D structures conditioned on input 2.5D views. To
minimize `dgan is to improve the performance of discriminator to distinguish fake and real reconstruction pairs. To
jointly optimize the generator, we assign weights β to `ae
and (1 − β) to `ggan . Overall, the loss functions for generator
and discriminator are as follows:
3.4
`g = β`ae + (1 − β)`ggan
(5)
`d = `dgan
(6)
Training
We adopt an end-to-end training procedure for the whole
network. To simultaneously optimize both generator and
discriminator, we alternate between one gradient decent
step on discriminator and then one step on generator. For
the WGAN-GP, λ is set as 10 for gradient penalty as in [76]. α
ends up as 0.85 for our modified cross entropy loss function,
while β is 0.2 for the joint loss function `g .
The Adam solver [81] is used for both discriminator and
generator with a batch size of 4. The other three Adam
parameters are set to default values. Learning rate is set
to 0.0001 for the discriminator and 0.0005 for the generator
in all epochs. As we do not use dropout or batch normalization, the testing phase is exactly the same as the training
stage. The whole network is trained on a single Titan X GPU
from scratch.
3.5
Data Synthesis
For the task of 3D dense reconstruction from a single depth
view, obtaining a large amount of training data is an obstacle. Existing real RGB-D datasets for surface reconstruction
suffer from occlusions and missing data and there is no
ground truth of complete and high resolution 2563 3D
shapes for each single view. The recent work 3D-EPN [11]
synthesizes data for 3D object completion, but their map
encoding scheme is the complicated TSDF which is different
from our network requirement.
To tackle this issue, we use the ShapeNet [72] database
to generate a large amount of training and testing data with
synthetically rendered depth images and the corresponding
complete 3D shape ground truth. Particularly, a subset of
object categories is selected for our experiments. For each
category, we generate training data from around 220 CAD
models, while synthesizing testing data from around 40
CAD models. For each CAD model, we create a virtual
depth camera to scan it from 125 different angles, 5 uniformly sampled views for each of roll, pitch and yaw space.
For each virtual scan, both a depth image and the corresponding complete 3D voxelized structure are generated
with regard to the same camera angle. That depth image
is simultaneously transformed to a partial 2.5D voxel grid
using virtual camera parameters. Then a pair of partial 2.5D
view and the complete 3D shape is synthesized. Overall,
6
around 20K training pairs and 4K testing pairs are generated for each 3D object category. All data are produced in
Blender.
Besides the large quantity of synthesized data, we also
collect real world data in order to test the proposed network.
We use a Microsoft Kinect camera to manually scan a set of
common objects, such as chairs, tables, etc., from multiple
angles. Then, we use ElasticFusion [47] to reconstruct the
full 3D shapes of the objects, as well as the camera pose
in each scan. The 3D objects are manually segmented from
the background. We then extract ground truth information
by aligning the full 3D objects with the partial 2.5D views.
It should be noted that, due to noise and quantization
artifacts of low-cost RGB-D sensors, and the inaccuracy of
the algorithm, the full 3D ground truth is not 100% accurate.
4
E VALUATION
In this section, we evaluate our 3D-RecGAN++ with comparison to the state of the art approaches and an ablation
study to fully investigate the proposed network.
4.1
Metrics
To evaluate the performance of 3D reconstruction, we consider the mean Intersection-over-Union (IoU) between predicted 3D voxel grids and their ground truth. The IoU for
an individual voxel grid is formally defined as follows:
PN
I(yi > p) ∗ I(y¯i )
h
IoU = P i=1
i
N
i=1 I I(yi > p) + I(y¯i )
where I(·) is an indicator function, yi is the predicted value
at the ith voxel, y¯i is the corresponding ground truth, p is the
threshold for voxelization, N is the total number of voxels
in a whole voxel grid. In all our experiments, p is set as
0.5. If the predicted value is over 0.5, it is more likely to be
occupied from the probabilistic aspect. The higher the IoU
value, the better the reconstruction of a 3D model.
4.2
Competing Approaches
We compare against three state of the art deep learning
based approaches for single depth view reconstruction. We
also compare against the generator alone in our network,
i.e. without the GAN, named as 3D-RecAE for short.
(1) 3D-EPN. In [11], Dai et al. proposed a neural network
to reconstruct the 3D shape up to a 323 voxel grid, after
which a high resolution shape is retrieved from an existing
3D shape database. For fair comparison, we only compared
with their neural network performance. Besides, occupancy
grid representation is used for the network training the
testing.
(2) Varley et al. In [2], a network was designed to
complete the 3D shape from a single 2.5D depth view for
robot grasping. The output of their network is a 403 voxel
grid.
(3) Han et al. In [23], a global structure inference network
and a local geometry refinement network are proposed to
complete a high resolution shape from a noisy shape. The
network is not originally designed for single depth view
reconstruction, but its output shape is up to a 2563 voxel
grid and is comparable to our network. For fair comparison,
the same occupancy grid representation is used for their network. It should be noted that their network involves many
convoluted designs, yet the training procedure is extremely
slow and inefficient due to many LSTMs involved.
(4) 3D-RecAE. As for our 3D-RecGAN++, we remove
the discriminator and only keep the generator to infer 3D
complete shape from a single depth view. This comparison
illustrates the benefits from adversarial learning.
4.3
Single-category Results
(1) Results. All networks are separately trained and
tested on four different categories with the same network
configurations. To fairly compare the IoU between different
approaches, we down sample all results of 2563 voxel grids
to 323 using max pooling with a stride of 8 along the three
axes. Table 1 shows the IoU comparison of all methods on
323 voxel grids, while Table 2 shows the IoU comparison of
[23] and our approaches on higher resolution of voxel grids.
Figure 5 shows the qualitative results of single category reconstruction. In this paper, the meshgrid function in Matlab
is used to plot all 3D shapes for better visualization.
(2) Analysis. The proposed 3D-RecGAN++ significantly
outperforms the competing approaches in terms of the IoU
at both lower (323 voxel grids) and higher resolutions
(2563 voxel grids). The 3D shapes generated by our 3DRecGAN++ are much more visually compelling than others
in terms of the shape accuracy and the geometrical details.
TABLE 1: Per-category IoU (323 voxel grids).
3D-EPN [11]
Varley et al. [2]
Han et al. [23]
3D-RecAE (ours)
3D-RecGAN++ (ours)
bench
0.758
0.653
0.611
0.800
0.806
chair
0.739
0.619
0.524
0.790
0.793
coach
0.834
0.818
0.505
0.858
0.868
table
0.772
0.678
0.615
0.808
0.821
TABLE 2: Per-category IoU (up to 2563 voxel grids).
Han et al. [23] (643 )
Han et al. [23] (1283 )
Han et al. [23] (2563 )
3D-RecAE (ours) (643 )
3D-RecAE (ours) (1283 )
3D-RecAE (ours) (2563 )
3D-RecGAN++ (ours) (643 )
3D-RecGAN++ (ours) (1283 )
3D-RecGAN++ (ours) (2563 )
4.4
bench
0.544
0.492
0.417
0.733
0.669
0.552
0.745
0.683
0.564
chair
0.469
0.432
0.389
0.736
0.693
0.622
0.741
0.699
0.628
coach
0.483
0.469
0.431
0.832
0.812
0.722
0.844
0.825
0.733
table
0.560
0.524
0.476
0.759
0.715
0.644
0.772
0.730
0.659
Multi-category Results
(1) Results. All networks are also trained and tested on
multiple categories without being given any class labels.
The networks are trained on four categories: {bench, chair,
coach, table}; and then tested separately on an individual
category. Table 3 shows the IoU comparison of all methods
on the resolution of 323 voxel grids, while Table 4 shows
the IoU comparison of [23] and our methods on higher
resolution of voxel grids. Figure 6 shows the qualitative
results of all approaches on multiple categories.
(2) Analysis. The proposed 3D-RecGAN++ significantly
outperforms the state of the art by a large margin in all
categories which are trained together on a single model.
7
2.5D input
(643)
3D-EPN
Varley et al.
Han et al.
Besides, the performance of our network trained on multiple
categories, does not degrade compared with training the
network on individual categories. This confirms that our
network has enough capacity and capability to learn diverse
features from multiple categories.
TABLE 3: Multi-category IoU (323 voxel grids).
3D-EPN [11]
Varley et al. [2]
Han et al. [23]
3D-RecAE (ours)
3D-RecGAN++ (ours)
4.5
3D-RecAE
3D-RecGAN++ Ground Truth
(2563)
(403)
(2563)
(2563)
(323)
Fig. 5: Qualitative results of per-category reconstruction from different approaches.
bench
0.758
0.681
0.531
0.789
0.806
chair
0.734
0.630
0.512
0.767
0.784
coach
0.843
0.823
0.469
0.850
0.863
table
0.772
0.691
0.596
0.795
0.806
Cross-category Results
(1) Results. To further investigate the generality of networks, we train all networks on {bench, chair, coach, table},
and then test them on another 6 totally different categories:
{airplane, car, faucet, guitar, gun, monitor}. For each of the
6 categories, it has 4625 single arbitrary views from random
selected 37 objects for testing, which is the same data size
as used in the previous {bench, chair, coach, table}. Table 5
(2563)
TABLE 4: Multi-category IoU (up to 2563 voxel grids).
Han et al. [23] (643 )
Han et al. [23] (1283 )
Han et al. [23] (2563 )
3D-RecAE (ours) (643 )
3D-RecAE (ours) (1283 )
3D-RecAE (ours) (2563 )
3D-RecGAN++ (ours) (643 )
3D-RecGAN++ (ours) (1283 )
3D-RecGAN++ (ours) (2563 )
bench
0.453
0.408
0.364
0.719
0.653
0.537
0.741
0.679
0.562
chair
0.451
0.416
0.390
0.710
0.665
0.595
0.730
0.688
0.618
coach
0.442
0.427
0.398
0.825
0.804
0.715
0.838
0.819
0.728
table
0.534
0.496
0.458
0.741
0.695
0.623
0.753
0.710
0.638
shows the IoU comparison of all approaches on 323 voxel
grids, while Table 6 shows the IoU comparison of [23] and
our approaches on higher resolution of voxel grids. Figure
7 shows the qualitative results of all methods on 6 unseen
categories.
We further evaluate the generality of our 3D-RecGAN++
on a specific category. Particularly, we conduct four groups
of experiments. In the first group, we train our 3DRecGAN++ on bench, then separately test on the remaining
3 categories: {chair, coach, table}. In the second group, the
8
2.5D input
(643)
3D-EPN
Varley et al.
Han et al.
3D-RecAE
3D-RecGAN++ Ground Truth
(323)
(403)
(2563)
(2563)
(2563)
Fig. 6: Qualitative results of multi-category reconstruction from different approaches.
TABLE 5: Cross-category IoU (323 voxel grids).
airplane
3D-EPN [11]
0.667
Varley et al. [2]
0.524
Han et al. [23]
0.466
3D-RecAE (ours)
0.670
3D-RecGAN++ (ours) 0.683
car
0.723
0.069
0.414
0.805
0.802
faucet
0.683
0.025
0.431
0.723
0.731
guitar
0.753
0.041
0.448
0.762
0.769
gun monitor
0.742 0.715
0.065 0.051
0.401 0.585
0.725 0.722
0.739 0.738
TABLE 6: Cross-category IoU (up to 2563 voxel grids).
Han et al. [23] (643 )
Han et al. [23] (1283 )
Han et al. [23] (2563 )
3D-RecAE (ours) (643 )
3D-RecAE (ours) (1283 )
3D-RecAE (ours) (2563 )
3D-RecGAN++ (ours) (643 )
3D-RecGAN++ (ours) (1283 )
3D-RecGAN++ (ours) (2563 )
airplane car faucet guitar gun monitor
0.373 0.372 0.356 0.341 0.296 0.517
0.318 0.350 0.319 0.286 0.245 0.470
0.248 0.299 0.284 0.225 0.198 0.414
0.602 0.774 0.678 0.700 0.665 0.661
0.547 0.745 0.645 0.643 0.605 0.614
0.365 0.557 0.520 0.413 0.381 0.529
0.615 0.769 0.687 0.709 0.682 0.680
0.559 0.740 0.653 0.651 0.625 0.633
0.379 0.555 0.529 0.424 0.399 0.547
network is trained on chair and separately tested on {bench,
coach, table}. Similarly, another two groups of experiments
are conducted. Basically, this experiment is to investigate
how well our approach learns features from one category
(2563)
and then generalizes to a different category, and vice versa.
Table 7 shows the cross-category IoU of our 3D-RecGAN++
trained on individual category over 2563 voxel grids.
(2) Analysis. The proposed 3D-RecGAN++ achieves
much higher IoU across the unseen categories than competing approaches. Our network not only learns rich features
from different object categories, but also is able to generalize
well to completely new types of categories. This implies
that our network may learn geometric relationships such
as lines, planes, curves which are common across various
object categories. It can be also observed that our model
trained on bench tends to be more general than others,
because bench is likely to have more general features to be
learned, while simple categories such as coach are unlikely
to consist of many general features that are shared across
different categories.
4.6
Real-world Experiment Results
(1) Results. Lastly, in order to evaluate the domain adaptation capability of the networks, we train all networks on
synthesized data of categories {bench, chair, coach, table},
9
2.5D input
(643)
3D-EPN
Varley et al.
Han et al.
3D-RecAE
3D-RecGAN++ Ground Truth
(323)
(2563)
(2563)
(403)
(2563)
Fig. 7: Qualitative results of cross-category reconstruction from different approaches.
TABLE 7: Cross-category IoU of 3D-RecGAN++ trained on
individual category (2563 voxel grids).
Group 1 (trained on bench)
Group 2 (trained on chair)
Group 3 (trained on coach)
Group 4 (trained on table)
bench
–
0.482
0.387
0.477
chair
0.510
–
0.460
0.495
coach
0.549
0.509
–
0.449
table
0.592
0.549
0.401
–
and then test them on real-world data collected by a Microsoft Kinect camera. The real-world data were collected
in different environments, including offices, homes, and
outdoor university parks, as shown in Figure 8. Compared
to synthesized data, real-world partial 2.5D views are noisier and largely incomplete. For each object, we randomly
selected 20 different 2.5D depth views for testing. Table 8
shows the IoU performance of all approaches using 323
voxel grids, while Table 9 compares the IoU of [23] and
(2563)
our approaches on higher resolutions. Figure 9 shows some
qualitative results for all methods.
(2) Analysis. There are two reasons why the IoU is
significantly lower compared with testing on the synthetic
dataset. First, the ground truth objects obtained from ElasticFusion are empty rather than solid, and are only occupied
on the surface. However, all networks predict dense and
solid voxel grids, so the interior of bulky objects like couches
is not matching. Secondly, the input 2.5D depth view from
real world dataset is noisy and incomplete, due to the
limitation of the RGB-D sensor (e.g., reflective surfaces,
outdoor light). In many cases, the input 2.5D view does not
capture the whole object and only contains a small part of
the object, which also leads to failure cases (e.g. the 6th row
in Figure 9) and a lower IoU scores overall. However, our
proposed network is still able to reconstruct reasonable 3D
dense shapes given the noisy and incomplete 2.5D input
10
Fig. 8: Real world objects for reconstruction.
depth views, while the competing algorithms (e.g. Varley et
al.) are not robust to real world noise and unable to generate
compelling results.
TABLE 8: Real-world multi-category IoU (323 voxel grids).
3D-EPN [11]
Varley et al. [2]
Han et al. [23]
3D-RecAE (ours)
3D-RecGAN++ (ours)
bench
0.350
0.043
0.291
0.352
0.347
chair
0.273
0.0.008
0.240
0.295
0.297
coach
0.173
0.038
0.159
0.199
0.201
table
0.246
0.009
0.228
0.268
0.268
TABLE 9: Real-world multi-category IoU (up to 2563 voxel
grids).
Han et al. [23] (643 )
Han et al. [23] (1283 )
Han et al. [23] (2563 )
3D-RecAE (ours) (643 )
3D-RecAE (ours) (1283 )
3D-RecAE (ours) (2563 )
3D-RecGAN++ (ours) (643 )
3D-RecGAN++ (ours) (1283 )
3D-RecGAN++ (ours) (2563 )
4.7
bench
0.184
0.117
0.070
0.231
0.145
0.071
0.230
0.147
0.074
chair
0.148
0.092
0.045
0.178
0.096
0.032
0.174
0.097
0.032
coach
0.101
0.067
0.040
0.107
0.055
0.024
0.109
0.057
0.025
table
0.126
0.069
0.031
0.148
0.071
0.024
0.146
0.069
0.023
Impact of Adversarial Learning
(1) Results. In all above experiments, the proposed 3DRecGAN++ tends to outperform the ablated network 3DRecAE which does not include the adversarial learning of
GAN part. In all visualization of experiment results, the
3D shapes from 3D-RecGAN++ are also more compelling
than 3D-RecAE. To further investigate how the adversarial
learning improves the final 3D results comparing with 3DRecAE, we calculate the mean precision and recall from the
above multi-category experiment results. Table 10 shows
the mean precision of 3D-RecGAN++ and 3D-RecAE on
individual categories using the network trained on multiple
categories, while Table 11 shows the mean recall.
(2) Analysis. It can be seen that the results of 3DRecGAN++ have much higher precision scores than 3DRecAE, which means 3D-RecGAN++ has much less false
positive estimations, while 3D-RecAE tends to estimate
much more false positives. Therefore, the estimated 3D
shapes from 3D-RecAE are likely to be ’fatter’ and ’bigger’, while 3D-RecGAN++ tends to output ’thinner’ shapes
with much more shape details being exposed. Both 3DRecGAN++ and 3D-RecAE can achieve extremely high recall scores (i.e. above 0.9), although 3D-RecGAN++ has
lower recall scores compared with 3D-RecAE. This means
both 3D-RecGAN++ and 3D-RecAE are capable of estimating almost all of the object shapes without too many false
negatives. In other words, the ground truth 3D shape tends
to be a subset of the estimated shape result.
TABLE 10: Multi-category mean precision (up to 2563 voxel
grids).
3D-RecAE (323 )
3D-RecAE (643 )
3D-RecAE (1283 )
3D-RecAE (2563 )
3D-RecGAN++ (323 )
3D-RecGAN++ (643 )
3D-RecGAN++ (1283 )
3D-RecGAN++ (2563 )
bench
0.823
0.755
0.688
0.564
0.856
0.798
0.737
0.608
chair
0.802
0.746
0.700
0.626
0.832
0.781
0.739
0.663
coach
0.868
0.843
0.822
0.730
0.888
0.866
0.847
0.751
table
0.825
0.773
0.727
0.651
0.845
0.797
0.755
0.676
TABLE 11: Multi-category mean recall (up to 2563 voxel
grids).
3D-RecAE (323 )
3D-RecAE (643 )
3D-RecAE (1283 )
3D-RecAE (2563 )
3D-RecGAN++ (323 )
3D-RecGAN++ (643 )
3D-RecGAN++ (1283 )
3D-RecGAN++ (2563 )
bench
0.952
0.939
0.930
0.921
0.934
0.915
0.900
0.887
chair
0.945
0.935
0.929
0.924
0.931
0.918
0.908
0.902
coach
0.978
0.975
0.974
0.972
0.968
0.964
0.961
0.959
table
0.956
0.947
0.940
0.935
0.946
0.934
0.925
0.918
Overall, with regard to experiments on per-category,
multi-category, and cross-category experiments, our 3DRecGAN++ outperforms others by a large margin, although
all other approaches can reconstruct reasonable shapes. In
terms of the generality, Varley et al. [2] and Han et al.
[23] are inferior because [2] uses a single fully connected
layers, instead of 3D ConvNets, for shape generation which
is unlikely to be general for various shapes, and [23] applies
LSTMs for shape blocks generation which is inefficient and
11
2.5D input
(643)
3D-EPN
Varley et al.
Han et al.
unable to learn general 3D structures. However, our 3DRecGAN++ is superior thanks to the generality of simple
yet efficient 3D autoencoder and the 3D convolutional discriminator. Besides, the 3D-RecAE tends to over estimate the
3D shape, while the adversarial learning of 3D-RecGAN++
is likely to remove the over-estimated parts, so as to leave
the estimated shape to be clearer with more shape details.
5
3D-RecAE
3D-RecGAN++ Ground Truth
(2563)
(2563)
(323)
(2563)
(403)
Fig. 9: Qualitative results of real world objects reconstruction from different approaches.
D ISCUSSION
Although our 3D-RecGAN++ achieves the state of the
art performance in 3D object reconstruction from a single
depth view, it has limitations. Firstly, our network takes
the volumetric representation of a single depth view as
input, instead of taking a raw depth image. Therefore,
a preprocessing of raw depth images is required for our
network. However, in many application scenarios such as
robot grasping, such preprocessing would be trivial and
straightforward given the depth camera parameters. Secondly, the input depth view of our network only contains a
clean object information without cluttered background. One
possible solution is to leverage an existing segmentation
algorithm such as Mask-RCNN [82] to clearly segment the
target object instance from the raw depth view.
6
(2563)
C ONCLUSION
In this work, we proposed a novel framework 3DRecGAN++ that reconstructs the full 3D structure of an
object from an arbitrary depth view. By leveraging the
generalization capabilities of autoencoders and generative
adversarial networks, our 3D-RecGAN++ predicts dense
and accurate 3D structures with fine details, outperforming
the state of the art in single-view shape completion for
individual object category. We further tested our network’s
ability to reconstruct multiple categories without providing
any object class labels during training or testing, and it
showed that our network is still able to predict precise
3D shapes. Besides, we investigated the network’s reconstruction performance on unseen categories, our proposed
approach can also predict satisfactory 3D structures. Finally,
our model is robust to real world noisy data and can infer
accurate 3D shapes although the model is purely trained
on synthesized data. This confirms that our network has
the capability of learning general 3D latent features of the
objects, rather than simply fitting a function for the training
datasets, and the adversarial learning of 3D-RecGAN++
learns to add geometric details for estimated 3D shapes. In
summary, our network only requires a single depth view to
recover a dense and complete 3D shape with fine details.
12
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
A. Sharma, O. Grau, and M. Fritz, “VConv-DAE : Deep Volumetric
Shape Learning Without Object Labels,” ECCV, 2016.
J. Varley, C. Dechant, A. Richardson, J. Ruales, and P. Allen, “Shape
Completion Enabled Robotic Grasping,” IROS, 2017.
R. A. Newcombe, S. Izadi, O. Hilliges, D. Molyneaux, D. Kim,
A. J. Davison, P. Kohli, J. Shotton, S. Hodges, and A. Fitzgibbon,
“KinectFusion: Real-time dense surface mapping and tracking,”
ISMAR, 2011.
M. Nießner, M. Zollhöfer, S. Izadi, and M. Stamminger, “Real-time
3D reconstruction at scale using voxel hashing,” ACM Transactions
on Graphics, vol. 32, no. 6, pp. 1–11, 2013.
F. Steinbrucker, C. Kerl, J. Sturm, and D. Cremers, “Large-Scale
Multi-Resolution Surface Reconstruction from RGB-D Sequences,”
ICCV, 2013.
A. Nealen, T. Igarashi, O. Sorkine, and M. Alexa, “Laplacian Mesh
Optimization,” SIGGRAPH, 2006.
W. Zhao, S. Gao, and H. Lin, “A robust hole-filling algorithm for
triangular mesh,” The Visual Computer, vol. 23, no. 12, pp. 987–997,
2007.
M. Kazhdan, M. Bolitho, and H. Hoppe, “Poisson Surface Reconstruction,” Symposium on Geometry Processing, 2006.
M. Kazhdan and H. Hoppe, “Screened poisson surface reconstruction,” ACM Transactions on Graphics, vol. 32, no. 3, pp. 1–13, 2013.
Z. Wu, S. Song, A. Khosla, F. Yu, L. Zhang, X. Tang, and J. Xiao,
“3D ShapeNets: A Deep Representation for Volumetric Shapes,”
CVPR, 2015.
A. Dai, C. R. Qi, and M. Nießner, “Shape Completion using 3DEncoder-Predictor CNNs and Shape Synthesis,” CVPR, 2017.
I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. WardeFarley, S. Ozair, A. Courville, and Y. Bengio, “Generative Adversarial Nets,” NIPS, 2014.
D. P. Kingma and M. Welling, “Auto-Encoding Variational Bayes,”
ICLR, 2014.
Z. Hu, Z. Yang, X. Liang, R. Salakhutdinov, and E. P. Xing,
“Controllable Text Generation,” ICML, 2017.
T. Karras, T. Aila, S. Laine, and J. Lehtinen, “Progressive Growing
of GANs for Improved Quality, Stability, and Variation,” ICLR,
2018.
X. Chen, Y. Duan, R. Houthooft, J. Schulman, I. Sutskever, and
P. Abbeel, “InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets,” NIPS, 2016.
T. D. Kulkarni, W. F. Whitney, P. Kohli, and J. B. Tenenbaum, “Deep
Convolutional Inverse Graphics Network,” NIPS, 2015.
E. Grant, P. Kohli, and M. V. Gerven, “Deep Disentangled Representations for Volumetric Reconstruction,” ECCV Workshops, 2016.
R. Girdhar, D. F. Fouhey, M. Rodriguez, and A. Gupta, “Learning
a Predictable and Generative Vector Representation for Objects,”
ECCV, 2016.
H. Huang, E. Kalogerakis, and B. Marlin, “Analysis and synthesis of 3D shape families via deep-learned generative models of
surfaces,” Computer Graphics Forum, vol. 34, no. 5, pp. 25–38, 2015.
J. Wu, C. Zhang, T. Xue, W. T. Freeman, and J. B. Tenenbaum,
“Learning a Probabilistic Latent Space of Object Shapes via 3D
Generative-Adversarial Modeling,” NIPS, 2016.
M. Mirza and S. Osindero, “Conditional Generative Adversarial
Nets,” arXiv, 2014.
X. Han, Z. Li, H. Huang, E. Kalogerakis, and Y. Yu, “HighResolution Shape Completion Using Deep Neural Networks for
Global Structure and Local Geometry Inference,” ICCV, 2017.
B. Yang, H. Wen, S. Wang, R. Clark, A. Markham, and N. Trigoni,
“3D Object Reconstruction from a Single Depth View with Adversarial Learning,” ICCV Workshops, 2017.
A. Monszpart, N. Mellado, G. J. Brostow, and N. J. Mitra, “RAPter:
Rebuilding Man-made Scenes with Regular Arrangements of
Planes,” ACM Transactions on Graphics, vol. 34, no. 4, pp. 1–12,
2015.
N. J. Mitra, L. J. Guibas, and M. Pauly, “Partial and Approximate
Symmetry Detection for 3D Geometry,” SIGGRAPH, 2006.
M. Pauly, N. J. Mitra, J. Wallner, H. Pottmann, and L. J. Guibas,
“Discovering structural regularity in 3D geometry,” ACM Transactions on Graphics, vol. 27, no. 3, p. 1, 2008.
I. Sipiran, R. Gregor, and T. Schreck, “Approximate Symmetry
Detection in Partial 3D Meshes,” Computer Graphics Forum, vol. 33,
no. 7, pp. 131–140, 2014.
P. Speciale, M. R. Oswald, A. Cohen, and M. Pollefeys, “A Symmetry Prior for Convex Variational 3D Reconstruction,” ECCV, 2016.
[30] S. Thrun and B. Wegbreit, “Shape from symmetry,” ICCV, 2005.
[31] Y. M. Kim, N. J. Mitra, D.-M. Yan, and L. Guibas, “Acquiring
3D Indoor Environments with Variability and Repetition,” ACM
Transactions on Graphics, vol. 31, no. 6, 2012.
[32] Y. Li, A. Dai, L. Guibas, and M. Nießner, “Database-Assisted Object Retrieval for Real-Time 3D Reconstruction,” Computer Graphics
Forum, vol. 34, no. 2, pp. 435–446, 2015.
[33] L. Nan, K. Xie, and A. Sharf, “A Search-Classify Approach for
Cluttered Indoor Scene Understanding,” ACM Transactions on
Graphics, vol. 31, no. 6, pp. 1–10, 2012.
[34] T. Shao, W. Xu, K. Zhou, J. Wang, D. Li, and B. Guo, “An
interactive approach to semantic modeling of indoor scenes with
an RGBD camera,” ACM Transactions on Graphics, vol. 31, no. 6,
pp. 1–11, 2012.
[35] Y. Shi, P. Long, K. Xu, H. Huang, and Y. Xiong, “Data-driven
contextual modeling for 3d scene understanding,” Computers &
Graphics, vol. 55, pp. 55–67, 2016.
[36] J. Rock, T. Gupta, J. Thorsen, J. Gwak, D. Shin, and D. Hoiem,
“Completing 3D Object Shape from One Depth Image,” CVPR,
2015.
[37] R. Hartley and A. Zisserman, Multiple View Geometry in Computer
Vision. Cambridge University Press, 2004.
[38] R. A. Newcombe, S. J. Lovegrove, and A. J. Davision, “DTAM:
Dense Tracking and Mapping in Real-time,” ICCV, 2011.
[39] S. Baker and I. Matthews, “Lucas-Kanade 20 Years On : A Unifying Framework : Part 1,” International Journal of Computer Vision,
vol. 56, no. 3, pp. 221–255, 2004.
[40] C. B. Choy, D. Xu, J. Gwak, K. Chen, and S. Savarese, “3DR2N2: A Unified Approach for Single and Multi-view 3D Object
Reconstruction,” ECCV, 2016.
[41] X. Di, R. Dahyot, and M. Prasad, “Deep Shape from a Low
Number of Silhouettes,” ECCV, 2016.
[42] Z. Lun, M. Gadelha, E. Kalogerakis, S. Maji, and R. Wang, “3D
Shape Reconstruction from Sketches via Multi-view Convolutional Networks,” 3DV, 2017.
[43] D. J. Rezende, S. M. A. Eslami, S. Mohamed, P. Battaglia, M. Jaderberg, and N. Heess, “Unsupervised Learning of 3D Structure from
Images,” NIPS, 2016.
[44] A. Kar, C. Häne, and J. Malik, “Learning a Multi-View Stereo
Machine,” NIPS, 2017.
[45] M. Ji, J. Gall, H. Zheng, Y. Liu, and L. Fang, “SurfaceNet: An Endto-end 3D Neural Network for Multiview Stereopsis,” ICCV, 2017.
[46] T. Whelan, J. McDonald, M. Kaess, M. Fallon, H. Johannsson, and
J. J. Leonard, “Kintinuous: Spatially Extended Kinectfusion,” RSS
Workshops, 2012.
[47] T. Whelan, S. Leutenegger, R. F. Salas-moreno, B. Glocker, and A. J.
Davison, “ElasticFusion : Dense SLAM Without A Pose Graph,”
RSS, 2015.
[48] G. Riegler, A. O. Ulusoy, H. Bischof, and A. Geiger, “OctNetFusion:
Learning Depth Fusion from Data,” 3DV, 2017.
[49] V. Blanz and T.Vetter, “Face Recognition based on Fitting a 3D
Morphable Model,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 25, no. 9, pp. 1063–1074, 2003.
[50] P. Dou, S. K. Shah, and I. A. Kakadiaris, “End-to-end 3D face
reconstruction with deep neural networks,” CVPR, 2017.
[51] A. Kar, S. Tulsiani, J. Carreira, and J. Malik, “Category-specific
object reconstruction from a single image,” CVPR, 2015.
[52] J. Gwak, C. B. Choy, M. Chandraker, A. Garg, and S. Savarese,
“Weakly supervised 3D Reconstruction with Adversarial Constraint,” arXiv, 2017.
[53] S. Tulsiani, T. Zhou, A. A. Efros, and J. Malik, “Multi-view Supervision for Single-view Reconstruction via Differentiable Ray
Consistency,” CVPR, 2017.
[54] X. Yan, J. Yang, E. Yumer, Y. Guo, and H. Lee, “Perspective
Transformer Nets: Learning Single-View 3D Object Reconstruction
without 3D Supervision,” NIPS, 2016.
[55] C. Kong, C.-H. Lin, and S. Lucey, “Using Locally Corresponding
CAD Models for Dense 3D Reconstructions from a Single Image,”
CVPR, 2017.
[56] A. Kurenkov, J. Ji, A. Garg, V. Mehta, J. Gwak, C. Choy, and
S. Savarese, “DeformNet: Free-Form Deformation Network for 3D
Shape Reconstruction from a Single Image,” NIPS, 2017.
[57] J. K. Murthy, G. V. S. Krishna, F. Chhaya, and K. M. Krishna,
“Reconstructing Vechicles from a Single Image : Shape Priors for
Road Scene Understanding,” ICRA, 2017.
[58] H. Christian, S. Tulsiani, and J. Malik, “Hierarchical Surface Prediction for 3D Object Reconstruction,” 3DV, 2017.
13
[59] M. Tatarchenko, A. Dosovitskiy, and T. Brox, “Octree Generating Networks: Efficient Convolutional Architectures for Highresolution 3D Outputs,” ICCV, 2017.
[60] A. Johnston, R. Garg, G. Carneiro, I. Reid, and A. v. d. Hengel,
“Scaling CNNs for High Resolution Volumetric Reconstruction
from a Single Image,” ICCV Workshops, 2017.
[61] C.-H. Lin, C. Kong, and S. Lucey, “Learning Efficient Point Cloud
Generation for Dense 3D Object Reconstruction,” AAAI, 2018.
[62] J. Wu, Y. Wang, T. Xue, X. Sun, W. T. Freeman, and J. B. Tenenbaum,
“MarrNet: 3D Shape Reconstruction via 2.5D Sketches,” NIPS,
2017.
[63] M. Firman, O. M. Aodha, S. Julier, and G. J. Brostow, “Structured
Prediction of Unobserved Voxels From a Single Depth Image,”
CVPR, 2016.
[64] S. Song, F. Yu, A. Zeng, A. X. Chang, M. Savva, and T. Funkhouser,
“Semantic Scene Completion from a Single Depth Image,” CVPR,
2017.
[65] W. Wang, Q. Huang, S. You, C. Yang, and U. Neumann, “Shape Inpainting using 3D Generative Adversarial Network and Recurrent
Convolutional Networks,” ICCV, 2017.
[66] C. Zou, E. Yumer, J. Yang, D. Ceylan, and D. Hoiem, “3D-PRNN:
Generating Shape Primitives with Recurrent Neural Networks,”
ICCV, 2017.
[67] C. Ledig, L. Theis, F. Huszar, J. Caballero, A. Cunningham,
A. Acosta, A. Aitken, A. Tejani, J. Totz, Z. Wang, and W. Shi,
“Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network,” CVPR, 2017.
[68] S. Reed, Z. Akata, X. Yan, L. Logeswaran, B. Schiele, and H. Lee,
“Generative Adversarial Text to Image Synthesis,” ICML, 2016.
[69] M. Gadelha, S. Maji, and R. Wang, “3D Shape Induction from 2D
Views of Multiple Objects,” 3DV, 2017.
[70] E. Smith and D. Meger, “Improved Adversarial Systems for 3D
Object Generation and Reconstruction,” CoRL, 2017.
[71] A. A. Soltani, H. Huang, J. Wu, T. D. Kulkarni, and J. B. Tenenbaum, “Synthesizing 3D Shapes via Modeling Multi-View Depth
Maps and Silhouettes with Deep Generative Networks,” CVPR,
2017.
[72] A. X. Chang, T. Funkhouser, L. Guibas, P. Hanrahan, Q. Huang,
Z. Li, S. Savarese, M. Savva, S. Song, H. Su, J. Xiao, L. Yi, and
F. Yu, “ShapeNet: An Information-Rich 3D Model Repository,”
arXiv, 2015.
[73] L. v. d. Maaten and G. Hinton, “Visualizing Data using t-SNE,”
Journal of Machine Learning Research, vol. 9, no. Nov, pp. 2579–2605,
2008.
[74] O. Ronneberger, P. Fischer, and T. Brox, “U-Net : Convolutional
Networks for Biomedical Image Segmentation,” MICCAI, 2015.
[75] M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein GAN,”
ICML, 2017.
[76] I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and
A. Courville, “Improved Training of Wasserstein GANs,” NIPS,
2017.
[77] M. Arjovsky and L. Bottou, “Towards Principled Methods for
Training Generative Adversarial Networks,” ICLR, 2017.
[78] Y. Mroueh, T. Sercu, and V. Goel, “McGAN: Mean and Covariance
Feature Matching GAN,” ICML, 2017.
[79] J. Bao, D. Chen, F. Wen, H. Li, and G. Hua, “CVAE-GAN: FineGrained Image Generation through Asymmetric Training,” ICCV,
2017.
[80] A. Brock, T. Lim, J. M. Ritchie, and N. Weston, “Generative
and Discriminative Voxel Modeling with Convolutional Neural
Networks,” NIPS Workshops, 2016.
[81] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” ICLR, 2015.
[82] K. He, G. Gkioxari, P. Dollar, and R. Girshick, “Mask R-CNN,”
ICCV, 2017.
| 1cs.CV
|
arXiv:1704.02043v2 [math.AG] 9 Nov 2017
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS AND
GROUPS OF PSEUDO-AUTOMORPHISMS
SERGE CANTAT AND YVES DE CORNULIER
A BSTRACT. Pseudo-automorphisms are birational transformations acting as regular automorphisms in codimension 1. We import ideas from geometric group
theory to study groups of birational transformations, and prove that a group of birational transformations that satisfies a fixed point property on CAT(0) cubical complexes is birationally conjugate to a group acting by pseudo-automorphisms on
some non-empty Zariski-open subset. We apply this argument to classify groups
of birational transformations of surfaces with this fixed point property up to birational conjugacy.
1. I NTRODUCTION
1.1. Birational transformations and pseudo-automorphisms. Let X be a quasiprojective variety, over an algebraically closed field k. Denote by Bir(X) the group
of birational transformations of X and by Aut(X) the subgroup of (regular) automorphisms of X. For the affine space of dimension n, automorphisms are invertible
transformations f : Ank → Ank such that both f and f −1 are defined by polynomial
formulas in affine coordinates:
f (x1 , . . . , xn ) = (f1 , . . . , fn ), f −1 (x1 , . . . , xn ) = (g1 , . . . , gn )
with fi , gi ∈ k[x1 , . . . , xn ]. Similarly, birational transformations of Ank are given by
rational formulas, i.e. fi , gi ∈ k(x1 , . . . , xn ).
Birational transformations may contract hypersurfaces. Roughly speaking, pseudo-automorphisms are birational transformations that act as automorphisms in
codimension 1. Precisely, a birational transformation f : X 99K X is a pseudoautomorphism if there exist Zariski-open subsets U and V in X such that X r U
and X r V have codimension ≥ 2 and f induces an isomorphism from U to V.
The pseudo-automorphisms of X form a group, which we denote by Psaut(X).
For instance, all birational transformations of Calabi-Yau manifolds are pseudoautomorphisms; and there are examples of such manifolds for which Psaut(X) is
Date: November 9, 2017.
2010 Mathematics Subject Classification. Primary 14E07, Secondary 14J50, 20F65.
1
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
2
infinite while Aut(X) is trivial (see [10]). Pseudo-automorphisms are studied in
Section 2.
Definition 1.1. Let Γ ⊂ Bir(X) be a group of birational transformations of an
irreducible projective variety X. We say that Γ is pseudo-regularizable if there
exists a triple (Y, U, ϕ) where
(1) Y is a projective variety and ϕ : Y 99K X is a birational map;
(2) U is a dense Zariski open subset of Y ;
(3) ϕ−1 ◦ Γ ◦ ϕ yields an action of Γ by pseudo-automorphisms on U.
More generally if α : Γ → Bir(X) is a homomorphism, we say that it is pseudoregularizable if α(Γ) is pseudo-regularizable.
One goal of this article is to use rigidity properties of commensurating actions,
a purely group-theoretic concept, to show that many group actions are pseudoregularizable. In particular, we exhibit a class of groups for which all actions by
birational transformations on projective varieties are pseudo-regularizable.
1.2. Property (FW). The class of groups we shall be mainly interested in is characterized by a fixed point property appearing in several related situations, for instance for actions on CAT(0) cubical complexes. Here, we adopt the viewpoint of
commensurated subsets. Let Γ be a group, and Γ × S → S an action of Γ on a set
S. Let A be a subset of S. One says that Γ commensurates A if the symmetric
difference
γ(A)4A = (γ(A) r A) ∪ (A r γ(A))
is finite for every element γ of Γ. One says that Γ transfixes A if there is a subset
B of S such that A4B is finite and B is Γ-invariant: γ(B) = B, for every γ in Γ.
A group Γ has Property (FW) if, given any action of Γ on any set S, all com√
mensurated subsets of S are automatically transfixed. For instance, SL 2 (Z[ 5]) and
SL 3 (Z) have Property (FW), but non-trivial free groups do not share this property.
Property (FW) is discussed in Section 3.
Let us mention that among various characterizations of Property (FW) (see [11]),
one is: every combinatorial action of Γ on a CAT(0) cube complex fixes some cube.
Another, for Γ finitely generated, is that all its infinite Schreier graphs are oneended.
1.3. Pseudo-regularizations. Let X be a projective variety. The group Bir(X)
does not really act on X, because there are indeterminacy points; it does not act
on the set of hypersurfaces either, because some of them may be contracted. As
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
3
˜
we shall explain, one can introduce the set Hyp(X)
of all irreducible and reduced
0
hypersurfaces in all birational models X 99K X (up to a natural identification).
Then there is a natural action of the group Bir(X) on this set, given by strict transforms. The rigorous construction of this action follows from a general categorical
framework, which is developed in Section 4. Moreover, this action commensurates
the subset Hyp(X) of hypersurfaces of X. This construction leads to the following
result.
Theorem A. Let X be a projective variety over an algebraically closed field. Let Γ
be a subgroup of Bir(X). If Γ has Property (FW), then Γ is pseudo-regularizable.
There is also a relative version of Property (FW) for pairs of groups Λ ≤ Γ,
which leads to a similar pseudo-regularization theorem for the subgroup Λ: this is
discussed in Section 6.4, with applications to distorted birational transformations.
Remark 1.2. Theorem A provides a triple (Y, U, ϕ) such that ϕ conjugates Γ to
a group of pseudo-automorphisms on the open subset U ⊂ Y . There are two extreme cases for the pair (Y, U) depending on the size of the boundary Y r U. If
this boundary is empty, Γ acts by pseudo-automorphisms on a projective variety Y .
If the boundary is ample, its complement U is an affine variety, and then Γ actually acts by regular automorphisms on U (see Section 2.4). Thus, in the study of
groups of birational transformations, pseudo-automorphisms of projective varieties
and regular automorphisms of affine varieties deserve specific attention.
1.4. Classification in dimension 2. In dimension 2, pseudo-automorphisms do not
differ much from automorphisms; for instance, Psaut(X) coincides with Aut(X) if
X is a smooth projective surface. Thus, for groups with Property (FW), Theorem A can be used to reduce the study of birational transformations to the study
of automorphisms of quasi-projective surfaces. Combining results of Danilov and
Gizatullin on automorphisms of affine surfaces with a theorem of Farley on groups
of piecewise affine transformations of the circle, we will be able to prove the following theorem.
Theorem B. Let X be a smooth, projective, and irreducible surface, over an algebraically closed field. Let Γ be an infinite subgroup of Bir(X). If Γ has Property
(FW), there is a birational map ϕ : Y 99K X such that
(1) Y is the projective plane P2 , a Hirzebruch surface Fm with m ≥ 1, or the
product of a curve C by the projective line P1 . If the characteristic of the
field is positive, Y is the projective plane P2k .
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
4
(2) ϕ−1 ◦ Γ ◦ ϕ is contained in Aut(Y ).
Remark 1.3. The group Aut(Y ) has finitely many connected components for all
surfaces Y listed in Assertion (1) of Theorem B. Thus, changing Γ into a finite
index subgroup Γ0 , one gets a subgroup of Aut(Y )0 . Here Aut(Y )0 denotes the
connected component of the identity of Aut(Y ); this is an algebraic group, acting
algebraically on Y .
Example 1.4. Groups with Kazhdan Property (T) satisfy Property (FW). Thus, Theorem B extends Theorem A of [8] and the present article offers a new proof of that
result.
√
Theorem B can also be applied to the group SL 2 (Z[ d]), where d ≥ 2 is a nonsquare positive integer. Thus, every action of this group on a projective surface by
birational transformations is conjugate to an action by regular automorphisms on
P2k , the product of a curve C by the projective line P1k , or a Hirzebruch surface.
Moreover, in this case, Margulis’ superrigidity theorem can be combined with Theorem B to get a more precise result, see §10.
Remark 1.5. In general, for a variety X one can ask whether Bir(X) transfixes
Hyp(X), or equivalently is pseudo-regularizable. For a surface X, this holds precisely when X is not birationally equivalent to the product of the projective line
with a curve. See §7.1 for more precise results.
1.5. Acknowledgement. This work benefited from interesting discussions with
Jérémy Blanc, Vincent Guirardel, Vaughan Jones, Christian Urech, and Junyi Xie.
2. P SEUDO - AUTOMORPHISMS
This preliminary section introduces useful notation for birational transformations
and pseudo-automorphisms, and presents a few basic results.
2.1. Birational transformations. Let X and Y be two irreducible and reduced
algebraic varieties over an algebraically closed field k. Let f : X 99K Y be a
birational map. Choose dense Zariski open subsets U ⊂ X and V ⊂ Y such
that f induces an isomorphism fU,V : U → V . Then the graph Gf of f is defined
as the Zariski closure of {(x, fU,V (x)) : x ∈ U } in X × Y ; it does not depend on
the choice of U and V . The graph Gf is an irreducible variety; both projections
u : Gf → X and v : Gf → Y are birational morphisms and f = v ◦ u−1 .
We shall denote by Ind(f ) the indeterminacy set of the birational map f .
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
5
Theorem 2.1 (Theorem 2.17 in [23]). Let f : X 99K Y be a rational map, with X
a normal variety and Y a projective variety. Then the indeterminacy set of f has
codimension ≥ 2.
Example 2.2. The transformation of the affine plane (x, y) 7→ (x, y/x) is birational,
and its indeterminacy locus is the line {x = 0}: this set of co-dimension 1 is
mapped “to infinity”. If the affine plane is compactified by the projective plane,
the transformation becomes [x : y : z] 7→ [x2 : yz : xz], with two indeterminacy
points.
Assume that X is normal; in particular, it is smooth in codimension 1. The jacobian determinant Jac(f )(x) is defined in local coordinates, on the smooth locus
of X, as the determinant of the differential dfx ; Jac(f ) depends on the coordinates,
but its zero locus does not. The zeroes of Jac(f ) form a hypersurface of the smooth
part of X; the zero locus of Jac(f ) will be defined as the Zariski closure of this
hypersurface in X. The exceptional set of f is the subset of X along which f is
not a local isomorphism onto its image; by a corollary of Zariski’s main theorem,
it coincides with the union of Ind(f ), the zero locus of Jac(f ), and additional parts
which are contained in the singular locus of X and have therefore codimension ≥ 2.
Its complement is the largest open subset on which f is a local isomorphism (see
[33, 36], for instance).
The total transform of a subset Z ⊂ X is denoted by f∗ (Z). If Z is not contained in Ind(f ), we denote by f◦ (Z) its strict transform, defined as the Zariski
closure of f (Z r Ind(f )). We say that a hypersurface W ⊂ Z is contracted if it is
not contained in the indeterminacy set and the codimension of its strict transform is
larger than 1.
2.2. Pseudo-isomorphisms. A birational map f : X 99K Y is a pseudo-isomorphism if one can find Zariski open subsets U ⊂ X and V ⊂ Y such that
(i) f realizes a regular isomorphism from U to V and
(ii) X r U and Y r V have codimension ≥ 2.
Pseudo-isomorphisms from X to itself are called pseudo-automorphisms (see
§ 1.2). The set of pseudo-automorphisms of X is a subgroup Psaut(X) of Bir(X).
Example 2.3. Start with the standard birational involution σn : Pnk 99K Pnk which
is defined in homogeneous coordinates by σn [x0 : . . . : xn ] = [x−1
: . . . : x−1
].
0
Q n
Blow-up the (n + 1) vertices of the simplex ∆n = {[x0 : . . . : xn ];
xi =
0}; this provides a smooth rational variety Xn together with a birational morphism
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
6
π : Xn → Pnk . Then, π −1 ◦ σn ◦ π is a pseudo-automorphism of Xn , and is an
automorphism if n ≤ 2.
Proposition 2.4. Let f : X 99K Y be a birational map between two (irreducible,
reduced) normal algebraic varieties. Assume that the codimension of the indeterminacy sets of f and f −1 is at least 2. Then, the following properties are equivalent:
(1) The birational maps f and f −1 do not contract any hypersurface.
(2) The jacobian determinants of f and f −1 do not vanish on the regular loci
of X r Ind(f ) and Y r Ind(f −1 ) respectively.
(3) For every smooth point q ∈ X r Ind(f ), f is a local isomorphism from a
neighborhood of q to a neighborhood of f (q), and the same holds for f −1 .
(4) The birational map f is a pseudo-isomorphism from X to Y .
Proof. Denote by g be the inverse of f . If the Jacobian determinant of f vanishes at
some (smooth) point of X r Ind(f ), then it vanishes along a hypersurface V ⊂ X.
If (1) is satisfied, the image of V is a hypersurface W in Y , and we can find a point
p ∈ V r Ind(f ) such that f (p) is not an indeterminacy point of g. Since the product
of the jacobian determinant of f at p and of g at f (p) must be equal to 1, we get
a contradiction. Thus (1) implies (2), and (2) is equivalent to (1). Now, assume
that (2) is satisfied. Then f does not contract any positive dimensional subset of
X reg r Ind(f ): f is a quasi-finite map from X reg r Ind(f ) to its image, and so is g.
Zariski’s main theorem implies that f realizes an isomorphism from X reg r Ind(f )
to Y r Ind(g) (see [33], Prop. 8.57). Thus, (2) implies (4) and (3). By assumption,
Ind(f ) and Ind(g) have codimension ≥ 2; thus, (3) implies (2). Since (4) implies
(1), this concludes the proof.
Example 2.5. Let X be a smooth projective variety with trivial canonical bundle
KX . Let Ω be a non-vanishing section of KX , and let f be a birational transformation of X. Then, f ∗ Ω extends from X r Ind(f ) to X and determines a new section
of KX ; this section does not vanish identically because f is dominant, hence it does
not vanish at all because KX is trivial. As a consequence, Jac(f ) does not vanish, f
is a pseudo-automorphism of X, and Bir(X) = Psaut(X). We refer to [10, 16] for
families of Calabi-Yau varieties with an infinite group of pseudo-automorphisms.
2.3. Projective varieties.
Proposition 2.6 (see [5]). Let f : X 99K Y be a pseudo-isomorphism between two
normal projective varieties. Then
(1) the total transform of Ind(f ) by f is equal to Ind(f −1 );
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
7
(2) f has no isolated indeterminacy point;
(3) if dim(X) = 2, then f is a regular isomorphism.
Proof. Let p ∈ X be an indeterminacy point of the pseudo-isomorphism f : X 99K
Y . Then f −1 contracts a subset C ⊂ Y of positive dimension on p. Since f and
f −1 are local isomorphisms on the complement of their indeterminacy sets, C is
contained in Ind(f −1 ). The total transform of a point q ∈ C by f −1 is a connected
subset of X that contains p and has dimension ≥ 1. This set Dq is contained in
Ind(f ) because f is a local isomorphism on the complement of Ind(f ); since p ∈
Dq ⊂ Ind(f ), p is not an isolated indeterminacy point. This proves Assertions (1)
and (2). The third assertion follows from the second one because indeterminacy
sets of birational transformations of projective surfaces are finite sets.
Let W be a hypersurface of X, and let f : X 99K Y be a pseudo-isomorphism.
The divisorial part of the total transform f∗ (W ) coincides with the strict transform
f◦ (W ). Indeed, f∗ (W ) and f◦ (W ) coincide on the open subset of Y on which f −1
is a local isomorphism, and this open subset has codimension ≥ 2.
Recall that the Néron-Severi group NS(X) is the free abelian group of codimension 1 cycles modulo cycles which are numerically equivalent to 0. Its rank is finite
and is called the Picard number of X.
Theorem 2.7. The action of pseudo-isomorphisms on Néron-Severi groups is functorial: (g ◦ f )∗ = g∗ ◦ f∗ for all pairs of pseudo-isomorphisms f : X 99K Y and
g : Y 99K Z. If X is a normal projective variety, the group Psaut(X) acts linearly
on the Néron-Severi group NS(X); this provides a morphism
Psaut(X) → GL (NS(X)).
The kernel of this morphism is contained in Aut(X) and contains Aut(X)0 as a
finite index subgroup.
As a consequence, if X is projective the group Psaut(X) is an extension of a
discrete linear subgroup of GL (NS(X)) by an algebraic group.
Proof. The first statement follows from the equality f∗ = f◦ on divisors. The
second follows from the first. To study the kernel K of the linear representation
Psaut(X) → GL (NS(X)), fix an embedding ϕ : X → Pm
k and denote by H the
m
polarization given by hyperplane sections in Pk . For every f in K, f∗ (H) is an
ample divisor, because its class in NS(X) coincides with the class of H. Now, a
theorem of Matsusaka and Mumford implies that f is an automorphism of X (see
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
8
[25] exercise 5.6, and [32]). To conclude, note that Aut(X)0 has finite index in the
kernel of the action of Aut(X) on NS(X) (see [31, 27]).
2.4. Affine varieties. The group Psaut(Ank ) coincides with the group Aut(Ank ) of
polynomial automorphisms of the affine space Ank : this is a special case of the
following proposition.
Proposition 2.8. Let Z be an affine variety. If Z is factorial, the group Psaut(Z)
coincides with the group Aut(Z).
Proof. Fix an embedding Z → Am
k . Rational functions on Z are restrictions of
m
rational functions on Ak . Thus, every birational transformation f : Z → Z is
given by rational formulas f (x1 , . . . , xm ) = (f1 , . . . , fm ) where each fi is a rational
function
pi
∈ k(x1 , . . . , xm );
fi =
qi
here, pi and qi are relatively prime polynomial functions. Since the local rings OZ,x
are unique factorization domains, we may assume that the hypersurfaces WZ (pi ) =
{x ∈ Z; pi (z) = 0} and WZ (qi ) = {x ∈ Z; qi (z) = 0} have no common
components. Then, the generic point of WZ (qi ) is mapped to infinity by f . Since f
is a pseudo-isomorphism, WZ (qi ) is in fact empty; but if qi does not vanish on Z, f
is a regular map.
3. G ROUPS WITH P ROPERTY (FW)
3.1. Commensurated subsets and cardinal definite length functions (see [11]).
Let G be a group, and G × S → S an action of G on a set S. Let A be a subset
of S. As in the Introduction, one says that G commensurates A if the symmetric
difference A4gA is finite for every element g ∈ G. One says that G transfixes A if
there is a subset B of S such that A4B is finite and B is G-invariant: gB = B for
every g in G. If A is transfixed, then it is commensurated. Actually, A is transfixed
if and only if the function g 7→ #(A4gA) is bounded on G.
A group G has Property (FW) if, given any action of G on a set S, all commensurated subsets of S are automatically transfixed. More generally, if H is a
subgroup of G, then (G, H) has relative Property (FW) if every commensurating action of G is transfixing in restriction to H. This means that, if G acts on a
set S and commensurates a subset A, then H transfixes automatically A. The case
H = G is Property (FW) for G.
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
9
We refer to [11] for a detailed study of Property (FW). The next paragraphs
present the two main sources of examples for groups with Property (FW) or its
relative version, namely Property (T) and distorted subgroups.
Remark 3.1. Property (FW) should be thought of as a rigidity property. To illustrate
this idea, consider a group K with Property (PW); by definition, this means that K
admits a commensurating action on a set S, with a commensurating subset C such
that the function g 7→ #(C4gC) has finite fibers. If G is a group with Property
(FW), then, every homomorphism G → K has finite image.
3.2. Property (FW) and Property (T). One can rephrase Property (FW) as follows: G has Property (FW) if and only if every isometric action on an “integral
Hilbert space” `2 (X, Z) has bounded orbits, where X is any discrete set.
A group has Property (FH) if all its isometric actions on Hilbert spaces have
fixed points. More generally, a pair (G, H) of a group G and a subgroup H ⊂ G
has relative Property (FH) if every isometric G-action on a Hilbert space has an
H-fixed point. Thus, the relative Property (FH) implies the relative Property (FW).
By a theorem of Delorme and Guichardet, Property (FH) is equivalent to Kazhdan’s Property (T) for countable groups (see [13]). Thus, Property (T) implies Property (FW).
Kazhdan’s Property (T) is satisfied by lattices in semisimple Lie groups all of
whose simple factors have Property (T), for instance if all simple factors have real
rank ≥ 2. For example, SL 3 (Z) satisfies Property (T).
Property (FW) is actually conjectured to hold for all irreducible lattices in semisimple Lie groups of real rank ≥ 2, such as SL 2 (R)k for k ≥ 2. (here, irreducible
means that the projection of the lattice modulo every simple factor is dense.) This
is known in the case of a semisimple Lie group admitting at least one noncompact
simple factor with Kazhdan’s Property (T), for instance in SO (2, 3) × SO (1, 4),
which admits irreducible lattices (see [12]).
3.3. Distortion. Let G be a group. An element g of G is distorted in G if there
exists a finite subset Σ of G generating a subgroup hΣi containing g, such that
limn→∞ n1 |g n |Σ = 0; here, |g|Σ is the length of g with respect to the set Σ. If G is
finitely generated, this condition holds for some Σ if and only if it holds for every
finite generating subset of G. For example, every finite order element is distorted.
Example 3.2. Let K be a field. The distorted elements of SL n (K) are exactly the
virtually unipotent elements, that is, those elements whose eigenvalues are all roots
of unity; in positive characteristic, these are elements of finite order. By results of
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
10
Lubotzky, Mozes, and Raghunathan (see [29, 28]), the same characterization
holds
√
in the group SL n (Z), as soon as n ≥ 3; it also holds in SL n (Z[ d]) when n ≥ 2
and d ≥ 2 is not a perfect square. In contrast, in SL 2 (Z), every element of infinite
order is undistorted.
Lemma 3.3 (see [11]). Let G be a group, and H a finitely generated abelian subgroup of G consisting of distorted elements. Then, the pair (G, H) has relative
Property (FW).
This lemma provides many examples. For instance, if G is any finitely generated
nilpotent group and G0 is its derived subgroup, then (G, G0 ) has relative Property
(FH); this result is due to Houghton, in a more general formulation encompassing
polycyclic groups (see [11]). Bounded generation by distorted unipotent elements
can also be used to obtain nontrivial examples of groups with √
Property (FW), including√the above examples SL n (Z) for n ≥ 3, and SL n (Z[ d]). The case of
SL 2 (Z[ d]) is particularly interesting because it does not have Property (T).
3.4. Subgroups of PGL 2 (k) with Property (FW). If a group G acts on a tree T
by graph automorphisms, then G acts on the set E of directed edges of T (T is
non-oriented, so each edge gives rise to a pair of opposite directed edges). Let Ev
be the set of directed edges pointing towards a vertex v. Then Ev 4Ew is the set
of directed edges lying in the segment between v and w; it is finite of cardinality
2d(v, w), where d is the graph distance. The group G commensurates the subset Ev
for every v, and #(Ev 4gEv ) = 2d(v, gv). As a consequence, if G has Property
(FW), then it has Property (FA) in the sense that every action of G on a tree has
bounded orbits. This argument can be combined with Proposition 5.B.1 of [11] to
obtain the following lemma.
Lemma 3.4 (See [11]). Let G be a group with Property (FW), then all finite index
subgroups of G have Property (FW), and hence have Property (FA). Conversely, if
a finite index subgroup of G has Property (FW), then so does G.
On the other hand, Property (FA) is not stable by taking finite index subgroups.
Lemma 3.5. Let k be an algebraically closed field and Λ be a subgroup of GL 2 (k).
(1) Λ has a finite orbit on the projective line if and only if it is virtually solvable,
if and only if its Zariski closure does not contain SL 2 .
(2) Assume that all finite index subgroups of Λ have Property (FA) (e.g., Λ has
Property FW). If the action of Λ on the projective line preserves a nonempty, finite set, then Λ is finite.
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
11
The proof of the first assertion is standard and omitted. The second assertion
follows directly from the first one.
In what follows, we denote by Z ⊂ Q the ring of algebraic integers (in some
fixed algebraic closure Q of Q).
Theorem 3.6 (Bass [2]). Let k be an algebraically closed field.
(1) If k has positive characteristic, then GL 2 (k) has no infinite subgroup with
Property (FA).
(2) Suppose that k has characteristic zero and that Γ ⊂ GL 2 (k) is a countable
subgroup with Property (FA), and is not virtually abelian. Then Γ acts irreducibly on k2 , and is conjugate to a subgroup of GL 2 (Z). If moreover
Γ ⊂ GL 2 (K) for some subfield K ⊂ k containing Q, then we can choose
the conjugating matrix to belong to GL 2 (K).
On the proof. The original statement [2, Theorem 6.5] yields this statement, except
the last fact, and assumes that Γ is contained in GL 2 (M ) with M a finitely generated
field. The latter condition is actually automatic: indeed, being a countable group
with Property (FA), Γ is finitely generated [35, §6, Th. 15], and one can choose K
to be the field generated by entries of a finite generating subset.
For the last assertion, we have Γ ∪ BΓB −1 ⊂ GL 2 (K) for some B ∈ GL 2 (k)
such that BΓB −1 ⊂ GL 2 (Z); we claim that this implies that B ∈ k∗ GL 2 (K). First,
since Γ is absolutely irreducible, this implies that BM2 (K)B −1 ⊂ M2 (K). The
conclusion follows from Lemma 3.7 below, which can be of independent interest.
Lemma 3.7. Let K ⊂ L be fields. Then the normalizer {B ∈ GL 2 (L) : BM2 (K)B −1 ⊂
M2 (K)} is reduced to L∗ GL 2 (K) = {λA : λ ∈ L∗ , A ∈ GL 2 (K)}.
Proof. Write
b 1 b2
B=
.
b3 b4
Since BAB −1 ∈ M2 (K) for the three elementary matrices A ∈ {E11 , E12 , E21 },
we deduce by a plain computation that bi bj /bk b` ∈ K for all 1 ≤ i, j, k, ` ≤ 4 such
that bk b` 6= 0. In particular, for all indices i and j such that bi and bj are nonzero,
the quotient bi /bj = bi bj /b2j belongs to K. It follows that B ∈ L∗ GL 2 (K).
Corollary 3.8. Let k be an algebraically closed field. Let C be a projective curve
over k, and let k(C) be the field of rational functions on the curve C. Let Γ be an
infinite subgroup of PGL 2 (k(C)). If Γ has Property (FA), then
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
12
(1) the field k has characteristic 0;
(2) there is an element of PGL 2 (k(C)) that conjugates Γ to a subgroup of
PGL 2 (Z) ⊂ PGL 2 (k(C)).
4. A CATEGORAL LIMIT CONSTRUCTION
The purpose of this section is to describe a general categorical construction,
which can be used to construct various actions of groups of birational transformations, such as Manin’s construction of the Picard-Manin space (see [30, 8]), as
well as the commensurating action which is the main construction of this paper. A
closely related construction is performed by V. Jones in [24] to construct representations of Thompson’s groups, although it does not directly apply here.
4.1. Categories of projective varieties. Here, in a category C, arrows between any
two objects X and Y are assumed to form a set HomC (X, Y ). Recall that a category
is small if its objects form a set, and is essentially small if it is equivalent to a small
category, or equivalently if there is a subset of the collection of objects meeting
every isomorphism class. A subcategory C of a category D is full if all arrows of D
between objects of C are also arrows of C.
Example 4.1. Our main example will be the following. Fix an algebraically closed
field k. Let V = Vk be the category whose objects are irreducible (reduced) projective k-varieties and whose arrows are birational morphisms. Let V ] be the category
with the same objects, but whose arrows are birational maps. Similarly, one can
consider the category VN of irreducible (reduced) normal projective varieties, with
arrows given by birational morphisms, and the category VN ] with the same objects
but whose arrows are all birational maps. By construction, VN is a full subcategory
of V, which is a subcategory of V ] .
4.2. Relative thinness and well-cofiltered categories. Given a category C and an
object X ∈ Ob(C), let us define the category CX whose objects are pairs (Y, f ) with
Y ∈ Ob(C) and f ∈ HomC (Y, X), and whose arrows (Y, f ) → (Z, g) are given by
arrows u ∈ HomC (Y, Z) such that g ◦ u = f . A category is thin if there is at most
one arrow between any two objects. Let us say that a category is relatively thin if
the category CX is thin for all X ∈ Ob(X).
Example 4.2. A category in which every arrow is invertible is relatively thin, and
so are all its subcategories. This applies to the categories of Example 4.1: the
category Vk] of birational maps between irreducible projective varieties, and to its
subcategory Vk , and similarly to VN ]k and its subcategory VN k .
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
13
Recall that a category is cofiltered if it satisfies the following two properties (a)
and (b):
(a) for any pair of objects X1 , X2 , there exists an object Y with arrows X1 ←
Y → X2 ;
(b) for any pair of objects X, Y and arrows u1 , u2 : X → Y , there exists an
object W and an arrow w : W → X such that u1 ◦ w = u2 ◦ w.
Note that (b) is automatically satisfied when the category is thin. We say that a
category C is well-cofiltered if it is relatively thin and for every object X ∈ Ob(C),
the category CX is cofiltered (note that we do not require C to be cofiltered).
Example 4.3. Coming again to the categories of Example 4.1, the category Vk is
essentially small and well-cofiltered. It is relatively thin, as mentioned in Example
4.2. To show that (Vk )X is cofiltered, consider two birational morphisms f1 : X1 →
X and f2 : X2 → X, and denote by h the composition f2−1 ◦ f1 . The graph Gh
is a projective subvariety of X1 × X2 . One can compose the projection of Gh
onto X1 with f1 (resp. onto X2 with f2 ) to get a birational morphism Gh → X;
this birational morphism is an object in (Vk )X that dominates f1 : X1 → X and
f2 : X2 → X, as in property (a).
The full subcategory VN k of Vk enjoys the same properties. When k has characteristic zero, the resolution of indeterminacies implies that its full subcategory of
non-singular varieties (and birational morphisms) is also well-cofiltered.
4.3. Filtering inductive limits.
4.3.1. We shall say that a category E admits filtering inductive limits if for every
small, thin and cofiltered category D and every contravariant functor F : D →
E, the colimit of F exists (and then it also exists when “small” is replaced with
“essentially small”). For example, the category of sets and the category of groups
admit filtering inductive limits (see [36], § 1.4, for colimits).
4.3.2. Let us consider an essentially small category C, a category E admitting
filtering inductive limits, and a contravariant functor F : C → E; we denote the
functor F by X 7→ FX on objects and u 7→ Fu on arrows. Assume that C is wellcofiltered. Then, for every object X ∈ Ob(C), we can restrict the functor F to CX
and take the colimit F̃X of this restriction F : CX → E. Roughly speaking, F̃X is
the inductive limit in E of all FY for Y ∈ CX . So, for every arrow u : Y → X in C,
there is an arrow in E, φu : FY → F̃X in E; and for every arrow v : Z → Y in CX ,
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
14
the following diagram commutes
FO Z
Fv
FY
φu◦v
φu
/
!
F̃X .
The colimits F̃X satisfy a universal property. To describe it, consider an object
E ∈ Ob(E), together with arrows ψY : FY → E for all Y ∈ CX , and assume that
for every arrow v : Z → Y in CX we have the relation ψZ ◦ Fv = ψY . Then, there
exists a unique arrow ψ : F̃X → E in E such that for every (Y, u) ∈ Ob(CX ) the
following diagram commutes:
F̃X
O
ψ
Fu
/
>E
ψY
FY .
This construction provides a bijection ΦX from the inductive limit lim
←−Y ∈CX HomE (FY , E)
to Hom(F̃X , E) whose reciprocal bijection maps an element ψ ∈ Hom(F̃X , E) to
the family of arrows (ψ ◦ Fu )(Y,u)∈CX .
4.3.3. We can now define the covariant functor α associated to F . At the level
of objects, α maps X ∈ Ob(C) to the limit F̃X . Let us now describe α at the
level of arrows. If we fix (Y, u) ∈ Ob(CX ), the family of arrows (φu◦v : FZ →
F̃X )(Z,v)∈Ob(CY ) corresponds under ΦY to an arrow αu : F̃Y → F̃X . For every
(Z, v) ∈ Ob(CY ), the following diagram commutes
F̃OY
φv
αu
/
F̃
= X
(4.1)
φu◦v
FZ ,
and this characterizes the arrow αu . If (W, w) ∈ CY , the uniqueness readily proves
that αu◦w = αu ◦ αw , that is, X 7→ F̃X , u 7→ αu is a covariant functor C → E,
denoted α, and called the relative colimit functor associated to F .
4.3.4. Note that the previous diagram can be essentially rewritten in the form of
the commutative square on the left of the next equation. The commutative diagram
on the right only refers to u; it is obtained by composing the left diagram with the
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
15
map Fu and by using the equalities φv ◦ Fv = φidY and φu ◦ Fu = φidX .
αu
F̃OY
/
φv
F̃X
O
Fv
/
F̃X
O
φidY
φu
FZ o
αu
F̃OY
φidX
FY o
FY
(4.2)
Fu
FX .
Lemma 4.4. Suppose that C is well-cofiltered. Then α maps arrows of C to invertible arrows (i.e. isomorphisms) of E.
Proof. Fix (Y, u) ∈ CX . The proof consists in constructing a map, and then show
that it is the inverse map of αu .
Consider (S, s) ∈ Ob(CX ). By assumption, in C we can find a commutative
diagram as the one on the left of the following equation; hence, in E we obtain the
diagram on the right, where g = φx ◦ Fw by definition.
Y o
u
Xo
x
D
s
Fx
FOY
w
φx
FD
O
Fw
Fu
S,
/
Fs
FX
/
/
F̃
= Y
g
FS .
A priori g depends on the choice of (D, x, w); let us show that it only depends on
(u, s) and, for that purpose, let us denote g temporarily by g = gD (x, w being
implicit). First consider the case of a commutative diagram as the one on the left
in the next equation; in E, this diagram induces the diagram depicted on the right,
where everything not involving gD or gD0 is commutative.
Y `o
x
x0
q
D0
u
D
Fx
FOY
Fx0
~
w
!
Fu
v
Fq
/4
F̃
<D Y
φx 0
F a
S,
φx
FO
= D
Fw
D0
w0
Xo
/
Fs
FX
Fw0
/
gD ,gD0
FS ,
Thus, by definition gD0 = φx0 ◦ Fw0 = φx ◦ Fq ◦ Fw0 = φx ◦ Fw = gD . Now consider,
more generally two objects D0 and D00 and four arrows forming a diagram in CX :
YO o
D00
/
D0
S;
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
16
we have to show that gD0 = gD00 . Since C is well-cofiltered, CX is thin and cofiltered,
and we can complete the previous diagram into the one on the left of the following
equation. Since this diagram is in CX which is a thin category, it is commutative;
which means that if we complete it with both composite arrows D → Y and both
composite arrows D → S, the resulting arrows coincide; the resulting diagram, on
the right of the equation, is a commutative one.
DO 0
DO 0
Y a
}
!
D
>
Y ao
S
}
D
/
!
>
S
D00 ;
D00 ;
Using the previous case, we deduce gD0 = gD = gD00 . Thus, we have seen that
gD does not depend on the choice of D; we now write it as gu,s . In particular,
when S ∈ (CX )Y , we can choose D = S (and w the identity); we thus deduce that
gu,s = φx (where s = u ◦ x).
x
w
Consider (T, t) ∈ CS and choose D ∈ Ob(CX ) with Y ← D → T in CX . Then
we have the diagram in E
FOY
Fx
/
FD
O
φx
Fw
FO T
gu,s◦t
Ft
FX
Fs
/
!
/
F̃
> Y
gu,s
FS ,
where the left rectangle is commutative as well as the upper right triangle; since
by definition gu,s = φx ◦ Ft◦w = φx ◦ Fw ◦ Ft = gu,s◦t ◦ Ft , the lower right
triangle is also commutative. So the family (gs : FS → F̃Y )(S,s)∈Ob(CX ) defines
an element gu : F̃X → F̃Y . Namely, for every (S, s) ∈ Ob(CX ) the following
diagram commutes
F̃X
O
φs
FS ,
and gu is characterized by this property.
gu
/
F̃
> Y
gu,s
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
17
We now combine this with the map αu , and make use of the same notation as the
one in Equations (4.1) and (4.2). When S = Z (so s = u ◦ v, gu,s = φv ), we obtain
the commutative diagram.
F̃Y a
αu
/
gu
F̃X
O
φs
φv
/
F̃
= Y
φv
FZ ,
Since this holds for all (Z, v) ∈ (CX )Y , the universal property of F̃Y implies that
gu ◦ αu is the identity of F̃Y .
On the other hand, turning back to the notation of the beginning of the proof,
both triangles in the following diagram are commutative
αu
F̃OY
φx
/
F̃O
= X
φu◦x
FD o
Fw
φs
FS ;
since gu,s = φx ◦ Fw , this implies that the right triangle of the following diagram is
commutative, the left-hand triangle from above also being commutative
F̃X a
gu
/
F̃OY
αu
/
F̃
= X
gu,s
φs
φs
FS .
Since this holds for all (S, s) ∈ CX , by the universal property of F̃X , we obtain that
αu ◦ gu is the identity of F̃X . This ends the proof that αu is invertible.
4.4. Good right-localization and extensions. Given a category D with a subcategory C with the same objects, we say that (C, D) is a good right-localization if (i)
every arrow u : X → Y in D admits an inverse u−1 : Y → X and (ii) every arrow
in D can be decomposed as g ◦ f −1 where f and g are arrows of C.
Lemma 4.5. Let (C, D) be a good right-localization. Then C is well-cofiltered.
Proof. Clearly any category in which all arrows are invertible is relatively thin. It
follows that D and its subcategory C are relatively thin. Now consider a pair of
objects (Y, u) and (Z, v) of CX . Then v −1 ◦ u is an arrow of D (because all arrows
s
t
are invertible in D), and it can be decomposed as Y ← W → Z, with s and t arrows
of C. By definition v ◦ t = u ◦ s determines an arrow W → X; endowing W with
the resulting composite arrow to X, the arrows s and t become arrows in CX .
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
18
Lemma 4.6. Let (C, D) be a good right-localization, and let E be another category.
Consider a (covariant) functor β from C to the category E mapping every arrow
to an invertible arrow of E. Then β has a unique extension to a functor from the
category D to the category E.
Proof. The uniqueness is clear. For the existence, consider an arrow u in D. We
wish to map u to β(g) ◦ β(f )−1 , where u = g ◦ f −1 . We have to prove that this does
not depend on the choice of (f, g). Thus write u = g1 ◦ f1−1 = g2 ◦ f2−1 . Since CX is
well-cofiltered (Lemma 4.5), we can produce a diagram as follows in C, where the
left “square” and the whole square are commutative:
f1
X`
~
YO 1
h1
g1
Z
f2
h2
Y
> 2
g2
X2
Then the right square is also commutative: indeed g1 ◦ h1 = (g1 ◦ f1−1 ) ◦ (f1 ◦ h1 ) =
(g2 ◦ f2−1 ) ◦ (f2 ◦ h2 ) = g2 ◦ h2 . Then
β(g1 ) ◦ β(f1 )−1 = β(g1 ) ◦ β(h1 ) ◦ β(h1 )−1 ◦ β(f1 )−1 = β(g1 ◦ h1 ) ◦ β(f1 ◦ h1 )−1
= β(g2 ◦ h2 ) ◦ β(f2 ◦ h2 )−1 = β(g2 ) ◦ β(h2 ) ◦ β(h2 )−1 ◦ β(f2 )−1 = β(g2 ) ◦ β(f2 )−1 ;
hence we can define without ambiguity β(u) = β(g) ◦ β(f )−1 .
We have to prove β(v) ◦ β(u) = β(v ◦ u) for any arrows u, v of D. This already
holds for u, v arrows of C. Write u = g ◦ f −1 , v = j ◦ h−1 with f , g, h, and j
arrows of C. Write h−1 ◦ g = t ◦ s−1 with s, t arrows of C. Then g ◦ s = h ◦ t,
so β(g) ◦ β(s) = β(g ◦ s) = β(h ◦ t) = β(h) ◦ β(t), which can be rewritten
β(h)−1 ◦ β(g) = β(t) ◦ β(s)−1 . In turn, we get
β(v ◦ u) = β(j ◦ h−1 ◦ g ◦ f −1 ) = β(j ◦ t ◦ s−1 ◦ f −1 ) = β(j ◦ t) ◦ β(f ◦ s)−1
= β(j) ◦ β(t) ◦ β(s)−1 ◦ β(f )−1 = β(j) ◦ β(h)−1 ◦ β(g) ◦ β(f )−1 = β(v) ◦ β(u).
Combining Lemmas 4.4, 4.5 and 4.6, we deduce:
Proposition 4.7. Consider a good right-localization (C, D) and a category E admitting filtering inductive limits. Let F be a contravariant functor from C to E. Then
the relative colimit functor α, defined by
X 7→ F̃X = −
lim
→Y →X F (Y ),
u 7→ αu ,
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
19
has a unique extension to a covariant functor from D to E.
5. I RREDUCIBLE HYPERSURFACES
Let X be a normal projective variety. In this chapter, we make use of the categorical construction of Section 4.3 to define an action of Bir(X) on the set of all
irreducible hypersurfaces in all “models” Y → X of X. We also sketch an application to the construction of Picard-Manin spaces in Section 5.5.
5.1. Localizations in categories of projective varieties. Consider the categories
V, V ] , VN , and VN ] from examples 4.1 and 4.3.
Proposition 5.1. Let k be an algebraically closed field.
(1) The categories V and VN are well-cofiltered (as defined in §4.2).
(2) The pairs (V, V ] ) (VN , VN ] ) are good right-localizations of categories (as
defined in §4.4).
Proof. Property (1) follows from Example 4.3. Let us now prove the second property. Clearly, every arrow in V ] is invertible. Let f : X 99K X 0 be a birational
g0
g
map. The graph Gf is an irreducible variety, both projections X ← Gf → X 0 are
g
g0
birational morphisms and f = g 0 ◦ g −1 . Since f = g 0 ◦ g −1 with X ← Gf → X 0 ,
we deduce that (V, V ] ) is a good right-localization.
For VN , we only need to compose with the normalization map Y → Gf to get
the result (see [36], §9.7).
To spare the reader from going into too much category theory, let us state explicitly Proposition 4.7 in this case:
Corollary 5.2. Let E be a category admitting filtering inductive limits. Consider a
contravariant functor F from VN to the category E. Then the covariant functor α,
]
X 7→ F̃X = −
lim
→Y →X F (Y ), u 7→ αu has a unique extension to a functor from VN
to E.
5.2. The functor of irreducible hypersurfaces. Let us construct the functor to
which we will apply Corollary 5.2. For X ∈ V, define Hyp(X) as the set of irreducible and reduced hypersurfaces of X.
Proposition 5.3. Let f : Y → X be a birational morphism between two irreducible
projective varieties (an arrow in V).
(1) The number of T ∈ Hyp(Y ) such that f (T ) is not a hypersurface is finite
(these T are precisely the hypersurfaces contracted by f ).
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
20
(2) For every S ∈ Hyp(X), the number of S 0 ∈ Hyp(Y ) such that f (S 0 ) = S is
positive and finite.
(3) If X is normal, this number is equal to 1, and the unique preimage S 0 of S
is the strict transform f ◦ (S) = (f −1 )◦ (S). In particular, S 7→ S 0 = f ◦ (S)
is injective and its image has finite complement: this complement Hyp(Y ) r
f ◦ (Hyp(X)) is the set of hypersurfaces T that are contracted by f .
Proof. Let U ⊂ Y be a Zariski-dense open subset on which f induces an isomorphism onto its image. Let F be the complement of U .
(1) If f (T ) is not a hypersurface, then T ⊂ F . So T is one of the irreducible
components of F , which leaves finitely many possibilities.
(2) Since Y is projective, f is surjective, and hence f (T ) = S, where T =
−1
f (S); T is a proper subvariety. Then at least one irreducible component T 0 of T
satisfies f (T 0 ) = S, and conversely, every S 0 ∈ Hyp(Y ) such that f (S 0 ) = S has to
be an irreducible component of T , hence there are finitely many of them.
(3) We now use Theorem 2.1: since X is normal and Y is projective, the indeterminacy set Ind(f −1 ) ⊂ X has codimension ≥ 2. Hence the strict transform
(f −1 )◦ (S) of S is well-defined and is equal to f −1 (X r Ind(f −1 ). The total transform of S by f −1 may contain additional components of codimension 1, but all of
them are contracted into Ind(f −1 ), which has codimension 2 (hence are not equal
to S). This proves that S 0 = (f −1 )◦ (S). Since f (S 0 ) = S, the map f ◦ := (f −1 )◦ is
injective. Moreover, by construction, its image is made of hypersurfaces which are
not contained in f −1 (Ind(f )). Since every element T ∈ Hyp(Y ) which is not contracted coincides with f ◦ (f (T )), the image of f ◦ is in fact equal to the complement
of the set of contracted hypersurfaces.
From Proposition 5.3, the map X 7→ Hyp(X) defines a contravariant functor
from VN to the category of sets, mapping an arrow f : Y → X to the (injective)
map
f ◦ : Hyp(X) → Hyp(Y ).
If F denoted the functor Hyp as in Section 4.3, then we would have FX = Hyp(X)
˜
and Ff = f ◦ . For X in the category VN , define Hyp(X)
as the filtering inductive
limit
˜
Hyp(X)
= lim
−→Y →X Hyp(Y ).
˜
By construction, X 7→ Hyp(X)
is a covariant functor from VN to the category
of sets. By Corollary 5.2, it has a unique extension to a functor from VN ] to the
˜
category of sets. The image of an object Y ∈ VN ] is denoted Hyp(Y
), and the
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
21
image of an arrow f : Y 99K Y 0 , that is of a birational map between two normal
projective varieties Y and Y 0 , is denoted by f• . By construction, f• is a bijection
0
˜
˜
from Hyp(Y
) to Hyp(Y
).
For an arrow u in VN , i.e. a birational morphism u : Y → X between two normal projective varieties, we rewrite the commutative square on the right of Equation (4.2) as
u•
˜
˜
Hyp(Y
) ∼ / Hyp(X)
O
O
iY
?
Hyp(Y ) o
u◦
?
iX
? _ Hyp(X).
The two injections iX and iY will simply be viewed as inclusions in what follows. So the bijection u−1
• extends the injection which is given by the strict trans◦
form u . Since the image of u◦ has finite complement, the symmetric difference
Hyp(Y )4u−1
• (Hyp(X)) is finite. This latter property passes to inverses and compositions; hence, for every birational map v : X 99K X 0 between normal irreducible
projective varieties, the symmetric difference Hyp(X)4v•−1 (Hyp(X 0 )) is finite. To
give a more precise statement, let us introduce the following notation: given a birational map v : X 99K X 0 between normal irreducible projective varieties, define
exc(v) by
exc(v) = # {S ∈ Hyp(X); v contracts S} .
This is the number of contracted hypersurfaces S ∈ Hyp(X) by v.
Proposition 5.4. Let v : X 99K X 0 be a birational transformation between normal
irreducible projective varieties. Let S be an element of Hyp(X).
(1) If S ∈ (v −1 )◦ Hyp(X 0 ), then v• (S) = v◦ (S) ∈ Hyp(X 0 ).
(2) If S ∈
/ (v −1 )◦ Hyp(X 0 ), then v◦ (S) has codimension ≥ 2 (i.e. v contracts
0
˜
S), and v• (S) is an element of Hyp(X
) r Hyp(X 0 ).
(3) The symmetric difference v• (Hyp(X))4Hyp(X 0 ) contains exc(v)+exc(v −1 )
elements.
Proof. Let U be the complement of Ind(v) in X 0 . Since, by Theorem 2.1, Ind(v)
has codimension ≥ 2, no S ∈ Hyp(X) is contained in Ind(v).
Let us prove (1). When v is a birational morphism the assertion follows from
Proposition 5.3. To deal with the general case, write v = g ◦ f −1 where f : Y → X
and g : Y → X 0 are birational morphisms from a normal variety Y . Since f is a
birational morphism, f • (S) = f ◦ (S) ⊂ Hyp(Y ); since S is not contracted by v,
g• (f ◦ (S)) = g◦ (f ◦ (S)) ∈ Hyp(X 0 ). Thus, v• (S) = g• (f • (S)) coincides with the
strict transform v◦ (S) ∈ Hyp(X 0 ).
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
22
Now let us prove (2), assuming thus that S ∈
/ (v −1 )◦ Hyp(X 0 ). Let S 00 ∈ Hyp(Y )
be the hypersurface (f −1 )• (S) = (f −1 )◦ (S). Then f (S 00 ) = S. If g◦ (S 00 ) is a
hypersurface S 0 , then (v −1 )◦ (S 0 ) = S, contradicting S ∈
/ (v −1 )◦ Hyp(X 0 ). Thus,
g contracts S 00 onto a subset S 0 ⊂ X 0 of codimension ≥ 2. Since S 0 = v◦ (S),
assertion (2) is proved.
Assertion (3) follows from the previous two assertions.
Example 5.5. Let g be a birational transformation of Pnk of degree d, meaning that
g ∗ (H) ' dH where H denotes a hyperplane of Pnk , or equivalently that g is defined
by n + 1 homogeneous polynomials of the same degree d without common factor of
positive degree. The exceptional set of g has degree (n+1)(d−1); as a consequence,
excPnk (g) ≤ (n+1)(d−1). More generally, if H is a polarization of X, then excX (g)
is bounded from above by a function that depends only on the degree degH (g) :=
(g ∗ H) · H dim(X)−1 .
˜
5.3. Action of Bir(X) on Hyp(X).
Let us now restrict the functor f 7→ f• to the
elements of Bir(X). The existence of this functor and Proposition 5.4 give the
following theorem.
Theorem 5.6. Let X be a normal projective variety. The group Bir(X) acts faith˜
fully by permutations on the set Hyp(X)
via the homomorphism
˜
Bir(X) → Perm(Hyp(X))
g
7→
g•
˜
This action commensurates the subset Hyp(X) of Hyp(X):
for every g ∈ Bir(X),
|g• (Hyp(X))4Hyp(X)| = exc(g) + exc(g −1 ).
The only thing that has not been proven yet is the fact that the homomorphism
˜
f ∈ Bir(X) 7→ Perm(Hyp(X))
is injective. But the kernel of this homomorphism
is made of birational transformations f such that f◦ (W ) = W for every irreducible
hypersurface W of X. Since X is projective, one can embed X in some projective
space Pm
k ; then, every point of X(k) is the intersection of finitely many irreducible
hyperplane sections of X: since all these sections are fixed by f , every point is fixed
by f , and f is the identity.
5.4. Products of varieties. Let X, Y be irreducible, normal projective varieties.
We consider the natural embedding of Bir(X) into Bir(X × Y ), given by the birational action f · (x, y) = (f (x), y), f ∈ Bir(X). There is a natural injection jY of
Hyp(X) into Hyp(X × Y ), given by S 7→ S × Y , which naturally extends to an
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
23
˜
˜
injection of Hyp(X)
into Hyp(X
× Y ); this inclusion is Bir(X)-equivariant. The
following proposition will be applied to Corollary 6.7.
Proposition 5.7. Let a group Γ act on X by birational transformations. Then Γ
˜
˜
transfixes Hyp(X) in Hyp(X)
if and only if it transfixes Hyp(X ×Y ) in Hyp(X
×Y ).
More precisely, the subset Hyp(X × Y ) r jY (Hyp(X)) is Bir(X)-invariant.
Proof. The reverse implication is immediate, since any restriction of a transfixing
action is transfixing. The direct implication follows from the latter statement, which
we now prove. Consider S ∈ Hyp(X × Y ) r jY (Hyp(X)). This means that S is an
irreducible hypersurface of X × Y whose projection to X is surjective. Now, for
γ ∈ Bir(X), γ induces an isomorphism between open dense subset U, V of X, and
hence between U × Y and V × Y ; in particular, γ does not contract S. This shows
that γ stabilizes Hyp(X × Y ) r jY (Hyp(X)).
5.5. Manin’s construction. Instead of looking at the functor X 7→ Hyp(X) from
the category of normal projective varieties to the category of sets, one can consider
the Néron-Severi functor X 7→ NS(X) from the category of smooth projective varieties to the category of abelian groups. In characteristic zero, or for surfaces in arbitrary characteristic, the resolution of singularities shows that smooth projective varieties, together with birational morphisms, form a good right localization of the category of smooth projective varieties with birational maps between them. Thus, one
˜
can construct a functor, the relative colimit of Néron-Severi groups, X 7→ NS(X)
˜
˜
that maps birational maps X 99K Y to group isomorphisms NS(X)
→ NS(Y
). In
dimension 2, this construction is known as the Picard-Manin space (see [8, 30]).
One may also replace NS(X) by other cohomology groups if they behave contravariantly with respect to birational morphisms (see [7] for instance).
6. P SEUDO - REGULARIZATION OF BIRATIONAL TRANSFORMATIONS
˜
In this section, we make use of the action of Bir(X) on Hyp(X)
to characterize
and study groups of birational transformations that are pseudo-regularizable, in the
sense of Definition 1.1. As before, k is an algebraically closed field.
6.1. An example. Consider the birational transformation f (x, y) = (x + 1, xy) of
P1k × P1k . The vertical curves Ci = {x = −i}, i ∈ Z, are exceptional curves for the
cyclic group Γ = hf i: each of these curves is contracted by an element of Γ onto a
point, namely f◦i+1 (Ci ) = (1, 0).
Let ϕ : Y 99K P1k × P1k be a birational map, and let U be a non-empty open subset
of Y . Consider the subgroup ΓY := ϕ−1 ◦ Γ ◦ ϕ of Bir(Y ). If i is large enough,
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
24
ϕ◦−1 (Ci ) is an irreducible curve Ci0 ⊂ Y , and these curves Ci0 are pairwise distinct,
so that most of them intersect U. For positive integers m, f i+m maps Ci onto (m, 0),
and (m, 0) is not an indeterminacy point of ϕ−1 if m is large. Thus, ϕ−1 ◦ f m ◦ ϕ
contracts Ci0 , and ϕ−1 ◦ f m ◦ ϕ is not a pseudo-automorphism of U. This argument
proves the following lemma.
Lemma 6.1. Let X be the surface P1k ×P1k . Let f : X 99K X be defined by f (x, y) =
(x + 1, xy), and let Γ be the subgroup generated by f ` , for some ` ≥ 1. Then the
cyclic group Γ is not pseudo-regularizable.
This shows that Theorem A requires an assumption on the group Γ. More generally, consider a subgroup Γ ⊂ Bir(X) such that Γ
(a) contracts a family of hypersurfaces Wi ⊂ X whose union is Zariski dense
(b) the union of the family of strict transforms f◦ (Wi ), for f ∈ Γ contracting
Wi , form a subset of X whose Zariski closure has codimension at most 1.
Then, Γ cannot be pseudo-regularized.
6.2. Characterization of pseudo-Isomorphisms. Recall that f• denotes the bi0
˜
˜
jection Hyp(X)
→ Hyp(X
) which is induced by a birational map f : X 99K X 0 .
Also, for any nonempty open subset U ⊂ X, we define Hyp(U ) = {H ∈ Hyp(X) :
H ∩ U 6= ∅}; it has finite complement in Hyp(X).
Proposition 6.2. Let f : X 99K X 0 be a birational map between normal projective
varieties. Let U ⊂ X and U 0 ⊂ X 0 be two dense open subsets. Then, f induces a
pseudo-isomorphism U → U 0 if and only if f• (Hyp(U )) = Hyp(U 0 ).
Proof. If f restricts to a pseudo-isomorphism U → U 0 , then f maps every hypersurface of U to a hypersurface of U 0 by strict transform. And (f −1 )◦ is an inverse
for f◦ : Hyp(U ) → Hyp(U 0 ). Thus, f• (Hyp(U )) = f◦ (Hyp(U ) = Hyp(U 0 ).
Let us now assume that f• (Hyp(U )) = Hyp(U 0 ). Since X and X 0 are normal,
Ind(f ) and Ind(f −1 ) have codimension ≥ 2 (Theorem 2.1).
Let fU,U 0 be the birational map from U to U 0 which is induced by f . The indeterminacy set of fU,U 0 is contained in the union of the set Ind(f ) ∩ U and the set
of points x ∈ U r Ind(f ) which are mapped by f in the complement of U 0 ; this
second part of Ind(fU,U 0 ) has codimension 2, because otherwise there would be an
irreducible hypersurface W in U which would be mapped in X 0 r U 0 , contradicting
the equality f• (Hyp(U )) = Hyp(U 0 ). Thus, the indeterminacy set of fU,U 0 has codimension ≥ 2. Changing f in its inverse f −1 , we see that the indeterminacy set of
fU−10 ,U : U 0 99K U 0 has codimension ≥ 2 too.
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
25
If fU,U 0 contracted an irreducible hypersurface W ⊂ U onto a subset of U 0 of
codimension ≥ 2, then f• (W ) would not be contained in Hyp(U 0 ) (it would corre0
˜
spond to an element of Hyp(X
)rHyp(X 0 ) by Proposition 5.4). Thus, fU,U 0 satisfies
the first property of Proposition 2.4 and, therefore, is a pseudo-isomorphism.
6.3. Characterization of pseudo-regularization. Let X be a (irreducible, reduced)
normal projective variety. Let Γ be a subgroup of Bir(X). Assume that the action
˜
˜
of Γ on Hyp(X)
fixes (globally) a subset A ⊂ Hyp(X)
such that
|A4Hyp(X)| < +∞.
In other words, A is obtained from Hyp(X) by removing finitely many hypersur˜
faces Wi ∈ Hyp(X) and adding finitely many hypersurfaces Wj0 ∈ Hyp(X)
r
0
Hyp(X). Each Wj comes from an irreducible hypersurface in some model πj : Xj →
X, and there is a model π : Y → X that covers all of them (i.e. π ◦ πj−1 is a morphism from Y to Xj for every j). Then, π ◦ (A) is a subset of Hyp(Y ). Changing X
into Y , A into π ◦ (A), and Γ into π −1 ◦ Γ ◦ π, we may assume that
(1) A = Hyp(X) r {E1 , . . . , E` } where the Ei are ` distinct irreducible hypersurfaces of X,
˜
(2) the action of Γ on Hyp(X)
fixes the set A.
In what follows, we denote by U the non-empty Zariski open subset X r ∪i Ei and
by ∂X the boundary X r U = E1 ∪ · · · ∪ E` ; ∂X is considered as the boundary of
the compactification X of U.
Lemma 6.3. The group Γ acts by pseudo-automorphisms on the open subset U. If
U is smooth and there is an ample divisor D whose support coincides with ∂X, then
Γ acts by automorphisms on U.
In this statement, we say that the support of a divisor D coincides with ∂X if
P
D = i ai Ei with ai > 0 for every 1 ≤ i ≤ `.
Proof. Since A = Hyp(U) is Γ-invariant, Proposition 6.2 shows that Γ acts by
pseudo-automorphisms on U.
Since D is an ample divisor, some positive multiple mD is very ample, and the
complete linear system |mD| provides an embedding of X in a projective space.
The divisor mD corresponds to a hyperplane section of X in this embedding, and
the open subset U is an affine variety because the support of D is equal to ∂X.
Proposition 2.8 concludes the proof of the lemma.
˜
By Theorem 5.6, every subgroup of Bir(X) acts on Hyp(X)
and commensu˜
rates Hyp(X). If Γ transfixes Hyp(X), there is an invariant subset A of Hyp(X)
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
26
for which A4Hyp(X) is finite. Thus, one gets the following characterization of
pseudo-regularizability (the converse being immediate).
Theorem 6.4. Let X be a normal projective variety over an algebraically closed
field k. Let Γ be a subgroup of Bir(X). Then Γ transfixes the subset Hyp(X) of
˜
Hyp(X)
if and only if Γ is pseudo-regularizable.
Of course, this theorem applies directly when Γ ⊂ Bir(X) has property (FW)
because Theorem 5.6 shows that Γ commensurates Hyp(X).
Remark 6.5. Assuming char(k) = 0, we may work in the category of smooth
varieties (see Example 4.3 and § 5.5). As explained in Remark 1.2 and Lemma 6.3,
there are two extreme cases, corresponding to an empty or an ample boundary B =
∪i Ei .
If U = Y , Γ acts by pseudo-automorphisms on the projective variety Y . As
explained in Theorem 2.7, Γ is an extension of a subgroup of GL (NS(Y )) by an
algebraic group (which is almost contained in Aut(Y )0 ).
If U is affine, Γ acts by automorphisms on U. The group Aut(U) may be huge
(for instance if U is the affine space), but there are techniques to study groups of
automorphisms that are not available for birational transformations. For instance
Γ is residually finite and virtually torsion free if Γ is a group of automorphisms
generated by finitely many elements (see [3]).
6.4. Distorted elements. Theorem 6.4 may be applied when Γ has Property (FW),
or for pairs (Λ, Γ) with relative Property (FW). Here is one application:
Corollary 6.6. Let X be an irreducible projective variety. Let Γ be a distorted
cyclic subgroup of Bir(X). Then Γ is pseudo-regularizable.
The contraposition is useful to show that some elements of Bir(X) are undistorted. Let us state it in a strong “stable” way.
Corollary 6.7. Let X be a normal irreducible projective variety and let f be an
element of Bir(X) such that the cyclic group hf i does not transfix Hyp(X) (i.e., f is
not pseudo-regularizable). Then the cyclic subgroup hf i is undistorted in Bir(X),
and more generally for every irreducible projective variety, the cyclic subgroup
hf × IdY i is undistorted in Bir(X × Y ).
The latter consequence indeed follows from Proposition 5.7. This can be applied
to various examples, such as those in Example 7.9.
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
27
7. I LLUSTRATING RESULTS
7.1. Surfaces whose birational group is transfixing. If X is a projective curve,
Bir(X) always transfixes Hyp(X), since it acts by automorphisms on a smooth
model of X. We now consider the same problem for surfaces, starting with the
following result, which holds in arbitrary dimension.
Proposition 7.1. Let X be a normal irreducible variety of positive dimension over
an algebraically closed field k. Then Bir(X × P1 ) does not transfix Hyp(X × P1 ).
Proof. We can suppose that X is affine and work in the model X × A1 . For
ϕ a nonzero regular function on X, define a regular self-map f of X × A1 by
f (x, t) = (x, ϕ(x)t). Denoting by Z(ϕ) the zero set of ϕ, we remark that f induces
an automorphism of the open subset (X r Z(ϕ)) × A1 . In particular, it induces a
permutation of Hyp((X r Z(ϕ)) × A1 ). Moreover, since f contracts the complement Z(ϕ) × A1 to the subset Z(ϕ) × {0}, which has codimension ≥ 2, its action
˜
on Hyp(X
× A1 ) maps the set of codimension 1 components of Z(ϕ) × A1 outside
M = Hyp(X × A1 ). Therefore M r f −1 (M ) is the set of irreducible components of
Z(ϕ) × A1 . Its cardinal is equal to the number of irreducible components of Z(ϕ).
When ϕ varies, this number is unbounded; hence, Bir(X × A1 ) does not transfix
Hyp(X × A1 ).
Varieties that are birational to the product of a variety and the projective line are
said to be ruled. Proposition 7.1 states that for any ruled irreducible projective
variety Y of dimension ≥ 2, Bir(Y ) does not transfix Hyp(Y ). The converse holds
for surfaces, by the following theorem.
Theorem 7.2. Let k be an algebraically closed field. Let X be an irreducible
normal projective surface over k. The following are equivalent:
(1)
(2)
(3)
(4)
Bir(X) does not transfix Hyp(X);
the Kodaira dimension of X is −∞;
X is ruled;
there is no irreducible projective surface Y that is birationally equivalent to
X, and such that Bir(Y ) = Aut(Y ).
Proof. The equivalence between (2) and (3) is classical (see [1]). The group Aut(Y )
˜
fixes Hyp(Y ) ⊂ Hyp(Y
), hence (1) implies (4). If the Kodaira dimension of X is
≥ 0, then X has a unique minimal model X0 , and Bir(X0 ) = Aut(X0 ). Thus, (4)
implies (2). Finally, Proposition 7.1 shows that (3) implies (1).
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
28
Theorem 7.3. Let X be an irreducible projective surface over an algebraically
closed field k. The following are equivalent:
(1) some finitely generated subgroup of Bir(X) does not transfix Hyp(X);
(2) some cyclic subgroup of Bir(X) does not transfix Hyp(X);
(3) • k has characteristic 0, and X is birationally equivalent to the product
of the projective line with a curve of genus 0 or 1, or
• k has positive characteristic, and X is a rational surface.
Example 7.4. Let k be an algebraically closed field that is not algebraic over a finite
field. Let t be an element of infinite order in the multiplicative group k∗ . Then the
birational transformation g of P2k given, in affine coordinates, by (x, y) 7→ (tx +
1, xy) does not transfix Hyp(P2k ). Indeed, it is easy to show that the hypersurface
C = {x = 0} satisfies, for n ∈ Z, f n (C) ∈ Hyp(P2k ) if and only if n ≤ 0.
Example 7.5. The example of non-transfixing element in Example 7.4 works under
a small restriction on k. Here is an example over an arbitrary algebraically closed
field k. Let L and L0 be two lines in P2k intersecting transversally at a point q.
Let f be a birational transformation of P2k that contracts L0 onto q and fixes the
line L. For instance, in affine coordinates, the monomial map (x, y) 7→ (x, xy)
contracts the y-axis onto the origin, and fixes the x-axis. Assume that there is an
open neighborhood U of q such that f does not contract any curve in U except the
line L0 . Let C be an irreducible curve that intersects L and L0 transversally at q.
Then, for every n ≥ 1, the strict transform f◦n (C) is an irreducible curve, and the
order of tangency of this curve with L goes to infinity with n. Thus, the degree of
f◦n (C) goes to infinity too and the f◦n (C) form an infinite sequence in Hyp(P2k ).
Now, assume that C is contracted by f −1 onto a point p, p ∈
/ Ind(f ), and p is
−1
−m
fixed by f . Then, for every m ≥ 1, f• (C) is not in Hyp(P2k ). This shows
that the orbit of C under the action of f• intersects Hyp(P2k ) and its complement
2
2
n
−m
˜
Hyp(P
k ) r Hyp(Pk ) on the infinite sets {f◦ (C) ; n ≥ 1} and {f• (C) ; m ≥ 1}.
In particular, f does not transfix Hyp(P2k ).
Since such maps exist over every algebraically closed field k, this example shows
that property (2) of Theorem 7.3 is satisfied for every rational surface X.
Proof. Trivially (2) implies (1).
Suppose that (3) holds and let us prove (2). The case X = P1 × P1 is already
covered by Lemma 6.1 in characteristic zero, and by the previous example in positive characteristic. The case X = C × P1 in characteristic zero, where C is an
elliptic curve, is similar. To see it, fix a point t0 ∈ C and a rational function ϕ on C
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
29
that vanishes at t0 . Then, since k has characteristic zero, one can find a translation
s of C of infinite order such that the orbit {sn (t0 ) : n ∈ Z} does not contain any
other zero or pole of ϕ (here we use that the characteristic of k is 0). Consider the
birational transformation f ∈ Bir(X) given by f (t, x) = (s(t), ϕ(t)x). Let H be
the hypersurface {t0 } × C. Then for n ∈ Z, we have (f• )n H ∈ Hyp(X) if and only
if n ≤ 0. Hence the action of the cyclic group hf i does not transfix Hyp(X).
Let us now prove that (1) implies (3). Applying Theorem 7.2, and changing
X to a birationally equivalent surface if necessary, we assume that X = C × P1
for some (smooth irreducible) curve C. We may now assume that the genus of C
is ≥ 2, or ≥ 1 in positive characteristic, and we have to show that every finitely
generated group Γ of Bir(X) transfixes Hyp(X). Since the genus of C is ≥ 1, the
group Bir(X) preserves the fibration X → C; this gives a surjective homomorphism
Bir(X) → Aut(C). Now let us fully use the assumption on C: if its genus is ≥ 2,
then Aut(C) is finite; if its genus is 1 and k has positive characteristic, then Aut(C)
is locally finite1, and in particular the projection of Γ on Aut(C) has a finite image.
Thus the kernel of this homomorphism intersects Γ in a finite index subgroup Γ0 .
It now suffices to show that Γ0 transfixes Hyp(X). Every f ∈ Γ0 has the form
f (t, x) = (t, ϕt (x)) for some rational map t 7→ ϕt from C to PGL 2 ; define Uf ⊂ C
as the open and dense subset on which ϕγ is regular: by definition, f restricts to an
automorphism of Uf × P1 . Let S be a finite generating subset of Γ0 , and let US be
the intersection of the open subsets Ug , for g ∈ S. Then Γ0 acts by automorphisms
on US × P1 and its action on Hyp(X) fixes the subset Hyp(US ). Hence Γ transfixes
Hyp(X).
It would be interesting to obtain characterizations of the same properties in dimension 3 (see Question 11.2).
7.2. Transfixing Jonquières twists. Let X be an irreducible normal projective surface and π a morphism onto a smooth projective curve C with rational connected
fibers. Let Birπ (X) be the subgroup of Bir(X) permuting the fibers of π. Since C
is a smooth projective curve, the group Bir(C) coincides with Aut(C) and we get a
canonical homomorphism rC : Birπ (X) → Aut(C).
The main examples to keep in mind are provided by P1 ×P1 , Hirzebruch surfaces,
and C × P1 for some genus 1 curve C, π being the first projection.
Let Hypπ (X) denote the set of irreducible curves which are contained in fibers
˜ π (X) = Hypπ (X) t (Hyp(X)
˜
˜
of π, and define Hyp
r Hyp(X)), so that Hyp(X)
=
1
Every finitely generated subgroup is finite.
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
30
˜ π (X) t (Hyp(X) r Hypπ (X)). An irreducible curve H ⊂ X is an element of
Hyp
Hyp(X) r Hypπ (X) if and only if its projection π(H) coincides with C; this curves
are said to be transverse to π.
˜
˜ π (X) t (Hyp(X) r Hypπ (X))
Proposition 7.6. The decomposition Hyp(X)
= Hyp
is Birπ (X)-invariant.
Proof. Let H ⊂ X be an irreducible curve which is transverse to π. Since Birπ (X)
acts by automorphisms on C, H can not be contracted by any element of Birπ (X);
more precisely, for every g ∈ Birπ (X), g• (H) is an element of Hyp(X) which is
transverse to π. Thus the set of transverse curves is Birπ (X)-invariant.
This proposition and the proof of Theorem 7.3 lead to the following corollary.
Corollary 7.7. Let G be a subgroup of Birπ (X). If π maps the set of indeterminacy
points of the elements of G into a finite subset of C, then G transfixes Hyp(X).
In the case of cyclic subgroups, we establish a converse under the mild assumption of algebraic stability. Recall that a birational transformation f of a smooth
projective surface is algebraically stable if the forward orbit of Ind(f −1 ) does not
intersect Ind(f ). By [14], given any birational transformation f of a surface X,
there is a birational morphism u : Y → X, with Y a smooth projective surface,
such that fY := u−1 ◦ f ◦ u is algebraically stable. If π : X → C is a fibration,
as above, and f is in Birπ (X), then fY preserves the fibration π ◦ u. Thus, we
may always assume that X is smooth and f is algebraically stable after a birational
conjugacy.
Proposition 7.8. Let X be a smooth projective surface, and π : X → C a rational
fibration. If f ∈ Birπ (X) is algebraically stable, then f transfixes Hyp(X) if, and
only if the orbit of π(Ind(f )) under the action of rC (f ) is finite.
For X = P1 × P1 , the reader can check (e.g., conjugating a suitable automorphism) that the proposition fails without the algebraic stability assumption.
Proof. Denote by A ⊂ Aut(C) the subgroup generated by rC (f ). Consider a fiber
F ' P1 which is contracted to a point q by f . Then, there is a unique indeterminacy
point p of f on F . If the orbit of π(q) under the action of A is infinite, the orbit of
q under the action of f is infinite too. Set qn = f n−1 (q) for n ≥ 1 (so that q1 = q);
this sequence of points is well defined because f is algebraically stable: for every
n ≥ 1, f is a local isomorphism from a neighborhood of qn to a neighborhood
˜
of qn+1 . Then, the image of F in Hyp(X)
under the action of f n is an element
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
31
˜
of Hyp(X)
r Hyp(X): it is obtained by a finite number of blow-ups above qn .
Since the points qn form an infinite set, the images of F form an infinite subset of
˜
Hyp(X)
r Hyp(X). Together with the previous corollary, this argument proves the
proposition.
Example 7.9. Consider X = P1 × P1 , with π(x, y) = x (using affine coordinates).
Start with fa (x, y) = (ax, xy), for some non-zero parameter a ∈ k. The action of
rC (fa ) on C = P1 fixes the images 0 and ∞ of the indeterminacy points of fa . Thus,
˜
fa transfixes Hyp(X)
by Corollary 7.7. Now, consider ga (x, y) = (ax, (x + 1)y).
Then, the orbit of −1 under multiplication by a is finite if and only if a is a root
of unity; thus, if a is not a root of unity, ga does not transfix Hyp(X). Section 6.1
provides more examples of that kind.
8. B IRATIONAL TRANSFORMATIONS OF SURFACES I
From now on, we work in dimension 2: X, Y , and Z will be smooth projective
surfaces over the algebraically closed field k. (In dimension 2, the resolution of
singularities is available in all characteristics, so that we can always assume the
varieties to be smooth.)
8.1. Regularization. In this section, we refine Theorem 6.4, in order to apply results of Danilov and Gizatullin. Recall that a curve C in a smooth surface Y has
normal crossings if each of its singularities is a simple node with two transverse
tangents. In the complex case, this means that C is locally analytically equivalent
to {xy = 0} (two branches intersecting transversally) in an analytic neighborhood
of each of its singularities.
Theorem 8.1. Let X be a smooth projective surface, defined over an algebraically
closed field k. Let Γ be a subgroup of Bir(X) that transfixes the subset Hyp(X) of
˜
Hyp(X).
There exists a smooth projective surface Z, a birational map ϕ : Z 99K X
and a dense open subset U ⊂ Z such that, writing the boundary ∂Z := Z r U as a
finite union of irreducible components Ei ⊂ Z, 1 ≤ i ≤ `, the following properties
hold:
(1) The boundary ∂Z is a curve with normal crossings.
(2) The subgroup ΓZ := ϕ−1 ◦ Γ ◦ ϕ ⊂ Bir(Z) acts by automorphisms on the
open subset U.
(3) For all i ∈ {1, . . . , `} and g ∈ ΓZ , the strict transform of Ei under the
action of g on Z is contained in ∂Z: either g◦ (Ei ) is a point of ∂Z or
g◦ (Ei ) is an irreducible component Ej of ∂Z.
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
32
(4) For all i ∈ {1, . . . , `}, there exists an element g ∈ ΓZ that contracts Ei to a
point g◦ (Ei ) ∈ ∂Z. In particular, Ei is a rational curve.
(5) The pair (Z, U) is minimal for the previous properties, in the following
sense: if one contracts a smooth curve of self-intersection −1 in ∂Z, then
the boundary stops to be a normal crossing divisor.
Before starting the proof, note that the boundary ∂Z may a priori contain an
irreducible rational curve E with a node.
Proof. We apply Theorem 6.4 (which works in positive characteristic too, because
X is a surface), and get a birational morphism ϕ0 : Y0 → X and an open subset U0
of Y0 that satisfy properties (1) and (3), except that we only know that the action of
Γ0 := ϕ−1
0 ◦ Γ ◦ ϕ0 on U0 is by pseudo-automorphisms (not yet by automorphisms).
We shall progressively modify the triple (Y0 , U0 , ϕ0 ) to obtain a surface Z with
properties (1) to (5).
Step 1.– First, we blow-up the singularities of the curve ∂Y0 = Y0 r U0 to get a
boundary that is a normal crossing divisor. This replaces the surface Y0 by a new
one, still denoted Y0 . This modification adds new components to the boundary ∂Y0
but does not change the fact that Γ0 acts by pseudo-automorphisms on U0 . Let `0 be
the number of irreducible components of Y0 r U0 .
Step 2.– Consider a point q in U0 , and assume that there is a curve Ei of ∂Y0
that is contracted to q by an element g ∈ Γ0 ; fix such a g, and denote by D the
union of the curves Ej such that g◦ (Ej ) = q. By construction, g is a pseudoautomorphism of U0 . The curve D does not intersect the indeterminacy set of g,
since otherwise there would be a curve C containing q that is contracted by g −1 .
And D is a connected component of ∂Y0 , because otherwise g maps one of the Ej
to a curve that intersects U0 . Thus, there are small neighborhoods W of D and
W 0 of q such that W ∩ ∂Y0 = D and g realizes an isomorphism from W r D to
W 0 r {q}, contracting D onto the smooth point q ∈ Y0 . As a consequence, there is
a birational morphism π1 : Y0 → Y1 such that
(1) Y1 is smooth
(2) π1 contracts D onto a point q1 ∈ Y1
(3) π1 is an isomorphism from Y0 r D to Y1 r {q1 }.
In particular, π1 (U0 ) is an open subset of Y1 and U1 = π1 (U0 ) ∪ {q1 } is an open
neighborhood of q1 in Y1 .
Then, Γ1 := π1 ◦ Γ0 ◦ π1−1 acts birationally on Y1 , and by pseudo-automorphisms
on U1 . The boundary ∂Y1 = Y1 r U1 contains `1 irreducible components, with
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
33
`1 < `0 (the difference is the number of components of D), and is a normal crossing
divisor because D is a connected component of ∂Y0 .
Repeating this process, we construct a sequence of surfaces πk : Yk−1 → Yk and
open subsets πk (Uk−1 ) ⊂ Uk ⊂ Yk such that the number of irreducible components
of ∂Yk = Yk r Uk decreases. After a finite number of steps (at most `0 ), we may
assume that Γk ⊂ Bir(Yk ) does not contract any boundary curve onto a point of the
open subset Uk . On such a model, Γk acts by automorphisms on Uk .
We fix such a model, which we denote by the letters Y , U, ∂Y , ϕ. The new birational map ϕ : Y 99K X is the composition of ϕ0 with the inverse of the morphism
Y0 → Yk . On such a model, properties (1) and (2) are satisfied. Moreover, (3)
follows from (2). We now modify Y further to get property (4).
Step 3.– Assume that the curve Ei ⊂ Y r U is not contracted by Γ. Let F be
the orbit of Ei : F = ∪g∈Γ g◦ (Ei ); this curve is contained in the boundary ∂Y of the
open subset U. Changing U into
U 0 = U ∪ (F r ∂Y r F ),
the group Γ also acts by pseudo-automorphisms on U 0 . This operation decreases
the number ` of irreducible components of the boundary. Thus, combining steps 2
and 3 finitely many times, we reach a model that satisfies Properties (1) to (4). We
continue to denote it by Y .
Step 4.– If the boundary ∂Y contains a smooth (rational) curve Ei of self-intersection −1, it can be blown down to a smooth point q by a birational morphism
π : Y → Y 0 ; the open subset U is not affected, but the boundary ∂Y 0 has one
component less. If Ei was a connected component of ∂Y , then U 0 = π(U) ∪ {q}
is a neighborhood of q and one replaces U by U 0 , as in step 2. Now, two cases may
happen. If the boundary ∂Y 0 ceases to be a normal crossing divisor, we come back
to Y and do not apply this surgery. If ∂Y 0 has normal crossings, we replace Y by
this new model. In a finite number of steps, looking successively at all (−1)-curves
and iterating the process, we reach a new surface Z on which all five properties are
satisfied.
Remark 8.2. One may also remove property (5) and replace property (1) by
(1’) The Ei are rational curves, and none of them is a smooth rational curve with
self-intersection −1.
But doing so, we may lose the normal crossing property. To get property (1’), apply
the theorem and argue as in step 4.
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
34
8.2. Constraints on the boundary. We now work on the new surface Z given by
Theorem 8.1. Thus, Z is the surface, Γ the subgroup of Bir(Z), U the open subset
on which Γ acts by automorphisms, and ∂Z the boundary of U.
Proposition 8.3 (Gizatullin, [17] § 4). There are four possibilities for the geometry
of the boundary ∂Z = Z r U.
(1)
(2)
(3)
(4)
∂Z is empty.
∂Z is a cycle of rational curves.
∂Z is a chain of rational curves.
∂Z is not connected; it is the disjoint union of finitely many smooth rational
curves of self-intersection 0.
Moreover, in cases (2) and (3), the open subset U is the blow-up of an affine surface.
Thus, there are four possibilities for ∂Z, which we study successively. We shall
start with (1) and (4) in sections 8.3 and 8.4. Then case (3) is dealt with in Section 8.5. Case (2) is slightly more involved: it is treated in Section 9.
Before that, let us explain how Proposition 8.3 follows from Section 5 of [17].
First, we describe the precise meaning of the statement, and then we explain how
the original results of [17] apply to our situation.
The boundary and its dual graph .– Consider the dual graph GZ of the boundary ∂Z. The vertices of GZ are in one to one correspondence with the irreducible
components Ei of ∂Z. The edges correspond to singularities of ∂Z: each singular point q gives rise to an edge connecting the components Ei that determine the
two local branches of ∂Z at q. When the two branches correspond to the same
irreducible component, one gets a loop of the graph GZ .
We say that ∂Z is a chain of rational curves if the dual graph is of type A` : ` is
the number of components, and the graph is linear, with ` vertices. Chains are also
called zigzags by Gizatullin and Danilov.
We say that ∂Z is a cycle if the dual graph is isomorphic to a regular polygon
with ` vertices. There are two special cases: when ∂Z is reduced to one component,
this curve is a rational curve with one singular point and the dual graph is a loop
(one vertex, one edge); when ∂Z is made of two components, these components
intersect in two distinct points, and the dual graph is made of two vertices with two
edges between them. For ` = 3, 4, . . ., the graph is a triangle, a square, etc.
Gizatullin’s original statement.– To describe Gizatullin’s article, let us introduce some useful vocabulary. Let S be a projective surface, and C ⊂ S be a curve;
C is a union of irreducible components, which may have singularities. Assume that
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
35
S is smooth in a neighborhood of C. Let S0 be the complement of C in S, and
let ι : S0 → S be the natural embedding of S0 in S. Then, S is a completion of
S0 : this completion is marked by the embedding ι : S0 → S, and its boundary is
the curve C. Following [17] and [18, 19], we only consider completions of S0 by
curves (i.e. S rι(S0 ) is of pure dimension 1), and we always assume S to be smooth
in a neighborhood of the boundary. Such a completion is
(i) simple if the boundary C has normal crossings;
(ii) minimal if it is simple and minimal for this property: if Ci ⊂ C is an
exceptional divisor of the first kind then, contracting Ci , the image of C is
not a normal crossing divisor anymore. Equivalently, Ci intersects at least
three other components of C. Equivalently, if ι0 : S0 → S 0 is another simple
completion, and π : S → S 0 is a birational morphism such that π ◦ ι = ι0 ,
then π is an isomorphism.
If S is a completion of S0 , one can blow-up boundary points to obtain a simple
completion, and then blow-down some of the boundary components Ci to reach a
minimal completion.
Now, consider the group of automorphisms of the open surface S0 . This group
Aut(S0 ) acts by birational transformations on S. An irreducible component Ei of
the boundary C is contracted if there is an element g of Aut(S0 ) that contracts Ei :
g◦ (Ei ) is a point of C. Let E be the union of the contracted components. In [17],
Gizatullin proves that E satisfies one of the four properties stated in Proposition 8.3;
moreover, in cases (2) and (3), E contains an irreducible component Ei with Ei2 > 0
(see Corollary 4, Section 5 of [17]).
Thus, Proposition 8.3 follows from the properties of the pair (Z, U, Γ): the open
subset U plays the role of S0 , and Z is the completion S; the boundary ∂Z is the
curve C: it is a normal crossing divisor, and it is minimal by construction. Since
every component of ∂Z is contracted by at least one element of Γ ⊂ Aut(U), ∂Z
coincides with Gizatullin’s curve E. The only thing we have to prove is the last
sentence of Proposition 8.3, concerning the structure of the open subset U.
First, let us show that E = ∂Z supports an effective divisor D such that D2 > 0
and D·F ≥ 0 for every irreducible curve. To do so, fix an irreducible component E0
of ∂Z with positive self-intersection. Assume that ∂Z is a cycle, and list cyclically
the other irreducible components: E1 , E2 , ..., up to Em , with E1 and Em intersecting
E0 . First, one defines a1 = 1. Then, one chooses a2 > 0 such that a1 E1 + a2 E2
intersects positively E1 , then a3 > 0 such that a1 E1 + a2 E2 + a3 E3 intersects
P
positively E1 and E2 , ..., up to m
i=1 ai Ei that intersects all components Ei , 1 ≤
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
36
i ≤ m − 1 positively. Since E02 > 0 and E0 intersects Em , one can find a coefficient
a0 for which the divisor
m
X
D=
ai Ei
i=0
2
satisfies D > 0 and D · Ei > 0 for all Ei , 0 ≤ i ≤ m. This implies that D
intersects every irreducible curve F non-negatively. Thus, D is big and nef (see
[26], Section 2.2). A similar proof applies when ∂Z is a zigzag.
Let W be the subspace of NS(X) spanned by classes of curves F with D · F = 0.
Since D2 > 0, Hodge index theorem implies that the intersection form is negative definite on W . Thus, Mumford-Grauert contraction theorem provides a birational morphism τ : Z → Z 0 that contracts simultaneously all curves F with class
[F ] ∈ W and is an isomorphism on Z rF ; in particular, τ is an isomorphism from a
neighborhood V of ∂Z onto its image τ (V) ⊂ Z 0 . The modification τ may contract
curves that are contained in U, and may create singularities for the new open subset
U 0 = τ (U), but does not modify Z near the boundary ∂Z. Now, on Z 0 , the divisor D0 = τ∗ (D) intersects every effecitve curve positively and satisfies (D0 )2 > 0.
Nakai-Moishezon criterion shows that D0 is ample (see [26], Section 1.2.B); consequently, there is an embedding of Z 0 into a projective space and a hyperplane section
H of Z 0 for which Z 0 r H coincides with U 0 . This proves that U is a blow-up of the
affine (singular) surface U 0 .
8.3. Projective surfaces and automorphisms. In this section, we (almost always)
assume that Γ acts by regular automorphisms on a projective surface X. This
corresponds to case (1) in Proposition 8.3. Our goal is the special case of Theorem B which is stated below as Theorem 8.8. We shall assume that Γ has property
(FW) in some of the statements (this was not a hypothesis in Theorem thm:FWregularization-surfaces).
We may, and shall, assume that X is smooth. We refer to [1, 4, 20] for the
classification of surfaces and the main notions attached to them.
8.3.1. Action on the Néron-Severi group. The intersection form is a non-degenerate
quadratic form qX on the Néron-Severi group NS(X), and Hodge index theorem
asserts that its signature is (1, ρ(X) − 1), where ρ(X) denotes the Picard number,
i.e. the rank of the lattice NS(X) ' Zρ .
The action of Aut(X) on the Néron-Severi group NS(X) provides a linear representation preserving the intersection form qX . This gives a morphism
Aut(X) → O (NS(X); qX ).
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
37
Fix an ample class a in NS(X) and consider the hyperboloid
HX = {u ∈ NS(X) ⊗Z R; qX (u, u) = 1 and qX (u, a) > 0}.
This set is one of the two connected components of {u; qX (u, u) = 1}. With the riemannian metric induced by (−qX ), it is a copy of the hyperbolic space of dimension
ρ(X) − 1; the group Aut(X) acts by isometries on this space (see [9]).
Proposition 8.4. Let X be a smooth projective surface. Let Γ be a subgroup of
Aut(X). If Γ has Property (FW), its action on NS(X) fixes a very ample class, the
image of Γ in O (NS(X); qX ) is finite, and a finite index subgroup of Γ is contained
in Aut(X)0 .
Proof. The image Γ∗ of Γ is contained in the arithmetic group O (NS(X); qX ).
The Néron-Severi group NS(X) is a lattice Zρ and qX is defined over Z. Thus,
O (NS(X);
qX ) is a standard arithmetic group in the sense of [6], § 1.1. The main results
of [6] imply that the action of Γ∗ on the hyperbolic space HX has a fixed point. Let
u be such a fixed point. Since qX is negative definite on the orthogonal complement
u⊥ of u in NS(X), and Γ∗ is a discrete group acting by isometries on it, we deduce
P
that Γ∗ is finite. If a is a very ample class, the sum γ∈Γ∗ γ ∗ (a) is an invariant, very
ample class.
The kernel K ⊂ Aut(X) of the action on NS(X) contains Aut(X)0 as a finite
index subgroup. Thus, if Γ has Property (FW), it contains a finite index subgroup
that is contained in Aut(X)0 .
8.3.2. Non-rational surfaces. In this paragraph, we assume that the surface X is
not rational. The following proposition classifies subgroups of Bir(X) with Property (FW); in particular, such a group is finite if the Kodaira dimension of X is
non-negative (resp. if the characteristic of k is positive). Recall that we denote by
Z ⊂ Q the ring of algebraic integers.
Proposition 8.5. Let X be a smooth, projective, and non-rational surface, over the
algebraically closed field k. Let Γ be an infinite subgroup of Bir(X) with Property (FW). Then k has characteristic 0, and there is a birational map ϕ : X 99K
C × P1k that conjugates Γ to a subgroup of Aut(C × P1k ). Moreover, there is a finite
index subgroup Γ0 of Γ such that ϕ ◦ Γ0 ◦ ϕ−1 , is a subgroup of PGL 2 (Z), acting
on C × P1k by linear projective transformations on the second factor.
Proof. Assume, first, that the Kodaira dimension of X is non-negative. Let π : X →
X0 be the projection of X on its (unique) minimal model (see [20], Thm. V.5.8).
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
38
The group Bir(X0 ) coincides with Aut(X0 ); thus, after conjugacy by π, Γ becomes
a subgroup of Aut(X0 ), and Proposition 8.4 provides a finite index subgroup Γ0 ≤ Γ
that is contained in Aut(X0 )0 . Note that Γ0 inherits Property (FW) from Γ.
If the Kodaira dimension of X is equal to 2, the group Aut(X0 )0 is trivial; hence
Γ0 = {IdX0 } and Γ is finite. If the Kodaira dimension is equal to 1, Aut(X0 )0
is either trivial, or isomorphic to an elliptic curve, acting by translations on the
fibers of the Kodaira-Iitaka fibration of X0 (this occurs, for instance, when X0 is the
product of an elliptic curve with a curve of higher genus). If the Kodaira dimension
is 0, then Aut(X0 )0 is also an abelian group (either trivial, or isomorphic to an
abelian surface). Since abelian groups with Property (FW) are finite, the group Γ0
is finite, and so is Γ.
We may now assume that the Kodaira dimension kod(X) is negative. Since X
is not rational, then X is birationally equivalent to a product S = C × P1k , where
C is a curve of genus g(C) ≥ 1. Denote by k(C) the field of rational functions on
the curveC. We fix a local coordinate x on C and denote the elements of k(C) as
functions a(x) of x. The semi-direct product Aut(C) n PGL 2 (k(C)) acts on S by
birational transformations of the form
a(x)y + b(x)
1
,
(x, y) ∈ C × Pk 7→ f (x),
c(x)y + d(x)
and Bir(S) coincides with this group Aut(C) n PGL 2 (k(C)); indeed, the first projection π : S → C is equivariant under the action of Bir(S) because every rational
map P1k → C is constant.
Since g(C) ≥ 1, Aut(C) is virtually abelian. Property (FW) implies that there
is a finite index, normal subgroup Γ0 ≤ Γ that is contained in PGL 2 (k(C)). By
Corollary 3.8, every subgroup of PGL 2 (k(C)) with Property (FW) is conjugate to a
subgroup of PGL 2 (Z) or a finite group if the characteristic of the field k is positive.
We may assume now that the characteristic of k is 0 and that Γ0 ⊂ PGL 2 (Z) is
infinite. Consider an element g of Γ; it acts as a birational transformation on the
surface S = C × P1k , and it normalizes Γ0 :
g ◦ Γ0 = Γ0 ◦ g.
Since Γ0 acts by automorphisms on S, the finite set Ind(g) is Γ0 -invariant. But a
subgroup of PGL 2 (k) with Property (FW) preserving a non-empty, finite subset of
P1 (k) is a finite group. Thus, Ind(g) must be empty. This shows that Γ is contained
in Aut(S).
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
39
8.3.3. Rational surfaces. We now assume that X is a smooth rational surface, that
Γ ≤ Bir(X) is an infinite subgroup with Property (FW), and that Γ contains a
finite index, normal subgroup Γ0 that is contained in Aut(X)0 . Recall that a smooth
surface Y is minimal if it does not contain any smooth rational curve of the first
kind, i.e. with self-intersection −1. Every exceptional curve of the first kind E ⊂
X is determined by its class in NS(X) and is therefore invariant under the action
of Aut(X)0 . Contracting such (−1)-curves one by one, we obtain the following
lemma.
Lemma 8.6. There is a birational morphism π : X → Y onto a minimal rational surface Y that is equivariant under the action of Γ0 ; Y does not contain any
exceptional curve of the first kind and Γ0 becomes a subgroup of Aut(Y )0 .
Let us recall the classification of minimal rational surfaces and describe their
groups of automorphisms. First, we have the projective plane P2k , with Aut(P2k ) =
PGL 3 (k) acting by linear projective transformations. Then comes the quadric P1k ×
P1k , with Aut(P1k × P1k )0 = PGL 2 (k) × PGL 2 (k) acting by linear projective transformations on each factor; the group of automorphisms of the quadric is the semidirect product of PGL 2 (k) × PGL 2 (k) with the group of order 2 generated by the
permutation of the two factors, η(x, y) = (y, x). Then, for each integer m ≥ 1, the
Hirzebruch surface Fm is the projectivization of the rank 2 bundle O ⊕ O(m) over
P1k ; it may be characterized as the unique ruled surface Z → P1k with a section C
of self-intersection −m. Its group of automorphisms is connected and preserves the
ruling. This provides a homomorphism Aut(Fm ) → PGL 2 (k) that describes the action on the base of the ruling, and it turns out that this homomorphism is surjective.
If we choose coordinates for which the section C intersects each fiber at infinity, the
kernel Jm of this homomorphism acts by transformations of type
(x, y) 7→ (x, αy + β(x))
where β(x) is a polynomial function of degree ≤ m. In particular, Jm is solvable.
In other words, Aut(Fm ) is isomorphic to the group
(GL 2 (k)/µm ) n Wm
where Wm is the linear representation of GL 2 (k) on homogeneous polynomials
of degree m in two variables, and µm is the kernel of this representation: it is
the subgroup of GL 2 (k) given by scalar multiplications by roots of unity of order
dividing m.
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
40
Lemma 8.7. Given the above conjugacy π : X → Y , the subgroup π ◦ Γ ◦ π −1 of
Bir(Y ) is contained in Aut(Y ).
Proof. Assume that the surface Y is the quadric P1k × P1k . Then, according to Theorem 3.6, Γ0 is conjugate to a subgroup of PGL 2 (Z) × PGL 2 (Z). If g is an element
of Γ, its indeterminacy locus is a finite subset Ind(g) of P1k × P1k that is invariant under the action of Γ0 , because g normalizes Γ0 . Since Γ0 is infinite and has Property
(FW), this set Ind(g) is empty (Lemma 3.5). Thus, Γ is contained in Aut(P1k × P1k ).
The same argument applies for Hirzebruch surfaces. Indeed, Γ0 is an infinite
subgroup of Aut(Fm ) with Property (FW). Thus, up to conjugacy, its projection
in PGL 2 (k) is contained in PGL 2 (Z). If it were finite, a finite index subgroup
of Γ0 would be contained in the solvable group Jm , and would therefore be finite
too by Property (FW); this would contradict |Γ0 | = ∞. Thus, the projection of
Γ0 in PGL (Z) is infinite. If g is an element of Γ, Ind(g) is a finite, Γ0 -invariant
subset, and by looking at the projection of this set in P1k one deduces that it is empty
(Lemma 3.5). This proves that Γ is contained in Aut(Fm ).
Let us now assume that Y is the projective plane. Fix an element g of Γ, and
assume that g is not an automorphism of Y = P2 ; the indeterminacy and exceptional
sets of g are Γ0 invariant. Consider an irreducible curve C in the exceptional set
of g, together with an indeterminacy point q of g on C. Changing Γ0 in a finite
index subgroup, we may assume that Γ0 fixes C and q; in particular, Γ0 fixes q, and
permutes the tangent lines of C through q. But the algebraic subgroup of PGL 3 (k)
preserving a point q and a line through q does not contain any infinite group with
Property (FW) (Lemma 3.5). Thus, again, Γ is contained in Aut(P2k ).
8.3.4. Conclusion, in Case (1). Putting everything together, we obtain the following particular case of Theorem B.
Theorem 8.8. Let X be a smooth projective surface over an algebraically closed
field k. Let Γ be an infinite subgroup of Bir(X) with Property (FW). If a finite index
subgroup of Γ is contained in Aut(X), there is a birational morphism ϕ : X → Y
that conjugates Γ to a subgroup ΓY of Aut(Y ), with Y in the following list:
(1) Y is the product of a curve C by P1k , the field k has characteristic 0, and a
finite index subgroup Γ0Y of ΓY is contained in PGL 2 (Z), acting by linear
projective transformations on the second factor;
(2) Y is P1k × P1k , the field k has characteristic 0, and ΓY is contained in
PGL 2 (Z) × PGL 2 (Z);
(3) Y is a Hirzebruch surface Fm and k has characteristic 0;
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
41
(4) Y is the projective plane P2k .
In particular, Y = P2k if the characteristic of k is positive.
Remark 8.9. Denote by ϕ : X → Y the birational morphism given by the theorem.
Changing Γ in a finite index subgroup, we may assume that it acts by automorphisms on both X and Y .
If Y = C × P1 , then ϕ is in fact an isomorphism. To prove this fact, denote by
ψ the inverse of ϕ. The indeterminacy set Ind(ψ) is ΓY invariant because both Γ
and ΓY act by automorphisms. From Lemma 3.5, applied to Γ0Y ⊂ PGL 2 (k), we
deduce that Ind(ψ) is empty and ψ is an isomorphism. The same argument implies
that the conjugacy is an isomorphism if Y = P1k × P1k or a Hirzebruch surface Fm ,
m ≥ 1.
Now, if Y is P2k , ϕ is not always an isomorphism. For instance, SL 2 (C) acts on
P2k with a fixed point, and one may blow up this point to get a new surface with an
action of groups with Property (FW). But this is the only possible example, i.e. X
is either P2k , or a single blow-up of P2k (because Γ ⊂ PGL 3 (C) can not preserve
more than one base point for ϕ−1 without loosing Property (FW)).
8.4. Invariant fibrations. We now assume that Γ has Property (FW) and acts by
automorphisms on U ⊂ X, and that the boundary ∂X = X rU is the union of ` ≥ 2
pairwise disjoint rational curves Ei ; each of them has self-intersection Ei2 = 0 and
is contracted by at least one element of Γ. This corresponds to the fourth possibility
in Gizatullin’s Proposition 8.3. Since Ei · Ej = 0, the Hodge index theorem implies
that the classes ei = [Ei ] span a unique line in NS(X), and that [Ei ] intersects
non-negatively every curve.
From Section 8.3.2, we may, and do assume that X is a rational surface. In
particular, the Euler characteristic of the structural sheaf is equal to 1: χ(OX ) = 1,
and Riemann-Roch formula gives
E12 − KX · E1
+ 1.
2
The genus formula implies KX ·E1 = −2, and Serre duality shows that h2 (X, E1 ) =
h0 (X, KX −E1 ) = 0 because otherwise −2 = (KX −E1 )·E1 would be non-negative
(because E1 intersects non-negatively every curve). From this, we obtain
h0 (X, E1 ) − h1 (X, E1 ) + h2 (X, E1 ) =
h0 (X, E1 ) = h1 (X, E1 ) + 2 ≥ 2.
Since E12 = 0, we conclude that the space H 0 (X, E1 ) has dimension 2 and determines a fibration π : X → P1k ; the curve E1 , as well as the Ei for i ≥ 2, are fibers
of π.
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
42
If f is an automorphism of U and F ⊂ U is a fiber of π, then f (F ) is a (complete)
rational curve. Its projection π(f (F )) is contained in the affine curve P1k r ∪i π(Ei )
and must therefore be reduced to a point. Thus, f (F ) is a fiber of π and f preserves
the fibration. This proves the following lemma.
Lemma 8.10. There is a fibration π : X → P1k such that
(1) every component Ei of ∂X is a fiber of π, and U = π −1 (V) for an open
subset V ⊂ P1k ;
(2) the generic component of π is a smooth rational curve;
(3) Γ permutes the fibers of π: there is a morphism ρ : Γ → PGL 2 (k) such that
π ◦ f = ρ(f ) ◦ π for every f ∈ Γ.
The open subset V ( P1k is invariant under the action of ρ(Γ); hence ρ(Γ) is
finite by Property (FW) and Lemma 3.5. Let Γ0 be the kernel of this morphism. Let
ϕ : X 99K P1k × P1k be a birational map that conjugates the fibration π to the first
projection τ : P1k × P1k → P1k . Then, Γ0 is conjugate to a subgroup of PGL 2 (k(x))
acting on P1k × P1k by linear projective transformations of the fibers of τ . From
Corollary 3.8, a new conjugacy by an element of PGL 2 (k(x)) changes Γ0 in an
infinite subgroup of PGL 2 (Z). Then, as in Sections 8.3.2 and 8.3.3 we conclude
that Γ becomes a subgroup of PGL 2 (Z) × PGL 2 (Z), with a finite projection on the
first factor.
Proposition 8.11. Let Γ be an infinite group with Property (FW), with Γ ⊂ Aut(U),
and U ⊂ Z as in case (4) of Proposition 8.3. There exists a birational map
ψ : Z 99K P1k × P1k that conjugates Γ to a subgroup of PGL 2 (Z) × PGL 2 (Z), with
a finite projection on the first factor.
8.5. Completions by zigzags. Two cases remain to be studied: ∂Z can be a chain
of rational curves (a zigzag in Gizatullin’s terminology) or a cycle of rational curves
(a loop in Gizatullin’s terminology). Cycles are considered in Section 9. In this
section, we rely on difficult results of Danilov and Gizatullin to treat the case of
chains of rational curves (i.e. case (3) in Proposition 8.3). Thus, in this section
(i) ∂X is a chain of smooth rational curves Ei
(ii) U = X r ∂X is an affine surface (singularities are allowed)
(iii) every irreducible component Ei is contracted to a point of ∂X by at least
one element of Γ ⊂ Aut(U) ⊂ Bir(X).
In [18, 19], Danilov and Gizatullin introduce a set of “standard completions” of
the affine surface U. As in Section 8.2, a completion (or more precisely a “marked
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
43
completion”) is an embedding ι : U → Y into a complete surface such that ∂Y =
Y r ι(U) is a curve (this boundary curve may be reducible). Danilov and Gizatullin
only consider completions for which ∂Y is a chain of smooth rational curves and
Y is smooth in a neighborhood of ∂Y ; the surface X provides such a completion.
Two completions ι : U → Y and ι0 : U → Y 0 are isomorphic if the birational map
ι0 ◦ι−1 : Y → Y 0 is an isomorphism; in particular, the boundary curves are identified
by this isomorphism. The group Aut(U) acts by pre-composition on the set of
isomorphism classes of (marked) completions.
Among all possible completions, Danilov and Gizatullin distinguish a class of
“standard (marked) completions”, for which we refer to [18] for a definition. There
are elementary links (corresponding to certain birational mappings Y 99K Y 0 ) between standard completions, and one can construct a graph ∆U whose vertices are
standard completions; there is an edge between two completions if one can pass
from one to the other by an elementary link.
Example 8.12. A completion is m-standard, for some m ∈ Z, if the boundary curve
∂Y is a chain of n + 1 consecutive rational curves E0 , E1 , . . ., En (n ≥ 1) such that
E02 = 0, E12 = −m, and Ei2 = −2 if i ≥ 2.
Blowing-up the intersection point q = E0 ∩ E1 , one creates a new chain starting
by E00 with (E00 )2 = −1; blowing down E00 , one creates a new (m + 1)-standard
completion. This is one of the elementary links.
Standard completions are defined by constraints on the self-intersections of the
components Ei . Thus, the action of Aut(U) on completions permutes the standard completions; this action determines a morphism from Aut(U) to the group of
isometries (or automorphisms) of the graph ∆U (see [18]):
Aut(U) → Iso (∆U ).
Theorem 8.13 (Danilov and Gizatullin, [18, 19]). The graph ∆U of all isomorphism
classes of standard completions of U is a tree. The group Aut(U) acts by isometries
of this tree. The stabilizer of a vertex ι : U → Y is the subgroup G(ι) of automorphisms of the complete surface Y that fix the curve ∂Y . This group is an algebraic
subgroup of Aut(Y ).
The last property means that G(ι) is an algebraic group that acts algebraically on
Y . It coincides with the subgroup of Aut(Y ) fixing the boundary ∂Y ; the fact that
it is algebraic follows from the existence of a G(ι)-invariant, big and nef divisor
which is supported on ∂Y (see the last sentence of Proposition 8.3).
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
44
The crucial assertion in this theorem is that ∆U is a simplicial tree (typically,
infinitely many edges emanate from each vertex). There are sufficiently many links
to assure connectedness, but not too many in order to prevent the existence of cycles
in the graph ∆U .
Corollary 8.14. If Γ is a subgroup of Aut(U) that has the fixed point property on
trees, then Γ is contained in G(ι) ⊂ Aut(Y ) for some completion ι : U → Y .
If Γ has Property (FW), it has Property (FA) (see Section 3.4). Thus, if it acts
by automorphisms on U, Γ is conjugate to the subgroup G(ι) of Aut(Y ), for some
zigzag-completion ι : U → Y . Theorem 8.8 of Section 8.3.3 implies that the action
of Γ on the initial surface X is conjugate to a regular action on P2k , P1k ×P1k or Fm , for
some Hirzebruch surface Fm . This action preserves a curve, namely the image of the
zigzag into the surface Y . The following examples list all possibilities, and conclude
the proof of Theorem B in the case of zigzags (i.e. case (3) in Proposition 8.3).
Example 8.15. Consider the projective plane P2k , together with an infinite subgroup
Γ ⊂ Aut(P2k ) that preserves a curve C and has Property (FW). Then, C must be a
smooth rational curve: either a line, or a smooth conic. If C is the line “at infinity”,
then Γ acts by affine transformations on the affine plane P2k r C. If the curve is the
conic x2 + y 2 + z 2 = 0, Γ becomes a subgroup of PO3 (k).
Example 8.16. When Γ is a subgroup of Aut(P1k × P1k ) that preserves a curve C
and has Property (FW), then C must be a smooth curve because Γ has no finite
orbit (Lemma 3.5). Similarly, the two projections C → P1k being equivariant with
respect to the morphisms Γ → PGL 2 (k), they have no ramification points. Thus,
C is a smooth rational curve, and its projections onto each factor are isomorphisms.
In particular, the action of Γ on C and on each factor are conjugate. From these
conjugacies, one deduces that the action of Γ on P1k × P1k is conjugate to a diagonal
embedding
γ ∈ Γ 7→ (ρ(γ), ρ(γ)) ∈ PGL 2 (k) × PGL 2 (k)
preserving the diagonal.
Example 8.17. Similarly, the group SL 2 (k) acts on the Hirzebruch surface Fm ,
preserving the zero section of the fibration π : Fm → P1k . This gives examples of
groups with Property (FW) acting on Fm and preserving a big and nef curve C.
Starting with one of the above examples, one can blow-up points on the invariant
curve C, and then contract C, to get examples of zigzag completions Y on which Γ
acts and contracts the boundary ∂Y .
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
45
9. B IRATIONAL TRANSFORMATIONS OF SURFACES II
In this section, U is a (normal, singular) affine surface with a completion X by a
cycle of ` rational curves. Every irreducible component Ei of the boundary ∂X =
X r U is contracted by at least one automorphism of U. Our goal is to classify
subgroups Γ of Aut(U) ⊂ Bir(X) that are infinite and have Property (FW): in fact,
we shall show that no such group exists. This ends the proof of Theorem B since
all the other possibilities of Proposition 8.3 have been dealt with in the previous
section.
Example 9.1. Let (A1k )∗ denote the complement of the origin in the affine line A1k ;
it is isomorphic to the multiplicative group Gm over k. The surface (A1k )∗ × (A1k )∗
is an open subset in P2k whose boundary is the triangle of coordinate lines {[x :
y : z]; xyz = 0}. Thus, the boundary is a cycle of length ` = 3. The group of
automorphisms of (A1k )∗ × (A1k )∗ is the semi-direct product
GL 2 (Z) n (Gm (k) × Gm (k));
it does not contain any infinite group with Property (FW).
9.1. Resolution of indeterminacies. Let us order cyclically the irreducible components Ei of ∂X, so that Ei ∩ Ej 6= ∅ if and only if i − j = ±1(mod `). Blowing
up finitely many singularities of ∂X, we may assume that ` = 2m for some integer
m ≥ 1; in particular, every curve Ei is smooth. (With such a modification, one
may a priori create irreducible components of ∂X that are not contracted by the
group Γ.)
Lemma 9.2. Let f be an automorphism of U and let fX be the birational extension
of f to the surface X. Then
(1) Every indeterminacy point of fX is a singular point of ∂X, i.e. one of the
intersection points Ei ∩ Ei+1 .
(2) Indeterminacies of fX are resolved by inserting chains of rational curves.
Property (2) means that there exists a resolution of the indeterminacies of fX ,
given by two birational morphisms : Y → X and π : Y → X with f ◦ = π, such
that π −1 (∂X) = −1 (X) is a cycle of rational curves. Some of the singularities of
∂X have been blown-up into chains of rational curves to construct Y .
Proof. Consider a minimal resolution of the indeterminacies of fX . It is given by
a finite sequence of blow-ups of the base points of fX , producing a surface Y and
two birational morphisms : Y → X and π : Y → X such that fX = π ◦ −1 .
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
46
π
f
F IGURE 1. A blow-up sequence creating two (red) branches. No branch
of this type appears for minimal resolution.
Since the indeterminacy points of fX are contained in ∂X, all necessary blow-ups
are centered on ∂X.
The total transform F = ∗ (∂X) is a union of rational curves: it is made of a
cycle, together with branches emanating from it. One of the assertions (1) and (2)
fails if and only if F is not a cycle; in that case, there is at least one branch.
Each branch is a tree of smooth rational curves, which may be blown-down onto
a smooth point; indeed, these branches come from smooth points of the main cycle
that have been blown-up finitely many times. Thus, there is a birational morphism
η : Y → Y0 onto a smooth surface Y0 that contracts the branches (and nothing
more).
The morphism π maps F onto the cycle ∂X, so that all branches of F are contracted by π. Thus, both and π induce (regular) birational morphisms 0 : Y0 → X
and π0 : Y0 → X. This contradicts the minimality of the resolution.
Let us introduce a family of surfaces
πk : Xk → X.
First, X1 = X and π1 is the identity map. Then, X2 is obtained by blowing-up the `
singularities of ∂X1 ; X2 is a compactification of U by a cycle ∂X2 of 2` = 2m+1
smooth rational curves. Then, X3 is obtained by blowing up the singularities of
∂X2 , and so on. In particular, ∂Xk is a cycle of 2k−1 ` = 2m+k−1 curves.
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
47
Denote by Dk the dual graph of ∂Xk : vertices of Dk correspond to irreducible
components Ei of ∂Xk and edges to intersection points Ei ∩ Ej . A simple blow-up
(of a singular point) modifies both ∂Xk and Dk locally as follows
F IGURE 2. Blowing-up one point.
˜
The group Aut(U) acts on Hyp(X)
and Lemma 9.2 shows that its action stabilizes
˜
the subset B of Hyp(X) defined as
n
o
˜
B = C ∈ Hyp(X)
: ∃k ≥ 1, C is an irreducible component of ∂Xk .
In what follows, we shall parametrize B in two distinct ways by rational numbers.
9.2. Farey and dyadic parametrizations. Consider an edge of the graph D1 , and
identify this edge with the unit interval [0, 1]. Its endpoints correspond to two adjacent components Ei and Ei+1 of ∂X1 , and the edge corresponds to their intersection q. Blowing-up q creates a new vertex (see Figure 2). The edge is replaced
by two adjacent edges of D2 with a common vertex corresponding to the exceptional divisor and the other vertices corresponding to (the strict transforms of) Ei
and Ei+1 ; we may identify this part of D2 with the segment [0, 1], the three vertices
with {0, 1/2, 1}, and the two edges with [0, 1/2] and [1/2, 1].
Subsequent blow-ups may be organized in two different ways by using either a
dyadic or a Farey algorithm (see Figure 3).
In the dyadic algorithm, the vertices are labelled by dyadic numbers m/2k . The
vertices of Dk+1 coming from an initial edge [0, 1] of D1 are the points {n/2k ; 0 ≤
n ≤ 2k } of the segment [0, 1]. We denote by Dyad(k) the set of dyadic numbers
n/2k ∈ [0, 1]; thus, Dyad(k) ⊂ Dyad(k + 1). We shall say that an interval [a, b] is
a standard dyadic interval if a and b are two consecutive numbers in Dyad(k) for
some k.
In the Farey algorithm, the vertices correspond to rational numbers p/q. Adjacent
vertices of Dk coming from the initial segment [0, 1] correspond to pairs of rational
numbers (p/q, r/s) with ps − qr = ±1; two adjacent vertices of Dk give birth to
a new, middle vertex in Dk+1 : this middle vertex is (p + r)/(q + s) (in the dyadic
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
48
algorithm, the middle vertex is the “usual” euclidean middle). We shall say that an
interval [a, b] is a standard Farey interval if a = p/q and b = r/s with ps − qr =
−1. We denote by Far(k) the finite set of rational numbers p/q ∈ [0, 1] that is given
by the k-th step of Farey algorithm; thus, Far(1) = {0, 1} and Far(k) is a set of
2k+1 rational numbers p/q with 0 ≤ p ≤ q. (One can check that 1 ≤ q ≤ Fib(k),
with Fib(k) the k-th Fibonacci number.)
0/1
0/1
1/1
0/1
1/1
0/1
1/3
1/1
0/1
1/3
1/1
0/1
1/2
0/1
0/1
1/4
1/3
1/2
1/3
1/2
2/5
3/5
3/4
1/1
1/2
1/4
1/2
1/4
1/8
1/1
3/4
1/2
3/8
1/1
3/4
5/8
1/1
7/8
F IGURE 3. On the left, the Farey algorithm. On the right, the dyadic one.
By construction, the graph D1 has ` = 2m edges. The edges of D1 are in one
to one correspondance with the singularities qj of ∂X1 . Each edge determines a
subset Bj of B; the elements of Bj are the curves C ⊂ ∂Xk (k ≥ 1) such that πk (C)
contains the singularity qj determined by the edge. Using the dyadic algorithm
(resp. Farey algorithm), the elements of Bj are in one-to-one correspondence with
dyadic (resp. rational) numbers in [0, 1]. Gluing these segments cyclically together
one gets a circle S1 , together with a nested sequence of subdivisions in `, 2`, . . .,
2k−1 `, . . . intervals; each interval is a standard dyadic interval (resp. standard Farey
interval) of one of the initial edges .
Since there are ` = 2m initial edges, we may identify the graph D1 with the circle
S1 = R/Z = [0, 1]/0'1 and the initial vertices with the dyadic numbers in Dyad(m)
modulo 1 (resp. with the elements of Far(m) modulo 1). Doing this, the vertices of
Dk are in one to one correspondence with the dyadic numbers in Dyad(k + m − 1)
(resp. in Far(k + m − 1)).
Remark 9.3. (a).– By construction, the interval [p/q, r/s] ⊂ [0, 1] is a standard
Farey interval if and only if ps − qr = −1, iff it is delimited by two adjacent
elements of Far(m) for some m.
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
49
(b).– If h : [x, y] → [x0 , y 0 ] is a homeomorphism between two standard Farey intervals mapping rational numbers to rational numbers and standard Farey intervals to
standard Farey intervals, then h is the restriction to [x, y] of a unique linear projective transformation with integer coefficients:
at + b
a b
, for some element
of PGL 2 (Z).
h(t) =
c d
ct + d
(c).– Similarly, if h is a homeomorphism mapping standard dyadic intervals to intervals of the same type, then h is the restriction of an affine dyadic map
u
h(t) = 2m t + n , with m, n ∈ Z.
2
In what follows, we denote by GFar the group of self-homeomorphisms of S1 =
R/Z that are piecewise PGL 2 (Z) mapping with respect to a finite decomposition of
the circle in standard Farey intervals [p/q, r/s]. In other words, if f is an element of
GFar , there are two partitions of the circle into consecutive intervals Ii and Ji such
that the Ii are intervals with rational endpoints, h maps Ii to Ji , and the restriction
f : Ii → Ji is the restriction of an element of PGL 2 (Z) (see [34], §1.5.1).
Theorem 9.4. Let U be an affine surface with a compactification U ⊂ X such that
∂X := X r U is a cycle of smooth rational curves. In the Farey parametrization of
˜
the set B ⊂ Hyp(X)
of boundary curves, the group Aut(U) acts on B as a subgroup
of GFar .
Remark 9.5. There is a unique orientation preserving self-homeomorphism of the
circle that maps Dyad(k) to Far(k) for every k. This self-homeomorphism conjugates GFar to the group GDya of self-homeomorphisms of the circle that are piecewise affine with respect to a dyadic decomposition of the circle, with slopes in ±2Z ,
and with translation parts in Z[1/2]. Using the parametrization of B by dyadic numbers, the image of Aut(U) becomes a subgroup of GDya .
Remark 9.6. The reason why we keep in parallel the dyadic and Farey viewpoints
is the following: the Farey viewpoint is more natural for algebraic geometers (this is
related to toric –i.e. monomial– maps and appears clearly in [21]), while the dyadic
viewpoint is more natural to geometric group theorists, because this is the classical
setting used in the study of Thompson groups (see [34], §1.5).
Proof. Lemma 9.2 is the main ingredient. Consider the action of the group Aut(U)
on the set B. Let f be an element of Aut(U) ⊂ Bir(X). Consider an irreducible
curve E ∈ B, and denote by F its image: F = f• (E) is an element of B by
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
50
Lemma 9.2. There are integers k and l such that E ⊂ ∂Xk and F ⊂ ∂Xl . Replacing
Xk by a higher blow-up Xm → X, we may assume that flm := πl−1 ◦ f ◦ πm is
regular on a neighborhood of the curve E (Lemma 9.2). Let qk be one of the two
singularities of ∂Xm that are contained in E, and let E 0 be the second irreducible
component of ∂Xm containing q. If E 0 is blown down by flm , its image is one of
the two singularities of ∂Xl contained in F (by Lemma 9.2). Consider the smallest
integer n ≥ l such that ∂Xn contains the strict transform F 0 = f• (E 0 ); in Xn , the
curve F 0 is adjacent to the strict transform of F (still denoted F ), and f is a local
isomorphism from a neighborhood of q in Xm to a neighborhood of q 0 := F ∩ F 0 in
Xn .
Now, if one blows-up q, the exceptional divisor D is mapped by f• to the exceptional divisor D0 obtained by a simple blow-up of q: f lifts to a local isomorphism
from a neighborhood of D to a neighborhood of D0 , the action from D to D0 being
given by the differential dfq . The curve D contains two singularities of ∂Xm+1 ,
which can be blown-up too: again, f lifts to a local isomorphism if one blow-ups
the singularities of ∂Xn+1 ∩ D0 . We can repeat this process indefinitely. Let us
now phrase this remark differently. The point q determines an edge of Dm , hence
a standard Farey interval I(q). The point q 0 determines an edge of Dn , hence another standard Farey interval I(q 0 ). Then, the points of B that are parametrized by
rational numbers in I(q) are mapped by f• to rational numbers in I(q 0 ) and this
map respects the Farey order: if we identify I(q) and I(q 0 ) to [0, 1], f• is the restriction of a monotone map that sends Far(k) to Far(k) for every k. Thus, on I(q), f•
is the restriction of a linear projective transformation with integer coefficients (see
Remark 9.3-(b)). This shows that f• is an element of GFar .
9.3. Conclusion. Consider the group G∗Dya of self-homeomorphisms of the circle
S1 = R/Z that are piecewise affine with respect to a finite partition of R/Z into
dyadic intervals [xi , xi+1 [ with xi in Z[1/2]/Z for every i, and satisfy
h(t) = 2mi t + ai
with mi ∈ Z and ai ∈ Z[1/2] for every i. This group is known as the Thompson
group of the circle, and is isomorphic to the group G∗Far of orientation-preserving
self-homeomorphisms in GFar (defined in §9.2).
Theorem 9.7 (Farley, Hughes [15, 22]). Every subgroup of G∗Dya (and hence of
GFar ) with Property (FW) is a finite cyclic group.
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
51
Indeed fixing a gap in an earlier construction of Farley [15]2, Hughes proved [22]
that GFar has Property PW, in the sense that it admits a commensurating action
whose associated length function is a proper map (see also Navas’ book [34]). This
implies the conclusion, because every finite group of orientation-preserving selfhomeomorphisms of the circle is cyclic.
Thus, if Γ is a subgroup of Aut(U) with Property (FW), it contains a finite index
˜
subgroup Γ0 that acts trivially on the set B ⊂ Hyp(X).
This means that Γ0 extends
as a group of automorphisms of X fixing the boundary ∂X. Since ∂X supports
a big and nef divisor, Γ0 contains a finite index subgroup Γ1 that is contained in
Aut(X)0 .
Note that Γ1 has Property (FW) because it is a finite index subgroup of Γ. It
preserves every irreducible component of the boundary curve ∂X, as well as its singularities. As such, it must act trivially on ∂X. When we apply Theorem 8.8 to Γ1 ,
the conjugacy ϕ : X → Y can not contract ∂X, because the boundary supports an
ample divisor. Thus, Γ1 is conjugate to a subgroup of Aut(Y ) that fixes a curve
pointwise. This is not possible if Γ1 is infinite (see Theorem 8.8 and the remarks
following it).
We conclude that Γ is finite in case (2) of Proposition 8.3.
√
10. B IRATIONAL ACTIONS OF SL 2 (Z[ d])
We develop here Example 1.4. If k is an algebraically closed field of characteristic√0, therefore containing Q, we denote by σ1 and σ2 the distinct embeddings
√
of Q( d) into k. Let j1 and j2 be the resulting embeddings of SL 2 (Z[ d]) into
SL 2 (k), and j = j1 × j2 the resulting embedding into
G = SL 2 (k) × SL 2 (k).
√
Theorem 10.1. Let Γ be a finite index subgroup of SL 2 (Z[ d]). Let X be an
irreducible projective surface over an algebraically closed field k. Let α : Γ →
Bir(X) be a homomorphism with infinite image. Then k has characteristic zero,
and there exist a finite index subgroup Γ0 of Γ and a birational map ϕ : Y 99K X
such that
(1) Y is the projective plane P2 , a Hirzebruch surface Fm , or C × P1 for some
curve C;
(2) ϕ−1 α(Γ)ϕ ⊂ Aut(Y );
2
The gap in Farley’s argument lies in Prop. 2.3 and Thm. 2.4 of [15].
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
52
(3) there is a unique algebraic homomorphism β : G → Aut(Y ) such that
ϕ−1 α(γ)ϕ = β(j(γ))
for every γ ∈ Γ0 .
Theorem B ensures that the characteristic of k is 0 and that (1) and (2) are satisfied. If Y is P2 or a Hirzebruch surface Fm , then Aut(Y ) is a linear algebraic
group. If Y is a product C × P1 , a finite index subgroup of Γ preserves the projection onto P1 , so that it acts via an embedding into the linear algebraic group
Aut(P1 ) = PGL 2 (k).
When k has positive characteristic, Y is the projective plane, and the Γ-action is
given by a homomorphism Γ → PGL 3 (k). Then we use the fact that for any n,
every homomorphism f : Γ → GL n (k) has finite image. Indeed, it is well-known
that GL n (k) has no infinite order distorted elements: elements of infinite order have
some transcendental eigenvalue and the conclusion easily follows. Since Γ has an
exponentially distorted cyclic subgroup, f has infinite kernel, and infinite normal
subgroups of Γ have finite index (Margulis normal subgroup theorem).
On the other hand, in characteristic zero we conclude the proof of Theorem 10.1
with the following lemma.
√
d). Consider the embedding j
Lemma 10.2.
Let
k
be
any
field
extension
of
Q(
√
of SL 2 (Z[ d]) into G = SL 2 (k) × SL 2 (k) given by the standard embedding into
the left-hand SL 2 and its Galois conjugate in the right-hand√SL 2 . Then for every
linear algebraic group H and homomorphism f : SL 2 (Z[ d]) → H(k), there
exists a unique homomorphism f¯ : G → H of k-algebraic groups such that the
homomorphisms f and f˜ ◦ j coincide on some finite index subgroup of Γ.
Proof. The uniqueness is a consequence of Zariski density of the image of j. Let
us prove the existence. Zariski density allows to reduce to the case when H = SL n .
First, the case k = R is given by Margulis’ superrigidity, along with the fact that
every continuous real representation of SL n (R) is algebraic. The case of fields
containing R immediately follows,
√ and in turn it follows for subfields of overfields
of R (as soon as they contain Q( d)).
11. O PEN PROBLEMS
11.1. Regularization and Calabi-Yau varieties.
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
53
Question 11.1. Let Γ be a group with Property (FW). Is every birational action
of Γ regularizable ? Here regularizable is defined in the same way as pseudoregularizable, but assuming that the action on U is by automorphisms (instead of
pseudo-automorphisms).
A particular case is given by Calabi-Yau varieties, in the strict sense of a simply connected complex projective manifold X with trivial canonical bundle and
hk,0 (X) = 0 for 0 < k < dim(X). For such a variety the group Bir(X) coincides with Psaut(X). One can then ask (1) whether every infinite subgroup Γ of
Psaut(X) with property (FW) is regularizable on some birational model Y of X
(without restricting the action to a dense open subset), and (2) what are the possibilities for such a group Γ.
11.2. Transfixing birational groups.
Question 11.2. For which irreducible projective varieties X
(1) Bir(X) does not transfix Hyp(X)?
(2) some finitely generated subgroup of Bir(X) does not transfix Hyp(X)?
(3) some cyclic subgroup of Bir(X) does not transfix Hyp(X).
We have the implications: X is ruled ⇒ (3) ⇒ (2) ⇒ (1). In dimension 2, we
have: ruled ⇔ (1) /⇒(2) ⇔ (3) (see §7.1). It would be interesting to find counterexamples to these equivalences in higher dimension, and settle each of the problems
raised in Question 11.2 in dimension 3.
11.3. The affine space. The group of affine transformations of A3C contains SL 3 (C),
and this group contains many subgroups with Property (FW). In the case of surfaces,
Theorem B shows that groups of birational transformations with Property (FW)
are contained in algebraic groups, up to conjugacy. The following question asks
whether this type of theorem may hold for Aut(A3C ).
Question 11.3. Does there exist an infinite subgroup of Aut(A3C ) with Property (FW)
that is not conjugate to a group of affine transformations of A3C ?
11.4. Length functions. Recall that a length function ` on a group G is quasigeodesic if there exists M > 0 such that for every n ≥ 1 and every g ∈ G with
−1
`(g) ≤ n, there exist 1 = g0 , g1 , . . . , gn = g in G such that `(gi−1
gi ) ≤ M for all i.
−1
Equivalently G, endowed with the distance (g, h) 7→ `(g h), is quasi-isometric to
a connected graph.
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
54
Question 11.4. Given an irreducible variety X, is the length function
g ∈ Bir(X) 7→ |Hyp(X)4gHyp(X)|
quasi-geodesic? In particular, what about X = P2 and the Cremona group Bir(P2 )?
R EFERENCES
[1] Wolf P. Barth, Klaus Hulek, Chris A. M. Peters, and Antonius Van de Ven. Compact complex
surfaces, volume 4 of Ergebnisse der Mathematik und ihrer Grenzgebiete. Springer-Verlag,
Berlin, second edition, 2004.
[2] Hyman Bass. Groups of integral representation type. Pacific J. Math., 86(1):15–51, 1980.
[3] Hyman Bass and Alexander Lubotzky. Automorphisms of groups and of schemes of finite type.
Israel J. Math., 44(1):1–22, 1983.
[4] Arnaud Beauville. Surfaces algébriques complexes. Société Mathématique de France, Paris,
1978. Avec une sommaire en anglais, Astérisque, No. 54.
[5] Eric Bedford and Kyounghee Kim. Dynamics of (pseudo) automorphisms of 3-space: periodicity versus positive entropy. Publ. Mat., 58(1):65–119, 2014.
[6] Nicolas Bergeron, Frédéric Haglund, and Daniel T. Wise. Hyperplane sections in arithmetic
hyperbolic manifolds. J. Lond. Math. Soc. (2), 83(2):431–448, 2011.
[7] Sébastien Boucksom, Charles Favre, and Mattias Jonsson. Valuations and plurisubharmonic
singularities. Publ. Res. Inst. Math. Sci., 44(2):449–494, 2008.
[8] Serge Cantat. Sur les groupes de transformations birationnelles des surfaces. Annals Math.,
(174), 2011.
[9] Serge Cantat. Dynamics of automorphisms of compact complex surfaces. In Frontiers in Complex Dynamics: In celebration of John Milnor’s 80th birthday, volume 51 of Princeton Math.
Series, pages 463–514. Princeton University Press, 2014.
[10] Serge Cantat and Keiji Oguiso. Birational automorphism groups and the movable cone theorem for Calabi-Yau manifolds of Wehler type via universal Coxeter groups. Amer. J. Math.,
137(4):1013–1044, 2015.
[11] Yves Cornulier. Group actions with commensurated subsets, wallings and cubings. ArXiv
1302.5982, (1):1–71, 2015.
[12] Yves Cornulier. Irreducible lattices, invariant means, and commensurating actions. Math. Z.,
279(1-2):1–26, 2015.
[13] Pierre de la Harpe and Alain Valette. La propriété (T ) de Kazhdan pour les groupes localement
compacts (avec un appendice de Marc Burger). Astérisque, (175):158, 1989. With an appendix
by M. Burger.
[14] J. Diller and C. Favre. Dynamics of bimeromorphic maps of surfaces. Amer. J. Math.,
123(6):1135–1169, 2001.
[15] Daniel S. Farley. Proper isometric actions of Thompson’s groups on Hilbert space. Int. Math.
Res. Not., (45):2409–2414, 2003.
[16] M. J. Fryers. The movable fan of the horrocks-mumford quintic. unpublished manuscript,
arXiv:math/0102055, pages 1–20, 2001.
[17] M. H. Gizatullin. Invariants of incomplete algebraic surfaces that can be obtained by means of
completions. Izv. Akad. Nauk SSSR Ser. Mat., 35:485–497, 1971.
[18] M. H. Gizatullin and V. I. Danilov. Automorphisms of affine surfaces. I. Izv. Akad. Nauk SSSR
Ser. Mat., 39(3):523–565, 703, 1975.
[19] M. H. Gizatullin and V. I. Danilov. Automorphisms of affine surfaces. II. Izv. Akad. Nauk SSSR
Ser. Mat., 41(1):54–103, 231, 1977.
COMMENSURATING ACTIONS OF BIRATIONAL GROUPS
55
[20] Robin Hartshorne. Algebraic geometry. Springer-Verlag, New York, 1977. Graduate Texts in
Mathematics, No. 52.
[21] John H. Hubbard and Peter Papadopol. Newton’s method applied to two quadratic equations in
C2 viewed as a global dynamical system. Mem. Amer. Math. Soc., 191(891):vi+146, 2008.
[22] Bruce Hughes. Local similarities and the haagerup property. with an appendix by daniel s.
farley. Groups Geom. Dyn., 3(2):299–315, 2009.
[23] Shigeru Iitaka. Algebraic geometry: an introduction to the birational geometry of algebraic
varieties, volume 76. Springer, 1982.
[24] Vaughan FR Jones. A no-go theorem for the continuum limit of a periodic quantum spin chain.
arXiv preprint arXiv:1607.08769, 2016.
[25] János Kollár, Karen E. Smith, and Alessio Corti. Rational and nearly rational varieties, volume 92 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 2004.
[26] Robert Lazarsfeld. Positivity in algebraic geometry. I, volume 48 of Ergebnisse der Mathematik
und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics. Springer-Verlag,
Berlin, 2004. Classical setting: line bundles and linear series.
[27] David I. Lieberman. Compactness of the Chow scheme: applications to automorphisms and
deformations of Kähler manifolds. In Fonctions de plusieurs variables complexes, III (Sém.
François Norguet, 1975–1977), volume 670 of Lecture Notes in Math., pages 140–186.
Springer, Berlin, 1978.
[28] Alexander Lubotzky, Shahar Mozes, and M. S. Raghunathan. The word and Riemannian metrics on lattices of semisimple groups. Inst. Hautes Études Sci. Publ. Math., (91):5–53 (2001),
2000.
[29] Alexander Lubotzky, Shahar Mozes, and MS Raghunathan. Cyclic subgroups of exponential
growth and metrics on discrete groups. Comptes Rendus Académie des Sciences Paris Série 1,
317:735–740, 1993.
[30] Yu. I. Manin. Cubic forms, volume 4 of North-Holland Mathematical Library. North-Holland
Publishing Co., Amsterdam, second edition, 1986. Algebra, geometry, arithmetic, Translated
from the Russian by M. Hazewinkel.
[31] T. Matsusaka. Polarized varieties, fields of moduli and generalized Kummer varieties of polarized abelian varieties. Amer. J. Math., 80:45–82, 1958.
[32] T. Matsusaka and D. Mumford. Two fundamental theorems on deformations of polarized varieties. Amer. J. Math., 86:668–684, 1964.
[33] J. S. Milne. Algebraic geometry. preprint, http://www.jmilne.org/math/CourseNotes/AG.pdf,
pages 1–221, 2017.
[34] Andrés Navas. Groups of circle diffeomorphisms. Chicago Lectures in Mathematics. University
of Chicago Press, Chicago, IL, spanish edition, 2011.
[35] Jean-Pierre Serre. Arbres, amalgames, SL2 . Société Mathématique de France, Paris, 1977.
Avec un sommaire anglais, Rédigé avec la collaboration de Hyman Bass, Astérisque, No. 46.
[36] Ravi Vakil. The Rising see, Foundations of Algebraic Geometry. webpage of the author,
math216.wordpress.com, 2017.
U NIV R ENNES , CNRS, IRMAR - UMR 6625, F-35000 R ENNES
CNRS AND U NIV LYON , U NIV C LAUDE B ERNARD LYON 1, I NSTITUT C AMILLE J ORDAN ,
43 BLVD . DU 11 NOVEMBRE 1918, F-69622 V ILLEURBANNE
E-mail address: [email protected]
E-mail address: [email protected]
| 4math.GR
|
Predictive Second Order Sliding Control of Constrained Linear Systems
with Application to Automotive Control Systems
arXiv:1802.01617v1 [cs.SY] 5 Feb 2018
Mohammad Reza Amini† , Mahdi Shahbakhti‡, and Jing Sun†
Abstract— This paper presents a new predictive second order
sliding controller (PSSC) formulation for setpoint tracking of
constrained linear systems. The PSSC scheme is developed by
combining the concepts of model predictive control (MPC)
and second order discrete sliding mode control. In order to
guarantee the feasibility of the PSSC during setpoint changes,
a virtual reference variable is added to the PSSC cost function
to calculate the closest admissible set point. The states of
the system are then driven asymptotically to this admissible
setpoint by the control action of the PSSC. The performance
of the proposed PSSC is evaluated for an advanced automotive
engine case study, where a high fidelity physics-based model
of a reactivity controlled compression ignition (RCCI) engine
is utilized to serve as the virtual test-bed for the simulations.
Considering the hard physical constraints on the RCCI engine
states and control inputs, simultaneous tracking of engine load
and optimal combustion phasing is a challenging objective to
achieve. The simulation results of testing the proposed PSSC on
the high fidelity RCCI model show that the developed predictive
controller is able to track desired engine load and combustion
phasing setpoints, with minimum steady state error, and no
overshoot. Moreover, the simulation results confirm the robust
tracking performance of the PSSC during transient operations,
in the presence of engine cyclic variability.
I. INTRODUCTION
Feasibility and stability of model predictive control (MPC)
scheme for a tracking control problem, where the desired
trajectory changes, are challenging to be guaranteed. Addition of a virtual reference, as an extra optimization variable,
to the cost function of the MPC has been proposed in the
literature [1]–[3] to solve this issue. This approach is similar
to the so-called command governor technique [4] for tracking
problem of constrained linear systems, in which a nonlinear
low-pass filter of the reference is added (the output of the
filter can be seen as the virtual reference) to ensure the
admissible evolution of the system to the reference. For
the proposed approach in [1]–[3], an additional term in the
cost function of the MPC reflects the difference between the
original and the virtual references. Moreover, an augmented
space consisting of the original system’s states and the
virtual reference, is used to formulate a terminal invariant
set to guarantee the persistent feasibility of the constrained
optimization problem, which is then utilized to prove the
asymptotic stability of model predictive controllers [3], [5].
The disadvantage of this approach is the increased complex† M. R. Amini and J. Sun are with the Department of Naval Architecture
and Marine Engineering, University of Michigan, Ann Arbor, MI 48109
USA. ({mamini,jingsun}@umich.edu)
‡ M. Shahbakhti is with the Department of Mechanical EngineeringEngineering Mechanics, Michigan Technological University, Houghton, MI
49931 USA. ([email protected])
ity of the invariant set computation due to extra variable
introduced to the optimization problem.
Sliding mode control (SMC) is proven to be a robust and
computationally efficient solution for tracking problems of
linear and nonlinear systems with a great deal of uncertainty [6]. The key feature of SMC is converting a high dimensional tracking control problem into a lower dimensional
stabilization control problem. On the other side, conventional
SMCs do not consider the constraints on the states and inputs
of the system. This is because the control input of the SMC
is mainly calculated based on the present information of the
system, and information from future events is not considered
in the controller formulation. A novel combination of MPC
and first order discrete SMC (DSMC) has been proposed
in [7] to handle state and input constraints within the sliding
controller by taking into account the future information via
a receding horizon sliding control (RHSC) scheme.
The concept of second order DSMC has been proposed in
the authors’ previous works [8], [9], where the asymptotic
stability of the second order sliding mode controller is also
proven. The main advantage of a second order DSMC in
comparison with the first order controller is its enhanced
smoothness in the tracking, and improved response to upcoming changes [8]–[10]. This is due to the fact that in a
second order DSMC, in addition to the sliding variable, its
derivative is also driven to zero [11].
The second order DSMC provides fast and computationally efficient tracking performance [12], [13]. However, as
mentioned earlier, the constraints on the states and inputs of
the system can not be incorporated in the controller design,
which may cause saturation in the controller system [7].
Therefore, formulating the second order sliding controller
in a predictive scheme, not only preserves the key features
of the sliding control and MPC, but also allows for handling
the input and state constraints within the sliding controller
formulation [14]. In this paper, a novel predictive second
order sliding controller (PSSC) is presented. The proposed
PSSC enables handling of the constraint on the inputs and
states by direct inclusion of the future information. Moreover,
the PSSC ensures asymptotic tracking of any admissible
piecewise setpoints, or the closest admissible setpoint, if the
target trajectory is not admissible.
The performance of the proposed PSSC is evaluated
for control of an advanced automotive engine including a
Reactivity Controlled Compression Ignition (RCCI). RCCI
technology offers significantly higher fuel conversion efficiency compared to conventional automotive engines through
creating optimum heat release shape by exploiting reactivity
gradients inside the combustion chamber. RCCI typically
runs with lean air-fuel mixture, and has low combustion
temperature, resulting in much lower NOx and PM engineout emissions, compared to conventional engines [15]. On the
other side, simultaneous combustion phasing and engine load
control is hard to achieve due to the strong nonlinearity and
internal coupling in the dynamics of RCCI engines, specifically during the transients. It will be shown in this paper that
the PSSC is able to handle the coupling within the RCCI
engine to achieve simultaneous engine load and combustion
phasing tracking, with consideration of the constraints on the
control signals during the engine transient operation.
The main contribution of this paper includes the first
development of a predictive second order sliding controller
(PSSC) for tracking of constrained linear systems. The PSSC
is formulated with respect to a novel invariant sliding domain
to ensure feasibility and stability of the controller. Moreover,
this paper presents the first application of a predictive sliding
controller for simultaneous load and combustion phasing
tracking in RCCI engines.
II. D ISCRETE S ECOND O RDER S LIDING C ONTROL
A linear multi-input multi-output system can be represented by the following discrete-time state space equations:
x(k + 1) = Ax(k) + Bu(k),
(1)
y(k) = Cx(k),
(2)
where x ∈ Rn , u ∈ Rm , y ∈ Rm are the state, input, and
output of the linear system, respectively. A ∈ Rn×n , B ∈
Rn×m , C ∈ Rm×n are the system matrices. Moreover, it
is assumed that the state and input of the linear system in
Eq. (1) are constrained: x ∈ X ⊆ Rn , u ∈ U ⊆ Rm . X and
U sets are restricted to have the origin in their interior and
be polyhedral [5]. The control objective is to drive the output
of the system (y) to its desired value (yd ) in the presence of
the state and input constraints. To this end, an error function
(e) is defined:
e(k) = y(k) − y d (k)
(3)
and the control objective is set to drive e to zero. Here, the
first order sliding function (s) definition of a linear system
with relative degree of d1 , · · · , dm from [5] is adopted to
define the second order sliding function (ξ), which will
be used later to formulate the predictive sliding controller.
Moreover, it is assumed that the linear system in Eq. (1) is
minimum-phase [16]. The first order sliding variable can be
defined with respect to Eq. (3) as follows:
s1 (k)
s(k) = ... =
(4)
sm (k)
α1,0 e1 (k) + α1,1 e1 (k + 1) + · · · + α1,l1 e1 (k + l1 )
..
.
αm,0 em (k) + αm,1 em (k + 1) + · · · + αm,lm em (k + lm )
6 0, i = 1, · · · , m. By
where li = di − 1, and αi,li =
substituting Eq. (3) into Eq. (4), s can be simplified as:
s(k) = Gx(k) − H(k)
where:
l1
P
j
(5)
α1,j A
c1
j=0
.
,
..
G=
lm
P
j
cm
αm,j A
(6)
j=0
α1,0 yd,1 (k) + · · · + α1,l1 yd,1 (k + l1 )
..
H(k) =
(7)
.
αm,0 yd,m (k) + · · · + αm,lm yd,m (k + lm )
where ci , i = 1, · · · , m is the ith row of C. The interesting
point about Eq. (6) is that ci Aj B = 0, i = 1, · · · , m, j =
1, · · · , li [5]. Therefore no term containing B appears in
Eq. (6). If ci Aj B 6= 0, then the ith output of the system
has a relative degree less than di , which violates the initial
statement of assuming the ith output to have a relative degree
of di , thus ci Aj B = 0.
In the discrete time, the second order sliding variable
(ξ(k)) is defined with respect to Eq. (4) as follows [9], [10]:
ξ(k) = s(k + 1) + βs(k)
(8)
where β ∈ Rm×m , which manipulates the rate of state
convergence to the sliding manifold, is chosen such that all
the eigenvalues are inside the unit circle [8]. The equivalent control input of the second order sliding controller is
achieved by solving the following equality [9]:
ξ(k + 1) = ξ(k) = 0.
(9)
Applying Eq. (9) leads to the following equivalent control
input (ueq (k)) of the second order DSMC:
ueq (k) =
(10)
−1
−(GB)
GA + βG x(k) − H(k + 1) + βH(k) .
Eq. (10) calculates a vector of control input signals (u)
based on the model matrices, current states (x), and also
current and future information of the desired trajectories
(yd,i (k), · · · , yd,i (k+li +1), i = 1, · · · , m). Moreover, since
it is assumed that the system in Eq. (1) has a relative degree
of d1 , · · · , dm , the existence of (GB)−1 is ensured [5], [16].
III. P REDICTIVE S ECOND O RDER S LIDING C ONTROL
In this section, a novel combination of the second order
sliding controller with MPC, as the predictive second order
sliding control (PSSC) scheme is presented for constrained
linear system tracking. To this end, following the procedure
proposed in [1], [5], first a terminal invariant second order
sliding domain is calculated. Next, this invariant set is used
as extended terminal constraint to formulate the PSSC.
In the first step, it is required to determine a terminal state
feedback control law, based on the second order sliding control, in order to characterize the steady states and inputs [17].
Moreover, the reference output (y d ) in Eq. (3) is replaced
with a virtual fixed but arbitrary reference (ỹ d ) to calculate
the invariant second order sliding domain, according to the
procedure in [1], [5]. By substituting ỹ d in Eq. (3), the
error becomes: ẽ(k) = y(k) − ỹ d . It was shown in [5] that
substitution of ẽ(k) in Eq. (5) results in the following first
order sliding function:
s(k) = Gx(k) − H̃ ỹ d
(11)
where:
l1
P
α1,j
j=0
H̃ =
..
.
lm
P
j=0
αm,j
Thus, the second order sliding function becomes:
ξ(k) = Gx(k + 1) − H̃ ỹ d + β Gx(k) − H̃ ỹ d ,
(12)
(24)
Eventually, based on the vectors defined in Eq. (22)-(24),
the PSSC, as a constrained finite-time control problem, is
formulated:
min
kΞ(k + 1)k2 + kỹ d − y d (k)k1
X(k),U (k),ỹd
subject to
(13)
GA + βG x(k) − I + β H̃ ỹ d .
(14)
Eq. (14) can be written in the following form:
ueq (k) = Kx(k) + Lỹ d
(22)
In a similar manner, the state and input vectors over the
prediction horizon are defined as:
X(k) = [x(k) · · · x(k + N )]
(23)
U (k) = [u(k) · · · u(k + N − 1)]
and, the equivalent control input from Eq. (10) is calculated
as:
ueq (k) = −(GB)−1
for the PSSC, as follows:
Ξ(k + 1) = [ξ(k + 1) · · · ξ(k + N )].
(15)
where, K = −(GB)−1 (GA + βG) and L = (GB)−1 (I +
β)H̃. It can be seen from Eq. (15) that the equivalent control
input has two terms, the standard state feedback term (Kx),
and Lỹ d , which accounts for changes in the reference.
In the second step, ueq from Eq. (15) is used as a terminal
control input to determine the invariant set for tracking. To
this end, the equivalent control input from Eq. (15) is plugged
into Eq. (1):
x(k + 1) = (A + BK)x(k) + BLỹ d .
(16)
The state (x) and ỹ d vectors are then augmented (w =
[x⊺ ỹ ⊺d ]⊺ ) with respect to Eq. (15) to define the equivalent
dynamic [1], [18] of the linear system in Eq. (1):
w(k + 1) = Aeq w(k)
(17)
s(i) = Gx(i) − H̃ ỹ d ,
i=k+1,··· ,k+N
ξ(i) = s(i + 1) + βs(i), i=k+1,··· ,k+N −1
x(i + 1) = Ax(i) + Bu(i), i=k,··· ,k+N −1
x(i) ∈ X , u(i) ∈ U, i=k,··· ,k+N −1
[x⊺ (k + N ) ỹ ⊺d ]⊺ ∈ T
(25)
ỹ d in Eq. (25) is now an optimization variable which
is chosen by the PSCC optimization algorithm. Moreover,
kỹ d − y d (k)k1 is an offset cost in order to account for the
deviation of ỹ d from y d [1], [5]. It should be noted that
the concept of second order sliding mode control is utilized
to calculate the invariant set, via the equivalent control input
ueq from Eq. (15), as a terminal constraint in the formulation
of the PSSC. Moreover, in the absence of the constraints, and
assuming a prediction horizon of N =1, the PSSC becomes
the ideal second order DSMC. The feasibility and stability
analysis of the proposed PSSC in Eq. (25) can be proven
by exploiting the results from [1], [5], [17], with respect to
the calculated invariant second order sliding domain from
Eq. (20). This analysis is skipped here, as it is out of the
scope of this paper. The next section centers on illustration
of the PSSC for a challenging automotive control problem.
IV. C ASE S TUDY: RCCI E NGINE C ONTROL
where:
A + BK
=
0m×n
BL
.
I m×m
and, the projected domain of T (Eq. (20)) on the original
state space (x) is called Z:
Z = projx (T ).
(21)
An RCCI engine, with highly nonlinear and internally
coupled dynamics is chosen for the case study in this
paper. The RCCI is a promising engine technology with
challenging control design issues. Its performance is sensitive
to operating conditions, and the combustion can become
unstable if not controlled properly. Moreover, the engine,
as a nonlinear dynamic plant, is subject to actuator and
state constraints. These features make the RCCI engine an
appropriate candidate for evaluating the proposed PSSC,
as it can be designed based on a linear structure, and be
implemented on the nonlinear plant while enforcing the
physical constraints.
A. Predictive Controller Formulation
In this section, the formulation of a predictive second order
sliding controller (PSSC) is presented by incorporating the
calculated invariant second order sliding domain for tracking
from Eq. (20). The second order sliding variable from
Eq. (13) over an N -step prediction horizon is constructed
A. High Fidelity Physics-Based RCCI Model
A high fidelity physics-based RCCI engine model
from [19], [20] is used to evaluate the performance of the
proposed PSSC. The model is developed, and has been experimentally validated, to predict start of combustion (SOC),
crank angle for 50% fuel burnt (CA50), and indicated mean
Aeq
(18)
Moreover, the state and input constraints can be re-written
for the augmented system in Eq. (17) as:
Weq = {w : [I n×n 0n×m ]w ∈ X , [K L]w ∈ U} (19)
Finally, the invariant set (T ) is calculated as:
T = {w : Akeq w ∈ Weq , ∀k ≥ 0 }
(20)
effective pressure (IMEP). The details of the high fidelity
physics-based RCCI dynamic model, specifications of the
RCCI engine (4-cylinder 2-liter GM engine), and experimental validation results of the model are available in [19].
The highly nonlinear nature of the physics-based RCCI
model makes it difficult to use in the design of linear
controllers. Thus, in order to design the PSSC for CA50
and IMEP tracking, first the high fidelity model is linearized
around an operation point. The linearized model has four
states in the state space:
x = [CA50 Tsoc Psoc IM EP ]⊺
(26)
where Psoc and Tsoc are pressure and temperature at SOC.
The output vector (y) of the linear model is:
y = [CA50 IM EP ]⊺
(27)
and the control input vector is:
u = [SOI F Q]⊺
(28)
where, FQ is the total injected fuel quantity, and SOI is
the start of injection. Since CA50 and IMEP are the only
measurable outputs of the RCCI model, a Kalman filter is
designed to estimate the Psoc and Tsoc , with respect to the
linearized model [19]. The Kalman filter is updated every
cycle based on the outputs of the high fidelity RCCI model.
Finally, the PSSC is formulated to track the desired setpoints
of IMEP and CA50 as shown in Fig. 1. Since both outputs
of the linearized model have a relative degree of one with
respect to either of the inputs, the vector of the first order
sliding function is defined as follows:
CA50(k) − CA50target (k)
s(k) =
(29)
IM EP (k) − IM EPtarget (k)
and, the second order sliding function is calculated with
respect to Eq. (8). The PSSC simulations on the RCCI
model are performed in MATLAB® /SIMULINK® by using
YALMIP [21] for formulating the optimization problem. The
predication horizon is set to N =5 engine cycles.
Fig. 1. Schematic of the designed predictive second order sliding controller
for adjusting RCCI combustion phasing (CA50) and load (IMEP).
B. Predictive Sliding Control of the RCCI Engine
Figs. 2 shows the performance of the proposed PSSC
and the second order DSMC in tracking variable CA50
and IMEP trajectories, where the high fidelity physics-based
RCCI model is considered as the plant according to Fig. 1.
The constraints on the control inputs are selected to avoid
unstable combustion. The matrix (β) of the second order
Fig. 2. CA50 and IM EP control using PSSC and second order DSMC.
Plant: High Fidelity Model. Operating conditions: P R= 20, Tin = 333.1
K, Pin = 95 kPa, Ne = 1000 RP M , F Qmin =15 mg/cycle.
sliding vector is the same for the PSSC and the second order
DSMC. However, there are fundamental differences between
the tracking performances of these two controllers. Since the
original reference (yd ) is feasible, the admissible reference
(ỹd ) is the same as yd . Thus, it is expected that by applying
the second order DSMC to the physics-based RCCI model,
the controller shows smooth and offset-free performance.
However, Fig. 2-a and b show that overshoots occur during
the early 10 cycles of the RCCI engine operation. Moreover,
there are steady state errors in CA50 and IMEP tracking,
specifically at higher CA50, and higher loads (IMEP). This
deviation in the second order DSMC performance can be
explained with respect to the error in the linearization process
of the highly nonlinear physics-based RCCI model around
specific operation point. By comparing the second order
DSMC with PSSC in Fig. 2-a and b, it can be observed that
the PSSC reacts to the upcoming changes in the reference
trajectories faster than the second order DSMC. This is
due to the fact that the PSSC takes into account the future
information via the receding horizon strategy.
By changing the constraints on the control signals, the performance of the tracking controller is considerably affected,
as the reference trajectories may not be feasible anymore.
Fig. 3 shows the influence of changing the operation bound
of F Qmin from 15 mg/cycle in Fig. 2, to 19 mg/cycle.
Due to the change in the control input constraint, the lower
IMEP level is not reachable. Since the second order DSMC
does not consider the future information and cannot handle
the constraints on the control input, the calculated F Q signal
saturates (Fig. 3-d), which results in a steady state error in
IMEP tracking (Fig. 3-b).
On the other hand, the proposed PSSC calculates the
feasible (ỹd ) with respect to actual references (yd ). Based
on the calculated feasible CA50 and IMEP references, the
PSSC puts the efforts to reach the desired trajectories with
Fig. 3. CA50 and IM EP control using PSSC and second order DSMC.
Operating conditions: P R= 20, Tin = 333.1 K, Pin = 95 kPa, Ne = 1000
RP M , F Qmin =19 mg/cycle.
lower steady state errors, and with no overshoots, compared
to the second order DSMC. The better performance of the
PSSC in tracking is due to its awareness of the reachable
references, despite the error in the linearized RCCI model,
and also the early reaction to the upcoming changes in the
desired trajectories.
The optimization algorithm of the PSSC determines the
feasible reference (ỹd ), and provides a baseline to compare
the behavior of the second order DSMC with the PSSC. The
tracking results in Fig. 3 show that while the PSSC tracks the
feasible IMEP reference accurately, the tracking performance
for CA50 is not affected due to the constraint on the F Q,
and the small steady state error is because of the error in the
RCCI model linearizion. On the other side, because of the
linearizion process error, and the saturation in the control
signal of the second order DSMC, both CA50 and IMEP
tracking are affected, and steady state errors can be seen for
both references, specifically at higher loads (IMEP). Additionally, it can be observed that overshoots happen during
the first 10 cycles of the second order DSMC operation,
while the PSSC provides overshoot-free performance for
both desired trajectories.
The states of the RCCI engine are highly coupled. Thus,
maintaining one of the outputs at a constant level, while the
other output is set to follow a variable reference is challenging to achieve. For example, it is common to maintain an
optimum CA50, while IMEP is changing. In order to evaluate the PSSC performance during these specific operating
conditions, the simulations are performed for two different
scenarios. In the first case, the objective is to maintain a
constant CA50, while the desired IMEP is changing (Fig. 4a1−4 ). In the second case, it is desired to change CA50,
while IMEP needs to be at a constant level of 600 kP a
(Fig. 4-b1−4 ). This represents situations on the RCCI engine
that CA50 needs to be changed due to engine-out emission
constraint, or pressure rise rate constraint.
It can be observed in Fig. 4 that despite the large over
shoot at the be beginning of the engine operation from the
second order DSMC, both PSSC and DSMC show acceptable
IMEP tracking results (Fig. 4-a2 ). However, Fig. 4-a1 depicts
that the DSMC is not able to maintain CA50 at the desired
level, and there is an average error of 2 CAD for reference
CA50 tracking from the second order DSMC. On the other
side, the PSSC shows better CA50 tracking results with an
average tracking error of 0.5 CAD.
Similarly in Fig. 4-b, the PSSC outperforms the second
order DSMC in maintaining the desired IMEP, while tracking the desired CA50 accurately. Specifically, it can be seen
that the second order DSMC fails to keep IMEP at 600 kP a
when the reference CA50 drops to 6 CAD at the 41th engine
cycle. Additionally, the predictive controller does not have
the large overshoots and steady state error as those seen in
the second order DSMC.
Finally, in order to evaluate the PSSC under the RCCI
engine cyclic variability and measurement noise, the PSCC
is studied in Fig. 5 for tracking the same CA50 and IMEP
trajectories evaluated in Fig. 4. It can be observed that under
the introduced measurement noise and cyclic variability, the
PSSC is able to achieve the desired loads (IMEP), while
maintaining CA50 at a constant level (Fig. 5-a1,2 ). Similarly,
Fig. 5-b1,2 illustrates good PSSC performance with average
tracking errors of 11.4 kPa and 0.4 CAD in maintaining the
engine at a constant IMEP, while desired CA50 changes.
V. S UMMARY
AND
C ONCLUSIONS
A new formulation of a predictive sliding controller based
on the concept of second order sliding mode, in combination with MPC, was presented in this paper for tracking
of constrained linear systems. A virtual reference, as the
admissible reference, was added to the cost function of
the predictive controller to account for the changes in the
setpoint. Since inclusion of the virtual reference requires
computation of a terminal set to guarantee the feasibility
of the proposed predictive second order sliding controller
(PSSC), an extended invariant second order sliding domain
was calculated based on the augmented state space of the
system states and the virtual reference. The performance
of the proposed PSSC was demonstrated in a multi-input
multi-output structure for a highly nonlinear and internally
coupled RCCI engine tracking problem. An experimentally
validated physics-based model of an RCCI engine was used
to design and assess the proposed PSSC for simultaneous
load (IMEP) and combustion phasing (CA50) tracking under
hard constraints on the control inputs of the engine. Comparing to the second order DSMC, the proposed PSSC showed
better tracking results, with minimum steady state error, and
no overshoot for different tracking scenarios. Moreover, the
simulation results confirm the robustness of the PSSC under
measurement noise and the engine cyclic variability.
ACKNOWLEDGMENT
Dr. Andreas Hansen from University of California, Berkeley, and
Mr. Akshat Raut from Michigan Tech are gratefully acknowledged
for their technical comments during the course of this study.
Fig. 4. CA50 and IM EP control using PSSC and second order DSMC for maintaining (a) constant CA50, and (b) constant IM EP . Operating
conditions: P R= 20, Tin = 333.1 K, Pin = 95 kPa, Ne = 1000 RP M , F Qmin =15 mg/cycle
(a1 )
CA50 [o aTDC]
12
10
10
8
Reference (yd )
PSSC (Constant CA50)
6
5
4
(a2 )
600
IMEP [kPa]
(b1 )
15
14
(b2 )
700
Reference (yd )
PSSC (Constant IMEP)
650
550
600
500
550
450
500
0
50
100
0
Engine Cycle No.
50
100
Engine Cycle No.
Fig.
5.
CA50
and
IM EP
control
using
PSSC
under measurement noise and cyclic variability for (a) constant CA50,
and (b) constant IM EP . Operating conditions: P R= 20, Tin = 333.1 K,
Pin = 95 kP a, Ne =1000 RP M , cyclic variability of 2o for CA50, and
25 kP a for IM EP , based on the RCCI engine experimental data in [19].
R EFERENCES
[1] D. Limón, I. Alvarado, T. Alamo, and E.F. Camacho. MPC for Tracking Piecewise Constant References for Constrained Linear Systems.
Automatica, 44(9):2382–2387, 2008.
[2] A. Ferramosca, D. Limón, I. Alvarado, T. Alamo, and E.F. Camacho.
MPC for Tracking with Optimal Closed-loop Performance. Automatica, 45(8):1975–1978, 2009.
[3] A. Ferramosca, D. Limón, I. Alvarado, T. Alamo, F. Castaño, and E.F.
Camacho. Optimal MPC for Tracking of Constrained Linear Systems.
International Journal of Systems Science, 42(8):1265–1276, 2011.
[4] E. Garone, S. Di Cairano, and I. Kolmanovsky. Reference and
command governors for systems with constraints: A ssurvey on theory
and applications. Automatica, 75:306–328, 2017.
[5] A. Hansen, Y. Li, and J.K. Hedrick. Invariant Sliding Domains for
Constrained Linear Receding Horizon Tracking Control. IFAC Journal
of Systems and Control, 2:12–17, 2017.
[6] J.-J.E. Slotine and W. Li. “Applied Nonlinear Control”. Prentice-hall
Englewood Cliffs, NJ, Chapters 7 & 8, 1991.
[7] A. Hansen and J. Hedrick. Receding Horizon Sliding Control for
Linear and Nonlinear Systems. ACC, 2015. Chicago, IL, USA.
[8] M. R. Amini, M. Shahbakhti, and S. Pan. MIMO First and Second
Order Discrete Sliding Mode Controls of Uncertain Linear Systems
Under Implementation Imprecisions. ASME DSCC, 2017. Tysons
Corner, VA, USA.
[9] M. R. Amini, M. Shahbakhti, S. Pan, and J. K. Hedrick. Discrete
Adaptive Second Order Sliding Mode Controller Design with Application to Automotive Control Systems with Model Uncertainties. ACC,
2017. Seattle, WA, USA.
[10] M. Mihoub, A. Nouri, and R. Abdennour. Real-Time Application of
Discrete Second Order Sliding Mode Control to a Chemical Reactor.
Control Engineering Practice, 17(9):1089–1095, 2009.
[11] T. Salgado-Jimenez, J. Spiewak, P. Fraisse, and B. Jouvencel. A Robust
Control Algorithm for AUV: Based on a High Order Sliding Mode.
OCEANS’04. MTTS/IEEE TECHNO-OCEAN’04, 2004. Kobe, Japan.
[12] H. Sira-Ramirez. Structure at Infinity, Zero Dynamics and Normal
Forms of Systems Undergoing Sliding Motions. International Journal
of Systems Science, 21(4):665–674, 1990.
[13] M.R. Amini, M. Shahbakhti, S. Pan, and K. Hedrick. Bridging the Gap
Between Designed and Implemented Controllers via Adaptive Robust
Discrete Sliding Mode Control. Control Eng. Practice, 59:1–15, 2017.
[14] A. Mozaffari, N.L. Azad, A. Hansen, and J.K. Hedrick. A Receding
Horizon Sliding Controller for Automotive Engine Coldstart: Design
and Hardware-in-the-Loop Testing with an Echo State Network HighFidelity Model. Asian Journal of Control, 18(4):1219–1238, 2016.
[15] D.A. Splitter, R.M. Hanson, S.L. Kokjohn, and R.D. Reitz. Improving
Engine Performance by Optimizing Fuel Reactivity with a Dual Fuel
PCCI Strategy. Conference on Thermo and Fluid Dynamic Processes
in Diesel Engines, 2010. Valencia, Spain.
[16] A.V. Kraev, A.I. Rogovskii, and V.V. Fomichev. On a Generalization
of Relative Degree. Differential Equations, 50(8):1122–1127, 2014.
[17] D. Limon, I. Alvarado, T. Alamo, and E.F. Camacho. MPC for
Tracking of Piece-Wise Constant References for Constrained Linear
Systems. 16th Triennial World Congress, 2005. Prague, Czech Rep.
[18] S. Spurgeon. Hyperplane Design Techniques for Discrete-Time Variable Structure Control Systems. Int. J. of Control, 55(2):445–456,
1992.
[19] A. Raut, M. Bidarvatan, H. Borhan, and M. Shahbakhti. Closed-Loop
Combustion Control of an RCCI Engine. ACC, 2018. Milwaukee, WI,
USA.
[20] N.N.T. Kondipati, J.K. Arora, M. Bidarvatan, and M. Shahbakhti.
Modeling, Design and Implementation of a Closed-Loop Combustion
Controller for an RCCI Engine. ACC, 2017. Seattle, WA, USA.
[21] J. Lofberg. YALMIP: A toolbox for modeling and optimization in
MATLAB. In 2004 IEEE Int. Symp. on Computer Aided Control Sys.
Design. IEEE, 2004. New Orleans, LA, USA.
| 3cs.SY
|
ON THE FINITENESS PROPERTIES OF LOCAL COHOMOLOGY
MODULES FOR REGULAR LOCAL RINGS
arXiv:1703.00756v1 [math.AC] 2 Mar 2017
MONIREH SEDGHI, KAMAL BAHMANPOUR AND REZA NAGHIPOUR†
Abstract. Let a denote an ideal in a regular local (Noetherian) ring R and let N be a
finitely generated R-module with support in V (a). The purpose of this paper is to show
that all homomorphic images of the R-modules ExtjR (N, Hai (R)) have only finitely many
associated primes, for all i, j ≥ 0, whenever dim R ≤ 4 or dim R/a ≤ 3 and R contains a
field. In addition, we show that if dim R = 5 and R contains a field, then the R-modules
ExtjR (N, Hai (R)) have only finitely many associated primes, for all i, j ≥ 0.
1. Introduction
In the present paper we continue the study of the finiteness properties of local cohomology modules for regular local rings. An interesting problem in commutative algebra
is determining when the set of associated primes of the i-th local cohomology module
Hai (R) of a Noetherian ring R with support in an ideal a of R, is finite. This question
was raised by Huneke in [11] at the Sundance Conference in 1990. Examples given by A.
Singh [25] (in the non-local case) and M. Katzman [16] (in the local case) show there exist
local cohomology modules of Noetherian rings with infinitely many associated primes.
However, in recent years there have been several results showing that this conjecture is
true in many situations. The first result were obtained by Huneke and Sharp. In fact,
Huneke and Sharp [12] (in the case of positive characteristic) have shown that, if R is a
regular local ring containing a field, then Hai (R) has only finitely many associated primes
for all i ≥ 0. Subsequently, G. Lyubeznik in [13] and [14] showed this result for unramified
regular local rings of mixed characteristic and in characteristic zero. Further, Lyubeznik
posed the following conjecture:
Conjecture. If R is a regular ring and a an ideal of R, then the local cohomology
modules Hai (R) have finitely many associated prime ideals for all i ≥ 0.
While this conjecture remains open in this generality, several nice results are now available, see [3, 10, 19]. In lower dimensional cases, Marley in [19] showed that the set of
associated prime ideals of the local cohomology modules Hai (M) is finite if R is a local ring
Key words and phrases. Associated prime, cofinite module, local cohomology, minimax module, regular
ring, weakly Laskerian module.
2010 Mathematics Subject Classification: 13D45, 14B15, 13H05.
This research was in part supported by a grant from Azarbaijan Shahid Madani University (No.
217/d/917).
†
Corresponding author: e-mail: [email protected] (Reza Naghipour).
1
2
M. SEDGHI, K. BAHMANPOUR AND R. NAGHIPOUR
of dimension d and a an ideal of R, M is a finitely generated R-module, in the following
cases: (1) d ≤ 3; (2) d ≤ 4 and R has an isolated singularity; (3) d = 5 and R is an
unramified regular local ring and M is torsion-free.
For a survey of recent developments on finiteness properties of local cohomology modules, see Lyubezniks interesting paper [15].
The purpose of this paper is to provide some results concerning the set of associated
primes of the local cohomology modules for a regular local ring, that almost results are
extensions of Marley’s results on local cohomology modules over the strong ring (i.e. the
regular local ring containing a field). Namely, we show that, for a finitely generated module
N over a regular local ring R with support in V (a), the R-modules ExtjR (N, Hai (R)) are
weakly Laskerian, for all i, j ≥ 0, whenever dim R ≤ 4 or dim R/a ≤ 3 and R contains
a field. In addition, we show that if dim R = 5, then, for all i, j ≥ 0, the R-module
ExtjR (N, Hai (R)) has only finitely many associated primes, when R contains a field.
We say that an R-module M is said to be weakly Laskerian if the set of associated
primes of any quotient module of M is finite [6]. Our main result in Section 2 is to
establish some finiteness results of local cohomology modules for a regular local ring R
with respect to an ideal a with dim R/a ≤ 3. More precisely, we prove the following.
Theorem 1.1. Let R be a regular local ring containing a field with dim R = d ≥ 3, and let
a be an ideal of R such that dim R/a ≤ 3. Then ExtjR (R/a, Hai (R)) is weakly Laskerian,
for all i, j ≥ 0.
The result of Theorem 1.1 is proved in Section 2. Pursuing this point of view further,
we obtain the following consequence of Theorem 1.1, which is an extension of Marley’s
result in [19].
Corollary 1.2. Let R be a regular local ring of dimension d ≤ 4 containing a field and
a an ideal of R. Then, for any finitely generated R-module N with Supp(N) ⊆ V (a), the
R-module ExtnR (N, Hai (R)) is weakly Laskerian, for all integers n, i ≥ 0.
It will be shown in Section 3 that the subjects of Section 2 can be used to prove a
finiteness result of local cohomology modules for a regular local ring of dimension 5. In
fact, we will generalize the main result of Marley for a regular local ring of dimension 5.
More precisely, we shall show that:
Theorem 1.3. Let R be a regular local ring containing a field with dim R = 5 and let a
be an ideal of R. Then the set AssR ExtnR (N, Hai (R)) is finite, for each finitely generated
R-module N with support in V (a) and for all integers n, i ≥ 0.
The proof of Theorem 1.3 is given in Section 3. The following proposition will be one
of our main tools for proving Theorem 1.3.
Proposition 1.4. Let R be a regular local ring of dimension d, and let a be an ideal of R
with height a = 1. Then the set AssR ExtiR (N, Ha1(R)) is finite for each finitely generated
R-module N with support in V (a) and for all integers i ≥ 0.
As a consequence of Theorem 1.3 we derive the following result.
ON THE FINITENESS PROPERTIES OF LOCAL COHOMOLOGY MODULES
3
Theorem 1.5. Let R be a regular local ring containing a field such that dim R ≤ 5, and
let a be an ideal of R. Then, for all integers n, i ≥ 0 and any finitely generated R-module
N with support in V (a), the set AssR ExtnR (N, Hai (R)) is finite.
Finally, in this section we will show that, if M is a finitely generated module over a
regular local ring R with dim R ≤ 4, then for any finitely generated R-module N with
support in V (a), the R-module ExtnR (N, Hai (M)) is weakly Laskerian, for all integers
i, n ≥ 0. In particular the set AssR Hai (M) is finite, for all integers i, n ≥ 0.
Hartshorne [9] introduced the notion of a cofinite module, answering in negative a
question of Grothendieck [8, Exposé XIII, Conjecture 1.1]. In fact, Grothendieck asked
if the modules HomR (R/a, Hai (M)) always are finitely generated for any ideal a of R and
any finitely generated R-module M. This is the case when a = m, the maximal ideal in
a local ring, since the modules Hmi (M) are Artinian. Hartshorne defined an R-module
M to be a-cofinite if the support of M is contained in V (a) and ExtiR (R/a, M) is finitely
generated for all i ≥ 0.
In [26], H. Zöschinger, introduced an interesting class of minimax modules, and he has
in [26, 27] given many equivalent conditions for a module to be minimax. An R-module
N is said to be a minimax module, if there is a finitely generated submodule L of N, such
that N/L is Artinian. The class of minimax modules thus includes all finitely generated
and all Artinian modules. Also, an R-module M is called a-cominimax if the support of
M is contained in V (a) and ExtiR (R/a, M) is minimax for all i ≥ 0. The concept of the
a-cominimax modules is introduced in [2].
Throughout this paper, R will always be a commutative Noetherian ring with non-zero
identity and a will be an ideal of R. For an R-module M, the i-th local cohomology
module of M with respect to a is defined as
Hai (M) = lim ExtiR (R/an , M).
−→
n≥1
We refer the readers to [7] and [5] for more details on local cohomology.
We shall use Min(a) to denote the set of all minimal primes of a. For each R-module
L, we denote by AsshR L the set {p ∈ AssR L : dim R/p = dim L}. Also, for any ideal b
of R, we denote {p ∈ Spec R : p ⊇ b} by V (b). Finally, for any ideal c of R, the radical
of c, denoted by Rad(c), is defined to be the set {x ∈ R : xn ∈ c for some n ∈ N}. For
any unexplained notation and terminology we refer the readers to [5] and [21].
2. Finiteness of local cohomology modules for ideals of small dimension
The purpose of this section is to study the finiteness properties of local cohomology
modules for a regular local ring R with respect to an ideal a of R with dim R/a ≤ 3. The
main goal is Theorem 2.7, which plays a key role in Section 3. This result extends a main
result of T. Marley [19].
The following lemmas and proposition will be needed in the proof of Theorem 2.7.
Theorem 2.1. (Huneke-Sharp and Lyubeznik). Let (R, m) be a regular local ring containing a field. Then for each ideal a of R and all integers i ≥ 0, the set of associated
primes of the local cohomology modules Hai (R) are finite.
4
M. SEDGHI, K. BAHMANPOUR AND R. NAGHIPOUR
Proof. See [12, 13, 14].
Lemma 2.2. Let (R, m) be a regular local ring containing a field with dim R = d ≥ 3.
Let p1 , . . . , pn be prime ideals of R such that dim R/pi = 3 for all i =
Tn1, . . . , n. Then
d
j
Ha (R) = 0 and Supp(Ha (R)) is finite, for j = d − 1, d − 2, where a := i=1 pi .
Proof. It follows easily from Lichtenbaum-Hartshorne and Grothendieck’s vanishing theorems that Had (R) = 0, Supp(Had−1 (R)) ⊆ {m} and for each q ∈ Supp(Had−2 (R)) we have
dim R/q ≤ 1. Therefore dim Supp(Had−2 (R)) ≤ 1, and so
Supp(Had−2 (R)) ⊆ AsshR Had−2 (R) ∪ {m}.
Now it follows from Lemma 2.1 that Supp(Had−2 (R)) is a finite set, as required.
Corollary 2.3. Let (R, m) and a be as in Lemma 2.2. Then ExtiR (R/a, Haj (R)) is a
weakly Laskerian R-module for all i ≥ 0 and for j = d − 1, d − 2.
Proof. The result follows easily from Lemma 2.2.
The next lemma was proved by Melkersson for a-cofiniteness. The proof given in [22,
Proposition 3.11] can be easily carried for weakly Laskerian modules.
Lemma 2.4. Let R be a Noetherian ring, a an ideal of R, and M an R-module such that
ExtiR (R/a, M) is a weakly Laskerian R-module for all i. If t is a non-negative integer
such that the R-module ExtiR (R/a, Haj (M)) is weakly Laskerian, for all i and all j 6= t,
then this is the case also when j = t.
Proposition 2.5. Let (R, m) be a regular local ring containing a field with dim R = d ≥ 3.
Let pT
1 , . . . , pn be prime ideals of R such that dim R/pi = 3, for all i = 1, . . . , n, and let
a := ni=1 pi . Then ExtiR (R/a, Haj (R)) is a weakly Laskerian R-module for all i, j ≥ 0.
Proof. If Hai (R) 6= 0, then it follows from Lemma 2.2 and [5, Theorems 6.1.2 and 6.2.7]
that i ∈ {d −3, d −2, d −1}. Whence, the assertion follows from Corollary 2.3 and Lemma
2.4.
The next result was proved by Kawasaki for finitely generated modules. The proof
given in [17, Lemma 1] can be easily carried for weakly Laskerian modules.
Lemma 2.6. Let R be a Noetherian ring, T an R-module, and a an ideal of R. Then the
following conditions are equivalent:
(i) ExtnR (R/a, T ) is weakly Laskerian for all n ≥ 0.
(ii) For any finitely generated R-module N with support in V (a), ExtnR (N, T ) is weakly
Laskerian for all n ≥ 0.
We now are prepared to prove the main theorem of this section, which shows that when
R is a regular local ring contains a field and a an ideal of R such that dim R/a ≤ 3, then
all homomorphic images of the R-modules ExtjR (R/a, Hai (R)) have only finitely many
associated primes.
Theorem 2.7. Let (R, m) be a regular local ring containing a field with dim R = d ≥ 3,
and let a be an ideal of R such that dim R/a ≤ 3. Then the R-module ExtjR (R/a, Hai (R))
is weakly Laskerian, for all i, j ≥ 0.
ON THE FINITENESS PROPERTIES OF LOCAL COHOMOLOGY MODULES
5
Proof. In view of [4, Corollaries 2.7 and 3.2], we may assume that dim R/a = 3. Then
we have grade a = d − 3. Hence, by virtue of [5, Theorems 6.1.2 and 6.2.7], Hai (R) = 0,
whenever i 6∈ {d−3, d−2, d−1}. Moreover, by [23, Corollary 3.3], the set Supp(Had−1 (R))
is finite, in view of Proposition 2.5, it is enough for us to show that the R-module
ExtjR (R/a, Had−3(R)) is weakly Laskerian. To this end, let
X := {p ∈ Min(a) | height p = d − 3}
and
Y := {p ∈ Min(a) | height p ≥ d − 2}.
T
Then X 6= ∅. Set b = p∈X p. If Y = ∅, then Rad(a) = b, and so the assertion follows
T
from Proposition 2.5. Therefore we may assume that Y 6= ∅. Let c = p∈Y p. Then
Rad(a) = b ∩ c and height c ≥ d − 2. Next, we show that height (b + c) ≥ d − 1. Suppose
the contrary is true. Then there exists a prime ideal q of R such that b + c ⊆ q and
height q = d − 2. Therefore there exist p1 ∈ X and p2 ∈ Y such that p1 + p2 ⊆ q. Since
height p2 ≥ d − 2, it follows that q = p2 , and so p1 & p2 , which is a contradiction (note
that p1 , p2 ∈ Min(a)).
Consequently, grade (b + c) ≥ d − 1 and grade c ≥ d − 2. Therefore, it follows from [5,
Theorem 6.2.7], that
d−3
d−2
Hb+c
(R) = 0 = Hb+c
(R)
and
Hcd−3 (R) = 0.
It now follows from Rad(a) = b ∩ c and the Mayer-Vietoris sequence (see [5, Theorem 3.2.3]), that Had−3 (R) ∼
= Hbd−3 (R), and so by Proposition 2.5, the R-module
j
ExtR (R/b, Had−3(R)) is weakly Laskerian, for all j ≥ 0.
On the other hand, since dim R/(b + c) ≤ 1, it is easy to see that
V (b + c) = AsshR R/(b + c) ∪ {m}.
Now, as
Supp(ExtiR (R/c, Hbd−3(R))) ⊆ V (b + c),
it follows that ExtiR (R/c, Hbd−3(R)) is a weakly Laskerian R-module, for all i ≥ 0. Also,
as
Supp(ExtiR (R/b + c, Hbd−3 (R))) ⊆ V (b + c),
analogous to the preceding, we see that the R-module
ExtiR (R/b + c, Hbd−3(R))
is also weakly Laskerian for all i ≥ 0. Now, the exact sequence
0 −→ R/ Rad(a) −→ R/b ⊕ R/c −→ R/b + c −→ 0,
induces the long exact sequence
· · · −→ ExtiR (R/b, Had−3 (R)) ⊕ ExtiR (R/c, Had−3(R)) −→ ExtiR (R/ Rad(a), Had−3 (R))
−→ ExtiR (R/b + c, Had−3(R)) −→ · · · ,
which shows that the R-module ExtiR (R/ Rad(a), Had−3 (R)) is weakly Laskerian, for all
i ≥ 0, and so it follows from Lemma 2.6 that, the R-module ExtiR (R/a, Had−3(R)) is
weakly Laskerian, for all i ≥ 0, as required.
6
M. SEDGHI, K. BAHMANPOUR AND R. NAGHIPOUR
Corollary 2.8. Let (R, m) be a regular local ring of dimension d ≥ 3 containing a field,
and let a be an ideal of R such that dim R/a ≤ 3. Then, for any finitely generated Rmodule N with Supp(N) ⊆ V (a), the R-module ExtnR (N, Hai (R)) is weakly Laskerian, for
all integers n, i ≥ 0.
Proof. The result follows from Theorem 2.7 and Lemma 2.6.
Corollary 2.9. Let (R, m) be a regular local ring of dimension d ≤ 4 containing a field,
and a an ideal of R. Then, for any finitely generated R-module N with Supp(N) ⊆ V (a),
the R-module ExtnR (N, Hai (R)) is weakly Laskerian, for all integers n, i ≥ 0.
Proof. Since R is regular local, so dim R/a = 4 if and only if a = 0. Thus, the assertion is
clear in the case of dim R/a = 4. Moreover, the case dim R/a = 3 follows from Corollary
2.8. Also, the case dim R/a = 2 follows from [4, Theorem 3.1] and Lemma 2.4. Finally, if
dim R/a ≤ 1, then the assertion follows from [4, Theorem 2.6] and [17, Lemma 1].
3. Finiteness of local cohomology modules for regular local rings of
small dimension
It will be shown in this section that the subjects of the previous section can be used to
prove the finiteness of local cohomology modules for a regular local ring R with dim R = 5.
The main result is Theorem 3.4. The following proposition will serve to shorten the proof
of that theorem. The following easy lemma will be used in Proposition 3.2.
f
g
Lemma 3.1. Let R be a Noetherian ring and let M ′ −→ M −→ M ′′ be an exact sequence
of R-modules such that M ′ is weakly Laskerian and M ′′ has only finitely many associated
primes. Then M has only finitely many associated primes.
Proof. The assertion follows from the exact sequence
0 −→ M ′ /Im g −→ M −→ Im g −→ 0,
by applying [21, Theorem 6.3].
Proposition 3.2. Let (R, m) be a d-dimensional regular local ring, and a an ideal of
R such that height a = 1. Then the set AssR ExtiR (N, Ha1 (R)) is finite, for each finitely
generated R-module N with support in V (a) and for all integers i ≥ 0.
Proof. Let
X := {p ∈ Min(a) | height p = 1} and Y := {p ∈ Min(a) | height p ≥ 2}.
T
Then X 6= ∅. Let b = p∈X p. Since R is a UFD, it follows from [21, Exercise 20.3] that
b is a principal ideal, and so there is an element x ∈ R such that b = Rx. Hence, if
Y = ∅, then Rad(a) = Rx, and so it follows from [18, Theorem 1] that, the R-module
ExtiR (N, Ha1(R)) is finitely generated. Thus the set AssR ExtiR (N, Ha1 (R)) is T
finite.
Therefore, we may assume that Y 6= ∅. Then Rad(a) = Rx ∩ c, where c = p∈Y p. It is
easy to see that height c ≥ 2 and height (b + c) ≥ 3. Whence, by using the Mayer-Vietoris
sequence it yields that Ha1 (R) ∼
= Hb1 (R). Therefore, by [18, Theorem 1], the R-module
1
Ha (R) is b-cofinite.
ON THE FINITENESS PROPERTIES OF LOCAL COHOMOLOGY MODULES
7
On the other hand, according to Artin-Rees lemma, there exists a positive integer n
such that
bn N ∩ Γb (N) = 0 = cn N ∩ Γc (N).
We claim that bn N ∩ cn N = 0. To this end, suppose that q ∈ AssR (bn N ∩ cn N). Then
q = (0 :R y), for some y(6= 0) ∈ bn N ∩ cn N. As Supp(N) ⊆ V (a), it follows that a ⊆ q,
and so b ∩ c ⊆ q. Thus b ⊆ q or c ⊆ q, and so y ∈ Γb (N) or y ∈ Γc (N). Furthermore,
since y ∈ bn N ∩ cn N, it follows that y = 0, which is a contradiction.
Now, since bn N ∩ cn N = 0, the exact sequence
0 −→ N −→ N/bn N ⊕ N/cn N −→ N/(bn + cn )N −→ 0,
induces the long exact sequence
· · · −→ ExtiR (N/bn N, Ha1(R)) ⊕ ExtiR (N/cn N, Ha1 (R)) −→ ExtiR (N, Ha1 (R))
n
n
1
−→ Exti+1
R (N/(b + c )N, Ha (R)) −→ · · · .
Since Ha1 (R) is b-cofinite and
(†)
Supp(N/(bn + cn )N) ⊆ Supp(N/bn N) ⊆ V (b),
it follows from [17, Lemma 1] that the R-modules
ExtiR (N/bn N, Ha1 (R)) and ExtiR (N/(bn + cn )N, Ha1(R))
are finitely generated for all i ≥ 0.
Next, let T := N/cn N and we show that ExtiR (T, Ha1 (R)) is also b-cofinite for all i ≥ 0.
To do this, since Supp(Γb (T )) ⊆ V (b) it follows from [17, Lemma 1] that the R-module
ExtiR (Γb (T ), Ha1(R)) is finitely generated for each i ≥ 0. Hence it is enough to show that
the R-module ExtiR (T /Γb (T ), Ha1(R)) is finitely generated. As T /Γb (T ) is b-torsion-free,
we therefore make the additional assumption that T is a b-torsion-free R-module. Then
in view of [5, Lemma 2.11], x is a non-zerodivisor on T , and so the exact sequence
x
0 −→ T −→ T −→ T /xT −→ 0
induces the long exact sequence
x
· · · −→ ExtiR (T /xT, Ha1 (R)) −→ ExtiR (T, Ha1 (R)) −→ ExtiR (T, Ha1 (R))
1
−→ Exti+1
R (T /xT, Ha (R)) −→ · · · .
(††)
Since Supp(T /xT ) ⊆ V (b) it follows from [17, Lemma 1] that ExtiR (T /xT, Ha1 (R)) is a
finitely generated R-module, for all i ≥ 0, (note that b = Rx). Consequently, it follows
from the exact sequence (††) that the R-modules
(0 :ExtiR (T,Ha1 (R)) x) and ExtiR (T, Ha1(R))/xExtiR (T, Ha1 (R))
are finitely generated and hence b-cofinite. Therefore it follows from Melkersson’s result
[22, Corollary 3.4] that the R-module ExtiR (T, Ha1 (R)) is b-cofinite for all i ≥ 0.
Now, let Ω := ExtiR (N/bn N, Ha1 (R)) ⊕ ExtiR (N/cn N, Ha1 (R)). Then for every finitely
generated submodule L of Ω the R-module Ω/L is also b-cofinite, and so the set AssR Ω/L
is finite. Now, it follows from the exact sequence (†) and Lemma 3.1 that the set
8
M. SEDGHI, K. BAHMANPOUR AND R. NAGHIPOUR
AssR ExtiR (N, Ha1 (R)) is finite, as required.
The next lemma was proved in [1] in the case R is local. The proof given in [1, Lemma
2.5] can be easily carried over Noetherian rings, so that we omit the proof.
Lemma 3.3. Let R be a Noetherian ring, x an element of R and a an ideal of R such
that a ⊆ Rad(Rx). Then, for any finitely generated R-module M, the R-homomorphism
x
Haj (M) −→ Haj (M) is an isomorphism, for each j ≥ 2.
We are now ready to state and prove the main theorem of this section.
Theorem 3.4. Let (R, m) be a five-dimensional regular local ring containing a field and
a an ideal of R. Then the set AssR ExtnR (N, Hai (R)) is finite, for each finitely generated
R-module N with support in V (a) and for all integers n, i ≥ 0.
Proof. Since dim R/a = 5 if and only if a = 0, the assertion is clear in this case. Hence we
consider the case when dim R/a ≤ 4. If dim R/a ≤ 3, then the result follows from the proof
of Corollary 2.8. Therefore, we may assume that dim R/a = 4. Then height(a) = 1 and in
view of the Lichtenbaum-Hartshorne vanishing theorem Ha5 (R) = 0. Whence Hai (R) = 0
whenever i 6∈ {1, 2, 3, 4}. Also, since by [23, Corollary 3.3], the set Supp(Ha4 (R)) is finite,
it follows from
Supp(ExtnR (N, Ha4 (R))) ⊆ Supp(Ha4 (R))
that the set AssR ExtnR (N, Ha4 (R)) is also finite. Consequently, in view of Proposition 3.2,
we may consider the cases i = 2, 3.
Case 1. i = 2.
Suppose that
X := {p ∈ Min(a) | height p = 1} and Y := {p ∈ Min(a) | height p ≥ 2}.
T
T
Then X 6= ∅ and Y 6= ∅. Let b = p∈X p and c = p∈Y p. Since R is a UFD, it follows
from [21, Ex. 20.3] that b is a principal ideal, and so there is an element x ∈ R such
that b = Rx. Moreover, Rad(a) = Rx ∩ c, height c ≥ 2, and in view of Proposition 3.2,
1
1
∼ 1
module. As
we have
S Ha (R) = Hb (R). Thus, by [18, Theorem 1], Ha (R) is a b-cofinite
S
c 6⊆ p∈X p, it follows that there is an element y ∈ c such that y 6∈ p∈X p. Now, in view
of [24, Corollary 3.5], there exists the exact sequence
1
1
2
0
2
0 −→ HRy
(HRx
(R)) −→ HRx+Ry
(R) −→ HRy
(HRx
(R)) −→ 0,
2
1
1
2
and so HRx+Ry
(R) ∼
(HRx
(R)), (note that HRx
(R) = 0).
= HRy
Using again [24, Corollary 3.5], to show that there exists the exact sequence
1
2
0
0 −→ HRy
(Ha1 (R)) −→ Ha+Ry
(R) −→ HRy
(Ha2 (R)) −→ 0,
(†)
and so it follows from Ha1 (R) ∼
= Hb1 (R) that
H 1 (H 1 (R)) ∼
(R).
= H 1 (H 1(R)) = H 1 (H 1 (R)) ∼
= H2
Ry
a
Ry
b
Ry
Rx
Also, since xy ∈ Rad(a), it follows that the R-module
using Lemma 3.3, it is easy to see that the R-module
Rx+Ry
Ha2 (R)
Ha2 (R)
is R(yx)-torsion. Hence
is Ry-torsion. That is
ON THE FINITENESS PROPERTIES OF LOCAL COHOMOLOGY MODULES
9
0
HRy
(Ha2 (R)) = Ha2 (R). Consequently, from the exact sequence (†) we get the following
exact sequence,
2
2
0 −→ HRx+Ry
(R) −→ Ha+Ry
(R) −→ Ha2 (R) −→ 0.
(††)
Furthermore, in view of Proposition 3.2, there exists a positive integer n such that the
sequence
0 −→ N −→ N/bn N ⊕ N/cn N −→ N/(bn + cn )N −→ 0
is exact, and so we obtain the long exact sequence
· · · −→ ExtiR (N/bn N, Ha2(R)) ⊕ ExtiR (N/cn N, Ha2 (R)) −→ ExtiR (N, Ha2 (R))
n
n
2
−→ Exti+1
R (N/(b + c )N, Ha (R)) −→ · · · .
(†††)
Since
Supp(N/(bn + cn )N) ⊆ Supp(N/bn N) ⊆ V (b),
it follows from Lemma 3.3 that
ExtjR (N/bn N, Ha2 (R)) = 0 = ExtjR (N/(bn + cn )N, Ha2 (R)),
for all j ≥ 0. Whence, the exact sequence (†††) implies that
j
Extj (N, H 2(R)) ∼
= Ext (N/cn N, H 2(R)).
R
a
R
a
Next, it is easy to see that height(Rx + Ry) = 2, and so height(a + Ry) = 2. Thus
dim R/(a + Ry) = 3, and hence as
Supp(N/cn N) ⊆ V (c) ⊆ V (a + Ry),
2
it follows from Corollary 2.8 that the R-module ExtjR (N/cn N, Ha+Ry
(R)) is weakly Laskerian, for all j ≥ 0.
Also, as grade(Rx + Ry) = 2, it follows from [5, Theorem 3.3.1] and [22, Proposition
2
3.11] that, the R-module HRx+Ry
(R) is Rx+Ry-cofinite. Now, by modifying the argument
2
of the proof of Proposition 3.2, one can see that the R-module ExtjR (N/cn N, HRx+Ry
(R))
j
n
2
is Rx-cofinite for all j ≥ 0. In particular, AssR ExtR (N/c N, HRx+Ry (R)) is a finite set,
for all j ≥ 0. Moreover, from the exact sequence (††), we deduce the long exact sequence
2
· · · −→ ExtjR (N/cn N, Ha+Ry
(R)) −→ ExtjR (N/cn N, Ha2 (R))
n
2
−→ Extj+1
R (N/c N, HRx+Ry (R)) −→ · · · .
Now using Lemma 3.1 and the above long exact sequence induced, it follows from the
isomorphism
j
ExtjR (N, Ha2(R)) ∼
= ExtR (N/cn N, Ha2(R)),
that the set AssR ExtjR (N, Ha2(R)) is finite.
Case 2. i = 3.
Let X 6= ∅, Y 6= ∅ and y be as in the case 1. Then using the same argument, it follows
from Lemma 3.3 that the R-modules Ha2 (R) and Ha3 (R) are Ry-torsion. Thus
0
1
(Ha3 (R)) = Ha3 (R).
HRy
(Ha2 (R)) = 0 and HRy
10
M. SEDGHI, K. BAHMANPOUR AND R. NAGHIPOUR
Therefore, using the exact sequence
1
3
0
0 −→ HRy
(Ha2 (R)) −→ Ha+Ry
(R) −→ HRy
(Ha3 (R)) −→ 0,
(†)
3
(see [24, Corollary 3.5]), we obtain that Ha3(R) ∼
(R).
= Ha+Ry
Moreover, it follows easily from the exact sequence
0 −→ N −→ N/bn N ⊕ N/cn N −→ N/(bn + cn )N −→ 0,
that
j
ExtjR (N, Ha3(R)) ∼
= ExtR (N/cn N, Ha3(R)),
3
for all j ≥ 0. Hence, for all j ≥ 0 the R-module ExtjR (N/cn N, Ha+Ry
(R)) is weakly
j
n
3
Laskerian, and so ExtR (N/c N, Ha (R)) is also a weakly Laskerian R-module. Therefore
ExtjR (N, Ha3(R)) is a weakly Laskerian R-module, and hence it has finitely many associated
primes, as required.
Theorem 3.5. Let (R, m) be a regular local ring containing a field such that dim R ≤ 5,
and let a be an ideal of R. Then, for all integers n, i ≥ 0 and any finitely generated
R-module N with support in V (a), the set AssR ExtnR (N, Hai (R)) is finite.
Proof. The assertion follows from Corollary 2.8 and Theorem 3.4.
The final theorem of this section shows that if R is a regular local ring with dim R ≤ 4,
then the R-module Hai (M) has finitely many associated primes, for all i ≥ 0 and for any
finitely generated M over R. Recall that, an R-module L is called an a-cominimax module
[2] if Supp(L) ⊆ V (a) and ExtiR (R/a, L) is minimax, for all i ≥ 0.
Theorem 3.6. Let (R, m) be a regular local ring such that dim R ≤ 4. Suppose that a
is an ideal of R and M a finitely generated R-module. Then for all integers i, j ≥ 0, the
R-module Extj (R/a, Hai (M)) is weakly Laskerian.
Proof. If dim R ≤ 3, then by virtue of [1, Theorem 2.12], the R-modules Hai (M) are acominimax, and so the R-module Extj (R/a, Hai (M)) is weakly Laskerian. Hence we may
assume that dim R = 4.
Now if dim R/a = 4, then a = 0 and so the result holds. Also, case dim R/a ≤ 2
follows from [4, Corollary 3.2]. Hence we may assume that dim R/a = 3. Then we have
heigth(a) = 1.
On the other hand, if dim M ≤ 2 then in view of [4, Corollary 2.7] and [22, Proposition
5.1], the R-module Hai (M) is a-cofinite, and so the result is clear. Also, in the case
of dim M = 3, the assertion follows from [22, Proposition 5.1], [23, Corollary 3.3], the
Grothendieck Vanishing Theorem and Lemma 2.4.
Therefore we may assume that dim M = 4. Then in view of [22, Proposition 5.1], [23,
Corollary 3.3], the Grothendieck Vanishing Theorem and Lemma 2.4, it is enough to show
that the R-module Extj (R/a, Ha2(M)) is weakly Laskerian, for all j ≥ 0. To this end, let
X := {p ∈ Min(a) | height p = 1} and Y := {p ∈ Min(a) | height p ≥ 2}.
ON THE FINITENESS PROPERTIES OF LOCAL COHOMOLOGY MODULES
11
T
Then X 6= ∅. Let b = p∈X p. If Y = ∅ then Rad(a) = b, and so Ha2 (M) = 0, as required.
Therefore we may assume that Y 6= ∅. ThenTRad(a) = b ∩ c, height c ≥ 2, and in view of
Theorem 2.7, height (b + c) ≥ 3, where c = p∈Y p.
Now, since Rad(a) = b ∩ c, in view of the Mayer-Vietoris sequence the sequence
f
g
2
3
Hb+c
(M) −→ Hc2 (M) −→ Ha2 (M) −→ Hb+c
(M)
(†)
is exact.
Since dim R/(b + c) ≤ 1 we have V (b + c) = Min(b + c), and so as
i
Supp(Hb+c
(M)) ⊆ V (b + c),
i
it follows that the set Supp(Hb+c
(M)) is finite, for i = 2, 3. Therefore, it follows from the
exact sequence (†) that the R-modules Ker f and Im g are weakly Laskerian. Thus, as
dim R/c ≤ 2 it follows from [4, Corollary 3.3] that the R-module ExtiR (R/c, Hc2 (M)/Ker f )
is weakly Laskerian for all i ≥ 0. Now, the exact sequence
0 −→ Hc2 (M)/Ker f −→ Hc2 (M) −→ Im g −→ 0,
induces the long exact sequence
ExtiR (R/c, Hc2 (M)/Ker f ) −→ ExtiR (R/c, Hc2(M)) −→ ExtiR (R/c, Im g),
for all i ≥ 0. Consequently, in view of Lemma 2.12, ExtiR (R/c, Hc2(M)) is a weakly
Laskerian R-module.
Finally, by using the exact sequence
0 −→ R/ Rad(a) −→ R/b ⊕ R/c −→ R/b + c −→ 0,
we get the long exact sequence
· · · −→ ExtiR (R/b + c, Ha2 (M)) −→ ExtiR (R/b ⊕ R/c, Ha2(M))
2
−→ ExtiR (R/ Rad(a), Ha2(M)) −→ Exti+1
R (R/b + c, Ha (M)) −→ · · · .
Now, by applying Lemma 2.10, we obtain that
ExtiR (R/b + c, Ha2(M)) = 0 = ExtiR (R/b, Ha2(M)),
for each i ≥ 0, and therefore
ExtiR (R/ Rad(a), Ha2(M)) ∼
= ExtiR (R/c, Ha2 (M)),
for each i ≥ 0.
Hence, the R-module ExtiR (R/ Rad(a), Ha2 (M)) is weakly Laskerian, for each i ≥ 0.
Thus, it follows from Lemma 2.6 that, the R-module ExtiR (R/a, Ha2(M)) is also weakly
Laskerian, for all i ≥ 0, and this completes the proof.
We end this section with a result which is a generalization of Corollary 2.8.
Corollary 3.7. Let the situation be as in Theorem 3.6. Then for any finitely generated
R-module N with support in V (a), the R-module ExtjR (N, Hai (M)) is weakly Laskerian, for
all integers i, j ≥ 0. In particular the set AssR ExtjR (N, Hai (M)) is finite, for all integers
i, j ≥ 0.
Proof. The assertion follows from Theorem 3.6 and Lemma 2.6.
12
M. SEDGHI, K. BAHMANPOUR AND R. NAGHIPOUR
Acknowledgments
The authors are deeply grateful to the referee for a very careful reading of the manuscript
and many valuable suggestions in improving the quality of the paper. We also would like
to thank from the Azarbaijan Shahid Madani University for the financial support (No.
217/d/917).
References
[1] N. Abazari and K. Bahmanpour, On the finiteness of Bass numbers of local cohomology modules, J.
Alg. Appl. 10 (2011), 783-791.
[2] J. A’zami, R. Naghipour and B. Vakili, Finiteness properties of local cohomology modules for aminimax modules, Proc. Amer. Math. Soc. 137 (2009), 439-448.
[3] K. Bahmanpour and R. Naghipour, Associated primes of local cohomology modules and Matlis duality, J. Algebra. 320 (2008), 2632-2641.
[4] K. Bahmanpour and R. Naghipour, Cofiniteness of local cohomology modules for ideals of small
dimension, J. Algebra. 321 (2009), 1997-2011.
[5] M.P. Brodmann and R.Y. Sharp, Local cohomology; an algebraic introduction with geometric applications, Cambridge University Press, Cambridge, 1998.
[6] K. Divaani-Aazar and A. Mafi, Associated primes of local cohomology modules, Proc. Amer. Math.
Soc. 133 (2005), 655-660.
[7] A. Grothendieck, Local cohomology, Notes by R. Hartshorne, Lecture Notes in Math. 862, Springer,
New York, 1966.
[8] A. Grothendieck, Cohomologie local des faisceaux coherents et théorémes de lefschetz locaux et
globaux (SGA2), North-Holland, Amsterdam, 1968.
[9] R. Hartshorne, Affine duality and cofiniteness, Invent. Math. 9 (1970), 145-164.
[10] M. Hellus, On the associated primes of a local cohomology module, J. Algebra. 237 (2001), 406419.
[11] C. Huneke, Problems on local cohomology, Free resolutions in commutative algebra and algebraic
geometry, Res. Notes Math. 2(1992), 93-108.
[12] C. Huneke and R.Y. Sharp, Bass numbers of local cohomology module, Trans. Amer. Math. Soc. 339
(1993), 765-779.
[13] G. Lyubeznik, Finiteness properties of local cohomology modules (an application of D-modules to
commutative algebra), Invent. Math. 113 (1993), 41-55.
[14] G. Lyubeznik, Finiteness properties of local cohomology modules for regular local rings of mixed
chartacteristic: the unramified case, Comm. Algebra, 28 (2000), 5867-5882.
[15] G. Lyubeznik, A partial survey of local cohomology, local cohomology and its applications, Lectures
Notes in Pure and Appl. Math. 226 (2002), 121-154.
[16] M. Katzman, An example of an infinite set of associated primes of a local cohomology module, J.
Algebra 252 (2002), 161-166.
[17] K.I. Kawasaki, On the finiteness of Bass numbers of local cohomology modules, Proc. Amer. Math.
Soc. 124 (1996), 3275-3279.
[18] K.I. Kawasaki, Cofiniteness of local cohomology modules for principal ideals, Bull. London Math.
Soc. 30 (1998), 241-246.
[19] T. Marley, The associated primes of local cohomology modules over rings of small dimension,
Manuscripta Math. 104 (2001), 519-525.
[20] T. Marley and J.C. Vassilev, Cofiniteness and associated primes of local cohomology modules, J.
Algebra 256 (2002), 180-193.
[21] H. Matsumura, Commutative ring theory, Cambridge Univ. Press, Cambridge, UK, 1986.
[22] L. Melkersson, Modules cofinite with respect to an ideal, J. Algebra, 285 (2005), 649-668.
[23] R. Naghipour and M. Sedghi, A characterization of Cohen-Macaulay modules and local cohomology,
Arch. Math. 87 (2006), 303-308.
ON THE FINITENESS PROPERTIES OF LOCAL COHOMOLOGY MODULES
13
[24] P. Schenzel, Proregular sequences, local cohomology and completion, Math. Scand. 92 (2003), 161180.
[25] A.K. Singh, P-torsion elements in local cohomology modules, Math. Res. Lett. 7 (2000), 165-176.
[26] H. Zöschinger, Minimax modules, J. Algebra, 102 (1986), 1-32.
[27] H. Zöschinger, Ü ber die maximalbedingung für radikalvolle untermoduln, Hokkaido Math. J. 17
(1988), 101-116.
Department of Mathematics, Azarbaijan Shahid Madani University, 53714-161, Tabriz,
Iran.
E-mail address: m [email protected]
E-mail address: [email protected]
Department of Mathematics, Faculty of Mathematical Sciences, University of Mohaghegh Ardabili, 56199-11367, Ardabil, Iran.
E-mail address: [email protected]
Department of Mathematics, University of Tabriz, 51666-16471, Tabriz, Iran.
E-mail address: [email protected]
E-mail address: [email protected]
| 0math.AC
|
Derivative coordinates for analytic tree fractals
and fractal engineering
Henk Mulder
[email protected]
Abstract
We introduce an alternative coordinate system based on derivative polar and spherical
coordinate functions and construct a root-to-canopy analytic formulation for tree fractals. We
develop smooth tree fractals and demonstrate the equivalence of their canopies with iterative
straight lined tree fractals. We then consider implementation and application of the analytic
formulation from a computational perspective. Finally we formulate the basis for
concatenation and composition of fractal trees as a basis for fractal engineering of which we
provide some examples.
1.
Introduction
Tree fractals are objects of interest and fascination but surprisingly, we have seen only limited application
of tree fractals as foundations for systems and engineering. One notable exception are L-systems as a model
for botanical structures such as plants and trees [1]. The preponderance of such fractals in nature [2] would
suggest that as scientists and engineers we are missing an opportunity for efficient system design.
One of the limitations of tree fractals is in their construction. In one form or another, they are mostly
iteratively constructed, i.e. the systematic branching and stacking of geometric primitives such as straight
line segments [3]. Such composed structures hamper our ability to analyze them effectively. What if we
had an analytical function that represented an entire tree fractal?
This question has motived the research presented here and has resulted in exactly such a function. In
this paper we present an analytic formulation of tree fractals and a foundation of derivative coordinates that
underpins them. We then use that formulation to derive some initial results including a formal description
of fractal engineering.
1
2.
Derivative coordinate functions
Theorem 1 Let a parametric path 𝑝(𝑠) for 𝑝: ℝ ↦ ℝ2 be defined by a pair of Cartesian coordinate
functions (𝑥(𝑠), 𝑦(𝑠)) for 𝑥, 𝑦: ℝ ↦ ℝ then 𝑝 may be expressed as a function of a pair of derivative
𝜕𝑟 𝜕𝜑
coordinate functions ( , ) for 𝑟, 𝜑: ℝ ↦ ℝ , that represent the relative polar coordinates at a point
𝜕𝑠 𝜕𝑠
(𝑥, 𝑦) on 𝑝. The path 𝑝 is then given by:
𝜕𝑟
𝜕𝜑
∫ cos (∫ 𝜕𝑠 𝑑𝑠) 𝑑𝑠
𝑥(𝑠)
𝑝(𝑠) = (
) = ( 𝜕𝑠
)
𝜕𝑟
𝜕𝜑
𝑦(𝑠)
∫ sin (∫ 𝑑𝑠) 𝑑𝑠
𝜕𝑠
𝜕𝑟 𝜕𝜑
and ( ,
𝜕𝑠 𝜕𝑠
(1)
𝜕𝑠
) are defined as
2
𝜕𝑟
𝜕𝑠
(𝜕𝜑
)=(
𝜕𝑠
√(𝜕𝑥) + (𝜕𝑦)
𝜕𝑠
𝜕𝑠
𝜕 tan−1
𝜕𝑦 𝜕𝑥
⁄
𝜕𝑠 𝜕𝑠
2
).
(2)
𝜕𝑠
Notation. In this paper we will see a lot of partial derivatives and to facilitate in line equations, from here
on we will represent these with the del operator ∇ and we remind ourselves that unless otherwise indicated,
𝜕𝑟
these are all 1-dimensional derivatives with respect to s, i.e. ∇r = where 𝑟: ℝ ↦ ℝ. So (1) becomes
𝜕𝑠
𝑝(𝑠) = (
𝑥(𝑠)
∫ ∇r cos(∫ ∇φ 𝑑𝑠) 𝑑𝑠
)=(
)
𝑦(𝑠)
∫ ∇r sin(∫ ∇φ 𝑑𝑠) 𝑑𝑠
(3)
and (2)
∇r
√(∇x)2 + (∇y)2
( )=(
)
∇φ
∇ tan−1 ∇y/∇x
(4)
We will refer to ∇r as the radial derivative coordinate and to ∇φ as the angular derivative coordinate. We
may refer to the derivative coordinates functions (∇r, ∇φ) simply as derivative coordinates, suggesting
derivative coordinate values for a specific value of the parametric variable 𝑠.
We will favor the Euler form of (3) for most of this paper,
𝑝(𝑠) = 𝑥(𝑠) + iy(s) = ∫ ∇r exp(∫ i∇φ 𝑑𝑠) 𝑑𝑠
(5)
but when developing the n-dimensional forms where n > 2, we will revert to the trigonometric form.
2
Proof. Substitute (4) into (3), noting that the integrals cancel out the derivatives with respect to the same
𝑥(𝑠)
𝑥(𝑠)
variable s, we find that (
)=(
).
𝑦(𝑠)
𝑦(𝑠)
The derivative coordinates (∇r, ∇φ) are purposefully written as derivatives since geometrically they
describe how a path evolves with respect to a point on the path, i.e. the relative radial and angular
evolution. For that reason we do not evaluate the integral ∫ ∇φ 𝑑𝑠 = φ to remind ourselves that ∇φ is our
coordinate, and not φ.
However, there are two cases when these integrals are interesting in their own right.
Corollary 1.1 Let a parametric path 𝑝(𝑠) be represented by derivative coordinates functions (𝛻𝑟, 𝛻𝜑) and
the path equation ∫ 𝛻𝑟 𝑒𝑥𝑝(∫ 𝑖𝛻𝜑 𝑑𝑠) 𝑑𝑠, then the path length 𝑙 between two points at 𝑠1 and 𝑠2 is
𝑠
𝑙 = ∫𝑠 2 ∇r 𝑑𝑠.
(6)
1
Proof. The derivative coordinate ∇r in (4) is defined as ∇r = √(∇x)2 + (∇y)2 , i.e. the differential arc
𝑠
length of 𝑝 at point (𝑥(𝑠), 𝑦(𝑠)), it follows that the arc length from 𝑠1 and 𝑠2 is ∫𝑠 2 ∇r 𝑑𝑠.
1
Corollary 1.2 Let a parametric path 𝑝(𝑠) be represented by derivative coordinates functions (𝛻𝑟, 𝛻𝜑) and
the path equation ∫ 𝛻𝑟 𝑒𝑥𝑝(∫ 𝑖𝛻𝜑 𝑑𝑠) 𝑑𝑠, then the absolute angle ∆𝛷 between the tangent lines at two
points at 𝑠1 and 𝑠2 is
𝑠
∆Φ = ∫𝑠 2 ∇φ 𝑑𝑠
(7)
1
𝜕𝑦
𝜕𝑥
Proof. From the definition (2) we note that ∫ ∇φ 𝑑𝑠 = tan−1 / . By eliminating s we find the anti𝜕𝑠 𝜕𝑠
derivative Φ = tan−1
follows that
𝑠2 .
𝜕𝑦
𝜕𝑥
𝑠2
∫𝑠 ∇φ 𝑑𝑠 =
1
which is the angle of the tangent at 𝑠 on the path 𝑝 with respect to x-axis. It
Φ2 − Φ1 , i.e. the absolute angle ∆Φ between the tangent lines at points 𝑠1 and
We will extend the path formulation of Theorem 1 to the third dimension.
Corollary 1.3. Let a parametric path 𝑝(𝑠) for 𝑝: ℝ ↦ ℝ3 be defined by the Cartesian coordinate functions
(𝑥(𝑠), 𝑦(𝑠), 𝑧(𝑠)) for 𝑥, 𝑦, 𝑧: ℝ ↦ ℝ then 𝑝 may be expressed as a function of a triple of derivative
coordinate functions, (𝛻𝑟(𝑠), 𝛻𝜑(𝑠), 𝛻𝜓(𝑠)) for 𝑟, 𝜑, 𝜓: ℝ ↦ ℝ that represent the relative spherical
coordinates at a point (𝑥, 𝑦, 𝑧) on 𝑝. The path 𝑝 is then given by:
∫ ∇r cos(∫ ∇φ 𝑑𝑠) sin(∫ ∇𝜓 𝑑𝑠) 𝑑𝑠
𝑥(𝑠)
𝑝(𝑠) = (𝑦(𝑠)) = ( ∫ ∇r sin(∫ ∇φ 𝑑𝑠) sin(∫ ∇𝜓 𝑑𝑠) 𝑑𝑠 )
𝑧(𝑠)
∫ ∇r cos(∫ ∇𝜓 𝑑𝑠) 𝑑𝑠
and (𝛻𝑟(𝑠), 𝛻𝜑(𝑠), 𝛻𝜓(𝑠)) are defined as
3
(8)
√(∇𝑥)2 + (∇𝑦)2 + (∇𝑧)2
∇r
(∇φ) = (
)
∇tan−1 ∇𝑦/∇𝑥
∇𝜓
−1
2
2
2
∇cos ∇𝑧/√(∇𝑥) + (∇𝑦) + (∇𝑧)
(9)
Proof. (9) and (10) are the standard spherical coordinate system. Substitution of (9) into (10) verifies that
both spherical and Cartesian forms are equal.
More generally, we can extend this to any higher dimension.
Corollary 1.4. Let a parametric path 𝑝(𝑠), 𝑝: ℝ ↦ ℝ𝑛 for 𝑛 > 3 be defined by Cartesian coordinate
functions (𝑥𝑖 (𝑠)|4 ≤ 𝑖 ≤ 𝑛, 𝑖 ∈ ℕ) where 𝑥𝑖 : ℝ ↦ ℝ then 𝑝 may be expressed as a function of derivative
coordinate functions, (𝛻𝑣𝑖 |4 ≤ 𝑖 ≤ 𝑛, 𝑖 ∈ ℕ) where 𝑣𝑖 : ℝ ↦ ℝ represent the relative spherical coordinates
on an n-sphere at a point (𝑥𝑖 |4 ≤ 𝑖 ≤ 𝑛, 𝑖 ∈ ℕ) on 𝑝.
Proof. We will not give a formal proof but note that as long as the relative evolution of a path p(s), locally
within a hyper spherical coordinate system is unrestricted, then the integral of the relative evolution will
describe any path p(s) for p: ℝ ↦ ℝ𝑛
3.
Application of derivative coordinate functions in tree fractals
This research was initiated with the purpose of finding an analytic formulation of tree fractals. Before we
can demonstrate the use of derivative coordinate functions, we need to introduce a formulation for branches
in tree fractals.
If we accept, or define it to be as such, that √𝑥 2 has two solutions (𝑥, −𝑥) for 𝑥 ≠ 0 and one solution
(0) for 𝑥 = 0, then 𝑓(𝑥) = √𝑥 2 is a multivalued function with a node point at 𝑥 = 0 where 𝑓(𝑥) is single
valued. Taking this one step further, a function √sin2 (𝑥) is a multivalued function with node points at
𝑥 = 𝑛𝜋 for 𝑛 ∈ ℤ where 𝑓(𝑥) is single valued. When we integrate such a function, each time we get to a
node where f(x) is single valued, the function becomes dual valued again, i.e. it forks. This is repeated at
every node point and the repetitive forking creates a multivalued tree. This multivalued integral
∫ √sin2 (𝑥)𝑑𝑥 is shown below.
4
Image 2: multivalued integral of 𝑢̌(𝑠)
Image 1: multivalued integral of √sin2 (𝑥)
We could use such an analytic multivalued integral as a mechanism for tree branching however, we prefer
to define a function that is linear.
1, −1 𝑠 ∉ 𝐵
Definition 2. We define a multivalued periodic unit function 𝑢̌(𝑠) such that 𝑢̌(𝑠) = {
where 𝐵
0 𝑠∈𝐵
is a branch point set, 𝐵 = {𝑠𝑖 |∀𝑠𝑖 : 𝑖 ∈ ℕ, 𝑠 ∈ ℝ}.
(9)
The multivalued integral of the multivalued periodic unit function is shown in image 2.
Theorem 2. Using suitable derivative functional coordinates where the angular coordinate 𝛻𝜑 is
multiplied by the multivalued periodic unit function 𝑢̌(𝑠) with a branch point set 𝐵 =
{𝑠𝑖 |∀𝑠𝑖 : 𝑖 ∈ ℕ, 𝑠 ∈ ℝ}, then any symmetric binary tree is described by the function
𝑝̌ (𝑠) = ∫ ∇r exp(∫ i𝑢̌∇φ 𝑑𝑠) 𝑑𝑠
(10)
Notation. The “check” symbol over 𝑝̌ indicates that this is a tree function. It may be pronounced as p-tree.
Proof. We will demonstrate that the tree is symmetric, binary and that the formulation can represent any
tree.
At each branch node 𝑝̌ (𝑠𝑖 ) where 𝑠𝑖 ∈ 𝐵, the direction of the two emanating branches is determined
uniquely by the term 𝑢̌∇φ which has the multi-values ∇φ and -∇φ in between the branch points such that
the respective branch paths 𝑝𝑎 and 𝑝𝑏 are given by:
𝑝(𝑠) = ∫ ∇r exp(∫ i∇φ 𝑑𝑠) 𝑑𝑠
and
5
(11)
𝑝(𝑠) = ∫ ∇r exp(− ∫ i∇φ 𝑑𝑠) 𝑑𝑠
(12)
We note that the only difference between these paths 𝑝𝑎 and 𝑝𝑏 is in the direction of the rotation ∫ ∇φ 𝑑𝑠
and − ∫ ∇φ 𝑑𝑠 but not in the quantity. It follows that these two branches are symmetric. Since all branches
are symmetric from the branch point from which they emanate and that at each branch point there are
precisely two branches, it follows that the tree is symmetric and binary.
We note that at a branch point s𝑖 , the multivalued periodic unit function is single valued, i.e. 𝑢̌ = 0 and
we shall consider that case in the following corollary.
Corollary 2.1 A tree defined by Theorem 2, 𝑝̌ (𝑠) = ∫ 𝛻𝑟 𝑒𝑥𝑝(∫ 𝑖𝑢̌𝛻𝜑 𝑑𝑠) 𝑑𝑠 is continuous at every point of
its domain 𝑠 ∈ ℝ, if and only if the derivative coordinates functions (𝛻𝑟, 𝛻𝜑) are continuous on
𝑟(𝑠), 𝜑(𝑠): ℝ ↦ ℝ.
Proof. It is clear that the tree is continuous between consecutive tree nodes since at these intervals of the
domain the path function 𝑝(𝑠) reduces to the path function of Theorem 1, 𝑝(𝑠) = ∫ ∇r exp(∫ i∇φ 𝑑𝑠) 𝑑𝑠.
At the branch points 𝑠𝑖 , we consider the limits of the integral of the angular coordinates of the right branch
𝑝𝑎 from below and from above
lim𝑠↑𝑠𝑖 ∫ ∇φ𝑎 𝑑𝑠 = lim𝑠↓𝑠𝑖 ∫ ∇φ𝑎 𝑑𝑠
(13)
and for the left branch 𝑝𝑏
lims↑si ∫ ∇φ𝑏 ds = lims↓si ∫ −∇φ𝑏 ds
(14)
it follows that from (13) and (14)
𝑝𝑎 (𝑠𝑖 ) = 𝑝𝑏 (𝑠𝑖 )
at the branch points and thus a tree defined by Theorem 2 is continuous on its domain.
(15)
Corollary 2.2 A tree defined by Theorem 2, 𝑝̌ (𝑠) = ∫ 𝛻𝑟 𝑒𝑥𝑝(∫ 𝑖𝑢̌𝛻𝜑 𝑑𝑠) 𝑑𝑠 is analytic and continuous on
its entire domain 𝑠 ∈ ℝ if and only if the derivative coordinates functions (𝛻𝑟, 𝛻𝜑) are analytic and
continuous on 𝑟(𝑠), 𝜑(𝑠): ℝ ↦ ℝ.
Proof. Since by Corollary 2.1 a tree defined by Theorem 2, 𝑝̌ (𝑠) = ∫ ∇r exp(∫ i𝑢̌∇φ 𝑑𝑠) 𝑑𝑠 is continuous
and (∇r, ∇φ) are analytic for 𝑟(𝑠), 𝜑(𝑠): ℝ ↦ ℝ, we note that we may write
𝑝̌ (𝑠) = 𝑓(∇r, ∇φ)
(16)
where f itself is a composite analytic function, it follows that a tree defined by Theorem 2 is analytic and
continuous on 𝑟(𝑠), 𝜑(𝑠): ℝ ↦ ℝ.
6
Lemma 2.3. A symmetric binary tree is fully defined by any of its branch paths, from the root to any of its
branch tips.
Proof. Starting from root, after the first branch node, the binary branches are symmetric i.e. one defines the
other and vice versa. This is the case for any two branches at a branch point. It follows that a symmetric
binary tree fractal is fully defined by any of the paths that start at the root and end at a branch tip.
Notation. From here on, when referring to a tree, we are referring to a binary symmetric tree unless
otherwise stated.
Corollary 2.4 Theorem 2 may define any tree.
Proof. By Lemma 2.3 we note that a tree is defined by any of its branch paths. If we select the branch path
where 𝑢̌(𝑠) = 1 for all its nodes, then the path 𝑝 is given by:
𝑝(𝑠) = ∫ ∇r exp(∫ i∇φ 𝑑𝑠) 𝑑𝑠
(17)
By Theorem 1 such a path function may represent any path 𝑝, it follows that Theorem 2 may describe any
tree.
Corollary 2.5. Using suitable derivative functional coordinates (𝛻𝑣𝑖 |3 ≤ 𝑖 ≤ 𝑛, 𝑖 ∈ ℕ) where 𝑣𝑖 : ℝ ↦
ℝ where the directional coordinates 𝛻𝑣𝑖 are multiplied by the multivalued periodic unit function 𝑢̌𝑑 (𝑠)
with a given branch set 𝐵𝑑 = {𝑠𝑖 |∀𝑠𝑖 : 𝑖 ∈ ℕ, 𝑠 ∈ ℝ}𝑑 for dimension 𝑑 ≥ 3, then any tree may be described
by the higher dimensional path functions defined under Corollaries 1.3 and 1.4.
Proof. We will not provide a proof but note that with addition that each additional dimension 𝑑 ≥ 3 we
have added a corresponding multivalued periodic unit function 𝑢̌𝑑 (𝑠) with an associated branch point set
𝐵𝑑 = {𝑠𝑖 |∀𝑠𝑖 : 𝑖 ∈ ℕ, 𝑠 ∈ ℝ}𝑑 . With this addition, the argument of the proof of Theorem 2 can be shown to
be valid for dimensions 𝑑 ≥ 3.
Lemma 2.6. A tree defined by Theorem 2 is bounded if the length of its branches is bounded over its
domain 𝑠 ∈ ℝ.
Proof. A tree defined by Theorem 2 is essentially an integral of free differential polar vectors. It is clear
that if the integral of the angular coordinate ∫ ∇φ 𝑑𝑠 = 𝑐 for some constant c, then the path represented by
𝑢̌(𝑠) = 1 for all branches on this longest possible path is a straight line segment at angle c with respect to a
Cartesian origin point. The length 𝑙 of this line segment is
𝑙 = ∫ ∇r 𝑑𝑠
(18)
And the tree is bounded by a circle centered on the root with radius l. Therefore a tree as defined by
Theorem 2 is bounded if ∫ ∇r 𝑑𝑠 is bounded.
In some cases we can be more precise than this and calculate the smallest boundary circle explicitly.
7
Lemma 2.7. A non-trivial tree defined by Theorem 2, i.e. where the derivative coordinates ∫ 𝛻𝑟 𝑑𝑠 ≠ 0 and
∫ 𝛻𝜑 𝑑𝑠 ≠ 𝑐 for some constant 𝑐, has a non-integer Hausdorff dimension that exceeds its topological
dimension.
Proof. We do not give a formal proof but note that the canopy of this tree is constructed from infinitely
many non-trivial branches within a bounded area. It follows that the canopy of this tree is infinitely
intricate. One may also note that regardless of the particular shape of the branches, their distinctness will
increase a box count from generation to generation when assessing the specific Haussdorf dimension for
this tree.
Theorem 3. A tree as defined by Theorem 2 is an exact self-similar fractal if the derivative coordinates
satisfy the property that between any two consecutive parent and child branches defined by 𝑠𝑎𝑏 ∈ [𝑎. . 𝑏]
and 𝑠𝑏𝑐 ∈ [𝑏. . 𝑐] where a,b,c are branch points, that the scaling between branch points is constant
∇r |𝑠=𝑏..𝑐 = 𝜌∇r |𝑠=𝑎..𝑏 and 0 < 𝜌 < 1
(19)
and the directional change is equivalent
∇φ |𝑠=𝑏..𝑐 = ∇φ |𝑠=𝑎..𝑏
(20)
𝑠𝑖−1 − 𝑠𝑖 = 𝑐 for 𝑐 > 0
(21)
and branch points 𝑠𝑖 are equidistant
Proof. We will follow the characteristics of fractals outlined by Falconer [4]. By Lemma 2.7, the bounded
binary symmetric tree has a Hausdorff dimension that exceeds its topological dimension. The conditions
that ∇r |𝑠=𝑏..𝑐 = 𝜌∇r |𝑠=𝑎..𝑏 and ∇φ |𝑠=𝑏..𝑐 = ∇φ |𝑠=𝑎..𝑏 for consecutive parent and child branches ensures
that parent and child are exactly self-similar. Finally, the definition of the tree by Theorem 2 is simple.
Corrolary 3.1. If either 𝛻𝑟 |𝑠=b..c = 𝜌𝛻𝑟 |𝑠=𝑎..𝑏 or 𝛻𝜑 |𝑠=b..c = 𝛻𝜑 |𝑠=𝑎..𝑏 , but not both, then a bounded
binary symmetric tree formulated by Theorem 2 for the branch point set 𝐵 = {𝑠𝑖 |∀𝑠𝑖 : 𝑖 ∈ ℕ, 𝑠 ∈ ℝ} is
quasy fractal.
Proof. No formal proof is given but it is clear from the condition ∇r |𝑠=𝑏..𝑐 = 𝜌∇r |𝑠=𝑎..𝑏 or ∇φ |𝑠=𝑏..𝑐 =
∇φ |𝑠=𝑎..𝑏 , that some but not all fractal features are carried from branch to branch across the generations.
Remark 3.2. The conditions of Theorem 3 and Corollary 3.1 on the self-similarity across branch
generations may be relaxed by allowing the sets 𝑠𝑎𝑏 and 𝑠𝑏𝑐 to be defined less rigidly between branch
points but across arbitrary intervals of 𝑠 as long as these intervals are repeated ad infinitum. The resulting
fractal will be exact self-similar.
8
4.
Computational considerations
4.1 Calculating curves in ℝ𝒏
The definition of the derivate coordinate functions as an alternative representation of a curve in ℝ𝑛
simplifies the calculation of visual rendering of curves and objects in ℝ2 and ℝ3 , in particular when these
curves are not available algebraically. Furthermore, the calculation of curves and objects in ℝ𝑛 , 𝑛 > 3 is
also simplified.
We recall that rendering curves and objects in computer graphics often involves drawing a point 𝑝2 at a
given distance ∆𝑟 and rotation ∆𝜑 with respect to some point 𝑝1 such that the vector ∆𝒗(∆𝑟, ∆𝜑) = 𝑝2 − 𝑝1 .
The usual approach is to translate (𝑻) 𝑝1 back to the origin, rotate (𝑹) it such that it is parallel to the 𝑥-axis
and sometimes we may need to scale (𝑺) it back to unit length as well. Then we add the new line segment
and reverse the transformations: re-scale (𝑺−𝟏 ), re-rotate (𝑹−𝟏 ), re-translate (𝑻−𝟏 ).
𝑝2 = (𝑝1 ∙ 𝑻 ⋅ 𝑹 ∙ 𝑺 + ∆𝒗) ∙ 𝑺−𝟏 ∙ 𝑹−𝟏 ∙ 𝑻−𝟏
(22)
These transformations are typically done with matrices or quaternions which in particular for higher
dimensions can add a computational complexity.
A similar calculation for a curve in ℝ2 using derivate coordinate functions where ∆𝒗 = (∆𝑟, ∆𝜑),
𝑝1 = (𝑥1 , 𝑦1 ) and the direction at that point is Φ1 , then using Theorem 1:
𝑝2 (𝑠) = x2 + iy2 = ∫ ∇r exp(∫ i∇φ 𝑑𝑠) 𝑑𝑠
(23)
which we implement as a Riemann sum
𝑝2 (𝑠) = x2 + iy2 = ∑ ∇r exp(∑ i∇φ ∆𝑠) ∆𝑠
(24)
𝑥 + ∆𝑟 cos(Φ1 + ∆𝜑)
𝑝2 = ( 1
).
𝑦1 + ∆𝑟 sin(Φ1 + ∆𝜑)
(25)
to calculate the next point
Arguably, all of the transformations represented by the matrices 𝑻, 𝑹, 𝑺 are embedded in equation (25).
The inverse matrices are avoided by the integrals that ‘remember’ the position and direction of the curve at
the last point. So what is the difference? This is exactly the point: the transformations are built into the
equation that describes the path that we are calculating and we only have to concern ourselves with the
derivate coordinate functions, not with transformations.
4.2 Tree fractals
4.2.1 Pseudo code for generating fractals
An example of usage of an analytical formulation of a fractal by Theorem 3 is shown in the pseudo code
below.
9
Initialize
1:
2:
3:
4:
5:
6:
7:
set initial position 𝒑 = [0,0]
set initial direction 𝛷 = 0
set path parametric variable 𝑠 = 0
set step increment ∆𝑠 = 11
̌ [s] 2,3
read from a file, or calculate from a function into arrays ∇r[s], ∇𝜑[𝑠] and 𝑢
4
EvaluateFractal(𝑠, 𝛷 , 𝒑¸ 𝑢̌𝑟𝑖𝑔ℎ𝑡 ) // first right branch
EvaluateFractal(𝑠, 𝛷 , 𝒑¸ 𝑢̌𝑙𝑒𝑓𝑡 ) // first left branch
Function 𝛷𝑛𝑒𝑥𝑡 , 𝒑𝒏𝒆𝒙𝒕 =FractalEquation( ∇r, ∇𝜑, 𝒑𝒍𝒂𝒔𝒕 , 𝛷𝑙𝑎𝑠𝑡 )
1: 𝛷𝑛𝑒𝑥𝑡 = 𝛷𝑙𝑎𝑠𝑡 + ∇𝜑
𝑥𝑙𝑎𝑠𝑡
2: [𝑦 ] = 𝑝𝑙𝑎𝑠𝑡
𝑙𝑎𝑠𝑡
𝑥𝑛𝑒𝑥𝑡 𝑥𝑙𝑎𝑠𝑡 + ∇r cos ∇𝜑 ∆𝑠
3: 𝒑𝒏𝒆𝒙𝒕 = [𝑦
]=[
] // refer equation (25)
𝑦𝑙𝑎𝑠𝑡 + ∇r sin ∇𝜑 ∆𝑠
𝑛𝑒𝑥𝑡
4: return 𝛷𝑛𝑒𝑥𝑡 , 𝒑𝒏𝒆𝒙𝒕
Recursive EvaluateFractal(𝒔, 𝜱𝒍𝒂𝒔𝒕 , 𝒑𝒍𝒂𝒔𝒕 ¸ 𝑢̌)
1: while s < smax ∧ 𝑢̌[𝑠] ≠ 0 do // test for eof and branch condition
̌ ∇𝜑[𝑠], 𝒑𝒍𝒂𝒔𝒕, 𝛷𝑙𝑎𝑠𝑡 )
2:
Φnext , 𝒑𝒏𝒆𝒙𝒕 =FractalEquation( ∇𝑟[𝑠], 𝑢
3:
PlotLine(𝒑𝒍𝒂𝒔𝒕, 𝒑𝒏𝒆𝒙𝒕 ) // use a canvas or write to coordinate file for offline plotting
4:
𝒑𝒍𝒂𝒔𝒕 = 𝒑𝒏𝒆𝒙𝒕
5:
𝛷𝑙𝑎𝑠𝑡 = 𝛷𝑛𝑒𝑥𝑡
6:
𝑠 = 𝑠 + ∆𝑠
7: end while
̌ (𝑠) was 0
8: if 𝑠 < 𝑠𝑚𝑎𝑥 then // this implies that 𝑢
9:
EvaluateFractal(𝑠, 𝛷𝑙𝑎𝑠𝑡 , 𝒑𝒍𝒂𝒔𝒕¸ 𝑢̌𝑟𝑖𝑔ℎ𝑡 [𝑠] // next right branch
10:
EvaluateFractal(𝑠, 𝛷𝑙𝑎𝑠𝑡 , 𝒑𝒍𝒂𝒔𝒕¸ 𝑢̌𝑙𝑒𝑓𝑡 [𝑠] // next left branch
11: end if
A rudimentary Javascript implementation of this program is available [6] as well as a more evolved Ruby
implementation with accessory functions (refer section 4.2.4) for color, line width and transparency [7].
b
b
If another value for ∆𝑠 is chosen, note that ∫a ∇r 𝑑𝑠 and ∫a ∇φ 𝑑𝑠 for a branch from a to b, should remain invariant
and ∇r and ∇φ will need to be scaled accordingly.
2
The multivalued periodic unit function 𝑢̌ is either 1 and -1, or it is 0. This may be implemented as an array with
two columns 𝑢̌𝑟𝑖𝑔ℎ𝑡 and 𝑢̌𝑙𝑒𝑓𝑡 . When 𝑢̌ = 0, we may set both 𝑢̌𝑟𝑖𝑔ℎ𝑡 and 𝑢̌𝑙𝑒𝑓𝑡 to zero and for the purposes of a
branch test agree that we can use either 𝑢̌𝑟𝑖𝑔ℎ𝑡 or 𝑢̌𝑙𝑒𝑓𝑡 .
3
In practice we will often implement a branch test based on intervals of s rather than 𝑢̌. In that case we define two
constants 𝑢̌𝑟𝑖𝑔ℎ𝑡 = 1 and 𝑢̌𝑟𝑖𝑔ℎ𝑡 = −1.
4
Continuous tree fractals never have a trunk; that is an artifice of straight line tree fractals where a trunk is defined
as an initial lone branch.
1
10
4.2.2 Examples of tree fractals
The use of derivate coordinate functions in tree fractals allows us to introduce new features. We will
substitute the traditional straight line segments by other line shapes, including analytic functions. For
example, if we set ∇φ = c for some constant c such that for a branch from node s𝑎 to s𝑏 , the overall
b
𝜋
2
𝑠⁄
∆𝑠
angular change Φ = ∫a ∇φ 𝑑𝑠 = for any branch and ∇r = ( ) and position the branch points in 𝑠 at
3
3
equidistant intervals of length ∆𝑠, the fractal function p̌(s) of Theorem 3 will generate the fractal below
(image 3) and the derivative coordinate functions ∇r and ∇φ and the absolute periodic unit function |𝑢̌(𝑠)|
that were used to generate it (image 4)
Image 3: Smooth fractal : Φ =
𝜋
3
2
and ∇r = ( )
𝑠⁄
∆𝑠
Image 4: corresponding ∇r, ∇φ and |𝑢̌(𝑠)|
3
For comparison, we generate a straight line fractal using the similar parameters that are only assigned at the
branch points and zero elsewhere. So,
set ∇φ = c𝜑 for some constant c𝜑 at the branch points and ∇φ = 0 elsewhere whilst ensuring
b
that ∫a ∇φ 𝑑𝑠 =
𝜋
and
3
𝑠
2 ⁄∆𝑠
set ∇r = c𝑟 ( )
3
for some constant c𝑟 at the branch points and ∇r = 0 elsewhere whilst
b
∫a ∇r 𝑑𝑠
ensuring that
is invariant with respect to the same integral in the preceding smooth
fractal (Image 3).
we will obtain the straight line fractal below that appears to have the same general form as the smooth
fractal above. In Theorem 4 that follows, we will demonstrate the equivalence of their camopies under
scaling, rotation and translation.
11
Image 5: Straight line fractal : Φ =
𝑠
2 ⁄∆𝑠
𝜋
3
and ∇r =
Image 6: Corresponding ∇r, ∇φ and |𝑢̌(𝑠)|
( )
3
It is worth pointing out that the fractals in images 3 and 5 are not the same size. The curvature of the
smooth fractal in image 3 effectively shrinks the fractal compared to the straight lined fractal of image 5.
The scale factor between these two fractals may be found by comparing the absolute distance of subsequent
branch points 𝑠1 and 𝑠2 for these two fractals 𝑝̌𝐴 and 𝑝̌𝐵 .
𝑠𝑐𝑎𝑙𝑒 =
𝑝̌𝐴 (𝑠1 )−𝑝̌𝐴 (𝑠2 )
𝑝̌𝐵 (𝑠1 )−𝑝̌𝐵 (𝑠2 )
(26)
Theorem 4. If two tree fractals 𝑝̌𝐴 (𝑠) and 𝑝̌𝐵 (𝑠) are exact self-similar, with the respective derivative
coordinate functions (𝛻𝑟𝐴 , 𝛻𝜑𝐴 ) and (𝛻𝑟𝐵 , 𝛻𝜑𝐵 ) then if between any consecutive branch points 𝑠1 and 𝑠2
the angular integral
𝑠
𝑠
1
1
2
2
∫𝑠 ∇φ𝐴 𝑑𝑠 = ∫𝑠 ∇φ𝐵 𝑑𝑠
(27)
and branch length
𝑠
𝑠
1
1
2
2
∫𝑠 ∇r𝐴 𝑑𝑠 = c ∫𝑠 ∇r𝐵 𝑑𝑠
(28)
for some constant 𝑐 then in the limit for 𝑠 → ∞, the canopies of these fractals are equivalent through
scaling, translation and rotation.
Proof. We consider two binary symmetric tree fractals with the respective derivative coordinate functions
(∇r𝑎 , ∇φ𝑎 ) and (∇r𝑏 , ∇φ𝑏 ) then if we can demonstrate that with a suitable scaling, rotation and translation
the corresponding node points between the two fractals are arbitrarily close at 𝑠 → ∞ then these fractals are
equivalent through scaling, translation and rotation.
12
First we scale the fractal using (26) to ensure that at least for two subsequent nodes 𝑎, 𝑏 the absolute
distance between these two nodes is equivalent. Since by (26) and (27) and the fact that both of these
fractals are exact self-similar, we are sure that the distance between any subsequent pair of corresponding
nodes on the two fractals is equivalent. By virtue of the fact that total angular change (27) between branch
nodes is equivalent and that absolute distance between consecutive branch nodes is as well, we can rotate
and translate one fractal such that symmetry lines between the branches overlap with the other. As shown
in image 7 there remains a distance 𝑑 between corresponding branch points.
Image 7: Distance between corresponding nodes on two similar fractals
Due to the exact self-similar property (19) we know that
𝑑2 = 𝑑1 𝜌 and 0 < 𝜌 < 1
(28)
Since the branch points 𝑠i are equidistant (21) when 𝑠 → ∞, the node generations 𝑖 → ∞ and the distance d
between two corresponding nodes on fractals 𝑝̌𝐴 (𝑠) and 𝑝̌𝐵 (𝑠) is
d = lim𝑖→∞ 𝑑1 𝜌𝑖
(29)
Since 0 < 𝜌 < 1 it follows that 𝑑 = 0. Hence the canopies are equivalent through scaling, translation and
rotation.
4.2.3 Branch point programming
Binary trees and tree fractals
In the previous example we stated that the branch points in 𝑠 were at equidistant intervals. Using Theorem
2, we can introduce branch points arbitrarily to create different trees. Below is a tree that has the same
derivative coordinate functions as the fractal in Image 3 but with arbitrary branch points along s.
13
Image 8: Smooth fractal : Φ =
𝜋
3
2
and ∇r = ( )
𝑠⁄
∆𝑠
Image 9: Corresponding ∇r, ∇φ and |𝑢̌(𝑠)|
3
It may appear that this tree has lost its self-similarity, however this branch point pattern may well repeat
itself so the tree may still be qualitative fractal. This is why we will sometimes refer to such trees as fractal.
n-ary trees and tree fractals
In practice, we do not have to limit ourselves to binary trees. We may create n-ary trees and fractals or we
can create trees with different numbers of branches at different generations of branch points. The theorems
1,2 3 and their corollaries may be adapted to include these cases but we will not do so in this paper.
When implementing such mixed n-ary trees and fractals, which may be in higher dimensions the
programming of branch points and their relationship with the number of branches, the appropriate
derivative coordinate functions and dimensional direction a systematic approach is advised. For some
applications where the branch geometry of primary interest an L-system grammar can be used very
effectively [1].
Other applications, such as in engineering, a modular approach that focuses on sub trees may be more
suitable.
4.2.4 Accessory functions
Theorems 1,2 and 3 are useful formulations for mathematical analysis. For applications in engineering they
are useful for creating and optimizing tree and fractal structures or even architectures. In such practical
environments we need to extend the model of derivative coordinate functions to include other features than
a tree or fractal skeleton.
Definition 4.1. We define a set of derivative accessory functions 𝛻𝑏𝑖 for 𝑐𝑖 : ℝ ↦ ℝ𝑛 , 𝑖, 𝑛 ∈ ℕ and 𝑛 > 0 in
which we include the derivative coordinate functions and a set of non-derivative accessory functions 𝑏𝑗 for
𝑐𝑗 : ℝ ↦ ℝ𝑘 , 𝑗, 𝑘 ∈ ℕ and 𝑘 > 0 to enhance the formulation of a tree or fractal defined in theorem 2. The
set of derivative accessory functions is:
𝐶(𝑠) = {∇𝑟, ∇𝜑, ∇𝑏0 ⋯ ∇𝑏𝑛 , 𝑐0 ⋯ 𝑐𝑘 }
14
(29)
The accessory functions, together with the fractal function 𝑝̌ (𝑠), specify the enhanced fractal
𝑃̌ (𝑠) = {𝑝̌ (𝑠), 𝐶(𝑠)}.
(30)
Derivative vs non-derivative accessory functions
Here we have defined derivative functions to ensure that they are dependent on their relative values, or
history, rather than an absolute dependence on 𝑠. We may need this for line width, tube diameter and other
relative accessory functions. We also have non-derivative functions that allow us to control absolute
accessory functions. This may include color and material for example.
Static accessory functions
The accessory functions are typically used to describe additional features. For graphical rendering this may
be line width, color, transparency or any other aspect that is required. We may see accessory functions that
define a type of material, a finish etc.
Interdependent accessory functions
Accessory functions may be interdependent. If we have an accessory function that represents the water
pressure in a system of fractal pipes, than it will depend on the size of the pipes which would be another
accessory function.
Dynamic accessory functions
In modelling we can also create accessory functions that are dynamical parameters such as temperature,
flow or torque etc.
Sensory accessory functions
Accessory functions may also depend on external parameters. One interesting example is an accessory
function that measures the distance from the tree or fractal to a perimeter. This quantity can then be used to
adjust the angular derivative coordinate function ∇𝜑 such that it bends a branch away from or towards the
perimeter.
4.2.5 Fractal engineering
We are now in a position to formally define a process for engineering trees and fractals. We aim to
concatenate trees such that all of their properties, whether geometric or otherwise connect without
undesired discontinuities.
Theorem 5. Two trees 𝑃̌1 and 𝑃̌2 as defined by Theorem 2 may be concatenated to create a tree 𝑃̌ with
preservation of continuity by sequencing their accessory functions:
𝑃̌ = 𝑃̌1 ≪ 𝑃̌2 ⟹
(31)
𝑃̌ = {𝑝̌ (𝑠), C1 (s1 ) + C2 (s2 )}
(32)
For the domain 𝑠 = {𝑠1 ∪ 𝑠2 } is continuous and {𝑠1 ∩ 𝑠2 } = ∅.
15
Proof. It is sufficient to note that the derivative accessory functions, including the coordinate functions are
specified as derivatives and that in evaluating their absolute function values, these derivative functions are
integrated over their derivative variable 𝑠. As long as we ensure that the integration constants ℎ1 and ℎ2 for
the definite integral of each derivate accessory function ∇𝑏1 and ∇𝑏2 is such that the resulting primitives
𝑏1 (𝑠) + ℎ1 = 𝑏2 (𝑠) + ℎ1 , then concatenation will be smooth for each accessory function.
The non-derivative accessory functions do not required to connect smoothly since they represent
absolute properties.
Example of fractal engineering
Starting with two well-known curves, a smooth Koch curve based on the straight lined fractal with
𝑠
−1
𝑜
parameters ∇r = (α) ⁄∆𝑠 where 𝛼 =
𝑜 and Φ = 144 , a “golden fractal” by Tara D. Taylor[5]
2 cos 144
Image 10: Smooth Koch curve
and a smooth H-fractal
Image 11: Smooth H-fractal
we can concatenate the accessory functions:
16
H-fractal
Koch curve
Image 11: concatenated derivative coordinate functions
we obtain an engineered hybrid.
Image 12: Concatenated (hybrid) H-Koch fractal
Modular design in fractal engineering
Theorem 4 opens up the door towards modular design in fractal engineering. All engineering disciplines
have in common that systems and structures are created from subsystems and substructures.
Componentization is at the core of engineering. Theorem 5 allows us to create trees by adding up sub-trees.
There is no limit to this process we can use it to create low level structures that gradually become more
useful as they combine into larger structures.
As yet, engineers have not found many applications for fractal structures, but this is likely a matter of
maturity in design thinking. Below are some examples of engineered fractals and trees. The first example
(Image 13) shows the use of sub systems, wheels and brackets. These subsystems in turn have subsystems
in the form of cogs and spokes.
17
Image 13: Fractal system (sculpture) created from subsystems (wheels)
The second example of an engineered tree with a degree of fractal symmetry is shown below. The
accessory functions of this object are a sequence of smooth functions, giving an organic impression. Note
that the five forms share a root whose transparency accessory function renders it invisible. Although there
is no use of subsystems in this tree, it is clear that elements of it could be reused as subsystems.
Image 14: Tree system (organic object) created from other sequenced forms (organs)
18
Such engineered fractals and trees created from components, suggest the opportunity for component
libraries and tools for combining them. Genetic Fractals [8] experimented with a rudimentary library and
programming language for creating such structures.
4.2.6 Applications
From an engineering perspective, fractals are a solution in search of a problem. Probably we will find many
such problems in time. At present, the author has focused on two applications.
Computer Graphics
The use of derivative coordinates and the analytic fractal formulation is a powerful and simple approach to
drawing and animating all manner of tree fractals in two or more dimensions. As with most fractals we see,
there is an artistic incentive in this. The smooth nature of these fractals, combined with creativity allows us
to explore new forms and ideas.
The image below shows a hybrid tree fractal that has 4 branches at the root, followed by 2 branches, 3
branches and a series of 6 branches. The accessory functions include color, branch width and transparency.
Image 15: Two dimensional hybrid tree fractal
The following image was adapted from an animation of a rotating four dimensional tree [9], i.e. it has
branches into the second, third and fourth dimension consecutively.
19
Image 16: Four dimensional tree fractal projected into three dimensional space
3D Design and 3D Printing
It is a small step to take three dimensional design into a 3D printing environment. There are some technical
hurdles such as the need to convert the structure and its features into a triangulated mesh and a suitable file
format such as STL. 3D printing itself has its limitations in printing resolution and choice and mixing of
materials.
More importantly, and perhaps more fundamentally, the popularization of 3D printing has to overcome
the 3D design obstacle. Traditional manufacturing relies on professional designers and sophisticated design
tools for the creation of meaningful designs. The limitations of design tools suitable for non-professionals
restrict 3D designs to simplistic and childlike quality. Although a lot of progress is made by the 3D design
industry [10], non-professionals are not fully equipped yet to manipulate non-trivial geometrical forms in
three dimensions on two dimensional screens. The modular approach of Theorem 4 introduces a one
dimensional paradigm for 3D design. We need not concern ourselves with coordinates and rotations in
three dimensions, we only have to decide how to sequence subcomponents (a one dimensional activity) and
the tree formulation of Theorem 2 will take care of the complexities.
The example below shows a 3D printed tree fractal whose branches bend and curl [11].
20
Image 17: 3D printed tree designed with the tree formulation of Theorem 3.
[image and 3D printing credit: PrintaBit.ch]
The last example below shows the wind blades of a hypothetical wind turbine where the wing ribs are a
subsystem of the overall tree object. As the tree is evaluated, an algorithm calculates the triangular surface
meshes required for 3D printers.
Image 18: design for wind blades with triangular mesh for 3D printing
21
5.
Conclusion
5.1 Summary
We reviewed the analytic foundation of derivative coordinates in Theorem 1 that allows us to construct any
path in n-dimensional space. We then created a model for branching using a multivalued unit function that
allowed us to extend Theorem 1 to Theorem 2, i.e. an analytic formulation for binary symmetric trees,
including smooth trees. Theorem 3 summarized the conditions for using this formulation for exact selfsymmetric tree fractals.
Using these formulations we reviewed the computational advantages of using derivative coordinates
and the tree and fractal formulations over traditional constructions using matrices. We then looked at a few
examples of fractals created with this approach and demonstrated equivalence between straight lined tree
fractals and fractals in Theorem 4.
We extended the model by introducing derivative accessory functions that allow us to model additional
features that we find in engineered systems, such as materials, finishes etc. This led to Theorem 5 that
allows us to do fractal engineering, i.e. building tree and fractal structures from components whilst ensuring
structural integrity and continuity. Finally we reviewed a few applications of the theory in computer
graphics and 3D printing.
5.2 Directions for further research
It would seem that there are a number of directions that the theory presented here can be further developed.
Three areas appear to stand out.
Analysis of smooth trees and fractals
Evidently, the ability to use analytical tools on the analytic formulation of trees and fractals should allow us
to explore existing and new features of tree fractals analytically rather than numerically.
Development of complex fractal structures in engineering
The use of fractal structures in engineering is still in its infancy. The analytic formulation of such structures
will allow us to marry structure with structural analysis and dynamics.
Use of derivative coordinates in mathematics
Derivative coordinates are independent from an absolute coordinate system. Beyond the use in trees and
fractals, they may be used in any system with localized behavior.
6.
References
[1]
Aristid Lindemayer, Przemyslaw Prusinkiewicz. The algorithmic beauty of plants. Springer-Verlag New
York, 1990
Benoit B. Mandelbrot. The fractal geometry of nature. W. H. Freeman and Company, (1982)
Benoit B. Mandelbrot, Michael Frame. The canopy and shortest path of a self-contacting fractal tree. The
Mathematical Intelligencer, (1999)
[2]
[3]
22
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
Kenneth Falconer. Fractal Geometry: Mathematical Foundations and Applications. John Wiley & Sons,
Ltd, 2003
Tara D Taylor, Finding gold in the forest: self-contacting symmetric binary fractal trees and the golden
ratio. Talk at joint AMS/MAA meeting, pp17-19, 2007
Smooth tree fractal explorer. http://www.geneticfractals.com/gf/GeneticFractalssmoothTreeFractalExplorer.html
2D genetic fractals maker’s guide. http://geneticfractals.org/2d-genetic-fractals-makers-guide/
3D design concept. http://geneticfractals.com/index.php
Animation of four dimensional hyper tree. https://www.youtube.com/watch?v=nCdaLnlHa8w
Autodesk design apps. http://www.123dapp.com/
3D printed tree fractal. http://www.nyce.ly/artworks/shadow-tree-ombrier-i
23
| 5cs.CE
|
IRRATIONAL BEHAVIOR OF ALGEBRAIC DISCRETE VALUATIONS
arXiv:1311.6047v2 [math.AC] 10 Jun 2014
SOUMYA DEEPTA SANYAL
1. Introduction
In this paper, we discuss the behavior of algebraic discrete valuations dominating a two
dimensional normal (Noetherian) local domain (R, mR ).
To a valuation ν of the quotient field K = QF (R) of R, we associate a valuation ring
Vν = {f ∈ K | ν(f ) ≥ 0}. This ring is quasi-local, with maximal ideal mν = {f ∈ K |
ν(f ) > 0}. We say that the valuation ν dominates (R, mR ) if Vν ⊃ R and R ∩ mν = mR .
We say that the valuation ν is discrete if its value group is order isomorphic to Z. In this
case Vν is Noetherian, and thus is local.
For n ∈ Z≥0 , let In = {f ∈ R | ν(f ) ≥ n}. The questions we address in this paper are: what is ℓR (R/In ) for n ≫ 0? How close is it to being a polynomial? What is
n)
? Is this limit well behaved?
limn→∞ ℓR (R/I
n2
There are two types of discrete valuations ν dominating R. The first case is that ν is divisorial, i.e., the residue field extension Vν /mν : R/mR is transcendental. By Abhyankar’s
inequality (Theorem 1 of [1]) it follows that the residue field extension given by a divisorial valuation ν dominating a normal local domain of dimension two is of transcendence
degree exactly one. The second case is that ν is algebraic, i.e., the residue field extension
Vν /mν : R/mR is algebraic.
In the divisorial case, it is known by work of Cutkosky and Srinivas in [5] that when R
is excellent and equicharacteristic of dimension 2, ℓR (R/In ) can be written as a quadratic
polynomial Q(n) plus a bounded function σ(n) for n ≫ 0. They further show that if
R/mR has characteristic zero, or is finite, then σ(n) is periodic for n ≫ 0 and they give an
example to show that there exist R, ν with characteristic R/mR > 0 such that σ(n) is not
eventually periodic. Finally, in their Example 6, they give an example of a divisorial valun)
is an irrational number.
ation ν dominating a 3 dimensional ring such that limn→∞ ℓR (R/I
n3
The analysis in dimension two of [5] left open the remaining case that ν is an algebraic
discrete valuation. In the case that Vν /mν is finite over R/mR , everything is well behaved.
In section 3 of this paper, we prove the following Theorem 1.1 and Corollary 1.2:
Theorem 1.1. Suppose that ν is an algebraic discrete valuation dominating a local domain
(R, mR ), and that Vν /mν is finite over R/mR . Then there exist c, b ∈ Z such that
ℓR (R/In ) = cn + b
for n ≫ 0, where c = |Vν /mν : R/mR |.
1
Corollary 1.2. Suppose that ν is an algebraic discrete valuation dominating a local domain (R, mR ), and that Vν /mν is finite over R/mR . Then
ℓR (R/In )
= 0.
n→∞
n2
lim
In section 4 of this paper, we show in Theorems 4.11 and 4.19, and Corollary 4.20, that
when we consider infinite algebraic residue field extensions, we have a more interesting
result:
Theorem 1.3. There exists a regular local ring (R, mR ) of dimension 2 and a discrete,
rank 1 valuation ν of the quotient field of R dominating R, such that the function ℓR (R/In )
cannot be written as a quasi-polynomial plus a bounded function for large integers n.
We remind the reader that a quasi-polynomial in the indeterminate n is an expression of
the form ad (n)nd + ad−1 (n)nd−1 + · · · + a1 (n)n + a0 (n), where d is the degree of the quasipolynomial, and the coefficients ad (n), ad−1 (n), . . . , a1 (n), a0 (n) are periodic functions of n.
Theorem 1.4. Let C ∈ R ∩ [0, 21 ] be given. There exists a regular local ring (R, mR )
of dimension 2 and a discrete, rank 1 valuation ν of the quotient field of R dominating
R, such that each of the functions ℓR (R/In ) and ℓR (In /In+1 ) is not a quasi-polynomial
plus a bounded function for large integers n and such that limn→∞ ℓR (Inn/In+1 ) = C and
n)
limn→∞ ℓR (R/I
= C2 .
n2
Corollary 1.5. There exists a regular local ring (R, mR ) of dimension 2 and a discrete,
rank 1 valuation ν of the quotient field of R dominating R, such that each of the functions
ℓR (R/In ) and ℓR (In /In+1 ) is not a quasi-polynomial plus a bounded function for large
n)
are irrational
integers n and such that each of limn→∞ ℓR (Inn/In+1 ) and limn→∞ ℓR (R/I
n2
(even transcendental) positive numbers.
Next, we discuss some questions that arise from the methods used in this paper. Since
{In } is a graded family of mR -primary ideals in the two dimensional local ring R, the limit
ℓR (R/In )
n→∞
n2
is known to exist in equicharacteristic regular local rings by Mustata [10], and they exist
in arbitrary regular local rings (even analytically irreducible local rings) by Cutkosky [2]
and [3]. In light of these results, we may ask the following question:
lim
Question 1.6. Which numbers C are realizable as limits
ℓR (R/In )
lim
n→∞
n2
from an algebraic discrete valuation on a regular local ring R of dimension two?
2
In this paper, we show that all the numbers in the real interval [0, 41 ] are realizable as
limits. Since I1 = mR , and I1n ⊂ In for all n, by comparison with the Hilbert-Samuel
multiplicity we have the upper bound
ℓR (R/In )
1
e(R)
= .
≤
2
n→∞
n
2
2
In all of our examples, we have the stronger result that
lim
ℓR (In /In+1 )
n→∞
n
(1)
lim
exists, and every number in the real interval [0, 12 ] can be obtained as a limit.
Thus we may ask the following question:
Question 1.7. Does limn→∞ ℓR (Inn/In+1 ) always exist for an algebraic discrete valuation
dominating a regular local ring of dimension two?
It is known that this limit does not generally exist if {In } is a filtration of mR -primary
ideals (Theorem 4.6 of [4]), so a positive answer to Question 1.7 would use special properties of the valuation ideals In .
Now we briefly discuss the methods used in this paper. In constructing our examples,
we make use of an algorithm of [7] that generalizes an algorithm of [11] for constructing
generating sequences of valuations. The algorithm in [7] is valid in arbitrary two dimensional regular local rings. This technique of generating sequences is also used in [8] and
[9] to find stable toric forms of extensions of associated graded rings along a valuation in
finite defectless extensions of algebraic function fields of dimension two.
We take as our ground field k = L ({σi }), where L is an arbitrary field and {σi }i∈Z+
is a set of algebraically independent elements over L. We take for our ring R := k[[x, y]]
and inductively define a generating sequence of our valuation, {Pi }i≥0 ⊂ R by P0 =
x, P1 = y, and Pi+1 = Pi2 − σi x2ri for i ≥ 1. Provided that the ri ∈ N are such that
r0 = 1 and ri+1 > 2ri for i ≥ 1, the algorithm ensures that there exists a unique discrete
valuation ν of QF (R) dominating R, such that ν(Pi ) = ri for i ∈ N. The residue field of
√
ν is naturally isomorphic to k({ σi }).
L
In the associated graded algebra s≥0 Is /Is+1 , the lengths of the graded components
Is /Is+1 are given by the number of distinct monomials in the generators {Pi } that have
value equal to s.
That is, a k-basis of Is /Is+1 is:
(2) Bs = {[P0n0 P1n1 · · · Pini ] | i ∈ N, n0 , . . . , ni ∈ N,
nj ∈ {0, 1} for 1 ≤ j ≤ i, and n0 + n1 r1 + · · · + ni ri = s}.
3
By multiplying basis elements through by an element of suitable ν-value in R and
checking linear independence conditions, we show that ℓR (Is /Is+1 ) is a non-decreasing
function of s. By choosing the ν-values ri of the generators, we can control the growth
of ℓR (Is /Is+1 ). By imposing a combinatorial condition on the sequence {ri }, that ri+1 >
r0 + r1 + · · · + ri for all i ≥ 0, we can ensure that there is a partition of R≥0 of the form
{[ak , bk )}k∈Z>0 , such that on every interval [ak , bk ), the function ℓR (Is /Is+1 ) is constant,
and takes on distinct values for every k. This part of the construction already shows that
the Hilbert function is highly non-polynomial.
Next, we show that for these examples, the limit
n)
as a consequence.
existence of ℓR (R/I
n2
ℓR (In /In+1 )
n
exists, and deduce the
The construction places the set of multiplicities (0, 14 ] in bijective correspondence with
[0, ∞). This is done by specifying the values of ri+1 − (2ri + 1). More precisely, to construct the valuation to have multiplicity B ∈ (0, 14 ], we set the difference ri+1 − (2ri + 1)
1
to be equal to the coefficient of 2i−1
in the 2-adic expansion of B2 − 2. For the remaining case, taking ri+1 −(2ri +1) = 2i−1 for i ≥ 0 gives an example where the multiplicity is 0.
2. Notation
Let N denote the set {0, 1, 2, . . . } and Z+ denote the set {1, 2, 3, . . . }. Suppose that ν
is a discrete valuation dominating a local ring (R, mR ). Let In = {f ∈ R | ν(f ) ≥ n}. If
f ∈ In , denote by [f ] its image in In /In+1 . If r, s ∈ N, [x] ∈ Ir /Ir+1 and [y] ∈ Is /Is+1 ,
denote by [x] · [y] their product (x + Ir+1 )(y + Is+1 ) ∈ Ir+s /Ir+s+1 in the associated graded
ring ⊕n∈N In /In+1 . In particular if c ∈ R/mR , we will write c · [x] ∈ Ir /Ir+1 . Also, if
c ∈ R/mR and [d] ∈ Vν /mν , then c · [d] = (c + mR )(d + mν ) ∈ Vν /mν . If R/mR ∼
= k ⊂ R,
then ℓR (Is /Is+1 ) = dimR/mR (Is /Is+1 ). In this case (which applies to this paper), we will
use the two notations interchangeably. Finally, we let Γν = {ν(f ) | f ∈ QF (R) \ {0}} and
S R (ν) = {ν(f ) | f ∈ R \ {0}} ⊂ Γν denote the value group of ν and the value semigroup
of ν respectively.
3. Results for finite residue field extensions
In this section, we prove Theorem 1.1 and Corollary 1.2 from the introduction.
Lemma 3.1. Suppose that ν is an algebraic discrete valuation dominating a local domain
(R, mR ). Let n ∈ N. Suppose that dimR/mR (In /In+1 ) = r. Then for all k ∈ S R (ν), we
have that dimR/mR (In+k /In+k+1 ) ≥ r.
Proof. Suppose that e1 , . . . , er are elements of In \ In+1 such that [e1 ] , . . . , [er ] form a
basis for In /In+1 over R/mR . Let k ∈ S R (ν) be given. Then there exists g ∈ R such
that ν(g) = k. Hence ge1 , . . . , ger are elements of In+k \ In+k+1 with nonzero residues
[ge1 ] , . . . , [ger ] in In+k /In+k+1 .
4
P
Suppose that ri=1 ci [gei ] = P
[0] in In+k /In+k+1
P
Pr , where ci ∈ R/mR for 1 ≤ i ≤ r. Then
r
r
ν(
c
ge
)
>
n
+
k.
Thus
ν(
c
e
)
=
ν(
i=1 i i
i=1 ci gei ) − ν(g) > n + k − k = n, whence
Pr i=1 i i
i=1 ci [ei ] = [0] in In /In+1 . Since [e1 ] , . . . , [er ] formed a basis for In /In+1 , we must have
ci = 0 for all 1 ≤ i ≤ r. Hence [ge1 ] , . . . , [ger ] are linearly independent over R/mR , and
so dimR/mR (In+k /In+k+1 ) ≥ r.
Lemma 3.2. Suppose that ν is an algebraic discrete valuation dominating a local domain
(R, mR ). There exists n0 ∈ N such that n ∈ S R (ν) for all n ≥ n0 .
Proof. Since ν is a valuation of the quotient field of R, S R (ν) = {ν(f ) | f ∈ R \ {0}}
generates Γν . To see this, let fg ∈ K ∗ , with f, g ∈ R \ {0}. Then ν( fg ) = ν(f ) − ν(g).
Hence, any element of Γν is the difference of two elements in S R (ν). In particular, since
Γν = Z, there exist t, u ∈ S R (ν) such that t − u = 1.
If u = 0, then t = 1, S R (ν) = N and the lemma follows by taking n0 = 0. Suppose
that u > 0 and let i ∈ N. Then there exist k, r ∈ N such that i = ku + r, 0 ≤ r < u, by
the division algorithm. Hence u2 + i = tr + u(u + k − r) ∈ S R (ν). Thus for all n ≥ u2 ,
n ∈ S R (ν). Thus the lemma follows by taking n0 = u2 .
Lemma 3.3. Suppose that ν is an algebraic discrete valuation dominating a local domain
(R, mR ), and that Vν /mν is finite over R/mR . For all n ∈ N, dimR/mR (In /In+1 ) ≤ |V /mν :
R/mR |.
Proof. Let r = |V /mν : R/mR |, and let f1 , . . . fr+1 be elements
h
iin In \ In+1 such that
fr+1
[f1 ], . . . , [fr+1 ] are distinct elements of In /In+1 . Then [1], . . . , f1 form distinct nonzero
residues hin iV /mν . Hence there exist ci ∈ R/mR for 1 ≤ i ≤ r + 1, not all zero, such that
Pr+1
Pr+1 fi
Pr+1
Pr+1
fi
i=1 ci f1 = [0]. Thus ν( i=1 ci f1 ) > 0, and so ν( i=1 ci fi ) > n. Hence
i=1 ci [fi ] =
[0], and [f1 ], . . . , [fr+1 ] are linearly dependent over R/mR . Thus dimR/mR (In /In+1 ) ≤ r.
Proposition 3.4. Suppose that ν is an algebraic discrete valuation dominating a local
domain (R, mR ), and that Vν /mν is finite over R/mR . There exists n1 ∈ N such that for
all n ≥ n1 , dimR/mR (In /In+1 ) = |V /mν : R/mR |.
Proof. Let |V /mν : R/mR | = r. For 1 ≤ i ≤ r − 1, there exist elements αi ∈ V \ mν ,
such that the elements [1] , [α1 ] , . . . , [αr−1 ] form a basis for V /mν over R/mR . For 1 ≤
i ≤ r − 1, we may write αi = fgii , where fi , gi ∈ R and ν(fi ) = ν(gi ) = ni . Next, define
Qr−1
Pr−1
g = i=1
gi . Then g ∈ R, and ν(g) = i=1
ni =: N . Further, for all 1 ≤ i ≤ r − 1,
Q
P
gfi
gαi = gi = fi · j6=i gj , so that gαi ∈ R, and in particular, ν(gαi ) = ν(fi ) + j6=i ν(gj ) =
P
ni + j6=i nj = N . Thus g ∈ IN \ IN +1 and gαi ∈ IN \ IN +1 for 1 ≤ i ≤ r − 1. In
particular, [g] and the elements [gαi ] for 1 ≤ i ≤ r − 1 are nonzero elements of IN /IN +1 .
Pr−1
Suppose that c0 [g] + i=1
ci [gαi ] = [0] in IN /IN +1 , where ci ∈ R/mR for 0 ≤ i ≤ r − 1.
Pr−1
Pr−1
Pr−1
Then ν(c0 g + i=1 ci gαi ) > N , whence ν(c0 + i=1
ci αi ) = ν(c0 g + i=1
ci gαi ) − ν(g) >
Pr−1
N − N = 0, so that c0 [1] + i=1 ci [αi ] = [0] in V /mν . Hence cj = 0 for 0 ≤ j ≤ r − 1,
5
and so [g] and the elements [gαi ], 1 ≤ i ≤ r − 1 are linearly independent over R/mR in
IN /IN +1 . Thus dimR/mR (IN /IN +1 ) ≥ r. By Lemma 3.3, dimR/mR (IN /IN +1 ) = r, and so
dimR/mR (IN /IN +1 ) = |V /mν : R/mR |.
Finally, suppose that n0 is as in Lemma 3.2. Then for all n ≥ n0 , we have that
n ∈ S R (ν). Hence dimR/mR (IN +n /IN +n+1 ) = |V /mν : R/mR | for all n ≥ n0 , by Lemma
3.1 and Lemma 3.3. Thus the lemma follows by taking n1 = N + n0 .
Now we prove Theorem 1.1 and Corollary 1.2 from the introduction. We remind the
reader of their statements:
Theorem 3.5. Suppose that ν is an algebraic discrete valuation dominating a local domain
(R, mR ), and that Vν /mν is finite over R/mR . Then there exist c, b ∈ Z such that
ℓR (R/In ) = cn + b
for n ≫ 0, where c = |V /mν : R/mR |.
Corollary 3.6. Suppose that ν is an algebraic discrete valuation dominating a local domain (R, mR ), and that Vν /mν is finite over R/mR . Then
ℓR (R/In )
= 0.
n→∞
n2
lim
Proof. In Proposition 3.4 we have shown that there exists n1 ∈ N such that ℓR (In /In+1 ) =
c for n ≥ n1 , where c = |V /mν : R/mR |. Observe that
ℓR (R/In ) =
k=n−1
X
ℓR (Ik /Ik+1 ).
k=0
Hence for n ≥ n1 , we have
ℓR (R/In ) =
k=n
1 −1
X
ℓR (Ik /Ik+1 ) +
k=0
k=n−1
X
ℓR (Ik /Ik+1 )
k=n1
and hence
ℓR (R/In ) =
k=n
1 −1
X
k=0
Thus
ℓR (R/In ) = cn +
ℓR (Ik /Ik+1 ) + [n − n1 ]c.
"k=n −1
1
X
k=0
#
ℓR (Ik /Ik+1 ) − n1 c .
This proves Theorem 1.1. Dividing both sides by n2 and taking the limit as n → ∞ gives
Corollary 1.2.
6
4. Main Results
In this section we prove results for algebraic residue field extensions of infinite degree.
Lemma 4.1. Let k = L ({σi }), where L is a field and {σi }i∈Z+ is a set of algebraically
√
independent elements over L. Let αi = σi for i ∈ Z+ . Let R be the power series ring
R := k[[x, y]]. Define elements Pi ∈ R by
P0 = x, P1 = y, and Pi+1 = Pi2 − σi x2ri for i ≥ 1.
Suppose that ri ∈ N are such that
r0 = 1 and ri+1 > 2ri for i ≥ 1.
Then there exists a unique valuation ν of the quotient field of R which dominates R, such
that
ν(Pi ) = ri for i ∈ N.
ν has the property that {Pi } satisfies the conclusions of Theorem 4.2. of [7]. The residue
field of ν is naturally isomorphic to k({αi }). We can take Ui = xri for i ≥ 1 in 4) of
Theorem 4.2. of [7], and then we obtain that
Pi
αi = r
xi
for i ≥ 1.
Proof. Let βi = ri for i ∈ N. Theorem 1.1 [7] and its proof give the existence of a
valuation ν associated to the {βi }i∈N and {αi }i∈Z+ , which dominates R. The proof (from
the middle of page 21 of [7]) determines ν by constructing a sequence of polynomials which
are a generating sequence determining the valuation ν. Now the sequence of polynomials
are in fact the generating sequence associated to ν by the algorithm of Theorem 4.2 [7]. If
we start the generating sequence with x and y, and we take the Ui in the construction to
be xri (which is consistent with the construction), we have that the generating sequence
constructed in the proof of Theorem 1.1 is in fact the {Pi }, and the algorithm of Theorem
4.2 produces the generating sequence {Pi }.
Remark 4.2. In the choice of the sequence {ri }i∈N determining the valuation ν in Lemma
4.1, the condition that ri+1 > 2ri for all i ≥ 1 is necessary for the conclusions of Lemma
4.1, by the use of the algorithm of [7].
Lemma 4.3. A k-basis of Is /Is+1 is
(3) Bs = {[P0n0 P1n1 · · · Pini ] | i ∈ N, n0 , . . . , ni ∈ N,
nj ∈ {0, 1} for 1 ≤ j ≤ i, and n0 + n1 r1 + · · · + ni ri = s}.
Proof. We first show that the distinct elements of Bs are linearly independent over k. Let
us fix an enumeration {γh } of the monomials
(4) {P0n0 P1n1 · · · Pini | i ∈ N, n0 , . . . , ni ∈ N,
nj ∈ {0, 1} for 1 ≤ j ≤ i, and n0 + n1 r1 + · · · + ni ri = s}.
7
Observe that {[γh ]} is an enumeration of the elements of Bs . By construction, ν(γ1 ) =
ν(γh ) = s for all h. Furthermore, the exponent i-tuples (n0 , . . . , ni ) associated to the
monomials γh are distinct for distinct h.
P
Suppose now that therePis an equation
ch [γh ] = 0 in Is /Is+1 , with ch ∈ k. Then
P
P
ν( ch γh ) > s, and so ν( ch γγh1 ) = ν( ch γh ) − ν(γ1 ) > s − s = 0. Thus we have an
P γh
equation
ch [ γ1 ] = 0 in Vν /mν . Now Theorem 4.2 (2) of [7] states that the {[ γγh1 ]} are
linearly independent over k. Thus we must have had ch = 0 for all h, and so the elements
of Bs were linearly independent over k.
Next, by reducing equation (36) modulo Is+1 in the proof of Theorem of 1.1 in [7], we
see that the elements of Bs generate Is /Is+1 over k.
Whence Bs is a k-basis for Is /Is+1 .
Definition 4.4. Let θ ∈ R≥0 ∪ {∞}.
1. If θ < ∞, then define an associated sequence {ej }j≥2 as follows: define e2 ∈ N
so that e22 ≤ θ < e22+1 , and define for every j ≥ 3, ej ∈ {0, 1} such that 0 ≤
P
1
1
θ − e2 21 − · · · − ej 2j−1
< 2j−1
. Further, for every j ≥ 2, define Rj := θ − jk=2 ek 21−k .
2. If θ = ∞, then define an associated sequence {ej }j≥2 by ej = 2j−1 for every j ≥ 2.
P
Notice in this case that ij=2 ej 21−j = i − 1.
In the sequel, we will have the following assumptions. We assume θ ∈ R≥0 ∪ {∞} given,
and we define the sequence {ri }i≥0 inductively by
(5)
r0 = 1, r1 = 1, and ri+1 = 2ri + 1 + ei+1 for all i ≥ 1.
Then the general term ri is given by the formula
(6)
r0 = 1 and ri = 2i − 1 +
for i ≥ 1, where {ej } is as in Definition 4.4.
i
X
ek 2i−k ,
k=2
Lemma 4.5. With our assumption (5), we have that
1. ri+1 > 2ri for all i ≥ 1,
2. 2ri > r0 + · · · + ri for all i ≥ 2,
Proof. Since ri+1 = 2ri +1+ei+1 > 2ri for all i ≥ 1 by construction, the first claim follows.
We verify the second claim by induction. By (5), r2 = 3 + e2 > 2 = r0 + r1 , and the
second claim holds for the case i = 2.
8
Now suppose that i > 2, and that the second claim holds for the case k = i−1. We have
that 2ri = ri + ri > ri + (r0 + · · · + ri−1 ), the last inequality by the inductive statement.
Hence the second claim holds for all i ≥ 2.
By the first claim of Lemma 4.5, the sequence {ri } defined by (5) satisfies the assumptions of Lemma 4.1. Let ν be the valuation defined by Lemma 4.1 with the sequence {ri }
defined by (5).
Remark 4.6. In the sequel, we will have inductively chosen the sequence {ri }i∈Z+ such
that
ri+1 ≥ max(2ri , r0 + r1 + · · · + ri ) + 1.
Lemma 4.7. For any i ∈ Z≥1 , let s ∈ (r0 + r1 + · · · + ri , ri+1 ). Then dimk (Is /Is+1 ) = 2i .
Proof. Define the following function:
N0 (i, s, n1 , . . . , ni ) = s − (n1 r1 + · · · + ni ri ),
where n1 , . . . , ni ∈ {0, 1}, as in Lemma 4.3. Since s > r1 + · · · + ri ≥ n1 r1 + · · · + ni ri ,
N0 (i, s, n1 , . . . , ni ) is nonnegative. Note that for any choice of n1 , . . . , ni , we have that
N (i,s,n1 ,...,ni ) n1
N0 (i, s, n1 , . . . , ni ) + n1 r1 + · · · + ni ri = s, so that [P0 0
P1 · · · Pini ] ∈ Bs .
Furthermore, since rj > s for j > i by Lemma 4.5, we have that n0 + n1 r1 + · · · +
nj−1 rj−1 + nj rj 6= s for any choice of j ≥ i + 1 and n0 , n1 , . . . , nj−1 , nj ∈ N with nj > 0.
Thus the elements of Bs are precisely the following:
N0 (i,s,n1 ,...,ni )
{[P0
P1n1 · · · Pini ] | n1 , . . . , ni ∈ {0, 1}}.
Whence dimk (Is /Is+1 ) = 2i .
Lemma 4.8. Let {ri }i≥0 ⊂ Z≥0 be a sequence, let max(2r2 , r0 +r1 +r2 ) = 2r2 > r0 +r1 +r2 ,
and let ri+1 > max(2ri , r0 +· · ·+ri ) for i ≥ 2. Then max(2ri , r0 +· · ·+ri ) = 2ri > r0 +. . . ri
for all i ≥ 2.
Proof. The case i = 2 is given by hypothesis. Now suppose that the statement holds for
some k ≥ 2. Then 2rk+1 = rk+1 + rk+1 > rk+1 + max(2rk , r0 + · · · + rk ) = rk+1 + 2rk >
rk+1 + (r0 + · · · + rk ). Hence max(2rk+1 , r0 + · · · + rk+1 ) = 2rk+1 > r0 + · · · + rk+1 , and
the result follows by induction.
Remark 4.9. Suppose {ri }i∈Z≥0 is constructed as in (6). Then ri+1 > max(2ri , r0 + · · · +
ri ) for all i ≥ 2 and the conclusions of Lemma 4.8 hold. By Remark 4.6 and Lemma 4.8
9
it is enough to show that 2r2 = 6 + 2e2 > 5 + e2 = r0 + r1 + r2 .
Lemma 4.10. For all i ∈ Z>0 we have that ri+1 − (r0 + · · · + ri ) ≥ i.
P
Pi
i+1−j and
Proof. From (6) we can deduce that ri+1 = 2i+1 − 1 + i+1
j=2 ej 2
j=0 rj =
P
P
j
i
2i+1 − 1 − i + j=2 k=2 ek 2j−k . Hence
ri+1 −
i
X
rj = i +
j=0
and it is enough to show that
i+1
X
k=2
Pi+1
ek 2i+1−k −
i+1−k
k=2 ek 2
Suppose first that θ < ∞. We have that
i+1
X
k=2
ek 2i+1−k −
= ei+1 + 2i
i
X
k=2
−
Pi
j=2
j
i X
X
j
i X
X
ek 2j−k
j=2 k=2
Pj
j−k
k=2 ek 2
≥ 0.
ek 2j−k
j=2 k=2
ek 21−k −
= ei+1 + 2i (θ − Ri ) −
≥ ei+1 + 2i (θ − Ri ) −
i
X
2j−1
j=2
i
X
j=2
i
X
j=2
j
X
ek 21−k
k=2
2j−1 (θ − Rj )
2j−1 (θ − Ri )
= ei+1 + 2i (θ − Ri ) − (2i − 2)(θ − Ri )
= ei+1 + 2(θ − Ri ) ≥ 0,
the penultimate inequality since for all 1 ≤ j ≤ i, we have that θ − Rj ≤ θ − Ri .
Suppose next that θ = ∞. We have that
i+1
X
i+1−k
ek 2
k=2
i
= ei+1 + 2
i
X
k=2
−
1−k
ek 2
j
i X
X
ek 2j−k
j=2 k=2
−
= ei+1 + 2i (i − 1) −
≥ ei+1 + 2i (i − 1) −
10
i
X
j=2
i
X
j=2
i
X
j=2
j−1
2
j
X
ek 21−k
k=2
2j−1 (j − 1)
2j−1 (i − 1)
= ei+1 + 2i (i − 1) − (2i − 2)(i − 1)
= ei+1 + 2(i − 1) ≥ 0.
Theorem 4.11. There exists a regular local ring (R, mR ) of dimension 2 and a discrete,
rank 1 valuation ν of the quotient field of R dominating R, such that the function α(n) =
ℓR (In /In+1 ) is not a quasi-polynomial plus a bounded function for large integers n.
Proof. Take R = k[[x, y]] and consider the valuation ν constructed by the generating
sequence of Lemma 4.1, along with the associated function α(n) = ℓR (In /In+1 ). We
will show that if ν has a generating sequence defined inductively by: ν(P0 ) = r0 = 1,
ν(P1 ) = r1 = 1 and ν(Pi+1 ) = ri+1 where {ri } is defined by (5), then α(n) cannot be the
sum of a quasipolynomial and a bounded function for large n.
Suppose by way of contradiction that there exists n0 > 0 such that
α(n) = A(n) + b(n),
for n ≥ n0 , where
A(n) = αd (n)nd + · · · + α1 (n)n + α0 (n),
the αi (n) are periodic functions of integral period, s is their common integral period, and
−M ≤ b(n) ≤ M is a bounded function. By defining b(n) = α(n) − A(n) for n ≤ n0 , we
can take n0 to be 1.
For each 0 ≤ p < s, consider the function
Ap (t) := A(p + st) = αd (p + st)(p + st)d + · · · + α1 (p + st)(p + st) + α0 (p + st).
The coefficients αd (p + st), . . . , α1 (p + st), α0 (p + st) are constant functions of t ∈ Z+ ,
and are each equal to αd (p), . . . , α1 (p), α0 (p), respectively. Expanding the powers of p + st
shows that these are polynomials of degree ≤ d in t. Thus we may consider s (not
necessarily distinct) polynomial functions of t ∈ Z+ :
Ap (t) = αd (p)(p + st)d + · · · + α1 (p)(p + st) + α0 (p),
and also s (not necessarily distinct) bounded functions of t ∈ Z+ :
bp (t) = b(p + st),
for 0 ≤ p < s.
For each 0 ≤ p < s, define fp (t) := Ap (t) + bp (t). For any function ψ(t), define
∆t ψ(t) := ψ(t) − ψ(t − 1). Thus for all 0 ≤ p < s, ∆t fp (t) = ∆t Ap (t) + ∆t bp (t) =
Ap (t) − Ap (t − 1) + bp (t) − bp (t − 1).
The sequence {ri }i∈Z+ satisfies ri+1 − (r0 + · · · + ri ) ≥ i. Hence ri+1 > r0 + · · · + ri + 2s
for all i sufficiently large. Then by Lemma 4.7, for each 0 ≤ p < s, ∆t fp (t) has infinitely
many zeroes in N. Therefore for each 0 ≤ p < s, there is an infinite increasing sequence
{tl }l∈N ⊂ N such that such that r0 + · · · + rl < p + (tl − 1)s < p + tl s < rl+1 and
∆t Ap (tl ) + ∆t bp (tl ) = 0; i.e.: ∆t Ap (tl ) = −∆t bp (tl ). Hence the sequence {∆t Ap (tl )}l∈N is
bounded (by −2M and 2M ).
11
For each 0 ≤ p < s, ∆t Ap (t) is a polynomial, and is a continuous function on R. If
∆t Ap (t) is nonconstant, then either liml→∞ ∆t Ap (tl ) = ∞ or liml→∞ ∆t Ap (tl ) = −∞.
Thus for each 0 ≤ p < s, ∆t Ap (t) is constant, and therefore Ap (t) is either constant or
linear. For 0 ≤ p < s, write Ap (t) = mp t + cp .
We cannot have mp ≤ 0 for any 0 ≤ p < s. By way of contradiction, suppose
that mp′ ≤ 0. Recall that bp′ (t) = fp′ (t) − mp′ t − cp′ . By Lemma 4.7, and since
ri+1 > r0 + · · · + ri + s for all i sufficiently large, we can find an increasing sequence
{tk }k∈N such that r0 + · · · + rk < p′ + tk s < rk+1 and fp′ (tk ) = 2k hold for all k ≫ 0.
Then for all k ≫ 0, we have bp′ (tk ) = 2k − mp′ tk + cp′ , and so letting k → ∞, we get that
bp′ (tk ) → ∞, contradicting that bp′ (t) was bounded. Thus, all the mp > 0 for 0 ≤ p < s.
Let 0 ≤ p < s be given. Then bp (t) = fp (t) − mp t − cp for t ∈ Z+ . Consider the
subsequence {is,k }k∈N such that is,k = s(k + 1) + 1. Then by Lemma 4.10, ris,k +1 − (r0 +
· · · + ris,k ) > s(k + 1). For each k ∈ N, define tk,max = max{t ∈ Z+ | p + st < ris,k +1 } and
tk,min = min{t ∈ Z+ | p + st > r0 + · · · + ris,k }. Then fp (t) = 2is,k for tk,min ≤ t ≤ tk,max
by Lemma 4.7, and tk,max − tk,min ≥ k. Thus for all k ∈ N,
bp (tk,max ) − bp (tk,min ) = mp (tk,min − tk,max ) ≤ −kmp .
Let N ∈ N be arbitrary and let k > mNp . Then bp (tk,max ) − bp (tk,min ) < −N . Now taking
N > 2M contradicts that −M ≤ b(n) ≤ M was bounded, since if −M ≤ b(n) ≤ M , then
−2M ≤ b(n1 ) − b(n0 ) ≤ 2M for any n0 , n1 ∈ Z+ .
Thus we see that for this choice of the sequence {ri }i∈Z+ , α(n) cannot be written as a
quasi-polynomial plus a bounded function. This proves the theorem.
Corollary 4.12. There exists a regular local ring (R, mR ) of dimension 2 and a discrete,
rank 1 valuation ν of the quotient field of R dominating R, such that the function ℓR (R/In )
cannot be written as a quasi-polynomial plus a bounded function for large integers n.
Proof. Let R and ν be as in the proof of Theorem 4.11, and suppose that ℓR (R/In ) =
P
l
Q(n) + σ(n) for n ≫ 0, where Q(n) = l=d
l=0 αl (n)n is a quasi-polynomial and σ(n) is a
bounded function, say |σ(n)| ≤ M . Then we have that:
ℓR (In /In+1 ) = ℓR (R/In+1 ) − ℓR (R/In ) = (Q(n + 1) − Q(n)) + (σ(n + 1) − σ(n))
Let s be the common integral period of the coefficients {αl (n)}0≤l≤d of Q(n). Then
p=d
l=d
X
X
l
Q(n + 1) − Q(n) =
αl (n + 1) − αp (n) np
p
p=0
l=p
is also a quasi-polynomial with coefficients of integral period s. Further, −2M ≤ σ(n +
1) − σ(n) ≤ 2M is a bounded function. Thus ℓR (In /In+1 ) is a quasi-polynomial plus a
bounded function, contradicting Theorem 4.11.
Proposition 4.13. Let α(n) = ℓR (In /In+1 ). Then the following recursive relation holds:
for any n ∈ Z>0 , consider the unique i ∈ Z≥0 such that n ∈ [ri , ri+1 ). Then we have that
α(n) = α(ri − 1) + min{α(ri − 1), α(n − ri )}.
12
Proof. Let n ∈ Z≥0 . Recall from Lemma 4.3 that a k-basis for In /In+1 is given by the set
Bn = {[P0n0 P1n1 . . . Pini ] | i ≥ 0, nk ∈ Z≥0 for k ≥ 0, nk ∈ {0, 1} for 1 ≤ k ≤ i, n0 +
i
X
nk rk = n}.
k=1
Thus we may canonically identify the set Bn with the set of tuples
(7)
Bn ∼
= {(n0 , n1 , . . . , ni ) ∈ N × {0, 1}i | n0 +
i
X
nk rk = n}.
k=1
where i is the unique j ∈ Z≥0 such that n ∈ [rj , rj+1 ).
Define Bn0 = {(n0 , n1 , . . . , ni ) ∈ Bn | ni = 0} and Bn1 = {(n0 , n1 , . . . , ni ) ∈ Bn | ni = 1}.
0
0 | = |B
Let us first show that |Bm
ri −1 |. Suppose that (n0 , n1 , . . . , ni ) ∈ Bn . Then
Pi−1
Pi−1
Pi−1
n0 + j=1 nj rj = n. By Lemma 4.5,
j=1 nj rj ≤
j=1 rj < ri ≤ n, hence n0 =
Pi
n − j=1 nj rj > n − ri ≥ 0 and so n0 − n + ri − 1 ≥ 0. Let us define a mapping
λ : Bn0 → Bri −1 by λ((n0 , n1 , . . . , ni−1 , ni )) = (n0 − n + ri − 1, n1 , . . . , ni−1 ). Notice that λ
is injective. For, suppose that λ((n0 , n1 , . . . , ni−1 , ni )) = (n0 − n + ri − 1, n1 , . . . , ni−1 ) =
(n′0 − n + ri − 1, n′1 , . . . , n′i−1 ) = λ((n′0 , n′1 , . . . , n′i−1 , n′i )). By (7) for the case m = ri − 1
we have that nj = n′j for 1 ≤ j ≤ i, whence n0 = n′0 and λ is injective. Furthermore, if (n0 , n1 , . . . , ni−1 ) ∈ Bri −1 , then (n0 + n − ri + 1, n1 , . . . , ni−1 , 0) ∈ Bn and
λ((n0 + n − ri + 1, n1 , . . . , ni−1 , 0)) = (n0 , n1 , . . . , ni−1 ), whence λ is surjective. Thus
λ gives a bijection between the finite sets Bn0 and Bri −1 , and |Bn0 | = |Bri −1 |.
Next, we shall show that
|Bn1 | =
|Bri −1 |
|Bn−ri |
if n − ri ≥ ri
if n − ri < ri
P
To this end, suppose that (n0 , n1 , . . . , ni−1 , ni = 1) ∈ Bn1 . Then n0 + i−1
j=1 rj nj = n − ri .
1
So, we have a (injective) mapping µ : Bn → Bn−ri , given by µ((n0 , n1 , . . . , ni−1 , ni = 1)) =
(n0 , n1 , . . . , ni−1 ). We claim that if n − ri < ri , then µ is a bijection between Bn1 and Bn−ri ,
0
.
and if n − ri ≥ ri , then µ is a bijection between Bn1 and Bn−r
i
P
Suppose that n − ri < ri . Then in any representation n − ri = n0 + ij=1 rj nj as in (7),
we must have ni = 0. It follows that µ is onto. Similarly, if n − ri ≥ ri , we have that µ
0
0
maps bijectively onto Bri −1 via the map
since n − ri < ri+1 . Further, Bn−r
is onto Bn−r
i
i
(n0 , n1 , . . . , ni−1 , ni = 0) → (n0 − n + 2ri − 1, n1 , . . . , ni−1 ).
Notice next that by Lemma 3.1, if n − ri ≥ ri , then |Bn−ri | ≥ |Bri −1 |, and if n − ri < ri ,
then |Bn−ri | ≤ |Bri −1 |. Thus we obtain that |Bn1 | = min{|Bn−ri |, |Bri −1 |}.
Hence the formula α(n) = α(ri − 1) + min{α(n − ri ), α(ri − 1)} holds.
Lemma 4.14. For all i ≥ 2, the following holds:
13
1. If θ < ∞, then
α(ri −1)
ri
2. If θ = ∞, then
α(ri −1)
ri
1
= 2+θ
+ εi , where 0 < εi :=
Pi
every i ≥ 1, Ri := θ − j=1 ej 2−j .)
= εi , where 0 < εi :=
Ri +21−i
.
(2+θ)[2+θ−Ri −21−i ]
(Recall that for
1
.
1+i−21−i
Proof. Suppose first that θ < ∞. Observe that
α(ri − 1)
2i−1
,
=
P
ri
2i − 1 + ij=2 ej 2i−j
by Lemma 4.7.
Hence,
α(ri − 1)
=
ri
2−
1
2i−1
+
1
Pi
1−j
j=2 ej 2
,
1
,
2 − 2i−1 + (θ − Ri )
1
,
=
Ri
1
(2 + θ)(1 − (2+θ)2
i−1 − 2+θ )
=
=
(1 −
1
1
(2+θ)2i−1
−
Ri
2+θ )
1
+ ( (2+θ)2
i−1 +
1
(2+θ)2i−1
(2 + θ)(1 −
−
Ri
2+θ )
Ri
2+θ )
,
1
+ εi .
2+θ
Lastly, observe that for each i ≥ 2, 0 ≤ θ − Ri . Hence −21−i ≤ θ − Ri − 21−i , and
2− 21−i ≤ 2+ θ − Ri − 21−i . Hence 2+ θ − Ri − 21−i > 0. Further, Ri ≥ 0, so εi > 0 follows.
=
Next suppose that θ = ∞. Observe that
by Lemma 4.7.
α(ri − 1)
2i−1
,
=
P
ri
2i − 1 + ij=2 ej 2i−j
Hence,
α(ri − 1)
=
ri
2−
=
1
2i−1
+
1
Pi
1−j
j=2 ej 2
,
1
,
2 − 2i−1 + (i − 1)
1
=
1 ,
1 + i − 2i−1
1
= εi .
1
εi
Observe that ε2 = 25 > 0. Suppose that i ≥ 3 is given, and εi−1 > 0. We have that
1
1
1
= εi−1
+ (1 + 2i−1
> 0. Hence εi > 0 for i ∈ N by induction.
) > εi−1
14
Corollary 4.15. 1. Suppose that θ < ∞. Then for all i ≥ 2,
limi→∞ εi = 0, and so
limi→∞ α(rrii−1)
=
2. Suppose that θ = ∞. Then for all i ≥ 2,
so
limi→∞ α(rrii−1)
1
2+θ .
α(ri −1)
ri
α(ri −1)
ri
>
1
2+θ .
Furthermore,
> 0. Furthermore, limi→∞ εi = 0, and
= 0.
Proof. Suppose that θ < ∞. Then the first claim follows since εi > 0 and the second claim
follows since 0 ≤ Ri < 21−i .
Next, suppose that θ = ∞. Then the first claim follows since εi > 0 and the second
1
claim follows since 1 + i − 2i−1
> i.
1
)n for all n ≥ 0.
Lemma 4.16. 1. Suppose that θ < ∞. Then we have that α(n) > ( 2+θ
2. Suppose that θ = ∞. Then we have that
α(n)
n
> 0 for all n ≥ 1.
Proof. First suppose that θ < ∞.
1
Observe first that if n = 0, then α(0) = 1 > ( 2+θ
)0, and if n = 1, then α(1) = 2 > 1 >
α(n)
1
2
2
2+θ (1). If n ∈ [r0 + r1 , r2 ) = [2, r2 ), then n = n ≥ r2
2
2
2
1
so α(n)
n ≥ 3+e2 ≥ 3+2θ > 4+2θ = 2+θ . Hence the result
=
2
3+e2 .
By construction, e2 ≤ 2θ,
holds for all n ∈ [0, r2 ).
Let n ≥ 2 be given, and consider the unique i ∈ Z≥1 such that n ∈ [ri , ri+1 ). We will
prove the result by induction on i. The case i = 1 has already been verified. Assume then
that i ≥ 2. We will use the formula
α(n) = α(ri − 1) + min{α(ri − 1), α(n − ri )},
and the observation that if
(8)
and
α(ri − 1) >
1
2+θ
ri
1
(9)
min{α(ri − 1), α(n − ri )} ≥
2+θ
then upon adding these two inequalities, we obtain
1
n.
α(n) >
2+θ
15
(n − ri ),
Observe that we already have (8) from Corollary 4.15.
We now check (9).
the inductive statement holds: that ri ≤ n < ri+1 , i ≥ 2,
Assume
1
and that α(m) > 2+θ m for all 0 ≤ m < ri .
If n − ri < ri , then by Lemma 3.1,
min{α(ri − 1), α(n − ri )} = α(n − ri ) >
1
2+θ
(n − ri ),
so (9) holds and we are done.
Suppose then that n − ri ≥ ri , in which case min{α(ri − 1), α(n − ri )} = α(ri − 1) by
Lemma 3.1. Observe that since ri ≤ n < ri+1 , we have 0 ≤ n − ri ≤ ri+1 − 1 − ri , so that
α(ri −1)
α(ri −1)
n−ri ≥ ri+1 −1−ri . We have
i
ri+1 − 1 − ri = 2 − 1 +
so
It follows that
i
X
ej 2i−j + ei+1 ,
j=2
α(ri − 1)
2i−1
.
≥
P
n − ri
2i − 1 + ij=2 ej 2i−j + ei+1
α(ri − 1)
1
≥
n − ri
2 + θ − Ri +
(ei+1 −1)
2i−1
−1
−1
1
≤ 0, or equivalently, ei+1
≤ Ri . But − 2i−1
≤
and it suffices to show that −Ri + ei+1
2i−1
2i−1
ei+1 −1
1−i
≤ 0 since ei+1 ∈ {0, 1} for i ≥ 2, and 0 ≤ Ri ≤ 2
by construction. Hence
2i−1
α(ri − 1)
1
,
≥
n − ri
2+θ
or
min{α(ri − 1), α(n − ri )} = α(ri − 1) ≥
and (9) holds.
1
2+θ
(n − ri )
Thus the result follows.
Finally, the result follows in the case θ = ∞ since for all n ≥ 1, we have α(n) ≥ 1, and
1
so α(n)
n ≥ n > 0.
Lemma 4.17. Let α(n) = ℓR (In /In+1 ) and let ε > 0 be given.
1. Suppose θ < ∞. Then there exists N (ε) such that for all i ≥ N (ε) and all n ∈ [ri , ri+1 ),
1
+ ε)n + α(rN (ε) − 1).
we have that α(n) ≤ ( 2+θ
2. Suppose θ = ∞. Then there exists N (ε) such that for all i ≥ N (ε) and all n ∈ [ri , ri+1 ),
we have that α(n) ≤ εn + α(rN (ε) − 1).
16
Proof. We first verify the result when θ < ∞.
We will use the formula
α(n) = α(ri − 1) + min{α(ri − 1), α(n − ri )},
and the observation that if
(10)
α(ri − 1) ≤
and
(11)
min{α(ri − 1), α(n − ri )} ≤
1
+ ε ri
2+θ
1
+ ε (n − ri ) + α(rN (ε) − 1),
2+θ
then upon adding these two inequalities, we obtain
1
α(n) ≤
+ ε n + α(rN (ε) − 1).
2+θ
Observe that by Corollary 4.15, for every ε > 0, there exists N (ε) such that for all i ≥
1
+ ε,
N (ε), εi < ε, where εi is as in Lemma 4.14. Thus for i ≥ N (ε), we have α(rrii−1) < 2+θ
or
(12)
α(ri − 1) <
1
+ ε ri .
2+θ
Let i ≥ N (ε) and suppose that n − ri ≥ ri . Then
(13)
min{α(ri − 1), α(n − ri )} = α(ri − 1) <
1
1
+ ε ri ≤
+ ε (n − ri ).
2+θ
2+θ
Thus adding (12) and (13) gives that
1
1
α(n) ≤
+ε n≤
+ ε n + α(rN (ε) − 1)
2+θ
2+θ
and hence the result for this case.
Suppose next that i ≥ N (ε) and n − ri < ri . We will prove this case by an inductive
argument.
First assume that i = N (ε) for the base case. Then
(14)
1
+ ε (n − ri ),
min{α(ri − 1), α(n − ri )} = α(n − ri ) ≤ α(rN (ε) − 1) ≤ α(rN (ε) − 1) +
2+θ
17
whence adding (12) and (14) gives that
1
+ ε n + α(rN (ε) − 1)
α(n) ≤
2+θ
and the result for the base case.
Next, assume the inductive statement: that n − ri < ri , thati > N (ε) and that for all
1
N (ε) ≤ j < i, and all m ∈ [rj , rj+1 ), we have α(m) ≤ 2+θ
+ ε m + α(rN (ε) − 1). Since
n − ri ∈ [rj , rj+1 ) for some j < i, we have by the inductive statement if j ≥ N (ε),
(15)
min{α(ri − 1), α(n − ri )} = α(n − ri ) ≤
1
+ ε (n − ri ) + α(rN (ε) − 1),
2+θ
and if j < N (ε), then α(n − ri ) ≤ α(rN (ε) − 1), and (15) still holds.
Thus adding (12) and (15) gives that
1
+ ε n + α(rN (ε) − 1)
α(n) ≤
2+θ
and the result for the inductive step.
Hence
α(n) = α(ri − 1) + min{α(ri − 1), α(n − ri )} ≤
for n ≥ rN (ε) .
1
+ ε n + α(rN (ε) − 1)
2+θ
Next, we verify the result when θ = ∞.
We will use the formula
α(n) = α(ri − 1) + min{α(ri − 1), α(n − ri )},
and the observation that if
(16)
and
(17)
α(ri − 1) ≤ εri
min{α(ri − 1), α(n − ri )} ≤ ε(n − ri ) + α(rN (ε) − 1),
then upon adding these two inequalities, we obtain
α(n) ≤ εn + α(rN (ε) − 1).
18
Observe that by Corollary 4.15, for every ε > 0, there exists N (ε) such that for all
i ≥ N (ε), εi < ε, where εi is as in Lemma 4.14. Thus for i ≥ N (ε), we have α(rrii−1) < ε,
or
(18)
α(ri − 1) < εri .
Let i ≥ N (ε) and suppose that n − ri ≥ ri . Then
(19)
min{α(ri − 1), α(n − ri )} = α(ri − 1) < εri ≤ ε(n − ri ).
Thus adding (18) and (19) gives that
α(n) ≤ εn ≤ εn + α(rN (ε) − 1)
and hence the result for this case.
Suppose next that i ≥ N (ε) and n − ri < ri . We will prove this case by an inductive
argument.
First assume that i = N (ε) for the base case. Then
(20)
min{α(ri − 1), α(n − ri )} = α(n − ri ) ≤ α(rN (ε) − 1) ≤ α(rN (ε) − 1) + ε(n − ri ),
whence adding (18) and (20) gives that
and the result for the base case.
α(n) ≤ εn + α(rN (ε) − 1)
Next, assume the inductive statement: that n − ri < ri , that i > N (ε) and that for
all N (ε) ≤ j < i, and all m ∈ [rj , rj+1 ), we have α(m) ≤ εm + α(rN (ε) − 1). Since
n − ri ∈ [rj , rj+1 ) for some j < i, we have by the inductive statement if j ≥ N (ε),
(21)
min{α(ri − 1), α(n − ri )} = α(n − ri ) ≤ ε(n − ri ) + α(rN (ε) − 1),
and if j < N (ε), then α(n − ri ) ≤ α(rN (ε) − 1), and (21) still holds.
Thus adding (18) and (21) gives that
α(n) ≤ εn + α(rN (ε) − 1)
and the result for the inductive step.
19
Hence
α(n) = α(ri − 1) + min{α(ri − 1), α(n − ri )} ≤ εn + α(rN (ε) − 1)
for n ≥ rN (ε) .
Corollary 4.18. Let α(n) = ℓR (In /In+1 ).
α(n)
n
limn→∞ α(n)
n
1. Suppose θ < ∞. Then limn→∞
2. Suppose θ = ∞. Then
=
1
2+θ .
= 0.
Proof. First suppose θ < ∞. Dividing both sides of the formula in Lemma 4.17(a) by n,
1
we see that for all ε > 0, we have that lim supn→∞ α(n)
n ≤ 2+θ + ε. Hence
lim sup
n→∞
α(n)
1
≤
.
n
2+θ
Now by Lemma 4.16(a), we see that
lim inf
n→∞
1
α(n)
≥
.
n
2+θ
Hence the result follows.
Next suppose θ = ∞. Dividing both sides of the formula in Lemma 4.17(b) by n, we
see that for all ε > 0, we have that lim supn→∞ α(n)
n ≤ ε. Hence
lim sup
n→∞
α(n)
≤ 0.
n
Now by Lemma 4.16(b), we see that
lim inf
n→∞
α(n)
≥ 0.
n
Hence the result follows.
Theorem 4.19. Let C ∈ R ∩ [0, 12 ] be given. There exists a regular local ring (R, mR )
of dimension 2 and a discrete, rank 1 valuation ν of the quotient field of R dominating
R, such that each of the functions ℓR (R/In ) and ℓR (In /In+1 ) is not a quasi-polynomial
plus a bounded function for large integers n and such that limn→∞ ℓR (Inn/In+1 ) = C and
n)
= C2 .
limn→∞ ℓR (R/I
n2
Proof. If 0 < C ≤ 12 , then let θ = C1 − 2 < ∞, and if C = 0, then let θ = ∞. In either case,
let {ej } be the associated sequence given by Definition 4.4, let the sequence {ri }i∈N be
given by (5), and let ν be the valuation associated to {ri } by Lemma 1. We have proved
the first claim in Corollary 4.12.
When 0 < C ≤ 12 , we have limn→∞
we have
limn→∞ α(n)
n
α(n)
n
=
1
2+θ
= C by Corollary 4.18(a). When C = 0,
= 0 = C by Corollary 4.18(b). Thus the second claim holds.
Pn−1
ℓR (Ij /Ij+1 ) = ℓR (R/I1 ) + j=1
ℓR (Ij /Ij+1 ). Hence
Pn−1
Pn−1
ℓR (R/I1 ) + j=1 ℓR (Ij /Ij+1 )
ℓR (R/I1 )
ℓR (R/In )
j=1 ℓR (Ij /Ij+1 )
=
=
+
.
n2
n2
n2
n2
Notice that ℓR (R/In ) =
Pn−1
j=0
20
By Lemma 5.1 of [6], limn→∞
ℓR (R/In )
n2
=
C
2
and the third claim holds.
Corollary 4.20. There exists a regular local ring (R, mR ) of dimension 2 and a discrete,
rank 1 valuation ν of the quotient field of R dominating R, such that each of the functions
ℓR (R/In ) and ℓR (In /In+1 ) is not a quasi-polynomial plus a bounded function for large
n)
are irrational
integers n and such that each of limn→∞ ℓR (Inn/In+1 ) and limn→∞ ℓR (R/I
n2
(even transcendental) positive numbers.
Proof. Since the proof of Theorem 4.19 does not depend on whether C > 0 is rational, we
may choose C ∈ (0, 12 ) \ Q in the first line of the proof of Theorem 4.19, and the result
follows.
References
[1] S. Abhyankar, ”On the valuations centered in a local domain”, Amer. J. Math. 78 (1956),
321-348.
[2] S. D. Cutkosky, ”Multiplicities associated to graded families of ideals”, Algebra and Number
Theory 7 (2013), 2059-2083,
[3] S. D. Cutkosky, ”Multiplicities of graded families of linear series and ideals.”,
arXiv:1301.5613 (2013).
[4] S. D. Cutkosky, ”Asymptotic Multiplicities”, arXiv:1311.1432v1 (2013).
[5] S. D. Cutkosky, V. Srinivas, ”On a problem of Zariski on dimensions of linear systems”,
Annals of Math. 137 (1993), 531-559.
[6] S. D. Cutkosky, K. Dalili, O. Kascheyeva, ”Growth of rank 1 valuation semigroups”, Communications in Algebra 38 (2010), 2768 - 2789.
[7] S. D. Cutkosky, P.A.Vinh, ”Valuation semigroups of two dimensional local rings”, Proceedings of the London Math. Soc. 108 (2014), 350-384.
[8] L. Ghezzi, H. T. Hà, O. Kashcheyeva, “Toroidalization of generating sequences in dimension
two function fields”, Journal of Algebra 301(2) (2006), 838-866.
[9] L. Ghezzi, O. Kashcheyeva, “Toroidalization of generating sequences in dimension two function fields of positive characteristic”, Journal of Pure and Applied Algebra 209(3) (2007),
631-649.
[10] M. Mustata,“On multiplicities of graded sequences of ideals”, Journal of Algebra 256 (2002),
229-249.
[11] M. Spivakovsky, “Valuations in function fields of surfaces,”, Amer. J. Math. 112 (1990),
107-156.
28 Mathematical Sciences Building, Department of Mathematics, University of Missouri,
Columbia, MO 65211, USA
E-mail address: [email protected]
21
| 0math.AC
|
SINGULAR DERIVED CATEGORIES OF Q-FACTORIAL
TERMINALIZATIONS AND MAXIMAL MODIFICATION ALGEBRAS
arXiv:1108.4518v3 [math.AG] 1 Apr 2014
OSAMU IYAMA AND MICHAEL WEMYSS
Abstract. Let X be a Gorenstein normal 3-fold satisfying (ELF) with local rings
which are at worst isolated hypersurface (e.g. terminal) singularities. By using the
singular derived category Dsg (X) and its idempotent completion Dsg (X), we give
necessary and sufficient categorical conditions for X to be Q-factorial and complete
locally Q-factorial respectively. We then relate this information to maximal modification algebras(=MMAs), introduced in [IW2], by showing that if an algebra Λ is
derived equivalent to X as above, then X is Q-factorial if and only if Λ is an MMA.
Thus all rings derived equivalent to Q-factorial terminalizations in dimension three
are MMAs. As an application, we extend some of the algebraic results in [BIKR] and
[DH] using geometric arguments.
Contents
1. Introduction
1
1.1. Overview
1
1.2. Results on Q-factoriality
2
1.3. Results on MMAs
3
1.4. Application to cAn singularities
3
1.5. Conjectures
4
2. Preliminaries
5
2.1. d-CY and d-sCY algebras
5
2.2. MM and CT modules
6
2.3. Homologically finite complexes
8
3. Singular derived categories and Q-factorial terminalizations
9
3.1. Singular derived categories of Gorenstein schemes with isolated singularities 9
3.2. Rigid-freeness and Q-factoriality
12
4. MMAs and NCCRs via tilting
13
4.1. The endomorphism ring of a tilting complex
13
4.2. Cohen–Macaulayness and crepancy
15
4.3. A counterexample in dimension four
18
5. cAn singularities via derived categories
19
5.1. Crepant modifications of cAn singularities
19
References
23
1. Introduction
1.1. Overview. The broader purpose of this work, in part a continuation of [IW2], is to
understand (and run) certain aspects of the minimal model program (=MMP) in dimension three using categorical and noncommutative techniques. As part of the conjectural
underlying picture, it is believed in dimension three (see 1.8 below) that the theory of
noncommutative minimal models (=MMAs below) should ‘control’ the geometry of commutative minimal models (=Q-factorial terminalizations) in the same way noncommutative crepant resolutions (=NCCRs) control the geometry of crepant resolutions [V04b]. If
this is to have any chance of being true, we must first be able to extend the well-known
characterization of smoothness by the singular derived category to be zero to a characterization of Q-factoriality. This involves moving from smooth schemes to singular ones,
1
2
OSAMU IYAMA AND MICHAEL WEMYSS
and it is unclear geometrically what form such a derived category characterization should
take.
However, from the noncommutative side, in the study of MMAs one homological
condition, which we call ‘rigid-freeness’, completely characterizes when ‘maximality’ has
been reached. The purpose of this paper is to show that this very same condition gives a
necessary condition for a scheme with isolated Gorenstein singularities to be Q-factorial.
If furthermore the singularities are hypersurfaces (e.g. Gorenstein terminal singularities)
this condition is both necessary and sufficient. The ability of the noncommutative side to
give new results purely in the setting of algebraic geometry adds weight to the conjectures
outlined in §1.5 below.
We now explain our results in more detail.
1.2. Results on Q-factoriality. We denote by CM R the category of CM (=maximal
Cohen-Macaulay) R-modules, and by CM R its stable category (see §2). As in [O03, O09],
we denote Dsg (X) := Db (coh X)/ per(X) (see §3.1 for more details). We call a functor
F : T → T ′ an equivalence up to direct summands if it is fully faithful and any object
s ∈ T ′ is isomorphic to a direct summand of F (t) for some t ∈ T . The first key technical
result is the following. Note that the latter part on the idempotent completion also follows
from work of Orlov [O09], and is well-known to experts [BK].
Theorem 1.1. (=3.2) Suppose that X is a Gorenstein scheme of dimension d satisfying
(ELF) (see 3.2 for full details and explanation), with isolated singularities {x1 , . . . , xn }.
Then there is a triangle equivalence
n
M
CM OX,xi
Dsg (X) ֒→
i=1
up to direct summands. Thus taking the idempotent completion gives a triangle equivalence
n
M
bX,xi .
CM O
Dsg (X) ≃
i=1
Recall that a normal scheme X is defined to be Q-factorial if for every Weil divisor
D, there exists n ∈ N for which nD is Cartier. If we can always take n = 1, we say that
X is locally factorial. Note that these conditions are Zariski local, i.e. X is Q-factorial
(respectively, locally factorial) if and only if OX,x is Q-factorial (respectively, factorial)
for all closed points x ∈ X. It is well–known (and problematic) that, unlike smoothness,
Q-factoriality cannot in general be checked complete locally. Thus we define X to be
bX,x
complete locally Q-factorial (respectively, complete locally factorial) if the completion O
is Q-factorial (respectively, factorial) for all closed points x ∈ X.
Now we recall a key notion from representation theory. Let T be a triangulated
category with a suspension functor [1]. An object a ∈ T is called rigid if HomT (a, a[1]) =
0. We say that T is rigid-free if every rigid object in T is isomorphic to the zero object.
Applying 1.1, we have the following result.
Theorem 1.2. (=3.10, 3.11) Suppose that X is a normal 3-dimensional Gorenstein
scheme over a field k, satisfying (ELF), with isolated singularities {x1 , . . . , xn }.
(1) If Dsg (X) is rigid-free, then X is locally factorial.
(2) If Dsg (X) is rigid-free, then X is complete locally factorial.
(3) If OX,xi are hypersurfaces for all 1 ≤ i ≤ n, then the following are equivalent:
(a) X is locally factorial.
(b) X is Q-factorial.
(c) Dsg (X) is rigid-free.
(d) CM OX,x is rigid-free for all closed points x ∈ X.
bX,xi are hypersurfaces for all 1 ≤ i ≤ n, then the following are equivalent:
(4) If O
(a)
(b)
(c)
(d)
X is complete locally factorial.
X is complete locally Q-factorial.
Dsg (X) is rigid-free.
bX,x is rigid-free for all closed points x ∈ X.
CM O
Q-FACTORIAL TERMINALIZATIONS AND MMAS
3
Thus the technical geometric distinction between a Q-factorial and complete locally
Q-factorial variety is explained categorically by the existence of a triangulated category
in which there are no rigid objects, but whose idempotent completion has many. Note
also that 1.2 demonstrates that by passing to the idempotent completion of Dsg (X) we
lose information on the global geometry (see e.g. 2.13), and so thus should be avoided.
1.3. Results on MMAs. The homological characterization of rigid-freeness in 1.2 originated in the study of modifying and maximal modifying modules (see 2.6 for definitions).
Recall that if M is a maximal modifying R-module, we say that EndR (M ) is a maximal
modification algebra (=MMA).
Proposition 1.3. (=2.14, 2.3) Suppose that R is normal, Gorenstein, equi-codimensional,
3-dimensional ring, and let M be a modifying R-module. Assume that EndR (M ) has only
isolated singularities. Then EndR (M ) is an MMA if and only if the category Dsg (EndR (M ))
is rigid-free.
Throughout this paper we will use the word variety to mean
• a normal integral scheme, of finite type over a field k, satisfying (ELF) (see §3.1).
Recall that if X and Y are varieties, then a projective birational morphism f : Y → X is
called crepant if f ∗ ωX = ωY (see §4 for more details). A Q-factorial terminalization of X
is a crepant projective birational morphism f : Y → X such that Y has only Q-factorial
terminal singularities. When Y is furthermore smooth, we call f a crepant resolution.
Maximal modification algebras were introduced in [IW2] with the aim of generalizing NCCRs to cover the more general situation when crepant resolutions do not exist.
Throughout this paper, we say that Y is derived equivalent to Λ if Db (coh Y ) is equivalent to Db (mod Λ). The following result ensures that all algebras derived equivalent to
a variety Y above are of the form EndR (M ), which relate the geometry to NCCRs and
MMAs.
Theorem 1.4. (=4.5) Let Y → Spec R be a projective birational morphism between ddimensional varieties. Suppose that Y is derived equivalent to some ring Λ. If Λ ∈ ref R,
then Λ ∼
= EndR (M ) for some M ∈ ref R.
Using this, we have the following result, which explains the geometric origin of the
definition of modifying modules:
Theorem 1.5. Let f : Y → Spec R be a projective birational morphism between ddimensional Gorenstein varieties. Suppose that Y is derived equivalent to some ring Λ,
then
(1) (=4.14) f is crepant ⇐⇒ Λ ∈ CM R.
(2) (=4.15) f is a crepant resolution ⇐⇒ Λ is an NCCR of R.
In either case, Λ ∼
= EndR (M ) for some M ∈ ref R.
We stress that 1.5(1) holds even if Y is singular. The following consequence of 1.2,
1.3 and 1.5 is one of our main results.
Theorem 1.6. (=4.16) Let f : Y → Spec R be a projective birational morphism, where
Y and R are both Gorenstein varieties of dimension three. Assume that Y has (at worst)
isolated singularities {x1 , . . . , xn } where each OY,xi is a hypersurface. If Y is derived
equivalent to some ring Λ, then the following are equivalent
(1) f is crepant and Y is Q-factorial.
(2) Λ is an MMA of R.
In this situation, all MMAs of R have isolated singularities, and are all derived equivalent.
As an important consequence, when k = C and Y has only terminal singularities and
derived equivalent to some ring Λ, then Y is a Q-factorial terminalization of Spec R if and
only if Λ is an MMA. See 4.17 for more details.
1.4. Application to cAn singularities. As an application of the above, we can extend some results of [BIKR] and [DH]. We let K denote an algebraically closed field of
characteristic zero, and suppose that f1 , . . . , fn ∈ m := (x, y) ⊆ K[[x, y]] are irreducible
polynomials, and let
R := K[[u, v, x, y]]/(uv − f1 . . . fn ).
4
OSAMU IYAMA AND MICHAEL WEMYSS
When K = C then R is a cAm singularity for m := ord(f1 . . . fn ) − 1 [BIKR, 6.1(e)].
Note that R is not assumed to be an isolated singularity, so f1 , . . . , fn are not necessarily
pairwise distinct. For each element ω in the symmetric group Sn , define
T ω := R ⊕ (u, fω(1) ) ⊕ (u, fω(1) fω(2) ) ⊕ . . . ⊕ (u, fω(1) . . . fω(n−1) ).
Our next result generalizes [BIKR, 1.5] and [DH, 4.2], and follows more or less immediately
from 1.6. This is stronger and more general since the results in [BIKR, DH] assume that
R is an isolated singularity, and only study cluster tilting objects.
Theorem 1.7. (=5.4) Let f1 , . . . , fn ∈ m := (x, y) ⊆ K[[x, y]] be irreducible polynomials
and R = K[[u, v, x, y]]/(uv − f1 . . . fn ). Then
(1) Each T ω (ω ∈ Sn ) is an MM R-module which is a generator. The endomorphism
rings EndR (T ω ) have isolated singularities.
(2) T ω is a CT R-module for some ω ∈ Sn ⇐⇒ T ω is a CT R-module for all ω ∈ Sn
⇐⇒ fi ∈
/ m2 for all 1 ≤ i ≤ n.
We remark that the proof of 1.7 also provides a conceptual geometric reason for the
condition fi ∈
/ m2 . We refer the reader to §5 for details.
1.5. Conjectures. One of the motivations for the introduction of MMAs is the following
conjecture, which naturally generalizes conjectures of Bondal–Orlov and Van den Bergh:
Conjecture 1.8. Suppose that R is a normal Gorenstein 3-fold over C, with only canonical (equivalently rational, since R is Gorenstein [R87, (3.8)]) singularities. Then
(1) R admits an MMA.
(2) All Q-factorial terminalizations of R and all MMAs of R are derived equivalent.
A special case of 1.8 is the long–standing conjecture of Van den Bergh [V04b, 4.6],
namely all crepant resolutions of Spec R (both commutative and non-commutative) are
derived equivalent. The results in this paper (specifically 1.6) add some weight to 1.8.
We note that 1.8 is true in some situations:
Theorem 1.9. Suppose that R is a normal Gorenstein 3-fold over C whose Q-factorial
terminalizations Y → Spec R have only one dimensional fibres (e.g. R is a terminal
Gorenstein singularity, or a cAn singularity in §5). Then Conjecture 1.8 is true.
Proof. The Q-factorial terminalizations are connected by a finite sequence of flops [K89],
so they are derived equivalent by [C02]. By [V04a] there is a derived equivalence between
Y and some algebra EndR (M ) with EndR (M ) ∈ CM R. Thus by 1.6 EndR (M ) is an
MMA. On the other hand, we already know that MMAs are connected by tilting modules
[IW2, 4.14], so they are all derived equivalent.
As it stands, the proof of 1.9 heavily uses the MMP. We believe that it should be
possible to prove 1.9 (for more general situations) without using the MMP, instead building
on the categorical techniques developed in this paper, and generalizing Van den Bergh’s
[V04a] interpretation of Bridgeland–King–Reid [BKR]. Indeed, it is our belief and longterm goal to show that many of the results of the MMP come out of our categorical picture
(along the lines of [B02], [BKR]), allowing us to both bypass some of the classifications
used in the MMP, and also run some aspects of the MMP in a more efficient manner.
Conventions. Throughout commutative rings are always assumed to be noetherian, and
R will always denote a commutative noetherian ring. All modules will be left modules, so
for a ring A we denote mod A to be the category of finitely generated left A-modules, and
Mod A will denote the category of all left A-modules. Throughout when composing maps
f g will mean f then g. Note that with these conventions HomR (M, X) is a EndR (M )module and HomR (X, M ) is a EndR (M )op -module. For M ∈ mod A we denote add M
to be the full subcategory consisting of direct summands of finite direct sums of copies
of M , and we denote proj A := add A to be the category of finitely generated projective
A-modules. We say that M ∈ mod R is a generator if R ∈ add M . For an abelian category
A, we denote by Db (A) the bounded derived category of A.
If X is a scheme, OX,x will denote the localization of the structure sheaf at the closed
bX,x the completion of OX,x at the unique maximal
point x ∈ X. We will denote by O
Q-FACTORIAL TERMINALIZATIONS AND MMAS
5
ideal. For us, locally will always mean Zariski locally, that is if we say that R has locally
only isolated hypersurface singularities, we mean that each Rm is an isolated hypersurface
singularity (or is smooth). When we want to discuss the completion, we will always refer
to this as complete locally.
Throughout, k will denote an arbitrary field. Rings and schemes will not be assumed
to be finite type over k, unless specified. When we say ‘over k’ we mean ‘of finite type
over k’.
Acknowledgements. Thanks are due to Vanya Cheltsov, Anne-Sophie Kaloghiros,
Raphael Rouquier and Ed Segal for many invaluable suggestions and discussions.
2. Preliminaries
For a commutative noetherian local ring (R, m) and M ∈ mod R, recall that the
depth of M is defined to be depthR M := inf{i ≥ 0 : ExtiR (R/m, M ) 6= 0}. We say that
M ∈ mod R is maximal Cohen-Macaulay (or simply, CM ) if depthR M = dim R. Now let
R be a (not necessarily local) commutative noetherian ring. We say that M ∈ mod R is
CM if Mp is CM for all prime ideals p in R, and we say that R is a CM ring if R is a CM
R-module. Denoting (−)∗ := HomR (−, R) : mod R → mod R, we say that X ∈ mod R is
reflexive if the natural map X → X ∗∗ is an isomorphism.
We denote ref R to be the category of reflexive R-modules, and we denote CM R to
be the category of CM R-modules. Throughout this section, we stress our convention
that commutative rings are always assumed to be noetherian, and R always denotes a
commutative noetherian ring.
2.1. d-CY and d-sCY algebras. When dealing with singularities, throughout this paper we use the language of d-sCY algebras, introduced in [IR, §3]. Although technical, it
provides the common language that links the commutative rings and the various noncommutative endomorphism rings that we will consider.
To do this, let R be a commutative ring with dim R = d and let Λ be a module-finite
R-algebra (i.e. an R-algebra which is a finitely generated
LR-module). For any X ∈ mod Λ,
denote by E(X) the injective hull of X, and put E := m∈Max R E(R/m). This gives rise
to Matlis duality D := HomR (−, E). Note that if R is over an algebraically closed field k,
then D coincides with Homk (−, k) on the category fl Λ by [O76, 1.1, 1.2]. We let Dbfl (Λ)
denote all bounded complexes with finite length cohomology.
Definition 2.1. For n ∈ Z, we call Λ n-Calabi-Yau (=n-CY) if there is a functorial
isomorphism
HomD(Mod Λ) (X, Y [n]) ∼
= D HomD(Mod Λ) (Y, X)
for all X ∈ Dbfl (Λ) and Y ∈ Db (mod Λ). Similarly we call Λ singular n-Calabi-Yau (=nsCY) if the above functorial isomorphism holds for all X ∈ Dbfl (Λ) and Y ∈ Kb (proj Λ).
We will use the following characterizations.
Proposition 2.2. (1) [IR, 3.10] A commutative ring R is d-sCY (respectively, d-CY) if
and only if R is Gorenstein (respectively, regular) and equi-codimensional, with dim R = d.
(2) [IR, 3.3(1)] Let R be a commutative d-sCY ring and Λ a module-finite R-algebra which
is a faithful R-module. Then Λ is d-sCY if and only if Λ ∈ CM R and HomR (Λ, R)m ∼
= Λm
as Λm -bimodules for all m ∈ Max R.
Suppose that R is a commutative d-sCY ring, and let Λ be a module-finite R-algebra
which is d-sCY. We say that X ∈ mod Λ is maximal Cohen-Macaulay (or simply, CM ) if
X ∈ CM R. We denote by CM Λ the category of CM Λ-modules. Then X ∈ mod Λ is CM
if and only if ExtiR (X, R) = 0 for all i ≥ 1, which is equivalent to ExtiΛ (X, Λ) = 0 for all
i ≥ 1 by the proof of [IR, 3.4(5)(i)]. We denote CM Λ to be the stable category, where we
factor out by those morphisms which factor through projective Λ-modules.
On the other hand, when Λ is a noetherian ring we denote
Dsg (Λ) := Db (mod Λ)/ Kb(proj Λ)
to be the singular derived category of Λ. Since any d-sCY algebra satisfies inj.dimΛ Λ = d
and inj.dimΛop Λ = d by [IR, 3.1(6)(2)], we have the following equivalence by a standard
theorem of Buchweitz [B86, 4.4.1(2)].
6
OSAMU IYAMA AND MICHAEL WEMYSS
Theorem 2.3. Suppose that R is a commutative d-sCY ring and Λ is a module-finite
R-algebra which is d-sCY, then there is a triangle equivalence
Dsg (Λ) ≃ CM Λ.
By 2.3 we identify Dsg (Λ) and CM Λ in the rest of this paper. Now we recall two
important notions.
Definition 2.4. [A84] Let R be a commutative d-sCY ring and Λ is a module-finite Ralgebra which is d-sCY.
(1) We say that Λ is non-singular if gl.dim Λ = d.
(2) We say that Λ has isolated singularities if gl.dim Λp = dim Rp for all non-maximal
primes p of R.
Proposition 2.5. Suppose that R is a commutative d-sCY ring and Λ is a module-finite
R-algebra which is d-sCY. Then
(1) Λ is non-singular if and only if CM Λ = 0.
(2) Λ has isolated singularities if and only if all Hom-sets in CM Λ are finite length Rmodules.
Proof. These are well-known (e.g. [Y90, 3.3], [A84]), but for convenience of the reader we
give the proof.
(1) See for example [IW2, 2.17(1)⇔(3)].
(2) (⇒) Since HomCM Λ (X, Y ) ∼
= Ext1Λ (Ω−1 X, Y ) and Ω−1 X ∈ CM Λ, we know that
SuppR HomCM Λ (X, Y ) consists of maximal ideals (e.g. [IW2, 2.6]). Thus the assertion
follows.
(⇐) Let p be a prime ideal of R with ht p < d and let X ∈ mod Λp . Certainly we can find
Y ∈ mod Λ with Yp ∼
= X and so consider a projective resolution
0 → K → Pd−1 → . . . → P1 → P0 → Y → 0
in mod Λ. Since Λ ∈ CM R, by localizing and using the depth lemma we see that K ∈
CM R, i.e. K ∈ CM Λ. Consider
0 → ΩK → Pd → K → 0,
(2.A)
∼
then since
= HomCM Λ (ΩK, ΩK) has finite length, it is supported only
on maximal ideals. Hence (2.A) splits under localization to p and so Kp is free. Thus
proj.dimΛp (X) < d. By Auslander–Buchsbaum, gl.dim Λp = dim Rp .
Ext1Λ (K, ΩK)
2.2. MM and CT modules. Recall that M ∈ ref R gives an NCCR Λ := EndR (M ) of
R if Λ ∈ CM R and gl.dim Λ = d [V04b]. In this case Λ is d-sCY by 2.8(2) below, and by
definition Λ is non-singular.
The following more general notions are quite natural from a representation theoretic
viewpoint.
Definition 2.6. [IW2] Let R be a d-sCY ring. Then
(1) M ∈ ref R is called a rigid module if Ext1R (M, M ) = 0.
(2) M ∈ ref R is called a modifying module if EndR (M ) ∈ CM R.
(3) We say M ∈ ref R is a maximal modifying (=MM) module if it is modifying and
further if M ⊕ Y is modifying for Y ∈ ref R, then Y ∈ add M . Equivalently,
add M = {X ∈ ref R | EndR (M ⊕ X) ∈ CM R}.
(4) We call M ∈ CM R a CT module if
add M = {X ∈ CM R | HomR (M, X) ∈ CM R}.
The following results will be used extensively. As in [IW2], if X ∈ mod R, we denote
fl X to be the largest finite length sub-R-module of X.
Lemma 2.7. [IW2, 2.7, 5.12] Suppose that R is a 3-sCY ring, and let M ∈ ref R.
(1) If M is modifying, then fl Ext1R (M, M ) = 0. The converse holds if M ∈ CM R.
(2) Assume that R is an isolated singularity. If M is modifying, then it is rigid. The
converse holds if M ∈ CM R.
Q-FACTORIAL TERMINALIZATIONS AND MMAS
7
Lemma 2.8. Suppose that R is a d-sCY normal domain, and M ∈ ref R. Let Λ =
EndR (M ). Then
(1) [IR, 2.4(3)] Λ ∼
= HomR (Λ, R) as Λ-bimodules.
(2) [IW2, 2.22(2)] M is modifying if and only if Λ is d-sCY .
We remark that the property CT can be checked complete locally [IW2, 5.5], that is
cm is a CT R
bm -module for all m ∈ Max R. In our
M is a CT R-module if and only if M
context in §3, this corresponds to the fact that a scheme X is non-singular if and only if
complete locally it is non-singular.
cm is
In contrast, the property of being MM cannot be checked complete locally. If M
b
an MM Rm -module for all m ∈ Max R, then M is an MM R-module. However the converse
is not true (see e.g. 2.13 below). This corresponds to the difference between Q-factorial
and complete locally Q-factorial singularities in §3.
We record here the following easy lemma, which shows the first link between modifying modules and factoriality.
Proposition 2.9. Suppose R is 3-sCY normal domain, with isolated singularities. Consider the statements
(1) R is an MM module.
(2) Every modifying R-module is projective.
(3) CM R is rigid-free.
(4) R is locally factorial.
(5) R is Q-factorial.
Then we have (1) ⇔ (2) ⇔ (3) ⇒ (4) ⇒ (5).
Proof. Since R is an isolated singularity, modifying modules are rigid by 2.7(2).
(1)⇒(2) Let M ∈ ref R be any modifying module. Since R is MM, there exists an exact
sequence
f
0 → F1 → F0 → M
with Fi ∈ add R such that f is a right (add R)-approximation, by [IW2, 4.12]. Then clearly
f is surjective, and so we have proj.dimR (M ) ≤ 1. By [AG, 4.10], if proj.dimR (M ) = 1
then Ext1R (M, M ) 6= 0, a contradiction. Hence M is projective.
(2)⇒(3) is clear, since HomCM R (M, M [1]) ∼
= Ext1R (M, M ).
(3)⇒(1) Suppose that EndR (R⊕X) ∈ CM R for some X ∈ ref R. Then X = HomR (R, X)
is a CM R-module which is a rigid object in CM R. Hence X is zero in CM R, i.e.
X ∈ add R.
(2)⇒(4) Since R is a normal domain, all members of the class group have R as endomorphism rings, and thus they are modifying R-modules. Hence every member of the class
group is projective, so the result follows.
(4)⇒(5) is clear.
Below we will use the following result from commutative algebra.
Theorem 2.10. [D1, 3.1(1)] Let S be a regular local ring of dimension four containing
a field k, and let R = S/(f ) be a hypersurface. If R is a Q-factorial isolated singularity,
then the free modules are the only modifying modules.
Proof. This is the equi-characteristic version of [D1, 3.1(1)], but since loc. cit. is in somewhat different language than used here, so we sketch Dao’s proof. Let N be a modifying
R-module. Since R is isolated, N is rigid by 2.7(2). On the other hand, by [D2], N is
a Tor-rigid R-module. These two facts imply, via a result of Jothilingham [J75, Main
Theorem] (see also [D1, 2.4]), that N is a free R-module.
Now we have the following result, which strengthens 2.9.
Theorem 2.11. Let R be a 3-sCY normal domain over k, which locally has only isolated
hypersurface singularities. Then the following are equivalent:
(1) R is an MM module.
(2) Every modifying R-module is projective.
(3) CM R is rigid-free.
8
OSAMU IYAMA AND MICHAEL WEMYSS
(4) R is locally factorial.
(5) R is Q-factorial.
Proof. By 2.9 we only have to show (5)⇒(1). Suppose that M ∈ ref R with EndR (R ⊕
M ) ∈ CM R. Then for any m ∈ Max R we have EndRm (Rm ⊕ Mm ) ∈ CM Rm with Rm
satisfying the assumptions of 2.10, hence Mm is a free Rm -module. Thus Mm ∈ add Rm
for all m ∈ Max R, so M ∈ add R [IW2, 2.26].
Remark 2.12. The corollary is false if we remove the isolated singularities assumption,
since then (4) does not necessarily imply (1). An example is given by R = C[x, y, z, t]/(x2 +
y 3 + z 5 ). Also, note that in the isolated singularity case it is unclear if the hypersurface
assumption is strictly necessary; indeed Dao conjectures that 2.10 still holds in the case of
complete intersections [D3, §4]. We remark that removing the hypersurface assumption
in 2.11 would allow us to bypass the use of the classification of terminal singularities later.
Example 2.13. Consider the element f = x2 + x3 + y 2 − uv in the ring C[u, v, x, y], and
bm is not an MM R
bm -module,
set R := C[u, v, x, y]/(f ). Then R is an MM R-module, but R
where m = (u, v, x, y).
√
Proof. Since R has an isolated singularity only at m := (u, v, x, y) and x + 1 does not
exist in Rm , we have that Rm is a factorial hypersurface singularity for all m ∈ Max R.
Hence if M is a modifying R-module, then Mm is a free Rm -module for any m ∈ Max R
by 2.10. Thus M is projective. √
bm is
For the last statement, since x + 1 exists (and is a unit) in the completion, R
bm ⊕ (u, x) is a modifying module.
isomorphic to C[[u, v, x, y]]/(uv − xy), for which R
Now we give a general categorical criterion for a given modifying module to be maximal, generalizing 2.9.
Proposition 2.14. Suppose that R is normal 3-sCY, let M be a modifying R-module and
set Λ := EndR (M ). If M is an MM R-module, then CM Λ is rigid-free. Moreover the
converse holds if Λ has isolated singularities.
Proof. By [IW2, 4.8(1)], M is an MM R-module if and only if there is no non-zero object
X ∈ CM Λ such that fl HomCM Λ (X, X[1]) = 0. This clearly implies CM Λ is rigid-free.
By 2.5(2), the converse is true if Λ has isolated singularities.
Remark 2.15. We conjectured in [IW2] that in dimension three MMAs always have
isolated singularities (see 4.16 for some evidence), so the key property from 2.14 is that
the stable category of CM modules is rigid-free. Note that Λ := EndR (M ) ∈ CM R is an
NCCR if and only if the stable category CM Λ is zero. Hence when passing from NCCRs
to MMAs, the categories Dsg (Λ) ≃ CM Λ are no longer zero, but instead rigid-free. This
motivates §3.
2.3. Homologically finite complexes. In this subsection we wish to consider the general setting of commutative rings, and so in particular the Krull dimension may at times be
infinite. The following is based on some Danish handwritten notes of Hans-Bjørn Foxby.
Theorem 2.16. Let R be a commutative ring (not necessarily of finite Krull dimension),
let Λ be a module-finite R-algebra, and let M ∈ mod Λ. If proj.dimΛp (Mp ) < ∞ for all
p ∈ Spec R, then proj.dimΛ (M ) < ∞.
Proof. Take a projective resolution
d
d
. . . → P1 →1 P0 →0 M → 0,
of M with finitely generated projective R-modules Pi , and set K0 := Im d0 = M and
Ki := Im di = Ker di−1 for i > 0. Now
proj.dimΛ (M ) < n ⇐⇒ ExtnΛ (M, Kn ) = 0,
and the right hand side is equivalent to the condition that SuppR ExtnΛ (M, Kn ) is empty.
Localizing at p ∈ Spec R, we have
proj.dimΛp (Mp ) ≥ n ⇐⇒ ExtnΛ (M, Kn )p = ExtnΛp (Mp , (Kn )p ) 6= 0
⇐⇒ p ∈ SuppR ExtnΛ (M, Kn ).
Q-FACTORIAL TERMINALIZATIONS AND MMAS
9
Since ExtnΛ (M, Kn ) is a finitely generated R-module, we have
SuppR ExtnΛ (M, Kn ) = V (AnnR (ExtnΛ (M, Kn ))).
For ease of notation set In = AnnR (ExtnΛ (M, Kn )), then
proj.dimΛp (Mp ) ≥ n ⇐⇒ p ∈ V (In ).
(2.B)
In particular, we have a decreasing sequence
V (I0 ) ⊇ V (I1 ) ⊇ V (I2 ) ⊇ . . . ,
so we have V (It ) = V (I
) = . . . for some t since the Spec R is noetherian by our assumpTt+1
∞
tion. If p ∈ V (It ) = n=t V (In ), then proj.dimΛp (Mp ) = ∞ by (2.B), a contradiction.
Consequently V (It ) = ∅, so we have ExttΛ (M, Kt ) = 0 and proj.dimΛ (M ) < t.
Corollary 2.17. Let R be a commutative ring and Λ a module-finite R-algebra. Then
(1) For every p ∈ Spec R, the algebra Λp has only finitely many simple modules.
(2) Let M ∈ mod Λ. Then proj.dimΛ (M ) < ∞ if and only if for all X ∈ mod Λ we have
ExtjΛ (M, X) = 0 for j ≫ 0.
Proof. (1) Λp is a module-finite Rp -algebra. There are only finitely many simple Λp modules since they are annihilated by pRp and hence they are modules over the finite
dimensional (Rp /pRp )-algebra Λp /pΛp, which has only finitely many simple modules.
(2) We only have to show ‘if’ part. For any p ∈ Spec R, we know that the localization
functor (−)p : mod Λ → mod Λp is essentially surjective, and that the completion functor
d : fl Λp → fl Λ
b p is an equivalence since the completion does not change finite length
(−)
bp is the sum of all simple Λ
b p -modules.
modules. By (1) there exists X ∈ mod Λ such that X
j
The assumptions imply that there exists t ≥ 0 such that ExtΛ (M, X) = 0 for all j > t.
cp , X
bp ) = 0 for all j > t. This implies proj.dim b (M
cp ) ≤ t since M
cp has a
Thus Extjb (M
Λp
Λp
minimal projective resolution. Thus we have proj.dimΛp (Mp ) ≤ t by applying [IW2, 2.26,
2⇔4] to the t-th syzygy of M . By 2.16, it follows that proj.dimΛ (M ) < ∞.
Let T be a triangulated category with a suspension functor [1]. Recall that x ∈ T is
called homologically finite if for all y ∈ T , HomT (x, y[i]) = 0 for all but finitely many i.
The following is well-known under more restrictive hypothesis:
Proposition 2.18. Let R be a commutative ring and Λ be a module-finite R-algebra.
Then the homologically finite complexes in Db (mod Λ) are precisely Kb (proj Λ).
Proof. Since every object in Kb (proj Λ) is homologically finite, we just need to show the
converse. If X is homologically finite, replace X by its projective resolution (P, d). The
truncation of P , namely τ≥i P , is quasi-isomorphic to X for small enough i. Fix such an
i. Thus, if we can show that Ker(di ) has finite projective dimension, certainly it follows
that X belongs to Kb (proj Λ).
Now by the usual short exact sequence given by truncation, there is a triangle
Q → X → Ker(di )[i] → Q[1]
where Q ∈ Kb (proj Λ). Since the first two are homologically finite, so is Ker(di ). In
particular, for all Y ∈ mod Λ, we have that ExtjΛ (Ker(di ), Y ) = 0 for large enough j. By
2.17(2), Ker(di ) has finite projective dimension, as required.
3. Singular derived categories and Q-factorial terminalizations
3.1. Singular derived categories of Gorenstein schemes with isolated singularities. Recall that a scheme X is called Gorenstein if OX,x is a Gorenstein local ring for all
closed points x ∈ X. We say that a scheme X satisfies (ELF) if X is separated, noetherian, of finite Krull dimension, such that coh X has enough locally free sheaves (i.e. for all
F ∈ coh X, there exists a locally free sheaf E and a surjection E ։ F ). This is automatic
in many situations, for example if X is quasi–projective over a commutative noetherian
ring [TT, 2.1.2(c), 2.1.3].
Definition 3.1. We say that F ∈ Qcoh X is locally free if X can be covered with open
sets U for which each F |U is a (not necessarily finitely generated) free OX |U -module.
10
OSAMU IYAMA AND MICHAEL WEMYSS
If F ∈ Qcoh X is locally free, then Fx is a free OX,x -module for all closed points
x ∈ X. The converse is true if F ∈ coh X.
As in [O03, §1] we consider
• the full triangulated subcategory per(X) of Db (coh X) consisting of objects which
are isomorphic to bounded complexes of finitely generated locally free sheaves in
Db (coh(X)).
• the full triangulated subcategory Lfr(X) of Db (Qcoh X) consisting of objects
which are isomorphic to bounded complexes of locally free sheaves in Db (Qcoh(X)).
Following Orlov [O03, O09], we denote
Dsg (X) := Db (coh X)/ per(X)
and
DSG (X) := Db (Qcoh X)/ Lfr(X).
Then the natural functor Dsg (X) → DSG (X) is fully faithful [O03, 1.13], and we regard
Dsg (X) as a full subcategory of DSG (X).
The aim of this subsection is to prove the following result, which plays a crucial role
in this paper. Note that part (2) also follows from work of Orlov [O09], and is well-known
to experts [BK].
Theorem 3.2. Suppose that X is a Gorenstein scheme of dimension d satisfying (ELF),
with isolated singularities {x1 , . . . , xn }.
Ln
Ln
(1) There is a triangle equivalence Dsg (X) → i=1 Dsg (OX,xi ) = i=1 CM OX,xi up to
summands, given by F 7→ (Fx1 , . . . , Fxn ).
Ln
bX,xi .
(2) Taking the idempotent completion gives a triangle equivalence Dsg (X) ≃ i=1 CM O
Since a key role in our proof of Theorem 3.2 is played by the category DSG (X), we
need to deal with infinitely generated modules. Let us start with recalling results by
Gruson–Raynaud [GR] and Jensen [J70].
Proposition 3.3. Let R be a noetherian ring of finite Krull dimension and M ∈ Mod R.
Then
(1) FPD(R) := sup{proj.dim(M ) | M ∈ Mod R and proj.dimR (M ) < ∞} = dim R.
(2) flat.dimR (M ) < ∞ implies proj.dimR (M ) < ∞.
Proof. (1) is due to Bass and Gruson–Raynaud [GR, II.3.2.6] (see also [F77, 3.2]).
(2) is due to Jensen [J70, Prop.6] (see also [F77, 3.3]).
We have the following immediate consequence.
Lemma 3.4. Let R be a noetherian ring of finite Krull dimension. Then for any M ∈
Mod R, proj.dimRm (Mm ) < ∞ for all m ∈ Max R ⇐⇒ proj.dimR (M ) ≤ dim R.
Proof. (⇐) is trivial
(⇒) The hypothesis implies that proj.dimRp (Mp ) < ∞ for all p ∈ Spec R. Thus by
3.3(1), we have proj.dimRp (Mp ) ≤ dim Rp ≤ dim R for all p ∈ Spec R. In particular
flat.dimRp (Mp ) ≤ dim R for all p ∈ Spec R. Now X ∈ Mod R is zero if Xp = 0 for
all p ∈ Spec R. This implies flat.dimR (M ) ≤ dim R since Tor groups localize also for
infinitely generated modules. By 3.3(2), M has finite projective dimension. Again by
3.3(1), we have proj.dimR (M ) ≤ dim R.
The following is easy:
Lemma 3.5. Let R be a commutative noetherian ring of finite Krull dimension, then
Kb (Free R) = Lfr(Spec R) = Kb (Proj R).
Proof. We have natural inclusions Kb (Free R) ⊆ Lfr(Spec R) ⊆ Kb (Proj R) since every
free module is clearly locally free, and further every locally free R-module is locally projective, thus projective by [GR, II.3.1.4(3)]. We only have to show Kb (Proj R) = Kb (Free R).
The proof is similar to [R89, 2.2] — by the Eilenberg swindle, if P ∈ Proj R, there exists F ∈ Free R such that P ⊕ F is free. Hence from every object of Kb (Proj R) we
can get to an object of Kb (Free R) by taking the direct sum with complexes of the form
0 → F → F → 0, which are zero objects in the homotopy category.
Corollary 3.6. Suppose that X satisfies (ELF). If F ∈ Qcoh X satisfies proj.dimOX,x (Fx ) <
∞ for all closed points x ∈ X, then F ∈ Lfr(X).
11
Q-FACTORIAL TERMINALIZATIONS AND MMAS
Proof. Since X satisfies (ELF), membership of Lfr(X) can be checked locally (see [O03,
1.7, proof of 1.14]). Thus we only have to show F |U ∈ Lfr(U ) for each affine open subset
U = Spec R of X. Since Fm is an Rm -module with finite projective dimension for all
m ∈ Max R, we have F |U ∈ Kb (Proj R) by 3.4. By 3.5, we have F |U ∈ Lfr(U ), so the
assertion follows.
We will also require the following well-known lemma, due to Orlov. If s is an object in
a triangulated category T , we denote thickT (s) to be the smallest triangulated subcategory
of T which is closed under direct summands and isomorphisms and contains s.
Lemma 3.7. (Orlov) Suppose that X satisfies (ELF), and has only isolated singularities
{x1 , . . . , xn }.LDenote the corresponding skyscraper sheaves by k1 , . . . , kn . Then Dsg (X) =
n
thickDsg (X) ( i=1 ki ).
Proof. By assumption the singular locus consists of a finite number of closed points, hence
is closed. It follows that Dsg (X) = thickDsg (X) (coh{x1 ,...,xn } X) by [C10, 1.2]. It is clear
Ln
that every coherent sheaf supported in {x1 , . . . , xn } belongs to thickDsg (X) ( i=1 ki ).
Now we are ready to prove 3.2.
gi
h
For each xi , consider a morphism fi := (Spec OX,xi → Spec Ri →i X) where Spec Ri is
some affine open subset of X containing xi . Let S := OX,x1 ⊕. . .⊕OX,xn and Y := Spec S.
Then the collection of the fi induce a morphism
`
f
Y = ni=1 Spec OX,xi → X.
The functor gi ∗ is just extension of scalars corresponding to a localization Ri → Rim , so
gi ∗ is exact and preserves quasi-coherence. Further hi is an affine morphism (since X is
separated), hence hi ∗ also is exact and preserves quasi-coherence. Hence each fi ∗ is exact
and preserves quasi-coherence. Thus we have an adjoint pair
f∗
Qcoh Y =
Qcoh X
n
L
Qcoh OX,xi
i=1
f∗
which are explicitly given by f ∗ F = (Fx1 , . . . , Fxn ) and f∗ (G1 , . . . , Gn ) =
These functors are exact, so induce an adjoint pair
f∗
Db (Qcoh Y ) =
Db (Qcoh X)
n
L
i=1
f∗
Ln
i=1
fi ∗ Gi .
Db (Qcoh OX,xi )
in the obvious way.
We will show f ∗ (Lfr(X)) ⊆ Lfr(Y ) and f∗ (Lfr(Y )) ⊆ Lfr(X), since this then induces
an adjoint pair
f∗
Ln
DSG (X)
DSG (Y ) = i=1 DSG (OX,xi ).
f∗
by [O03, 1.2]. It is clear that f ∗ takes locally free sheaves to projective modules, and so
f ∗ (Lfr(X)) ⊆ Lfr(Y ). On the other hand, each fi ∗ is an affine, flat morphism so fi ∗ OX,xi
is a flat OX -module. Further, each fi ∗ preserves sums, so takes projective OX,xi -modules
to flat OX -modules. Consequently, it follows that f∗ takes a projective S-module P to
a sheaf f∗ (P ) for which f∗ (P )x is a flat OX,x -module for all closed points x ∈ X. By
3.3(2) we have proj.dimOX,x (f∗ (P )x ) < ∞ for all closed points x ∈ X, hence by 3.6
f∗ (P ) ∈ Lfr(X) holds. Thus we have f∗ (Lfr(Y )) ⊆ Lfr(X).
Now denote the skyscraper sheaves in X corresponding to theL
singular points x1 , . . . , xn
n
by k1 , . . . , kn . By 3.7 we know that Dsg (X) = thickDsg (X) ( i=1 ki ). This implies
Ln
Dsg (X) ⊆ thickDSG (X) ( i=1 ki ).
Let α : 1 → f∗ ◦f ∗ be the unit. Since f∗ ◦f ∗ (ki ) = f∗ (0, . . . , 0, ki , 0, . . . , 0) = fL
i ∗ ki = ki
for all 1 ≤ i ≤ n, we have that α : 1 → f∗ ◦ f ∗ is an isomorphism on thickDSG (X) ( ni=1 ki ).
Ln
In particular f ∗ is fully faithful on thickDSG (X) ( i=1 ki ). Since f ∗ clearly takes Dsg (X)
L
to Dsg (Y ), and Dsg (X) ⊆ thickDSG (X) ( ni=1 ki ), we deduce that
f∗
Dsg (X) −→ Dsg (Y )
12
OSAMU IYAMA AND MICHAEL WEMYSS
L
is fully faithful. On the other hand, since Dsg (Y ) = thickDsg (Y ) (f ∗ ni=1 ki ) holds by 3.7,
we have Dsg (Y ) = thickDsg (Y ) (f ∗ (Dsg (X))). This immediately implies that f ∗ : Dsg (X) →
Dsg (Y ) is an equivalence up to direct summands (e.g. [N01, 2.1.39]).
The first statement (1) in the theorem now follows from the well-known equivalence
Dsg (OX,xi ) ≃ CM OX,xi due to Buchweitz [B86, 4.4.1]. The second statement (2) follows
bX,xi [KMV, A.1].
from (1), since the idempotent completion of CM OX,xi is CM O
The first corollary of 3.2 is the following alternative proof of [O03, 1.24], which we
will use later:
Corollary 3.8. Suppose that X is a d-dimensional Gorenstein scheme over k, satisfying
(ELF), with isolated singularities. Then all Hom-sets in Dsg (X) are finite dimensional
k-vector spaces.
Proof. By 3.2(1), it is enough to show that all Hom-sets in CM OX,xi are finite dimensional
k-vector spaces. Since each OX,xi is isolated, all Hom-sets in CM OX,xi are finite length
OX,xi -modules by 2.5(2). But each OX,xi is a localization of a finitely generated kalgebra, so by the Nullstellensatz its residue field is a finite extension of k. Thus the
assertion follows.
3.2. Rigid-freeness and Q-factoriality. In this section we give our main result which
characterizes the Q-factorial property in terms of the singular derived category, and then
relate this to MMAs.
Definition 3.9. Suppose that X is a Gorenstein scheme. We say that F ∈ coh X is a
Cohen–Macaulay (=CM) sheaf if Fx is a CM OX,x -module for all closed points x ∈ X.
Under the assumption that X is Gorenstein, F ∈ coh X is a CM sheaf if and only if
Exti (F , OX )x (= ExtiOX,x (Fx , OX,x )) = 0
for all closed points x ∈ X and all i > 0. Thus F ∈ coh X is a CM sheaf if and only if
Exti (F , OX ) = 0 for all i > 0.
The following is a generalization of 2.9(3)⇒(4).
Proposition 3.10. Suppose that X is a Gorenstein normal scheme satisfying (ELF), of
dimension three which has only isolated singularities {x1 , . . . , xn }. If Dsg (X) is rigid-free,
then X is locally factorial.
Proof. We need to show that any reflexive sheaf G on X of rank one is locally free. For
all closed points x ∈ X, we have a reflexive OX,x -module Gx of rank one. Since OX,x is a
normal domain, we have EndOX,x (Gx ) = OX,x . Since OX,x is an isolated singularity, we
have Ext1OX,x (Gx , Gx ) = 0 by 2.7(2).
Since X satisfies (ELF), there exists an exact sequence 0 → K → V → G → 0 with a
locally free sheaf V. Localizing at x, we have an exact sequence
0 → Kx → Vx → Gx → 0
(3.C)
of OX,x -modules with a projective OX,x -module Vx . Since Gx ∈ ref OX,x , we have Kx ∈
CM OX,x . Since Ext1OX,x (Gx , Gx ) = 0, we have EndOX,x (Kx ) ∈ CM OX,x by applying
[IW2, 4.10] to (3.C). Thus Ext1OX,x (Kx , Kx ) = 0 holds, again by 2.7(2).
Since Kx ∈ CM OX,x , we have by 3.2(1)
HomDsg (X) (K, K[1]) ∼
=
n
M
i=1
Ext1OX,xi (Kxi , Kxi ) = 0.
Since Dsg (X) is rigid-free by our assumption, we have K ∈ per(X). Thus Kx is a CM
OX,x -module of finite projective dimension. By the Auslander–Buchsbaum equality, Kx
has to be a projective OX,x -module. By (3.C), we have proj.dimOX,x Gx ≤ 1 for all x ∈ X.
If proj.dimOX,x Gx = 1, then Ext1OX,x (Gx , Gx ) 6= 0 by [AG, 4.10], a contradiction. Thus Gx
is a projective OX,x -module for all x ∈ X. Hence G is locally free.
Now we are ready to state our main result, which gives a relationship between factoriality of schemes and rigid-freeness of their singular derived categories.
Q-FACTORIAL TERMINALIZATIONS AND MMAS
13
Theorem 3.11. Suppose that X is a normal 3-dimensional Gorenstein scheme over k,
satisfying (ELF), with isolated singularities {x1 , . . . , xn }.
(1) If OX,xi are hypersurfaces for all 1 ≤ i ≤ n, then the following are equivalent:
(a) X is locally factorial.
(b) X is Q-factorial.
(c) Dsg (X) is rigid-free.
(d) CM OX,x is rigid-free for all closed points x ∈ X.
bX,xi are hypersurfaces for all 1 ≤ i ≤ n, then the following are equivalent:
(2) If O
(a)
(b)
(c)
(d)
X is complete locally factorial.
X is complete locally Q-factorial.
Dsg (X) is rigid-free.
bX,x is rigid-free for all closed points x ∈ X.
CM O
Proof. (1) (a)⇒(b) is clear. (b)⇒(d) follows from 2.11(5)⇒(3). (d)⇒(c) follows from
3.2(1). (c)⇒(a) follows from 3.10.
(2) The proof is identical to the proof in (1).
4. MMAs and NCCRs via tilting
The aim of this section is to give explicit information on rings that are derived
equivalent to certain varieties. Our main results (4.5, 4.14, 4.15 and 4.16) show a strong
relationship between the crepancy of birational morphisms and the property that the
rings are Cohen-Macaualy. In particular, we show that all algebras derived equivalent to
Q-factorial terminalizations in dimension three are MMAs.
Throughout this section, we continue to use the word variety to mean a normal
integral scheme, of finite type over a field k, satisfying (ELF). Some of the results below
remain true with weaker assumptions. If X is a variety, we denote DX := g ! k to be the
dualizing complex of X [H66, Ch V §10], where g : X → Spec k is the structure morphism.
If further X is CM of dimension d, we will always use ωX to denote DX [−d], and refer to
it as the geometric canonical. Thus, although canonical modules and dualizing complexes
are not unique, throughout this section ωX has a fixed meaning. Note that crepancy is
defined with respect to this canonical, i.e. f : Y → X is called crepant if f ∗ ωX = ωY .
4.1. The endomorphism ring of a tilting complex. We call V ∈ per(Y ) a pretilting
complex if HomD(Y ) (V, V[i]) = 0 for all i 6= 0, and a tilting complex if it is a pretilting
complex satisfying thickD(Y ) (V) = per(Y ). If Y is derived equivalent to Λ, there exists a
tilting complex V such that Λ ∼
= EndD(Y ) (V).
The following two results are well-known.
Lemma 4.1. Suppose that X is a scheme over a field k satisfying (ELF). Suppose that
X is derived equivalent to some ring Λ, then
(1) per(X) ≃ Kb (proj Λ).
(2) Dsg (X) ≃ Dsg (Λ).
Proof. (1) Clearly the derived equivalence Db (coh X) ≃ Db (mod Λ) induces an equivalence
between full subcategories of homologically finite complexes. We observed in 2.18 that
Kb (proj Λ) are precisely the homologically finite complexes in Db (mod Λ). On the other
hand, per(X) are precisely the homologically finite complexes in Db (coh X) by [O05, 1.11].
Thus the assertion follows.
(2) follows immediately from (1).
Corollary 4.2. Suppose that X is a scheme over a field k satisfying (ELF). Suppose that
X is derived equivalent to some ring Λ, which is a d-sCY algebra over a normal d-sCY
k-algebra. Then
(1) X is smooth if and only if Λ is non-singular (in the sense of 2.4).
(2) If X has only isolated singularities, then Λ has isolated singularities.
Proof. The derived equivalence induces Dsg (X) ≃ Dsg (Λ) by 4.1. It is very well-known
that X is smooth if and only if Dsg (X) = 0, so (1) follows from 2.5(1). (2) is immediate
from 3.8 and 2.5(2).
14
OSAMU IYAMA AND MICHAEL WEMYSS
In the rest of this subsection we suppose that we have a projective birational morphism f : Y → Spec R and that Y is derived equivalent to some ring Λ, and we investigate
when Λ has the form EndR (M ) for some M ∈ ref R. We need the following well-known
lemma.
Lemma 4.3. Let Y → Spec R be a projective birational morphism between d-dimensional
normal integral schemes. Let p ∈ Spec R and consider the following pullback diagram:
Y′
i
Y
g
Spec Rp
f
j
Spec R
(1) If p is a height one prime, then g is an isomorphism.
(2) If V is a pretilting complex of Y with Λ := EndD(Y ) (V), then Λ is a module-finite
R-algebra and i∗ V is a pretilting complex of Y ′ with Λp ∼
= EndY ′ (i∗ V).
Proof. (1) Since Y and Spec R are integral schemes and f is a projective birational morphism, it is well-known that f is the blowup of some ideal I of R [Liu, 8.1.24]. But
blowups are preserved under flat base change, and in fact g is the blowup at the ideal
Ip [Liu, 8.1.14]. Now since R is normal Rp is a discrete valuation ring, so every ideal is
principal. Thus g is the blowup of a Cartier divisor, and so is an isomorphism by the
universal property of blowing up.
(2) Since coh Y is an R-linear category, so is its derived category, and hence Λ has the
structure of an R-algebra. Since V is a pretilting complex, we have
Λ = RHomY (V, V) = Rf∗ RHomY (V, V).
(4.D)
Now since V is perfect, RHomY (V, V) is a bounded complex of coherent sheaves. Further,
since f is proper, f∗ preserves coherence and so Rf∗ : Db (coh Y ) → Db (mod R). Therefore
Rf∗ RHomY (V, V) = Λ is a finitely generated R-module.
Lastly, applying j ∗ to both sides of (4.D) gives
Λp
= j ∗ Rf∗ RHomY (V, V) = Rg∗ i∗ RHomY (V, V) = Rg∗ RHomY ′ (i∗ V, i∗ V)
= RHomY ′ (i∗ V, i∗ V),
where the second equality is flat base change, and the third holds since i is flat and V is
coherent [H66, II.5.8]. Thus we have the assertion.
We will also need the following result of Auslander–Goldman. A proof can be found
in [IW2, 2.11].
Proposition 4.4. [AG] Let R be a normal domain, and let Λ be a module-finite R-algebra.
Then the following conditions are equivalent:
(1) There exists M ∈ ref R such that Λ ∼
= EndR (M ) as R-algebras.
(2) Λ ∈ ref R and further Λp is Morita equivalent to Rp for all p ∈ Spec R with ht p = 1.
This leads to the following, which is an analogue of [IW2, 4.6(1)].
Theorem 4.5. Let Y → Spec R be a projective birational morphism between d-dimensional
normal integral schemes, and suppose that Y is derived equivalent to some ring Λ. Then
Λ is a module-finite R-algebra. If moreover Λ ∈ ref R, then Λ ∼
= EndR (M ) for some
M ∈ ref R.
Proof. Since Λ ∼
= EndD(Y ) (V) for some tilting complex V, it is a module-finite R-algebra
by 4.3(2). Now consider a height one prime p ∈ Spec R, then by base change we have a
pullback diagram
Y′
i
g
Spec Rp
Y
f
j
Spec R
6 0
By 4.3(1) g is an isomorphism, and by 4.3(2) Λp = RHomY ′ (i∗ V, i∗ V). Note that i∗ V =
since Λ ∈ ref R, so since R is normal necessarily Λ is supported everywhere. Since Rp is a
Q-FACTORIAL TERMINALIZATIONS AND MMAS
15
local ring, the only perfect complexes x with HomD(Rp ) (x, x[i]) = 0 for all i > 0 are shifts
of projective modules [RZ, 2.12]. Thus g∗ i∗ V ∼
= Rpa [b] for some a ∈ N and b ∈ Z. Hence
a
∼
Λp = EndRp (Rp ), which is Morita equivalent to Rp . This holds for all height one primes,
so by 4.4 the assertion follows.
4.2. Cohen–Macaulayness and crepancy. Let f : Y → Spec R be a projective birational morphism such that Y is derived equivalent to some ring Λ. In this section we show
that f is crepant if and only if Λ ∈ CM R. To do this requires the following version of
Grothendieck duality.
Theorem 4.6. Suppose that f : Y → Spec R is a projective morphism, then
RHomY (F , f ! G) = RHomR (RΓ F , G)
for all F ∈ D(Qcoh Y ), and G ∈ D + (Mod R).
Proof. The sheafified duality theorem [N96, §6, 6.3] reads
Rf∗ RHomY (F , f ! G) = RHomR (Rf∗ F , G).
In this situation, Rf∗ = RΓ, and since Spec R is affine, global and local hom agree.
The following is also well-known [V04a, 3.2.9].
Lemma 4.7. Suppose that f : Y → Spec R is a projective birational morphism, where Y
and R are both Gorenstein varieties of dimension d. Then
(1) f ! OR is a line bundle.
(2) f is crepant if and only if f ! OR = OY .
Proof. Since R is Gorenstein ωR is a line bundle and thus is a compact object in D(Mod R).
!
∗
!
Hence by [N96, p227–228] we have f ! ωR = Lf ∗ ωR ⊗L
Y f OR = f ωR ⊗Y f OR and so
ωY = DY [−d] = f ! DR [−d] = f ! ωR = f ∗ ωR ⊗Y f ! OR .
Since both ωY and f ∗ ωR are line bundles, f ! OR = (f ∗ ωR )−1 ⊗Y ωY is a line bundle.
Moreover f is crepant if and only if f ∗ ωR = ωY if and only if f ! OR = OY .
The following result show that crepancy implies that Λ is Cohen-Macaulay.
Lemma 4.8. Suppose that f : Y → Spec R is a crepant projective birational morphism,
where Y and R are both Gorenstein varieties of dimension d. Then EndD(Y ) (V) ∈ CM R
for any pretilting complex V of Y .
Proof. We have
∼
RHomY (V, V) ∼
= RHomY (V, V ⊗L
Y OY ) = RHomY (RHomY (V, V), OY ).
Applying RΓ to both sides, we have
RHomY (V, V)
∼
=
RHomY (RHomY (V, V), OY )
∼
=
RHomY (RHomY (V, V), f ! OR )
∼
=
RHomR (RHomY (V, V), R).
4.7
4.6
Since V is pretilting RHomY (V, V) = EndD(Y ) (V), so the above isomorphism reduces to
EndD(Y ) (V) = RHomR (EndD(Y ) (V), R).
Hence applying H i to both sides, we obtain ExtiR (EndD(Y ) (V), R) = 0 for all i > 0. Thus
EndD(Y ) (V) ∈ CM R, as required.
To show the converse of 4.8, namely Λ is Cohen–Macaulay implies crepancy, will
involve Serre functors. However, since we are in the singular setting these are somewhat
more subtle than usual. The following is based on [G06, 7.2.6].
16
OSAMU IYAMA AND MICHAEL WEMYSS
Definition 4.9. Suppose that Y → Spec R is a morphism where R is CM ring with a
canonical module CR . We say that a functor S : per(Y ) → per(Y ) is a Serre functor
relative to CR if there are functorial isomorphisms
RHomR (RHomY (F , G), CR ) ∼
= RHomY (G, S(F ))
in D(Mod R) for all F , G ∈ per(Y ). If Λ is a module-finite R-algebra, we define Serre
functor S : Kb (proj Λ) → Kb (proj Λ) relative to CR in a similar way.
Remark 4.10. We remark that since we are working in the non-local CM setting, canonical modules are not unique. Although crepancy is defined with respect to the geometric
canonical ωR , and [G06] defines Serre functors with respect to ωR , there are benefits of
allowing the flexibility of different canonical modules. Bridgeland–King–Reid [BKR, 3.2]
had technical problems when the geometric canonical is not trivial, since they first had
to work locally (where the canonical is trivial), and then extend this globally. Below, we
avoid this problem by considering Serre functors with respect to the canonical module R,
using the trick in 4.13(2).
The following two lemmas are standard.
Lemma 4.11. Suppose that S and T are two Serre functors relative to the same canonical
CR . Then S and T are isomorphic.
Proof. There are functorial isomorphisms
RHomY (G, S(F )) ∼
= RHomY (G, T(F ))
for all F , G ∈ per(Y ), which after applying H 0 give functorial isomorphisms
Homper(Y ) (−, S(−)) ∼
= Homper(Y ) (−, T(−))
Since S and T take values in per(Y ), we may use Yoneda’s lemma to conclude that S and
T are isomorphic.
Lemma 4.12. Suppose that Y → Spec R is a projective birational morphism between
varieties, and Y is derived equivalent to Λ. Then any Serre functor S : Kb (proj Λ) →
Kb (proj Λ) relative to CR induces a Serre functor S′ : per(Y ) → per(Y ) relative to CR
Proof. It is enough to show that we have a triangle equivalence F : per(Λ) → per(Y )
with a functorial isomorphism RHomΛ (A, B) ∼
= RHomY (F A, F B) in D(R) for all A, B ∈
per(Λ), since then for a quasi-inverse E of F we have that S′ := F ◦S◦E : per(Y ) → per(Y )
enjoys functorial isomorphisms
RHomR (RHomY (F A, F B), CR ) ∼
= RHomR (RHomΛ (A, B), CR )
∼
= RHomΛ (B, SA)
∼
= RHomY (F B, F (SA))
∼
= RHomY (F B, S′ (F A))
in D(R). Let V be a tilting complex of Y , and let A be the DG endomorphism R-algebra
of V .
(i) Let Cdg (A ) be the DG category of DG A -modules. We denote by pretr(A ) the
smallest DG subcategory of Cdg (A ) which is closed under [±1] and cones and contains
A . Since the DG category perdg (Y ) of perfect complexes of Y is pretriangulated, there
exists a fully faithful DG functor V ⊗A − : pretr(A ) → perdg (Y ) which induces a triangle
equivalence G : per(A ) → per(Y ) [K06, Section 4.5]. In particular, we have a functorial
isomorphism Hom•A (A, B) ∼
= Hom•Y (V ⊗A A, V ⊗A B) of DG R-modules for all A, B ∈
pretr(A ), where we denote by Hom•A and Hom•Y the Hom-sets in our DG categories.
Thus we have a functorial isomorphism RHomA (A, B) ∼
= RHomY (GA, GB) in D(R) for
all A, B ∈ per(A ).
(ii) Let f : A → B be a quasi-isomorphism of DG R-algebras. Then the DG functor
B⊗A − : pretr(A ) → pretr(B) gives a triangle equivalence H : per(A ) → per(B). For all
A, B ∈ pretr(A ), we have a quasi-isomorphism Hom•A (A, B) → Hom•B (B ⊗A A, B ⊗A B)
of DG R-modules. In particular, we have a functorial isomorphism RHomA (A, B) ∼
=
RHomB (HA, HB) in D(R) for all A, B ∈ per(A).
(iii) Since V is a tilting complex, we have quasi-isomorphisms A ≤0 → A and A ≤0 → Λ
Q-FACTORIAL TERMINALIZATIONS AND MMAS
17
of DG R-algebras, where A ≤0 is a DG sub R-algebra (· · · → A −1 → Ker d0 → 0 → · · · )
of A . Combining (i) and (ii), we have the desired assertion.
The following is our key observation.
Lemma 4.13. Suppose that f : Y → Spec R is a projective birational morphism, where
Y and R are both Gorenstein varieties.
(1) − ⊗Y f ! OR : per(Y ) → per(Y ) is a Serre functor relative to R.
(2) f is crepant if and only if id : per(Y ) → per(Y ) is a Serre functor relative to R.
Proof. (1) We know by 4.7 that f ! OR is a line bundle, so it follows that tensoring by f ! OR
gives a functor − ⊗Y f ! OR : per(Y ) → per(Y ). Further, we have functorial isomorphisms
4.6
RHomY (G, F ⊗Y f ! OR ) ∼
= RHomY (RHomY (F , G), f ! OR ) ∼
= RHomR (RHomY (F , G), OR )
in D(Mod R) for all F , G ∈ per(Y ). Thus − ⊗Y f ! OR is a Serre functor relative to R.
(2) By (1) and 4.11, id is a Serre functor relative to R if and only if − ⊗Y f ! OR = id as
functors per(Y ) → per(Y ). This is equivalent to f ! OR = OY , which is equivalent to that
f is crepant by 4.7.
The following explains the geometric origin of the definition of modifying modules,
and is the first main result of this subsection.
Theorem 4.14. Let f : Y → Spec R be a projective birational morphism between ddimensional Gorenstein varieties. Suppose that Y is derived equivalent to some ring Λ,
then the following are equivalent.
(1) f is crepant.
(2) Λ ∈ CM R.
(3) id : per(Y ) → per(Y ) is a Serre functor relative to R.
In this case Λ ∼
= EndR (M ) for some M ∈ ref R.
Proof. (3)⇔(1) is shown in 4.13(2), and (1)⇒(2) is shown in 4.8.
(2)⇒(3) Let
−
−
S := RHomR (Λ, R) ⊗L
Λ − : D (mod Λ) → D (mod Λ).
By [IR, 3.5(2)(3)], there exists a functorial isomorphism
∼ RHomR (RHomΛ (B, A), R)
RHomΛ (A, S(B)) =
(4.E)
in D(R) for all A ∈ Db (mod Λ) and all B ∈ Kb (proj Λ).
Now suppose that Λ ∈ CM R. Since it is reflexive, Λ ∼
= EndR (M ) for some M ∈ ref R
by 4.5. Then we have
Λ∼
= HomR (Λ, R) ∼
= RHomR (Λ, R)
in D(Λ ⊗R Λop ) where the first isomorphism holds by 2.8(1), and the second since Λ ∈
CM R. Thus we have S ∼
= id, and (4.E) shows that id : Kb (proj Λ) → Kb (proj Λ) is
a Serre functor relative to R. By 4.12 it follows that id : per(Y ) → per(Y ) is a Serre
functor relative to R.
When further Y is smooth, we can strengthen 4.14.
Corollary 4.15. Let f : Y → Spec R be a projective birational morphism between ddimensional Gorenstein varieties. Suppose that Y is derived equivalent to some ring Λ,
then the following are equivalent.
(1) f is a crepant resolution of Spec R.
(2) Λ is an NCCR of R.
Proof. By 4.14 f is crepant if and only if Λ ∈ CM R. Assume that this is satisfied, then
again by 4.14 Λ ∼
= EndR (M ) for some M ∈ ref R. By 4.2(1) Y is smooth if and only if Λ
is non-singular, which means that Λ is an NCCR of R.
We are now in a position to relate maximal modification algebras and Q-factorial
terminalizations in dimension three. The following is the second main result of this subsection.
18
OSAMU IYAMA AND MICHAEL WEMYSS
Theorem 4.16. Let f : Y → Spec R be a projective birational morphism, where Y and R
are both Gorenstein varieties of dimension three. Assume that Y has (at worst) isolated
singularities {x1 , . . . , xn } where each OY,xi is a hypersurface. If Y is derived equivalent
to some ring Λ, then the following are equivalent.
(1) f is crepant and Y is Q-factorial.
(2) Λ is an MMA of R.
In this situation, all MMAs of R are derived equivalent and have isolated singularities.
Proof. By 4.14 f is crepant if and only if Λ ∈ CM R. Assume that this is satisfied, then
again by 4.14 Λ ∼
= EndR (M ) for some M ∈ ref R. Thus Dsg (Y ) ≃ CM Λ by 4.1, 2.8(2)
and 2.3. Since the singularities of Y are isolated, Λ has isolated singularities by 4.2(2).
Now by 3.11(1), Y is Q-factorial if and only if Dsg (Y ) ≃ CM Λ is rigid-free, which by 2.14
holds if and only if Λ is an MMA. Lastly, all maximal modification algebras are derived
equivalent [IW2, 4.15], thus all are derived equivalent to Y and hence all have isolated
singularities by 4.2(2).
Remark 4.17. When k = C and Y has only terminal singularities, the geometric assumptions in 4.16 are satisfied since terminal singularities are isolated for 3-folds [KM,
5.18], and Gorenstein terminal singularities are Zariski locally hypersurfaces [R83, 0.6(I)].
This shows that when k = C and Y has only terminal singularities, Y is a Q-factorial
terminalization of Spec R if and only if Λ is an MMA.
Remark 4.18. We also remark that if Conjecture 1.8 is true, then the Y in 4.16 always
admit a tilting complex. Note that the corresponding statement of 1.8 in dimension four
is false, see 4.20.
Note that all the results in this subsection remain valid in the complete local setting.
In particular 4.19 below corresponds to 4.16, and will be used in §5. Here, a morphism
f : Y → X = Spec R is called crepant if f ∗ ωX = ωY holds (as before), where now X has
a unique canonical sheaf ωX = OR , and we choose the canonical sheaf ωY := f ! ωX on Y .
Corollary 4.19. Suppose f : Y → Spec R is a projective birational morphism between 3dimensional Gorenstein normal integral schemes, satisfying (ELF), where R is a complete
local ring containing a copy of its residue field. Suppose further that Y has (at worst)
isolated singularities {x1 , . . . , xn } where each OY,xi is a hypersurface. If Y is derived
equivalent to some ring Λ, then the following are equivalent.
(1) Y is complete locally Q-factorial and f is crepant.
(2) Y is Q-factorial and f is crepant.
(3) Λ is an MMA of R.
In this situation, all MMAs of R are derived equivalent and have isolated singularities.
Proof. (2)⇔(3) is similar to 4.16 (1)⇔(2).
(1)⇔(2) By 3.11(1)(2), we only have to show that Dsg (Y ) = Dsg (Y ) holds. By 4.14
Λ∼
= EndR (M ) ∈ CM R for some M ∈ ref R. We have Dsg (Y ) ≃ CM Λ by 4.1, 2.8(2) and
2.3. The endomorphism ring EndΛ (X) of any X ∈ mod Λ is again a module-finite algebra
over a complete local ring R. Thus any idempotent in EndΛ (X) is an image of some
idempotent in EndΛ (X) by [CR, 6.5, 6.7], which corresponds to some direct summand
of X. Therefore the category CM Λ ≃ Dsg (Y ) is idempotent complete, and we have the
assertion.
4.3. A counterexample in dimension four. Here we show that we cannot always
expect the setup of 4.16 and 4.15 to hold in higher dimension. This puts severe limitations
on any general homological theory that covers dimension four.
The following is an extension of an example of Dao [D1, 3.5].
Theorem 4.20. Let R := C[x0 , x1 , x2 , x3 , x4 ]/(x50 + x41 + x42 + x43 + x44 ). This has a crepant
resolution, which we denote by Y → Spec R. Then there is no algebra Λ that is derived
equivalent to Y .
Proof. Note first that R is an isolated singularity, with unique singular point at the origin.
A projective crepant resolution exists by [Lin, Thm. A.4]. Suppose Λ is derived equivalent
to Y , then by 4.15 Λ ∼
= EndR (M ) is an NCCR of R. Completing with respect to the
Q-FACTORIAL TERMINALIZATIONS AND MMAS
19
c) is an NCCR of R.
b But End b (M
c) ∈
maximal ideal of the origin, this implies that EndRb (M
R
b with M
c ∈ ref R,
b so since R
b is a 4-dimensional local isolated hypersurface, M
c must
CM R
c) ∼
b
be free [D1, 2.7(3)]. But this forces EndRb (M
M
(
R),
which
is
a
contradiction
since
= n
b
Mn (R) has infinite global dimension.
5. cAn singularities via derived categories
We now illustrate the results of the previous section to give many examples of maximal
modification algebras. The following result is due to Shepherd–Barron (unpublished).
Proposition 5.1. Let R = C[[u, v, x, y]]/(uv − f (x, y)) be an isolated cAn singularity.
Then the following are equivalent
(1) There does not exist a non-trivial crepant morphism Y → Spec R.
(2) f (x, y) is irreducible.
(3) R is factorial.
(4) R is Q-factorial.
(5) R is an MM R-module.
Proof. (1)⇒(2) If f (x, y) factors non-trivially as f = f1 f2 then blowing up the ideal (u, f1 )
yields a non-trivial crepant morphism (as in the calculation in §5.1). Hence f (x, y) cannot
factor.
(2)⇒(3) This is easy (see e.g. proof of [IW3, 5.9]).
(3)⇒(4) This is clear.
(4)⇒(5) This follows from 2.11(5)⇒(1).
(5)⇒(1) Since R is an isolated cDV singularity of type A (see e.g. [BIKR, 6.1(e)]), it
is a terminal singularity by [R83, 1.1]. If there exists a non-trivial crepant morphism
f : Y → Spec R, then f must have one-dimensional fibres. Hence by [V04a, 3.2.10] there
exists a non-projective modifying module. Since R is a MM R-module, this contradicts
2.9.
5.1. Crepant modifications of cAn singularities. In this section we work over K, an
algebraically closed field of characteristic zero. Let
R := K[[u, v, x, y]]/(uv − f (x, y)),
with f ∈ m := (x, y) ⊆ K[[x, y]]. We let f = f1 . . . fn be a factorization into prime
elements of K[[x, y]]. We restrict to this complete local setting since it simplifies the proof
of the main theorem 5.2. A similar version of 5.2 is true when R is not complete local,
though this requires a much more complicated proof [W12].
We remark that when K = C, R is a cAm singularity for m := ord(f ) − 1 (see e.g.
[BIKR, 6.1(e)]). This is terminal if and only if R is an isolated singularity [R83, 1.1] if
and only if (fi ) 6= (fj ) for all i 6= j.
For any subset I ⊆ {1, . . . , n} we denote
Y
fI :=
fi and TI := (u, fI )
i∈I
which is an ideal of R. For a collection of subsets
∅ ( I1 ( I2 ( ... ( Im ( {1, 2, ..., n},
we say that F = (I1 , . . . , Im ) is a flag in the set {1, 2, . . . , n}. We say that the flag F is
maximal if n = m + 1. Given a flag F = (I1 , . . . , Im ), we define
m
M
T F := R ⊕
TIj .
j=1
On the other hand, geometrically, given a flag F = (I1 , . . . , Im ) we define a scheme X F
as follows:
20
OSAMU IYAMA AND MICHAEL WEMYSS
First, let X F1 → Spec R be the blowup of the ideal (u, fI1 ) in R = K[[u, v, x, y]]/(uv −
f ). Then X F1 is covered by two open charts, given explicitly by
R1 :=
K[[u, v, x, y]][V1 ]
!
v − V1 ffI
1
uV1 − fI1
and R2 :=
K[[u, v, x, y]][U1 ]
!.
u − fI1 U1
U1 v − ffI
1
fI1
u
and U1 = fuI glue to give a copy of P1 inside X F1 (possibly
The new coordinates V1 =
1
moving in a family), which maps to the origin of Spec R.
f
Next, let X F2 → X F1 be the blowup of the divisor (U1 , fII2 ) in the second co-ordinate
1
f
chart R2 . Note that the zero set of the divisor (U1 , fII2 ) does not intersect the first co1
ordinate chart R1 , so R1 is unaffected under the second blowup. Locally above R2 , the
calculation to determine the structure of X F2 is similar to the above. Thus X F2 is covered
by the three affine open sets
K[[u, v, x, y]][V1 ]
!
v − V1 ffI
1
uV1 − fI1
and
K[[u, v, x, y]][U1 , V2 ]
u − fI1 U1
v − f V2
fI
2
U1 V2 −
and
K[[u, v, x, y]][U2 ]
!
u − fI2 U2
U2 v − ffI
2
fI2
fI1
The U2 and the V2 coordinates again glue to produce another P1 (again which might move
in a family) which maps to the origin of R2 and hence to the origin of R.
f
Continuing by blowing up the ideal (U2 , fII3 ), in this way we obtain a chain of pro2
jective birational morphisms
X Fm → X Fm−1 → . . . → X F1 → Spec R,
and we define X F := X Fm . See 5.5 later for a picture of this process.
Note that X F , being projective over the base Spec R, automatically satisfies (ELF)
[TT, 2.1.3], and by inspection of the charts, X F is a normal integral Gorenstein scheme,
so we can apply the results in the previous section. Also note that if we complete each of
the above affine open sets at the origin we obtain
K[[u, V1 , x, y]]
,
(uV1 − fI1 )
K[[U1 , V2 , x, y]]
(U1 V2 −
fI2
fI1
)
,
...
K[[Um−1 , Vm , x, y]]
(Um−1 Vm −
Theorem 5.2. Given a flag F = (I1 , . . . , Im ), denote X
is derived equivalent to EndR (T F ).
fIm
fIm−1
F
)
K[[Um , v, x, y]]
.
(Um v − fIf )
,
and T
m
F
as above. Then X F
Proof. In the explicit calculation for X F above, the preimage of the unique closed point n
of Spec R is a chain of P1 ’s (some of which move in a family), in a type Am configuration.
F
given as follows: let C = f −1 (n).
Now by [V04a, Thm. B], there is a tilting bundle on X S
Giving C the reduced scheme structure, write Cred = i∈I Ci , and let Li denote the line
bundle on X F such that Li · Cj = δij . If the multiplicity of Ci in C is equal to one, set
Mi := Li [V04a, 3.5.4], else define Mi to be given by the extension
0 → Ori −1 → Mi → Li → 0
L
associated to a minimal set of ri − 1 generators of H 1 (X F , Li−1 ). Then O ⊕ ( i∈I Mi )
is a tilting bundle on X F [V04a, 3.5.5].
In our situation, let Ci be the curve in X F , above the origin of Spec R, which in
the process of blowing up first appears in X Fi (see 5.5 for a picture of this). We now
claim that all the curves Ci have multiplicity one. If some Ci had multiplicity greater
than one, then Mi would be an indecomposable bundle with rank greater than one [V04a,
3.5.4]. But this would imply, by [V04a, 3.2.9], that its global sections H 0 (Mi ) := Mi is an
indecomposable CM R-module of rank greater than one, such that EndR (Mi ) ∈ CM R.
But this L
is impossible (see e.g. [IW3, 5.24]). Hence all curves have multiplicity one, and
m
so O ⊕ ( i=1 Li ) is a tilting
on X F . L
Lm bundle
m
∼
Li )) = EndR (R ⊕ ( i=1 H 0 (Li ))), and so it remains to show
Now End
LmX F (O0 ⊕ ( i=1
that R ⊕ ( i=1 H (Li )) ∼
= T F . In fact, we claim that H 0 (Li ) ∼
= (u, fIi ). But it is easy
21
Q-FACTORIAL TERMINALIZATIONS AND MMAS
to see (for example using the Čech complex) that the rank one CM module H 0 (Li ) is
generated as an R-module by u and fIi .
So as to match our notation with [BIKR] and [DH], we can (and do) identify maximal
flags with elements of the symmetric group Sn . Hence we regard each ω ∈ Sn as the
maximal flag
{ω(1)} ⊂ {ω(1), ω(2)} ⊂ . . . ⊂ {ω(1), . . . , ω(n − 1)}.
The following is simply a special case of 5.2.
Corollary 5.3. The scheme X ω is derived equivalent to EndR (T ω ) for all ω ∈ Sn .
Moreover the completions of singular points of X ω are precisely K[[u, v, x, y]]/(uv − fi ) for
all 1 ≤ i ≤ n such that fi ∈ (x, y)2 .
This yields the following generalization of [BIKR, 1.5] and [DH, 4.2], and provides a
conceptual reason for the condition fi ∈
/ m2 .
Corollary 5.4. Let f1 , . . . , fn ∈ m := (x, y) ⊆ K[[x, y]] be irreducible polynomials and
R = K[[u, v, x, y]]/(uv − f1 . . . fn ). Then
(1) Each T ω (ω ∈ Sn ) is an MM R-module which is a generator. The endomorphism
rings EndR (T ω ) have isolated singularities.
(2) T ω is a CT R-module for some ω ∈ Sn ⇐⇒ T ω is a CT R-module for all ω ∈ Sn
⇐⇒ fi ∈
/ m2 for all 1 ≤ i ≤ n.
Proof. (1) EndR (T ω ) is derived equivalent to a scheme whose (complete local) singularities
are listed in 5.3. Since each fi is irreducible, these are all factorial by 5.1. Hence the result
follows from 4.19.
(2) T ω is CT if and only if Γ := EndR (T ω ) is an NCCR [IW2, 5.4]. By 4.2(1) this occurs if
and only if X ω is smooth. But by 5.3, this happens if and only if each uv = fi is smooth,
which is if and only if each fi ∈
/ m2 .
For an algebraic non-derived category proof of the above, see [IW3, §5]. In fact the
commutative algebraic method in [IW3] is stronger, since it also gives that there are no
other MM (respectively CT) generators. Also, the methods in [IW3] work over more
general fields.
Remark 5.5. It is useful to visualize the above in an example.
y
inc
TI1
XF
f3
f1
inc
f4
u
R
inc
R⊕TI1 ⊕TI2 ⊕TI3
TI3
u
x
ϕ3
TI2
f2
TI1
inc
inc
f2
f1
R⊕TI1 ⊕TI2
X F2
f3 f4
u
R
x
u
ϕ2
x
X F1
y
x
f1
u
R
y
TI2
TI1
inc
f2 f3 f4
u
R⊕TI1
y
ϕ1
x
Spec R
v
R
y
R
u
22
OSAMU IYAMA AND MICHAEL WEMYSS
Consider uv = f1 f2 f3 f4 and choose maximal flag F = (I1 := {1} ⊂ I2 := {1, 2} ⊂
I3 := {1, 2, 3}). Thus TI1 = (u, f1 ), TI2 = (u, f1 f2 ) and TI3 = (u, f1 f2 f3 ). Moreover we
assume, so that we can draw an accurate picture, that (f2 ) = (f3 ), and (f3 ) 6= (f4 ), so
that uv = f2 f3 f4 is non-isolated, but uv = f3 f4 is isolated.
On the geometric side we have drawn the non-local picture; the complete local picture
is obtained by drawing a tubular neighbourhood around all the red curves, and restricting
to the origin in Spec R.
Now on X F1 , the black dot is the singularity uv = f1 , whereas the yellow dot is
uv = f2 f3 f4 . On X F2 , the latter singularity splits into uv = f2 (middle black dot) and
uv = f3 f4 (yellow). The left hand black dot is still uv = f1 . On X F , the black dots
are (reading left to right) uv = f1 , uv = f2 , uv = f3 and uv = f4 . All yellow dots
correspond to non-Q-factorial singularities, and all black dots correspond to isolated Qfactorial singularities, possibly smooth. The yellow dots with squiggles through them are
non-isolated singularities.
The red curves are the P1 ’s which map to the origin in Spec R. The right hand curve
in X F2 moves in a family (represented as blue lines) and ϕ2 contracts the whole family;
consequently ϕ2 contracts a divisor. However both ϕ1 and ϕ3 contract a single curve but
no divisor and so are flopping contractions. Note that Spec R and X F1 have canonical
singularities, but X F2 and X F have only terminal singularities.
The precise form of the quiver relies on the (easy) calculation in [IW3, 5.33], but this
can be ignored for now. Note that the geometric picture will change depending both on
the choice of polynomials, and their ordering. Blowing up in a different order can change
which curves move in families, and also normal bundles of the curves. Of course, this
depends on the choice of the polynomials too. On the algebraic side, if we change the
polynomials or their order then the number of loops change, as do the relations. Changing
the ordering corresponds to mutation (see [IW3, 5.31]).
On each level i the geometric space X Fi is derived equivalent to the corresponding
algebra by 5.2 . The top space has only Q-factorial terminal singularities, hence the top
algebra is an MMA.
Remark 5.6. A version of 5.4 is actually true in the non-local case (see [W12]), but we
explain here why the non-local case is a much more subtle problem.
For example, in the case R = C[u, v, x, y]/(uv − xy(y − 2)), define the modules
M1 := (u, x), L1 := (u, x(y − 2)) and L2 := (u, xy). Then we can show that both
M := R ⊕ M1 and L := R ⊕ L1 ⊕ L2 give NCCRs by checking complete locally. Also
we can show add M = add L by checking complete locally [IW2, 2.26]). Thus EndR (M )
and EndR (L) are Morita equivalent. Since dim R = 3 and R has an NCCR, by [V04b]
we obtain a derived equivalence between these NCCRs and some crepant resolution of
Spec R.
Hence our picture in 5.5 becomes
x(y−2)
u
inc
L2
y−2
u
xy
u
R
y
u
y
L1
R
x
u
inc
M1
y
y(y−2)
u
inc
ϕ1
x
v
R
y
u
with everything in the top row being derived equivalent. This all relies on the fact that
we a priori know that we have an NCCR (because we can check this complete locally) and
so we can use [V04b]. In a similar case uv = (x2 + y 3 )y(y − 2) (where no NCCR exists),
we cannot play this trick, since we do not know if MMAs can be detected locally, and also
since completion destroys Q-factoriality we cannot reduce to the complete local setting.
Q-FACTORIAL TERMINALIZATIONS AND MMAS
23
The other way to try to prove a non-local version of 5.4 is to try and find an explicit
tilting complex starting with the geometry, as in 5.2. Although a tilting bundle exists
abstractly for one-dimensional fibres [V04a], to describe it explicitly is delicate, since for
example we have to deal with issues such as curves being forced to flop together. We refer
the reader to [W12] for more details.
References
[A84]
[AG]
[B02]
[BKR]
[BH]
[B86]
[BIKR]
[BK]
[C02]
[C10]
[CR]
[D1]
[D2]
[D3]
[DH]
[F77]
[G06]
[GR]
[H66]
[H77]
[HVdB]
[IR]
[IW1]
[IW2]
[IW3]
[J70]
[J75]
[K88]
[K06]
[KMV]
[K89]
[KM]
[Lin]
M. Auslander, Isolated singularities and existence of almost split sequences. Representation theory, II (Ottawa, Ont., 1984), 194–242, Lecture Notes in Math., 1178, Springer, Berlin, 1986.
M. Auslander and O. Goldman, Maximal orders, Trans. Amer. Math. Soc. 97 (1960), 1–24.
T. Bridgeland, Flops and derived categories. Invent. Math. 147 (2002), no. 3, 613–632
T. Bridgeland, A. King and M. Reid, The McKay correspondence as an equivalence of derived
categories, J. Amer. Math. Soc. 14 (2001), no. 3, 535–554.
W. Bruns and J. Herzog, Cohen-Macaulay rings. Rev. ed., Cambridge Studies in Advanced Mathematics. 39 xiv+453 pp.
R. -O. Buchweitz, Maximal Cohen–Macaulay modules and Tate–cohomology over Gorenstein
rings, preprint, 1986.
I. Burban, O. Iyama, B. Keller and I. Reiten, Cluster tilting for one-dimensional hypersurface
singularities, Adv. Math. 217 (2008), no. 6, 2443–2484.
I. Burban and M. Kalck, Singularity category of a non-commutative resolution of singularities,
Adv. Math. 231 (2012), no. 1, 414–435.
J-C. Chen, Flops and equivalences of derived categories for threefolds with only terminal Gorenstein singularities, J. Differential Geom. 61 (2002), no. 2, 227–261.
X-W. Chen, Unifying two results of Orlov on singularity categories, Abh. Math. Semin. Univ.
Hambg. 80 (2010), no. 2, 207–212.
C. W. Curtis and I. Reiner, Methods of Representation Theory. Vol. I. With Applications to
Finite Groups and Orders. Wiley, New York (1990). Reprint of the 1981 original.
H. Dao, Remarks on non-commutative crepant resolutions of complete intersections, Adv. Math.
224 (2010), no. 3, 1021–1030.
H. Dao, Decent intersection and Tor-rigidity for modules over local hypersurfaces, Trans. Amer.
Math. Soc. 365 (2013), no. 6, 2803–2821.
H. Dao, Picard groups of punctured spectra of dimension three local hypersurfaces are torsion-free,
Compos. Math. 148 (2012), no. 1, 145–152.
H. Dao and C. Huneke, Vanishing of Ext, cluster tilting modules and finite global dimension of
endomorphism rings, Amer. J. Math. 135 (2013), no. 2, 561–578.
H–B. Foxby, Isomorphisms between complexes with applications to the homological theory of
modules, Mathematica Scandinavica, 40, (1977) no.1, 5–19.
V. Ginzburg, Calabi–Yau algebras, arXiv:math/0612139.
L. Gruson and M. Raynaud, Critères de platitude et de projectivité. Techniques de ‘platification’
d’un module, Invent. Math. 13, (1971) 1–89.
R. Hartshorne, Residues and duality, Lecture Notes in Mathematics, No. 20 Springer-Verlag,
Berlin-New York 1966 vii+423 pp.
R. Hartshorne, Algebraic geometry, Graduate Texts in Mathematics, No. 52. Springer-Verlag,
1977. xvi+496 pp.
L. Hille and M. Van den Bergh, Fourier-Mukai transforms, Handbook of tilting theory, 147–177,
London Math. Soc. Lecture Note Ser., 332, Cambridge Univ. Press, Cambridge, 2007.
O. Iyama and I. Reiten, Fomin-Zelevinsky mutation and tilting modules over Calabi-Yau algebras,
Amer. J. Math. 130 (2008), no. 4, 1087–1149.
O. Iyama and M. Wemyss, The classification of special Cohen Macaulay modules, Math. Z. 265
(2010), no.1, 41–83.
O. Iyama and M. Wemyss, Maximal Modifications and Auslander–Reiten Duality for Non-isolated
Singularities, arXiv:1007.1296, to appear Invent. Math.
O. Iyama and M. Wemyss, Reduction of triangulated categories and Maximal Modification Algebras for cAn singularities, arXiv:1304.5259.
C. U. Jensen, On the vanishing of lim(i) J. Alg. 15, (1970) 151–166.
←−
P. Jothilingam, A note on grade, Nagoya Math. J. 59 (1975), 149 –152.
Y. Kawamata, Crepant blowing-up of 3-dimensional canonical singularities and its application
to degenerations of surfaces, Ann. of Math. (2) 127 (1988), no. 1, 93–163.
B. Keller, On differential graded categories. International Congress of Mathematicians. Vol. II,
151–190, Eur. Math. Soc., Zurich, 2006.
B. Keller, D. Murfet, M. Van den Bergh, On two examples by Iyama and Yoshino, Compos.
Math. 147 (2011), no. 2, 591–612.
J. Kollár, Flops, Nagoya Math. J. 113 (1989), 15–36.
J. Kollár and S. Mori, Birational geometry of algebraic varieties, Cambridge Tracts in Math.,
134, Cambridge Univ. Press, Cambridge, 1998.
H.W. Lin, On crepant resolution of some hypersurface singularities and a criterion for UFD,
Trans. Amer. Math. Soc., 354 (2002), no.5, 1861–1868.
24
OSAMU IYAMA AND MICHAEL WEMYSS
[Liu]
Q. Liu, Algebraic geometry and arithmetic curves, Translated from the French by Reinie Erné.
Oxford Graduate Texts in Mathematics, 6. Oxford Science Publications. Oxford University Press,
Oxford, 2002. xvi+576 pp. ISBN: 0-19-850284-2.
[L09]
J. Lipman, Notes on derived functors and Grothendieck duality, Foundations of Grothendieck
duality for diagrams of schemes, 1–259, Lecture Notes in Math., 1960, Springer, Berlin, 2009.
[N96] A. Neeman, The Grothendieck duality theorem via Bousfield’s techniques and Brown representability, J. Amer. Math. Soc. 9 (1996), no. 1, 205–236.
[N01] A. Neeman, Triangulated categories, Annals of Mathematics Studies, vol. 148, Princeton University Press, Princeton, NJ, 2001.
[O76] A. Ooishi, Matlis duality and the width of a module, Hiroshima Math. J. 6 (1976), 573–587.
[O03] D. Orlov, Triangulated categories of singularities and D-branes in Landau-Ginzburg models, Proc.
Steklov Inst. Math. 2004, no. 3 (246), 227–248.
[O05] D. Orlov, Triangulated categories of singularities and equivalences between Landau-Ginzburg
Models, Sb. Math. 197 (2006), no. 11–12, 1827–1840.
[O09] D. Orlov, Formal completions and idempotent completions of triangulated categories of singularities, Adv. Math. 226 (2011), no. 1, 206–217.
[R83] M. Reid, Minimal models of canonical 3-folds, Algebraic varieties and analytic varieties (Tokyo,
1981), 131–180, Adv. Stud. Pure Math., 1, North-Holland, Amsterdam, 1983.
[R87] M. Reid, Young person’s guide to canonical singularities, Algebraic geometry, Bowdoin, 1985
(Brunswick, Maine, 1985), 345–414, Proc. Sympos. Pure Math., 46, Part 1, Amer. Math. Soc.,
Providence, RI, 1987.
[R89] J. Rickard, Morita theory for derived categories, J. London Math. Soc. (2) 39 (1989), 436–456.
[R98] P. Roberts, Multiplicities and Chern classes in local algebra, Cambridge Tracts in Mathematics,
133. Cambridge University Press, Cambridge, 1998. xii+303 pp.
[R08] R. Rouquier, Dimensions of triangulated categories, J. K-Theory 1 (2008), no. 2, 193–256.
[RZ]
R. Rouquier and A. Zimmermann, Picard groups for derived module categories, Proc. London
Math. Soc. (3) 87 (2003), no. 1, 197–225.
[S89]
O. Solberg, Hypersurface singularities of finite Cohen-Macaulay type. Proc. London Math. Soc.
(3) 58 (1989), no. 2, 258–280.
[TT]
R. W. Thomason and T. Trobaugh, Higher algebraic K-theory of schemes and of derived categories, The Grothendieck Festschrift, Vol. III, 247–435, Progr. Math., 88, Birkhäuser Boston,
Boston, MA, 1990.
[V04a] M. Van den Bergh, Three-dimensional flops and noncommutative rings, Duke Math. J. 122
(2004), no. 3, 423–455.
[V04b] M. Van den Bergh, Non-commutative crepant resolutions, The legacy of Niels Henrik Abel, 749–
770, Springer, Berlin, 2004.
[W10] M. Wemyss, The GL(2, C) McKay correspondence, Math. Ann. 350, no. 3, (2011), 631–659.
[W12] M. Wemyss, MMAs via key varities, in preparation.
[Y90] Y. Yoshino, Cohen-Macaulay modules over Cohen-Macaulay rings, London Mathematical Society
Lecture Note Series, 146. Cambridge University Press, Cambridge, 1990.
Osamu Iyama, Graduate School of Mathematics, Nagoya University, Chikusa-ku, Nagoya,
464-8602, Japan
E-mail address: [email protected]
Michael Wemyss, School of Mathematics, James Clerk Maxwell Building, The King’s
Buildings, Mayfield Road, Edinburgh, EH9 3JZ, UK.
E-mail address: [email protected]
| 0math.AC
|
1
Greening Geographical Power Allocation for
Cellular Networks
arXiv:1802.06285v1 [cs.IT] 17 Feb 2018
Yanju Gu
Abstract
Harvesting energy from nature (solar, wind etc.) is envisioned as a key enabler for realizing
green wireless networks. However, green energy sources are geographically distributed and the power
amount is random which may not enough to power a base station by a single energy site. Burning
brown energy sources such as coal and crude oil, though companied with carbon dioxide emission,
provides stable power. In this paper, without sacrificing communication quality, we investigate how
to perform green energy allocation to abate the dependence on brown energy with hybrid brown
and green energy injected in power networks. We present a comprehensive framework to characterize
the performance of hybrid green and brown energy empowered cellular network. Novel performance
metric “bits/tonCO2 /Hz” is proposed to evaluate the greenness of the communication network. As
green energy is usually generated from distributed geographical locations and is time varying, online
geographical power allocation algorithm is proposed to maximize the greenness of communication
network considering electricity transmission’s physical laws i.e., Ohm’s law and Kirchhoff’s circuit
laws. Simulations show that geographically distributed green energy sources complement each other by
improving the communication capacity while saving brown energy consumption. Besides, the penetration
of green energy can also help reduce power loss on the transmission breaches.
I. I NTRODUCTION
A. Context and Motivation
Telecommunications industry consumes 2% of the total electricity generation worldwide and
base stations energy consumption taking account more than 80% [1]. Energy reduction in base
stations has been studied in many ways: such as hardware design (e.g., more energy efficient
power amplifiers [2] and topological management (e.g., the deployment of relays and/or micro
BSs [3], [4]). Many base station equipment manufacturers have begun to offer a number of
cost friendly solutions to reduce power demands of base stations. For example, Nokia Networks
February 20, 2018
DRAFT
2
launches more than 20 products and services on its energy efficient base stations at Mobile World
Congress 2015.
For the cellular networks with multiple base stations, traffic-driven base station switchingon/off strategy has been proposed [5] to turn off a base station one by one that will minimally
affect the network. Similar idea has also been further studied in [6], [7] with distributed control
algorithm design. Instead of putting BSs into sleep, [8] tactically reduces the coverage (and the
power usage) of each BS, and strategically places microcells (relay stations) to offload the traffic
transmitted to/from base stations in order to save total power consumption.
In stead of simply saving base stations consumption, other efforts have been put in empowering
celluar networks with green energy such as sunlight and wind energy. In contrast, power source
emits CO2 while burned has negative impact on environment is referred as brown energy.
The fossil fueled power plants have great negative impacts on the environment as they emit a
large part of man-made greenhouse gas to the atmosphere. It has been estimated [9] that cellular
networks empowered by brown energy will consume so much energy that their CO2 emissions
equivalent will be 0.4 percent of the global value by 2020. CO2 is a heat-trapping “greenhouse”
gas which represents a negative externality on the climate system. Generation of electricity relies
heavily on brown energy. Countries such as Australia, China, India, Poland and South Africa
produce over two-thirds of their electricity and heat through the combustion of coal. To control
the use of brown energy, CO2 has already been priced and thereby giving a financial value to
each tonne of emissions saved. A sufficiently high carbon price also promotes investment in
clean, low-carbon technologies. In France, the new carbon tax is 25 US dollars per tonne of
CO2 and In Switzerland US the price is even 34.20 per tonne CO2 The total estimated income
from the carbon tax would have been between 4.5 billion euros annually.
In contrast, non-combustion energy sources such as sunlight and wind energy not convert
hydrocarbons to CO2 . Once the system is paid for, the owner of a renewable energy system will
be producing their own renewable electricity for essentially no cost and can sell the excess to the
local utility at a profit. Therefore, green cellar network benefits both environment protection and
economy cost. To build a green cellular network, we need to first improve base station which
dominates the contribution factor to overall energy consumption.
However, green energy has special properties different from traditional brown energy, and
how to efficiently stylize green energy is challenging. As green energy harvested is usually
random, most of existing works focus on designing new transmission strategies that can best
February 20, 2018
DRAFT
3
take advantage of and adapt to the random energy arrivals [10]. Optimal packet scheduling is
proposed in [11] to adaptively change the transmission rate according to the traffic load and
available energy, such that the time by which all packets are delivered is minimized.
However, green energy sources are geographically distributed, and the power amount is random
and may not enough to power a base station. For example, solar converters can deliver power
only during sunshine hours. During the night or on cloudy days, other energy sources have to
cover the power demand instead of solar power. Besides, as large solar panels are super expensive
and take up considerable space, middle size solar powers are always adopted and geographically
distributed.
B. Contributions and Organization of the Work
Microgrid is the key enabler for deep integration of renewable energy sources. It is intelligent,
reliable, and environmentally friendly. In this paper, we consider the case that base stations are
empowered by microgrid. Brown energy is a stable power source injected in the micogrid when
renewable energy is not enough. We investigate how to perform green energy allocation to abate
the dependence on brown energy without sacrificing communication quality. As geographically
distributed green energy need to be delivered to base stations to meet the power demand, power
flow needs to be performed considering power transmission’s physical laws constraints i.e.,
Ohm’s law and Kirchhoff’s circuit laws.
To solve above problems, in this paper we have made the following major contributions.
•
In this paper, we systematically study the hybrid energy sources (brown and green energy)
powered cellular network. To evaluate the greenness of a cellular network, we define the
ratio of spectrum efficiency over the total CO2 consumed (unit:bits/tonCO2 /Hz) as evaluation
metric, and the tradeoff between brown energy consumption and information data rate is
established.
•
The problem of optimal green power allocation to different base stations is modeled to
minimize the consumption of brown energy needed As the power allocation is achieved via
power flow over microgrid, power networks physical limits is considered. More importantly,
electricity transmission’s physical laws constraints i.e., Ohm’s law and Kirchhoff’s circuit
laws are considered.
•
Green energy is time varying in nature. Besides, due to measurement error, the exact amount
of available green energy is hard to know. Stochastic online green energy power allocation
February 20, 2018
DRAFT
4
algorithm is proposed and it is analytically shown that the expected brown energy consumed
converge to the minimum brown energy needed.
The rest of the paper is organized as follows. The network topology of cellular networks
empowered by microgrid with geographically distributed green energy is shown in Section II.
The brown energy consumption minimization problem is modelled and formulated in Section III.
The online stochastic power allocation algorithm is presented in Section IV. Simulation results of
online power allocation for real power networks are illustrated in Section V. Finally, conclusions
are given in Section VI. Notations throughout this paper are introduced in in Table I.
II. S YSTEM M ODEL
AND
G REEN C OMMUNICATION M ETRIC
The system model of wireless cellular networks empowered by brown and green energy is
shown in Fig. 1. Brown energy is assumed delivered by the power plant and green energy
generated by wind turbines and solar panels are geographically distributed. Both brown and
green energy are transmitted along the microgrid. The components in this system model are
detailed as follows:
•
Base station: A base station is a centralized device used to provide wireless communication
services to mobile units [12]. The wireless base station is the power consumption device.
From the power consumption perspective, the base station has two parts: the communication
unit and the cooling systems. The communication unit communication with mobile users
Fig. 1. A green communication networks in the proposed scheme that integrate multiple base stations powered by distributed
green energy. Brown energy from main grid is supplied to the communication networks via a bus of common coupling.
February 20, 2018
DRAFT
5
TABLE I
N OTATIONS
Symbol
Description
N
Total number of buses.
B,
Set of buses connected to base stations.
G,
Set of buses connected to green energy sources.
Cn
Communication capacity of base station n.
gn
Harvested energy at bus n.
pm
Power consumption at bus m.
E0
Brown energy imported from main grid.
vn
Voltage at bus n.
E0
Brown energy injected at bus 0.
Ploss
Total power loss on transmission line.
E (·)
Expectation over random variable.
t
Index of time slots.
T
Number of time slots.
at certain transmission power and communication mechanism e.g., 3G, 4G to meet the
quality-of-service of mobile users [13]. The cooling subsystem, including air conditioning
and fans, maintains an appropriate operation temperature at the BS. Each BS spends more
than 50% power on the cooling system. The base stations are connected to the micro-grid
to obtain stable power supply.
•
Microgrid: A microgrid is a localized grouping of electricity/power sources and loads [14],
[15], which consists of buses and branches, where a bus can represent a generator or a load
substation, and a branch can stand for distribution line. It connects to the main grid at a
bus of common coupling that maintains voltage at the same level as the main grid, which
provides the microgrid stable power source. Besides, the increased market penetration of
Fig. 2. Equivalent transmission branch between bus m and bus n with admittance ym,n . k and j are the other two buses
connected to bus n.
February 20, 2018
DRAFT
6
distributed green generation installed, such as solar, photovoltaic and wind, has been the
advent of an micro-grid. Voltage difference between different buses pushes current through
the branches to supply energy towards the load bus.
•
Green power source: Renewable energy provided from natural resources such as sunlight
and wind is environmentally friendly green energy. Though clean, no carbon emissions
renewable energy is not a panacea. Renewable energy is less stable and reliable compared
to fossil fuel. Though energy like solar power and wind power is abundant and unlimited,
the sun sets or disappears behind clouds, sometimes the wind stops blowing, causing
fluctuations. Besides, renewable energy exists over wide geographical areas and thus are
connected to different buses. Thus renewable energy needs to transmit to base stations at
different locations.
III. B ROWN E NERGY C ONSUMPTION M INIMIZATION : MODELING, AND
FORMULATION
In the previous section, we have discussed the system architecture for green energy allocation.
In this section, we first provide the metric for our geographically green energy allocation problem.
We then present the details for modeling the total brown energy cost, power transmission
constraints, and server service delay constraint as well as formulating the total brown energy
minimization problem. Based on detailed discussions, we show that the constrained optimization problem equivalent to a multiple objective optimization problem. The problem is proved
analytically only has Pareto optimal solution.
Considering a general cellular network consists of N base stations, and each base stations
provides Cn bit/Hz communication capacity. Power plants burns natural gas fuel source to meet
power demand base stations. In the meantime, green energy, such as solar power and wind power
is utilize to abate the dependent on brown energy consumption at the power plant.
Suppose among all power consumption for base stations, E0 kWh power is generated by the
brown energy and the carbon dioxide factor is η tonsCO2 /kWh. Green energy, such as solar
power and wind power is utilized to abate the dependent on brown energy.
Though metric like bit/Joule/Hz has been proposed to evaluate the greenness of a communication network [1], one can not evaluate how efficiently we utilize green energy. The emission
of CO2 can be exploited to distinguish the consumption of brown and green energy. More
February 20, 2018
DRAFT
7
specifically, let the ηE0 be the brown energy injected to the microgrid, the efficiency of utilizing
green energy to power the base station networks can be evaluated by the metric
PN
Cn
f = n=1 .
ηE0
(1)
The unit of function f is bits/tonCO2 /Hz. Next, we model the power flow over micro grid to
understand how geographical green energy can be allocated to different base stations.
A. Formulation of Power Flow over Micro-Grid
The power-flow problem is the computation of voltages and current in a power grid under
power line physical constraint. The power grid consists of buses and branches, where a bus
can represent a generator or a load substation, and a branch stands for a distribution line. Let
{0, 1, 2, . . . N} denote the set of bus indexes, where bus 0 is the bus of common coupling
connected to the main grid directly. In microgrid, this bus is a slack bus with fixed voltage
magnitude denoted by U.
Let G denote the set of Ng green energy sites and B denote the set of Nb base stations
connected to different buses according to their geographical locations. Two neighboring buses
are connected by a transmission line which are represented by a two-port model with certain
admittance as shown in Fig. 2.
1) Brown Energy from Main Grid: The amount of power consumption depends on the type
of base station and its communication load. Let pn denote the power consumption by the base
station connected to load bus n, and gm denote the energy generation by green energy source
connect to generation bus m. In physics, the law of conservation of energy states that the total
energy of an isolated system remains constant which is said to be conserved. Therefore, the
brown energy needs to be imported from main bus can be expressed as
X
X
E0 =
pn −
gm + Ploss 1 ,
n∈B
(2)
m∈G
where Ploss is the total branch losses due to physical heating of the lines by current flow. For
the reason that electricity delived in the micrgrid obeys Kirchhoff’s voltage law which states
that the sum of voltages around any closed loop in a circuit must be zero, the power loss must
related to the power generations and consumptions at each bus. We next investigate the explicit
expression of Ploss .
1
In this paper, we only consider the E0 > 0 case.
February 20, 2018
DRAFT
8
The admittances of all the branches in the network are denoted by an admittance matrix
Y , where the lowercase ymn indicates that matrix element that connects buses m and n. Let
X , Y −1 −
1
Y −1 11T Y −1 ,
1T Y −1 1
where p = [pT , −g T ]T
the power loss Ploss of the microgrid is given by [16]
1
Ploss ({pn }∀n∈B ) = 2 pT R{X}p,
(3)
U
with p , [p1 , ...pNb ]T and g , [g1 , ...gNg ]T 2 . Due to the fact that
elements in the first line and the first column of R{X} are all zeros, R{X} can be denoted
0 0 0
by block matrix R{X} = 0 BT M , where submatrix B, M and G are with compatible
0 M
G
dimensions such that (3) can be equivalently written as
1
Ploss ({pn }∀n∈B ) = 2 (g T Gg − 2g T M p + pT Bp).
U
From (4), one can notice that the power loss is quadratic function of pn for all n ∈ B.
(4)
2) Constraints at Buses and Branches: voltage stability played a vital role for the power
networks. For example, in the August 14, 2003, blackout in the Northeastern U.S. is due to
voltage instability. In microgrid, voltage regulator which in fact is a transformer tap is installed
for controlling or supporting voltage. Denote the voltage at bus n is vn , then it is constrained
in the safe range as |vnmin | ≤ |vn | ≤ vnmax . Let v , [v1 , . . . , vN ]T from (3) with the Ohm’s law,
from (3) one can obtain [16]
1
[(Gg − M p)T , (N T g − Bp)T ]T ,
U
is an all 1 vector with length N.
|v({pn }∀n∈B )| = U1N +
where 1N
(5)
The other constraint need to be considered for microgrid is that base station power consumption
at each bus should not exceed the maximum value. Mathematically, we can denote that |pmin
n | ≤
for n = 1, ...Nb .
pn ≤ pmax
n
Therefore, we formulate the brown energy minimization problem as
X
X
pn −
gm + Ploss ({pn }∀n∈B ),
min
{pn }∀n∈B
s.t.
∀n∈B
(6a)
m∈G
min
|pmin
n | ≤ pn ≤ pn ,
|vnmin | ≤ |vn ({pn }∀n∈B )| ≤ vnmin
∀n ∈ B
∀n ∈ B ∪ G.
(6b)
(6c)
It can be observed from (6), the objective function is in quadratic form [17], [18] and the
constrain domain is convex.
2
Only real power is considered in this model and reactive power is neutralized by the reactive compensation unit that is
installed at base station and renewable energy generator.
February 20, 2018
DRAFT
9
B. Communication Model
An important metric for characterizing any communication is the communication capacity
which is defined as the maximum amount of information that can be transmitted as a function
of available bandwidth given a constraint on transmitted power channel [19]. For the nth base
station, the total energy consumption pn includes the transmitted power and the rest that is due to
other components such as air conditioning, data processor, and circuits, which can be generally
modeled as a constant power pc,n > 0. Thus the transmitted power can be denoted by pn − pc,n .
[5].
Pertaining to topology constraints [20], [21] between base station and users, communications
can be categorized into uplink and downlink. An uplink is communications used for the transmission of signals from an user to the base station. As base station only performs received signal
processing for the uplink communication, the corresponding power consumption is assumed
incorporated in pc,n . Thus pn is mainly used for downlink transmission.
Next, we investigate the power consumption for downland transmission. Without loss of
generality, we consider a flat fading model with composite channel fading (including both largescale and small-scale fading) from base station n to user k denoted by hn,k . Let xn,k denote the
data symbol transmitted from base station n, then the received signal yn,k is expressed as
yn,k = (pn − pc,n )xn,k hn,k + wn,k ,
(7)
where the transmitted power to user k is p̃n,k joules/symbol, hn,k denotes the channel fading
between base station n and user k and and w ∼ N (0, σn2 ) is i.i.d. Gaussian noise.
The capacity Cn of a channel provides the performance limit: reliable communication can
be attained at any rate Rk < Cn ; reliable communication is impossible at rates Rk > Cn . In
the multiuser case, the consent capacity region which is the set of all Rk such that all users
can reliably communicate at theire own rate Rk , respectively. We have the multi-user data rate
bounds [22]
(pn − pc,n )|hn |2
,
(8)
σn2
where |hn | = arg minhn,k |hn,k |2 . (9) indicates that increasing the transmission power pn − pc,n
Cn = log2 1 +
can increase the transmission capacity. To achieve the lowest rate, the transmission power is
requested to be
pn ≥
(2Cn − 1)σn2
+pc,n .
|hn,k |2
{z
}
|
(9)
pn
February 20, 2018
DRAFT
10
Note that information of σn2 and |hn | is estimated at user k and feeded back to base station n
via control channel. This feedback communication is a standard process in nowadays wireless
communication systems for adaptive power control. We then have
X
Cn ,
max
{pn }∀n∈B
s.t.
(10a)
n∈B
∀n ∈ B
pn ≥ pn + pc,n ,
(10b)
min
|pmin
n | ≤ pn ≤ pn ,
∀n ∈ B
(10c)
|vnmin | ≤ |vn ({pn }∀n∈B )| ≤ vnmin ,
∀n ∈ B ∪ G
(10d)
It can be simply approved that the above communication capacity optimization is a convex
optimization problem.
C. Green Metric for Communication Networks and Pareto Optimal Solvation
Substituting the geographical power allocation model in (6) and the communication capacity
model in (10) in the greenness metric (1) we have
PN
C
P n=1 n
η= P
max
,
{pn }∀n∈B
η( n∈B pn − m∈G gm + Ploss ({pn }∀n∈B ))
s.t.
∀n ∈ B
pn ≥ pn + pc,n ,
(11a)
(11b)
min
|pmin
n | ≤ pn ≤ pn ,
∀n ∈ B
(11c)
|vnmin | ≤ |vn ({pn }∀n∈B )| ≤ vnmin ,
∀n ∈ B ∪ G
(11d)
The above problem is equivalent to multicriteria optimization problem and can be formulated
as follows
Obj 1 : max
{pn }∀n∈B
Obj 2 :
min
{pn }∀n∈B
s.t.
C=
N
X
log2 1 +
n=1
E0 =
X
n∈B
pn −
X
m∈G
pn |hn |2
,
σn2
(12a)
gm + Ploss ({pn }∀n∈B ),
(12b)
pn ≥ pn + pc,n ,
min
|pmin
n | ≤ pn ≤ pn ,
|vnmin| ≤ |vn | ≤ vnmin .
∀n ∈ B
∀n ∈ B
∀n ∈ B ∪ G
(12c)
(12d)
(12e)
It can be easily noticed that these two objective functions are conflicting objectives. For example,
when the total brown power is minimized (i.e., pn = 0), the communication capacity is also 0
February 20, 2018
DRAFT
11
and is not maximized. So there does not appear to exist an optimal solution in our problem that
optimizes both objectives simultaneously. Then we can only obtain the Pareto optimal solution
when investigating the multiple objectives optimization problem.
Instead of solving (12) directly, let consider a simpler single objective optimization problem
for a given C0 ,
E0 ({pn }∀n∈B )
min
{pn }∀n∈B
=
X
n∈B
s.t.
N
X
pn −
X
m∈G
log2 1 +
n=1
(13a)
gm + Ploss ({pn }∀n∈B ),
pn |hn |2
≥ C0 ,
σn2
pn ≥ pn + pc,n ,
(13b)
∀n ∈ B
(13c)
min
|pmin
n | ≤ pn ≤ pn ,
∀n ∈ B
(13d)
|vnmin | ≤ |vn ({pn }∀n∈B )| ≤ vnmin ,
∀n ∈ B ∪ G
(13e)
which is named as one shot optimization. We show in the Appendix that if E0 = E0∗ is an
optimal solution for a given value C0 = C ∗ in (13), then (C ∗ , E0∗ ) is a Pareto optimal solution
to (12).
Lemma 1 Let E0∗ be an optimal solution to (13) for a given value of C0 = C ∗ , then (E0∗ , C ∗ )
is a Pareto optimal solution to (12).
Proof 1 See Appendix.
It can be easily shown that (13) is a convex optimization problem and therefore can be solved
efficiently and globally using the interior-point methods. However, in practice the power of
green energy gn is random and the measurements of gn exists errors and lead to error in power
allocation. In the next section, a convergence guaranteed stochastic renewable energy allocation
will be discussed.
IV. O NLINE S TOCHASTIC P OWER A LLOCATION
A. Average Brown Energy Consumption Optimization
In microgrid, Algorithm 1 is implemented in a central control unit for power allocation via
gathering the green energy gm and other parameters in the micrograms and cellular networks.
February 20, 2018
DRAFT
12
However, real-time exact value of gm is difficult to be obtained due to measurements errors
and communication delays from distributed green energy sites to the central control unit. Rather
than implementing the unreliable and possibly obsolete instantaneous decision of pn of (13),
we consider a robust stochastic control scheme which is to minimize the average brown energy
consumption as
min
{pn (t)}∀n∈B
=E
s.t.
N
X
n=1
f ({gn (t)}∀n∈B )
X
n∈B
pn (t) −
log2 1 +
X
(14a)
gm (t) + Ploss ({pn (t)}∀n∈B ) ,
m∈G
pn (t)|hn |2
≥ C0 ,
σn2
pn ≥ pn (t) + pc,n (t),
min
|pmin
n | ≤ pn (t) ≤ pn (t),
∀n ∈ B
∀n ∈ B
|vnmin | ≤ |vn ({pn (t)}∀n∈B )| ≤ vnmin , ∀n ∈ B ∪ G
(14b)
(14c)
(14d)
(14e)
In the subsequent, we use pn (t) and gm (t) to represent the time varying power consumption and
green power generation of base station n and green energy site m.
B. Stochastic Online Power Allocation Algorithm Design and Analysis
Leveraging recent advances in online stochastic convex optimization, the above problem
(14) can be solved under stochastic approximation framework. In this paper, we use Bregmen
projection based mirror decent algorithm to iterate the objective variable p(t) that gradually
converges to the optimization point of the expectation functions in (14).
Let f ∗ (y) , supx∈domf (xT y−f (x)) denote the conjugate function of f (x), and F (·) denotes a
continuous differentiable function that is α-strongly convex w.r.t. Euclid norm k·k. The Bregmen
divergence associated with F (·) is defined as [23]
BF (x, y) , F (x) − F (y) − (x − y)T ▽F (y),
(15)
where BF satisfies BF (x, y) ≥ α2 kx − yk2 for some α > 0. The online mirror decent method
is described as the following two steps iterations: First, the gradient is performed in the dual
space rather than in the primal space as
ωp(t+1) = ▽F ∗ (▽F (p(t)) − δ▽f (p(t), pg )),
February 20, 2018
(16)
DRAFT
13
where F ∗ (·) denotes the dual function of F (·). and δ is the step size. The second step is the
projection step defined by the Bregman divergence associated to F (·) as
p(t + 1) = arg min BF (x, ωp (t + 1)),
x∈A
(17)
where A is the feasible domain of x. Intuitively, the above mirror decent algorithm minimizes
a first-order approximation of the function f (·) at the current iterate ωp(t) while forcing the next
iterate ωp(t+1) to lie close to ωp(t) . The step size δ controls the trade-off between these two.
To obtain closed form solution in the iteration, we use standard norm as the Bregmen divergence function i.e., F (x) = 12 kxk22. Substituting the expression into (15) and after simple
calculation, we obtain BF (x, y) = 12 kx−yk2 . By Holder’s inequality, and a simple optimization
of a quadratic polynomial, one has F ∗ (y) = supx∈D (xT y − 12 kxk2 ) ≤ supx∈D (kxkkyk −
1
kxk2 )
2
= 21 kx − yk2 . Notice that the inequality above is in fact an equality by definition of the
dual norm. Thus F ∗ (y) = y. Moreover, as ▽F (x) = x, we have ▽F (x)∗ = x.
Then the update with the gradient in (16) can be computed as
ωp(t+1) = p(t) − η▽f (p(t), g).
(18)
▽f (p(t), g) = −2g T M + 2p(t)T B.
(19)
where
Then the update with the projection step in (17) is derived as
1
p(t + 1) = arg min kx − ωp (t + 1)k2 ,
2
N
X
pn |hn |2
s.t.
log2 1 +
≥ C0 ,
σn2
n=1
pn ≥ pn + pc,n ,
(20a)
(20b)
∀n ∈ B
min
|pmin
n | ≤ pn ≤ pn ,
(20c)
∀n ∈ B
|vnmin | ≤ |vn ({pn }∀n∈B )| ≤ vnmin .
(20d)
∀n ∈ B ∪ G
(20e)
To summarize, the stochastic online power allocation can be solved by an iterative algorithm
with two steps in each iteration. The first step can be easily computed using the closed form of
the gradient as in (18) The second step is to compute (20). Obviously, (20) is a simple convex
optimization problem, which can be solved by interior point method. The stochastic online power
allocation algorithm is summarized in Algorithm 1.
February 20, 2018
DRAFT
14
Algorithm 1 Stochastic Online Power Allocation
1: Initialize: p(1) is computed by one shot optimization in (13).
2:
for t ∈ {1, 2, . . . , T } do
3:
Compute the gradient ωp(t+1) according to (18).
4:
Update the gradient according to (19).
5:
Compute p(t + 1) via minimizing BF (x, ωp (t + 1)) according to (20).
6:
end for
Notice that the proposed online power allocation algorithm does not depend on any measurement error or communication delay distribution assumption. It rather utilizes real-time communication and power data to infer the unknown statistics. Next we evaluate the performance of
the online power allocation algorithm.
Property 1 Let {p̃n (t)}∀n∈B and {p̂n (t)}∀n∈B are the minimizers of (13) and (14) respectively,
it holds that
k
E0 ({p̃n (t)}∀n∈B ) − f ({p̂n (t)}∀n∈B ) ≤ √ ,
T
(21)
where k is a constant irrespective of T .
Proposition 1 guarantees that the expected brown power consumption converges to the opti√
mum stochastic solution at the rate of O(1/ T ). Therefore the online power allocation algorithm
in Table 1 has sublinear accurate error.
V. PERFORMANCE EVALUATION
In this section, we present and discuss simulation results, in comparison with numerical results
pertaining to the previously developed analysis.
A. Parameter Setting
The 37 bus test feeder model [24] is used for the simulation of micogrid. It represents an
actual underground radial distribution feeder in California. The schematic view of this network
can be seen in Fig. 7. In this 37-bus network, bus 799 is the bus of common coupling. Stable
brown energy is injected to the microgrid when green energy is not enough. Green energy sites
are geographically distributed at buses 742, 725, 735 731 and 741. Each other buses is load
February 20, 2018
DRAFT
15
bus and is linked to a base station. The maximum power rate at each bus is set as 1.3MW.
Solar power generation sites of 12MW are placed respectively on bus 742, 725 and 735 and the
wind power generation sites of 8MW are placed respectively on bus 731 and 741. The green
energy data set are from Smart∗ Microgrid Data Set [25]. Besides, we generate the data set for
base station power consumption by adopting Huawei DBS3900 base station power consumption
parameters in the simultion [26]. User traffic data are loaded according to [27] for the simulation
and the total user number variation is given in Fig. 4. In general while users increases, the power
consumption increases.
For each operation interval, the microgrid control center collects green energy generation
gm from generation buses. It is assumed that the renewable generation data is observed with
a white Gaussian noise of 20% of its actual value due to measurement error. Besides, the
communication channel state information hn for all n ∈ B and noise variance σn2 are obtained
in the communication networks and send to the microgrid control center for implement the
proposed online stochastic power allocation algorithm.
B. Experimental Results
While the user traffic increases more power is needed to support base station for information
P
transmission. As the transmission power increases, the communication capacity n∈B Cn will
increase. Assume frequency reuse is adopted and therefore there is no inter-cell interference
beween neighboring cells. To investigate the advantage of green energy for the communication
system, base stations communication power at each base station is normalized by
1
(Ploss + E0 ).
Nb
Fig. 5 and Fig. 6 show the capacities variation of base stations at bus 720 and 730, respectivly. It
is clear the proposed online stochastic power allocation algorithm provides larger communication
capacity then the one shot solution. Thus it utilizes the green energy more efficiently and therefore
abates the dependence on brown energy from the main grid. Besides, it can be observed that
online stochastic algorithm provides more stable capacity then one shot algorithm.
Power loss on the microgrid is due to physical heating on the transmission branches caused
by electricity delivering. Though it is unavoidable, it is preferred to lower the power loss as
much as possible. Distributed green energy also decrease the power loss on the microgrid. We
evaluate the total power loss for the proposed stochastic online power allocation algorithm and
one shot optimization algorithm. Besides, the power loss is also simulated according to (4) in
the case that there is no green energy distributed in the microgrid and all energy is brown energy
February 20, 2018
DRAFT
16
Fig. 3. Schematic diagram of the IEEE 37-bus test feeder cases [24]. Bus 799 is the bus of common coupling that links the
microgrid to main power grid. Solar power harvesting sites are located at buses 742, 725, 735 and wind power harvesting sites
are located at buses 731 and 741. These green energy sites are denoted by green triangles. Other buses are load buses connected
to base stations.
750
User number
700
650
600
550
0
500
1000
Time slot
1500
2000
Fig. 4. Dynamic user number versus time slot.
imported from the main grid. Fig. 7 shows the total power loss cost reductions by comparing
results of the proposed online stochastic power allocation algorithm (Algorithm 1) and the one
shot solution (13). It shows that while geographically green energy incorporated, powr flow is
more efficiently as the power source is not far from the consumption bus. In contrast, with only
brown energy from main grid, the power loss increases greatly.
February 20, 2018
DRAFT
17
Base station capacity at bus 720
1
One shot solution
Onlilne stochastic algorithm
0.8
0.6
0.4
0.2
0
0
200
400
600
Time slot
800
1000
Fig. 5. Normalized communication capacities of base station empowered by bus 720 versus time slot.
Base station capacity at bus 730
1
0.8
0.6
0.4
0.2
0
0
One shot solution
Onlilne stochastic algorithm
200
400
600
Time slot
800
1000
Fig. 6. Normalized communication capacities of base station empowered by bus 730 versus time slot.
VI. C ONCLUSIONS
In this paper, we have investigated optimal green energy allocation to abate the dependence
on brown energy without sacrificing communication quality. Leveraging the fact that the green
energy power amount generated from distributed geographical locations varies, the framework
of green energy cooperation to empower cellular networks has been achieved by formulate the
problem as maximizing data rate per ton CO2 consumed. Electricity transmission’s physical laws
constraints i.e., Ohm’s law and Kirchhoff’s circuit laws, are considered for the power allocation.
The exact real-time amount of green energy is difficult to obtain due to the noise contaminated
measurement and possible communication delay from energy harvesting site to control unite,
and therefore degrades the power allocation performance. We therefore have proposed stochastic
online power allocation algorithm which is robust to the green energy uncertainty. It is shown
that the online algorithm converges to the optimum stochastic data rate per ton CO2 . Finally, we
February 20, 2018
DRAFT
18
6
10
x 10
One shot solution
Powered by brown energy only
Onlilne stochastic algorithm
8
Power loss
6
4
2
0
0
500
1000
Time slot
1500
2000
Fig. 7. Power loss versus time slot.
have conducted data trace-driven experiments. The results show that geographically distributed
green energy complement each other by improving the communication capacity while saving
brown energy consumption from main grid. We also come to some key findings such as with
penetration of green energy, power loss on the transmission breaches can be reduced.
A PPENDIX
The proof is based on contradiction. Suppose (E0∗ , C ∗ ) is not a Pareto optimal solution to
(12). Then there must exist a feasible solution (E0 , C) to (12) such that C > C ∗ and E0 < E0∗ .
Based on this feasible solution (E0 , C), we can construct another feasible solution to (13) as
follows. We keep power consumption unchanged, but decrease C to C ∗ through scheduling or
interference management. Clearly, (E0 , C ∗ ) is also a feasible solution. So corresponding to the
same C ∗ value, we have two feasible solution (E0 , C ∗ ) and (E0∗ , C ∗ ) and that E0 < E0∗ . This
contradicts to the fact that E0∗ is an optimal solution to (13) under a given C ∗ value. This
completes our proof.
R EFERENCES
[1] Z. Hasan, H. Boostanimehr, and V. Bhargava, “Green cellular networks: A survey, some research issues and challenges,”
Communications Surveys Tutorials, IEEE, vol. 13, no. 4, pp. 524–540, Fourth 2011.
[2] IEEE
PES
Distribution
System
Analysis
Subcommittee
,
“IEEE
PES
Test
Feeders,”
http://www.huawei.com/en/about-huawei/publications/communicate/hw-082748.htm.
[3] K. Son, E. Oh, and B. Krishnamachari, “Energy-aware hierarchical cell configuration: From deployment to operation,” in
Computer Communications Workshops (INFOCOM WKSHPS), 2011 IEEE Conference on, April 2011, pp. 289–294.
February 20, 2018
DRAFT
19
[4] J. Du and Y.-C. Wu, “Network-wide distributed carrier frequency offsets estimation and compensation via belief
propagation,” IEEE Transactions on Signal Processing, vol. 61, no. 23, pp. 5868–5877, 2013.
[5] C. Peng, S.-B. Lee, S. Lu, H. Luo, and H. Li, “Traffic-driven power saving in operational 3g cellular networks,” in
Proceedings of the 17th Annual International Conference on Mobile Computing and Networking, ser. MobiCom ’11, 2011,
pp. 121–132.
[6] E. Oh, K. Son, and B. Krishnamachari, “Dynamic base station switching-on/off strategies for green cellular networks,”
Wireless Communications, IEEE Transactions on, vol. 12, no. 5, pp. 2126–2136, May 2013.
[7] J. Du and Y.-C. Wu, “Fully distributed clock skew and offset estimation in wireless sensor networks,” in IEEE International
Conference on Acoustics, Speech and Signal Processing (ICASSP), 2013, pp. 4499–4503.
[8] C. Gao, W. Zhang, J. Tang, C. Wang, S. Zou, and S. Su, “Relax, but do not sleep: A new perspective on green wireless
networking,” in INFOCOM, 2014 Proceedings IEEE, April 2014, pp. 907–915.
[9] A. Bianzino, C. Chaudet, D. Rossi, and J. Rougier, “A survey of green networking research,” Communications Surveys
Tutorials, IEEE, vol. 14, no. 1, pp. 3–20, First 2012.
[10] O. Ozel, K. Tutuncuoglu, J. Yang, S. Ulukus, and A. Yener, “Transmission with energy harvesting nodes in fading wireless
channels: Optimal policies,” Selected Areas in Communications, IEEE Journal on, vol. 29, no. 8, pp. 1732–1743, September
2011.
[11] J. Yang and S. Ulukus, “Optimal packet scheduling in an energy harvesting communication system,” Communications,
IEEE Transactions on, vol. 60, no. 1, pp. 220–230, January 2012.
[12] J. Du and Y.-C. Wu, “Distributed clock skew and offset estimation in wireless sensor networks: Asynchronous algorithm
and convergence analysis,” IEEE Transactions on Wireless Communications, vol. 12, no. 11, pp. 5908–5917, 2013.
[13] J. Du, X. Liu, and L. Rao, “Proactive doppler shift compensation in vehicular cyber-physical systems,” IEEE/ACM
Transactions on Networking, 2018.
[14] J. Du, S. Ma, Y.-C. Wu, and H. V. Poor, “Distributed hybrid power state estimation under PMU sampling phase errors,”
IEEE Transactions on Signal Processing, vol. 62, no. 16, pp. 4052–4063, 2014.
[15] Q. Wang, X. Liu, J. Du, and F. Kong, “Smart charging for electric vehicles: A survey from the algorithmic perspective,”
IEEE Communications Surveys & Tutorials, vol. 18, no. 2, pp. 1500–1517, 2016.
[16] S. Bolognani and S. Zampieri, “On the existence and linear approximation of the power flow solution in power distribution
networks,” Power Systems, IEEE Transactions on, vol. PP, no. 99, pp. 1–10, 2015.
[17] J. Du, S. Ma, Y.-C. Wu, S. Kar, and J. M. Moura, “Convergence analysis of distributed inference with vector-valued
Gaussian belief propagation,” arXiv preprint arXiv:1611.02010, 2016.
[18] J. Du, S. Kar, and J. M. Moura, “Distributed convergence verification for Gaussian belief propagation,” Asilomar Conference
on Signals, Systems and Computers (ASILOMAR), arXiv preprint arXiv:1711.09888, 2017.
[19] K. Cai, X. Li, J. Du, Y.-C. Wu, and F. Gao, “Cfo estimation in ofdm systems under timing and channel length uncertainties
with model averaging,” IEEE Transactions on Wireless Communications, vol. 9, no. 3, pp. 970–974, 2010.
[20] J. Du, S. Ma, Y.-C. Wu, S. Kar, and J. M. Moura, “Convergence analysis of the information matrix in Gaussian
belief propagation,” IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), arXiv preprint
arXiv:1704.03969, 2017.
[21] ——, “Convergence analysis of belief propagation for pairwise linear Gaussian models,” IEEE Global Conference on
Signal and Information Processing (GlobalSIP), arXiv preprint arXiv:1706.04074, 2017.
[22] D. Tse and P. Viswanath, Fundamentals of Wireless.
Cambridge, 2009.
[23] M. Grotschel, S. O. Krumke, and J. Rambau, Online Optimizatio of Large Scale Systems.
February 20, 2018
Springer, 2001.
DRAFT
20
[24] IEEE
PES
Distribution
System
Analysis
Subcommittee
,
“IEEE
PES
Test
Feeders,”
http://www.ewh.ieee.org/soc/pes/dsacom/testfeeders/index.html, [Online; accessed 20-July-2015].
[25] A. M. S. Barker, D. Irwin, E. Cecchet, P. Shenoy, and J. Albrecht, “Smart: An open data set and tools for enabling
research in sustainable homes,” in Computer Communications Workshops (INFOCOM WKSHPS), 2011 IEEE Conference
on, August 2012, pp. 289–299.
[26] HUAWEI
TECHNOLOGIES
CO.,
LTD.
,
“Huawei
3900
Series
Base
Station
Product
Description,”
http://enterprise.huawei.com/ilink/enenterprise/download/HW 323282.
[27] K. Zheng, Y. Zheng, N. Yuan, and S. Shang, “On discovery of gathering patterns from trajectories,” in Data Engineering
(ICDE), 2013 IEEE 29th International Conference on, April 2013, pp. 242–253.
February 20, 2018
DRAFT
| 7cs.IT
|
arXiv:cs/0601038v1 [cs.LO] 10 Jan 2006
Under consideration for publication in Theory and Practice of Logic Programming
1
Constraint-based Automatic Verification of
Abstract Models of Multithreaded Programs
GIORGIO DELZANNO
Dipartimento di Informatica e Scienze dell’Informazione, Università di Genova
via Dodecaneso 35, 16146 Genova - Italy
(e-mail: [email protected])
submitted 17 December 2003; revised 13 April 2005; accepted 15 January 2006
Abstract
We present a technique for the automated verification of abstract models of multithreaded
programs providing fresh name generation, name mobility, and unbounded control.
As high level specification language we adopt here an extension of communication finitestate machines with local variables ranging over an infinite name domain, called TDL
programs. Communication machines have been proved very effective for representing communication protocols as well as for representing abstractions of multithreaded software.
The verification method that we propose is based on the encoding of TDL programs
into a low level language based on multiset rewriting and constraints that can be viewed as
an extension of Petri Nets. By means of this encoding, the symbolic verification procedure
developed for the low level language in our previous work can now be applied to TDL
programs. Furthermore, the encoding allows us to isolate a decidable class of verification
problems for TDL programs that still provide fresh name generation, name mobility, and
unbounded control. Our syntactic restrictions are in fact defined on the internal structure
of threads: In order to obtain a complete and terminating method, threads are only allowed
to have at most one local variable (ranging over an infinite domain of names).
KEYWORDS: Constraints, Multithreaded Programs, Verification.
1 Introduction
Andrew Gordon (Gordon 2001) defines a nominal calculus to be a computational
formalism that includes a set of pure names and allows the dynamic generation of
fresh, unguessable names. A name is pure whenever it is only useful for comparing
for identity with other names. The use of pure names is ubiquitous in programming
languages. Some important examples are memory pointers in imperative languages,
identifiers in concurrent programming languages, and nonces in security protocols.
In addition to pure names, a nominal process calculus should provide mechanisms
for concurrency and inter-process communication. A computational model that provides all these features is an adequate abstract formalism for the analysis of multithreaded and distributed software.
2
Giorgio Delzanno
The Problem Automated verification of specifications in a nominal process calculus
becomes particularly challenging in presence of the following three features: the possibility of generating fresh names (name generation); the possibility of transmitting
names (name mobility); the possibility of dynamically adding new threads of control
(unbounded control). In fact, a calculus that provides all the previous features can
be used to specify systems with a state-space infinite in several dimensions. This
feature makes difficult (if not impossible) the application of finite-state verification
techniques or techniques based on abstractions of process specifications into Petri
Nets or CCS-like models. In recent years there have been several attempts of extending automated verification methods from finite-state to infinite-state systems
(Abdulla and Nylén 2000; Kesten et al. 2001). In this paper we are interested in investigating the possible application of the methods we proposed in (Delzanno 2001)
to verification problems of interest for nominal process calculi.
Constraint-based Symbolic Model Checking In (Delzanno 2001) we introduced a
specification language, called MSR(C), for the analysis of communication protocols
whose specifications are parametric in several dimensions (e.g. number of servers,
clients, and tickets as in the model of the ticket mutual exclusion algorithm shown
in (Bozzano and Delzanno 2002)). MSR(C) combines multiset rewriting over first
order atomic formulas (Cervesato et al. 1999) with constraints programming. More
specifically, multiset rewriting is used to specify the control part of a concurrent
system, whereas constraints are used to symbolically specify the relations over local data. The verification method proposed in (Delzanno 2005) allows us to symbolically reason on the behavior of MSR(C) specifications. To this aim, following
(Abdulla et al. 1996; Abdulla and Nylén 2000) we introduced a symbolic representation of infinite collections of global configurations based on the combination of
multisets of atomic formulas and constraints, called constrained configurations.1
The verification procedure performs a symbolic backward reachability analysis by
means of a symbolic pre-image operator that works over constrained configurations
(Delzanno 2005). The main feature of this method is the possibility of automatically
handling systems with an arbitrary number of components. Furthermore, since we
use a symbolic and finite representation of possibly infinite sets of configurations,
the analysis is carried out without loss of precision.
A natural question for our research is whether and how these techniques can be
used for verification of abstract models of multithreaded programs.
Our Contribution In this paper we propose a sound, and fully automatic verification
method for abstract models of multithreaded programs that provide name generation, name mobility, and unbounded control. As a high level specification language
we adopt here an extension with value-passing of the formalism of (Ball et al. 2001)
1
Notice that in (Abdulla et al. 1996; Abdulla and Nylén 2000) a constraint denotes a symbolic
state whereas we use the word constraint to denote a symbolic representation of the relation of
data variables (e.g. a linear arithmetic formula) used as part of the symbolic representation of
sets of states (a constrained configuration).
Constraint-based Verification of Abstract Multithreaded Programs
3
based on families of state machines used to specify abstractions of multithreaded
software libraries. The resulting language is called Thread Definition Language
(TDL). This formalism allows us to keep separate the finite control component of a
thread definition from the management of local variables (that in our setting range
over a infinite set of names), and to treat in isolation the operations to generate
fresh names, to transmit names, and to create new threads. In the present paper we
will show that the extension of the model of (Ball et al. 2001) with value-passing
makes the model Turing equivalent.
The verification methodology is based on the encoding of TDL programs into a
specification in the instance MSRN C of the language scheme MSR(C) of(Delzanno 2001).
MSRN C is obtained by taking as constraint system a subclass of linear arithmetics
with only = and > relations between variables, called name constraints (N C). The
low level specification language MSRN C is not just instrumental for the encoding
of TDL programs. Indeed, it has been applied to model consistency and mutual
exclusion protocols in (Bozzano and Delzanno 2002; Delzanno 2005). Via this encoding, the verification method based on symbolic backward reachability obtained
by instantiating the general method for MSR(C) to NC-constraints can now be applied to abstract models of multithreaded programs. Although termination is not
guaranteed in general, the resulting verification method can succeed on practical
examples as the Challenge-Response TDL program defined over binary predicates
we will illustrated in the present paper. Furthermore, by propagating the sufficient
conditions for termination defined in (Bozzano and Delzanno 2002; Delzanno 2005)
back to TDL programs, we obtain an interesting class of decidable problems for abstract models of multithreaded programs still providing name generation, name
mobility, and unbounded control.
Plan of the Paper In Section 2 we present the Thread Definition Language (TDL)
with examples of multithreaded programs. Furthermore, we discuss the expressiveness of TDL programs showing that they can simulate Two Counter Machines. In
Section 3, after introducing the MSRN C formalism, we show that TDL programs
can be simulated by MSRN C specifications. In Section 4 we show how to transfer
the verification methods developed for MSR(C) to TDL programs. Furthermore, we
show that safety properties can be decided for the special class of monadic TDL
programs. In Section 5 we address some conclusions and discuss related work.
2 Thread Definition Language (TDL)
In this section we will define TDL programs. This formalism is a natural extension
with value-passing of the communicating machines used by (Ball et al. 2001) to
specify abstractions of multithreaded software libraries.
Terminology Let N be a denumerable set of names equipped with the relations =
and 6= and a special element ⊥ such that n 6= ⊥ for any n ∈ N . Furthermore, let
V be a denumerable set of variables, C = {c1 , . . . , cm } a finite set of constants, and
L a finite set of internal action labels. For a fixed V ⊆ V, the set of expressions is
4
Giorgio Delzanno
defined as E = V ∪ C ∪ {⊥} (when necessary we will use E(V ) to explicit the set of
variables V upon which expressions are defined). The set of channel expressions is
defined as Ech = V ∪ C. Channel expressions will be used as synchronization labels
so as to establish communication links only at execution time.
A guard over V is a conjunction γ1 , . . . , γs , where γi is either true, x = e or x 6= e
with x ∈ V and e ∈ E for i : 1, . . . , s. An assignment α from V to W is a conjunction
like xi := ei where xi ∈ W , ei ∈ E(V ) for i : 1, . . . k and xr 6= xs for r 6= s. A
message template m over V is a tuple m = hx1 , . . . , xu i of variables in V .
Definition 1
A TDL program is a set T = {P1 , . . . , Pt } of thread definitions (with distinct names
for local variables control locations). A thread definition P is a tuple hQ, s0 , V, Ri,
where Q is a finite set of control locations, s0 ∈ Q is the initial location, V ⊆ V is
a finite set of local variables, and R is a set of rules. Given s, s′ ∈ Q, and a ∈ L, a
rule has one of the following forms2 :
a
• Internal move: s −−→ s′ [γ, α], where γ is a guard over V , and α is an assignment from V to V ;
a
• Name generation: s −−→ s′ [x := new], where x ∈ V , and the expression new
denotes a fresh name;
a
• Thread creation: s −−→ s′ [run P ′ with α], where P ′ = hQ′ , t, W, R′ i ∈ T , and
α is an assignment from V to W that specifies the initialization of the local
variables of the new thread;
e!m
• Message sending: s −−→ s′ [γ, α], where e is a channel expression, m is a message template over V that specify which names to pass, γ is a guard over V ,
and α is an assignment from V to V .
e?m
• Message reception: s −−−→ s′ [γ, α], where e is a channel expression, m is a
message template over a new set of variables V ′ (V ′ ∩ V = ∅) that specifies
the names to receive, γ is a guard over V ∪ V ′ and α is an assignment from
V ∪ V ′ to V .
Before giving an example, we will formally introduce the operational semantics of
TDL programs.
2.1 Operational Semantics
In the following we will use N to indicate the subset of used names of N . Every
constant c ∈ C is mapped to a distinct name nc 6= ⊥ ∈ N , and ⊥ is mapped to ⊥.
Let P = hQ, s, V, Ri and V = {x1 , . . . , xk }. A local configuration is a tuple p =
hs′ , n1 , . . . , nk i where s′ ∈ Q and ni ∈ N is the current value of the variable xi ∈ V
for i : 1, . . . , k.
A global configuration G = hN, p1 , . . . , pm i is such that N ⊆ N and p1 , . . . , pm are
local configurations defined over N and over the thread definitions in T . Note that
2
In this paper we keep assignments, name generation, and thread creation separate in order to
simplify the presentation of the encoding into MSR.
Constraint-based Verification of Abstract Multithreaded Programs
5
there is no relation between indexes in a global configuration in G and in T ; G is
a pool of active threads, and several active threads can be instances of the same
thread definition.
Given a local configuration p = hs′ , n1 , . . . , nk i, we define the valuation ρp as
ρp (xi ) = ni if xi ∈ V , ρp (c) = nc if c ∈ C, and ρp (⊥) = ⊥. Furthermore, we
say that ρp satisfies the guard γ if ρp (γ) ≡ true, where ρp is extended to constraints in the natural way (ρp (ϕ1 ∧ ϕ2 ) = ρp (ϕ1 ) ∧ ρp (ϕ2 ), etc.).
The execution of x := e has the effect of updating the local variable x of a thread
with the current value of e (a name taken from the set of used values N ). On the contrary, the execution of x := new associates a fresh unused name to x. The formula
run P with α has the effect of adding a new thread (in its initial control location)
to the current global configuration. The initial values of the local variables of the
generated thread are determined by the execution of α whose source variables are
the local variables of the parent thread. The channel names used in a rendez-vous
are determined by evaluating the channel expressions tagging sender and receiver
rules. Value passing is achieved by extending the evaluation associated to the current configuration of the receiver so as to associate the output message of the sender
to the variables in the input message template. The operational semantics is given
via a binary relation ⇒ defined as follows.
Definition 2
Let G = hN, . . . , p, . . .i, and p = hs, n1 , . . . , nk i be a local configuration for P =
hQ, s, V, Ri, V = {x1 , . . . , xk }, then:
a
• If there exists a rule s −−→ s′ [γ, α] in R such that ρp satisfies γ, then G ⇒
hN, . . . , p′ , . . .i (meaning that only p changes) where p′ = hs′ , n′1 , . . . , n′k i,
n′i = ρp (ei ) if xi := ei is in α, n′i = ni otherwise, for i : 1, . . . , k.
a
• If there exists a rule s −−→ s′ [xi := new] in R, then G ⇒ hN ′ , . . . , p′ , . . .i
′
′
′
′
where p = hs , n1 , . . . , nk i, ni is an unused name, i.e., n′i ∈ N \ N , n′j = nj
for every j 6= i, and N ′ = N ∪ {n′i };
a
• If there exists a rule s −−→ s′ [run P ′ with α] in R with P ′ = hQ′ , t0 , W, R′ i,
W = {y1 , . . . , yu }, and α is defined as y1 := e1 , . . . , yu := eu then G ⇒
hN, . . . , p′ , . . . , qi (we add a new thread whose initial local configuration is q)
where p′ = hs′ , n1 , . . . , nk i, and q = ht0 , ρp (e1 ), . . . , ρp (eu )i.
• Let q = ht, m1 , . . . , mr i (distinct from p) be a local configuration in G associated with P ′ = hQ′ , t0 , W, R′ i.
e!m
e′ ?m′
Let s −−→ s′ [γ, α] in R and t −−−−→ t′ [γ ′ , α′ ] in R′ be two rules such that
m = hx1 , . . . , xu i, m′ = hy1 , . . . , yv i and u = v (message templates match).
We define σ as the value passing evaluation σ(yi ) = ρp (xi ) for i : 1, . . . , u,
and σ(z) = ρq (z) for z ∈ W ′ .
Now if ρp (e) = ρp (e′ ) (channel names match), ρp satisfies γ, and that σ
satisfies γ ′ , then hN, . . . , p, . . . , q, . . .i ⇒ hN, . . . , p′ , . . . , q′ , . . .i where p′ =
hs′ , n′1 , . . . , n′k i, n′i = ρp (v) if xi := v is in α, n′i = ni otherwise for i : 1, . . . , k;
q′ = ht′ , m′1 , . . . , m′r i, m′i = σ(v) if ui := v is in α′ , m′i = mi otherwise for
i : 1, . . . , r.
6
Giorgio Delzanno
Definition 3
An initial global configuration G0 has an arbitrary (but finite) number of threads
with local variables all set to ⊥. A run is a sequence G0 G1 . . . such that Gi ⇒ Gi+1
for i ≥ 0. A global configuration G is reachable from G0 if there exists a run from
G0 to G.
Example 1
Let us consider a challenge and response protocol in which the goal of two agents
Alice and Bob is to exchange a pair of new names hnA , nB i, the first one created
by Alice and the second one created by Bob, so as to build a composed secret key.
We can specify the protocol by using new names to dynamically establish private
channel names between instances of the initiator and of the responder. The TDL
program in Figure 1 follows this idea. The thread Init specifies the behavior of the
initiator. He first creates a new name using the internal action f resh, and stores
it in the local variable nA . Then, he sends nA on channel c (a constant), waits for
a name y on a channel with the same name as the value of the local variable nA
(the channel is specified by variable nA ) and then stores y in the local variable
mA . The thread Resp specifies the behavior of the responder. Upon reception of
a name x on channel c, he stores it in the local variable nB , then creates a new
name stored in local variable mB and finally sends the value in mB on channel with
the same name as the value of nB . The thread M ain non-deterministically creates
new thread instances of type Init and Resp. The local variable x is used to store
new names to be used for the creation of a new thread instance. Initially, all local
variables of threads Init/Resp are set to ⊥. In order to allow process instances to
participate to several sessions (potentially with different principals), we could also
add the following rule
restart
stopA −−−−−→ initA [nA := ⊥, mA := ⊥]
In this rule we require that roles and identities do not change from session to
session.3 Starting from G0 = hN0 , hinit, ⊥ii, and running the Main thread we can
generate any number of copies of the threads Init and Resp each one with a unique
identifier. Thus, we obtain global configurations like
hN, hinitM , ⊥i,
hinitA , i1 , ⊥, ⊥i, . . . , hinitA , iK , ⊥, ⊥i,
hinitB , iK+1 , ⊥, ⊥i, . . . , hinitB , iK+L , ⊥, ⊥i i
where N = {⊥, i1, . . . , iK , iK+1 , . . . , iK+L } for K, L ≥ 0. The threads of type Init
and Resp can start parallel sessions whenever created. For K = 1 and L = 1 one
possible session is as follows.
Starting from
h{⊥, i1 , i2 }, hinitM , ⊥i, hinitA , i1 , ⊥, ⊥i, hinitB , i2 , ⊥, ⊥ii
3
By means of thread and fresh name creation it is also possible to specify a restart rule in which
a given process takes a potential different role or identity.
Constraint-based Verification of Abstract Multithreaded Programs
7
T hread Init(local idA , nA , mA );
f resh
initA −−−−→ genA
c!hnA i
genA −−−−→ waitA
n ?hyi
waitA −−A
−−−→ stopA
[nA := new]
[true]
[mA := y]
T hread Resp(local id, nB , mB );
c?hxi
initB −−−→ genB
[nB := x]
f resh
genB −−−−→ readyB
n !hm i
[mB := new]
B
readyB −−B−−−−
→ stopB
[true]
T hread M ain(local x);
id
initM −−→ create
newA
create −−−−→
initM
new
→ initM
create −−−−B
[x := new]
[run Init with idA := x, nA := ⊥, mA := ⊥, x := ⊥]
[run Resp with idB := x, nB := ⊥, mB := ⊥, x := ⊥B ]
Fig. 1. Example of thread definitions.
if we apply the first rule of thread Init to hinitA , i1 , ⊥, ⊥i we obtain
h{⊥, i1 , i2 , a1 }, hinitM , ⊥i, hgenA , i1 , a1 , ⊥i, hinitB , i2 , ⊥, ⊥ii
where a1 is the generated name (a1 is distinct from ⊥, i1 , and i2 ). Now if we apply
the second rule of thread Init and the first rule of thread Resp (synchronization
on channel c) we obtain
h{⊥, i1 , i2 , a1 }, hinitM , ⊥i, hwaitA , i1 , a1 , ⊥i, hgenB , i2 , a1 , ⊥ii
If we apply the second rule of thread Resp we obtain
h{⊥, i1 , i2 , a1 , a2 }, hinitM , ⊥i, hwaitA , i1 , a1 , ⊥i, hreadyB , i2 , a1 , a2 ii
Finally, if we apply the last rule of thread Init and Resp (synchronization on
channel a1 ) we obtain
h{⊥, i1 , i2 , a1 , a2 }, hinitM , ⊥i, hstopA , i1 , a1 , a2 i, hstopB , i2 , a1 , a2 ii
Thus, at the end of the session the thread instances i1 and i2 have both a local
copy of the fresh names a1 and a2 . Note that a copy of the main thread hinitM , ⊥i
is always active in any reachable configuration, and, at any time, it may introduce
new threads (either of type Init or Resp) with fresh identifiers. Generation of fresh
names is also used by the threads of type Init and Resp to create nonces. Furthermore, threads can restart their life cycle (without changing identifiers). Thus,
in this example the set of possible reachable configurations is infinite and contains
configurations with arbitrarily many threads and fresh names. Since names are
stored in the local variables of active threads, the local data also range over an
infinite domain.
✷
8
Giorgio Delzanno
2.2 Expressive Power of TDL
To study the expressive power of the TDL language, we will compare it with the
Turing equivalent formalism called Two Counter Machines. A Two Counters Machine configurations is a tuple hℓ, c1 = n1 , c2 = n2 i where ℓ is control location taken
from a finite set Q, and n1 and n2 are natural numbers that represent the values
of the counters c1 and c2 . Each counter can be incremented or decremented (if
greater than zero) by one. Transitions combine operations on individual counters
with changes of control locations. Specifically, the instructions for counter ci are as
follows
Inc: ℓ1 : ci := ci + 1; goto ℓ2 ;
Dec: ℓ1 : if ci > 0 then ci := ci − 1; goto ℓ2 ; else goto ℓ3 ;
A Two Counter Machine consists then of a list of instructions and of the initial
state hℓ0 , c1 = 0, c2 = 0i. The operational semantics is defined according to the intuitive semantics of the instructions. Problems like control state reachability are
undecidable for this computational model.
The following property then holds.
Theorem 1
TDL programs can simulate Two Counter Machines.
Proof
In order to define a TDL program that simulates a Two Counter Machine we
proceed as follows. Every counter is represented via a doubly linked list implemented
via a collection of threads of type Cell and with a unique thread of type Last
pointing to the head of the list. The i-th counter having value zero is represented
as the empty list Cell(i, v, v), Last(i, v, w) for some name v and w (we will explain
later the use of w). The i-th counter having value k is represented as
Cell(i, v0 , v0 ), Cell(i, v0 , v1 ), . . . , C(i, vk−1 , vk ), Last(i, vk , w)
for distinct names v0 , v1 , . . . , vk . The instructions on a counter are simulated by
sending messages to the corresponding Last thread. The messages are sent on
channel Zero (zero test), Dec (decrement), and Inc (increment). In reply to each
of these messages, the thread Last sends an acknowledgment, namely Y es/N o for
the zero test, DAck for the decrement, IAck for the increment operation. Last
interacts with the Cell threads via the messages tstC, decC, incC acknowledged
by messages z/nz, dack. iack. The interactions between a Last thread and the Cell
threads is as follows.
Zero Test Upon reception of a message hxi on channel Zero, the Last thread with
local variables id, last, aux checks that its identifier id matches x - see transition
from Idle to Busy - sends a message hid, lasti on channel tstC directed to the cell
pointed to by last (transition from Busy to W ait), and then waits for an answer. If
the answer is sent on channel nz, standing for non-zero, (resp. z standing for zero)
- see transition from W ait to AckN Z (resp. AckZ) - then it sends its identifier on
Constraint-based Verification of Abstract Multithreaded Programs
9
Thread Last(local id, last, aux);
(Zero test)
Zero?hxi
Idle −−−−−−→ Busy
[id = x]
tstC!hid,lasti
Busy −−−−−−−−−→ W ait
nz?hxi
W ait −−−−→ AckN Z
[id = x]
z?hxi
W ait −−−→ AckZ
[id = x]
Y es!hidi
AckZ −−−−−→ Idle
No!hidi
AckN Z −−−−−→ idle
(Decrement)
Dec?hxi
Idle −−−−−→ Dbusy
[id = x]
decC!hid,lasti
DBusy −−−−−−−−−→ DW ait
dack?hx,ui
DW ait −−−−−−−→ DAck
[id = x, last := u]
DAck!hidi
DAck −−−−−−−→ Idle
(Increment)
Inc?hxi
Idle −−−−−→ IN ew
[id = x]
new
[aux := new]
run
[run Cell with idc := id; prev := last; next := aux]
IN ew −−−→ IRun
IRun −−→ IAck
IAck!hidi
IAck −−−−−−→ Idle
[last := aux]
Fig. 2. The process defining the last cell of the linked list associated to a counter
channel N o (resp. Y es) as an acknowledgment to the first message - see transition
from AckN Z (resp. Z) to Idle. As shown in Fig. 3, the thread Cell with local
variables idc, prev, and next that receives the message tstC, i.e., pointed to by a
thread Last with the same identifier as idc, sends an acknowledgment on channel
z (zero) if prev = next, and on channel nz (non-zero) if prev 6= next.
Decrement Upon reception of a message hxi on channel Dec, the Last thread with
local variables id, last, aux checks that its identifier id matches x (transition from
Idle to Dbusy), sends a message hid, lasti on channel decC directed to the cell
pointed to by last (transition from Busy to W ait), and then waits for an answer.
10
Giorgio Delzanno
Thread Cell(local idc, prev, next);
(Zero test)
tstC?hx,ui
idle −−−−−−−→ ackZ [x = idc, u = next, prev = next]
tstC?hx,ui
idle −−−−−−−→ ackN Z [x = idc, u = next, prev 6= next]
z!hidci
ackZ −−−−→ idle
nz!hidci
ackN Z −−−−−→ idle
(Decrement)
dec?hx,ui
idle −−−−−−→ dec [x = idc, u = next, prev 6= next]
dack!hidc,previ
dec −−−−−−−−−−→ idle
Fig. 3. The process defining a cell of the linked list associated to a counter
If the answer is sent on channel dack (transition from DW ait to DAck) then it
updates the local variable last with the pointer u sent by the thread Cell, namely
the prev pointer of the cell pointed to by the current value of last, and then sends
its identifier on channel DAck to acknowledge the first message (transition from
DAck to Idle).
As shown in Fig. 3, a thread Cell with local variables idc, prev, and next that
receives the message decC and such that next = last sends as an acknowledgment
on channel dack the value prev.
Increment To simulate the increment operation, Last does not have to interact with
existing Cell threads. Indeed, it only has to link a new Cell thread to the head of
the list (this is way the Cell thread has no operations to handle the increment
operation). As shown in Fig. 2 this can be done by creating a new name stored in
the local variable aux (transition from IN ew to IRun) and spawning a new Cell
thread (transition from IRun to IAck) with prev pointer equal to last, and next
pointer equal to aux. Finally, it acknowledges the increment request by sending its
identifier on channel IAck and updates variable last with the current value of aux.
Two Counter Machine Instructions We are now ready to use the operations provided by the thread Last to simulate the instructions of a Two Counter Machine.
As shown in Fig. 4, we use a thread CM with two local variables id1 , id2 to represent the list of instructions of a 2CM with counters c1 , c2 . Control locations of the
Two Counter Machines are used as local states of the thread CM . The initial local
state of the CM thread is the initial control location. The increment instruction
on counter ci at control location ℓ1 is simulated by an handshaking with the Last
thread with identifier idi : we first send the message Inc!hidi i, wait for the acknowledgment on channel IAck and then move to state ℓ2 . Similarly, for the decrement
Constraint-based Verification of Abstract Multithreaded Programs
11
Thread CM (local id1 , id2 );
..
.
(Instruction : ℓ1 : ci := ci + 1; goto ℓ2 ; )
Inc!hid i
ℓ1 −−−−−−i→ waitℓ1
IAck!hxi
waitℓ1 −−−−−−→ ℓ2 [x = idi ]
..
.
(Instruction : ℓ1 : ci > 0 then ci := ci − 1; goto ℓ2 ; else goto ℓ3 ; )
Zero!hid i
ℓ1 −−−−−−−i→ waitℓ1
NZAck?hxi
waitℓ1 −
−−−−−−−
→ decℓ1 [x = idi ]
Dec!hid i
decℓ1 −−−−−−i→ wdecℓ1
DAck?hyi
wdecℓ1 −−−−−−→ ℓ2 [y = idi ]
ZAck?hxi
waitℓ1 −−−−−−→ ℓ3 [x = idi ]
..
.
Fig. 4. The thread associated to a 2CM.
Thread Init(local nid1 , p1 , nid2 , p2 );
f reshId
init −−−−−→ init1 [nid1 := new]
f reshP
init1 −−−−−→ init2 [p1 := new]
runC
init2 −
−−−→ init3 [run Cell with idc := nid1 ; prev := p1 ; next := p1 ]
runL
init3 −−−→ init4 [run Last with idc := nid1 ; last := p1 ; aux := ⊥]
f reshId
init4 −−−−−→ init5 [nid2 := new]
f reshP
init5 −−−−−→ init6 [p2 := new]
runC
init6 −
−−−→ init7 [run Cell with idc := nid2 ; prev := p2 ; next := p2 ]
runL
init7 −−−→ init8 [run Last with idc := nid2 ; last := p2 ; aux := ⊥]
runCM
init8 −−−−−→ init9 [run 2CM with id1 := nid1 ; id2 := nid2 ]
Fig. 5. The initialization thread.
instruction on counter ci at control location ℓ1 we first send the message Zero!hidi i.
If we receive an acknowledgment on channel N ZAck we send a Dec request, wait
for completion and then move to ℓ2 . If we receive an acknowledgment on channel
ZAck we directly move to ℓ3 .
12
Giorgio Delzanno
Initialization The last step of the encoding is the definition of the initial state of the
system. For this purpose, we use the thread Init of Fig. 5. The first four rules of Init
initialize the first counter: they create two new names nid1 (an identifier for counter
c1 ) and p1 , and then spawn the new threads Cell(nid1 , p1 , p1 ), Last(nid1 , p1 , ⊥).
The following four rules spawns the new threads Cell(nid2 , p2 , p2 ), Last(nid2 , p2 , ⊥).
After this stage, we create a thread of type 2CM to start the simulation of the instructions of the Two Counter Machines. The initial configuration of the whole
system is G0 = hinit, ⊥, ⊥i. By construction we have that an execution step from
hℓ1 , c1 = n1 , c2 = n2 i to hℓ2 , c1 = m1 , c2 = m2 i is simulated by an execution run going from a global configuration in which the local state of thread CM is hℓ1 , id1 , id2 i
and in which we have ni occurrences of thread Cell with the same identifier idi
for i : 1, 2, to a global configuration in which the local state of thread CM is
hℓ2 , id1 , id2 i and in which we have mi occurrences of thread Cell with the same
identifier idi for i : 1, 2. Thus, every executions of a 2CM M corresponds to an execution of the corresponding TDL program that starts from the initial configuration
G0 = hinit, ⊥, ⊥i.
As a consequence of the previous theorem, we have the following corollary.
Corollary 1
Given a TDL program, a global configurations G, and a control location ℓ, deciding
if there exists a run going from G0 to a global configuration that contains ℓ (control
state reachability) is an undecidable problem.
3 From TDL to MSRN C
As mentioned in the introduction, our verification methodology is based on a translation of TDL programs into low level specifications given in MSRN C . Our goal is
to extend the connection between CCS and Petri Nets (German and Sistla 1992)
to TDL and MSR so as to be able to apply the verification methods defined in
(Delzanno 2005) to multithreaded programs. In the next section we will summarize
the main features of the language MSRN C introduced in (Delzanno 2001).
3.1 Preliminaries on MSRN C
N C-constraints are linear arithmetic constraints in which conjuncts have one of
the following form: true, x = y, x > y, x = c, or x > c, x and y being two variables
from a denumerable set V that range over the rationals, and c being an integer.
The solutions Sol of a constraint ϕ are defined as all evaluations (from V to Q)
that satisfy ϕ. A constraint ϕ is satisfiable whenever Sol(ϕ) 6= ∅. Furthermore, ψ
entails ϕ whenever Sol(ψ) ⊆ Sol(ϕ). N C-constraints are closed under elimination
of existentially quantified variables.
Let P be a set of predicate symbols. An atomic formula p(x1 , . . . , xn ) is such that
p ∈ P, and x1 , . . . , xn are distinct variables in V. A multiset of atomic formulas is
indicated as A1 | . . . | Ak , where Ai and Aj have distinct variables (we use variable
renaming if necessary), and | is the multiset constructor.
Constraint-based Verification of Abstract Multithreaded Programs
13
In the rest of the paper we will use M, N , . . . to denote multisets of atomic formulas,
ǫ to denote the empty multiset, ⊕ to denote multiset union and ⊖ to denote multiset
difference. An MSRN C configuration is a multiset of ground atomic formulas, i.e.,
atomic formulas like p(d1 , . . . , dn ) where di is a rational for i : 1, . . . , n.
An MSRN C rule has the form M −→ M′ : ϕ, where M and M′ are two (possibly
empty) multisets of atomic formulas with distinct variables built on predicates in
P, and ϕ is an N C-constraint. The ground instances of an MSRN C rule are defined
as
Inst(M −→ M′ : ϕ) = {σ(M) −→ σ(M′ ) | σ ∈ Sol(ϕ)}
where σ is extended in the natural way to multisets, i.e., σ(M) and σ(M′ ) are
MSRN C configurations.
An MSRN C specification S is a tuple hP, I, Ri, where P is a finite set of predicate
symbols, I is finite a set of (initial) MSRN C configurations, and R is a finite set of
MSRN C rules over P.
The operational semantics describes the update from a configuration M to one of its
possible successor configurations M′ . M′ is obtained from M by rewriting (modulo
associativity and commutativity) the left-hand side of an instance of a rule into the
corresponding right-hand side. In order to be fireable, the left-hand side must be
included in M. Since instances and rules are selected in a non deterministic way, in
general a configuration can have a (possibly infinite) set of (one-step) successors.
Formally, a rule H −→ B : ϕ from R is enabled at M via the ground substitution
σ ∈ Sol(ϕ) if and only if σ(H) 4 M. Firing rule R enabled at M via σ yields the
new configuration
M′ = σ(B) ⊕ (M ⊖ σ(H))
We use M ⇒MSR M′ to denote the firing of a rule at M yielding M′ .
A run is a sequence of configurations M0 M1 . . . Mk with M0 ∈ I such that
Mi ⇒MSR Mi+1 for i ≥ 0. A configuration M is reachable if there exists M0 ∈ I
∗
∗
such that M0 ⇒MSR M, where ⇒MSR is the transitive closure of ⇒MSR . Finally, the successor and predecessor operators P ost and P re are defined on a set
of configurations S as P ost(S) = {M′ |M ⇒MSR M′ , M ∈ S} and P re(S) =
{M|M ⇒MSR M′ , M′ ∈ S}, respectively. P re∗ and P ost∗ denote their transitive
closure.
As shown in (Delzanno 2001; Bozzano and Delzanno 2002), Petri Nets represent a
natural abstractions of MSRN C (and more in general of MSR rule with constraints)
specifications. They can be encoded, in fact, in propositional MSR specifications
(e.g. abstracting away arguments from atomic formulas).
3.2 Translation from TDL to MSRN C
The first thing to do is to find an adequate representation of names. Since all we
need is a way to distinguish old and new names, we just need an infinite domain
in which the = and 6= relation are supported. Thus, we can interpret names in N
14
Giorgio Delzanno
either as integer of as rational numbers. Since operations like variable elimination
are computationally less expensive than over integers, we choose to view names
as non-negative rationals. Thus, a local (TDL) configuration p = hs, n1 , . . . , nk i
is encoded as the atomic formula p• = s(n1 , . . . , nk ), where ni is a non-negative
rational. Furthermore, a global (TDL) configuration G = hN, p1 , . . . , pm i is encoded
as an MSRN C configuration G•
p•1 | . . . | p•m | f resh(n)
where the value n in the auxiliary atomic formula f resh(n) is an rational number
strictly greater than all values occurring in p•1 , . . . , p•m . The predicate f resh will
allow us to generate unused names every time needed.
The translation of constants C = {c1 , . . . , cm }, and variables is defined as follows:
x• = x for x ∈ V, ⊥• = 0, c•i = i for i : 1, . . . , m. We extend ·• in the natural way
on a guard γ, by decomposing every formula x 6= e into x < e• and x > e• . We will
call γ • the resulting set of N C-constraints.4
Given V = {x1 , . . . , xk }, we define V ′ as the set of new variables {x′1 , . . . , x′k }.
Now, let us consider the assignment α defined as x1 := e1 , . . . , xk := ek (we add
assignments like xi := xi if some variable does not occur as target of α). Then, α•
is the N C-constraint x′1 = e•1 , . . . , x′k = e•k .
The translation of thread definitions is defined below (where we will often refer to
Example 1).
Initial Global Configuration Given an initial global configuration consisting of the
local configurations hsi , ni1 , . . . , niki i with nij = ⊥ for i : 1, . . . , u, we define the
following MSRN C rule
init → s1 (x11 , . . . , x1k1 ) | . . . | su (xu1 , . . . , xuku ) | f resh(x) :
x > C, x11 = 0, . . . , xuku = 0
here C is the largest rational used to interpret the constants in C.
For each thread definition P = hQ, s0 , V, Ri in T with V = {x1 , . . . , xk } we translate
the rules in R as described below.
a
Internal Moves For every internal move s −−→ s′ [γ, α], and every ν ∈ γ • we define
s(x1 , . . . , xk ) → s′ (x′1 , . . . , x′k ) : ν, α•
a
Name Generation For every name generation s −−→ s′ [xi := new], we define
^
s(x1 , . . . , xk ) | f resh(x) → s′ (x′1 , . . . , x′k ) | f resh(y) : y > x′i , x′i > x,
x′j = xj
j6=i
f resh
For instance, the name generation initA −−−−→ genA [n := new] is mapped into the
MSRN C rule initA (id, x, y)| f resh(u) −−→ genA (id′ , x′ , y ′ ) | f resh(u′ ) : ϕ where ϕ
4
As an example, if γ is the constraint x = 1, x 6= z then γ • consists of the two constraints
x = 1, x > z and x = 1, z > x.
Constraint-based Verification of Abstract Multithreaded Programs
15
is the constraint u′ > x′ , x′ > u, y ′ = y, id′ = id. The constraint x′ > u represents
the fact that the new name associated to the local variable n (the second argument
of the atoms representing the thread) is fresh, whereas u′ > x′ updates the current
value of f resh to ensure that the next generated names will be picked up from
unused values.
Thread Creation Let P = hQ′ , t0 , V ′ , R′ i and V ′ = {y1 , . . . , yu }. Then, for every
a
thread creation s −−→ s′ [run P with α], we define
s(x1 , . . . , xk ) → s′ (x′1 , . . . , x′k ) | t(y1′ , . . . , yu′ ) : x′1 = x1 , . . . , x′k = xk , α• .
new
A
initM [run Init with id := x, . . .] of Example 1.
E.g., consider the rule create −−−−→
Its encoding yields the MSRN C rule create(x) −−→ initM (x′ ) | initA (id′ , n′ , m′ ) : ψ,
where ψ represents the initialization of the local variables of the new thread x′ =
x, id′ = x, n′ = 0, m′ = 0.
Rendez-vous The encoding of rendez-vous communication is based on the use of
constraint operations like variable elimination. Let P and P ′ be a pair of thread
definitions, with local variables V = {x1 , . . . , xk } and V ′ = {y1 , . . . , yl } with V ∩
e′ ?m′
e!m
V ′ = ∅. We first select all rules s −−→ s′ [γ, α] in R and t −−−−→ t′ [γ ′ , α′ ] in R′ ,
such that m = hw1 , . . . , wu i, m′ = hw1′ , . . . , wv′ i and u = v. Then, we define the new
MSRN C rule
s(x1 , . . . , xk ) | t(y1 , . . . , yl ) → s′ (x′1 , . . . , x′k ) | t′ (y1′ , . . . , yl′ ) : ϕ
for every ν ∈ γ • and ν ′ ∈ γ ′• such that the NC-constraint ϕ obtained by eliminating w1′ , . . . , wv′ from the constraint ν ∧ ν ′ ∧ α• ∧ α′• ∧ w1 = w1′ ∧ . . . ∧ wv = wv′
nA ?hyi
is satisfiable. For instance, consider the rules waitA −−−−−→ stopA [mA := y] and
nB !hmB i
readyB −−−−−−→ stopB [true]. We first build up a new constraint by conjoining the
NC-constraints y = mB (matching of message templates), and nA = nB , m′A =
y, n′A = nA , m′B = mB , n′B = nB , id′1 = id1 , id′2 = id2 (guards and actions of
sender and receiver). After eliminating y we obtain the constraint ϕ defined as
nB = nA , m′A = mB , n′A = nA , m′B = mB , n′B = nB , id′1 = id1 , id′2 = id2 defined
over the variables of the two considered threads. This step allows us to symbolically
represent the passing of names. After this step, we can represent the synchronization of the two threads by using a rule that simultaneously rewrite all instances
that satisfy the constraints on the local data expressed by ϕ, i.e., we obtain the
rule
waitA (id1 , nA , mA )| readyB (id2 , nB , mB ) −→
stopA (id′1 , n′A , m′A ) | stopB (id′2 , n′B , m′B ) : ϕ
The complete translation of Example 1 is shown in Fig. 6 (for simplicity we have
applied a renaming of variables in the resulting rules). An example of run in the
resulting MSRN C specification is shown in Figure 7. Note that, a fresh name is
selected between all values strictly greater than the current value of f resh (e.g. in
the second step 6 > 4), and then f resh is updated to a value strictly greater than
all newly generated names (e.g. 8 > 6 > 4).
16
Giorgio Delzanno
init −−→ f resh(x) | initM (y) : x > 0, y = 0.
f resh(x) | initM (y) −−→ f resh(x′ ) | create(y ′ ) : x′ > y ′ , y ′ > x.
create(x) −−→ initM (x′ ) | initA (id′ , n′ , m′ ) : x′ = x, id′ = x, n′ = 0, m′ = 0.
create(x) −−→ initM (x′ ) | initB (id′ , n′ , m′ ) : x′ = x, id′ = x, n′ = 0, m′ = 0.
initA (id, n, m)| f resh(u) −−→ genA (id, n′ , m) | f resh(u′ ) : u′ > n′ , n′ > u.
genA (id1 , n, m)| initB (id2 , u, v) −−→ waitA (id1 , n, m) | genB (id′2 , u′ , v ′ ) : u′ = n, v ′ = v
genB (id, n, m)| f resh(u) −−→ readyB (id, n, m′ ) | f resh(u′ ) : u′ > m′ , m′ > u.
waitA (id1 , n, m)| readyB (id2 , u, v) −−→ stopA (id1 , n, m′ ) | stopB (id2 , u, v) : n = u, m′ = v.
stopA (id, n, m) −−→ initA (id′ , n′ , m′ ) : n′ = 0, m′ = 0, id′ = id.
stopB (id, n, m) −−→ initB (id′ , n′ , m′ ) : n′ = 0, m′ = 0, id′ = id.
Fig. 6. Encoding of Example 1: for simplicity we embed constraints like x = x′ into
the MSR formulas.
init ⇒ . . . ⇒ f resh(4) | initM (0) | initA (2, 0, 0) | initB (3, 0, 0)
⇒ f resh(8) | initM (0) | genA (2, 6, 0) | initB (3, 0, 0)
⇒ f resh(8) | initM (0) | waitA (2, 6, 0) | genB (3, 6, 0)
⇒ . . . ⇒ f resh(16) | initM (0) | waitA (2, 6, 0) | genB (3, 6, 0) | initA (11, 0, 0)
Fig. 7. A run in the encoded program.
Let T = hP1 , . . . , Pt i be a collection of thread definitions and G0 be an initial
global state. Let S be the MSRN C specification that results from the translation
described in the previous section.
Let G = hN, p1 , . . . , pn i be a global configuration with pi = hsi , vi1 , . . . , viki i, and
let h : N ❀ Q+ be an injective mapping. Then, we define G• (h) as the MSRN C
configuration
s1 (h(v11 ), . . . , h(v1k1 )) | . . . | sn (h(vn1 ), . . . , h(vnkn )) | f resh(v)
where v is a the first value strictly greater than all values in the range of h. Given
an MSRN C configuration M defined as s1 (v11 , . . . , v1k1 ) | . . . | sn (vn1 , . . . , vnkn )
with sij ∈ Q+ , let V (M) ⊆ Q+ be the set of values occurring in M. Then, given a
bijective mapping f : V (M) ❀ N ⊆ N , we define M• (f ) as the global configuration hN, p1 , . . . , pn i where pi = hsi , f (vi1 ), . . . , f (viki )i.
Based on the previous definitions, the following property then holds.
Theorem 2
For every run G0 G1 . . . in T with corresponding set of names N0 N1 . . ., there exist
sets D0 D1 . . . and bijective mappings h0 h1 . . . with hi : Ni ❀ Di ⊆ Q+ for i ≥ 0,
such that init G•0 (h0 )G•1 (h1 ) . . . is a run of S. Vice versa, if init M0 M1 . . . is a
run of S, then there exist sets N0 N1 . . . in N and bijective mappings f0 f1 . . . with
fi : V (Mi ) ❀ Ni for i ≥ 0, such that M•0 (f0 )M•1 (f1 ) . . . is a run in T .
Proof
We first prove that every run in T is simulated by a run in S.
Let G0 . . . Gl be a run in T , i.e., a sequence of global states (with associated set
Constraint-based Verification of Abstract Multithreaded Programs
17
of names N0 . . . Nl ) such that Gi ⇒ Gi+1 and Ni ⊆ Ni+1 for i ≥ 0.
We prove that it can be simulated in S by induction on its length l.
Specifically, suppose that there exist sets of non negative rationals D0 . . . Dl and
bijective mappings h0 . . . hl with hi : Ni ❀ Di for 0 ≤ i ≤ l, such that
c0 (h0 ) . . . G
cl (hl )
init G
is a run of S. Furthermore, suppose Gl ⇒ Gl+1 .
We prove the thesis by a case-analysis on the type of rule applied in the last step
of the run.
Let Gl = hNl , p1 , . . . , pr i and pj = hs, n1 , . . . , nk i be a local configuration for the
thread definition P = hQ, s, V, Ri with V = {x1 , . . . , xk } and ni ∈ Nl for i : 1, . . . , k.
a
Assignment Suppose there exists a rule s −−→ s′ [γ, α] in R such that ρpj satisfies γ,
Gl = hNl , . . . , pj , . . .i ⇒ hNl+1 , . . . , p′j , . . .i = Gl+1 Nl = Nl+1 , p′j = hs′ , n′1 , . . . , n′k i,
and if xi := yi occurs in α, then n′i = ρpj (yi ), otherwise n′i = ni for i : 1, . . . , k.
The encoding of the rule returns one M SRN C rule having the form
s(x1 , . . . , xk ) → s′ (x′1 , . . . , x′k ) : γ ′ , α
b
for every γ ′ ∈ γ
b.
cl (hl ) is a multiset of atomic formulas that contains the
By inductive hypothesis, G
formula s(hl (n1 ), . . . , hl (nk )).
Now let us define hl+1 as the mapping from Nl to Dl such that hl+1 (n′i ) = hl (nj ) if
xi := xj is in α and hl+1 (n′i ) = 0 if xi := ⊥ is in α. Furthermore, let us the define
the evaluation
σ = hx1 7→ hl (n1 ), . . . , xk 7→ hl (nk ), x′1 7→ hl+1 (n′1 ), . . . , x′k 7→ hl+1 (n′k )i
Then, by construction of the set of constraints b
γ and of the constraint α
b, it follows
that σ is a solution for γ ′ , α
b for some γ ′ ∈ b
γ . As a consequence, we have that
s(n1 , . . . , nk ) → s′ (n′1 , . . . , n′k )
is a ground instance of one of the considered M SRN C rules.
cl (hl ), if we apply a rewriting step
Thus, starting from the M SRN C configuration G
we obtain a new configuration in which s(n1 , . . . , nk ) is replaced by s′ (n′1 , . . . , n′k ),
c
[
and all the other atomic formulas in G
l+1 (hl+1 ) are the same as in Gl (hl ). The
[
resulting M SRN C configuration coincides then with the definition of G
l+1 (hl+1 ).
Creation of new names Let us now consider the case of fresh name generation.
a
Suppose there exists a rule s −−→ s′ [xi := new] in R, and let n 6∈ Nl , and suppose
hNl , . . . , pj , . . .i ⇒ hNl+1 , . . . , p′j , . . .i where Nl+1 = Nl ∪ {v}, p′j = hs′ , n′1 , . . . , n′k i
where n′i = n, and n′j = nj for j 6= i.
We note than that the encoding of the previous rule returns the M SRN C rule
s(x1 , . . . , xk ) | f resh(x) → s′ (x′1 , . . . , x′k ) | f resh(x′ ) : ϕ
where ϕ consists of the constraints y > x′i , x′i > x and x′j = xj for j 6= i. By
cl (hl ) is a multiset of atomic formulas that contains the
inductive hypothesis, G
18
Giorgio Delzanno
formulas s(hl (n1 ), . . . , hl (nk )) and f resh(v) where hl is a mapping into Dl , and v
is the first non-negative rational strictly greater than all values occurring in the
formulas denoting processes.
Let v be a non negative rational strictly greater than all values in Dl . Furthermore,
let us define v ′ = v + 1 and Dl+1 = Dl ∪ {v, v ′ }.
Furthermore, we define hl+1 as follows hl+1 (n) = hl (n) for n ∈ Nl , and hl+1 (n′i ) =
hl+1 (n) = v ′ . Furthermore, we define the following evaluation
σ
= h x 7→ v, x1 7→ hl (n1 ), . . . , xk 7→ hl (nk ),
x′ 7→ v ′ , x′1 7→ hl+1 (n′1 ), . . . , x′k 7→ hl+1 (n′k ) i
Then, by construction of σ and α
b, it follows that σ is a solution for α
b. Thus,
s(n1 , . . . , nk ) | f resh(v) → s′ (n′1 , . . . , n′k ) | f resh(v ′ )
is a ground instance of the considered M SRN C rule.
cl (hl ), if we apply a rewriting step we
Starting from the M SRN C configuration G
obtain a new configuration in which s(n1 , . . . , nk ) and f resh(v) are substituted by
[
s′ (n′1 , . . . , n′k ) and f resh(v ′ ), and all the other atomic formulas in G
l+1 (hl+1 ) are
c
the same as in Gl (hl ). We conclude by noting that this formula coincides with the
[
definition of G
l+1 (hl+1 ).
For sake of brevity we omit the case of thread creation whose only difference from
the previous cases is the creation of several new atoms instead (with values obtained
by evaluating the action) of only one.
Rendez-vous Let pi = hs, n1 , . . . , nk i and pj = ht, m1 , . . . , mu i two local configurations for threads P 6= P ′ , ni ∈ Nl for i : 1, . . . , k and mi ∈ Nl for i : 1, . . . , u.
c!m
c?m′
Suppose s −−→ s′ [γ, α] and t −−−→ t′ [γ ′ , α′ ], where m = hxi1 , . . . , xiv i, and m′ =
hy1 , . . . , yv i ( all defined over distinct variables) are rules in R.
Furthermore, suppose that ρpi satisfies γ, and that ρ′ (see definition of the operational semantics) satisfies γ ′ , and suppose that Gl = hNl , . . . , pi , . . . , pj , . . .i ⇒
hNl+1 , . . . , p′i , . . . , p′j , . . .i = Gl+1 , where Nl+1 = Nl , p′i = hs′ , n′1 , . . . , n′k i, p′j =
ht′ , m′1 , . . . , m′u i, and if xi := e occurs in α, then n′i = ρpi (e), otherwise n′i = ni
for i : 1, . . . , k; if ui := e occurs in α′ , then m′i = ρ′ (e), otherwise m′i = mi for
i : 1, . . . , u.
cl (hl ) is a multiset of atomic formulas that contains the
By inductive hypothesis, G
formulas s(hl (n1 ), . . . , hl (nk )) and t(hl (m1 ), . . . , hl (mu )).
Now, let us define hl+1 as the mapping from Nl to Dl such that hl+1 (n′i ) = hl (nj )
if xi := xj is in α, hl+1 (m′i ) = hl (mj ) if ui := uj is in α′ , hl+1 (n′i ) = 0 if xi := ⊥ is
in α, hl+1 (m′i ) = 0 if ui := ⊥ is in α′ .
Now, let us define σ as the evaluation from Nl to Dl such that
σ = σ1 ∪ σ2
σ1 = hx1 7→ hl (n1 ), . . . , xk 7→ hl (nk ), u1 7→ hl (m1 ), . . . , uu 7→ hl (mu )i
σ2 = hx′1 7→ hl+1 (n′1 ), . . . , x′k 7→ hl+1 (n′k ), u′1 7→ hl+1 (m′1 ), . . . , u′u 7→ hl+1 (m′u )i.
Then, by construction of the sets of constraints γ
b, γb′ , α
b and αb′ it follows that σ is
′
′
′
a solution for the constraint ∃w1 . . . . ∃wp .θ ∧ θ ∧ α
b ∧ αb′ ∧ w1 = w1′ ∧ . . . ∧ wp = wp′
for some θ ∈ γ
b and θ′ ∈ γb′ . Note in fact that the equalities wi = wi′ express the
Constraint-based Verification of Abstract Multithreaded Programs
19
passing of values defined via the evaluation ρ′ in the operational semantics.
As a consequence,
s(n1 , . . . , nk ) | t(m1 , . . . , mu ) → s′ (n′1 , . . . , n′k ) | t′ (m′1 , . . . , m′u )
is a ground instance of one of the considered M SRN C rules.
cl (hl ), if we apply a rewriting
Thus, starting from the M SRN C configuration G
step we obtain a new configuration in which s(n1 , . . . , nk ) has been replaced by
s′ (n′1 , . . . , n′k ), and t′ (m′1 , . . . , m′k ) has been replaced by t(m′1 , . . . , m′u ), and all the
cl (hl ). This formula coincides with the definition
other atomic formulas are as in G
[
of Gl+1 (hl+1 ).
The proof of completeness is by induction on the length of an MSR run, and by
case-analysis on the application of the rules. The structure of the case analysis is
similar to the previous one and it is omitted for brevity.
4 Verification of TDL Programs
Safety and invariant properties are probably the most important class of correctness
specifications for the validation of a concurrent system. For instance, in Example
1 we could be interested in proving that every time a session terminates, two instances of thread Init and Resp have exchanged the two names generated during
the session. To prove the protocol correct independently from the number of names
and threads generated during an execution, we have to show that from the initial configuration G0 it is not possible to reach a configuration that violates the
aforementioned property. The configurations that violate the property are those in
which two instances of Init and Resp conclude the execution of the protocol exchanging only the first nonce. These configurations can be represented by looking
at only two threads and at the relationship among their local data. Thus, we can
reduce the verification problem of this safety property to the following problem:
Given an initial configuration G0 we would like to decide if a global configuration that contains at least two local configurations having the form hstopA , i, n, mi
and hstopB , i′ , n′ , m′ i with n′ = n and m 6= m′ for some i, i′ , n, n′ , m, m′ is reachable. This problem can be viewed as an extension of the control state reachability
problem defined in (Abdulla and Nylén 2000) in which we consider both control locations and local variables. Although control state reachability is undecidable (see
Corollary 1), the encoding of TDL into MSRN C can be used to define a sound and
automatic verification methods for TDL programs. For this purpose, we will exploit
a verification method introduced for MSR(C) in (Delzanno 2001; Delzanno 2005).
In the rest of this section we will briefly summarize how to adapt the main results
in (Delzanno 2001; Delzanno 2005) to the specific case of MSRN C .
Let us first reformulate the control state reachability problem of Example 1 for
the aforementioned safety property on the low level encoding into MSRN C . Given
the MSRN C initial configuration init we would like to check that no configuration
in P ost∗ ({init}) has the following form
{stopA (a1 , v1 , w1 ), stopB (a2 , v2 , w2 )} ⊕ M
20
Giorgio Delzanno
for ai , vi , wi ∈ Q i : 1, 2 and an arbitrary multiset of ground atoms M. Let us call
U the set of bad MSRN C configurations having the aforementioned shape. Notice
that U is upward closed with respect to multiset inclusion, i.e., if M ∈ U and
M 4 M′ , then M′ ∈ U . Furthermore, for if U is upward closed, so is P re(U ).
On the basis of this property, we can try to apply the methodology proposed in
(Abdulla and Nylén 2000) to develop a procedure to compute a finite representation
R of P re∗ U ). For this purpose, we need the following ingredients:
1. a symbolic representation of upward closed sets of configurations (e.g. a set
of assertions S whose denotation [[S]] is U );
2. a computable symbolic predecessor operator SP re working on sets of formulas
such that [[SP re(S)]] = P re([[S]]);
3. a (decidable) entailment relation Ent to compare the denotations of symbolic
representations, i.e., such that Ent(N, M ) implies [[N ]] ⊆ [[M ]]. If such a relation Ent exists, then it can be naturally extended to sets of formulas as
follows: EntS (S, S ′ ) if and only if for all N ∈ S there exists M ∈ S ′ such that
Ent(N, M ) holds (clearly, if Ent is an entailment, then EntS (S, S ′ ) implies
[[S]] ⊆ [[S ′ ]]).
The combination of these three ingredients can be used to define a verification
methods based on backward reasoning as explained next.
Symbolic Backward Reachability Suppose that M1 , . . . , Mn are the formulas of our
assertional language representing the infinite set U consisting of all bad configurations. The symbolic backward reachability procedure (SBR) procedure computes a
chain {Ii }i≥0 of sets of assertions such that
I0 = {M1 , . . . , Mn }
Ii+1 = Ii ∪ SP re(Ii ) for i ≥ 0
The procedure SBR stops when SP re produces only redundant information, i.e.,
EntS (Ii+1 , Ii ). Notice that EntS (Ii , Ii+1 ) always holds since Ii ⊆ Ii+1 .
Symbolic Representation In order to find an adequate represention of infinite sets of
MSRN C configurations we can resort to the notion of constrained configuration introduced in (Delzanno 2001) for the language scheme MSR(C) defined for a generic
constraint system C. We can instantiate this notion with N C constraints as follows.
A constrained configuration over P is a formula
p1 (x11 , . . . , x1k1 ) | . . . | pn (xn1 , . . . , xnkn ) : ϕ
where p1 , . . . , pn ∈ P, xi1 , . . . , xiki ∈ V for any i : 1, . . . n and ϕ is an N C-constraint.
.
The denotation a constrained configuration M = (M : ϕ) is defined by taking the
upward closure with respect to multiset inclusion of the set of ground instances,
namely
[[M ]] = {M′ | σ(M) 4 M′ , σ ∈ Sol(ϕ)}
Constraint-based Verification of Abstract Multithreaded Programs
21
This definition can be extended to sets of MSRN C constrained configurations with
disjoint variables (we use variable renaming to avoid variable name clashing) in the
natural way.
In our example the following set SU of MSRN C constrained configurations (with
distinct variables) can be used to finitely represent all possible violations U to the
considered safety property
SU = { stopA (i1 , n1 , m1 ) | stopB (i2 , n2 , m2 ) : n1 = n2 , m1 > m2
stopA (i1 , n1 , m1 ) | stopB (i2 , n2 , m2 ) : n1 = n2 , m2 > m1 }
Notice that we need two formulas to represent m1 6= m2 using a disjunction
of > constraints. The MSRN C configurations stopB (1, 2, 6) | stopA (4, 2, 5), and
stopB (1, 2, 6) | stopA (4, 2, 5) | waitA (2, 7, 3) are both contained in the denotation
of SU . Actually, we have that [[SU ]] = U . This symbolic representation allows us to
reason on infinite sets of MSRN C configurations, and thus on global configurations
of a TDL program, forgetting the actual number or threads of a given run.
To manipulate constrained configurations, we can instantiate to N C-constraints
the symbolic predecessor operator SP re defined for a generic constraint system in
(Delzanno 2005). Its definition is also given in Section Appendix A in Appendix.
From the general properties proved in (Delzanno 2005), we have that when applied
to a finite set of MSRN C constrained configurations S, SP reN C returns a finite set
of constrained configuration such that [[SP reN C (S)]] = P re([[S]]), i.e., SP reN C (S)
is a symbolic representation of the immediate predecessors of the configurations in
the denotation (an upward closed set) of S. Similarly we can instantiate the generic
entailment operator defined in (Delzanno 2005) to MSRN C constrained configurations so as to obtain an a relation Ent such that EntN C (N, M ) implies [[N ]] ⊆ [[M ]].
Based on these properties, we have the following result.
Proposition 1
Let T be a TDL program with initial global configuration G0 , Furthermore, let
S be the corresponding MSRN C encoding. and SU be the set of MSRN C constrained configurations denoting a given set of bad TDL configurations. Then,
init 6∈ SP re∗N C (SU ) if and only if there is no finite run G0 . . . Gn and mappings h0 , . . . , hn from the names occurring in G to non-negative rationals such
that init• G•0 (h0 ) . . . G•n (hn ) is a run in S and G•n (hn ) ∈ [[U ]].
Proof
Suppose init 6∈ SP re∗N C (U ). Since [[SP reN C (S)]] = pre([[S]]) for any S, it follows
that there cannot exist runs initM0 . . . Mn in S such that Mn ∈ [[U ]]. The thesis
then follows from the Theorem 2.
As discussed in (Bozzano and Delzanno 2002), we have implemented our verification procedure based on M SR and linear constraints using a CLP system with linear
arithmetics. By the translation presented in this paper, we can now reduce the verification of safety properties of multithreaded programs to a fixpoint computation
built on constraint operations. As example, we have applied our CLP-prototype
to automatically verify the specification of Fig. 6. The unsafe states are those described in Section 4. Symbolic backward reachability terminates after 18 iterations
22
Giorgio Delzanno
and returns a symbolic representation of the fixpoint with 2590 constrained configurations. The initial state init is not part of the resulting set. This proves our
original thread definitions correct with respect to the considered safety property.
4.1 An Interesting Class of TDL Programs
The proof of Theorem 1 shows that verification of safety properties is undecidable for TDL specifications in which threads have several local variables (they
are used to create linked lists). As mentioned in the introduction, we can apply the sufficient conditions for the termination of the procedure SBR given in
(Bozzano and Delzanno 2002; Delzanno 2005) to identify the following interesting
subclass of TDL programs.
Definition 4
A monadic TDL thread definition P = hQ, s, V, Ri is such that V is at most a
singleton, and every message template in R has at most one variable.
A monadic thread definition can be encoded into the monadic fragment of MSRN C
studied in (Delzanno 2005). Monadic MSRN C specifications are defined over atomic
formulas of the form p or p(x) with p is a predicate symbol and x is a variable, and
on atomic constraints of the form x = y, and x > y. To encode a monadic TDL
thread definitions into a Monadic MSRN C specification, we first need the following
observation. Since in our encoding we only use the constant 0, we first notice that
we can restrict our attention to MSRN C specifications in which constraints have no
constants at all. Specifically, to encode the generation of fresh names we only have
to add an auxiliary atomic formula zero(z), and refer to it every time we need to
express the constant 0. As an example, we could write rules like
init −−→ f resh(x) | initM (y) | zero(z) : x > z, y = z
for initialization, and
create(x) | zero(z) −−→ initM (x′ ) | initA (id′ , n′ , m′ ) | zero(z) :
x′ = x, id′ = x, n′ = z, m′ = z, z ′ = z
for all assignments involving the constant 0. By using this trick an by following the
encoding of Section 3, the translation of a collection of monadic thread definitions
directly returns a monadic MSRN C specification. By exploiting this property, we
obtain the following result.
Theorem 3
The verification of safety properties whose violations can be represented via an
upward closed set U of global configurations is decidable for a collection T of
monadic TDL definitions.
Proof
Let S be the MSRN C encoding of T and SU be the set of constrained configuration
such that SU = U . The proof is based on the following properties. First of all, the
MSRN C specification S is monadic. Furthermore, as shown in (Delzanno 2005),
Constraint-based Verification of Abstract Multithreaded Programs
23
the class of monadic MSRN C constrained configurations is closed under application
of the operator SP reN C . Finally, as shown in (Delzanno 2005), there exists an
entailment relation CEnt for monadic constrained configurations that ensures the
termination of the SBR procedure applied to a monadic MSRN C specification.
Thus, for the monadic MSRN C specification S, the chain defined as I0 = SU ,
Ii+1 = Ii ∪ SP re(Ii ) always reaches a point k ≥ 1 in which CEntS (Ik+1 , Ik ), i.e.
[[Ik ]] is a fixpoint for P re. Finally, we note that we can always check for membership
of init in the resulting set Ik .
As shown in (Schnoebelen 2002), the complexity of verification methods based on
symbolic backward reachability relying on the general results in (Abdulla and Nylén 2000;
Finkel and Schnoebelen 2001) is non primitive recursive.
5 Conclusions and Related Work
In this paper we have defined the theoretical grounds for the possible application
of constraint-based symbolic model checking for the automated analysis of abstract
models of multithreaded concurrent systems providing name generation, name mobility, and unbounded control. Our verification approach is based on an encoding
into a low level formalism based on the combination of multiset rewriting and
constraints that allows us to naturally implement name generation, value passing,
and dynamic creation of threads. Our verification method makes use of symbolic
representations of infinite set of system states and of symbolic backward reachability. For this reason, it can be viewed as a conservative extension of traditional
finite-state model checking methods. The use of symbolic state analysis is strictly
related to the analysis methods based on abstract interpretation. A deeper study
of the connections with abstract interpretation is an interesting direction for future
research.
Related Work The high level syntax we used to present the abstract models of
multithreaded programs is an extension of the communicating finite state machines
used in protocol verification (Bochmann 1978), and used for representing abstraction of multithreaded software programs (Ball et al. 2001). In our setting we enrich
the formalism with local variables, name generation and mobility, and unbounded
control. Our verification approach is inspired by the recent work of Abdulla and
Jonsson. In (Abdulla and Jonsson 2003), Abdulla and Jonsson proposed an assertional language for Timed Networks in which they use dedicated data structures
to symbolically represent configurations parametric in the number of tokens and
in the age (a real number) associated to tokens. In (Abdulla and Nylén 2000), Abdulla and Nylén formulate a symbolic algorithm using existential zones to represent the state-space of Timed Petri Nets. Our approach generalizes the ideas of
(Abdulla and Jonsson 2003; Abdulla and Nylén 2000) to systems specified via multiset rewriting and with more general classes of constraints. In (Abdulla and Jonsson 2001),
the authors apply similar ideas to (unbounded) channel systems in which messages
can vary over an infinite name domain and can be stored in a finite (and fixed a
24
Giorgio Delzanno
priori) number of data variables. However, they do not relate these results to multithreaded programs. Multiset rewriting over first order atomic formulas has been proposed for specifying security protocols by Cervesato et al. in (Cervesato et al. 1999).
The relationships between this framework and concurrent languages based on process algebra have been recently studied in (Bistarelli et al. 2005). Apart from approaches based on Petri Net-like models (as in (German and Sistla 1992; Ball et al. 2001)),
networks of finite-state processes can also be verified by means of automata theoretic techniques as in (Bouajjani et al. 2000). In this setting the set of possible
local states of individual processes are abstracted into a finite alphabet. Sets of
global states are represented then as regular languages, and transitions as relations
on languages. Differently from the automata theoretic approach, in our setting
we handle parameterized systems in which individual components have local variables that range over unbounded values. The use of constraints for the verification
of concurrent systems is related to previous works connecting Constraint Logic
Programming and verification, see e.g. (Delzanno and Podelski 1999). In this setting transition systems are encoded via CLP programs used to encode the global
state of a system and its updates. In the approach proposed in (Delzanno 2001;
Bozzano and Delzanno 2002), we refine this idea by using multiset rewriting and
constraints to locally specify updates to the global state. In (Delzanno 2001), we
defined the general framework of multiset rewriting with constraints and the corresponding symbolic analysis technique. The language proposed in (Delzanno 2001) is
given for a generic constraint system C (taking inspiration from CLP the language
is called M SR(C)). In (Bozzano and Delzanno 2002), we applied this formalism to
verify properties of mutual exclusion protocols (variations of the ticket algorithm)
for systems with an arbitrary number of processes. In the same paper we also formulated sufficient conditions for the termination of the backward analysis. The
present paper is the first attempt of relating the low level language proposed in
(Delzanno 2001) to a high level language with explicit management of names and
threads.
Acknowledgments The author would like to thank Ahmed Bouajjani, Andrew Gordon, Fabio Martinelli, Catuscia Palamidessi, Luca Paolini, and Sriram Rajamani
and the anonymous reviewers for several fruitful comments and suggestions.
References
Abdulla, P. A., Cerāns, K., Jonsson, B., and Tsay, Y.-K. 1996. General Decidability Theorems for Infinite-State Systems. In Proceedings 11th Annual International
Symposium on Logic in Computer Science (LICS’96). IEEE Computer Society Press,
New Brunswick, New Jersey, 313–321.
Abdulla, P. A. and Jonsson, B. 2001. Ensuring Completeness of Symbolic Verification
Methods for Infinite-State Systems. Theoretical Computer Science 256, 1-2, 145–167.
Abdulla, P. A. and Jonsson, B. 2003. Model checking of systems with many identical
timed processes. Theoretical Computer Science 290, 1, 241–264.
Abdulla, P. A. and Nylén, A. 2000. Better is Better than Well: On Efficient Verification
of Infinite-State Systems. In Proceedings 15th Annual International Symposium on
Constraint-based Verification of Abstract Multithreaded Programs
25
Logic in Computer Science (LICS’00). IEEE Computer Society Press, Santa Barbara,
California, 132–140.
Ball, T., Chaki, S., and Rajamani, S. K. 2001. Parameterized Verification of Multithreaded Software Libraries. In 7th International Conference on Tools and Algorithms
for Construction and Analysis of Systems (TACAS 2001), Genova, Italy, April 2-6,.
LNCS, vol. 2031. Springer-Verlag, 158–173.
Bistarelli, S., Cervesato, I., Lenzini, G., and Martinelli, F. 2005. Relating multiset
rewriting and process algebras for security protocol analysis. Journal of Computer
Security 13, 1, 3–47.
Bochmann, G. V. 1978. Finite state descriptions of communicating protocols. Computer
Networks 2, 46–57.
Bouajjani, A., Jonsson, B., Nilsson, M., and Touili, T. 2000. Regular Model Checking. In Proceedings 12th International Conference on Computer Aided Verification
(CAV’00), E. A. Emerson and A. P. Sistla, Eds. LNCS, vol. 1855. Springer-Verlag,
Chicago, Illinois, 403–418.
Bozzano, M. and Delzanno, G. 2002. Algorithmic verification of invalidation-based
protocols. In 14th International Conference on Computer Aided Verification, CAV ’02.
Lecture Notes in Computer Science, vol. 2404. Springer.
Cervesato, I., Durgin, N., Lincoln, P., Mitchell, J., and Scedrov, A. 1999. A
Meta-notation for Protocol Analysis. In 12th Computer Security Foundations Workshop
(CSFW’99). IEEE Computer Society Press, Mordano, Italy, 55–69.
Delzanno, G. 2001. An Assertional Language for Systems Parametric in Several Dimensions. In Verification of Parameterized Systems - VEPAS 2001. ENTCS, vol. 50.
Delzanno, G. 2005. Constraint Multiset Rewriting. Tech. Rep. TR-05-08, Dipartimento
Informatica e Scienze dell’Informazione, Università di Genova, Italia.
Delzanno, G. and Podelski, A. 1999. Model checking in CLP. In Proceedings 5th
International Conference on Tools and Algorithms for Construction and Analysis of
Systems (TACAS’99). Lecture Notes in Computer Science, vol. 1579. Springer-Verlag,
Amsterdam, The Netherlands, 223–239.
Finkel, A. and Schnoebelen, P. 2001. Well-Structured Transition Systems Everywhere!
Theoretical Computer Science 256, 1-2, 63–92.
German, S. M. and Sistla, A. P. 1992. Reasoning about Systems with Many Processes.
Journal of the ACM 39, 3, 675–735.
Gordon, A. D. 2001. Notes on nominal calculi for security and mobility. In Foundations
of Security Analysis and Design, Tutorial Lectures. Lecture Notes in Computer Science,
vol. 2171. Springer, 262–330.
Kesten, Y., Maler, O., Marcus, M., Pnueli, A., and Shahar, E. 2001. Symbolic
model checking with rich assertional languages. Theoretical Computer Science 256, 1,
93–112.
Schnoebelen, P. 2002. Verifying Lossy Channel Systems has Nonprimitive Recursive
Complexity. Information Processing Letters 83, 5, 251–261.
Appendix A Symbolic Predecessor Operator
Given a set of MSRN C configurations S, consider the MSRN C predecessor operator
P re(S) = {M|M ⇒MSR M′ , M′ ∈ S}. In our assertional language, we can define a
symbolic version SP reN C of P re defined on a set S containing MSRN C constrained
26
Giorgio Delzanno
multisets (with disjoint variables) as follows:
SP reN C (S) = { (A ⊕ N : ξ) |
(A −→ B : ψ) ∈ R, (M : ϕ) ∈ S,
M′ 4 M, B ′ 4 B,
(M′ : ϕ) =θ (B ′ : ψ), N = M ⊖ M′ ,
ξ ≡ (∃x1 . . . . xk .θ)
and x1 , . . . , xk are all variables not in A ⊕ N }.
where =θ is a matching relation between constrained configurations that also takes
in consideration the constraint satisfaction, namely
(A1 | . . . | An : ϕ) =θ (B1 | . . . | Bm : ψ)
provided m = n and there exists a permutation j1 , . . . , jn of 1, . . . , n such that
Vn
the constraint θ = ϕ ∧ ψ ∧ i=1 Ai = Bji is satisfiable; here p(x1 , . . . , xr ) =
q(y1 , . . . , ys ) is an abbreviation for the constraints x1 = y1 ∧ . . . ∧ xr = ys if p = q
and s = r, f alse otherwise.
As proved in (Delzanno 2005), the symbolic operator SP reN C returns a set of
MSRN C constrained configurations and it is correct and complete with respect to
P re, i.e., [[SP reN C (S)]] = P re([[S]]) for any S. It is important to note the difference
between SP reN C and a simple backward rewriting step.
For instance, given the constrained configurations M defined as p(x, z) | f (y) : z >
y and the rule s(u, m) | r(t, v) → p(u′ , m′ ) | r(t′ , v ′ ) : u = t, m′ = v, v ′ =
v, u′ = u, t′ = t (that simulates a rendez-vous (u, t are channels) and value passing
(m′ = v)), the application of SP re returns s(u, m) | r(t, v) | f (y) : u = t, v > y as
well as s(u, m) | r(t, v) | p(x, z) | f (y) : u = t, x > y (the common multiset here is
ǫ).
| 2cs.AI
|
The Shape of Art History in the Eyes of the Machine∗
Marian Mazzone2
Bingchen Liu1
Diana Kim1
Mohamed Elhoseiny1
The Art and Artificial Intelligence Laboratory
1
Department of Computer Science, Rutgers University, NJ, USA
2
Department of Art History, College of Charleston, SC, USA
arXiv:1801.07729v2 [cs.AI] 12 Feb 2018
Ahmed Elgammal1†
Abstract
How does the machine classify styles in art? And how does it relate to art historians’ methods for analyzing style? Several studies have shown the ability of the machine to learn and
predict style categories, such as Renaissance, Baroque, Impressionism, etc., from images of
paintings. This implies that the machine can learn an internal representation encoding discriminative features through its visual analysis. However, such a representation is not necessarily
interpretable. We conducted a comprehensive study of several of the state-of-the-art convolutional neural networks applied to the task of style classification on 77K images of paintings,
and analyzed the learned representation through correlation analysis with concepts derived
from art history. Surprisingly, the networks could place the works of art in a smooth temporal
arrangement mainly based on learning style labels, without any a priori knowledge of time
of creation, the historical time and context of styles, or relations between styles. The learned
representations showed that there are few underlying factors that explain the visual variations
of style in art. Some of these factors were found to correlate with style patterns suggested
by Heinrich Wölfflin (1846-1945). The learned representations also consistently highlighted
certain artists as the extreme distinctive representative of their styles, which quantitatively confirms art historian observations.
Keywords: Artificial Intelligence — Computational Art History
∗
This paper is an extended version of a paper that will be published on the 32nd AAAI conference on Artificial
Intelligence, to be held in New Orleans, USA, February 2-7, 2018
†
Corresponding author: Ahmed Elgammal [email protected]
1
Art history could prove “to contain unexpected potentialities as a predictive science” – George
Kubler, The Shape of Time: Remarks on the History of Things, 1962.
1
Introduction
Style is central to the discipline of art history. The word “style” is used to refer to the individual
way or manner that someone makes or does something, for example Rembrandt’s style of painting.
Style also refers to groups of works that have a similar typology of characteristics, such as the
Impressionist style, or High Renaissance style. Art historians identify, characterize, and define
styles based on the evidence of the physical work itself, in combination with an analysis of the
cultural and historical features of the time and place in which it was made. Although we see style,
and we all know that it exists, there is still no central, agreed upon theory of how style comes about,
or how and why it changes. Some of the best scholars of art history have written persuasively about
the importance of style to the discipline, and the concomitant difficulty of defining or explaining
what it is and why it changes [1, 2]. Up to now connoisseurship has proven to be one of the
most effective means to detect the styles of various artists, and differentiate style categories and
distinctions in larger movements and periods.
Recent research in computer vision and machine learning have shown the ability of the machine to learn to discriminate between different style categories such as Renaissance, Baroque,
Impressionism, Cubism, etc., with reasonable accuracy, (e.g. [3, 4, 5, 6]). However, classifying
style by the machine is not what interests art historians. Instead, the important issues are what
machine learning may tell us about how the characteristics of style are identified, and the patterns
or sequence of style changes. The ability of the machine to classify styles implies that the machine has learned an internal representation that encodes discriminative features through its visual
analysis of the paintings. However, it is typical that the machine uses visual features that are not
interpretable by humans. This limits the ability to discover knowledge out of these results.
Our study’s emphasis is on understanding how the machine achieves classification of style,
what internal representation it uses to achieve this task, and how that representation is related to art
history methodologies for identifying styles. To achieve such understanding, we utilized one of the
key formulations of style pattern and style change in art history, the theory of Heinrich Wölfflin
(1846-1945). Wölfflin’s comparative approach to formal analysis has become a standard method of
art history pedagogy. Wölfflin chose to separate form analysis from discussions of subject matter
and expression, focusing on the “visual schema” of the works, and how the “visible world crystallized for the eye in certain forms” [7]. Wölfflin identified pairs of works of art to demonstrate
style differences through comparison and contrast exercises that focused on key principles or features. Wölfflin used his method to differentiate the Renaissance from the Baroque style through
five key visual principles: linear/painterly, planar/recessional, closed form/open form, multiplicity/unity, absolute clarity/relative clarity. Wölfflin posited that form change has some pattern of
differentiation, such that style types and changes can only come into being in certain sequences.
With advances in computer vision and machine learning and the availability of comprehensive
datasets of images, we are now positioned to approach the history of art as a predictive science, and
relate its means of determining questions of style to machine results. Lev Manovich has argued
that the use of computational methods, while providing a radical shift in scale, in fact continues the
humanities’ traditional methodologies [8]. It was nearly impossible to apply and empirically test
2
Wölfflin’s methods of style differentiation and analysis before developments in computer science.
No human being would assemble the number of examples needed to prove the value of his methods
for finding discriminative features. Nor could anyone amass the dataset necessary to demonstrate
the usefulness of his model for processing style description and sequencing beyond the immediate
examples of Renaissance and Baroque via his five pairs. We chose Wölfflin’s theory because of his
emphasis on formal, discriminative features and the compare/contrast logic of his system, qualities
that make it conducive to machine learning. Today, art historians use a wide variety of methods that
are not only solely focused on form, but for the type of analysis of this paper Wölfflin’s approach
is useful.
Methodology
Deep convolutional neural networks have recently played a transformative role in advancing artificial intelligence [9]. We evaluated a large number of state-of-the-art deep convolutional neural
network models, and variants of them, trained to classify styles. We focused on increasing the
interpretability of the learned presentation by forcing the machine to achieve classification with
a reduced number of variables without sacrificing classification accuracy. We then analyzed the
achieved representations through linear and nonlinear dimensionality reduction of the activation
space, visualization, and correlation analysis with time and with Wölfflin’s pairs. We used a collection of 77K digitized paintings to train, validate and test the models. We utilized two sets of
digitized paintings for visualization and correlation analysis of the achieved representations. In
particular, we used variants of AlexNet [10], VGGNet [11], ResNet [12], which were originally
developed for the task of object categorization for the ImageNet challenge [13] and each of them
raised the state of the art in that task when they were introduced. We adapted these networks
for classifying 20 style classes. Our study included varying the training strategies (training from
scratch on art data vs. using pre-trained models and fine-tuning them on art data), varying the
network architecture, and data augmentation strategies.
Main Results
Trained to predict styles, based only on noisy discrete style labels, without being given any notion
of time, the machine encoded art history in a smooth chronology. The learned representation can
be explained based on a handful of factors. The first two modes of variations are aligned with the
concepts of linear vs. painterly and planer vs. recessional suggested by Heinrich Wölfflin (18461945), and quantitatively explain most of the variance in art history, where temporal progression
correlates radially across these modes (See Figure 1). The importance of these results is that they
show that the selected art historian’s theories about style change can be quantifiably verified using
scientific methods. The results also show that style, which appears to be a subjective issue, can be
computationally modeled with objective means.
3
A
Monet
Renoir
Sisley
Cézanne
Cézanne
Marc
Picasso
Sisley
Picasso
Gris
Manet
Braque
Goya
Rubens
Malevich
Velázquez
Rembrandt
Warhol
Rembrandt
Van Eyck
Raphael
Malevich
Da Vinci
B
C
Rothko
Mondrian
Correla'on)with)
Wölfflin's)concepts)
Raphael
2%
1%
!0.6% !0.5% !0.4% !0.3% !0.2% !0.1%
0%
0.1% 0.2% 0.3% 0.4%
1%
2%
Absolute%vs%Rela6ve%
Clarity%
0.0015%
0.198%
Mul6plicity%vs%Unity%
!0.0302%
0.2267%
Planer%vs%Recession%
!0.4946%
!0.0736%
Close%vs%Open%
0.2448%
0.1724%
Linear%vs%Painterly%
!0.2336%
0.3547%
Figure 1: Modes of variations of the activation subspace showing smooth temporal transition and correlation with
Wölfflin’s concepts. (A) First and second modes of variations of the AlexNet+2 model with paintings color-coded
by date of creation. The first mode (the horizontal axis) seems to correlate with figurative art, which was dominant
till Impressionism, vs. non-figurative, distorted figures, and abstract art that dominate 20th century styles. Another
interpretation for this dimension is that it reflects Wölfflin’s concept of plane (to the right) vs. recession (to the left).
This axis correlates the most with Wölfflin’s concept of plane vs. recession with -0.50 PCC. The vertical axis correlates
with the linear (towards the bottom) vs. painterly (towards the top) concept (0.36 PCC). (B) The angular coordinate
exhibits strong correlation with time (PCC of 0.69). (C) Correlation with Wölfflin’s concepts.
4
2
Detailed Methodology
2.1
Challenges with Art Style Classification
In contrast to the typical object classification in images, the problem of style classification in art has
a different set of challenges that we highlight here. Style is not necessarily correlated with subject
matter, which corresponds to existence of certain objects in the painting. Style is mainly related
to the form and can be correlated with features at different levels, low, medium and high-level. As
a result, it is not necessary that networks which perform better for extracting semantic concepts,
such as object categories, would be perform as well in style classification. In the literature of object
classification, deeper networks were shown to perform better [11, 12] since they facilitate richer
representations to be learned at different levels of features. We do not know if a deeper network
will necessarily be better in the style classification domain. This remains something to discover
through our empirical study. However, the challenge, in the context of art style classification, is
the lack of images on a scale of magnitude similar to ImageNet (million of images). The largest
publicly available dataset, which we use, is only in the order of 80K images. This limitation is due
to copyright issue, which is integral to the domain of art. Moreover, collecting annotation in this
domain is hard since it requires expert annotators and typical crowd sourcing annotators are not
qualified.
Another fundamental difference is that styles do no lend themselves to discrete mutually exclusive classes as in supervised machine learning. Style transition over time is typically smooth, and
style labels are after-the-fact concepts imposed by art historians, sometimes centuries later. Paintings can have elements that belongs to multiple styles, and therefore not necessarily identifiable
with a unique style class.
Paintings also come in wide variety of sizes. The canvas can span a couple of hundreds of
square feet or can be as small as a couple of square inches. The aspect ratio can vary very significantly as well. A typical CNN requires its input to be resized to a fixed size retinal array. This
bound to introduce geometric distortions, which can affect the composition, and loss of details related to surface texture carrying essential information about brush strokes. Both composition and
brush strokes are essential to identify style. In this paper we do no address solving these issues.
We mainly report the behavior of the studied model for style classification despite their limitation
to address these issues.
2.2
Datasets
Training-testing Set: We trained, validated, and tested the networks using paintings from the publicly available WikiArt dataset1 . This collection (as downloaded in 2015) has images of 81,449
paintings from 1,119 artists ranging from the fifteenth century to contemporary artists. Several
prior studies on style classification used subsets of this dataset (e.g. [4, 5, 6]). Originally WikiArt
has 27 style classes. For the purpose of our study we reduced the number of classes to 20 classes by
merging fine-grained style classes with small number of images, for example we merged cubism,
analytical cubism, and synthetic cubism into one cubism class. Table 1 shows a list of the 20 style
classes used and their mapping from the original WikiArt classes. We also ignored the symbolism
1
Wikiart dataset http://www.wikiart.org
5
style class since in the Wikiart collection this class contains various paintings from different styles
that erroneously labeled as “symbolism”. In general, by visual inspection we notice that the style
labels in the WikiArt collection is noisy and not accurate. However, it remains the largest publicly available collection. Other available collections, which might be cleaner in their labels, are
typically much smaller in size for the purpose of training a deep network. We excluded from the
collections images of sculptures and photography. The total number of images used for training,
validation, and testing are 76,921 images. We split the data into training (85%), validation (9.5%)
and test sets (5.5%).
Table 1: Style classes used in learning the models
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
Style Class
Early Renaissance
High Renaissance
Mannerism and Late Renaissance
Northern Renaissance
Baroque
Rococo
Romanticism
Impressionism
Post-Impressionism
Realism
Art Nouveau
Cubism
Expressionism
Fauvism
Abstract-Expressionism
Color field painting
Minimalism
Naı̈ve art-Primitivism
Ukiyo-e
Pop-art
Total number of Images
Number of Images
1391
1343
1279
2552
4241
2089
7019
13060
6965
11528
4334
2562
6736
934
2881
1615
1337
2405
1167
1483
76921
Merged styles
Post Impressionism+Pointillism
Realism + Contemporary Realism + New Realism
Cubism + Analytical Cubism + Synthetic Cubism
Abstract Expressionism & Action Painting
Visualization Set I: We used another smaller dataset containing 1485 images of paintings from
the Artchive dataset2 to analyze and visualize the representation. Previous researches that have
used this dataset (e.g. [5]). We denote this dataset in this paper as the “Visualization dataset I”.
While the WikiArt collection is much bigger in size, this dataset contains a better representation
of the important works of western art from 1400-2000AD by 60 artists. Therefore, we mainly use
it to visualize and analyze the learned representation. We excluded from the collections images
of sculptures and images containing partial details of paintings. We also collected art historian’s
rating annotations (scale of 1 to 5) for each of the Wölfflin’s pairs for 1000 paintings from this data
set and use it in our correlation analysis.
Visualization Set II: We also used 62K painting from the Wikiart dataset for visualization and
analysis of the representations. We only included paintings that have date annotations for the
purposes of visualization and temporal correlation. The two chosen visualization sets have complementary properties. While the Artchive visualization set represents knowledge about influential
and important artists, the Wikiart is more arbitrarily chosen based on what is available in the public
2
Artchive dataset http://www.artchive.com
6
domain. The Artchive dataset lacks paintings from the 18th century and does not have broad sampling of early 19th paintings or post-war 20th century. In contrast, the Wikiart dataset is lacking
certain works of art that art historians would consider important, nevertheless it densely samples a
broad range of styles. In addition, the Wikiart dataset has a large bias towards 20th century art and
post WWII works.
2.3
Studied Deep Learning Models:
We performed a comprehensive comparative study on several deep convolutional networks, adapted
for the task of style classification. For the purpose of this paper, we report the results of three main
networks: AlexNet [10], VGGNet [11], and ResNet [12], as well as variants of them. All these
models were originally developed for the task of object recognition for the ImageNet challenge [13]
and each of them raised the state of the art when they were introduced.
For the non-expert reader, here we briefly summarize the main features of each of these models.
Deep convolutional networks in general consist of a sequence of layers of artificial neural units
of different types. Convolutional layers apply learned filters (templates) to each location of its
input image. These convolutional layers are interleaved with pooling layers, which aggregate the
responses of the convolution layers. Typically, the sequence of convolution and pooling layers
results in re-representing the visual information in the image as responses to a large number of
learned filters applied to wider and wider regions of the image as the information propagates deeper
in the network. Such filters are learned or tuned by the machine in response to the task in hand.
Finally the responses are passed to a sequence of fully connected layers that acts as a classifier.
AlexNet [10] architecture consists of five consecutive convolution layers, interleaved with some
pooling layers, followed by three fully connected layers, resulting in 1000 nodes representing the
object classes in ImageNet. The convolutional filters have different sizes at each layer, starting
from big filters (11x11) and reduced the filter size at following layers. In contrast VGGNet [11]
adapted an architecture of fixed size filters (3x3) over a deeper sequence of convolutional layers.
Residual Networks (ResNet [12]) introduced shortcut connection between the convolution layers
outputs and later layers, which results in much deeper architectures, reaching over 150 layers.
In our study, in general for all the models, the final softmax layer, originally designed for the
1000 classes in ImageNet, was removed and replaced with a layer of 20 softmax nodes, one for
each style class. Two modes of training were evaluated, 1) training the models from scratch on
the Wikiart data set, described above; 2) using a pre-trained model with ImageNet data and finetuned on the Wikiart data. Fine-tuning is the standard practice when adapting well-performing
pre-trained models to a different domain.
2.4
Increasing the interpretability of the representation
Having a large number of nodes at the fully connected layer allows the representation to project the
data into a very high dimensional space where classification would be easy (especially in our case
with only 20 classes), without forcing similar paintings across styles to come closer in the representation. To increase the interpretability of the representation, we force the network to achieve
classification through a lower dimension representation. To achieve this, after training the network
(whether from scratch or through fine-tuning), two more fully connected layers were added with
a reduced number of nodes. These reduced dimensional layers force the representation to use a
7
smaller number of degrees of freedom, which in turn forces paintings across styles to come closer
in the representation based on their similarity. In particular, we added two layers with 1024 and
512 nodes to all the models, and the models were then fine-tuned to adjust the weights for the new
layers. As will be shown later, adding these dimensionality reduction layers did not affect the classification accuracy. The experiments showed that gradually reducing the number of nodes in the
fully connected layers forced the network to achieve a “smoother” and interpretable representation.
It is important to emphasize that to achieve this effect without reducing the accuracy; the two new
layers have to be added after the network is trained, then models are fine-tuned. Training the full
architecture with the two extra layers, whether from scratch of fine-tuned, typically doesn’t result
in converging to similar accuracy.
We analyzed the modes of variations in the activation of each layer of the learned representation using Principle component analysis (PCA) [14]. We also analyzed the activation space using
Independent Component Analysis (ICA) [15]. We analyzed the nonlinear manifold of activations
through Laplacian Eigen Embedding (LLE) [16]. We chose these two techniques as widely used
representatives of linear and nonlinear dimensionality reduction techniques, each providing a different insight about the learned representation. We also performed correlation analysis between
the dimensions of the activation space and time as well as ground truth for Wölfflin’s concepts.
3
Quantitative Comparative Results
Table 2 shows the classification accuracy of different models using both pre-training with finetuning, and training from scratch. In all cases, the pre-trained and fine-tuned networks achieved
significantly better results than their counterparts that are trained from scratch (7% to 18% increase). This is not surprising and consistent with several models that adapted and fine-tuned
pre-trained networks for different domains. However, learned filters when the network trained
from scratch on the art domain were significantly different from the ones typically learned on ImageNet, which typically shows Gabor-like and blob-like filters. Figure 2 shows visualization of
the filters of AlexNet when trained on ImageNet compared to same filters trained on WikiArt for
style classification. While it is hard to interpret the filters trained for style classification, we do not
observe oriented-edge-like filters except for a horizontal edge filter. This emphasizes the difference
in nature between the problems and suggests that the better performance of the fine-tuned models
could be out-performed if sufficient data is available to train a style-classification network from
scratch on art data only.
Table 2: Comparison of classification of different models and different training methodologies
Network
AlexNet
AlexNet+2
VGGNet
VGGNet+2
ResNet
ResNet+2
Architecture
Original:5 Conv layers+3 FC layers
Adding 2 reduced FC layers
13 Conv layers + 3 FC layers
Adding 2 reduced FC layers
152 Conv layers
50 Conv layers
152 Conv layers +2 reduced FC layers
50 Conv layers + 2 reduced FC layers
Trained from scratch
48.1%
47.2%
51.6 %
55.2%
Pre-trained & Fine-tuned
58.2%
58.3%
60.1%
62.8%
63.7%
45.0%
60.2%
48.0%
8
Figure 2: Comparison of the filters learned on object classification vs style classification. Left:
Typical filters of AlexNet trained on ImageNet for object classification. Center and Right: Filters
of AlexNet trained for style classification using 11x11 and 7x7 dim filters.
Increasing the depth of the network only added no more than 5% to the accuracy from AlexNet
with 5 convolutional layers to ResNet with 152 convolutional layers. In the case for learning from
scratch, increasing the depth did not improve the results where a ResNet with 50 layers performed
worse than an AlexNet with only 5 convolution layers. VGGNet with 13 convolutional layers
performed only 2-3% better than AlexNet. Increasing the depth of VGGNet did not improve the
results. This limited gain in performance with increase in depth, in conjunction with the difference
in the learned filters, suggests that a shallow network might be sufficient for style classification
along with better filter design.
Effect of adding layers with reduced dimensionality:
The experiments showed that adding extra fully connected layers while gradually reducing the
number of nodes in them forces the networks to achieve a “smoother” and more interpretable representation. Having a large number of nodes at the fully connected layers allows the representation
to project the data into a very high dimensional space where classification would be easy (specially
in our case with only 20 classes), without necessarily enforcing similar paintings across styles to
come closer in the representation based on their similarity. We quantified this phenomenon by examining the dimensionality of the subspace of the activation (using Principle Component Analysis
(PCA [14]) of the visualization dataset using two measures: 1) The number of components needed
to preserve 95% of the variance. 2) The variance retained with the first two PCA dimensions. We
also evaluated the accuracy of the expanded models to see if the added reduced layers resulted in
any loss of accuracy. In most of the cased the added layers enhanced the accuracy (see Table 2)
Table 3 shows that adding two reduced-dimension layers effectively and consistently reduced
the dimensionality of the subspace of the data while preserving the classification accuracy. The
reduction is significant for AlexNet where 9 or 10 dimensions retained 95% of the variance compared to 201 and 397 dimensions for the cases of fine-tuned and learned from scratch networks
respectively, with around 60% of the variance retained in the first two dimension. Interestingly,
the representation achieved by VGGNets already has reduced dimension subspaces compared to
9
Table 3: The effect of adding two reduced dimension layers on the representation (subspace dimensionality and variance)
Model
Training Strategy
number of nodes
4096
AlexNet
Original Architecture
subspace dim 1 retained variance 2
201
21.71
Adding two dimensionality reduction layers
number of nodes subspace dim1 retained variance2
512
9
59.64
Pre-trained
& Finetuned
AlexNet From Scratch
4096
397
35.62
VGGNet Pre-trained
4096
55
49.52
& Finetuned
VGGNet From Scratch
4096
36
51.16
ResNet
Pre-trained
20483
491
17.53
& Finetuned
1 Subspace dim: Number of principle components cumulatively retaining 95% of variance.
2 Retained variance: Percentage of variance retained by the first two principle components.
3 ResNet does not have FC layers. This is the number of the nodes in the last pooling layer.
100
512
512
10
7
62.78
66.87
512
512
7
6
72.52
73.71
100
90
90
AlexNet+2
Finetuned
80
80
AexNet+2
From
Scratch
70
50
AlexNet
Finetuned
60
AexNet
From
Scratch
50
VGG+2
Finetuned
60
70
VGG
Finetuned
40
VGG
From
scratch
30
VGG+2
From
scratch
ResNet
Finetuned
20
40
ResNet+2
Finetuned
10
30
0
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
9
10
Figure 3: Cumulative retained variance by the first 10 PCA dimensions of the activation space at
last FC layer (512 nodes). Left: models after adding the dimensionality reductions layers. Right:
the original models.
the AlexNet and ResNet. However, adding the reduced dimension layers for VGG significantly
lowered the subspace dimension (only 7 dimensions retain 95% of the variance) while improving
its classification accuracy between 2-4%. The maximum reduction in subspace dimensionality was
in ResNet where the dimension of the subspace retaining 95% of the variance was reduced from
491 to only 6 with 74% of the variance in the first two dimensions. Figure 3 shows the cumulative
retained variance for the first 10 PCA dimensions of the activation subspace for all the models
before and after adding the two reduced dimensions layers.
4
Shape of Art History: Interpretation of the Representation
This section focuses on analyzing, visualizing, and interpreting the activation space induced by the
different networks after trained to classify style.
We define the activation space of a given fully connected layer as the output of that layer prior
to the rectified linear functions. In particular, in this paper, we show the analysis of activation of the
last reduced dimension fully connected layer prior to the final classification layer, which consists
of 512 nodes. We use the activations before the rectified linear functions in all the networks.
10
Few Factors Explain the Characteristics of Styles:
The learned representation by the machine shows that there are a few underlying factors that can
explain the characteristics of different styles in art history. Using Principle Component Analysis,
we find that only fewer than 10 modes of variations can explain over 95% of the variance in the
visualization set in all of the studied models with additional reduced fully connected layers. In
ResNet and VGGNet the number of these modes is as low as 6 and 7 respectively. In all of the
networks, the first two modes of variations explained from 60% to 74% from the variance in the
various models in visualization set I (Table 3). Moreover, it is clear from the visualizations of both
the linear and nonlinear embedding of the activation manifold in various models that art dated prior
to 1900 lie on a plane (subspace of dimension 2).
Consistent results are achieved by analyzing the 62K painting from the Wikiart data set where
it was found that subspaces of dimensions 10, 9 and 7 retain 95% of the variance of the activation
for AlexNet+2, VGGNet+2, ResNet+2 respectively (Table 5). The consistency of results in all
the studied networks and the two datasets (varying in size from ≈ 1500 to 62K paintings) imply
that the existence of a small number of the underlying factors explaining the representation is an
intrinsic property of art history, and not just an artifact of the particular dataset or model.
We will start our interpretation of these dimensions by investigating the time correlation with
these dimensions, followed by correlation analysis with Wölfflin’s concepts.
Table 4: Temporal correlation with the first two PCA dimensions and the first LLE dimensions of
the activation space in different models
model
AlexNet+2
VGGNet+2
ResNet+2
training
Fine-tuned
From scratch
Fine-tuned
From scratch
Fine-tuned
Pearson correlation coefficient with time
1st PCA dim 2nd PCA dim Radial 1st LLE dim
0.4554
0.5545 0.6944
0.7101
-0.5797
0.2829 0.6697
0.6723
-0.2462
0.5316 0.5659
-0.4012
0.5781
0.3165 0.6239
-0.6532
-0.6559
0.4579 0.7712
0.8130
Smooth Temporal Evolution of Style:
All the learned representations show smooth transition over time. Figure 4 shows the paintings
in the visualization set I, projected into the first two modes of variations of the activations of the
last reduced dimension fully connected layer for different networks. The paintings are color-coded
based on their date of creation. The plots consistently show that the networks learned smooth
temporal representations that reflect the historical progress in style. This is despite the fact that the
networks are trained only with images and their discrete style labels. No information was provided
about when each painting was created, when each style took place, which artist created which
painting, nor how styles are related (such as style x is similar to style y, or came after or before
style z). Despite the lack of all this information, the learned representations are clearly temporally
smooth and reflect high level of correlation with time.
Visualizing these modes of variations reveals very interesting ways of how the networks arrange
the data in a temporal way. For example, for the case of AlexNet+2 model, as shown in Figure 1-A,
11
Figure 4: First two modes of variations of different learned representations and their temporal
correlations. Every point is a painting, color-coded by the year of creation. Dataset: VS-I.
the data is arranged in the plot in a radial clock-wise way around the center to make a complete
circle in this 2D projection starting with Renaissance and ending with Abstract Art. We can see
the progress following the plot in a clock-wise way from Italian and Northern Renaissance at
the bottom, to Baroque, to Neo-classicism, Romanticism, reaching to Impressionism at the top
12
followed by Post impressionism, Expressionism and Cubism. The cycle completes with Abstract
and Pop Art. The explanation of how the network arranged the representation this way will become
clearer next when we show the results of nonlinear activation manifold analysis.
We quantified the correlation with time in different ways. We measure the Pearson Correlation
Coefficients (PCC) with each of the principle components, as well as with each dimension of
the nonlinear manifold embedding. Table 4 shows the temporal correlation using PCC with the
first two PCA dimensions and the first LLE embedding dimension for different models. We also
computed the correlation of the radial coordinate – of each painting in the projection to the first
two principle components– with time. The temporal correlation obtained in the representation is
consistent among the different networks.
Taking a fine-tuned AlexNet+2 as example, the first two dimensions have very strong correlation with time, 0.46 and 0.55 Pearson correlation coefficients respectively. A linear combination
of these two dimensions with convex weights proportional to their Pearson correlation coefficients
results in a single dimension that has 0.71 correlations with time (Figure 1-B). This cycle around
the plot from Renaissance till abstraction and Pop art suggests that the angular coordinate around
the plot would reflect correlation with time. In fact the angular coordinates have a PCC of 0.69
with time. The strongest temporal correlation is in the case of ResNet with 0.76 PCC radial correlations. This conclusion is consistent among all the networks that we tested. This conclusion
shows that style changes smoothly over time and proves that noisy discrete style labels are enough
for the networks to recover temporal arrangement, due mainly to visual similarities as encoded
through the learned representation in a continuous activation space.
Strong temporal correlation also exists with the modes of variations in Visualization set-II.
However, in this case, it is not necessarily the first two modes of variation that are the ones with
the highest temporal correlation. This is due to the bias in this dataset towards 20th century work.
Table 5 shows the PCC with time for each of the dimensions for three fine tuned networks.
Table 5: Analysis of the modes of variations of the activations on 62K paintings from the Wikiart
collection in three models and their correlation with time (Pearson Correlation Coefficients) . The
top two time-correlated modes are highlighted in bold.
Subspace Dim 1
Retained Variance 2
Correlation with time
1
2
3
4
5
6
7
AlexNet+2
10
51.13
0.12 0.66 0.24 -0.43 -0.23 0.04 0.08
VGGNet+2
9
43
-0.15 -0.14 -0.46 -0.07 -0.59 0.28 0.02
ResNet+2
7
66.80
-0.03 -0.72 0.20 0.22 0.30 -0.01 0.06
1 Subspace Dim: Number of principle components retaining 95% of the variance
2 Retained Variance: Variance retained by first two dimensions
8
-0.05
0.07
9
-0.03
0.02
10
0.00
Interpretation of the Modes of Variations - Relation to Wölfflin’s Pairs:
Visualizing the modes of variations reveals very interesting ways of how the networks were able to
consistently capture evolution and characteristics of styles. Here we take the fine-tuned AlexNet+2
as example, however similar results can be noticed in other models. The first mode (the horizontal
axis in Figure 1) seems to correlate with figurative art, which was dominant till Impressionism,
vs. non-figure, distorted figures, and abstract art that dominates 20th century styles. Another
interpretation for this dimension is that it reflects Wölfflin’s concept of planar (to the right) vs.
13
recession (to the left). This axis correlates the most with Wölfflin’s concept of planar vs recession
with -0.50 Pearson correlation coefficient. To a lesser degree, this horizontal axis correlates with
closed vs. open (0.24 PCC) and linear vs. painterly (-0.23 PCC). This quantitative correlation can
be clearly noticed by looking at the sampled paintings shown in Figure 1, where we can see that
the horizontal axis is characterized by planar, open, and linear form to the right (as in Cubism,
and Abstract) vs. recession, closed, and painterly to the left (as appears in some of the Baroque
paintings at the extreme left)
Table 6: Correlation with Wölfflin’s concepts. Pearson Correlation Coefficient of the first two PC
dimensions and the first two LLE dimensions of the activation space and Wölfflin’s concepts. The
concepts with maximum correlation with each dimension are shown.
Model
AlexNet+2
VGGNet+2
ResNet+2
Training
Fine-tuned
From scratch
Fine-tuned
From scratch
Fine-tuned
Pearson correlation coefficient(absolute values) with Wölfflin’s concepts
1st PCA dim 2nd PCA dim 1st LLE dim
2nd LLE dim
Planar vs.
Linear vs.
Planar vs.
and Linear vs.
Recession
Painterly
Recession
Painterly
0.4946
0.3579
-0.501
0.3216
0.5129
0.3272
0.4930
-0.3111
0.3662
0.2638
0.4512
0.2646
0.4621
0.4000
0.4897
0.3174
0.5314
0.4795
0.5251
0.4158
The second mode of variations (the vertical axis) has 0.36 Pearson correlation coefficient with
the linear (towards the bottom) vs. painterly (towards the bottom) concept. We can clearly see
the smooth transition from linear form in Renaissance at the bottom towards more painterly form
in Baroque to the extreme case of painterly at Impressionism at the top. Then we can see the
transition back to linear form in abstract and Pop art styles. Projecting the data into these two
dominant modes of variations, which are aligned with plane vs. recession and linear vs. painterly,
gives an explanation to why this representation correlates with time in a radial fashion.
The correlation between the first two modes of variations and the concepts of planar vs. recession and linear vs. painterly is consistent among all the representations learned by all tested
networks, whether pre-trained or trained from scratch. In all cases the first mode of variations correlates the most with the concept of planar vs. recession while the second mode correlates the most
with linear vs. painterly. Table 6 shows the Pearson correlation coefficients with these concepts.
Full correlation results with all dimensions are shown in the appendix.
The fourth and fifth dimensions of AlexNet+2 representation spread away strongly the Renaissance vs Baroque styles and put other styles in perspective to them (Figure 5). The fifth dimension
in particular (and the fourth dimension to a lesser degree) correlates with relative vs. absolute
clarity, unity vs. multiplicity, open vs. closed, and painterly vs. linear form from top to bottom
(with PCC 0.36, 0.30, 0.28, 0.16 respectively). This is consistent with Wölfflin’s theory since
he suggested exactly the same contrast between these concepts to highlight the difference between
Renaissance and Baroque. In Figure 5-A , the Renaissance style appears at the top (absolute clarity,
multiplicity, closed, linear form) while the Baroque appears at the bottom (relative clarity, unity,
and open, painterly form). We can see in the figure that Impressionism and Cubism are at the top
half of the plot since they share many of these same concepts with Baroque. The fourth dimension
seems to separate Impressionism and Cubism to the right from abstraction and Pop art to the left.
14
A
Rousseau
Van Eyck
Durer
Van Eyck
Van Eyck Mantegna
Da Vinci
Raphael
Rousseau Warhol
Raphael
Klimt
Warhol
Da Vinci
Lichtenstein
Sisley
Goya
Braque
Vermeer
B
Delacroix
Rembrandt
Picasso
Monet
Renoir
Velázquez
0.3%
0.2%
0.1%
Linear%vs%Painterly%
Close%vs%Open%
0%
4%
5%
!0.1%
Planer%vs%Recession%
Mul>plicity%vs%Unity%
Absolute%vs%Rela>ve%Clarity%
!0.2%
!0.3%
!0.4%
Figure 5: Separation between Renaissance and Baroque: (A) The fourth and fifth modes of variations show separation between Renaissance and Baroque. (B) Correlation between these two
modes of variations and Wölfflin’s concepts, which confirms with Wölfflin’s hypothesis.
Looking at the 4th, 5th, and 6th dimensions of the embedded activation manifold, we can notice
two orthogonal planes characterizing art prior to 1900 (Figure 10). One plane spans Renaissance
vs. Baroque while an orthogonal plane spans Impressionism vs. Post-Impressionism.
15
A
Renoir
Rothko
Monet
Van Gogh
Impressionism
Johns
Mondrain
Cézanne
Raphael
Post -Impressionism
Baroque
Cubism
Rousseau
Rembrandt
Picasso
Northern Renaissance
Velazquez
Durer
Van Eyck
B
Impressionism
C
D
E
Rubens
Baroque
Pop Art
Naïve art/Primitivism
Rousseau
F
G
H
Figure 6: The representation could discover distinctive artists for each style. (A)The first three modes
of variations of the VGGNet activations. Distinctive artists representing each style are identified by the
representation and pulled away from the cloud at the center. (B-E) Factorization of the activation space
using Independent Component Analysis into 7 maximally independent axes, which show alignment with
styles (more details in Figure 7). (F-H) The top three modes of variations in the VGG network activation
of 62K works of art from the Wikiart collection (projected pairwise as dimensions 1-2, 2-3, 3-1 from left to
right).
16
Discovering Limitations of Wölfflin’s Concepts:
Interestingly, not all modes of variations explaining the data correlate with Wölfflin’s concepts. In
all learned representations, one of the first five modes of variation always has close-to-zero linear
correlation with all Wölfflin’s concepts. A notable example is the fifth dimension of the embedded
activation manifolds, which separates Impressionism from Post-Impressionism, and has almost
zero linear correlation with Wölfflin’s pairs (Figure 10). This implies that the separation between
these styles is not interpretable in terms of Wölfflin’s pairs.
Identification of Representative Artists:
Visualizing the different representations shows that certain artists were consistently picked by the
machine as the distinctive representatives of their styles, as they were the extreme points along
the dimensions aligned with each style. This is visible in the first three modes of variations of the
representation learned by the VGGNet, shown in Figure6-A, which retains 77% of the variance .
Besides the observable temporal progress, the representation separates certain styles and certain
artists within each style distinctively from the cloud at the center as non-orthogonal axes.
We can see the Northern Renaissance in the yellow ellipse with the majority of the paintings
sticking out being by Van Eyck and Dürer. The Baroque in the black ellipse is represented by
Rubens, Rembrandt, and Velzquez. The orange ellipse is Impressionism and at its base are Pissarro,
Caillebotte, and Manet as the least painterly of the type, ending with Monet and Renoir as most
painterly on the end of the spike. The two red circles are Post-Impressionism, and in particular
one is dominated by Van Gogh, and the other by Cézanne who forms the base for the spike of
Cubism in the light blue ellipse. This spike is dominated by Picasso, Braque, and Gris; and goes
out to the most abstract Cubist works. Most interestingly the representation separates Rousseau, as
marked in the green ellipse, which is mainly dominated by his work. Consistent results can be seen
when plotting the embedding of the activation of the whole 62K paintings in the Wikiart dataset
collection as can be seen in Figure 6 F-H.
Applying Independent Component Analysis (ICA) [15, 17] to the activation space results in
finding non-maximally independent non-Gaussian orthogonal axes3 . Figure 6 B-E shows the ICA
factorization of the VGG network activation space. Here, the same subspace that retains 95% of
the variance is factorized by ICA. This results in 7 maximally independent components shown in
Figure 6-E.
We also show each of these components against the time axis in Figure 7. Some components
are aligned with specific styles (A: Baroque, C: Cubism, D: Post-Impressionism, E: abstract art, F:
Renaissance) and other components contrast several styles (B and F). This factorization facilitates
quantitatively discovering the artists who represent each style, as their works appear as the extreme
points on the axes of each style. It also highlights which styles are opposite to each other along
each of these axes 4 .
3
Independent Component Analysis (ICA) transforms activation to a new embedding space in which transformed
data become statistically independent as much as possible and the ICA achieve this by finding non-orthogonal basis in
original activation space maximizing independence criteria.
4
Since the ICA algorithm (Fast ICA [17]) used in this experiment has a limitation to capture over-completeness,
which implies existence greater number of fat tails than dimension of the space, our experiment found only some of
axes are aligned with styles. However, still the results show that ICA has merits to find some of interesting art historical
facts and provide quantitatively measurement of them, such as what art works or which artists are most representative
17
A
Impressionism
B
C
Cubism
Baroque
D
Van Eyck
Post Impressionism
Durer
Rousseau
Lichtenstein
Warhol
E
F
Abstract art
H
G
Renaissance
0.4%
0.3%
Linear%vs%Painterly%
0.2%
Close%vs%Open%
0.1%
Planer%vs%Recession%
0%
1%
Raphael
El-Greco
2%
3%
4%
5%
6%
7%
Mul@plicity%vs%Unity%
!0.1%
Rothko,
Johns
!0.2%
Absolute%vs%Rela@ve%
Clarity%
!0.3%
Figure 7: Independent component factorization of the activation of VGG network representation
achieved by ICA. A-G: Each independent component vs time (1400-2000AD). H: Correlation
(PCC) with Wölfflin’s concepts. Some components are aligned with specific styles (A: Baroque,
C: Cubism, D: Post-Impressionism, E: abstract art, F: Renaissance) and other components contrast
several styles (B and G).
The fact that the representation highlights a certain representative artist or artists for each style –
among hundreds or thousands of paintings by many artists in each style – emphasizes quantitatively
the significance of these particular artists in defining the styles they belong to.
Interpretations of the Activation Manifold and Understanding Influences:
Using nonlinear dimensionality reduction of the activation allows us to achieve an embedding of
the activation manifold of the visualization data sets, which reveals interesting observations as
well. In particular we used Local Linear Embedding (LLE) [16]. LLE, as other similar techniques,
is based on building a graph of data points, where each point is connected to its nearest neighbors,
and using such graph to embed the data to a low dimensional space where it can be visualized. In
our case the data points are the activation of a given painting as it goes through the deep network,
and measured at the last fully connected layer, and this results in an embedding of the activation
within each of styles.
18
manifold. Controlling the number of neighbors of each painting in constructing the graph allows
for controlling connections between paintings. If this parameter is set to a small number, this
results in accentuating distinctive paintings and trends. If we set this parameter to a large number,
we allow more paintings to connect and we can capture the overall trend.
The first dimension of the LLE representation of all the models exhibits strong temporal correlation, as shown in Table 4. For example, the first LLE dimension in AlexNet+2 representation
has 0.71 PCC with time that of ResNet+2 has 0.81 PCC with time. Also, similar to the modes of
variations, the first two LLE dimensions shows strong correlation with Wölfflin’s concepts planar
vs. recession and linear vs. painterly respectively, see Table 6. Looking at the representation
learned by AlexNet+2, as an example, we can notice that on the surface of the embedded activation manifold, different styles are arranged in the learned representation in a way that reflects
their temporal relation as well as stylistic similarity (see Figure 8-A)5 . The temporal relation between styles is not just linear historical order, since elements of some styles can appear later in
other styles. For example we can see linear temporal progress from Renaissance, to Baroque, to
Romanticism, to Realism to Impressionism. However we can also see direct connection between
Renaissance to Post-Impressionism and 20th century styles. This can be explained since the use of
linear (vs. painterly) and planar (vs. recessional) concepts in these later styles echo these elements
from Renaissance style(Fig 8-B).
One interesting connection that is captured in this representation is the arrangement of Cézanne’s
work in a way that connects Impressionism to Cubism. We can clearly see in Figures 8-B and 9A that Cézanne’s work acting as a bridge between Impressionism at one side and Cubism and
Abstract at the other side. Art historians consider Cézanne to be a key figure in the style transition towards Cubism and the development of abstraction in the 20th century art. This bridge of
Cézanne’s painting in the learned representation is quite interesting because that is a quantifiable
connection in the data, not just a metaphorical term. We can see branching at Post-Impressionism
where Cézanne’s work clearly separates from the other Post impressionist and expressionist works
towards the top. This branch continues to evolve until it connects to early Cubist works by Picasso
and Braque, as well as abstract works by Kandinsky.
Figures 8-B and 9-B show another interesting connection between the Renaissance and modern art as captured by the learned representation. Despite the fact that the structure reflects smooth
temporal progression, it is interesting to see outlier to this progression. In particular there are some
High Renaissance, Northern Renaissance and Mannerist paintings that stick out of the Renaissance
cluster to the left and connect to art from late 19th and early 20th centuries. This is because frequent similarity between art works across time results in pulling influential works of art out of
order and placing them closer to the art they may have influenced. We can see in the figure that
the works that stick out of the Renaissance cluster at the left and connect to modernity are mainly
dominated by some paintings by El-Greco and some paintings by Dürer. Among the paintings by
El-Greco that significantly stick out are Laocoön, Saint Ildefonso, View of Toledo, and Pietà. We
can also see works by Raphael, Mantegna, and Michelangelo in this group as well.
We can accentuate this representation by the reducing the number of neighbors of each painting while constructing the manifold. This small neighborhood construction results in lining up
distinctive paintings in thin structures to accentuate trends. This results of this version shown in
5
The representation is achieved by setting the neighborhood size to 100 to capture the overall trend.
19
Figure 8: Interesting connections discovered in the activation manifold. (A) Example of activation
manifold of AlexNet+2. Paintings color coded by date on the surface of the activation manifold
showing smooth transition over time. (B) Transitions between art movements and important connections. (C) Accentuated version of the manifold highlighting five major styles: Renaissance,
Baroque, Impressionism, Cubism, and abstract.
20
Figure 9: Interesting connections. (A) Cézanne’s bridge: we can see branching at PostImpressionism where Cézanne’s work clearly separates from the other Post-Impressionist and Expressionist works towards the top. This branch continues to evolve till it connects to early Cubist
works by Picasso and Braque, as well as abstracts works by Kandinsky. All thumbnails without
labels in this plot are by Cézanne. (B) The connection between Renaissance and modern styles
is marked by the outliers in the temporal progress patterns by certain works by El-Greco, Dürer,
Raphael, Mantegna, and Michelangelo.
Figure 8-C6 . In this visualization, five distinct styles are clearly accentuated among all the twenty
styles that the machine learned: Renaissance, Baroque, Impressionism, Cubism, and Abstract. We
show the representative paintings at the tip of each of the five styles, which are in order Raphael’s
St. Catherine (1507), Rembrandt’s self-portrait (1640), Monet’s Water Lilies under clouds (1903),
Braque’s Violin and Candlestick (1910), and Malevich’s Red Square (1915). Both Cézanne’s
bridge and El-Greco/Dürer’s bridge are accentuated in this representation as well.
Interestingly, both Cézanne’s connection and El-Greco/Dürer connection appears consistently
in the various representations learned by different networks, however manifested in different forms.
Another interesting observation is that Cubism appears as very unique style that sticks out of as a
singularity in most of the representations.
5
Conclusions and Discussion
In this paper we presented the result of a comprehensive study of training various deep convolutional neural networks for the purpose of style classification. We mainly emphasized analyzing the
learned representation to get an understanding of how the machine achieves style classification.
The different representations learned by the machine using different networks shared striking
similarities in many aspects. It is clear that a small numbers of factors can explain most of the
6
The accentuated representation is achieved by setting the neighborhood size to 25 to reduce the connectivity and
accentuate the visualization.
21
A
B
Impressionism
Renaissance
Baroque
Post-Impressionism
C
D
Baroque
Baroque
Impressionism
Post-Impressionism
0.3%
E
0.2%
Linear%vs%Painterly%
0.1%
Close%vs%Open%
Planer%vs%Recession%
0%
4%
5%
6%
Mul?plicity%vs%Unity%
!0.1%
Absolute%vs%Rela?ve%
Clarity%
!0.2%
!0.3%
Figure 10: Two orthogonal planes characterize art prior to 1900 in the learned Representation.
Renaissance-Baroque plane seems orthogonal to the Impressionism-Post-Impressionism plane in
the embedding of the activation manifold of fine-tuned AlexNet+2 model is shown. (A,B) Two
different view of the 4th, 5th, 6th dimensions are shown. (C) the 4th and 6th dimensions span
the Renaissance-Baroque differences. (D) 5th and 6th dimensions. The 5th dimension spans the
Impressionism-Post-Impressionism differences. Interestingly, the 5th dimension has very small
correlation with Wölfflin’s concepts. (E) Correlation with Wölfflin’s concepts
22
variance in the studied datasets in all the trained models (between 7 and 10 factors explains 95%
of the variance), which indicates that a small number of factors encapsulate the characterization of
different styles in art history.
The results also indicate that the learned representation shows a smooth temporal transition between styles, which the machine discovered without any notion of time given at training. Most
interestingly, studying the modes of variations in the representation showed a radial temporal
progress, with quantifiable correlation with time, starting at Renaissance, to Baroque, progressing all the way to Impressionism, Post impression, Cubism and closing the loop with 20th century
styles such as Abstract and Pop Art coming back close to Renaissance. By studying the correlation
between the modes of variations and Wölfflin’s suggested pairs, it was found that, consistently, all
the learned models have the first mode of variation correlating the most with the concept of plane
vs. recession while the second mode of variation correlates the most with the concept of linear
vs. painterly. This correlation explains the radial temporal progress and the loop closure between
Renaissance and 20th century styles since they share linearity and planarity in their form.
We also studied the activation manifold of the different representation learned by the machine,
which also reveal smooth temporal progress captured by the representation as well as correlation
with Wölfflin’s concepts. Studying the activation manifolds with different neighborhood structure
allowed us to discover different interesting connections in the history of art in a quantifiable way,
such as the role of Cézanne’s work as a bridge between a Impressionism and Cubism-Abstract
art. Another interesting connection is the connection between Renaissance and modern styles such
as Expressionism, Abstract-Expressionism, through the works of El-Greco, Dürer, Raphael, and
others.
Visualizing the different representations shows that certain artists were consistently picked by
the machine as the distinctive representatives of the styles they belong to as they were the extreme
points along the dimensions aligned with each style. For example, such distinctively representative artists are Van Eyck and Dürer for Northern Renaissance, Raphael for Italian Renaissance,
Rembrandt and Rubin for Baroque, Monet for Impressionism, Cézanne and Van Gogh for Post
Impressionism, Rousseau for Nave-Primitivism, Picasso and Braque for Cubism, and Malevich
and Kandinsky for Abstract. While this is quite known for art historians, the machine discovered
and highlighted these artists, amongst many others in each style, as the distinctive representatives
of their styles without any prior knowledge and in a quantifiable way.
The networks are presented by raw colored images, and therefore, they have the ability to
learn whatever features suitable to discriminate between styles, which might include compositional features, contrast between light and dark, color composition, color contrast, detailed brush
strokes, subject matter related concepts. In particular, networks pre-trained on object categorization datasets might suggest potential bias towards choosing subject-matter-related features for classification. However, visualizing the learned representations reveals that the learned representations
rule out subject matter as a basis for discrimination. This is clear from noticing the loop closure
between Renaissance style, which is dominated with religious subject matter, and modern 20th
century styles, such as Abstract, Pop art, and others. In contrast, this loop closure suggests that the
basis of discrimination is related to concepts related to the form as suggested by Wölfflin.
The implication of the networks ability to recover a smooth temporal progression through the
history of art, in absence of any temporal cues given at training, and in absence of any temporal constraints other than putting paintings of the same style closer to each other to achieve classification,
suggests that visual similarity is the main factor that forces this smooth temporal representation to
23
evolve, which in turns echoes the smooth transition of style in the history of art.
The results of this study highlight the potential role that data science and machine learning
can play in the domain of art history by approaching art history as a predictive science to discover
fundamental patterns and trends not necessarily apparent to the individual human eye. The study
also highlights the usefulness of re-visiting the formal methods in art history pioneered by art historians such as Wölfflin, in the age of data science using tools from computer vision and machine
learning. Finally, the study offers insights into the characteristics and functions of style for art
historians, confirming existing knowledge in an empirical way, and providing machine-produced
patterns and connections for further exploration.
References
[1] M. Schapiro, Style, Anthropology Today. 287-312 (1953).
[2] E. Gombrich, Style, International Encyclopedia of the Social Sciences 15, 352-361 (1968).
[3] L. Shamir et al, Impressionism, Expres-sionism, Surrealism: Automated Recognition of
Painters and Schools of Art, ACM Trans. Applied Perc., 7(2), (2010).
[4] S. Karayev, et al, Recognizing image style, arXiv:1311.3715 (2013).
[5] B. Saleh, K. Abe, R. S. Arora, A. Elgammal, Toward automated discovery of artistic influence, Multimedia Tools and Applications 75, 3565-3591 (2016).
[6] B. Saleh, A. Elgammal, Large-scale classification of fine-art paintings: learning the right
metric on the right features, Int. Journal of Digital Art 2, 71-93(2016).
[7] H. Wölfflin, Principles of Art History: The Problem of the Development of Style in Later Art,
(Trans MD Hottinger, New York: Dover Publications, Inc. (1950).
[8] L. Manovich, Data science and digital art history, Int. J. for Digital Art Hist. 1 (2015)
[9] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature, 521(7553), 436-444 (2015).
[10] A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with deep convolutional
neural networks, Adv. in Neur. Info. Process. Sys., 1097-1105 (2012).
[11] K. Simonyan, A. Zisserman. Very deep convolutional networks for large-scale image recognition, arXiv:1409.1556 (2014).
[12] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, Proc. of the
IEEE Conf. on Comp. Vis. and Patt. Recog. (2016).
[13] O. Russakovsky, et al, ImageNet Large Scale Visual Recognition Challenge, Int. Journal of
Computer Vision (2015).
[14] I. Jolliffe, Principle component analysis, John Wiley & Sons, Ltd, (2002).
24
[15] A. Hyvarinen, J. Karhunen, E. Oja, Independent Component Analysis, J. Wiley New York,
(2001).
[16] S. T. Roweis, L. K. Saul, Nonlinear dimensionality reduction by locally linear embedding,
Science 290.5500, 2323-2326 (2000).
[17] A. Hyvrinen, Fast and robust fixed-point algorithms for independent component analysis,
IEEE Trans. on Neur. Net. 10.3: 626-634 (1999).
25
| 1cs.CV
|
ON TORSION-FREE NILPOTENT LOOPS
arXiv:1610.06663v1 [math.GR] 21 Oct 2016
JACOB MOSTOVOY, JOSÉ M. PÉREZ-IZQUIERDO, AND IVAN P. SHESTAKOV
Abstract. We show that a torsion-free nilpotent loop (that is, a loop nilpotent with respect
to the dimension filtration) has a torsion-free nilpotent left multiplication group of, at most,
the same class. We also prove that a free loop is residually torsion-free nilpotent and that the
same holds for any free commutative loop. Although this last result is much stronger than
the usual residual nilpotency of the free loop proved by Higman, it is proved, essentially, by
the same method.
1. Introduction
It has been argued [7] that the usual definition of nilpotency in loops, given by Bruck [2]
is too weak for the most important properties of nilpotent groups to hold. Indeed, contrary
to what happens in the associative case, the successive quotients of Bruck’s lower central
series of a general finitely generated loop are not finitely generated; moreover, they do not
carry any algebraic structure similar to that of a Lie ring.
A different version of the lower central series, designed to eliminate these drawbacks, was
introduced in [7] under the name of the commutator-associator filtration (in much greater
generality, this definition was later independently stated by Hartl and Loiseau [4]). An
important feature of this filtration is its close relationship with the dimension filtration
which is defined via the powers of the augmentation ideal in the loop ring. By a theorem
of Jennings, If G is a group, the dimension filtration of G collapses after nth term if and
only if G is torsion-free nilpotent of class n; the same holds for loops if by nilpotency one
understands the nilpotency with respect to the commutator-associator filtration [8]. It is
therefore reasonable to call a loop torsion-free nilpotent of class n if its dimension filtration
collapses after the nth term.
In the present paper we prove two statements that show the usefulness of this notion.
Firstly, we show that a torsion-free nilpotent loop of class n has a torsion-free nilpotent left
multiplication group of class at most n. This fails to hold for torsion-free nilpotent loops in
the sense of Bruck, a counterexample being the free nilpotent (in the sense of Bruck) loop
of class two, whose left multiplication group is not nilpotent. We stress that the nilpotency
class of the left multiplication group may be strictly less than the class of the loop; examples
of such situation will be given.
Our second (and most important) result says that a free loop is residually torsion-free
nilpotent (and, in particular, residually nilpotent with respect to the commutator-associator
2010 Mathematics Subject Classification. 20N05.
Key words and phrases. Loops, nilpotent loops, torsion-free nilpotent loops.
The authors acknowledge the support by the Spanish Ministerio de Ciencia e Innovación (MTM201345588-C3-3-P); J. M. Pérez-Izquierdo also thanks support by PHBP14/00110; J. Mostovoy was also supported by the CONACYT grant 168093-F; I. P. Shestakov also acknowledges support by FAPESP, processo
2014/09310-5 and CNPq, processos 303916/2014-1 and 456698/2014-0.
1
filtration). This statement is much stronger than the assertion that free loops are residually
nilpotent in the sense of Bruck, proved by Higman in [5], although our proof is an application
of the same arguments (Higman was well aware that his methods were stronger than what
was necessary for the study of Bruck’s lower central series). A similar result is then proved
for free commutative loops.
Our main technical tool is the non-associative modification of the Magnus map from the
free group on k letters to the group of units in the ring of non-commutative power series
in k variables. We will use some very basic facts about loops and about non-associative
Hopf algebras as described in [11], although we shall give some definitions in order to fix the
notation.
2. Preliminaries
2.1. Torsion-free nilpotent loops. A loop L is a set with a unital, not necessarily associatve, product, such that both left and right multiplications
La : L → L,
x 7→ ax
Ra : L → L, x 7→ xa
by any a ∈ L are bijective. In a loop one defines the operations of left and the right division
by setting
a\x = L−1
a x
and
x/a = Ra−1 x
respectively. With each loop L one associates its left multiplication group LMlt(L): this is
the subgroup of the permutation group of L considered as a set, which is generated by the
La for all a ∈ L.
Extending the loop product in a loop L by linearity to the Q-vector space spanned by L,
one obtains the loop algebra QL. In the loop algebra, the augmentation ideal I is the kernel
of the homomorphism QL → Q which sends each element of L to 1. The nth power I n of
the augmentation ideal is the ideal spanned by the products of at least n elements of I. The
nth dimension subloop of L over Q is defined as
Dn L = {g ∈ L | g − 1 ∈ I n }.
We shall refer to these subloops simply as to dimension subloops, without any mention of the
field Q. For each n the subloop Dn L is normal (and fully invariant) in L and each quotient
Dn L/Dn+1 L is torsion-free [9]. The loop L is called torsion-free nilpotent of class n if Dn+1 L
is trivial while Dn L is not.
2.2. Free loops and the Magnus map. Let x = {x1 , . . . , xn }. The set W (x) of words
on x is defined recursively so as to consist of (a) all elements of x and e; (b) all expressions
of the form uv, u\v and u/v where u and v belong to W (x). The set W (x) is the free
algebra on x with one nullary operation e and three binary operations. The set Comp(w) of
components of a word w ∈ W (x) is defined by setting
Comp(e) = {e},
Comp(xi ) = {e, xi } when xi ∈ x
2
and
Comp(w) = Comp(u) ∪ Comp(v) ∪ {w} if w = uv, u\v or u/v.
A word in W (x) is reduced if none of its components is of the form
u\(uv), u(u\v), (uv)/v, (u/v)v, u/(v\u), (u/v)\u, ev, e\v, u/e, ue, u/u, or v\v.
The free loop F (x) is the quotient of the algebra W (x) by the relations
eu = u = ue, u\(uv) = v = u(u\v) and (uv)/v = u = (u/v)v.
Each element of F (x) can be represented by a unique reduced word so that W (x) ⊂ F (x);
this is a corollary of the Evans Normal Form Theorem [3].
Now, let X = {X1 , . . . , Xn }. Denote by Q{X} the free non-associative algebra on X
over Q and let Q{X} be its completion with respect to the degree; elements of Q{X} are
non-associative formal power series in the Xi . The power series with non-zero constant term
×
form a loop Q{X} under multiplication. Explicitly, the operation of the left division in
this loop is determined by
(1 + B)\(1 + A) = 1 + (A − B) − B(A − B) + B(B(A − B)) − . . .
for any power series A and B with no constant term. Notice that in the case when A
consists only of terms with right-normed parentheses and B = Xi is one of the generators,
(1 + B)\(1 + A) also contains terms with right-normed parentheses only.
The Magnus map is the homomorphism
×
M : F (x) → Q{X} ,
xi 7→ 1 + Xi .
Extending the Magnus map by linearity to any r ∈ QF (x) we see that M(r) has a zero
constant term if and only if r lies in the augmentation ideal of QF (x). This implies the
following:
Lemma 1 ([9]). An element w ∈ F (x) lies in Dn F (x) if and only if
M(w) = 1 + terms of degree n and higher.
In particular, in order to prove that F (x) is residually torsion-free nilpotent, it is sufficient
to show that the Magnus map is injective. This is totally analogous to the associative case
where the Magnus map is used to prove that the free group is residually torsion-free nilpotent
[6]. However, while in the associative case this is established by a straightforward argument
involving the syllable length of w, the non-associative version is somewhat more complex.
The Magnus map can be defined, in an entirely analogous fashion, for free commutative
loops. Here the algebra Q{X} should be replaced by the free commutative non-associative
algebra on X. We shall prove that the commutative Magnus map is also injective, and this
will imply that free commutative loops are residually torsion-free nilpotent.
3. The left multiplication group
Theorem 2. If L is a torsion-free nilpotent loop of class n, the group LMlt(L) is nilpotent
of class at most n.
3
We shall see that there are examples where the class of LMlt(L) is strictly smaller than
the class of L. On the other hand, when L is a group, we have LMlt(L) = L and, hence, the
class of LMlt(L) is the same as the class of L.
The key to the proof of Theorem 2 is the following:
Lemma 3. For all a ∈ L and F ∈ γn LMlt(L) we have F (a) ≡ a mod Dn L.
Proof. First, let us consider the case when L is the free loop generated by a, x1 , . . . , xn , and
F is the n-fold commutator [Lx1 , [. . . , [Lxn−1 , Lxn ]...]]. Then M(F (a)) contains right-normed
terms only. By dropping the parentheses we get the associative Magnus map of the element
[x1 , [. . . , [xn−1 , xn ]...]] · a
in the free group generated by a, x1 , . . . , xn ; this series is of the form
1 + a + terms of degree n and greater.
This implies that M(F (a)) is also of the same form. Indeed, the homomorphism of the free
non-associative algebra on a certain set of generators onto the free associative algebra on
the same generators which “forgets the parentheses” maps the space of the right-normed
monomials isomorphically onto the free associative algebra. This proves the lemma in this
case. The exact same argument works in the case when L is the free loop generated by
a, x1 , . . . , xm , with m ≥ n, and F is a product of arbitrary n-fold commutators in the Lxi .
Now, each element F (a) with F ∈ γn LMlt(L) in an arbitrary loop L is a homomorphic
image of Fe(a) where Fe is a product of arbitrary n-fold commutators in the Lxi in the free
loop generated by a, x1 , . . . , xm , for some m. Since the dimension subloops are respected by
loop homomorphisms, the lemma follows.
Theorem 2 is now an easy consequence of Lemma 3. Indeed, if L is torsion-free nilpotent
of class n, we have Dn+1 L = {1}. Then, by Lemma 3, for F ∈ γn+1 LMlt(L) and any a ∈ L
we have F (a) = a, which means that F is the identity in LMlt(L). Now, let us consider
several examples.
Proposition 4. Let L be the commutative loop whose elements are pairs of integers (p, q)
with
′
p p
′ ′
′
′
.
(p, q) · (p , q ) = p + p , q + q +
2
2
Then L is torsion-free nilpotent of class 4, with
D2 L = D3 L = D4 L = {(0, q) | q ∈ Z},
while LMlt(L) is nilpotent of class 3.
Proof. From what follows it should become clear that L is the reduction of the free commutative loop on one generator modulo D5 and the argument is based on the Magnus map.
Let L2 = {(0, q) | q ∈ Z} ⊂ L. It is clear that L2 is a normal subloop of L (in fact, it is
the centre of L) and the quotient L/L2 is abelian and torsion-free. It follows that Di L ⊂ L2
for i > 1.
It can be seen directly that (0, 1) ∈ D4 L. Indeed, if X = (1, 0) − 1 ∈ I, we have
(0, 1) − 1 = (4, 0)\((4, 1) − (4, 0)) = (4, 0)\(X 2X 2 − X 3 X),
4
which is readily seen to lie in I 4 ; see also Lemma 1 of [9]. (We can write X 3 here since, by
commutativity, X 2 X = XX 2 ). Since (0, 1) generates L2 , we see that D4 L = L2 .
Let us now show that D5 L is trivial. Let Qc {X} be the free commutative non-associative
algebra on one generator X and I k the ideal consisting of non-associative polynomials in X
without terms of degree < k. As a a vector space, Qc {X}/I 5 is spanned by the monomials
1, X, X 2 , X 3 , X 3 X and X 2 X 2 . Consider the map
L → Qc {X}/I 5
given by
p
p
p
3
2
− q X 3 X + qX 2 X 2
X +
X +
(p, q) 7→ 1 + pX +
4
3
2
mod I 5 .
A direct computation shows that it is an injective homomorphism of L into the loop of units
in Qc {X}/I 5 . Extend it to the loop algebra QL by linearity; the augmentation ideal I ⊂ QL
is then mapped to the ideal I/I 5 . In particular, if g ∈ D5 L, we see that g − 1 maps to zero,
which implies that g = 1.
It remains to see that LMlt(L) is nilpotent of class three. On one hand, notice that
L(p,q) = L(p,0) L(0,q) and that L(0,q) lies in the centre of LMlt(L) for all q. On the other hand,
we have
1
−1
[L(p,0) , L(r,0) ](a, b) = L−1
(p,0) L(r,0) L(p,0) L(r,0) (a, b) = (a, b + apr(p − r)).
2
Furthermore,
1
[L(s,0) , [L(p,0) , L(r,0) ]](a, b) = (a, b − prs(p − r))).
2
It follows that [L(s,0) , [L(p,0) , L(r,0) ]] lies in the center of LMlt(L) and, hence, LMlt(L) is
three-step nilpotent.
Proposition 5. Let L be the loop whose elements are pairs of integers (p, q) with
p ′
′ ′
′
′
p .
(p, q) · (p , q ) = p + p , q + q +
2
Then L is torsion-free nilpotent of class 3, with
D2 L = D3 L = {(0, q) | q ∈ Z},
while LMlt(L) is nilpotent of class 2.
The proof of this proposition is very similar to that of Proposition 4, only in this example
the loop in question is isomorphic to the reduction of the free non-commutative loop on one
generator modulo D4 ; we omit it.
Finally, we have the following fact which shows the radical difference between nilpotency
in the sense of Bruck for torsion-free loops and torsion-free nilpotency for loops:
Proposition 6. Let L = F/F3 be the free nilpotent (in the sense of Bruck) loop of class 2
on a countable number of generators. Then LMlt(L) is not nilpotent.
We shall prove this assertion in the next section. Note that, by [14], for any nilpotent, in
the sense of Bruck, loop L the left multiplication group LMlt(L) is solvable.
5
4. Higman’s construction
Here we state several results of [5] which will be of importance.
Let F = F {x} be the free loop on the set x = {x1 , . . . , xn }, α : F → L a surjective
homomorphism and N = ker α. The subloop [N, F ] is defined as smallest subloop of N
which is normal in F and such that N/[N, F ] is in the centre of F/[N, F ]. In particular, if
N = Fk , the kth term of Bruck’s lower central series, then, by definition, [N, F ] = Fk+1 .
Define the abelian group A to be freely generated by two types of symbols:
• xi for each xi ∈ x;
• hl1 , l2 i for each pair l1 , l2 ∈ L \ {e}.
On the set L × A consider the following product:
(l1 , a1 )(l2 , a2 ) = (l1 l2 , a1 + a2 + hl1 , l2 i),
where hl, ei = he, li = 0 for all l ∈ L. With this product, L × A becomes a loop, which we
denote by (L, A). In particular,
(l1 , a1 )/(l2 , a2 ) = (l1 /l2 , a1 − a2 − hl1 /l2 , l2 i),
(l2 , a2 )\(l1 , a1 ) = (l2 \l1 , a1 − a2 − hl2 , l2 \l1 i).
Consider the homomorphism
δ : F → (L, A)
w 7→ (α(w), ψ(w))
where ψ(xi ) = xi ∈ A.
Lemma 7 ([5], Lemma 3).
ker(δ) = [N, F ].
This lemma has the following corollary:
Lemma 8. γn−1 LMlt(F/F3 ) is non-trivial.
Proof. Let N = [F, F ] so that L = F/N is a free abelian group on x1 , . . . , xn (strictly
speaking, we should write α(xi ) instead of xi but the notation can take some abuse) and
[N, F ] = F3 .
By Lemma 7 the image of the map δ : F → (L, A) is isomorphic to F/F3 . Write
y = [Lxn , [Lxn−1 , . . . [Lx3 , Lx2 ]]](x1 ).
Then δ(y) = (x1 , a) with a 6= x1 . Indeed, it is easy to see (by induction on n, for instance)
that the term hx2 , x1 i appears in a with coefficient 1. This implies that the iterated commutator [Lxn , [Lxn−1 , . . . [Lx3 , Lx2 ]]] is not the identity and, hence, γn−1 LMlt(F/F3 ) 6= {1}.
Proposition 6 follows.
The main technical result of Higman’s paper [5] is the following:
Lemma 9 ([5], Lemma 6). Let w, w ′ ∈ W (x) be reduced words such that
• w ′ 6∈ Comp(w),
• α(w) = α(w ′),
6
• if u, v are words in Comp(w) ∪ Comp(w ′ ) such that α(u) = α(v) then either u = v
or {u, v} = {w, w ′}.
Then there exists a generator xi or hl1 , l2 i of A whose coefficient in ψ(w) is zero, and in
ψ(w ′ ) is ±1.
This statement has a consequence which will be of immediate importance for us:
Lemma 10 ([5], Corollary 1). Let S be a finite set of reduced words containing together with
any word all of its components. If the images of the elements of S in F/N under the natural
projection map are not all distinct, the number of elements in the image of S in F/[N, F ] is
strictly greater than the number of elements in the image of S in F/N.
5. Injectivity of the Magnus map
Let Fc (x) be the free commutative loop on the set x = {xi } and Qc {X} the free commutative non-associative algebra on the corresponding set X = {Xi }. The commutative version
of the Magnus map sends Fc (x) into the loop of invertible elements in the completion of
Qc {X} and is defined exactly in the same way as the the usual Magnus map: xi is sent to
1 + Xi .
Theorem 11. The Magnus maps
M : F (x) → Q{X}
×
and
Mc : Fc (x) → Qc {X}
×
are injective.
In particular, the free loops F (x) and Fc (x) are residually torsion-free nilpotent.
5.1. Group-like elements and the modified Magnus map. Recall that in a Hopf algebra a primitive element is an element x satisfying
∆(x) = x ⊗ 1 + 1 ⊗ x
and a group-like element is an element g such that
∆(g) = g ⊗ g
and ǫ(g) = 1. The set G(H) of all group-like elements in a non-associative Hopf algebra H
is, actually, a loop [13].
The free algebra Q{X} is a Hopf algebra whose coproduct is defined by the condition
that the generators are primitive. The only group-like element in Q{X} is 1; however, the
completion Q{X} has many group-like elements. A group-like element e(X) in the algebra
Q{X} of power series in one non-associative and non-commutative variable X with a nonzero coefficient at X is called a base for logarithms or a non-associative exponential series
(see [10, 12]). An example of a base for logarithms is the usual exponential series endowed
with the right-normed parentheses. Given e(X) there exists another series loge (1 + X) such
that loge (e(X)) = X and e(loge (1 + X)) = X.
7
It will be convenient to use a slightly different version of the Magnus map by choosing a
base e(X) for logarithms and setting
×
M′ : F (x) → G(Q{X}) ⊂ Q{X} ,
xi 7→ e(Xi ).
The map Xi 7→ e(Xi ) − 1 determines an algebra homomorphism Q{X} → Q{X} which
sends M(w) to M′ (w) for any w ∈ F = F (x). Therefore, if M′ is injective so is M.
Let N = ker M′ and L = F/N. We shall prove that N coincides with the kernel of
Higman’s homomorphism F → (L, A) and this, by Lemmas 7 and 10, will imply that N
is trivial. For this purpose, we shall embed (L, A) into a bigger loop, namely the loop of
group-like elements of a Hopf-algebraic version of Higman’s construction (L, A).
5.2. Higman’s construction for Hopf algebras. Let m(X) be the set of all non-associative
monomials on X of degree ≥ 1 and T the set of symbols
{t1 , . . . , tn } ⊔ {t(m1 , m2 ) | m1 , m2 ∈ m(X)}
with degrees
|t1 | = . . . = |tn | = 1 and |t(m1 , m2 )| = |m1 | + |m2 |,
where |X1 | = . . . |Xn | = 1. Write Q[T ] for the usual commutative and associative algebra
on T with the structure of a Hopf algebra obtained by declaring all elements of T to be
primitive.
Define a linear map
t : Q{X} ⊗ Q{X} → spanhT i
by setting t(m1 ⊗ m2 ) = t(m1 , m2 ), t(m1 ⊗ 1) = t(1 ⊗ m2 ) = t(1 ⊗ 1) = 0 and let
t∗ : Q{X} ⊗ Q{X} → Q[T ]
be the coalgebra morphism
t∗ (µ ⊗ µ′ ) = ǫ(µ)ǫ(µ′ )1 +
X 1
t(µ(1) , µ′(1) ) . . . t(µ(k) , µ′(k) ).
k!
k≥1
We shall use the same notation t and t∗ for the extensions of these maps to maps between
the respective completions of Q{X} ⊗ Q{X}, spanhT i and Q[T ] with respect to the degree.
The maps t and t∗ are related in the following way: for any g, g ′ ∈ G(Q[X]) we have
∞
X
1
t(g ⊗ g ′ )k = exp(t(g ⊗ g ′)) ∈ Q[T ],
t (g ⊗ g ) =
k!
k≥0
∗
′
where exp is the usual exponential series.
Lemma 12. The elements exp(t1 ), . . . exp(tn ) and t∗ (g ⊗ g ′), where g, g ′ ∈ G(Q[X])\{1},
freely generate a multiplicative abelian subgroup of G(Q[T ]).
Proof. Assume that
′ dm
exp(t1 )e1 · · · exp(tn )en t∗ (g1 ⊗ g1′ )d1 · · · t∗ (gm ⊗ gm
) =1
8
′
for some g1 , g1′ , . . . , gm , gm
∈ G(Q{X}) \ {1} with (gi , gi′ ) 6= (gj , gj′ ) if i 6= j, and some integers
e1 , . . . , en , d1 , . . . , dm . Then
′
exp(e1 t1 + · · · + en tn + d1 t(g1 ⊗ g1′ ) + · · · + dm t(gm ⊗ gm
)) = 1
and
′
e1 t1 + · · · en tn + d1 t(g1 ⊗ g1′ ) + · · · + dm t(gm ⊗ gm
) = 0.
Since {t1 , . . . , tn } and {t(m1 ⊗m2 ) | m1 , m2 ∈ m(X)} are algebraically independent, we have
e1 = · · · = en = 0 and
′
′
t(d1 g1 ⊗ g1′ + · · · + dm gm ⊗ gm
) = d1 t(g1 ⊗ g1′ ) + · · · + dm t(gm ⊗ gm
) = 0.
The kernel of t : Q{X} ⊗ Q{X} → spanhT i is
Q{X} ⊗ 1 + 1 ⊗ Q{X}
so that
′
d1 · g1 ⊗ g1′ + · · · + dm · gm ⊗ gm
= µ⊗1+1⊗η
′
∈ G(Q{X})\{1} and since different group-like
for some µ, η ∈ Q{X}. As g1 , g1′ , . . . , gm , gm
elements are linearly independent [1, Theorem 2.1.2], we see that d1 = · · · = dm = 0.
Now, define a new product on Q{X} ⊗ Q[T ] by
X
(5.1)
(x ⊗ α)(y ⊗ β) =
x(1) y(1) ⊗ t∗ (x(2) ⊗ y(2) )αβ.
With this product, both Q{X}⊗Q[T ] and Q{X} ⊗ Q[T ] are connected bialgebras. Consider
the homomorphism
f′ : F (x) → G(Q{X} ⊗ Q[T ])
M
xi 7→ e(Xi ) ⊗ exp(ti ),
where e is the same base for logarithms used for defining M′ . The formula
Xi 7→ loge (e(Xi ) ⊗ exp(ti ))
defines an injective algebra homomorphism
φ : Q{X} → Q{X} ⊗ Q[T ]
f′ . In particular, for any words w1 , w2 ∈ F (x) the equality
and we have φ ◦ M′ = M
M′ (w1 ) = M′ (w2 )
f′ (w1 ) = M
f′ (w2 ).
holds if an only if M
Now, by definition, L is a subloop of G(Q{X}) and Lemma 12 gives an natural injective
homomorphism of (L, A) into G(Q{X} ⊗ Q[T ]). Moreover, on L the map φ coincides with
the map L → (L, A) and, therefore, as mentioned before, Lemmas 7 and 10 imply that M′
is injective.
9
5.3. The commutative case. The proof of the injectivity of Mc involves minor changes:
• denote by mc (X) the set of all monomials of degree ≥ 1 in the commuting variables
X1 , . . . , Xn ;
• define the set T as {t1 , . . . , tn } ⊔ {t(m1 , m2 ) | m1 , m2 ∈ mc (X)} where we assume
that t(m1 , m2 ) = t(m2 , m1 ) for all m1 , m2 ∈ mc (X);
• define the coalgebra morphism
t∗ : Qc {X} ⊗ Qc {X} → Q[T ]
as before. The symmetry of t implies t∗ (µ ⊗ µ′ ) = t∗ (µ′ ⊗ µ) for all µ, µ′ ∈ Qc {X};
• write Qc {X} ⊗ Q[T ] for the commutative bialgebra with the product given by (5.1);
• note that Lemma 12 trivially holds when Q{X} is replaced by Qc {X};
f′ : Fc {x} → G(Qc {X} ⊗ Q[T ]) be the homo• let M′c : Fc {x} → G(Qc {X}) and M
c
′
f′ (xi ) = e(Xi ) ⊗ exp(ti ) for some
morphisms determined by Mc (xi ) = e(Xi ) and M
c
base for logarithms e(X);
• fix a total order on W (x) and call a word in W (x) reduced if none of its components
is the left-hand side of any of the equations
uv = vu (u < v)
v\(uv) = u
ev = v
ue = u
u/v = v\u
u\(uv) = v
u(u\v) = v
(v\u)\u = v
(u\v)u = v
e\v = v
v\v = e;
• in the definition of the loop (L, A), take the symbols hl1 , l2 i to be symmetric in the
sense that hl1 , l2 i = hl2 , l1 i.
With these new definitions Lemma 9 (that is, Higman’s Lemma 6) remains true, and the
injectivity of Mc follows from it in the very same fashion as in the non-commutative case.
For the sake of completeness, we include the proof of the commutative version of Lemma 9
here, although it follows Higman’s original proof very closely.
5.4. Proof of the commutative version of Lemma 9. First, we observe that w ′ 6= e
since w ′ 6∈ Comp(w).
The case when w ′ = xi is straightforward: xi = w ′ 6∈ Comp(w) implies that the coefficient
of xi in ψ(w ′) is 1 while in ψ(w) it is 0. Therefore, we only have to consider two cases: (1)
w ′ = uv and (2) w ′ = u\v, where both u and v are reduced words. As w ′ is reduced, we
have u 6= e. In fact, α(u) 6= 1, the unit element of L, since α(u) = 1 = α(e) implies, by
hypothesis, that either u = e, a contradiction, or {u, e} = {w, w ′}, which is not possible
since w ′ 6= u, e.
In what follows, we will often consider two elements, say c1 and c2 , in Comp(w ′ )∪Comp(w)
with α(c1 ) = α(c2 ). By hypothesis, this implies that either c1 = c2 or {w ′, w} = {c1 , c2 }. It
will be clear that w ′ 6= c1 , c2 (usually c1 and c2 will be either e, components of w or components of w ′ different from w ′ ) so we will conclude that c1 = c2 without further explanation.
10
5.4.1. Case w ′ = uv. First we observe that α(v) 6= 1. Indeed, α(v) = 1 = α(e) would imply
that v = e so that w ′ = ue is not reduced, a contradiction. This ensures that hα(u), α(v)i
is a generator of A. We will prove that the coefficient of hα(u), α(v)i in ψ(τ ) is zero for any
τ ∈ {w, u, v}, which will allow us to conclude that the coefficient of hα(u), α(v)i in ψ(w) is
zero while in ψ(w ′ ) it equals to one. To this end, assume that the coefficient of hα(u), α(v)i
in ψ(τ ) is non-zero and observe that w ′ 6∈ Comp(τ ). There are two possibilities:
(1) τ has a component τ1 τ2 with {α(τ1 ), α(τ2 )} = {α(u), α(v)}.
• If α(τ1 ) = α(u) and α(τ2 ) = α(v) then τ1 = u and τ2 = v. Thus τ = uv = w ′ , a
contradiction.
• If α(τ1 ) = α(v) and α(τ2 ) = u then τ1 = v, τ2 = u and τ = vu. Both τ and w ′
are reduced so u ≥ v and v ≥ u. Hence u = v and τ = uu = w ′ , a contradiction.
(2) τ has a component τ1 \τ2 with {α(τ1 ), α(τ1 \τ2 )} = {α(u), α(v)}.
• If α(τ1 ) = α(u) and α(τ1 \τ2 ) = α(v) then τ1 = u and
– either τ1 \τ2 = v, which implies that w ′ = uv = τ1 (τ1 \τ2 ) is not reduced, a
contradiction,
– or w ′ = τ1 \τ2 and w = v, which is again a contradiction since w ′ 6∈
Comp(τ ).
• If α(τ1 ) = α(v) and α(τ1 \τ2 ) = α(u) then τ1 = v and τ1 \τ2 = u. Thus w ′ =
(τ1 \τ2 )τ1 is not reduced, a contradiction.
5.4.2. Case w ′ = u\v. We first observe that α(u\v) 6= 1. Indeed, α(u\v) = 1 would imply
α(u) = α(v) so that u = v and w ′ = u\u is not reduced, a contradiction. This ensures
that hα(u), α(u\v)i is a generator of A. We will prove that the coefficient of hα(u), α(u\v)i
in ψ(τ ) is zero for any τ ∈ {w, u, v}, which will allow us to conclude that the coefficient
of hα(u), α(u\v)i in ψ(w) is zero while it is −1 in ψ(w ′). To this end, assume that the
coefficient of hα(u), α(u\v)i in ψ(τ ) is non-zero and observe that w ′ 6∈ Comp(τ ). There are
two possibilities:
(1) τ has a component τ1 τ2 with {α(τ1 ), α(τ2 )} = {α(u), α(u\v)}.
• If α(τ1 ) = α(u) and α(τ2 ) = α(u\v) then τ1 = u and α(v) = α(u)α(w ′) =
α(u)α(τ2 ) = α(τ1 τ2 ). Thus
– either v = τ1 τ2 , which implies w ′ = τ1 \(τ1 τ2 ), a contradiction,
– or w ′ = τ1 τ2 and w = v, which is not possible since w ′ 6∈ Comp(τ ).
• If α(τ1 ) = α(u\v) and α(τ2 ) = α(u) then τ2 = u and α(τ1 τ2 ) = α(τ1 )α(τ2 ) =
α(τ2 )α(τ1 ) = α(u)α(u\v) = α(v). Thus, τ1 τ2 = v and w ′ = u\v = τ2 \(τ1 τ2 ) is
not reduced, a contradiction.
(2) τ has a component τ1 \τ2 with {α(τ1 ), α(τ1 \τ2 )} = {α(u), α(u\v)}.
• If α(τ1 ) = α(u) and α(τ1 \τ2 ) = α(u\v) then τ1 = u and α(v) = α(u)α(u\v) =
α(τ1 )α(τ1 \τ2 ) = α(τ2 ) so τ2 = v. Hence w ′ = u\v = τ1 \τ2 ∈ Comp(τ ), a
contradiction.
• If α(τ1 ) = α(u\v) and α(τ1 \τ2 ) = α(u) then u = τ1 \τ2 and α(v) = α(u)α(u\v) =
α(u\v)α(u) = α(τ1 )α(τ1 \τ2 ) = α(τ2 ). Thus, τ2 = v and w ′ = u\v = (τ1 \τ2 )\τ2
is not reduced, a contradiction.
References
[1] E. Abe, Hopf algebras. Cambridge University Press, Cambridge-New York (1980).
[2] R.H. Bruck, A survey of binary systems. Springer Verlag, Berlin-Göttingen-Heidelberg (1958).
11
[3] T. Evans, On multiplicative systems defined by generators and relations. I. Normal form theorems,
Proc. Cambridge Phil. Soc. 47, (1951), 637–49.
[4] M. Hartl and B. Loiseau, On actions and strict actions in homological categories, Theory and Applications of Categories, 27, no. 15, (2013), 347–392.
[5] G. Higman, The lower central series of a free loop, Q. J. Math., 14 , no. 1, (1963), 131–140.
[6] W. Magnus, A. Karrass, D. Solitar, Donald, Combinatorial group theory: Presentations of groups
in terms of generators and relations. Interscience Publishers [John Wiley & Sons, Inc.], New YorkLondon-Sydney (1966) Heldermann Verlag, Berlin, 1990.
[7] J. Mostovoy, The Notion of Lower Central Series for Loops. In Sabinin, L., Sbitneva, L., Shestakov, I.
(eds.) Non-associative algebra and its applications. pp. 291–298. Chapman & Hall/CRC-Press, Boca
Raton, FL (2006)
[8] J. Mostovoy, Nilpotency and Dimension Filtration for Loops. Commun. Algebra 36, (2008),1565–1579.
[9] J. Mostovoy, J.M. Pérez Izquierdo, Dimension Filtration on Loops. Isr. J. Math. 158, (2007), 105–118.
[10] J. Mostovoy, J.M. Pérez Izquierdo, I. Shestakov, Nilpotent Sabinin algebras, J. Algebra, 419, (2014),
95–123.
[11] J. Mostovoy, J.M. Pérez Izquierdo, I. Shestakov, Hopf algebras in non-associative Lie theory, Bull.
Math. Sci. 4, no. 1, (2014),129-174.
[12] J. Mostovoy, J.M. Pérez Izquierdo, I. Shestakov, A Non-associative Baker-Campbell-Hausdorff formula,
arXiv:1605.00953 [math.RA].
[13] J.M. Pérez Izquierdo, Algebras, hyperalgebras, nonassociative bialgebras and loops, Adv. Math. 208,
834–876 (2007)
[14] C.R.B. Wright, On the multiplication group of a loop, Illinois J. Math. 13, no. 4, (1969), 660–673.
Departamento de Matemáticas, CINVESTAV-IPN, Col. San Pedro Zacatenco, Ciudad de
México, C.P. 07360, Mexico
E-mail address: [email protected]
Departamento de Matemáticas y Computación, Universidad de La Rioja, Edificio CCT C/ Madre de Dios, 53, 26006, Logroño, La Rioja, Spain
E-mail address: [email protected]
Instituto de Matematica e Estatstica, Universidade de São Paulo, Caixa Postal 66281,
São Paulo, SP 05311-970, Brazil
E-mail address: [email protected]
12
| 4math.GR
|
Graphettes: Constant-time determination of graphlet and
orbit identity including (possibly disconnected) graphlets up
to size 8
arXiv:1708.04341v1 [cs.DS] 14 Aug 2017
Adib Hasan1 , Po-Chien Chung2 , Wayne Hayes2* ,
1 Ananda Mohan College, Mymensingh, Bangladesh
2 Dept. of Computer Science, University of California, Irvine, California, USA
* Corresponding author: [email protected]
Abstract
Graphlets are small connected induced subgraphs of a larger graph G. Graphlets are
now commonly used to quantify local and global topology of networks in the field.
Methods exist to exhaustively enumerate all graphlets (and their orbits) in large
networks as efficiently as possible using orbit counting equations. However, the number
of graphlets in G is exponential in both the number of nodes and edges in G.
Enumerating them all is already unacceptably expensive on existing large networks, and
the problem will only get worse as networks continue to grow in size and density. Here
we introduce an efficient method designed to aid statistical sampling of graphlets up to
size k = 8 from a large network. We define graphettes as the generalization of graphlets
allowing for disconnected graphlets. Given a particular (undirected) graphette g, we
introduce the idea of the canonical graphette K(g) as a representative member of the
isomorphism group Iso(g) of g. We compute the mapping K, in the form of a lookup
table, from all 2k(k−1)/2 undirected graphettes g of size k ≤ 8 to their canonical
representatives K(g), as well as the permutation that transforms g to K(g). We also
compute all automorphism orbits for each canonical graphette. Thus, given any k ≤ 8
nodes in a graph G, we can in constant time infer which graphette it is, as well as which
orbit each of the k nodes belongs to. Sampling a large number N of such k-sets of nodes
provides an approximation of both the distribution of graphlets and orbits across G,
and the orbit degree vector at each node.
Author summary
Graphlets are small subgraphs of a larger network. They have been used extensively for
over a decade in the analysis of social, biological, and other networks. Unfortunately it
is extremely expensive to exhaustively enumerate all graphlets appearing in a large
graph, requiring days or weeks of computer time for recent large networks. Here we
introduce a novel method for statistically sampling graphlets from large graphs. The
time required does not depend upon the size of the input network, but instead upon the
number of samples desired. In addition, existing methods only look at graphlets up to
size 5 or 6; we allow graphlets up to size 8, which significantly improves on the
sensitivity and specificity of network analysis. Our method will allow graphlets to be
efficiently utilized to analyze networks of arbitrary size going into the future.
PLOS
1/13
Fig 1. All (connected) graphlets of sizes k = 3, 4, 5 nodes, and their automorphism
orbits; within each graphlet, nodes of equal shading are in the same orbit. The
numbering of these graphlets and orbits were created by hand [8] and do not correspond
to the automatically generated numbering used in this paper. The figure is taken
verbatim from [16].
Introduction
Network comparison is a growing area of research. In general the problem of complete
comparison of large networks is intractable, being an N P -complete problem [1]. Thus,
approximate heuristics are needed. Networks have been compared for statistical
similarity from a high-level using simple, easy-to-calculate measures such as the degree
distribution, clustering co-efficients, network centrality, among many others [2, 3]. While
more sophisticated methods such as spectral analysis [4, 5] and topological indices [6]
have been useful, the study of small subnetworks such as motifs [7] and graphlets [8, 9]
have become popular. They have been used extensively to globally classify highly
disparate types of networks [10] as well as to aid in local measures used to align
networks [11–14].
A graphlet is a small, connected, induced subgraph g of a larger graph G. Given a
particular graphlet g, the automorphism orbits of g are the sets of nodes that are
topologically identical to each other inside g. Graphlets and their automorphism orbits
with up to k = 5 nodes were first introduced in 2004 [8], and are depicted in Fig 1.
Recently, automated methods have been created that can enumerate, in a larger graph,
all graphlets and their automorphism orbits up to graphlet size k = 5 [15] and
subsequently to any k [16], although the latter authors only applied it up to k = 6.
Unfortunately, we have found that these methods take a very long time (hours to days)
even just to count graphlets up to size k = 5 on some large biological networks, such as
those in BioGRID [17]. It is not clear that such methods, especially for even larger k,
PLOS
2/13
will be applicable to the coming age of ever bigger networks, since the total number of
graphlets appearing in a large network tends to increase exponentially with both k (the
graphlet size) and n (the number of nodes in the large network). Eventually, an
exhaustive enumeration of all graphlets appearing in a large network may become
infeasible simply due to the number of graphlets that need to be enumerated, even
under the optimization of using orbit counting equations. On the other hand, graphlets
are too useful to abandon as a method of quantifying the topological structure of graphs.
An achievable alternative for a large network G is to statistically sample its graphlets
rather than exhaustively enumerate them. Additionally, such sampling could be useful
with the recent advent of comprehensive biological network databases [18]: each sampled
graphlet would act as a seed for local matching between larger networks, similar to how
k-mers (short sequences of length k) are used for seed-and-extend sequence matching in
BLAST [19].
To efficiently create a statistical sample of graphlets in a large network G, one must
be able to take an arbitrary set of k nodes from G, and efficiently (preferably in
constant time) determine both which graphlet is represented, as well as the
automorphism orbits of each of the k nodes. Here, we solve this problem both by
enumerating all graphlets (and their disconnected counterparts, which we term
graphettes) and their automorphism orbits up to graphettes of size k = 8. We present a
method that creates a lookup table that can quickly determine the graphette identity of
any k nodes, as well as their automorphism orbits. Since the lookup table required
significant time to pre-compute for k = 7 (a few hours on a single core) and k = 8
(hundreds of CPU weeks on a cluster), we provide the actual lookup tables for these
values of k online at http://github.com/Neehan/Faye.
Materials and methods
Definitions and notations
Given a graph G on n nodes, a k-graphette is a (not necessarily connected) induced
subgraph g on any set of k nodes of G. There are many ways one could choose the k
nodes, for example (i) choosing k nodes uniformly at random from G, or (ii) performing
a local search around some node u. We expect the former to be useful only in dense
networks, while the latter is probably more useful in sparse networks because most
random sets of k nodes in a sparse graph will be highly disconnected and thus not very
informative. One could also (iii) perform edge-based selection (with local expansion) to
ensure dense regions are sampled more frequently than sparse regions [20]; still other
methods have been suggested [21].
Given a set of k nodes, we wish to quickly ascertain which graphette is represented,
and which automorphism orbits each of the k nodes belong to. To do that we need a
canonical list of graphettes and their orbits, and a fast way to determine which
canonical graphette is represented by any permutation of k nodes. Here we demonstrate
how, if k is fixed and relatively small (k ≤ 8 in our case), this can be accomplished in
constant time by pre-computing and storing a lookup table indexed by a bit vector
representation of the lower triangular matrix of the (undirected) adjacency matrix of
the induced subgraph. Given such an index, the value associated with that index
identifies the canonical graphette (a canonical ordering of the nodes for that graphette).
We also pre-compute the automorphism orbits of all the canonical graphettes. Thus, by
reversing the lookup table we can, in constant time, infer the orbit identity of each of
the k nodes in that k-graphette. As a corrollary, we can also update the (statistically
sampled) graphette orbit degree vector of each of the k nodes, similar to the graphlet
degree vector [9].
PLOS
3/13
We use the following abbreviations and notations throughout:
G(V, E) The Graph with nodes V and edges E
V(G) The set of nodes of graph G
E(G, u, v) The boolean value denoting connectivity between nodes u and
v of graph G
⇐⇒, iff If and only if
|S| The number of elements in set S.
Adj(G) The adjacency matrix representation of graph G
Aut(G) The set of automorphisms of graph G
K(g) Canonical isomorph of graphette g
1
2
6
2
7
1
4
7
6
0
5
3
9
2
3
1
7
0
0
9
8
9
3
5
4
8
4
5
6
8
Fig 2. Three isomorphic representations of the Petersen graph.
Canonization of graphettes
If graphs G and H are isomorphic, it essentially means they are exactly the same graph,
but drawn differently. For example, Fig 2 shows three different drawings of the Petersen
graph. Technically, an isomorphism between networks G and H is a permutation
π : V(G) → V(H) so that
E(G, u, v) ⇐⇒ E(H, π(u), π(v)),
w
x
w
y
x
w
x
y
w
y
x
y
Fig 3. All the possible 3-graphettes.
Consider a 3-graphette with nodes w, x and y. There are only 4 possible such
graphettes, depicted in Fig 3. However, by permuting the order of the nodes, each of
these graphettes can be represented by several isomorphic variants. In order to
1
determine if two graphettes are isomorphic,
we will represent its (undirected) graph
with the lower-triangle of its adjacency matrix. We will place this lower-triangular
PLOS
4/13
w
w
x
Matrix:
Bit vector:
x
y
z
y
w
0
1
0
x
0
0
100
x
y
0
w
x
y
z
y
w
0
0
0
x
0
1
001
x
y
0
x
y
z
y
w
0
0
1
x
y
0
0
0
010
Fig 4. All 3-graphettes with exactly one edge; the canonical one is the one with lowest
integer representation (the middle one in this case). Each of them is placed in a lookup
table indexed by the bit vector representation of its adjacency matrix, pointing at the
canonical one. In this way we can determine that it is the one-edge 3-graphette in
constant time.
matrix into a bit vector, resulting in a representation similar to existing ones for orbit
identification [16].
We now describe the idea of a canonical representative of each isomorph. To provide
an explicit example, consider Fig 4, depicting the three isomorphic configurations of the
3-graphette that has exactly one edge. In order to determine that these graphettes are
all isomorphic, we take the bit vector representation depicted, and define the
lowest-numbered bitvector among all the isomorphs as the canonical representative. All
the other isomorphs in the lookup table point to it. In this way, every graph on 3 nodes
can be efficiently mapped to its canonical 3-isomorph.
We also automatically determine the number of automorphism orbits (see below) for
each canonical isomorph. Table 1 represents, for various values of k, the number of bits
1
b(k) required to store the lower-triangular matrix
of all graphettes on k nodes (i.e., the
length of the bit vector used to store this matrix); the resulting total number possible
representations of k nodes (which is simply 2b(k) ); the number of canonical isomorphs
N C(k); and the number of canonical automorphism orbits. Note that, to map each
possible set of k nodes to their canonical isomorphs, the lookup table has 2b(k) entries,
and each entry has a value between 0 and N C(k) − 1. Note that for k up to 8, the
graphettes can be stored in 32 bits. In that case, the maximum space required will be
32 × 228 = 1 GB. This is as far as we go, for now. Moore’s Law suggests that we may be
able to go to k = 9 within a few years, and to k = 10 in perhaps a decade or two.
We note that the most expensive part of our algorithm is creating the lookup table
between an arbitrary set of k nodes, to the canonical graphette represented by those k
nodes; in the absence of a requirement for this lookup table, one could use orbit
counting equations [16] to generate automorphism orbits up to k = 12.
Generating the lookup table from non-canonical to canonical
graphettes
Assume the large graph G has n nodes labeled 0 through n − 1, and pick an arbitrary
set of k nodes U = {u0 , u1 , . . . , uk−1 }. Create the subgraph g induced on the nodes in
U ⊆ V(G), and let its bit vector representation B be of the form lower-triangular matrix
described in Fig 4. We now describe how to create the lookup table that maps any such
PLOS
5/13
k
1
2
3
4
5
6
7
8
9
10
11
12
bits
b(k)
0
1
3
6
10
15
21
28
36
45
55
66
#Graphs
2b(k)
1
2
8
64
1K
32 K
2M
256 M
64 G
32 T
32 P
64 E
Space
b(k)2b(k)
0
0.25 B
3B
48 B
1.25 KB
60 KB
5.25 MB
896 MB
288 GB
180 TB
220 PB
528 EB
#Canonicals
N C(k)
1
2
4
11
34
156
1044
12346
274668
12005168
1018997864
165091172592
#Orbits
1
2
6
20
90
544
5096
79264
2208612
113743760
10926227136
1956363435360
required to store the
Table 1. For each value of k: the number of bits b(k) = k(k−1)
2
lower-triangle of the adjacency matrix for an undirected k-graphette; the number of
such k-graphettes counting all isomorphs which is just 2b(k) ; the number of canonical
k-graphettes (this will be the number of unique entries in the above lookup table [22],
and up to k = 8, 14 bits is sufficient); and the total number of unique automorphism
orbits (up to k = 8, 17 bits is sufficient) [27]. Note that up to k = 8, together the
lookup table for canonical graphettes and their canonical orbits fits into 31 bits,
allowing storage as a single 4-byte integer, with 1 bit to store whether the graphette is
connected (i.e., also a graphlet). The suffixes K, M, G, T, P, and E represent exactly
210 , 220 , 230 , 240 , 250 and 260 , respectively.
B to its canonical representative.
We iterate through all 2b(k) bit vectors in order; for each value B, we check to see if
it is isomorphic to any of the previously found canonical graphettes; if so, the lookup
table value is set to the previously found canonical graphette; otherwise we have a new,
previously unseen canonical graphette and the lookup table value is set to itself (B).
When checking for isomorphism between B and all previously found canonical
graphettes, we use a relatively simple brute force approach. If the degree distribution of
the two graphettes are different, we can immediately discard the pair as non-isomorphic;
otherwise we resort to cycling through every permutation of the nodes checking each
pair for graph equality, which has worst-case running time of k 2 k!. The total run time
to compute the lookup table for a particular value k is thus bounded above by
k 2 k! · N C(k) · 2b(k) , where k! is the maximum number of permutations we need to check
if a non-canonical matches an existing canonical, k 2 is the worst-case running time to
check if 2 specific permutations of k-graphettes are isomorphic, there are at most N C(k)
canonicals to check against [22], and 2b(k) = 2n(n−1)/2 is the total number of undirected
graphs on k nodes. More sophisticated approaches exist [23], which may more easily
allow higher values of k.
This process can also be parallelized, which is what we did for k = 8. Essentially, we
can split the 2b(k) non-canonical graphettes into m sets of about 2b(k) /m graphettes
each, and then spread the computation across m machines. For each of the m sets Si ,
we loop through all graphettes in that set and mark out which are isomorphic to each
other. For each set Si , we will find a set Ti of lowest-numbered “temporary” canonical
graphettes in Si , along with the map T C : Si → Ti of which graphettes in Si map to
each temporary canonical in Ti . That is, for each graphette g ∈ Si , ∃h ∈ Ti for which
the temporary canonical T C(g) = h. Finally, once all the m sets have been evaluated in
this way, a second stage passes through all the Ti , i = 0, . . . , m − 1, merging the
PLOS
6/13
temporary canonicals together into a final, global list of canonical graphettes, while also
propagating these globally lowest-numbered canonicals back up through the m
temporary canonical maps, so each graphette g globally maps to the globally
lowest-numbered canonical; we call this process sifting for canonicals, and it may
require several iterations to globally find the final list of canonicals. In this way we ran
k = 8 in about a week across 600 cores, for a total of 600 CPU-weeks. This process
could probably be made more efficient with smarter isomorphism checking [23, 24].
Graph automorphism and orbits
An isomorphism π : V(g) → V(g) (from a graph g to itself) is called an automorphism.
While an isomorphism is just a permutation of the nodes, it is called an
automorphism if it results in exactly the same labeling of the nodes in the same
order—in other words exactly the same adjacency matrix. The set of all automorphisms
of g will be called Aut(g).
An automorphism orbit, or just orbit, of g is a minimally sized collection of nodes
from V(g) that remain invariant under every automorphism of g [25]. There can be
more than one automorphism orbit, and each orbit can have anywhere from 1 to k
member nodes; refer again to Fig 1 for some examples. More formally, a set of nodes ω
constitute an orbit of g iff:
1. For any node u ∈ ω and any automorphism π of g, u ∈ ω ⇐⇒ π(u) ∈ ω.
2. if nodes u, v ∈ ω, then there exists an automorphism π of g and a γ > 0 so that
π γ (u) = v.
Now, we shall prove a few relevant results that will be useful later for automatically
enumerating the orbits.
Proposition 1. For each node u ∈ V(g) and each automorphism π : V(g) → V(g),
there exists an integer λ > 0 such that π λ (u) = u.
Proof. Because π is an automorphism,
u ∈ V(g) =⇒ π(u) ∈ V(g)
=⇒ π 2 (u) ∈ V(g)
..
.
=⇒ π i (u) ∈ V(g), ∀i ∈ N.
Since |V(g)| is finite and π is bijective, the conclusion obviously follows.
We shall call the set of nodes
Cπ (u) = {u, π(u), . . . , π λ−1 (u)}
the cycle of u under automorphism π, where λ is the smallest positive integer such that
π λ (u) = u.
Note that λ is not unique since π λ (u) = π 2λ (u) = · · · = u. Also, π, u, and λ are tied
together into triples such that knowing any two determines the third.
Corollary 1.1. π maps every node ∈ Cπ (u) to a node (possibly same) ∈ Cπ (u).
Corollary 1.2. In any automorphism π of g, every node appears in exactly one cycle.
PLOS
7/13
In other words, the cycles π creates are disjoint. (However, the cycles from different
automorphisms might not be so.) Hence, it makes sense to say splitting an
automorphism into its cycles. For example consider the permutation π = (201354) of
(012345). Since π(0) = 2, π(2) = 1, π(1) = 0, the nodes (012) form a cycle. Now start
with the next node, 3. π(3) = 3. So, (3) is another cycle. Finally, π(4) = 5, π(5) = 4, so,
(45) form another cycle. Hence, the permutation (201354) is split into three cycles,
namely (012), (3), (45).
Proposition 2. The orbits are disjoint. (In other words, each node appears in exactly
one orbit.)
Proof. Assume the contrary, i.e., a node u ∈ V(g) appears in two different orbits ω1 and
ω2 . According to the second condition, for any other node v ∈ ω1 , there exists an
automorphism π of g and a γ so that π γ (u) = v. However, from the first condition,
u ∈ ω2 =⇒ π(u) ∈ ω2
=⇒ π 2 (u) ∈ ω2
..
.
=⇒ π γ (u) ∈ ω2
=⇒ v ∈ ω2
Therefore, every node v ∈ ω1 also belongs to ω2 . Hence, ω1 ⊆ ω2 .
Following the same logic, ω2 ⊆ ω1 , implying ω1 = ω2 . ⇒⇐
Corollary 2.1. Each cycle appears in exactly one orbit, which completely contains that
cycle.
Proof. If an orbit ω partially contains a cycle Cπ (u), then ω is not invariant under
automorphism π, as π will map some node in ω (and Cπ (u)) to another node outside ω
(but still in Cπ (u)) according to corollary 1.1, contradicting our definition of orbits.
Since two orbits are disjoint, Cπ (u) must appear only in ω, and in none of the other
orbits.
These statements are enough to be able to find all orbits of each graphette, as we
now demonstrate.
Automatically enumerating all orbits of a graph
From the propositions in the previous section, an algorithm to enumerate the orbits can
be constructed like this:
1. Generate all automorphisms of g.
2. Split each automorphism into its cycles.
3. Merge the cycles from different automorphisms to form orbits.
Generating all automorphisms of g
Referring to Algorithm 1, the function generateAutomorphisms() applies every
possible permutation of V(g) over Adj(g). Each permutation creates an isomorph of
Adj(g). If Adj(g) is unchanged under some permutation π, then by definition, π is an
automorphism of g. Hence it is saved into Aut(g).
Two optimization strategies are employed:
PLOS
8/13
Algorithm 1 Automatically enumerating automorphism orbits of a graph
function generateAutomorphisms(Graph g)
Aut(g) = {} // Find the automorphisms of g
for each permutation π of V(g) do
apply π over Adj(g)
if Adj(g) == π(Adj(g)) then put π in Aut(g)
end if
end for
end function
function generateCycles(automorphism π)
C = {}
for node u in π do
if u is not visited then
mark u visited
new cycle Cπ (u) = {}
node v = π(u)
while v != u do
put v in Cπ (u)
mark v visited
v = π(v)
end while
put Cπ (u) in C
end if
end for
end function
function enumerateOrbits(C(g))
for each node u ∈ V(g) do ω(u) = u
end for
for cycle c ∈ C(g) do
let ωmin = ∞
for node u ∈ c do ωmin = min(ωmin , ω(u))
end for
for node u ∈ c do ω(u) = ωmin
end for
end for
end function
PLOS
9/13
1. No node is mapped to another node with unequal degree.
2. An automorphism of graph g is also an automorphism of its complement graph g 0 .
In practice, this algorithm generates all automorphisms of all the canonical
graphettes up to size 8 in a matter of seconds. Nevertheless, for additional speed up in
higher sizes, modern sophisticated automorphism detection algorithms [23, 24] may be
used.
Splitting automorphisms into cycles
An automorphism π of g is basically a permutation of nodes of g. Hence, to split π into
cycles, we can repeatedly apply π over every node u ∈ π and remember the nodes u
transforms into. This forms the cycle with node u, i.e. Cπ (u), which is saved in C. After
first visit, each node is marked visited to prevent more visits.
Merging cycles to enumerate orbits
Suppose C(g) is the set of all cycles resulting from all the automorphisms of g.
To enumerate orbits from it, first each node u is colored with a unique color
ω(u) = u. Then ω(u) is continuously updated to reflect the current color of u, as the
nodes belonging to same orbits are gradually colored by identical color.
For the nodes of each cycle c ∈ C(g), we save their minimum color in ωmin , and then
color all of them with ωmin . After coloring all the cycles in this way, nodes belonging to
same orbits get the same color, and hence, get enumerated.
Proof of correctness of Algorithm 1
Here we prove that Algorithm 1 determines every orbit of g.
Suppose a set ω is among the final sets generated by Algorithm 1. We shall prove ω
is an orbit of g by showing that it follows the two properties of orbits:
1. Let a node u ∈ ω form the cycle Cπ (u) under automorphism π. The
generateCycles function will apply π repeatedly until it finds a λ so that
π λ (u) = u and will therefore determine Cπ (u). Since the enumerateOrbits
function assigned u to ω, it had also assigned all nodes in Cπ (u) to ω. Hence
u ∈ ω ⇐⇒ π(u) ∈ ω.
2. Suppose nodes u, v ∈ ω. Then, either they belonged to a cycle from which they
were assigned to a mutual set ω in enumerateOrbits function, or there is a
third node w so that w shares separate cycles with u and v under different
automorphisms π1 and π2 . In the first case, u and v already belong to a common
cycle. In the second case, assume π1γ1 (w) = u and π2γ2 (w) = v. Consider the
permutation φ = π2γ2 ◦ π1−γ1 . Since composition of two automorphisms is an
automorphism [26], φ is also an automorphism. And notice that
φ(u) = π2γ2 π1−γ1 (u) = π2γ2 (w) = v
implying u and v belong to a common cycle under φ.
Therefore, ω is indeed an orbit of g. Since each node was given a unique orbit color in
the beginning of enumerateOrbits, every orbit of g will be eventually found by
Algorithm 1.
PLOS
10/13
Results and discussion
Using the algorithms described herein, we have enumerated all possible graphlets,
including the generalization of disconnected counterparts called graphettes, up to size
k = 8. The code and data can be found in http://github.com/Neehan/Faye.
(Note that the github code uses the upper triangle matrix, though we intend to convert
it to use the lower tringle as that representation has already been established [16].) We
have also enumerated all orbits up to size k = 8. More importantly to the statistical
sampling technique described in the Introduction, we have used a bit-vector
representation of all possible adjacency matrices of all possible sets of up to k = 8 nodes
and created a lookup table from the 2k(k−1)/2 k-sets to their canonical graphette
representatives. This allows us to determine, in constant time, the graphette represented
by these k nodes, as well as the automorphism orbits of each nodes. This allows efficient
estimation of both the global distribution of graphlets and orbits, as well as an
estimation of the graphlet (or orbit) degree vector for each node in a large graph G.
Although the lookup tables for k > 8 are at present too big to compute or store, we
could also use NAUTY or SAUCY to enumerate all the canonical graphettes up to size
k = 12, and use our orbit generation code Algorithm 1 to determine all the orbits in all
graphettes up to size k = 12. We have verified that previous results are consistent with
ours in terms of the number of distinct graphettes [22] and orbits [27] determined, as
displayed in Table 1.
In future work we will study which statistical sampling techniques most efficiently
produce a good estimate of the complete graphlet and local (per-node) degree vectors.
We also intend to study how this method may aid in cataloging of graphlets for
database network queries, or in non-alignment network comparison [10]. Finally, there
may be ways to combine our method with those of orbit counting equations [15, 16] to
more efficiently produce samples of orbit counts.
Acknowledgments
We thank Sridevi Maharaj, Dillon Kanne, and the anonymous referees for several
helpful suggestions on presentation.
References
1. Cook SA. The Complexity of Theorem-proving Procedures. In: Proceedings of
the Third Annual ACM Symposium on Theory of Computing. STOC ’71. New
York, NY, USA: ACM; 1971. p. 151–158. Available from:
http://doi.acm.org/10.1145/800157.805047.
2. Newman M. Networks: an introduction. 2010. United Slates: Oxford University
Press Inc, New York. 2010; p. 1–2.
3. Emmert-Streib F, Dehmer M, Shi Y. Fifty years of graph matching, network
alignment and network comparison. Information Sciences. 2016;346:180–197.
4. Wilson RC, Zhu P. A study of graph spectra for comparing graphs and trees.
Pattern Recognition. 2008;41(9):2833–2841.
5. Thorne T, Stumpf MP. Graph spectral analysis of protein interaction network
evolution. Journal of The Royal Society Interface. 2012; p. rsif20120220.
6. Dehmer M, Emmert-Streib F, Shi Y. Interrelations of graph distance measures
based on topological indices. PloS one. 2014;9(4):e94985.
PLOS
11/13
7. Milo R, Shen-Orr S, Itzkovitz S, Kashtan N, Chklovskii D, Alon U. Network
motifs: simple building blocks of complex networks. Science.
2002;298(5594):824–827.
8. Pržulj N, Corneil DG, Jurisica I. Modeling interactome: scale-free or geometric?
Bioinformatics. 2004;20(18):3508–3515.
9. Pržulj N. Biological network comparison using graphlet degree distribution.
Bioinformatics. 2007;23(2):e177–e183.
10. Yaveroğlu ÖN, Malod-Dognin N, Davis D, Levnajic Z, Janjic V, Karapandza R,
et al. Revealing the hidden language of complex networks. Scientific reports.
2014;4:4547.
11. Kuchaiev O, Milenković T, Memišević V, Hayes W, Pržulj N. Topological
network alignment uncovers biological function and phylogeny. Journal of The
Royal Society Interface. 2010;7(50):1341–1354. doi:10.1098/rsif.2010.0063.
12. Malod-Dognin N, Pržulj N. L-GRAAL: Lagrangian Graphlet-based Network
Aligner. Bioinformatics. 2015;doi:10.1093/bioinformatics/btv130.
13. Saraph V, Milenković T. MAGNA: maximizing accuracy in global network
alignment. Bioinformatics. 2014;30(20):2931–2940.
14. Mamano N, Hayes W. SANA: Simulated Annealing far outperforms many other
search algorithms for biological network alignment. Bioinformatics. 2017;0(0):8.
15. Hočevar T, Demšar J. A combinatorial approach to graphlet counting.
Bioinformatics. 2014;30(4):559–565. doi:10.1093/bioinformatics/btt717.
16. Melckenbeeck I, Audenaert P, Michoel T, Colle D, Pickavet M. An Algorithm to
Automatically Generate the Combinatorial Orbit Counting Equations. PLoS
ONE. 2016;11(1). doi:http://dx.doi.org/10.1371/journal.pone.0147078.
17. Chatr-aryamontri A, Breitkreutz BJ, Heinicke S, Boucher L, Winter A, Stark C,
et al. The BioGRID interaction database: 2013 update. Nucleic Acids Research.
2013;41(D1):D816–D823. doi:10.1093/nar/gks1158.
18. Pillich RT, Chen J, Rynkov V, Welker D, Pratt D. NDEx: A Community
Resource for Sharing and Publishing of Biological Networks. Protein
Bioinformatics: From Protein Modifications and Networks to Proteomics. 2017; p.
271–301.
19. Camacho C, Coulouris G, Avagyan V, Ma N, Papadopoulos JS, Bealer K, et al.
BLAST+: architecture and applications. BMC Bioinformatics. 2009;10:421.
20. Rahman M, Bhuiyan MA, Al Hasan M. Graft: An efficient graphlet counting
method for large graph analysis. IEEE Transactions on Knowledge and Data
Engineering. 2014;26(10):2466–2478.
21. Pržulj N, Corneil DG, Jurisica I. Efficient estimation of graphlet frequency
distributions in protein–protein interaction networks. Bioinformatics.
2006;22(8):974–980.
22. Sloane N. Online Encyclopedia of Integer Sequences (OEIS);. Available from:
http://oeis.org/A000088.
23. Mckay BD. Nauty; 2010. Available from:
http://users.cecs.anu.edu.au/˜bdm/nauty.
PLOS
12/13
24. Codenotti P, Katebi H, Sakallah KA, Markov IL. Conflict Analysis and
Branching Heuristics in the Search for Graph Automorphisms. In: Tools with
Artificial Intelligence (ICTAI). IEEE; 2013.
25. Gross JL. Graph Theory – Lecture 2: Structure and Representation — Part A;.
Available from:
http://www.cs.columbia.edu/˜cs4203/files/GT-Lec2.pdf.
26. Automorphism of a group;. Available from: https:
//groupprops.subwiki.org/wiki/Automorphism_of_a_group.
27. Sloane N. Online Encyclopedia of Integer Sequences (OEIS);. Available from:
http://oeis.org/A000666.
PLOS
13/13
| 8cs.DS
|
Reinforcement Learning in a large scale photonic Recurrent Neural Network
J. Bueno1 , S. Maktoobi2 , L. Froehly2 , I. Fischer1 , M. Jacquot2 , L. Larger2 , D. Brunner2,∗
arXiv:1711.05133v2 [cs.NE] 15 Nov 2017
2
1
Instituto de Fı́sica Interdisciplinar y Sistemas Complejos, IFISC (UIB-CSIC),
Campus Universitat de les Illes Balear, E-07122 Palma de Mallorca, Spain. and
FEMTO-ST Institute/Optics Department, CNRS & University Bourgogne Franche-Comté,
15B avenue des Montboucons, 25030 Besanon Cedex, France.
∗
Corresponding author: [email protected]
Photonic Neural Network implementations have been gaining considerable attention as a potentially disruptive future technology. Demonstrating learning in large scale neural networks is essential
to establish photonic machine learning substrates as viable information processing systems. Realizing photonic Neural Networks with numerous nonlinear nodes in a fully parallel and efficient learning
hardware was lacking so far. We demonstrate a network of up to 2500 diffractively coupled photonic
nodes, forming a large scale Recurrent Neural Network. Using a Digital Micro Mirror Device, we
realize reinforcement learning. Our scheme is fully parallel, and the passive weights maximize energy
efficiency and bandwidth. The computational output efficiently converges and we achieve very good
performance.
I.
INTRODUCTION
Multiple concepts of Neural Networks have initiated
a revolution in the way we process information. Deep
Neural Networks outperform humans in challenges previously deemed unsolvable by computers [1]. Among others, these systems are now capable to solve non-trivial
computational problems in optics [2]. At the same time,
Reservoir Computing (RC) emerged as a Recurrent Neural Network (RNN) concept [3]. Initially, RC received
substantial attention due to excellent prediction performance achieved with minimal optimization effort. However, quickly it was realized that RC is highly attractive
for analogue hardware implementations [4, 5].
As employed by the machine learning community, Neural Networks (NNs) consist of a large number of nonlinear nodes interacting with each other. Evolving the
NNs’ state requires performing vector-matrix products
with possibly millions of entries. Neural Network concepts therefore fundamentally benefit from parallelism.
Consequently, photonics was identified as an attractive
alternative to electronic implementation [6, 7]. Early
implementations were bulky and suffered from lack of
adequate technology and NN concepts. This recently
started to change, firstly because RC enabled a tremendous complexity-reduction of analog electronic and photonic RNNs’ [5, 8–11]. In addition, integrated photonic
platforms have matured and integrated photonic Neural
Networks are now feasible [12]. Various demonstrations
how a particular network of neurons can be implemented
have been realized in hardware. Yet, Neural Networks
consisting of numerous photonic nonlinear nodes combined with photonically implemented learning were so
far only demonstrated in a delay systems controlled by a
Field Programmable Gate Array [13]. Due to the timemultiplexing, delay system NNs fundamentally require
such auxiliary infrastructure and computational speed
suffers due to their serial nature.
While networks with multiple nodes are more challenging to implement, they offer key advantages in terms of
parallelism, speed and for realizing the essential vectormatrix products. Here, we demonstrated a network of
up to 2025 nonlinear network nodes, where each node
is a pixel of a Spatial Light Modulator (SLM). Recurrent and complex network connections are implemented
using a Diffractive Optical Element (DOE), an intrinsically parallel and passive device [14]. Simulations based
on the angular spectrum of plane waves show that the
concept is scalable to well over 20.000 nodes. In a photonic RNN with N =900 nodes we implement learning
using a digital micro-mirror device (DMD). The DMD is
intrinsically parallel as well and, once weights have been
trained, passive and energy efficient. Both, the coupling
and learning concepts’ bandwidth and power consumption is not impacted by the system’s size, offering attractive scaling properties. Here we apply such a passive
and parallel readout layer to an analogue hardware RNN,
and introduce learning strategies improving performance
of such systems. Using reinforcement learning we implement timeseries prediction with excellent performance.
Our findings open the door to novel and versatile photonic Neural Network concepts.
II.
NONLINEAR NODES AND DIFFRACTIVE
NETWORK
Figure 1(a) conceptually illustrates our RNN. Information enters the system via a single input node, from
where it is injected into a recurrently connected network
of nonlinear nodes. The computational result is provided
at the single output node after summing the network’s
state according to weight matrix WDM D . Following the
RC concept, one can choose the input and recurrent internal weights randomly [3]. Here, we create a complex
and recurrently connected network using imaging which
is spatially structured via a DOE, resulting in the internal connectivity matrix WDOE [14].
In Fig. 1(b) we schematically illustrate our experimental setup. The polarization of an illumination laser
2
of SLM (12.5 µm) and camera (5.2 µm). The detected
sat
is
state is xi = (αEi )2 with α = IGS
sat · N D. Here, I
the camera’s saturation intensity and N D the transmission through a neutral density filter (ND) always selected
such that the the dynamical range of the camera is fully
exploited, while avoiding over-exposure (maximum gray
scale GS=255). After multiplication with scalar β, we
add a constant phase offset θi and send the resulting matrix back to the SLM. Ignoring the DOE’s effect, each
pixel therefore corresponds to an Ikeda map:
xi (n + 1) = GS cos2 (βxi (n) + θi ).
FIG. 1. (a) Schematic illustration of a recurrent neural network. (b) Experimental setup. The NN state encoded on the
SLM is imaged onto the camera (CAM) through a polarizing
beam splitter (PBS) and a diffractive optical element (DOE).
In this way, we realize a recurrently coupled network of Ikeda
maps. A digital micromirror device (DMD) creates a spatially
modulated image of the SLM’s state. The integrated state,
obtained via superimposing the modulated intensities, is detected, creating the photonic RNN’s output. (c) The RNN’s
coupling matrix established by the DOE, with the inset showing a zoom into a smaller region.
(Thorlabs LP660-SF20, λ=661.2 nm, Ibias =89.69 mA,
T =23 ℃) is adjusted to s-polarization and the polarizing beam splitter cube (PBS) therefore reflects all light
towards the SLM. By focusing the illumination laser
onto the first microscope objective’s (MO1, Nikon CFI
Plan Achro 10X) back focal plane, the SLM (Hamamatsu
X13267-01) is illuminated by a plane wave. The λ2 -plate
in front of MO1 is adjusted such that the SLM operates
in intensity modulation mode. Consequently the optical
field transmitted through the PBS (p-polarization) for
each SLM pixel i is modulated according to
Ei = Ei0 cos(2π
xi
),
κSLM
(1)
where Ei0 and xSLM
are the illumination and gray scale
i
value of pixel i, respectively. κSLM is the SLM’s conversion factor between pixel gray scale and polarization
rotation angle.
Ignoring for now the DOE’s effect for explanatory purposes, the transmitted field is imaged (MO2, Nikon CFI
Plan Achro 10X) on a mirror, and double passing through
a λ4 -plate results in a s-polarized field. The PBS therefore reflects the entire optical field, which is consecutively
imaged (MO2, Nikon CFI Plan Fluor 4X) on the camera (CAM, Thorlabs DCC1545M). We rescale the camera
image via linear interpolation to fit the number of pixels
of the SLM. This step is necessary due to (i) an imaging magnification of 2.5, and (ii) the different pixel sizes
(2)
We write the SLM state as vector x(n + 1), yet in the
experiment this state corresponds to a square array of
SLM pixels.
Illumination wavelength, DOE (HOLOOR MS-443650-Y-X), as well as MO1 and MO2 were chosen such
that the spacing between diffractive orders matches the
pixel-spacing of the SLM. Therefore, upon adding the
DOE to the beam path, the optical field on the camPN DOE
era becomes EiC = j Wi,j
Ej , where WDOE is the
networks coupling matrix created by the DOE. In Fig.
1(c) we show the experimentally obtained W DOE for a
network of 30×30 nodes. Upon inspection of the inset
one can see that locally connectivity strengths vary significantly. This is due to each pixel illuminating a DOE
area comparable to the DOE’s lowest spatial frequency.
As this area shifts slightly from pixel to pixel, the intensity distribution between diffractive orders varies. This
intended effect inherently creates the heterogeneous photonic network topology needed for computation [3]. Finally, the photonic RNN’s state x(n + 1) is given by
xi (n + 1) = GS · cos2 (β · α
N
X
2
DOE
Wi,j
Ej
j
+ γWiinj u(n + 1) + θi ). (3)
Here u(n + 1) is the information to be injected into the
RNN and γ is the signal injection strength. Matlab is
used to control all instruments, update the network state
and to inject the input information weighted by matrix
γWiinj . The overall update rate of the entire system is
∼5 Hz. Currently, the maximum size of networks we
can realize consist of ∼2500 nodes, which is limited by
the imaging setup’s field of view and not by the concept
itself.
III.
NETWORK READOUT WEIGHTS
Having created a recurrent photonic neural network
driven by external data, the final to information processing is to adjust the system such that it performs the
required computation. This is typically achieved by modifying connection weights according to some learning routine. Inspired by the RC concept, we constrain learning
3
(b)
(a)
employ a version of reinforcement learning. The learning
input signal is injected after inverting the weight assigned
to one node. The error εk of signal ykout (n + 1) obtained
for configuration WkDM D is then compare to the error
εk−1 , where k is the index of learning iterations. If the
error is reduced, we keep DMD configuration WkDM D ,
DM D
if not, we revert back to Wk−1
and invert a different
weight. The weight to be updated is determined by the
largest entrys Wkselect,max position lk according to
h
FIG. 2. (a) Learning curves for a photonic RNN with two
phase offsets φi across the network. A fraction of (1 − µ)
nodes have a response function with a negative slope (Θ0 =
0.17π), the remaining nodes experience a response function
with a positive slope (Θ0 = 0.43π). (b) An almost symmetric
distribution of phase offsets, resulting in positive and negative
node responses, benefits computational performance.
induced weight adjustment to the readout layer. Our 900
RNN nodes are spatially distributed, and we therefore
can use a simple lens (Thorlabs AC254-400-B) to image
the SLM’s state onto a commercial array of micro mirrors (DLi4120 XGA, pitch 13.68 µm). Micro mirrors can
be flipped between ±12 °, such that for -12 °the optical
signal is directed to a detector. The detectors photo current then corresponds to the RNN output. With WDM D
as the readout weight vector, the RNN output becomes
y out (n + 1) = δ WDM D (1 − x(n + 1)) .
(4)
Here, δ relates the power recorded by the power meter
(Thorlabs S150C and PM100A) to the SLM state. The
signal directed towards the DMD is orthogonally polarized compared to the one directed to the camera, resulting in xDM D (n) = 1 − x(n). In the experiment weight
vector WDM D corresponds to a square matrix, which can
be seen in Fig. 1(b). The image labeled DMD shows a
typical configuration of the DMD with WDM D . As the
contribution of each node is either on or off, WDM D consists of Boolean entries only. Weights are not temporal
modulations as in delay system implementations of RC
[13], and therefore can be implemented by passive attenuations in reflection or transmission. Such passive processes are energy efficient and typically do not results in
a bandwidth limitation. In this specific implementation,
once trained, mirrors can simply remain in their position,
and if mechanically fixed, would not consume additional
energy. Also, readout eq. 4 is optically performed for all
elements in parallel.
IV.
PHOTONIC LEARNING
It is now the task of a learning algorithm to tailor
WDM D such that signal y out (n + 1) approximates a target value as good as possible. In our experiment, we
Wkselect =rand(N ) · Wbias ,
i
lk , Wkselect,max = max(Wkselect ),
DM D
Wk,l
=
k
DM D
¬(Wk−1,l
),
k
(5)
(6)
(7)
1
+ Wbias , Wlbias
= 0.
(8)
k
N
rand(N ) creates a random vector with N entries, and
at the start W1DM D and Wbias are randomly initialized.
Wbias acts as a bias vector, whose values are increased
by N1 each learning iteration while the bias belonging to
the most recently updated weight is set to zero. This results in a randomized selection process with a bias away
from inverting recently updated weights. In simulations
we found that reinforcement learning including such a
bias showed significantly faster learning convergence. As
a task to be performed by our system, we chose nonlinear time series prediction. The injected signal u(n + 1)
is the chaotic Mackey-Glass (MG) sequence [3], and the
RNN’s learning target is y T (n + 1) = u(n + 2), the onetime-step-prediction of the MG system. Parameters of
the temporal MG sequence where identical to [15], using
an integration step size of 0.1. For determining the error
εk we discarded the first 30 data points due to their transient nature. The RNNs remaining output sequence was
then inverted, its offset subtracted and normalized by its
standard deviation, creating signal ỹkout . The error was
measured by εk = σ(y T − ỹkout ), where σ is the standard
deviation and εk therefore corresponds to the normalized
mean square error (NMSE).
At this stage we would like to stress a significant difference between neural networks emulated on digital electronic computers and our photonic hardware implementation. In our system, all connection weights are positive, and WDM D is boolean. This restricts the functional space available for approximating the targeted
input-output transformation. As a result, first evaluations of the learning procedure and prediction of the
MG series suffered from minor performance. However,
we were able to mitigate this restriction by harnessing the non-monotonous slope of the cos2 nonlinearity. We randomly divided the offest phases θi |i=1...N ,
resulting in nodes with negative and positive slope of
their response function. We chose Θ0 = 42=0.17π
ˆ
and
Θ0 + ∆Θ = 106=0.43π,
ˆ
respectively, with a probability of 1 − µ for θi = Θ0 . As RNN-states and WDOE
entries are exclusively positive, the nonlinear transformation of nodes with θi = Θ0 is predominantly along a
positive slope, for θi = Θ0 + ∆Θ along a negative slope.
Wbias =
4
larger by a factor of 2.2 relative to a delay RC based on
a semiconductor laser [15] and by 6.5 relative to a MachZehnder modulator based setup [16]. These comparisons
have to be evaluated in the light of the significantly increased level of hardware implementation in our current
setup. In [15, 16], readout weights were applied digitally
in an off-line procedure using weights with double precision. In [16] a strong impact of digitization resolution on
the computational performance was identified, suggesting that ε can be significantly reduced by increasing the
resolution of WDM D .
FIG. 3.
(a) Learning performance at optimal parameters
(β=0.8, γ=0.4, µ=0.45). (b) The photonic RNN’s predicted
output in µW (blue line) can hardly be differentiated from
the prediction target (red dots). Prediction error ε is given
by the yellow dashed data.
This enables the reinforcement learning procedure to select from nonlinear transformations with positive and
negative slopes. We used feedback β = 0.8 and injection gain γ = 0.4, and learning curves for various ratios
(µ = [0.25, 0.35, 0.45, 0.5]) are shown in Fig. 2 (a). They
reveal a strong impact of this symmetry breaking. Optimum performance for each µ is shown in 2(b). Best
performance is found for an RNN operating around almost equally distributed operating points at µ = 0.45.
This demonstrates that the absence of negative weights
in WDM D , WDOE and x can be partially compensated
for by incorporating nonlinear transformations with positive as well as negative slopes. This result is of high
significance for optical neural networks, which, e.g. motivated by robustness considerations, renounce making
use of the optical phase to implement negative weights.
We further optimized our system’s performance by
scanning the remaining parameters β and γ. In Fig. 3
(a) we show the error convergence under optimized global
conditions for a training sample size of 500 steps (blue
stars). The error efficiently reduces, and finally stabilizes
at ε ≈ 0.013. Considering learning is limited to Boolean
readout weights this is an excellent result. After training, the prediction performance is evaluated further on
a sequence of 4500 datapoints consecutive data points
which were not part of the training dataset. As indicated by the red line in the same panel, the testing error
matches the training error. We can therefore conclude
that our photonic RNN successfully generalized the underlying target system’s properties. The excellent prediction performance can be appreciated in Fig. 3 (b).
Data belonging to the left y-axis (blue line) shows the
recorded output power, while on the right y-axis (red
dots) we show the normalized prediction target signal. A
difference between both is hardly visible, and the prediction error ε (yellow dashed line) is small. Down-sampling
the injected signals by 3 creates condition identical to
[15, 16]. Under these conditions our error (ε = 0.042) is
V.
CONCLUSION
We demonstrated a photonic RNN consisting of hundreds of photonic nonlinear nodes and the implementation of photonic reinforcement learning. Using a simple
Boolean valued readout implemented with a DMD, we
trained our system to predict the chaotic MG sequence.
The resulting prediction error is very low despite of the
Boolean readout weights.
In our work we demonstrate how symmetry breaking inside the RNN can compensate for exclusively positive intensities in our analogue neural networks systems. These results resolve a complication of general
importance to neural networks implemented in analog
hardware. Hardware-implemented networks and readout
weights based on physical devices open the door to a
new class of experiments, i.e. evaluating the robustness
and efficiency of learning strategies in fully implemented
analogue neural networks. The final step, a photonic realization of the input, should be straight forward, as it
only requires a complex spatial distribution of the input information. Finally, our system is not limited to
the reported slow opto-electronic system. Extremely fast
all-optical systems can be realized employing the same
concept since we intentionally implemented a 4f architecture to allow for self-coupling [14].
FUNDING INFORMATION
The authors acknowledge the support of the Region
Bourgogne Franche-Comté. This work was supported
by Labex ACTION program (contract ANR-11-LABX0001-0) and via the Volkswagen Foundation NeuroQNet
project.
ACKNOWLEDGMENTS
The authors would like to thank Christian Markus Dietrich for valuable contributions to earlier versions of the
setup.
5
[1] Y. LeCun, Y. Bengio, and G. Hinton, Nature 521, 436
(2015).
[2] A. Sinha, J. Lee, S. Li, and G. Barbastathis, Optica 4,
1117 (2017).
[3] H. Jaeger and H. Haas, Science (New York, N.Y.) 304,
78 (2004).
[4] K. Vandoorne, W. Dierckx, B. Schrauwen, D. Verstraeten, R. Baets, P. Bienstman, and J. Van Campenhout, Optics Express 16, 11182 (2008).
[5] L. Appeltant, M. C. Soriano, G. Van der Sande, J. Danckaert, S. Massar, J. Dambre, B. Schrauwen, C. R. Mirasso,
and I. Fischer, Nature Communications 2, 468 (2011).
[6] K. Wagner and D. Psaltis, Applied Optics 26, 5061
(1987).
[7] C. Denz, Optical Neural Networks (Springer Vieweg,
Wiesbaden, 1998).
[8] F. Duport, B. Schneider, A. Smerieri, M. Haelterman,
and S. Massar, Optics Express 20, 22783 (2012).
[9] Y. Paquot, F. Duport, A. Smerieri, J. Dambre,
B. Schrauwen, M. Haelterman, and S. Massar, Scien-
tific Reports 2, 287 (2012).
[10] L. Larger, M. C. Soriano, D. Brunner, L. Appeltant, J. M.
Gutierrez, L. Pesquera, C. R. Mirasso, and I. Fischer,
Optics Express 20, 3241 (2012).
[11] D. Brunner, M. C. Soriano, C. R. Mirasso, and I. Fischer,
Nature Communications 4, 1364 (2013).
[12] Y. Shen, N. C. Harris, S. Skirlo, M. Prabhu, T. BaehrJones, M. Hochberg, X. Sun, S. Zhao, H. Larochelle,
D. Englund, and M. Soljacic, Nature Photonics 11, 441
(2017), arXiv:1610.02365.
[13] P. Antonik, M. Haelterman, and S. Massar, Cognitive
Computation , 1 (2017).
[14] D. Brunner and I. Fischer, Optics Letters 40, 3854
(2015).
[15] J. Bueno, D. Brunner, M. Soriano, and I. Fischer, Optics
Express 25, 2401 (2016).
[16] M. C. Ortı́n, S.and Soriano, L. Pesquera, D. Brunner,
D. San-Martı́n, I. Fischer, C. R. Mirasso, and J. M.
Gutiérrez, Scientific Reports 5, 14945 (2015).
| 9cs.NE
|
Decentralized Autonomous Navigation
Strategies for Multi-Robot Search and Rescue
by
arXiv:1605.04368v1 [cs.RO] 14 May 2016
Ahmad Baranzadeh
2016
Abstract
Use of multi-robot systems has many advantages over single robot systems in various applications. However, it comes with its own complexity and challenges. In this report, we try to
improve the performance of existing approaches for search operations in multi-robot context.
We propose three novel algorithms that are using a triangular grid pattern, i.e., robots certainly
go through the vertices of a triangular grid during the search procedure. The main advantage
of using a triangular grid pattern is that it is asymptotically optimal in terms of the minimum
number of robots required for the complete coverage of an arbitrary bounded area. Therefore,
using the vertices of this triangular grid coverage guarantees complete search of a region as well
as shorter searching time. We use a new topological map which is made and shared by robots
during the search operation. We consider an area that is unknown to the robots a priori with
an arbitrary shape, containing some obstacles. Unlike many current heuristic algorithms, we
give mathematically rigorous proofs of convergence with probability 1 of the algorithms. The
computer simulation results for the proposed algorithms are presented using a simulator of real
robots and environment. We evaluate the performance of the algorithms via experiments with
real Pioneer 3DX mobile robots. We compare the performance of our own algorithms with
three existing algorithms from other researchers. The results demonstrate the merits of our
proposed solution.
A further study on formation building with obstacle avoidance for a team of mobile robots
is presented in this report. We propose a robust decentralized formation building with obstacle
avoidance algorithm for a group of mobile robots to move in a defined geometric configuration.
Furthermore, we consider a more complicated formation problem with a group of anonymous
robots; these robots are not aware of their position in the final configuration and need to
reach a consensus during the formation process. We propose a randomized algorithm for the
anonymous robots that achieves the convergence to a desired configuration with probability
1. We also propose a novel obstacle avoidance rule, used in the formation building algorithm.
A mathematically rigorous proof of the proposed algorithm is given. The performance and
applicability of the proposed algorithm are confirmed by the computer simulation results.
ii
Contents
1 Introduction
1.1 Multi-Robot Systems . . . . . . . . .
1.2 Multi-Robot Search and Rescue . . .
1.3 Communication and Coordination . .
1.4 Centralized vs. Decentralized . . . .
1.5 Consensus Variables Rule . . . . . .
1.6 Mapping . . . . . . . . . . . . . . . .
1.7 Formation Building . . . . . . . . . .
1.8 Contributions of This Report . . . .
1.8.1 Main Contribution Highlights
1.9 Report Outline . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2 Consensus Variables Locating Algorithm
2.1 Problem statement . . . . . . . . . . . .
2.2 Consensus Variables Locating Algorithm
2.3 Simulation Results . . . . . . . . . . . .
2.4 Summary . . . . . . . . . . . . . . . . .
3 Random Triangular Grid-Based Search
3.1 Distributed Random Search Algorithm
3.1.1 Searching the Whole Area . . .
3.1.2 Searching for Targets . . . . . .
3.1.3 Patrolling . . . . . . . . . . . .
3.1.4 Robots’ Motion . . . . . . . . .
3.2 Simulation Results . . . . . . . . . . .
3.2.1 Searching the Whole Area . . .
3.2.2 Searching for Targets . . . . . .
3.3 Summary . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
15
15
19
22
23
Algorithm
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
25
25
26
29
29
30
30
31
34
36
.
.
.
.
.
.
.
38
38
39
42
43
43
44
44
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4 Semi-Random Triangular Grid-Based Search Algorithm
4.1 Distributed Semi-Random Search Algorithm . . . . . . . .
4.1.1 Searching the Whole Area . . . . . . . . . . . . . .
4.1.2 Searching for Targets . . . . . . . . . . . . . . . . .
4.1.3 Patrolling . . . . . . . . . . . . . . . . . . . . . . .
4.1.4 Robots’ Motion . . . . . . . . . . . . . . . . . . . .
4.2 Simulation Results . . . . . . . . . . . . . . . . . . . . . .
4.2.1 Searching the Whole Area . . . . . . . . . . . . . .
iii
.
.
.
.
.
.
.
.
.
.
2
2
3
6
8
8
9
9
11
12
13
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4.3
4.2.2 Searching for Targets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5 Modified Triangular-Grid-Based Search Algorithm
5.1 Distributed Triangular Grid-Based Search Algorithm . . . . . . . . . .
5.1.1 Searching the Whole Area . . . . . . . . . . . . . . . . . . . . .
5.1.2 Searching for Targets . . . . . . . . . . . . . . . . . . . . . . . .
5.1.3 Patrolling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.4 Robots’ Motion . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.1 Searching the Whole Area . . . . . . . . . . . . . . . . . . . . .
5.2.2 Searching for Targets . . . . . . . . . . . . . . . . . . . . . . . .
5.3 Experiments with Real Robots . . . . . . . . . . . . . . . . . . . . . . .
5.4 Comparison Between Presented Search Algorithms and Other Methods
5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
6 Formatiom Building with Obstacle Avoidance
6.1 Multi-Robot System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2 Formation Building . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3 Formation Building with Anonymous Robots . . . . . . . . . . . . . . . . . .
6.4 Obstacle Avoidance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.5.1 Obstacle Avoidance . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.5.2 Formation Building with Obstacle Avoidance . . . . . . . . . . . . .
6.5.3 Formation Building with Anonymous Robots and Obstacle Avoidance
6.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7 Conclusions
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
52
52
53
56
57
57
57
58
60
65
68
69
.
.
.
.
.
.
.
.
.
72
72
74
79
81
84
85
86
89
89
92
1
Chapter 1
Introduction
The past three decades have seen increasingly rapid advances in robotics. Also, mobile robots
area; usually referred to unmanned aerial vehicles (UAVs), unmanned ground vehicles (UGVs),
unmanned underwater vehicles, driverless cars, etc., has been a major area of interest within
the field of robotics. Using mobile robots has recently become very popular in various tasks.
Mobile robots are employed to achieve from very simple home tasks like vacuum cleaning [1,2] to
some important and complicated jobs like finding a dangerous odour source in a warehouse [3].
They have been used to help or replace human in various applications such as unmanned aircraft
systems (UAS) [4–6], factory automation [7,8], mining [9–11], home and office assistance [12,13],
interactive guide systems [14], search and rescue operations [15, 16], search and exploration in
hazardous environments [17], and sensor networks [18, 19].
A single-robot system refers to only one individual robot which can model itself, the environment and their interaction [20]. Use of a single robot in applications mentioned above has
been studied by many researchers, and this field of research is well developed. NASA Mars
Pathfinder [21], Sony Aibo, a robotic dog capable of seeing and walking [22], and Boston Dynamics BigDog that is able to walk on icy terrain and recover its balance when kicked from
the side [23], are famous examples of mobile robots. Despite advances in single-robot systems
that enable them to overcome many obstacles and achieve a lot of operations, there are still
some tasks that are inherently impossible or too complicated to be done by an individual robot.
Furthermore, there are also some tasks that might not be timely or economically efficient to
be done by a single robot. For instance, two keys in different parts of an environment must be
activated concurrently, cannot be accomplished by a single robot [24].
1.1
Multi-Robot Systems
On the other side, using a team consisting of some autonomous mobile robots instead of only a
single robot for aforementioned tasks, is a newer approach that has attracted many researchers
in recent years [25–28]. Though, using multi-robot systems may cause some new challenging
problems in contrast to single-robot systems, they have a lot of advantages that assure the
researchers to have an increased interest in this topic. Furthermore, recent developments in
the fields of microelectronics, measurement and sensors, wireless communication networks,
and computer hardware and software, have led to a renewed interest in multi-robot systems.
Compared to single-robot systems, multi-robot systems have several potential advantages which
are [29–33]:
2
• More reliable; as a result of information sharing among robots, any failure in one robot
would not cause a failure of the whole system.
• Robust and fault-tolerant; using multi-robot systems can reduce errors in odometry, communication, sensing and so on.
• Scalable; with growing amount of work, the algorithm applied to multi-robot systems still
works properly.
• Flexible; they can adopt themselves with any changes in the environment.
• Lower cost; using some simple robots is cheaper than a single complex and powerful robot.
• A better spatial distribution.
• A better overall system performance in terms of time or the energy consumption required
completing a task.
Despite aforementioned advantages, multi-robot systems consist of many different and yet
related research topics and problems. Communication and coordination of the robots during the
operation, localization, and mapping are some of the most significant problems of multi-robot
systems which have been discussed in the next sections.
1.2
Multi-Robot Search and Rescue
In the field of multi-robot systems, search and rescue, exploration, foraging and patrolling are
some different terms that are closely related. A considerable amount of literature has been published on these topics [34–39]. Although these topics may seem distinct, their related problems
in multi-robot systems are almost similar so that they can be classified in the same category.
All of these problems can be done by a single robot, and there are many articles have been
published about that. But as outlined before, using a team of robots has many advantages
over a single robot that make it powerful so that it can accomplish a search task more efficient in terms of time and cost. Multi-robot exploring systems have numerous applications
such as exploring an entire region for unknown number of targets [40, 41], search and rescue
operations [42], surveyance a building to detect potential intruders (e.g., museums and laboratories) [43], patrolling in outdoor environments (e.g., a country borders) [44], intruder detection
and perimeter protection [45, 46].
This report intends to present some algorithms which can be applied in the search and
rescue operations (SAR) [47, 48] as well as other exploring tasks. Clearly, it is significant to
save human lives in a disaster such as an earthquake, a plane crash, a missed boat in the ocean
and a fire in an industrial warehouse. In most of such cases, search and rescue operations are
usually difficult or even impossible to be performed by humans. Therefore, using a team of
robots would be helpful to carry out the operation to increase the survival rate and to decrease
the risk of the operation. Besides, in a search and rescue operation, the search area must
be completely explored ensuring all possible targets are detected. Our suggested algorithms
thoroughly cover this goal in the sense that they guarantee exploring all points of the search
area. In the following, we mention some of the studies in this field carried out in recent years
by other researchers.
3
Beard and McLain [49] proposed a dynamic programming approach for a team of cooperating UAVs to explore a region as much as possible while avoiding regions of hazards. Vincent
and Rubin [50] used predefined swarm patterns to address the problem of cooperative search
strategies for UAVs searching for moving targets in hazardous environments. Yang et al. [51]
used distributed reinforcement learning for multi-agent search wherein the robots learn the
environment online and store the information as a map and utilise that to compute their route.
Baxter et al. [52] described two implementations of a potential field sharing multi-robot
system; pessimistic and optimistic. They considered that robots perform no reasoning and are
purely reactive in nature. They pointed out that potential field sharing has a positive impact on
robots involved in a search and rescue problem. Marjovi et al. [53] presented a method for fire
searching by a multi-robot team. They proposed a decentralized frontier based exploration by
which the robots explore an unknown environment to detect fire sources. In [15], Guarnieri et
al. presented a search and rescue system named HELIOS team, consisting of five tracked robots
for urban search And rescue. Two units are equipped with manipulators for the accomplishment
of particular tasks, such as the handling of objects and opening doors; the other three units,
equipped with cameras and laser range finders, are utilized to create virtual 3D maps of the
explored environment. The three units can move autonomously while collecting the data by
using a collaborative positioning system (CPS).
Sugiyama et al. [54] investigated the autonomous chain network formation by multi-robot
rescue systems. According to Sugiyama et al. [54], chain networks connecting a base station
and rescue robots are essential to reconnoiter distant spaces in disaster areas, and the chains
must be formed to assure communications among them and must be transformed if the target of
exploration changes. They adopted autonomous classification of robots into search robots and
relay robots so that the robots act according to the behavior algorithms of each class of robot
to form chain network threading the path to the distant spaces. In [55], Luo et al. proposed
an approach that employs a team consisting of ground and aerial vehicles simultaneously. The
ground vehicle is used for the purpose of environment mapping, the micro aerial vehicle is
used simultaneously for the purpose of search and localization with a vertical camera and a
horizontal camera, and two other micro ground vehicles with sonar, compass and colour sensor
used as the back-up team.
Lewis and Sycara [56] demonstrated the evolution of an experimental human-multi-robot
system for urban search and rescue in which operators and robots collaborate to search for
victims. Macwan and Nejat [57] presented a modular methodology for predicting a lost person’s
behavior (motion) for autonomous coordinated multi-robot wilderness search and rescue. They
asserted introducing new concept of isoprobability curves, which represents a unique mechanism
for identifying the targets probable location at any given time within the search area while
accounting for influences such as terrain topology, target physiology and psychology, clues
found, etc. Mobedi and Nejat [58] developed an active 3-D sensory system that can be used
in robotic rescue missions to map these unknown cluttered urban disaster environments and
determine the locations of victims.
Sugiyama et al. [59] proposed a system procedure for a multi-robot rescue system that
performs real-time exploration over disaster areas. The proposed system procedure consists of
the autonomous classification of robots into search and relay types, and behavior algorithms
for each class of robot. Searching robots explore the areas, and relay robots act as relay
terminals between searching robots and the base station. The rule of the classification and
the behavior algorithm refer to the forwarding table of each robot constructed for ad hoc
networking. The table construction is based on DSDV (destination-sequenced distance vector)
4
routing that informs each robot of its topological position in the network and other essentials.
In [60], Liu et al. proposed a hierarchical reinforcement learning (HRL) based semi-autonomous
control architecture for rescue robot teams to enable cooperative learning between the robot
team members. They claimed that the HRL-based control architecture allows a multi-robot
rescue team to collectively make decisions regarding which rescue tasks need to be carried out
at a given time, and which team member should execute them to achieve optimal performance
in exploration and victim identification.
In [61], Cipolleschi et al. proposed a system that exploits semantic information to push
robots to explore areas that are relevant; according to a priori information provided by human
users. That semantic information, which embedded in a semantic map, associates spatial
concepts (like rooms and corridors) with metric entities, showing its effectiveness to improve
total explored area.
Two main issues of multi-robot exploration are the exploration strategy employed to select the most convenient observation locations the robots should reach in a partially known
environment and the coordination method employed to manage the interferences between the
actions performed by robots [62]. To determine the effect of each issues, Amigoni et al. [62]
studied a search and rescue setting in which different coordination methods and exploration
strategies are implemented and their contributions to an efficient exploration of indoor environments are comparatively evaluated. Their results showed that the role of exploration strategies
dominates that of coordination methods in determining the performance of an exploring multirobot system in a highly structured indoor environment, while the situation is reversed in a
less structured indoor environment.
To our knowledge, there is not a lot of publications about the grid-based search by multirobot systems. A preliminary related work on this topic was undertaken by Spires et al. [63]
wherein they proposed space filling curves such as Hilbert curves for geographical search for
targets by multiple robots. Enns et al. [64] addressed the problem of searching a wide area
for targets using multiple autonomous air vehicles. They considered a scenario where viewing
from two perpendicular directions is needed to confirm a target. They employed a simple rule
wherein the unmanned air vehicles move in lanes to detect the targets.
In recent years, there has been an increasing amount of literature on using the Voronoi
diagram for multi-robot systems [65–67]. Wurm et al. [36] addressed the problem of exploring an
unknown environment with a team of mobile robots. They proposed an approach to distribute
the robots over the environment by partitioning the space into segments using Voronoi diagram.
Haumann et al. [68] proposed a frontier based approach for multi-robot exploration wherein
using a Voronoi partition of the environment, each robot autonomously creates and optimizes
the objective function to obtain a collision-free path in a distributed fashion. In [69], authors
used Voronoi decomposition of the map to build a connected graph of a place for generating
the set of exploration goals. Bhattacharya et al. [70] presented a distributed algorithm that
computes the generalized Voronoi tessellation of non-convex environments in real-time for use
in feedback control laws for cooperative coverage control in unknown non-convex environments.
Yang et al. [71] proposed a decentralized control algorithm of swarm robots for target search
and trapping inspired by bacteria chemotaxis; by dividing the target area into Voronoi cells.
Guruprasad and Ghose [72–74] studied employing Voronoi diagram for multi-robot deploy and
search.
However, such approaches that use Voronoi diagram for multi-robot search have some weakness. For example, they need a lot of resources for computation to obtain a reasonable performance that depends on the number of robots. Also, they are inefficient and quiet complex to
5
implement in practice [29]. This study makes a major contribution to research on multi-robot
systems by demonstrating some new methods of grid-based search to fill such gaps. In particular, our proposed methods do not need much computational resources, and are feasible to
implement in real systems.
Regarding the shape of search area, there are a lot of research conducted in multi-robot
systems with various assumptions about the search area [52, 73, 75, 76]. However, few of them
propose a comprehensive solution for the shape of the environment and obstacles. For instance,
in [52] a structured environment has been assumed as the search area whereas an unstructured
environment has been considered in [75]. In [73], the search area has been assumed a region without any obstacles while in [76], the operation has been done in an environment with
obstacles.
1.3
Communication and Coordination
It is apparent that the strength of multi-robot systems rises when the robots cooperate to
complete a task. Communication and coordination, as a means of interaction among robots, is
essential to the accomplishment of teamwork effectively [77]. The primary objective of coordination is to have a team of autonomous robots work together efficiently to achieve a collective
team behavior through local interaction. Recent advances in science and technology in the
fields of microelectronics, computing, communication and control systems have made it feasible
to deploy a large number of autonomous robots to work cooperatively to accomplish civilian
and military missions with the capability to significantly improve the operational effectiveness,
reduce the costs, and provide additional degrees of redundancy. Coverage control, flocking,
formation control and consensus are some of the problems in cooperative control of multi-robot
systems [78].
Multi-robot coordination has been primarily inspired from nature as there are various coordination behaviours in animals. For instance, flocking in a school of fish to avoid predators
and obstacles, flocking in a group of migrating birds, or a swarm of insects looking for food
are some famous examples of coordinating behaviour of animals [79]. A seminal study in this
area is the work of Reynolds in the 1980s, wherein the flocking behaviour of animals has been
simulated [80], which shows that a mathematical study of the animals flocking can be used as
a framework for cooperative control of multi-agent systems.
Today we have a large number of industrial systems which are good examples of multi-agent
coordination. An example is the deployment of sensors in an unknown environment [81] or
clock synchronization [82] in wireless sensor networks. Formation flying of satellites is another
example, which can be employed, e.g., for space interferometers and military surveillance [83,
84]. Multi-robot teams competitions, e.g., Robocup, are further useful samples of cooperation
control of multi-agent systems [85].
Fig. 1.1 shows the configuration of multi-robot systems [78]. As depicted in this figure,
there are three main components in multi-robot systems, which are: agent dynamics, interagent interactions and cooperative control laws. Describing agents dynamics depends on the
complexity of the agents, and in most cases, a first-order or second-order dynamics is used to
model the agents. In some circumstances, the agents are described by nonlinear models, and
some uncertainties or disturbances are taken into account.
Note that the agents in a multi-agent system are dynamically decoupled; therefore, they
must be coupled through another way enabling them to cooperate. Exchanging information
6
Agent Dynamics
Cooperative Control Laws
Inter-Agent Interactions
Figure 1.1: Configuration of multi-agent systems
through a communication network is a method by which the agents can interact with each
other. The information which is exchanged among the agents can be their odometry data, the
data of the environment they explore, e.g., maps, and the information about the targets they
detect. Direct sensing is another way by which the agents can communicate, for example, a
robot can detect the other robots’ positions by its sensors such as sonars, lasers or cameras. It is
very beneficial to represent the topology of the communication between the robots themselves
and with a center in a multi-robot system by a graph [86]. To put it simply, the agents of
a multi-robot system are equal to the nodes in the related graph, and an edge between two
nodes in the graph implies that there is a communication between two related robots. As the
communication between robots can be one-way or two-way, the related graph is directed or
undirected. Furthermore, if the robots are always connected via a communication network, the
related graph is static meaning that its edges are not time varying. On the other hand, due to
the limited communication range of the robots, they sometimes get disconnected. Thus, the
corresponding graph will be dynamic, i.e., the edges of the graph vary over time.
In multi-robot systems, cooperative control laws are applied via a control center or distributed among all agents; in agreement with the agent dynamics and the interaction topology.
Therefore, coordination control in a multi-robot system is classified into two main categories:
centralized and decentralized which will be discussed in the next section. Also, we describe
some related problems of cooperative control such as formation control, consensus and flocking
later in this chapter. For more detail on these topics and issues of multi-robot cooperation
control, see [87–91] and the references therein.
Networked control systems are another increasingly important area in multi-robot systems [92]. It has already been common to consider the control and communication as two
separate parts in a system. Usually in control theory, we consider that dynamical systems are
connected via ideal channels, whereas communication theory studies data transmission over
non-ideal channels. Actually, networked control systems are the combination of these two
theories. Since the control signals in multi-robot systems are transmitted through imperfect
communication channels with a limited capacity, it is more practical to study them as networked control systems [93–97]. In this report, we consider a multi-robot system as a network,
and a decentralized coordination control is applied by which the robots share their information
via wireless communication; therefore, our multi-robot system is an example of a networked
7
control system with various communication limitations.
1.4
Centralized vs. Decentralized
For a cooperative control problem, designing an appropriate controller to carry out the desired
team goal is the primary task [89]. Centralized approach and decentralized (or distributed)
approach are two major ways to control multi-robot systems [98–101]. In centralized control
systems , there is no communication between robots at all, and the robots are controlled by a
central unit which is responsible for the team coordination. Although multi-robot centralized
systems are easier to implement and apply, decentralized control has important advantages so
that it has recently brought researchers to pay more attention to it. Also in some applications,
it is hard or even impossible to use a centralized control; then, using a decentralized system
is inevitable. Several constraints are imposed by a centralized control such as limited wireless
communication range, the short sensing capability of sensors, high cost, limited resources and
energy, and large size of vehicles to manage and control or even infeasible to implement [89].
In decentralized control, there is not any control center; thus, the algorithm is such that
the control procedure is distributed among the agents. In other words, all the robots in a
decentralized control system are autonomous robots which are controlled by relatively simple
control rules [102, 103]. Thus, decentralized control only relies on local information of a robot
itself and the information that it receives from the other robots of the team; usually the
information from its neighbours. The decentralized control approach has many advantages
in achieving cooperative team performances, specifically with low operational costs, less system
requirements, high robustness, strong adaptivity, and flexible scalability; therefore, has been
widely recognized and appreciated [89].
In decentralized control of multi-robot systems, the primary goal typically is to have the
whole team working in a cooperative way throughout a distributed law. Here, cooperative
means an interaction among all the robots in the team via sharing their information. Therefore,
one of the main issues in decentralized control systems is how the robots share their information
among themselves. When the robots of a team collaborate to achieve a goal based on an
algorithm, they need to share their information. That information can be their motion’s data
like position and velocity, the maps they build during the search procedure or the information
about the targets they detect. Moreover, in multi-robot systems, a control algorithm must
consider some practical limitations such as memory usage [104], processing speed [105] and
communication bandwidth [106]. One shortcoming of decentralized control is that some robots
cannot predict the team behavior based only on the available local information from their
neighbours; so a team behavior cannot be controlled. In addition, in a complicated task where
the goal is to optimize a global cost, the decentralized control may suffer from the lack of
competence. Therefore, it would be interesting to assess a balance between decentralization
and centralization so as to achieve further improvement in overall performance.
1.5
Consensus Variables Rule
Consensus problem in networks is a cooperative behavior used in many fields of studies such
as sensor networks [107–109], mobile communication systems, unmanned vehicles, distributed
computing as well as multi-robot systems [110]. The first serious discussions and analysis of
8
consensus problems started with distributed computing in computer science [111, 112]. Vicsek
presented a model for studying collective behavior of interacting self-propelled particles, which
is a basis for modeling of many animals behaviours such as a school of fish, flock of birds
and swarm of insects [113, 114]. Later, a mathematical justification for the Vicsek model was
proposed in [115], and a general framework of the consensus problem for networks of integrators
was provided in [116]. There is a large volume of published studies describing the role of
consensus problems; see, e.g., [89, 90, 117, 118] and the references therein.
In the case of multi-robot systems, the consensus is an important and fundamental problem
that means reaching a consensus on an amount for one or more variables. For example, the
robots communicate with each other to agree on a particular value for their speed and direction. They share their information and also exchange their calculations based on a rule called
consensus algorithm [119]. Consensus algorithms are widely used in multi-robot systems such
as formation control [120], flocking [121], and coverage control [122, 123].
In this report, we apply an algorithm based on consensus variables, by which the robots
eventually reach an agreement on a common triangular grid. Hence, they can be located on
some vertices of that grid and begin the search operation by moving through the vertices of the
grid. Furthermore, in another algorithm proposed for formation building, a control rule based
on the consensus variables rule is employed.
1.6
Mapping
Mapping and map merging are two other challenging problems in multi-robot systems. In
single robot systems, a robot makes its map and utilizes it only by itself of course since there
is nobody else to share with. On the other hand, in multi-robot systems, each robot makes its
map which is used by the other robots, getting better performance in an operation. Therefore,
map merging is essential for multi-robot systems; however, it is a challenging issue due to the
different coordinate systems of the robots. Researchers proposed several methods for mapping
and map merging in multi-robot systems. One solution is based on a global or local positioning
system [124] which is not available in many conditions. Another approach assumes that all
robots know their positions in a shared map [17]. There are also methods that merge maps of
different robots to build a shared map of the environment [125, 126].
Geometric grid based [127] and topological maps [128] are two main methods for map
making of mobile robots. Although geometric grid based method is more accurate for the case
where the grid size is small enough, it needs considerable memory and takes a remarkable time
to process; especially for map merging in multi-robot systems. Thus, in multi-robot systems,
topological maps are more efficient and feasible. In this report, we use a kind of topological
map, described in detail in Chapter 2.
1.7
Formation Building
In the field of multi-robot systems, formation control has been the subject of many classic
studies in the past two decades. The purpose of formation control is to drive a group of agents
to some desired states; for example, to build a geometric pattern. There are some multi-robot
applications such as remote sensing, patrolling, search and rescue wherein it is helpful if the
robots move with a desired geometric formation. Therefore, the research about distribution
9
formation control of a team of mobile robots has become a new challenging area for researchers
in recent years; see, e.g., [129–133] and the references therein. The notable difference between
general distributed control approach and the approaches that are used for a team of mobile
robots is that in the second case, there is no dynamic coupling among the robots; meaning that
the robots do not directly affect each other. Based on the distributed control approach, each
robot uses the information provided by its nearest neighbouring robots in order to update its
linear and angular velocities at discrete time instants. There are some articles on this topic
that proposed a distributed control algorithm by which the robots will eventually move with
the same heading and speed; see, e.g., [134–136]. The more challenging problem is to apply a
distributed control algorithm to force the robots to move so that they finally build a desired
geometric pattern. Furthermore, formation building in the existence of obstacles is even a more
difficult problem.
Many of the articles presented in this area consider a simple linear model for the motion
of the robots without constraints on the control inputs; see, e.g., [137–139]. In particular,
these simple models do not consider the essential standard constraints on the angular and
linear velocities. Indeed, all actual vehicles such as Unmanned ground vehicles (UGVs) and
unmanned aerial vehicles (UAVs) have standard hard constraints on their angular and linear
velocities [140]. For example, with no constraints on the angular velocity, it may become a
large value which results in a small turning radius for the robots that is impossible to obtain
with actual robots. Therefore, the linear system approaches considered without constraints
on the control inputs are not applicable. In [106], an algorithm of flocking for a group of
wheeled robots described by the unicycle model with hard constraints on angular and linear
velocities was proposed; however, the much more challenging problem of formation building was
not considered. We consider a more difficult problem of formation building where a nonlinear
model with hard constraints on the angular and linear velocities describing the robots.
Communication between the robots is another issue that was considered in many of the
papers in this area. However, most of such publications consider leader-follower in the team;
therefore, they must consider quite restrictive classes of robot communication graphs; see,
e.g., [141, 142]. In some other papers, robots communication graph is assumed to be minimally
rigid [143, 144] or time-invariant and connected [145] which is also quite restrictive.
The existence of obstacles in the environment is the other problem that has been considered
in a lot of recent research on mobile robots control methods; see, e.g., [44, 146–152] and the
references therein. The problem of existence of obstacles in the environment is not considered
in most references mentioned above for formation building methods [153]. However, some
papers assume obstacles in the environment and include an obstacle avoidance method in their
proposed algorithms; see, e.g., [154–156] and the references therein. In [154], it was assumed
that the shape of the obstacle is convex and known to the robots. Obstacle avoidance strategy
in [155] is based on the concept of impedance with fictitious forces. The technique for obstacle
avoidance of mobile robots in [156] relies on a rotational potential field.
In this report, we consider the problem of distributed control of a team of autonomous
mobile robots in which the robots finally move with the same direction and speed in a desired
geometric pattern while avoiding the obstacles. We propose a distributed motion coordination
control algorithm so that the robots collectively move in a desired geometric pattern from
any initial position while avoiding the obstacles on their routes. Furthermore, we present
the algorithm of formation building with anonymous robots meaning that the robots do not
know their final position in the desired geometric configuration at the beginning; but, using
a randomized algorithm, they will eventually reach a consensus on their positions. Also, we
10
propose a new obstacle avoidance method by which the robots maintain a given distance to
the obstacles as well as mathematical justification of the proposed method. There are various
applications for the suggested formation control algorithm such as sweep coverage [100,110,157],
border patrolling [158], mine sweeping [159], ocean floor monitoring [160], and exploration in a
sea floor [161].
1.8
Contributions of This Report
In this report, we propose some new methods for distributed control of a multi-robot team
which its duty is to search all or a part of an unknown region. The aim of the search can be
either finding some targets in the region, patrolling the region frequently or putting some signs
on particular points in the region. The suggested algorithms use a triangular grid pattern, i.e.,
robots certainly go through the vertices of a triangular grid during the search operation. To
ensure that all the vertices of the region’s covering grid are visited by the robots, they must have
a common triangular grid. In order to achieve that, we use a two-stage algorithm. In the first
stage, robots apply an algorithm according to the consensus variables rule to deploy themselves
on the vertices of a common triangular grid which covers the region. In the second stage, they
begin searching the area by moving between the vertices of the common triangular grid. There
are various scenarios can be used in the second stage. We propose three different methods
for the second stage, namely, random triangular grid-based search algorithm, semi-random
triangular grid-based search algorithm and modified triangular-grid-based search algorithm,
which are presented in Chapters 3, 4 and 5, respectively.
The proposed algorithms are partly based on ideas from [162] where a distributed random
algorithm for self-deployment of a network of mobile sensors has been presented. In [162],
authors have proposed a method to deploy a team of mobile sensors on the vertices of a
triangular grid that covers the entire area of interest. Indeed, in the proposed search algorithms,
we benefit from the triangular grid coverage to cover the entire area. It is clear that any
triangular grid coverage of a region is a complete blanket coverage of that region. The main
advantage of deploying robots in a triangular grid pattern is that it is asymptotically optimal
regarding the minimum number of robots required for the complete coverage of an arbitrary
bounded area [163]. Therefore, using the vertices of this triangular grid coverage, which is
applied in this report, guarantees search of the whole region.
One of the advantages of the proposed algorithms is the method by which the robots make
and share their maps and use them for exploration. We use a kind of topological map, described
in detail in the next chapters. Furthermore, we consider a region with an arbitrary shape that
contains some obstacles. Also we assume the area is unknown to the robots a priori. In
addition, mathematically rigorous proofs of convergence with probability 1 of the algorithms
are given. Moreover, our algorithms are implemented and tested using Mobilesim, a simulator
of the real robots and environment. Mobilesim is a powerful simulator which considers robots’
real circumstances such as dynamics of motion, encoders, sonar and also borders and obstacles
of the environment. We also test one of the algorithms via experiments by real robots. In fact,
we confirm the performance of the proposed algorithm with experiments with Adept Pioneer
3DX wheeled mobile robots in a real world environment.
A further study on networked multi-robot formation building algorithm is presented in this
report. We consider the problem of distributed control of a team of autonomous mobile robots
in which the robots finally move with the same direction and speed in a desired geometric
11
pattern while avoiding the obstacles. We propose a distributed motion coordination control
algorithm so that the robots collectively move in a desired geometric pattern from any initial
position while avoiding the obstacles on their routes. In the proposed method, the robots
have no information on the shape and position of the obstacles and only use range sensors
to obtain the information. We use standard kinematic equations for the robots with hard
constraints on the linear and angular velocities. There is no leader in the team, and the robots
apply a distributed control algorithm based on the local information they obtain from their
nearest neighbours. We take the advantage of using the consensus variables approach that
is a known rule in multi-agent systems. Also, an obstacle avoidance technique based on the
information from the range sensors is used. Indeed, we propose a new obstacle avoidance
method by which the robots maintain a given distance to the obstacles. We consider quite
general class of robot communication graphs which are not assumed to be time-invariant or
always connected. Furthermore, we present the algorithm of formation building with obstacle
avoidance with anonymous robots meaning that the robots do not know their final position
in the desired geometric configuration at the beginning; but, using a randomized algorithm,
they eventually reach a consensus on their positions. Mathematically rigorous proofs of the
proposed control algorithms are given, and the effectiveness of the algorithms are illustrated
via computer simulations.
1.8.1
Main Contribution Highlights
The main contributions of this report are summarized as follows:
• Three novel algorithms are proposed for search with multi-robot systems. The detailed
descriptions of the proposed algorithms are presented.
• The suggested algorithms use a triangular grid pattern, i.e., robots certainly go through
the vertices of a triangular grid during the search procedure. The main advantage of
using a triangular grid pattern is that it is asymptotically optimal in terms of the minimum number of robots required for the complete coverage of an arbitrary bounded area.
Therefore, using the vertices of this triangular grid coverage, what is applied in this report, guarantees complete search of all the region as well as better performance in terms
of search time.
• We use a new kind of topological map which robots make and share during the search
operation.
• We consider a region with an arbitrary shape that contains some obstacles; also, we
assume the area is unknown to the robots a priori.
• Unlike many existing heuristic methods, we give mathematically rigorous proofs of convergence with probability 1 of the proposed algorithms.
• We present an extensive simulation study for the proposed algorithms using a powerful
simulator of real robots and environment. The results confirm the effectiveness and
applicability of the proposed algorithms.
12
• To evaluate the performance of the algorithms, the experiment results with real Pioneer
3DX mobile robots are presented for one of the search algorithms with detailed descriptions and explanations. The results demonstrate the features of the proposed algorithms
and their performance with real systems.
• We compare the proposed search algorithms with each other and also with three algorithms from other researchers. The comparison shows the strength of our algorithms over
the other existing algorithms.
• The problem of formation building for a group of mobile robots is considered. A robust decentralized formation building with obstacle avoidance algorithm for a group of
mobile robots to move in a defined geometric configuration is proposed. Furthermore,
we consider a more complicated formation problem with a group of anonymous robots
where the robots are not aware of their position in the final configuration, and have to
reach a consensus during the formation process. We propose a randomized algorithm for
the anonymous robots which achieves the convergence to the desired configuration with
probability 1.
• We give mathematically rigorous proofs of convergence of the proposed algorithms for
formation building and also computer simulation results to confirm the performance and
applicability of our proposed solution.
• A novel obstacle avoidance rule is proposed which is used in the formation building
algorithms.
1.9
Report Outline
The remainder of this report is organised as follows:
In Chapter 2, we describe the problem statement, defining the search problem in details and
shedding some lights on terms, assumptions and definitions used in this report. Then, the first
stage of the proposed search algorithms is presented that is the consensus variables locating
algorithm by which the robots will be located on the vertices of a common triangular grid. A
mathematically proof for the presented algorithm is given as well as some simulation results.
Whenever all the robots are located on the vertices of a common triangular grid, the search
operation starts by moving the robots between the vertices of the common triangular grid. The
first method of search that we propose is the random triangular grid-based search algorithm
presented in Chapter 3. Using this algorithm, the robots randomly move through the vertices
of the common triangular grid during the search operation. Therefore, a complete search of
the whole area is guaranteed. We give a mathematically rigorous proof of convergence of the
presented algorithm as well as the computer simulation results to demonstrate that the algorithm is effective and practicable. In a similar way, in Chapter 4, the semi-random triangular
grid-based search algorithm is presented along with a mathematically rigorous proof of the algorithm and computer simulation results. In Chapter 5, the modified triangular grid-based search
algorithm is given, and a mathematically rigorous proof of the algorithm along with computer
simulation results are presented. In addition, the experiment results with Pioneer 3DX wheeled
mobile robots are presented to confirm the performance of our suggested algorithm. Finally, a
comparison between the proposed triangular grid-based search algorithms in Chapters 3, 4 and
13
5 against three other algorithms are given. Chapter 6, describes a distributed motion coordination control algorithm so that the robots collectively move in a desired geometric pattern from
any initial position while avoiding the obstacles on their way. Also, a randomized algorithm
for the anonymous robots which achieves the convergence to the desired configuration is presented. Mathematically rigorous proofs of the proposed control algorithms are given, and the
effectiveness of the algorithms are confirmed via computer simulations. Chapter 7 summarizes
the work that is presented and discusses possible future research projects as an extension of
the presented work.
14
Chapter 2
Consensus Variables Locating
Algorithm
Consensus problem in networks is a cooperative behavior used in many fields of studies such as
sensor networks, mobile communication systems, unmanned vehicles, distributed computing as
well as multi-robot systems. In the case of multi-robot systems, it means reaching a consensus
on an amount for one or more variables. For example, the robots communicate with each other
to agree on a specific value for their speed and direction. They share their information and also
exchange their calculations based on a rule called consensus algorithm [119].
In the next three chapters, we present some search algorithms regarding that an area is
explored by a team of mobile robots to explore the entire area or to find some targets. The
presented algorithms are grid-based search algorithms meaning that the robots certainly pass
through the vertices of a grid; a triangular grid in our algorithms. Therefore, for the first step,
all the robots must be located on a common triangular grid to start the exploring operation.
To accomplish that, we apply an algorithm based on consensus variables, by which the robots
eventually reach an agreement on a common triangular grid. Hence, they can be located on
some vertices of that grid and begin the search operation by moving through the vertices of the
grid.
In this chapter, the problem statement is described firstly, which defines the problem of
search in details. We also give some terms, assumptions and definitions used in this chapter and
Chapters 3, 4 and 5. Then, the first stage of the proposed search algorithms is presented that
is ”consensus variables locating algorithm” by which the robots will be located on the vertices
of a common triangular grid. A mathematically rigorous proof of the presented algorithm is
given as well as some simulation results.
2.1
Problem statement
Consider a planar and bounded area R. Also, consider a few number of obstacles O1 ,O2 ,. . . Om
inside the area R (see Fig. 2.1). The goal is to search the whole area by a few autonomous
mobile robots in order to find some targets. The number of targets may be known or unknown
to the robots. We use a distributed algorithm to drive the robots inside the search area as well
as avoiding the obstacles and borders. The algorithm is such that the robots follow a pattern to
search. The proposed pattern is a triangular grid so that the robots search the area by moving
through the vertices of that triangular grid. The grid consists of equilateral triangles with sides
15
Figure 2.1: A triangular grid (dotted area) covers the search area
r. Fig. 2.2 shows this triangular grid that covers the searching area.
It is assumed that the robots are equipped with sensors to detect the targets, and these
sensors have a circular sensing area with radius rs . As shown in Fig. 2.2, the small circles of
radius rs with centers located on the vertices of the triangular grid are the sensing ranges of
the robots. It means that robot i can gather information of or detect a target if it is located in
a disk of center pi (k) with radius rs defined by Di,rs (k) := {p ∈ R2 : kp − pi (k)k ≤ rs }, where k
indicates discrete time instances; k = 0, 1, 2, . . ., pi (k) ∈ R2 denotes the Cartesian coordinates
of robot i at time k and k · k denotes the Euclidean norm. As demonstrated in Fig. 2.2,
to have an optimal search operation, the common areas of the robots’
sensing regions (small
√
circles) must be minimum. To achieve this goal, we assume r = 3rs . However, this is for an
ideal and optimal case but in practice where there is sensor
√ noise and/or position uncertainty,
the triangles sides r can be chosen a little smaller than 3rs that reduces the possibility of
unexplored areas.
One of the main benefits of multi-robot systems is that robots share their information in
order to improve the strength of the search algorithm in terms of time and cost. The information
includes the position of robots, maps, explored areas and detected targets. We assume that the
robots share their information via a wireless communication so that a robot always sends new
information it obtains to the other robots and also listens to receive new information from them.
Due to the limited communication range of the robots, we assume rc as the communication
range that is the same for all the mobile robots. It means that a robot can only receive
information from the robots which are located not further than rc . Therefore, the range of
communication for a robot can be defined as the disk of Di,rc (k) := {p ∈ R2 : kp − pi (k)k ≤ rc }.
In Fig. 2.2, the big circle of radius rc is the communication range of robot A, meaning that
robot A can communicate with robot B but not with robot C. The multi-robot system under
consideration is an example of networked control systems in which coordinates and heading of
neighbouring robots can be estimated from distance based measurements using robust Kalman
state estimation and model validation techniques via limited communication with each robots
neighbours; see, e.g., [164–167].
Definition 2.1.1 Robot j is a neighbour of robot i at time k if it is located on the disk Di,rc (k).
So, N i (k) = {j : pj ∈ Di,rc (k), j ∈ {1, 2, . . . , n}, j 6= i} is the set of all neighbours of robot i at
time k. Also, |N i (k)| denotes the number of its neighbours.
16
Communication range
Sensing range
Triangular grid
Node
A
B
C
Figure 2.2: Robots on vertices of a triangular grid; small circles are robots’ sensing ranges and
big circle is the communication range of robot A
It is essential for the robots in any search operation to make the map of the searching
area when it has not been available already to the robots. That helps the robots to keep the
information of the search operation in their memory for future use or to send it to a center
or to the other robots of the search team. The problem of mapping is a challenging problem
especially when there is a team of robots instead of only one robot. The main problem is
merging the maps which are made by different robots. It is much easier for robots to make
maps and merge them in the case that there is a central control station or when robots can use a
central positioning system like GPS. On the other hand, without a central positioning system,
robots have to make their map themselves in their own coordinate systems. Furthermore,
the information of maps needs lots of memory also processing of maps for merging is a time
consuming task. In such cases , i.e., distributed systems, it is useful to employ topological maps
instead of grid geometry maps that need a considerable amount of memory.
In this report, we take advantage of using a kind of topological map that robots make and
share among themselves. It is also assumed that the searching area is unknown to all the robots
a priori. Therefore, robots have to make the map of the area during the search operation.
To define the problem, we utilize definitions of [122]. Also, to clarify some terms in the rest
of this report, and to state our theoretical results, we introduce a number of assumptions and
definitions.
Assumption 2.1.1 The area R is bounded and connected; also, the obstacles On are nonoverlapping, closed, bounded, and linearly connected sets for any n ≥ 0.
Definition 2.1.2 Let O := ∪On for all n ≥ 0; then, introduce W := { p ∈ R: p ∈
/ O}.
17
As mentioned before, the robots can detect borders of the area R and all the obstacles
therein. Accordingly, W is actually those points in the area R that the robots can go there
during the search operation. Also, our search procedure is so that robots go through vertices of
a triangular grid that covers this area and cuts the plane into equilateral triangles. A triangular
grid can be determined by its vertices; therefore, when we say a triangular grid, it means the
set of all the vertices of it. It is obvious that there exists an infinite number of triangular grids.
Definition 2.1.3 Consider T , one of the all possible triangle grids that covers the area R.
T̂ := T ∩ W is called a triangular grid set in W (see Fig. 2.1).
Simply, T̂ is the set of all vertices of the covering triangular grid which robots must visit during
the search operation. Since we use topological maps, T̂ is actually the map of W.
Definition 2.1.4 A set consisting of all the vertices of a triangular grid in an area is called a
map of that area.
In fact, what the robots actually save in their memory as maps, are the coordinates of the
vertices. A robot can also put some tags on these vertices to assign some attributes like visited,
unvisited, occupied, etc. The robots share their maps with their neighbours too.
The relationships between robots can be defined by an undirected graph G(k). We assume
that any robot of the multi-robot team is a node of the graph G(k) at time k, i.e., i in
VG = {1, 2, . . . , n} , the node set of G(k), is related to robot i. In addition, robot i is a
neighbour of robot j at time k if and only if there is an edge between the nodes i and j of
graph G(k) where i 6= j. Therefore, the problem of communication among the team of robots
equals the problem of the connectivity of the related graph. It is undeniable that it does not
need for robot i to be the neighbour of robot j to get the information from it. The information
can be transferred through the other robots which connect these robots in the related graph.
Fig. 2.2 shows this condition in which robot A cannot directly communicate with robot C but
can do it through robot B. To guarantee the connectivity of the graph, we accept the following
assumption [115].
Assumption 2.1.2 There exists an infinite sequence of contiguous, non-empty, bounded, timeintervals [kj , kj+1 ), j = 0, 1, 2, . . ., starting at k0 = 0, such that across each [kj , kj+1 ), the union
of the collection {G(k) : k ∈ [kj , kj+1 )} is a connected graph.
Since the main goal of the proposed algorithm is to search an area in such a way that
robots certainly go through the vertices of a triangular grid, firstly, robots have to be located
on some vertices of that grid. Thus, we divide our algorithm into two stages. In the first stage,
robots make a common triangular grid in order to be finally located on some vertices of it. To
achieve this goal, we apply consensus variables method which is a known distributed method
for multi-robot systems [117]. In the second stage, robots start to search the area based on
moving between the vertices of the created grid which is common among all the members of the
team. In this chapter, we introduce the first stage of our search algorithms named consensus
variables locating algorithm. Then, we propose three different algorithms for the second stage
of our algorithms in Chapters 3, 4 and 5.
18
2.2
Consensus Variables Locating Algorithm
We propose some two-stage search algorithms so that in the first stage, using a consensus
variables rule, robots make a triangular grid that is common among all the members of the
team [122]. In the beginning, robots are located anywhere in the area W. Each robot can
assign its position and heading angle respect to its own coordinate system. The center of a
robot at the starting point is assumed to be the origin of its coordinate system, and its heading
vector as the x-axis. To define a unique triangular grid with equilateral triangles in a plane,
we only need a point and an angle. Thus, the point q which is any vertex of the grid together
with the angle θ that is the angle of the grid, uniquely defines the triangular grid T̂ [q, θ] that
covers the area W (see Fig. 2.1). We consider that at the beginning, the triangular grid that a
robot makes for itself at each vertex, is based on the position and heading of the robot that are
q and θ, respectively. Accordingly, each robot has its own grid that is different from the other
robots’ grids. To combine these different grids to a unique grid which will be common among
all robots, we apply the consensus variables approach.
We assume that at any time k, robot i has two consensus variables; qi (k) and θi (k) on which
it builds its triangular grid T̂ [qi (k), θi (k)]. At first, these consensus variables are not the same
for different robots, so their triangular grids are not the same as well. Using the proposed
algorithm will bring the consensus variables qi (k) and θi (k) from different values of qi (0) and
θi (0) to the same values of q0 and θ0 for all the robots. That is, a common triangular grid for
all the members of the team is built based on q0 and θ0 which are the same for all.
Assumption 2.2.1 The initial values of the consensus variables θi satisfy θi (0) ∈ [0, π) for
all i = 1, 2, . . . , n.
Definition 2.2.1 Consider p be a point on the plane; also, q and θ are a vertex and an angle
that build the triangular grid T̂ [q, θ]; then, C[q, θ](p) will be the closest vertex of T̂ [q, θ] to p.
If there is more than one vertex, any of them can be chosen.
Now, we propose the following rules as the consensus variables locating algorithm:
θi (k + 1) =
θi (k) + Σj∈Ni (k) θj (k)
;
1 + |Ni (k)|
qi (k + 1) =
qi (k) + Σj∈Ni (k) qj (k)
1 + |Ni (k)|
(2.1)
pi (k + 1) = C[qi (k), θi (k)](pi (k))
(2.2)
The rule (2.1) causes that the robots reach consensus on heading using θi and gain consensus
on phase shift using qi . Based on q and θ calculated by each robot using rule (2.1), a robot
makes a triangular grid for itself, which is used in rule (2.2). Fig. 2.3 illustrates the rule (2.1).
Suppose that there are four robots in the environment. At time k, robots R2 and R3 are
located in the communication range of robot R1 but robot R4 is out of the range. Therefore,
only robots R2 and R3 are the neighbours of robot R1. As a result, robot R1 updates variable
θ1 using the information from R2 and R3, i.e., θ1 (k + 1) will be the average of θ1 , θ2 and θ3 at
time k (see Fig. 2.3(a)). In a similar way, the variable q is updated as shown in Fig. 2.3(b).
Rule (2.2) means that whenever a robot makes its triangular grid, it will move to the nearest
vertex on it. Fig. 2.4 demonstrates this stage in which robot i located at pi (k) at time k, makes
19
3(k)
1(k 1)
R3
1 ( k )
R1
rc
4(k)
2(k)
R2
R4
(a)
q3 (k )
rc
R3
R1
R1
q1 (k )
q1 (k 1)
q4 (k )
q2 (k )
R4
R2
(b)
Figure 2.3: Updating consensus variables θ and q using the information from the neighbours
20
the triangular grid T̂ [q, θ] using θi (k) and qi (k). Then, it moves to the nearest vertex of the
grid, which will be the position of robot i at time k + 1, i.e., pi (k + 1). Since rule (2.1) brings
the same q and θ for all the robots, they eventually will build a common triangular grid, and
they all will be located on its vertices.
Tˆ[q i , i ]
p (k)
i
p (k 1)
i
q (k )
i
i (k )
21.0°
Figure 2.4: A robot moves to the nearest vertex of its triangular grid
Remark 2.2.1 The robots initially do not have a common coordinate system; otherwise, the
consensus building problem would be trivial. Therefore, each robot has consensus variables θi (k)
and qi (k) in its own coordinate system. However, each robot knows the bearing and the distance
to each of its neighbouring robots. Using this information, at any time instance k, robot i sends
to a neighbouring robot j the consensus variables θi (k) and qi (k) re-calculated in the coordinate
system with the line (pi , pj ) as the x-axis, pj as the origin, and the angle θi (k) is measured from
this axis in counter-clockwise direction. Using this information, each robot can re-calculate the
sums (2.1) at each time step in its own coordinate system [122].
Theorem 2.2.1 Suppose that Assumptions 2.1.1, 2.1.2 and 2.2.1 hold, and the mobile robots
move according to the distributed control law (2.1), (2.2). Then, there exists a triangular grid
set T̂ such that for any i = 1, 2, . . . , n, there exists a τ ∈ T̂ such that lim pj (k) = τ .
k→∞
Proof of Theorem 2.2.1: Assumption 2.1.1 and the update rule (2.1) guarantee that
there exist a θ0 and q0 such that
θi (k) → θ0 , qi (k) → q0 ∀i = 1, 2, . . . , n
(2.3)
(see [115]). Furthermore, the update rule (2.2) guarantees that pi (k + 1) ∈ T̂ [qi (k), θi (k)].
Therefore, this and (2.3) guarantee that lim pi (k) = τ where τ ∈ T̂ [q0 , θ0 ]. This completes
k→∞
the proof of Theorem 2.2.1.
21
2.3
Simulation Results
To verify the suggested algorithm, computer simulations are employed. The region W is considered to be searched by a few robots (see Fig. 2.1). We suppose a multi-robot team with three
robots which are randomly located in the region W with random initial values of angles. The
goal is locating the robots on the vertices of a triangular grid by applying algorithm (2.1),(2.2).
Table 2.1: Simulation Parameters
Linear speed (default)
Linear speed (Maximum)
Angular speed (default)
Angular speed (Maximum)
Linear acceleration
Angular acceleration
Linear deceleration
Angular deceleration
0.4
0.4
1.3
1.74
0.3
1.74
0.3
1.74
Localization method
Localization origin
Odometry
[0 0 0]
Odometry error
Number of sonars
Minimum view
Maximum view
Field of view
Noise
Position (x, y θ)
m/s
m/s
radian/s
radian/s
m/s2
radian/s2
m/s2
radian/s2
Localization
x, y, θ
Slip in x, y and θ (Uniform random distribution),
[ 0.0075 0.0075 0.0075 ]
proportional to velocity
Sonar
16
0.1
meter
5
meter
30
degree
0.0005
meter
0.069, 0.136, 90
0.114, 0.119, 50
0.148, 0.078, 30
0.166, 0.027, 10
0.166, -0.027, -10
0.148, -0.078, -30
0.114, -0.119, -50
0.069, -0.136, -90
m,m, degree
-0.157, -0.136, -90
-0.203 -0.119 -130
-0.237, -0.078, -150
-0.255, -0.027, -170
-0.255, 0.027, 170
-0.237, 0.078, 150
-0.203, 0.119, 130
-0.157, 0.136, 90
To simulate the algorithm, MobileSime, a simulator of mobile robots developed by Adept
MobileRobots, is used. We also use Visual C++ for programming and ARIA, a C++ library
that provides an interface and framework for controlling the robots. In addition, Pioneer 3DX
is selected as type of the robots, and simulation parameters are given in Table 2.1. To prevent
collisions between the robots and to avoid the obstacles and borders, an obstacle avoidance
algorithm is applied using functions provided in ARIA library. Furthermore, to avoid hitting
and sticking to the borders, we assume a margin near the borders such that the robots do not
pass it.
The simulation results are shown in Fig. 2.5. Fig. 2.5(a) displays the initial position of the
22
(a)
(b)
Figure 2.5: Robots’ locations; (a) Initial locations of robots, (b) After applying the consensus
variables locating algorithm
robots at k = 0, which are randomly distributed in the area. Applying algorithm (2.1),(2.2)
will result in locating the robots on the vertices of a common triangular grid. Routes of the
robots and their final position depicted in Fig. 2.5(b) shows that they are eventually located
at the desired places; on the vertices of a common triangular grid, at time 1m47s.
2.4
Summary
In this chapter, the problem of searching an area by a team of mobile robots using grid-based
algorithms has been presented. Some terms, definitions and assumptions have been defined
that are used in the proposed algorithms. It has been confirmed that for a grid-based search,
the robots must be located on the vertices of a common grid, first. That is a consensus rule
named consensus variables locating algorithm has been employed as the first stage of the search
algorithms, which locates the robots on the vertices of a common grid among all the robots; a
triangular grid in our proposed algorithm. A mathematically rigorous proof of convergence of
the presented algorithm has been demonstrated. Furthermore, the computer simulation results
23
using MobileSim, a powerful simulator of real robots and environment, have been presented
to confirm that the algorithm is effective and practicable. In the next three chapters, three
grid-based search algorithms will be proposed. The locating algorithm presented in this chapter
will be employed as the first stage of those algorithms.
24
Chapter 3
Random Triangular Grid-Based Search
Algorithm
As described in Chapter 2, applying consensus variables locating algorithm as the first stage
of a search algorithm, locates all the robots on the vertices of a common triangular grid. After
that, the search operation is started by moving the robots between the vertices of the common
triangular grid. The first method that we apply for search is ”random triangular grid-based
search algorithm”. By this method, the robots randomly move between the vertices of the
common triangular grid so that in each step they only move to the one of the six neighbouring
vertices. A mathematically rigorous proof of convergence with probability 1 of the algorithm
is given. Moreover, our algorithm is implemented and simulated using a simulator of the real
robots and environment. The other methods will be presented in Chapters 4 and 5.
3.1
Distributed Random Search Algorithm
In order to search an area by a team of mobile robots using the vertices of a grid as the exploring
points, we need to locate the searching mobile robots on the vertices of a common grid among
the robots. Consensus variables locating algorithm (described in Chapter 2) can be the first
stage of the suggested search algorithm, which locates all the robots on the vertices of a common
triangular grid T̂ (see Fig. 2.1). The next step will be search of the area W based on moving
the robots between the vertices of the covering grid of the area. In this chapter, we propose
a random triangular grid-based search algorithm. Suppose a robot located on a vertex of the
common triangular grid. Consequently, it can explore the surrounding area using its sensors,
and that depends on the sensing range of its sensors. After exploring that area, the robot moves
to another point which can be one of the six neighbouring vertices in the triangular grid. As
the first method, we suppose that selecting the neighbouring vertex is random. In this regard,
there are a few scenarios can be considered to search the area. The first scenario is exploring the
whole area which can be applicable when the robots are searching for an undetermined number
of targets. Therefore, to detect all possible targets, the team of robots must search the whole
area. Patrolling of the area is the other application for this scenario where the robots should
move continuously to detect the possible intruders to the area. In the case of given number of
targets which is our second scenario, the search operation should be stopped whenever all the
targets are detected without searching the whole area.
25
3.1.1
Searching the Whole Area
To make sure that the whole area is explored by the team of the robots, each vertex in the
triangular covering grid set of the area W must be visited at least one time by a member of the
team. Consider T̂ is a triangular covering grid of W, and also each vertex of T̂ has been visited
at least one time by a robot of the team. This guarantees that the area W has completely been
explored by the multi-robot team. Since the robots do not have any map at the beginning,
they need to do map making during the search operation so that their maps will gradually be
completed.
Definition 3.1.1 Let S
T̂i (k) be the set of all the vertices of T̂ have been detected by robot i at
time k. Then, T̂ (k) = T̂i (k) will be the map of the area W detected by the team of the robots
until time k.
Note that a detected vertex is different from an explored vertex. These terms are defined in
detail in the following definitions.
Definition 3.1.2 A detected vertex means that vertex is detected by a robot using map making,
and it is in the map of that robot though it might be visited or not by the robots.
Definition 3.1.3 An explored vertex is a vertex that is visited by, at least, one member of the
team.
Definition 3.1.4 The map of robot i at time k, Mi (k), is the set of the vertices in T̂ detected
by robot i itself or received from other robots by which they are detected until time k.
Definition 3.1.5 Suppose a Boolean variable Vτ (k) which defines the state of vertex τ ∈ T̂ (k)
at time k. Vτ (k) = 1 if the vertex τ has already been visited by, at least, one of the robots,
otherwise Vτ (k) = 0.
Assumption 3.1.1 The triangular grid set T̂i (k); k = 0, 1, ... is a connected set. That means
that if τ ∈ T̂i (k), then, at least, one of the six nearest neighbours of τ also belongs to T̂i (k).
Let ℵ(pi (k)) be a set containing all the closest vertices to pi (k) on the triangular grid T̂ (k);
also, consider |ℵ(pi (k))| as the number of elements in ℵ(pi (k)). It is clear that 1 ≤ |ℵ(pi (k))| ≤ 6.
In addition, assume ν be a randomly opted element of ℵ(pi (k)).
Consider at time k robot i is located at point pi (k), and it wants to go to the next vertex.
The following rule is proposed as the random triangular grid-based search algorithm:
(
ν
if |M̂i (k)| =
6 0 with probability |ℵ(pi1(k))|
(3.1)
pi (k + 1) =
pi (k) if |M̂i (k)| = 0
where M̂i (k) = {m ∈ Mi (k); Vm (k) = 0} is the set of all elements of Mi (k) have not been
visited before, and |M̂i (k)| denotes the number of elements in M̂i (k).
Applying the rule (3.1) ensures that the area T̂ is completely explored and every vertex of
it is visited at least one time by a robot of the team.
Theorem 3.1.1 Suppose that all assumptions hold, and the mobile robots move according to
the distributed control law (3.1). Then, for any number of robots, with probability 1 there exists
a time k0 ≥ 0 such that Vτ (k0 ) = 1; ∀τ ∈ T̂ .
26
Proof of Theorem 3.1.1: The algorithm 3.1 defines an absorbing Markov chain which
contains many transient states and a number of absorbing states that are impossible to leave.
Transient states are all the vertices of the triangular grid T̂ which have been visited by the
robots during the search procedure. On the other hand, absorbing states are the vertices where
the robots stop at the end of the search operation. Using the algorithm 3.1, a robot goes to
the vertices where may have not been visited yet. Therefore, the number of transient states
will eventually decrease. This continues until the number of the robots is equal to the number
of unvisited vertices which will be the absorbing states. It is also clear that these absorbing
states can be reached from any initial states, with a non-zero probability. This implies that
with probability 1, one of the absorbing states will be reached. This completes the proof of
Theorem 3.1.1.
In Fig. 3.1, the flowchart of the proposed algorithm is presented that shows how our
decision-making approach is implemented. At the first step, robots start making their maps
using their sonar. Each robot, based on the vertex on which it is located, assumes some
probable neighbouring vertices on the common triangular grid. The number of these probable
neighbouring vertices and their distance to the robot depend on the robot’s sonar range. Then,
the robot uses the sonar to detect its surrounding environment including borders and obstacles.
If any of those probable neighbouring vertices is located outside the borders or blocked by an
obstacle, it will be ignored. The rest of those probable neighbouring vertices will be added
to the map of the robot. This step is repeated every time that the robot occupies a vertex.
In order to avoid sticking in borders or obstacles, we consider a margin near the borders and
obstacles that depends on the size of the robot. If a vertex on the map is closer to the borders
or obstacles less than the margin, it will be eliminated from the robot’s map.
Whenever a robot makes any changes to its map, it sends the new map to the other robots by
transmitting packets. On the other side, whenever a robot receives a packet of data, it extracts
the new vertices from the received map and adds them to its map. Since the communication
range of the robots is limited, if two robots are far from each other, they cannot directly
communicate but can do it via other robots. In other words, since there is a connected network
of robots, each robot has the role of a hub in the network in order to share the maps among
the robots. Therefore, all the robots that are connected and make a network have a common
map. In the second phase of the algorithm when the robots have a common triangular grid
map, a robot can go far from the other robots and be disconnected from the team for a while.
In this case, sharing maps between disconnected robot and the others is paused until the robot
returns back to the communication range of the team again.
In the next step, each robot randomly chooses one of the nearest neighbouring vertices in
the map and goes there. Since we assume that the map of a robot is a connected set, there
exists always at least one neighbouring vertex, and at most six vertices. If the robot reaches the
target vertex, it marks that vertex as an explored vertex in its map and sends it to the other
neighbouring robots as well. However, because of some practical issues, maybe it is impossible
to reach the target vertex at a limited time or even maybe the target vertex is fake that has
been wrongly created during mapping. To avoid such problems, we include a factor of time.
Since the robot knows its location and the location of the target, it can estimate the time
needed to achieve the goal based on the distance to the target and velocity of the robot. Here,
we consider the parameter ET as the expected time to reach the target, that is a factor of the
estimated time. This coefficient that has a value greater than one, actually reflects the effects
of a non-straight route because of the shape of the search area and also existing obstacles or
other robots on the robot’s path. If the travelling time were more than ET, the robot would
27
Start
No
Is there any unexplored vertex
in the map?
Stop
Yes
Are vertex already in the
map?
Mapping
Detect new vertices in
the area using sonar
Yes
No
Add vertex to the map
Ignore it
Select one of the nearest
vertices in the map randomly
Go to the selected vertex,
Use obstacle avoidance
No
No
Yes
Yes
Set the vertex as an
explored vertex in the map
Is time > ET?
Delete the vertex from the
map
send map to other robots
Received map from other
robots?
Yes
Map merging
No
Communication &
Sharing
Has the selected vertex been
reached?
Navigation & Guidance
Send updated map to other
robots
Figure 3.1: The flowchart of the suggested algorithm; searching the whole area
ignore that target vertex, delete it from its map, and send the updated map to the other robots.
Since the area has borders and obstacles, the robots should avoid them while they are
searching for the targets. That is why we have to use an obstacle avoidance in our algorithm.
In addition, because of practical problems such as sonar and encoders accuracy and slipping
28
of the robots, there might be the difference between the actual position of a robot and the
coordinates of the vertices stored in its map. Therefore, if a robot is closer to a target vertex
than a specified distance, we consider the goal has been achieved.
3.1.2
Searching for Targets
When the robots are looking for some targets in the area, they should continue the search
operation till all the targets are detected. If the number of the targets is not specified, they
have to search the entire area like what described in Section 3.1.1. When robots know the
number of the targets, they do not need to explore the entire area but until all the targets are
detected. Suppose T = {T1 , T2 , . . . , Tnt } be the set of nt static targets should be detected by
the robots. As it was stated before, we assume the robots equipped with sensors by which the
targets can be detected whenever they are close enough to the robots. The distance by which
the robots must be close to the targets to be able to detect them is the sensing range of the
robots (rs ).
Definition 3.1.6 Suppose a Boolean variable VTj (k) which defines the state of target Tj at
time k. VTj (k) = 1 if the target Tj has been detected by at least one of the robots, otherwise
Tj (k) = 0.
Then, we modify the rule (3.1) as the following rule to ensure that the search operation
stops after finding all the targets.
(
ν
if ∃Tj ∈ T; VTj (k) = 0
pi (k + 1) =
(3.2)
pi (k) if ∀Tj ∈ T; VTj (k) = 1
Theorem 3.1.2 Suppose that all assumptions hold, and the mobile robots move according to
distributed control law (3.2). Then, for any number of robots and any number of targets, with
probability 1 there exists a time k0 ≥ 0 such that ∀j ; VTj (k0 ) = 1.
Proof of Theorem 3.1.2: Proof is similar to the proof of Theorem 3.1.1.
The flowchart in Fig. 3.1 can also be used to describe this operation. Fig. 3.2 depicts the
procedure we use to implement this algorithm. Most procedures are the same as the previous
one; therefore, we ignore their description. We only need to change the condition that stops
the operation in the algorithm. The operation will be stopped if all the targets are detected.
Also, the robots send the information about the detected targets to the other members of the
team.
3.1.3
Patrolling
The above algorithms are appropriate for the cases when the robots should break the search
operation; for instance, when the aim is finding some predetermined objects or labeling some
vertices of a grid. In cases where a permanent search is needed; for example, in applications
such as continuously patrolling or surveying a region, the algorithm should be modified such
that the search procedure maintains. To achieve that, we modify control law 3.1 as follows:
1
;
(3.3)
|ℵ(pi (k))|
hence, the robots do not stop and continuously move between the vertices of the grid.
pi (k + 1) = ν with probability
29
Start
Is there still any target to
detect?
Yes
No
Stop
Mapping
Navigation & Guidance
Assign detected targets in
the map
Communication &
Sharing
Figure 3.2: The flowchart of the suggested algorithm; searching for targets
3.1.4
Robots’ Motion
To have a more real estimation of time that the robots spend to search an area to detect
the targets, we apply motion dynamics in our simulations. Besides, since the environment is
unknown to the team, it is essential to use an obstacle avoidance in the low-level control of
the robots. As a result, we use a method of reactive potential field control in order to avoid
obstacles [168].
3.2
Simulation Results
To verify the suggested algorithm, computer simulations are employed. The region W is considered to be searched by a few robots (see Fig. 2.1). We suppose a multi-robot team consisting
some mobile robots which are randomly located in the region W with random initial values
of angles. The goal is to search the whole area or to find some targets by the robots using
proposed random triangular grid-based search algorithm.
To simulate the algorithm, MobileSime, a simulator of mobile robots developed by Adept
MobileRobots, is used. We also use Visual C++ for programming and ARIA, a C++ library
that provides an interface and framework for controlling the robots. In addition, Pioneer 3DX
is selected as the type of the robots. Robots’ parameters in the simulations are given in Table
2.1. Since this simulator simulates the real robots along with all conditions of a real world,
the results of the simulations would be obtained in the real world experiments with the real
robots indeed. Furthermore, to prevent collisions between robots and to avoid the obstacles and
borders, an obstacle avoidance algorithm is applied using functions provided in ARIA library.
Moreover, to avoid hitting and sticking to the borders, we assume a margin near the borders
such that the robots do not pass it.
30
(a)
(b)
(c)
(d)
Figure 3.3: Robots’ trajectories after applying the second stage of the algorithm; (a) After
2m19s, (b) After 7m38s, (c) After 16m50s, (d) After 29m44s
3.2.1
Searching the Whole Area
As mentioned in Chapter 2, a two-stage algorithm is used to achieve the goal. First, the rule
(2.1),(2.2) is applied which uses consensus variables in order to drive the robots to the vertices
of a common triangular grid. In Fig. 3.3(a), the beginning of the robots’ trajectories are the
position of the robots after applying the first stage of the search algorithm, i.e., rule (2.1),(2.2)
(see Fig. 2.5).
The second stage of the algorithm, i.e., the algorithm 3.1, is applied whenever the first stage
is completed. Fig. 3.3 demonstrates the result of applying algorithm 3.1 on a team of three
robots. As seen in Fig. 3.3, the robots go through the vertices of the common triangular grid
based on the proposed algorithm until the whole area is explored by the robots. Fig. 3.3(a),
Fig. 3.3(b) and Fig. 3.3(c) display the trajectories of the robots at times 2m19s, 7m38s and
16m50s after applying algorithm 3.1, respectively. Fig. 3.3(d) shows trajectories of the robots
at time 29m44s when the search operation has been completed. It is obvious that the area W
is completely explored by the robots such that each vertex of the covering triangular grid is
occupied at least one time by the robots. In this case study, we assume that the sides of the
equilateral triangles are 2 meter (sensing range of the robots is √23 m) and the communication
range between the robots is 10 meter. The area of the region W is about 528 m2 .
Although any number of robots can be used to search the whole area, it is evident that more
robots complete the operation in a shorter time. However, more number of robots certainly
increases the cost of the operation. The question is, how many robots should be used in order
31
to optimize both the time and cost. It seems, that is somehow dependent on the shape of the
region and the obstacles and also the sides of the triangles. In order to have a better view
of a relation between the number of robots and the search duration, twenty simulations with
different number of robots have been done. We consider teams consisting of one to fifteen
robots. Then, minimum, maximum, average and standard deviation are calculated for the
search duration of each team. Table 3.1 displays the results of these simulations that are also
depicted in Fig. 3.4.
Table 3.1: Duration of search
No of Robots 1
2
3
Min
74.16
40.42 27.86 21.21 16.43 16.85 15.15 15.91 14.51 12.47 11.71 10.11 10.35 12.51 11.44
Max
108.64 69.13 43.68 35.59 31.43 28.48 24.27 24.80 23.08 23.17 22.52 22.50 19.40 19.15 17.93
Average
87.88
52.09 36.43 30.51 25.56 21.35 19.54 19.69 18.94 17.52 16.44 16.23 15.14 15.52 14.64
STD
9.67
7.83
4.17
4
4.18
5
3.74
6
2.73
7
2.51
8
2.52
9
2.22
10
3.17
11
2.36
12
2.89
13
2.06
14
1.95
15
1.94
Figure 3.4: Duration of Search Vs. Number of Robots
As depicted in this figure, the search time decreases by increasing the number of robots. It
is noticeable that after increasing the number of robots to a specific number, the search duration almost remains constant. Indeed, increasing the number of robots increases the probable
collision between robots during the operation. Consequently, robots have to turn each other
to avoid the collision, and that is the main reason which increases the search time. Therefore, increasing the number of robots more than a specific value (seven robots in this case)
32
(a)
(b)
(c)
(d)
(e)
(f)
Figure 3.5: Robots’ trajectories in the case of searching for three targets. (a) One robot detects
three targets in 52m25s, (b) Two robots detect the targets in 24m4s, (c) Three robots detect
targets in 14m44s, (d) Four robots detect the targets in 14m57s, (e) Five robots detect the
targets in 8m27s and (f) Six robots detect the targets in 8m15s.
33
will be ineffectual in terms of time and should be avoided to save cost. As obvious from Fig.
3.4, increasing the number of robots more than seven improves the search time less than five
minutes.
3.2.2
Searching for Targets
To demonstrate that how the algorithm works in the case of search for a given number of
targets, we consider an area the same as in section 3.2.1 with three targets therein. The
targets, shown by green disks in Fig. 3.5, are located in different parts of the area. We evaluate
the presented algorithm to figure out how the robots can find the targets using multi-robot
teams with different number of robots. As depicted in Fig. 3.5, the results of simulation by
the teams consisting of one to six robots have been presented. A target is assumed detected
whenever it lies in the sensing range of a robot of the team. If there is just one robot in the
team, all the targets should be detected by that robot; thus, it will take longer time in compare
with the cases that there are more robots in the team.
As shown in Fig. 3.5(a), a robot searches for three targets using the presented search
algorithm and they all have been detected after 52m25s. Fig. 3.5(b) shows the same case with
two robots in the team so that a robot has detected one target and the other one has detected
two other targets in 24m4s. It should be mentioned that in Fig. 3.5, only the paths of the
robots after making consensus have been displayed. In Fig. 3.5(c), three robots search for three
targets and each robot finds one of them in 14m44s. As shown in that figure, when a robot
detects a target, it continues the search operation until all targets are detected by the team.
Fig. 3.5(d)-(f) show the search operation using 4-6 robots, respectively.
What is very noticeable in this case is that it is expected decreasing the time of detecting
the targets by increasing the number of robots while it has not occurred in some instances. For
example, when the number of robots has been increased form three to four, the time of detecting
the targets has been increased form 14m44s to 14m57s. To explain why this happens, we should
consider the fact that the time needed to detect the targets depends on many parameters not
only on the number of robots. The shape of the area and obstacles therein, the initial position
of the robots in the area and also the relative distance of the robots and targets are significant
parameters that affect the time of search. The other serious parameter must be considered, is
the nature of the search algorithm that is random. That is we might have different paths for
the same cases.
To discover more about how the number of robots affects on the search duration, we do
more simulations for each case. For example, Fig. 3.6 shows the paths of the robots in the
case that three robots are looking for three targets similar to the previous instance. It shows
that we have different paths thus different search durations. While in the first simulation, the
targets are detected in 14m44s, it takes 13m41 in the second and 19m11s in the third one which
is much more than the first one. Fig. 3.7 shows the results of three different simulations using
five robots. As depicted in that figure, the period of search for these simulations are 8m27s,
11m46s and 14m21s. Comparing to the case with three robots, it is obvious that the differences
between periods of search in this case are less than the case of the team with three robots.
That is an expected result because more robots means more coverage of the search area; hence,
the chance of detecting targets increases.
In order to have a better view of the relation between the number of robots and the search
duration, twenty simulations with different number of robots have been done. We consider
teams consisting of one to ten robots. Then, minimum, maximum and average of search
34
(a) Targets are detected in 14m44s
(b) Targets are detected in 13m41s
(c) Targets are detected in 19m11s
Figure 3.6: Robots’ trajectories in the case of searching for three targets by three robots.
Table 3.2: Time of detecting targets
No of Robots
1
2
3
4
5
6
7
8
9
10
Min
39.13
19.14
13.74
11.21
7.97
5.62
5.38
6.23
5.41
5.04
Max
76.38
42.67
25.87
20.49
14.35
12.87
11.97
10.95
10.67
9.83
Average
55.72
29.97
19.86
15.23
11.79
9.31
8.83
8.67
8.26
7.01
STD
9.29
5.64
3.63
2.50
1.79
1.84
1.72
1.13
1.43
1.26
duration for simulations with each team as well as their standard deviation are calculated.
Table 3.2 displays the results of these simulations that are also depicted in Fig. 3.8. The
results show that although the average of search duration to detect the targets decreases by
increasing the number of robots, the standard deviation of the team with less number of robots
is significantly high. Simply, the number of robots should be proportional to the search area
to have an adequate chance to detect targets in an acceptable time.
35
(a) Targets are detected in 8m27s
(b) Targets are detected in 11m46s
(c) Targets are detected in 14m21s
Figure 3.7: Robots’ trajectories in the case of searching for three targets by five robots.
3.3
Summary
In this chapter, we have developed a distributed control algorithm, namely, random triangular
grid-based search algorithm, to drive a multi-robot team to explore an unknown area. We
have used a triangular grid pattern and a two-stage random search algorithm for the control law so that the robots randomly move through the vertices of the triangular grid during
the search operation. Therefore, a complete search of the whole area has been guaranteed.
A mathematically rigorous proof of convergence of the presented algorithm has been demonstrated. Furthermore, the computer simulation results using MobileSim, a powerful simulator
of real robots and environment, have been presented to show that the algorithm is effective and
practicable.
36
Figure 3.8: Time of target detection Vs. Number of robots
37
Chapter 4
Semi-Random Triangular Grid-Based
Search Algorithm
In Chapter 3, the random triangular grid-based search algorithm was presented. In that
method, the robots randomly move between the vertices of a triangular grid so that in each step
they only move to the one of the closest neighbouring vertices. In a similar way, we present the
second method of search, namely, ”semi-random triangular grid-based search algorithm” in this
chapter. By this method, the robots still randomly move between the vertices of a triangular
grid so that in each step they move to the one of the closest neighbouring vertices, but only to
those vertices which have not been visited by the robots yet. If all the neighbouring vertices
have already been visited, one of them will randomly be selected. A mathematically rigorous
proof of convergence with probability 1 of the algorithm is given. Moreover, our algorithm is
implemented and simulated using a simulator of the real robots and environment. The third
method of search will be presented in the next chapter.
4.1
Distributed Semi-Random Search Algorithm
To search an area by a team of mobile robots using the vertices of a grid as the exploring
points, we need to locate the searching mobile robots on the vertices of a common grid among
the robots. Consensus variables locating algorithm (described in Chapter 2) can be the first
stage of the suggested search algorithm, which locates all the robots on the vertices of a common
triangular grid T̂ (see Fig. 2.1). The next step will be search of the area W based on moving
the robots between the vertices of the covering grid of the area. In this chapter, we propose
a semi-random triangular grid-based search algorithm. Suppose a robot located on a vertex
of the common triangular grid. Consequently, it can explore the surrounding area using its
sensors, and that depends on the sensing range of its sensors. After exploring that area, the
robot moves to another point which can be one of the six neighbouring vertices in the triangular
grid. As the second method, we suppose that selecting the neighbouring vertex is random but
from those neighbouring vertices which have not been visited yet by any of the robots. If all
the neighbouring vertices have been visited already, then on of them will be randomly selected.
In this regard, there are a few scenarios can be considered to search the area. The first scenario
is exploring the whole area which can be applicable when the robots are searching for an
undetermined number of targets. Therefore, to detect all possible targets, the team of robots
must search the whole area. Patrolling of the area is the other application for this scenario
38
where the robots should move continuously to detect the possible intruders to the area. In the
case of given number of targets which is our second scenario, the search operation should be
stopped whenever all the targets are detected without searching the whole area,.
4.1.1
Searching the Whole Area
To make sure that the whole area is explored by the team of the robots, each vertex in the
triangular covering grid set of the area W must be visited at least one time by a member of the
team. Consider T̂ is a triangular covering grid of W, and also each vertex of T̂ has been visited
at least one time by a robot of the team. This guarantees that the area W has completely been
explored by the multi-robot team. Since the robots do not have any map at the beginning,
they need to do map making during the search operation so that their maps will gradually be
completed.
Definition 4.1.1 Let T̂i (k)Sbe as the set of all the vertices of T̂ have been detected by robot i
at time k. Then, T̂ (k) = T̂i (k) will be the map of the area W detected by the team of the
robots until time k.
Note that a detected vertex is different from an explored vertex. These terms are defined in
detail in the following definitions.
Definition 4.1.2 A detected vertex means that vertex is detected by a robot using map making,
and it is in the map of that robot though it might be visited or not by the robots.
Definition 4.1.3 An explored vertex is a vertex that is visited by, at least, one member of the
team.
Definition 4.1.4 The map of robot i at time k, Mi (k), is the set of the vertices in T̂ detected
by robot i itself or received from other robots by which they are detected until time k.
Definition 4.1.5 Suppose a Boolean variable Vτ (k) which defines the state of vertex τ ∈ T̂ (k)
at time k. Vτ (k) = 1 if the vertex τ has already been visited by, at least, one of the robots,
otherwise Vτ (k) = 0.
Assumption 4.1.1 The triangular grid set T̂i (k); k = 0, 1, ... is connected. That means if
τ ∈ T̂i (k), then, at least, one of the six nearest neighbours of τ also belongs to T̂i (k).
Let ℵ(pi (k)) be a set containing all the closest vertices to pi (k) on the triangular grid T̂ (k);
also, consider |ℵ(pi (k))| as the number of elements in ℵ(pi (k)). It is clear that 1 ≤ |ℵ(pi (k))| ≤ 6.
In addition, assume ν be a randomly opted element of ℵ(pi (k)). Moreover, Let ℵ̂(pi (k)) be a
set containing all the closest vertices to pi (k) on the triangular grid T̂ (k) which have not been
visited yet also consider |ℵ̂(pi (k))| as the number of elements in ℵ̂(pi (k)). It is clear that
1 ≤ |ℵ̂(pi (k))| ≤ 6. Furthermore, assume ν̂ be a randomly opted element of ℵ̂(pi (k)).
Consider at time k robot i is located at point pi (k), and it wants to go to the next vertex.
The following rule is proposed as the semi-random triangular grid-based search algorithm:
if |M̂i (k)| =
6 0 & |ℵ̂(pi (k))| =
6 0 with probability
ν̂
pi (k + 1) = ν
if |M̂i (k)| =
6 0 & |ℵ̂(pi (k))| = 0 with probability
pi (k) if |M̂i (k)| = 0
39
1
|ℵ̂(pi (k))|
1
|ℵ(pi (k))|
(4.1)
where M̂i (k) = {m ∈ Mi (k); Vm (k) = 0} is the set of all elements of Mi (k) have not been
visited before, and |M̂i (k)| denotes the number of elements in M̂i (k).
Applying the rule (4.1) ensures that the area T̂ is completely explored, and every vertex of
it is visited at least one time by a robot of the team.
Theorem 4.1.1 Suppose that all assumptions hold, and the mobile robots move according to
the distributed control law (4.1). Then, for any number of robots, with probability 1 there exists
a time k0 ≥ 0 such that Vτ (k0 ) = 1; ∀τ ∈ T̂ .
Proof of Theorem 4.1.1: The algorithm 4.1 defines an absorbing Markov chain which
contains many transient states and some absorbing states that are impossible to leave. Transient
states are all the vertices of the triangular grid T̂ which have been visited by the robots during
the search procedure. On the contrary, absorbing states are the vertices where the robots stop
at the end of search. Using the algorithm 4.1, a robot goes to the vertices where may have
not been visited yet. Therefore, the number of transient states will eventually decrease. That
continues until the number of robots is equal to the number of unvisited vertices which will be
the absorbing states. It is also clear that these absorbing states can be reached from any initial
states, with a non-zero probability. This implies that with probability 1, one of the absorbing
states will be reached. This completes the proof of Theorem 4.1.1.
In Fig. 4.1, the flowchart of the proposed algorithm is presented that shows how our
decision-making approach is implemented. At the first step, robots start making their maps
using their sonar. Each robot, based on the vertex on which it is located, assumes some
probable neighbouring vertices on the common triangular grid. The number of these probable
neighbouring vertices and their distance to the robot depend on the robot’s sonar range. Then,
the robot uses the sonar to detect its surrounding environment including borders and obstacles.
If any of those probable neighbouring vertices is located outside the borders or blocked by an
obstacle, it will be ignored. The rest of those probable neighbouring vertices will be added
to the map of the robot. This step is repeated every time that the robot occupies a vertex.
In order to avoid sticking in borders or obstacles, we consider a margin near the borders and
obstacles that depends on the size of the robot. If a vertex on the map is closer to the borders
or obstacles less than the margin, it will be eliminated from the robot’s map.
Whenever a robot makes any changes to its map, it sends the new map to the other robots by
transmitting packets. On the other side, whenever a robot receives a packet of data, it extracts
the new vertices from the received map and adds them to its map. Since the communication
range of the robots is limited, if two robots are far from each other, they cannot directly
communicate but can do it via other robots. In other words, since there is a connected network
of robots, each robot has the role of a hub in the network in order to share the maps among
the robots. Therefore, all the robots that are connected and make a network have a common
map. In the second phase of the algorithm when the robots have a common triangular grid
map, a robot can go far from the other robots and be disconnected from the team for a while.
In this case, sharing maps between the disconnected robot and the others is paused until the
robot returns back to the communication range of the team again.
In the next step, each robot randomly chooses one of the nearest unvisited neighbouring
vertices in the map and goes there. If all the neighbouring vertices have been visited already,
one of them is randomly selected. Since we assume that the map of a robot is a connected
set, there exists always at least one neighbouring vertex, and at most six vertices. If the robot
reaches the target vertex, it marks that vertex as an explored vertex in its map and sends it to
40
Start
No
Is there any unexplored vertex
in the map?
Stop
Yes
Mapping
Detect new vertices in
the area using sonar
Yes
Are vertex already in the
map?
No
Add vertex to the map
Ignore it
Have all the nearest vertices
been explored?
No
Yes
Select one of the nearest
unexplo red vertices in the
map randomly
Select one of the nearest
vertices in the map randomly
Navigation & Guidance
Send updated map to other
robots
Go to the selected vertex,
Use obstacle avoidance
No
Has the selected vertex been
reached?
No
Is time > ET?
Yes
Yes
Delete the vertex from the
map
send map to other robots
Received map from other
robots?
Yes
Map merging
No
Communication &
Sharing
Set the vertex as an
explored vertex in the map
Figure 4.1: The flowchart of the suggested algorithm; searching the whole area
41
the other neighbouring robots as well. However, because of some practical issues, maybe it is
impossible to reach the target vertex at a limited time or even maybe the target vertex is fake
that has been wrongly created during mapping. To avoid such problems, we include a factor of
time. Since the robot knows its location and the location of the target, it can estimate the time
needed to achieve the goal based on the distance to the target and velocity of the robot. Here,
we consider the parameter ET as the expected time to reach the target, that is a factor of the
estimated time. This coefficient that has a value greater than one, actually reflects the effects
of a non-straight route because of the shape of the search area and also existing obstacles or
other robots on the robot’s path. If the travelling time were more than ET, the robot would
ignore that target vertex, delete it from its map, and send the updated map to the other robots.
Since the area has borders and obstacles, the robots should avoid them while they are
searching for the targets. That is why we have to use an obstacle avoidance in our algorithm.
Moreover, because of practical problems like sonar and encoders accuracy and slipping of the
robots, there might be a difference between the actual position of a robot and the coordinates of
the vertices stored in its map. Therefore, if a robot is closer to a target vertex than a specified
distance, we consider the goal has been achieved.
4.1.2
Searching for Targets
When the robots are looking for some targets in the area, they should continue the search
operation till all the targets are detected. If the number of the targets is not specified, they
have to search the entire area like what described in Section 4.1.1. When robots know the
number of the targets, they do not need to explore the entire area but until all the targets are
detected. Suppose T = {T1 , T2 , . . . , Tnt } be the set of nt static targets must be detected by
the robots. As it was stated before, we assume the robots equipped with sensors by which the
targets can be detected whenever they are close enough to the robots. The distance by which
the robots must be close to the targets to be able to detect them is the sensing range of the
robots (rs ).
Definition 4.1.6 Suppose a Boolean variable VTj (k) which defines the state of target Tj at
time k. VTj (k) = 1 if the target Tj has been detected by at least one of the robots, otherwise
Tj (k) = 0.
Therefore, we modify the rule (4.1) as the following rule to ensure that the search operation
stops after finding all the targets.
if |M̂i (k)| =
6 0 & |ℵ̂(pi (k))| =
6 0 & (∃Tj ∈ T; VTj (k) = 0)
ν̂
pi (k + 1) = ν
if |M̂i (k)| =
6 0 & |ℵ̂(pi (k))| = 0 & (∃Tj ∈ T; VTj (k) = 0)
pi (k) if ∀Tj ∈ T; VTj (k) = 1
(4.2)
Theorem 4.1.2 Suppose that all assumptions hold, and the mobile robots move according to
distributed control law (4.2). Then, for any number of robots and any number of targets, with
probability 1 there exists a time k0 ≥ 0 such that ∀j ; VTj (k0 ) = 1.
Proof: Proof is similar to the proof of Theorem 4.1.1.
The flowchart in Fig. 4.1 can also be used to describe this operation. Fig. 4.2 depicts the
procedure we use to implement this algorithm. Most procedures are the same as the previous
42
one; therefore, we ignore their description. We only need to change the condition that stops
the operation in the algorithm. The operation will be stopped if all the targets are detected.
Also, the robots send the information about the detected targets to the other members of the
team.
Start
Is there still any target to
detect?
Yes
No
Stop
Mapping
Navigation & Guidance
Assign detected targets in
the map
Communication &
Sharing
Figure 4.2: The flowchart of the suggested algorithm; searching for targets
4.1.3
Patrolling
The above algorithms are appropriate for the cases when the robots should break the search
operation; for instance, when the aim is finding some predetermined objects or labeling some
vertices of a grid. In cases where a permanent search is needed; for example, in applications
such as continuously patrolling or surveying a region, the algorithm should be modified such
that the search procedure maintains. That is, we modify control law 4.1 as follows:
pi (k + 1) =
(
ν̂
if |ℵ̂(pi (k))| =
6 0 with probability
ν
if |ℵ̂(pi (k))| = 0 with probability
1
|ℵ̂(pi (k))|
1
|ℵ(pi (k))|
(4.3)
hence, the robots do not stop and continuously move between the vertices of the grid.
4.1.4
Robots’ Motion
To have a more real estimation of time that the robots spend to search an area to detect the
targets, we apply motion dynamics in our simulations. In addition, since the environment is
unknown to the team, it is essential to use an obstacle avoidance in the low level control of
the robots. As a result, we use a method of reactive potential field control in order to avoid
obstacles [168].
43
4.2
Simulation Results
To verify the suggested algorithm, computer simulations are employed. The region W is considered to be searched by a few robots (see Fig. 2.1). We suppose a multi-robot team with
some autonomous mobile robots which are randomly located in the region W with random
initial values of angles. The goal is to search the whole area by the robots using proposed
semi-random triangular grid-based search algorithm.
To simulate the algorithm, MobileSime, a simulator of mobile robots developed by Adept
MobileRobots, is used. We also use Visual C++ for programming and ARIA, a C++ library
that provides an interface and framework for controlling the robots. In addition, Pioneer 3DX
is selected as the type of the robots. Robots parameters in the simulations are given in Table
2.1. Since this simulator simulates the real robots along with all conditions of a real world,
the results of the simulations would be obtained in the real world experiments with the real
robots indeed. Furthermore, to prevent collisions between robots and to avoid the obstacles and
borders, an obstacle avoidance algorithm is applied using functions provided in ARIA library.
Moreover, to avoid hitting and sticking to the borders, we assume a margin near the borders
such that the robots do not pass it.
4.2.1
Searching the Whole Area
As mentioned in Chapter 2, a two-stage algorithm is used to achieve the goal. First, the
algorithm (2.1),(2.2) is applied which uses consensus variables in order to drive the robots to
the vertices of a common triangular grid. In Fig. 4.3(a), the beginning of the robots’ trajectories
are the position of the robots after applying the first stage of the search algorithm, i.e., rule
(2.1),(2.2) (see Fig. 2.5).
The second stage of the algorithm, i.e., the algorithm 4.1, is applied whenever the first stage
is completed. Fig. 4.3 demonstrates the result of applying algorithm 4.1 on a team of three
robots. As seen in Fig. 4.3, the robots go through the vertices of the common triangular grid
based on the proposed algorithm until the whole area is explored by the robots. Fig. 4.3(a),
Fig. 4.3(b) and Fig. 4.3(c) display the trajectories of the robots at times 3m48s, 8m5s and
15m47s after applying algorithm 4.1, respectively. Fig. 4.3(d) shows trajectories of the robots
at time 18m3s when the search operation has been completed. It is obvious that the area W
is completely searched by the robots such that each vertex of the covering triangular grid is
occupied at least one time by the robots. In this case study, we assume that the sides of the
equilateral triangles are 2 meter (sensing range of the robots is √23 m) and the communication
range between the robots is 10 meter. The area of the region W is about 528 m2 .
Although any number of robots can be used to search the whole area, it is obvious that
more robots complete the operation in a shorter time. However, more number of robots certainly increases the cost of the operation. The question is, how many robots should be used in
order to optimize both the time and cost. It seems, that is somehow dependent on the shape
of the region and the obstacles and also the sides of the triangles. In order to have a better
view of the relation between the number of robots and the search duration, twenty simulations
with different number of robots have been done. We consider teams consisting of one to fifteen
robots. Then, minimum, maximum, average and standard deviation are calculated for the
search duration of each team. Table 4.1 displays the results of these simulations that are also
depicted in Fig. 4.4.
44
(a)
(b)
(c)
(d)
Figure 4.3: Robots’ trajectories after applying the second stage of the algorithm; (a) After
3m48s, (b) After 8m5s, (c) After 15m47s, (d) After 18m3s
Table 4.1: Duration of search
No of Robots 1
2
3
4
5
6
7
8
9
10
11 12 13 14 15
Min
36.51 18.40 15.28 11.21 10.05 8.03 6.47 6.56 6.13 5.82 5.81 5.74 4.80 4.69 4.37
Max
48.62 32.05 21.06 17.56 15.36 14.25 12.61 11.33 10.52 10.37 9.65 9.11 8.59 8.62 8.51
Average
41.57 25.52 18.29 14.69 12.56 10.81 10.02 9.24 9.05 8.43 8.21 7.55 7.56 7.10 7.03
STD
3.53 2.95 1.93 1.92 1.29 1.46 1.46 1.34 1.10 1.09 1.06 1.01 1.02 0.97 0.96
As depicted in this figure, the search time decreases by increasing the number of robots. It
is noticeable that after increasing the number of robots to a specific number, the search duration almost remains constant. Indeed, increasing the number of robots increases the probable
collision between robots during the operation. Consequently, the robots have to turn each other
to avoid the collision, and that is the main reason which increases the search time. Therefore,
increasing the number of robots more than a particular value (seven robots in this case) will
be ineffectual in terms of time and should be avoided to save cost. As apparent from Fig. 4.4,
increasing the number of robots more than seven improves the search time only about three
minutes.
45
Figure 4.4: Duration of Search Vs. Number of Robots
4.2.2
Searching for Targets
To demonstrate that how the algorithm works in the case of search for a given number of
targets, we consider an area the same as in section 4.2.1 with three targets therein. The
targets, shown by green disks in Fig. 4.5, are located in different parts of the area. We evaluate
the presented algorithm to figure out how the robots can find the targets using multi-robot
teams with different number of robots. As depicted in Fig. 4.5, the results of simulation by
the teams consisting of one to six robots have been presented. A target is assumed detected
whenever it lies in the sensing range of a robot of the team. If there is just one robot in the
team, all the targets should be detected by that robot; thus, it will take longer time in compare
with the cases that there are more robots in the team.
As shown in Fig. 4.5(a), a robot searches for three targets using the presented search
algorithm and they all have been detected after 17m13s. Fig. 4.5(b) shows the same case with
two robots in the team so that a robot has detected one target and the other one has detected
two other targets in 12m24s. It should be mentioned that in Fig. 4.5, only the paths of the
robots after making consensus have been displayed. In Fig. 4.5(c) three robots search for three
targets and each robot finds one of them in 5m49s. As shown in that figure, when a robot
detects a target, it continues the search operation until all targets are detected by the team.
Fig. 4.5(d)-(f) show the search operation using 4-6 robots, respectively.
What is very noticeable in this case is that, it is expected decreasing the time of detecting
the targets by increasing the number of robots while it has not occurred in some instances. For
example, when the number of robots has been increased from three to four, the time of detecting
46
(a)
(b)
(c)
(d)
(e)
(f)
Figure 4.5: Robots’ trajectories in the case of searching for three targets. (a) One robot detects
three targets in 17m13s, (b) Two robots detect the targets in 12m24s, (c) Three robots detect
targets in 5m49s, (d) Four robots detect the targets in 7m33s, (e) Five robots detect the targets
in 4m8s and (f) Six robots detect the targets in 5m19s.
47
the targets has been increased from 5m49s to 7m33s. To explain why this happens, we should
consider the fact that the time needed to detect the targets depends on many parameters not
only on the number of robots. The shape of the area and obstacles therein, the initial position
of the robots in the area and also the relative distance of the robots and targets are significant
parameters that affect the time of search. The other serious parameter must be considered is
the nature of the search algorithm that is semi-random. That is we might have different paths
for the same cases.
(a) Targets are detected in 5m49s
(b) Targets are detected in 4m33s
(c) Targets are detected in 14m21s
Figure 4.6: Robots’ trajectories in the case of searching for three targets by three robots.
To discover more about how the number of robots affects on the search duration, we do
more simulations for each case. For example, Fig. 4.6 shows the paths of the robots in the case
that three robots are looking for three targets similar to the previous instance. It shows that we
have different paths thus different search durations. While in the first simulation, the targets
are detected in 5m49s, it takes 4m33s in the second and 14m21s in the last one, which is much
more than the second one. Fig. 4.7 shows the results of three different simulations using five
robots. As depicted in that figure, the period of search for these simulations are 4m8s, 4m37s
and 7m33s. Comparing to the case with three robots, it is obvious that the differences between
the periods of search in this instance are less than the case of the team with three robots. That
is an expected result because more robots means more coverage of the search area; hence, the
chance of detecting targets increases.
48
(a) Targets are detected in 4m8s
(b) Targets are detected in 4m37s
(c) Targets are detected in 7m33s
Figure 4.7: Robots’ trajectories in the case of searching for three targets by five robots.
49
Table 4.2: Time of detecting targets
No of Robots
1
2
3
4
5
6
7
8
9
10
Min
13.45
7.26
4.13
3.59
2.72
2.49
2.07
2.01
1.95
1.72
Max
29.50
19.95
14.35
10.70
7.55
6.08
5.23
5.15
4.97
4.95
Average
21.99
11.36
8.41
6.95
4.44
3.96
3.46
3.30
3.02
3.13
STD
5.79
3.11
2.34
1.96
1.12
1.02
0.85
0.94
0.77
0.82
Figure 4.8: Time of target detection Vs. Number of robots
In order to have a better view of the relation between the number of robots and the search
duration, twenty simulations with different number of robots have been done. We consider
teams consisting of one to ten robots. Then, minimum, maximum and average of search duration for simulations with each team as well as their standard deviation are calculated. Table
4.2 displays the results of these simulations that are also depicted in Fig. 4.8. The results show
that although the average of search duration to detect the targets decreases by increasing the
number of robots, the standard deviation of the team with less number of robots is significantly
high. Simply, the number of robots should be proportional to the search area to have an adequate chance to detect targets in a desirable time.
50
4.3
Summary
In this chapter, we have developed a distributed control algorithm, namely, semi-random triangular grid-based search algorithm, to drive a multi-robot team to explore an unknown area.
We have used a triangular grid pattern and a two-stage semi-random search algorithm for the
control law so that the robots randomly move through the vertices of the triangular grid during
the search operation. Therefore, a complete search of the whole area has been guaranteed.
A mathematically rigorous proof of convergence of the presented algorithm has been demonstrated. Furthermore, the computer simulation results using MobileSim, a powerful simulator
of real robots and environment, have been presented to demonstrate that the algorithm is
effective and practicable.
51
Chapter 5
Modified Triangular-Grid-Based Search
Algorithm
In Chapters 3 and 4, two methods for exploring an unknown environment using a team of mobile
robots were presented. In those methods, the robots randomly move between the vertices of
a common triangular grid so that in each step they only move to the one of the neighbouring
vertices. In this chapter, we present the third method of search, namely, ”modified triangular
grid-based search algorithm”. By this method, the robots are not confined to move to the closest
neighbouring vertices. Instead, they move to the nearest unvisited vertex. A mathematically
rigorous proof of convergence with probability 1 of the algorithm is given. Moreover, our
algorithm is implemented and simulated using a simulator of the real robots and environment
and also tested via experiments with Adept Pioneer 3DX wheeled mobile robots. Finally, a
comparison between our three proposed algorithms and three algorithms from other researchers
is given.
5.1
Distributed Triangular Grid-Based Search Algorithm
To search an area by a team of mobile robots using the vertices of a grid as the exploring
points, we need to locate the searching mobile robots on the vertices of a common grid among
the robots. Consensus variables locating algorithm (described in Chapter 2) can be the first
stage of the suggested search algorithm, which locates‘ all the robots on the vertices of a
common triangular grid T̂ (see Fig. 2.1). The next step will be search of the area W based
on moving the robots between the vertices of the covering grid of the area. In this chapter, we
propose a triangular grid-based search algorithm. Suppose a robot located on a vertex of the
common triangular grid. Consequently, it can explore the surrounding area using its sensors,
and that depends on the sensing range of its sensors. After exploring that area, the robot
moves to another point which can be one of the unexplored vertices in the triangular grid. As
a modified version of the previous methods presented in Chapters 3 and 4, we suppose that
a robots selects the nearest unvisited vertex as the next destination in every step. If there is
more than one vertex, any of them can be chosen randomly. In this regard, there are various
scenarios can be considered to search the area. The first scenario is exploring the whole area
which can be applicable when the robots are searching for an undetermined number of targets.
Therefore, to detect all possible targets, the team of the robots must search the whole area.
Patrolling of the area is the other application for this scenario where the robots should move
52
continuously to detect the possible intruders to the area. In the case of given number of targets
which is our second scenario, the search operation should be stopped without searching the
whole area; whenever all the targets are detected.
5.1.1
Searching the Whole Area
To make sure that the whole area is explored by the team of the robots, each vertex in the
triangular covering grid set of the area W must be visited at least one time by a member of the
team. Consider T̂ is a triangular covering grid of W, and also each vertex of T̂ has been visited
at least one time by a robot of the team. This guarantees that the area W has completely been
explored by the multi-robot team. Since the robots do not have any map at the beginning,
they need to do map making during the search operation so that their maps will gradually be
completed.
Definition 5.1.1 Let S
T̂i (k) be the set of all the vertices of T̂ have been detected by robot i at
time k. Then, T̂ (k) = T̂i (k) will be the map of the area W detected by the team of the robots
until time k.
Note that a detected vertex is different from an explored vertex. These terms are defined in
detail in the following definitions.
Definition 5.1.2 A detected vertex means that vertex is detected by a robot using map making,
and it is in the map of that robot though it might be visited or not by the robots.
Definition 5.1.3 An explored vertex is a vertex that is visited by at least one member of the
team.
Definition 5.1.4 The map of robot i at time k, Mi (k), is the set of those vertices in T̂ detected
by robot i itself or received from other robots by which they are detected until time k.
Definition 5.1.5 Suppose a Boolean variable Vτ (k) which defines the state of vertex τ ∈ T̂ (k)
at time k. Vτ (k) = 1 if the vertex τ has already been visited by at least one of the robots,
otherwise Vτ (k) = 0.
Assumption 5.1.1 The triangular grid set T̂i (k); k = 0, 1, ... is connected. That means if
τ ∈ T̂i (k), then at least one of the six nearest neighbours of τ also belongs to T̂i (k).
Consider at time k robot i is located at point pi (k), and it wants to go to the next vertex.
The following rule is proposed as the modified triangular grid-based search algorithm:
(
m̂i (k) if |M̂i (k)| =
6 0
pi (k + 1) =
(5.1)
pi (k) if |M̂i (k)| = 0
where M̂i (k) = {m ∈ Mi (k); Vm (k) = 0} is the set of all elements of Mi (k) have not been
visited before, |M̂i (k)| denotes the number of elements in M̂i (k) and m̂i (k) is the vertex in
M̂i (k) nearest to robot i at time k.
Applying the rule (5.1) ensures that the area T̂ is completely explored, and every vertex of
it is visited at least one time by a robot of the team.
53
Theorem 5.1.1 Suppose that all assumptions hold, and the mobile robots move according to
the distributed control law (5.1). Then, for any number of robots, with probability 1 there exists
a time k0 ≥ 0 such that Vτ (k0 ) = 1; ∀τ ∈ T̂ .
Proof of Theorem 5.1.1: The algorithm 5.1 defines an absorbing Markov chain which
contains many transient states and a number of absorbing states that are impossible to leave.
Transient states are all the vertices of the triangular grid T̂ which have been occupied already
by the robots during the search procedure. On the other hand, absorbing states are the vertices
where the robots stop at the end of the search operation. Using the algorithm 5.1, a robot goes
to the vertices where may have not been visited yet. Therefore, the number of transient states
will eventually decrease. That continues until the number of robots is equal to the number of
unvisited vertices which will be the absorbing states. It is also explicit that these absorbing
states can be reached from any initial states, with a non-zero probability. This implies that
with probability 1, one of the absorbing states will be reached. This completes the proof of
Theorem 5.1.1.
In Fig. 5.1, the flowchart of the proposed algorithm is presented that shows how our
decision-making approach is implemented. At the first step, robots start making their maps
using their sonar. Each robot, based on the vertex on which it is located, assumes some
probable neighbouring vertices on the common triangular grid. The number of these probable
neighbouring vertices and their distance to the robot depend on the robot’s sonar range. Then,
the robot uses the sonar to detect its surrounding environment including borders and obstacles.
If any of those probable neighbouring vertices is located outside the borders or blocked by an
obstacle, it will be ignored. The rest of those probable neighbouring vertices will be added
to the map of the robot. This step is repeated every time that the robot occupies a vertex.
In order to avoid sticking in borders or obstacles, we consider a margin near the borders and
obstacles that depends on the size of the robot. If a vertex on the map is closer to the borders
or obstacles less than the margin, it will be eliminated from the robot’s map.
Whenever a robot makes any changes to its map, it sends the new map to the other robots by
transmitting packets. On the other side, whenever a robot receives a packet of data, it extracts
the new vertices from the received map and adds them to its map. Since the communication
range of the robots is limited, if two robots are far from each other, they cannot directly
communicate but can do it via other robots. In other words, since there is a connected network
of robots, each robot has the role of a hub in the network in order to share the maps among
the robots. Therefore, all the robots that are connected and make a network have a common
map. In the second phase of the algorithm when the robots have a common triangular grid
map, a robot can go far from the other robots and be disconnected from the team for a while.
In this case, sharing maps between the disconnected robot and the others is paused until the
robot returns back to the communication range of the team again.
In the next step, each robot chooses the nearest unexplored vertex in the map and goes there.
Whenever a robot visits a vertex, it marks that as an explored vertex in its map. Therefore,
all the robots know which vertices have not been explored yet. To find the nearest vertex that
has not been explored yet, each robot searches the map stored in its memory. This vertex has
probably been detected by the robot itself or added to robot’s map via sharing map among
robots. After finding the nearest unexplored vertex, the robot moves to reach there. If the
robot reaches the target vertex, it marks that vertex as an explored vertex in its map and sends
it to the other neighbouring robots as well. However, because of some practical issues, maybe it
is impossible to reach the target vertex at a limited time or even maybe the target vertex is fake
54
Start
No
Is there any unexplored vertex
in the map?
Stop
Yes
Are vertex already in the
map?
Mapping
Detect new vertices in
the area using sonar
Yes
No
Add vertex to the map
Ignore it
Find in the map the nearest
unexplored vertex
Go to the selected vertex,
Use obstacle avoidance
No
No
Yes
Yes
Set the vertex as an
explored vertex in the map
Is time > ET?
Delete the vertex from the
map
send map to other robots
Received map from other
robots?
Yes
Map merging
No
Communication &
Sharing
Has the selected vertex been
reached?
Navigation & Guidance
Send updated map to other
robots
Figure 5.1: The flowchart of the suggested algorithm; searching the whole area
that has been wrongly created during mapping. To avoid such problems, we include a factor of
time. Since the robot knows its location and the location of the target, it can estimate the time
needed to achieve the goal based on the distance to the target and velocity of the robot. Here,
we consider the parameter ET as the expected time to reach the target, that is a factor of the
55
estimated time. This coefficient with a value greater than one, actually reflects the effects of a
non-straight route because of the shape of the search area and also existing obstacles or other
robots on the robot’s path. If the traveling time were more than ET, the robot would ignore
that target vertex, delete it from its map, and send the updated map to the other robots.
Since the area has borders and obstacles, the robots should avoid them while they are
searching for the targets. That is why we have to use an obstacle avoidance in our algorithm.
In addition, because of practical problems like sonar and encoders accuracy and slipping of the
robots, there might be a difference between the actual position of a robot and the coordinates of
the vertices stored in its map. Therefore, if a robot is closer to a target vertex than a specified
distance, we consider the goal has been achieved.
5.1.2
Searching for Targets
When the robots are looking for some targets in the area, they should continue the search
operation till all the targets are detected. If the number of the targets is not given, they have
to search the entire area like what described in Section 5.1.1. When robots know the number
of the targets, they do not need to explore the entire area but until all the targets are detected.
Suppose T = {T1 , T2 , . . . , Tnt } be the set of nt static targets should be detected by the robots.
As it was stated before, we assume the robots equipped with sensors by which the targets can
be detected whenever they are close enough to the robots. The distance by which the robots
must be close to the targets to be able to detect them is the sensing range of the robots (rs ).
Definition 5.1.6 Suppose a Boolean variable VTj (k) which defines the state of target Tj at
time k. VTj (k) = 1 if the target Tj has been detected by at least one of the robots, otherwise
Tj (k) = 0.
Therefore, we modify the rule (5.1) as the following rule to ensure that the search operation
stops after finding all the targets.
(
m̂i (k) if ∃Tj ∈ T; VTj (k) = 0
pi (k + 1) =
pi (k) if ∀Tj ∈ T; VTj (k) = 1
(5.2)
Theorem 5.1.2 Suppose that all assumptions hold, and the mobile robots move according to
distributed control law (5.2). Then, for any number of robots and any number of targets, with
probability 1 there exists a time k0 ≥ 0 such that ∀j ; VTj (k0 ) = 1.
Proof: Proof is similar to the proof of Theorem 5.1.1.
The flowchart in Fig. 5.1 can also be used to describe this operation. Fig. 5.2 depicts the
procedure we use to implement this algorithm. Most procedures are the same as the previous
one; therefore, we ignore their description. We only need to change the condition that stops the
operation in the algorithm. The operation will be stopped when all the targets are detected.
Also, the robots send the information about the detected targets to the other members of the
team.
56
Start
Is there still any target to
detect?
Yes
No
Stop
Mapping
Navigation & Guidance
Assign detected targets in
the map
Communication &
Sharing
Figure 5.2: The flowchart of the suggested algorithm; searching for targets
5.1.3
Patrolling
The above algorithms are appropriate for the cases when the robots should break the search
operation; for instance, when the aim is finding some predetermined objects or labeling some
vertices of a grid. In cases where a permanent search is needed; for example, in applications
such as continuously patrolling or surveying a region, the algorithm should be modified such
that the search procedure maintains. That can be done by periodically changing the states
of the vertices in the maps of the robots; hence, the robots consider those vertices as the
targets again. To archive that, the state of the previously visited vertices should be changed
to unvisited in the map of the robots. Based on some conditions such as the size of the area
and the number of the robots, the period of changing the state of vertices can be chosen.
5.1.4
Robots’ Motion
To have a more real estimation of time that the robots spend to search an area to detect the
targets, we apply motion dynamics in our simulations. In addition, since the environment is
unknown to the team, it is essential to use an obstacle avoidance in the low-level control of
the robots. As a result, we use a method of reactive potential field control in order to avoid
obstacles [168].
5.2
Simulation Results
To verify the suggested algorithm, computer simulations are employed. The region W is considered to be searched by a few robots (see Fig. 2.1). We suppose a multi-robot team of some
autonomous mobile robots which are randomly located in the region W with random initial
57
values of angles. The goal is to search the whole area by the robots using proposed grid-based
search algorithm.
To simulate the algorithm, MobileSime, a simulator of mobile robots developed by Adept
MobileRobots, is used. We also use Visual C++ for programming and ARIA, a C++ library
that provides an interface and framework for controlling the robots. In addition, Pioneer 3DX
is selected as type of the robots. Robots parameters in the simulations are given in Table
2.1. Since, this simulator simulates the real robots along with all conditions of a real world,
the results of the simulations would be obtained in the real world experiments with the real
robots indeed. Furthermore, to prevent collisions between robots and to avoid the obstacles and
borders, an obstacle avoidance algorithm is applied using functions provided in ARIA library.
Moreover, to avoid hitting and sticking to the borders, we assume a margin near the borders
such that the robots do not pass it.
(a)
(b)
(c)
(d)
Figure 5.3: Robots’ trajectories after applying the second stage of the algorithm; (a) After
3m21s, (b) After 5m45s, (c) After 7m26s, (d) After 8m23s
5.2.1
Searching the Whole Area
As mentioned in Chapter 2, a two-stage algorithm is used to achieve the goal. First, the
algorithm (2.1),(2.2) is applied which uses consensus variables in order to drive the robots to
the vertices of a common triangular grid. In Fig. 5.3(a), the beginning of the robots’ trajectories
58
are the position of the robots after applying the first stage of the search algorithm, i.e., rule
(2.1),(2.2) (see Fig. 2.5).
The second stage of the algorithm, i.e., the algorithm 5.1, is applied whenever the first stage
is completed. Fig. 5.3 demonstrates the result of applying algorithm 5.1 on a team of three
robots. As seen in Fig. 5.3, the robots go through the vertices of the common triangular grid
based on the proposed algorithm until the whole area is explored by the robots. Fig. 5.3(a),
Fig. 5.3(b) and Fig. 5.3(c) display the trajectories of the robots at times 3m21s, 5m45s and
7m26s after applying algorithm 5.1, respectively. Fig. 5.3(d) shows trajectories of the robots
at time 8m23s when the search operation has been completed. It is obvious that the area W
is completely explored by the robots such that each vertex of the covering triangular grid is
occupied at least one time by the robots. In this case study, we assume that the sides of the
equilateral triangles are 2 meter (sensing range of the robots is √23 m) and the communication
range between the robots is 10 meter. The area of the region W is about 528 m2 .
Fig. 5.4 shows another simulation result but for an area with three obstacles with different
shapes. It shows that the algorithm indeed works with any number of obstacles.
Although any number of robots can be used for the search operation, it is obvious that more
robots complete the operation in a shorter time. However, more robots certainly increases the
cost of the operation. The question is, how many number of robots should be used in order
to optimize both the time and cost. It seems, that may be somehow dependent on the shape
of the region and the obstacles and also the sides of the triangles. In order to have a better
view of the relation between the number of robots and the search duration, twenty simulations
with different number of robots have been done. We consider teams consisting of one to fifteen
robots. Then, minimum, maximum, average and standard deviation are calculated for the
search duration of each team. Table 5.1 displays the results of these simulations that are also
depicted in Fig. 5.5.
Figure 5.4: Searching an area with three obstacles
As depicted in this figure, the search time decreases by increasing the number of robots. It
is noticeable that after increasing the number of robots to a specific number, the search duration almost remains constant. Indeed, increasing the number of robots increases the probable
collision between robots during the operation. Consequently, robots have to turn each other
to avoid the collision, and that is the main reason which increases the search time. Therefore,
59
Table 5.1: Duration of search
No of Robots 1
2
3
4
5
6
7
8
9
10 11 12 13 14 15
Min
16.98 9.55 7.03 5.68 5.09 4.23 3.72 3.66 3.44 3.38 3.18 3.12 2.98 3.01 2.88
Max
23.02 13.61 9.98 7.90 6.98 5.82 5.05 4.96 4.77 4.65 4.38 4.26 4.12 4.17 3.76
Average
20.51 11.62 8.56 6.73 5.97 5.00 4.42 4.37 4.09 4.00 3.83 3.71 3.56 3.57 3.31
STD
1.77 1.20 0.98 0.54 0.62 0.52 0.37 0.40 0.42 0.36 0.35 0.33 0.34 0.41 0.25
Figure 5.5: Duration of Search Vs. Number of Robots
increasing the number of robots more than a specific value (six robots in this case) will be
ineffectual in terms of time and should be avoided to save cost. As obvious from Fig. 5.5,
increasing the number of robots more than six, improves the search time less than one minute.
5.2.2
Searching for Targets
To demonstrate that how the algorithm works in the case of search for a given number of
targets, we consider an area the same as in section 5.2.1 with three targets therein. The
targets, shown by green disks in Fig. 5.6, are located in different parts of the area. We evaluate
the presented algorithm to figure out how the robots can find the targets using multi-robot
teams with different number of robots. As depicted in Fig. 5.6, the results of simulation by
the teams consisting of one to six robots have been presented. A target is assumed detected
whenever it lies in the sensing range of a robot of the team. If there is just one robot in the
60
(a)
(b)
(c)
(d)
(e)
(f)
Figure 5.6: Robots’ trajectories in the case of searching for three targets. (a) One robot detects
three targets in 12m49s, (b) Two robots detect the targets in 4m3s, (c) Three robots detect
targets in 3m32s, (d) Four robots detect the targets in 3m50s, (e) Five robots detect the targets
in 1m51s and (f) Six robots detect the targets in 2m15s.
61
team, all the targets should be detected by that robot; thus, it will take longer time in compare
with the cases that there are more robots in the team.
As shown in Fig. 5.6(a), a robot searches for three targets using the presented search
algorithm and they all have been detected after 12m49s. Fig. 5.6(b) shows the same case
with two robots in the team so that a robot has detected one target and the other one has
detected two other targets in 4m3s. It should be mentioned that in Fig. 5.6, only the paths
of the robots after making consensus have been displayed. In Fig. 5.6(c), three robots search
for three targets and each robot finds one of them in 3m32s. As shown in that figure, when a
robot detects a target, it continues the search operation until all the targets are detected by
the team. Fig. 5.6(d)-(f) show the search operation using 4-6 robots, respectively.
What is very noticeable in this case is that it is expected decreasing the time of detecting
the targets by increasing the number of robots while it has not occurred in some instances. For
example, when the number of robots has been increased from three to four, the time of detecting
the targets has been increased from 3m32s to 3m50s. To explain why this happens, we should
consider the fact that the time needed to detect the targets depends on many parameters not
only on the number of robots. The shape of the area and obstacles therein, the initial position
of the robots in the area and also the relative distance of the robots and targets are significant
parameters that affect the time of search. The other serious parameter must be considered
is the nature of the search algorithm that is semi-random. Indeed, a robot always chooses
the nearest unexplored vertex as its destination vertex, but sometimes there are a few vertices
which can be chosen as the nearest, and the robot randomly selects one of them. That is we
might have different paths for the same cases.
To discover more about how the number of robots affects on the search duration, we do
more simulations for each case. For example, Fig. 5.7 shows the paths of the robots in the case
that two robots are looking for three targets similar to the previous instance. It shows that we
have different paths thus different search durations. While in the first simulation, the targets
are detected in 4m3s, in the last one it occurs in 7m41s which is much more than the first one.
Fig. 5.8 shows the results of three different simulations using five robots. As depicted in that
figure, the period of search for these simulations are 1m51s, 1m58s and 2m33s. Comparing to
the case with two robots, it is obvious that the differences between periods of search in this
case are less than the case of the team with two robots. That is an expected result because
more robots means more coverage of the search area; hence, the chance of detecting targets
increases.
To have a better view of the relation between the number of robots and the search duration, twenty simulations with different number of robots have been done. We consider teams
consisting of one to ten robots. Then minimum, maximum and average of search duration
for simulations with each team as well as their standard deviation are calculated. Table 5.2
displays the results of these simulations that are also depicted in Fig. 5.9. The results show
that although the average of search duration to detect the targets decreases by increasing the
number of robots, the standard deviation of the team with less number of robots is significantly high. Simply, the number of robots should be proportional to the search area to have
an adequate chance to detect the targets in an acceptable time.
62
(a) Targets are detected in 4m3s
(b) Targets are detected in 6m44s
(c) Targets are detected in 7m41s
Figure 5.7: Robots’ trajectories in the case of searching for three targets by two robots.
63
(a) Targets are detected in 1m51s
(b) Targets are detected in 1m58s
(c) Targets are detected in 2m33s
Figure 5.8: Robots’ trajectories in the case of searching for three targets by five robots.
64
Table 5.2: Time of detecting targets
No of Robots
1
2
3
4
5
6
7
8
9
10
Min
7.82
3.95
2.35
2.05
1.73
1.54
1.45
1.38
1.32
1.27
Max
15.49
8.02
5.05
4.21
3.63
3.42
2.88
2.56
2.19
2.01
Average
11.95
5.99
3.61
3.02
2.58
2.34
2.07
1.80
1.70
1.60
STD
2.03
1.37
0.79
0.71
0.69
0.61
0.47
0.36
0.30
0.24
Figure 5.9: Time of target detection Vs. Number of robots
5.3
Experiments with Real Robots
In this section, the experiment results with real robots are presented. We use Pioneer 3DX
mobile robots to implement the proposed algorithm. Pioneer 3DX is one of the world’s most
popular research mobile robots. These robots are equipped with encoders and sonar by which
the localization and mapping can be done during the experiments. We do experiment with
three robots in a surrounded area. The workspace is an irregularly shaped fenced area of about
16.5 m2 with a rectangular obstacle of about 0.8 m2 (see Fig. 5.12). The goal is searching the
whole area by three robots using the proposed algorithm.
We have used sonar to detect area boundaries, obstacles and other robots. A method of
reactive potential field control is applied to avoid collisions with obstacles. Moreover, other
robots are viewed as obstacles by a robot. We have used ArNetworking, a library provided by
Adept MobileRobots for communication between robots. It works with ARIA, the library we
65
have applied to control the robots. The robots exchange the map of the environment that they
make during the search operation and the vertices which they visit.
Figure 5.10: Pioneer 3DX robot used in the experiments
(a)
(b)
(c)
(d)
Figure 5.11: Three robots exploring a whole area
The snapshots of the experiment results are demonstrated in Fig. 5.11. First, the robots
are randomly placed in the field with an obstacle . After applying the algorithm, the robots
start to do the first stage of the algorithm, i.e., locating themselves on the vertices of a common
triangular grid using consensus variables locating algorithm. Fig. 5.11(a) shows the result of
this stage. Then, the robots begin to search the area based on the second stage of the algorithm
66
that is modified triangular grid-based search algorithm presented in this chapter. Fig. 5.11(b)Fig. 5.11(d) are snapshots during this stage. The overall paths taken by the robots in the
experiment are depicted in Fig. 5.12. The results show that the area has been completely
explored based on the proposed algorithm. Note that the robots do not necessarily reach the
end points at the same time.
Table 5.3: search duration in the
No of Robots 1
Min
57
Max
66
Average
61.6
STD
4.04
experiments (seconds)
2
3
35
18
41
22
37.8 19.4
2.39 1.67
Also, to have a better view of a relation between the number of robots and the search
duration in the experiments, five runs with one, two and three robots have been done. Table
5.3 displays the calculated values for the minimum, maximum, average and standard deviation
of the search duration in the experiments.
Note that since the area and the triangles sides in the simulations and the experiments
are different, and we had only three robots in the experiments, an exact comparison between
results is not possible. But as the Tables 5.1 and 5.3 show, the results from the simulations
and experiments give similar outcomes. For example, in the simulations, three robots explore a
528 m2 area in about 546 seconds (with exploring speed of 0.97 m2 /s), and in the experiments
three robots explore a 16.5 m2 area in about 19.4 seconds (with exploring speed of 0.85 m2 /s).
We have used 1 m for the sides of triangles in the experiments, while in the simulations, they
have been 2 m; therefore, the robots have had more stops in the experiments thus the lower
speed of exploring.
Figure 5.12: Robots’ trajectories in the experiment
67
5.4
Comparison Between Presented Search Algorithms
and Other Methods
To evaluate the proposed triangular grid-based search algorithms in Chapters 3, 4 and this
chapter against other algorithms, we compare them with each other also with three more
algorithms: fixed-length random, Levy random-walk and Levy random-walk with potential
field algorithms [169]. In [169], the simulation results for these three algorithms have been
given in a plot of search time versus the number of robots. They consider a 400 m2 regular
square area to be explored by a team of mobile robots with a speed of 60 cm/s. We have
considered an irregular shape with an area of about 528 m2 , and the robots with a maximum
speed of 40 cm/s.
First, we compare our presented algorithm in this chapter with algorithms presented in
Chapters 3 and 4. All algorithms are grid-based such that the robots certainly go through the
vertices of a triangular grid. In the algorithm of random triangular grid-based search, presented
in Chapter 3, robots go to the one of the six nearest vertices in each step regardless that the
vertex has been explored already or not; therefore, we have a pure random grid-based search
algorithm. On the other side, in the algorithm of semi-random triangular grid-based search
presented in Chapter 4, a robot goes to the one of the unexplored neighbouring vertices (each
vertex has at most six neighbouring vertices). If all the nearest neighbouring vertices have been
visited, one of them is randomly selected. It is clear that a semi-random algorithm is faster
than a pure random algorithm but requires more resources and processing. Our third proposed
algorithm, i.e., modified triangular grid-based search algorithm presented in this chapter, drives
the robots to move to the nearest unexplored vertex in the area in each step. In this algorithm,
the robots are not constrained to only move to the one of the six nearest neighbouring vertices
in the grid; but, they can move to the nearest unexplored vertex anywhere in the area.
Fig. 5.13 displays a comparison between our presented search algorithms using data from
Tables 3.1, 4.1 and 5.1. As shown in Fig. 5.13, semi-random triangular grid-based search
algorithm is faster than random triangular grid-based search algorithm, and modified triangular
grid-based search algorithm is much faster than both. Numerically, the speed of search by semirandom triangular grid-based search algorithm is about fifty percent more than by random
triangular grid-based search algorithm, and the speed of search by modified triangular gridbased search algorithm is more than fifty percent greater than by random triangular grid-based
search algorithm. Also, a comparison between the standard deviation of these algorithms shows
that standard deviation in modified triangular grid-based search algorithm is much less than
the other methods meaning that in this algorithm, repeating the simulation for a team of robots
does not have a considerable effect on the time of search.
Second, we compare our three algorithms with three other algorithms: fixed-length random,
Levy random-walk and Levy random-walk with potential field algorithms. Table 5.4 shows the
search time for these algorithms and our proposed algorithms that also depicted in Fig. 5.14.
Note that the search times given in Table 5.4 are approximated values extracted from the graph
given in [169], and also average values from Tables 3.1, 4.1 and 5.1. Table 5.5 shows the speed
of search versus the number of robots for the algorithms that also depicted in Fig. 5.15. Since
the size of the areas and speed of the robots in our simulations and the simulations in [169] are
different, they have been normalized; then, we can compare them. As shown in Fig. 5.15, the
fastest algorithm is our algorithm presented in this chapter, i.e., modified triangular grid-based
search algorithm; more than two times faster than the fastest algorithm presented in [169]. Fig.
68
Figure 5.13: Comparison between presented search algorithms
5.15 also shows that semi-random triangular grid-based search algorithm presented in Chapter
4, is almost as fast as Levy random-walk and potential field algorithm, the fastest algorithm
in [169]; but, still slightly faster. Finally, Fig. 5.15 shows that random triangular grid-based
search algorithm presented in Chapter 3, is slightly slower than Levy random-walk algorithm
but noticeably faster than the fixed-length random algorithm; about two times faster.
5.5
Summary
In this chapter, we have developed a distributed control algorithm to drive a multi-robot team
to explore an unknown area. We have used a triangular grid pattern and a two-stage algorithm
for the control law so that the robots move through the vertices of the grid during the search
procedure. Therefore, a complete search of the whole area has been guaranteed. A mathematically rigorous proof of convergence of the presented algorithm has been demonstrated.
Furthermore, the computer simulation results using MobileSim, a powerful simulator of real
robots and environments, have been presented to confirm that the algorithm is effective and
practicable. Also, the experiments with Pioneer 3DX wheeled mobile robots have been done to
confirm the performance of our suggested algorithm. The presented results of the experiments
with real robots show that the algorithm is quite practical. Finally, a comparison between
the proposed triangular grid-based search algorithms in Chapters 3, 4 and this chapter against
three other algorithms have been given.
69
Table 5.4: Comparison Between Presented Search Algorithms and Other Methods (Search time,
minute)
@
Search
@
@
@
Algorithm
@
@
@
@
@
Fixedlength
random
Levy
randomwalk
Levy
randomwalk and
Potential
Field
Random
Triangular
GridBased
SemiRandom
Triangular
GridBased
Modified
TriangularGridBased
1
253.3
114.7
115.7
87.9
41.6
20.5
5
87.5
19.5
13.9
25.6
12.6
6.0
10
38.8
14.7
9.3
17.5
8.4
4.0
@
@
@
@
@
@
Number of
robots
@
@
@
@
@
Figure 5.14: Comparison between presented search algorithms and other methods
70
Table 5.5: Comparison Between Presented Search Algorithms and Other Methods (Searching
speed, m2 /min)
@
Search
@
@
@
Algorithm
@
@
@
@
@
Fixedlength
random
Levy
randomwalk
Levy
randomwalk and
Potential
Field
Random
Triangular
GridBased
SemiRandom
Triangular
GridBased
Modified
TriangularGridBased
1
1.1
2.3
2.3
3.0
6.4
13.0
5
3.0
13.7
19.2
10.4
21.2
44.7
10
6.9
18.2
28.8
15.2
31.6
66.7
@
@
@
@
@
Number of
robots
@
@
@
@
@
@
Figure 5.15: Comparison between presented search algorithms and other methods
71
Chapter 6
Formatiom Building with Obstacle
Avoidance
In this chapter, we propose a distributed motion coordination control algorithm for a team
of mobile robots so that the robots collectively move in a desired geometric pattern from any
initial position while avoiding the obstacles on their routes. In the proposed method, the
robots have no information on the shape and position of the obstacles and only use range
sensors to obtain the information. We use standard kinematic equations for the robots with
hard constraints on the linear and angular velocities. There is no leader in the team and the
robots apply a distributed control algorithm based on the local information they obtain from
their nearest neighbours. We take the advantage of using the consensus variables approach
that is a known rule in multi-agent systems. Also, we propose a randomized algorithm for the
anonymous robots, which achieves the convergence to the desired configuration with probability
1. Furthermore, we propose a novel obstacle avoidance technique based on the information from
the range sensors. Mathematically rigorous proofs of the proposed control algorithms are given,
and the effectiveness of the algorithms are illustrated via computer simulations.
6.1
Multi-Robot System
We consider a system consisting of n autonomous mobile robots labeled 1 through n moving in
a plane. The kinematic equations of motion for the robots are given by
ẋi (t) = vi (t) cos(θi (t))
ẏi (t) = vi (t)sin(θi (t))
(6.1)
θ̇i (t) = ωi (t)
for all i = 1, 2, . . . , n, where (xi (t), yi (t)) are the Cartesian coordinates of robot i at time
t, and θi (t) is its orientation with respect to the x-axis measured in the counter-clockwise
direction. Also, vi (t), the speed of the robot, and ωi (t), its angular velocity, are the control
inputs. Note that model (6.1) is a very common model, and many mobile agents (UGVs, UAVs,
missiles, etc.) can be described by this model [170–175]. Furthermore, we need the following
practical constraints:
− ω max ≤ ωi (t) ≤ ω max
72
∀t ≥ 0
(6.2)
V m ≤ vi (t) ≤ V M
∀t ≥ 0
(6.3)
for all i = 1, 2, . . . , n. Here, ω max > 0 and 0 < V m < V M are given constants.
Moreover, let zi (t) be the vector of the robots’ coordinates and Vi (t) as the robots’ velocity
vector defined by
xi (t)
vi (t)cos(θi (t))
zi (t) :=
, Vi (t) :=
(6.4)
yi (t)
vi (t)sin(θi (t))
for all i = 1, 2, . . . , n.
We assume that the robots share their information via a wireless communication at discrete
time instants k = 0, 1, 2, . . .. Due to limited communication range of the robots, we assume rc
as the communication range for all the mobile robots, meaning that a robot can only receive
information from the robots which are located not farther than rc .
Definition 6.1.1 Robot j is the neighbour of robot i at time k if and only if it is located on the
disk of radius rc with the center of robot i’s position. Also, let N i (k) be the set of all neighbours
of the robot i at time k, and |N i (k)| be the number of elements in N i (k).
The relationship among the robots can be defined by an undirected graph G(k). We assume
that any robot of the multi-robot team is a node of the graph G(k) at time k, i.e., i in VG =
{1, 2, . . . , n}, the node set of G(k), is related to robot i. In addition, robot i is the neighbour
of robot j at time k if and only if there is an edge between the nodes i and j of the graph G(k)
where i 6= j. Therefore, the problem of communication among the robots equals the problem
of the connectivity of the related graph. Note that robot i does not need to be the neighbour of
robot j to get the information from. The information is transferred through the other robots
which connect these robots in the related graph. We will also need the following assumption.
Assumption 6.1.1 There exists an infinite sequence of contiguous, non-empty, bounded timeintervals [kj , kj+1 ), j = 0, 1, 2, . . ., starting at k0 = 0, such that across each [kj , kj+1 ), the union
of the collection {G(k) : k ∈ [kj , kj+1 )} is a connected graph.
To achieve the common heading and speed of formation, we use the consensus variables
θ̃i (k) and ṽi (k), respectively. Also, we need a common origin of coordinates of the formation
for the multi-robot system; therefore, x̃i (k) and ỹi (k) are used as the consensus variables for
the coordinates of the robots. In other words, the robots start with different initial values
of consensus variables x̃i (0), ỹi (0), θ̃i (0) and ṽi (0), and each robot calculates these consensus
variables at any time k such that eventually the consensus variables converge to some consensus
values which define a common speed and orientation in a common coordinate system.
Assumption 6.1.2 The initial values of the consensus variables θ̃i satisfy θ̃i (0) ∈ [0, π) for
all i = 1, 2, . . . , n.
Assumption 6.1.3 The information on the other robots that is available to robot i at time k
is the coordinates (xj (k), yj (k)), and the consensus variables θ̃j (k), x̃j (k), ỹj (k) and ṽj (k) for
all j ∈ Ni (k).
In practice, the coordinates of neighbouring robots can be obtained using Kalman state
estimation via limited capacity communication channels [176].
73
6.2
Formation Building
We propose the following rules for updating the consensus variables θ̃i (k), x̃i (k), ỹi (k) and ṽi (k)
:
P
θ̃i (k) +
θ̃i (k + 1) =
x̃i (k + 1) =
ỹi (k + 1) =
ṽi (k + 1) =
θ̃j (k)
j∈Ni (k)
1 + |Ni (k)|
P
xi (k) + x̃i (k) +
(xj (k) + x̃j (k))
j∈Ni (k)
1 + |Ni (k)|
P
yi (k) + ỹi (k) +
(yj (k) + ỹj (k))
j∈Ni (k)
1 + |Ni (k)|
P
ṽi (k) +
ṽj (k)
− xi (k + 1)
(6.5)
− yi (k + 1)
j∈Ni (k)
1 + |Ni (k)|
Based on rule (6.5), the mobile robots use the consensus variables to achieve a consensus
on the heading, speed and origin of the coordinate system of the formation.
Lemma 6.2.1 Suppose that Assumptions 6.1.1 and 6.1.2 hold, and the consensus variables are
updated according to the decentralized control rule (6.5). Then, there exist constants θ̃0 , X̃0 , Ỹ0
and ṽ0 such that
lim θ̃i (k) = θ̃0
k→∞
lim ṽi (k) = ṽ0
k→∞
lim (xi (k) + x̃i (k)) = X̃0
(6.6)
k→∞
lim (yi (k) + ỹi (k)) = Ỹ0
k→∞
for all i = 1, 2, . . . , n. Furthermore, the convergence in (6.6) is exponentially fast.
The statement of Lemma 6.2.1 immediately follows from the main result of [115]. Note that
the constants θ̃0 , X̃0 , Ỹ0 and ṽ0 are the same for all the robots.
Definition 6.2.1 A navigation law is said to be globally stabilizing with initial conditions
(xi (0), yi (0), θi (0)), i = 1, 2, . . . , n and the given values of configuration C = {X1 , X2 , . . . , Xn , Y1 , Y2 , . . .
if there exists a Cartesian coordinate system and ṽ0 such that the solution of the closed-loop
system (6.1) with these initial conditions and the proposed navigation law in this Cartesian
coordinate system satisfies:
lim (xi (t) − xj (t)) = Xi − Xj
t→∞
lim (yi (t) − yj (t)) = Yi − Yj
t→∞
74
(6.7)
and
lim θi (t) = 0
t→∞
lim vi (t) = ṽ0
(6.8)
t→∞
for all 1 ≤ i 6= j ≤ n. where X1 , X2 , . . . , Xn , Y1 , Y2 , . . . , Yn are given constants.
Rule (6.8) means that as t → ∞ all the robots will finally move in the same direction
along the x-axis with the same speed. Furthermore, rule (6.7) indicates that a geometric
configuration of the robots given by C will be obtained. For instance, if we have four robots
and C = {0, 0, 2, 2, 0, 1, 0, 1}, then the geometric formation of the robots will be a rectangle of
sides 1 and 2.
Since we use the discrete time consensus variables θ̃i (k), x̃i (k), ỹi (k) and ṽi (k) updated according to (6.5), we need to define the corresponding piecewise constant continuous time variables as
θ̃i (t)
x̃i (t)
ỹi (t)
ṽi (t)
:= θ̃i (k) ∀t ∈ (k, k + 1)
:= x̃i (k) ∀t ∈ (k, k + 1)
:= ỹi (k) ∀t ∈ (k, k + 1)
:= ṽi (k) ∀t ∈ (k, k + 1).
(6.9)
For any time t and any robot i, we consider a Cartesian coordinate system with the x-axis
in the direction θ̃i (t) (according to the definition (6.9), θ̃i (t) is piecewise constant). In other
words, in this coordinate system θ̃i (t) = 0 and xi (t), yi (t) are now coordinates of robot i in this
system. Notice that we now formulate our decentralized control law for each robot in its own
coordinate system. Since according to Lemma 6.2.1, θ̃i (k) converges to the same value for all
i, all these robots’ coordinate systems converge to the same coordinate system in which (6.7)
holds.
Assumption 6.2.1 Let c > 0 be any constant such that
c>
2V M
.
ω max
(6.10)
We assume that the constant c and also the configuration C are known to all the robots.
Introduce the functions hi (t) as
h(t) := (xi (t) + x̃i (t)) + Xi + tṽi (t)
(6.11)
for all i = 1,2,. . . , n. Also, introduce two-dimensional vector gi (t) as
x
g (t)
gi (t) : = iy
gi (t)
where
75
(6.12)
Y-Axis
X-Axis
Figure 6.1: Vectors geometry
hi (t) + c if xi (t) ≤ hi (t)
xi (t) + c if xi (t) > hi (t)
gix (t)
:=
giy (t)
:= (yi (t) + ỹi (t)) + Yi
(6.13)
and two-dimensional vector di (t) as
di (t) := gi (t) − zi (t)
(6.14)
for all i = 1,2,. . . , n, where zi (t) is defined by (6.4).
Now, we introduce the following decentralized control law:
vi (t) =
ωi (t) = ω
V M if xi (t) ≤ hi (t)
V m if xi (t) > hi (t)
max
(6.15)
sign(ψi (t))
for all i = 1, 2, . . . , n, where ψi (t) is the angle between Vi (t) and di (t) measured from Vi (t)
in the counter-clockwise direction, i.e.,
ψi (t) = ∠(Vi (t), di (t))
(6.16)
(see Fig. 6.1), and sign(·) is defined by
−1 if α < 0
0
if α = 0
sign(α) :=
1
if α > 0
We also need the following assumption.
76
(6.17)
Assumption 6.2.2 The initial robots’ speeds satisfy
V m < vi (0) < V M
for all i = 1, 2, . . . , n.
Notice that Assumption 6.2.2 is just slightly stronger than the requirement (6.3) for t = 0
where non-strict inequalities are required.
The proposed algorithm is based on robots’ headings and coordinates which, of course,
depend on initial conditions. Therefore, the proposed law depends on initial conditions on
robots’ headings and coordinates. The connectivity of the multi-robot formation is maintained
due to Assumption 6.1.1 which is a standard assumption in numerous papers on multi- agent
systems; see, e.g., [106, 115] and the references therein.
Now, we are in a position to present the main result of this section.
Theorem 6.2.1 Consider the autonomous mobile robots described by the equations (6.1) and
the constraints (6.2), (6.3). Let C = {X1 , X2 , . . . , Xn , Y1 , Y2 , . . . , Yn } be a given configuration. Suppose that Assumptions 6.1.1, 6.1.2, 6.2.1 and 6.2.2 hold. Then, the decentralized
control law (6.5), (6.15) is globally stabilizing with any initial conditions and the configuration
C.
Proof of Theorem 6.2.1: Let 1 ≤ i ≤ n. We consider a fictitious target Ti moving on the
plane with coordinates gi (t) defined by (6.13). Furthermore, introduce another fictitious target
T̃i moving on the plane with coordinates g̃i (t) defined by
x
g̃ (t)
g̃i (t) : = iy
(6.18)
g̃i (t)
where
X0 + Xi + tṽ0 + c if xi (t) ≤ X̃0 + Xi + tṽ0
xi (t) + c if xi (t) > X̃0 + Xi + tṽ0
g̃ix (t)
:=
g̃iy (t)
:= Ỹ0 + Yi
(6.19)
It immediately follows from Lemma 6.2.1 that
lim (g̃i (t) − gi (t)) = 0.
t→∞
(6.20)
Moreover, this convergence is exponentially fast. Let ψi (t) be the angle between the velocity
vector Vi (t) of robot i and the line-of-sight between the robot and Ti ; and βi (t) be the angle
between the velocity vector ViT (t) of Ti and the line-of-sight from robot i to Ti (see Fig.6.1).
It is well-known (see, e.g., [177]) that the following equation holds:
ψ̇i (t) =
where d˜i (t) is defined as
kV T (t)k sin βi (t)
kVi (t)k sin ψi (t)
− ωi (t) − i
kd˜i (t)k
kd˜i (t)k
d˜i (t) := g̃i (t) − zi (t),
(6.21)
(6.22)
zi (t) is defined by (6.4), and k · k denotes the standard Euclidean vector norm. It obviously
follows from (6.19),(6.22) that
77
t
Figure 6.2: Sliding mode solution
kd˜i (t)k ≥ c ∀t ≥ 0.
(6.23)
Furthermore, (6.19) implies that
ViT (t)
ViT x (t)
=
ViT y (t)
= 0.
Tx
(Vi (t)
;
:=
ViT y (t)
ṽ0 if xi (t) ≤ X0 + Xi + tṽ0
vi (t) if xi (t) > X0 + Xi + tṽ0
(6.24)
It follows from (6.24) and (6.3) that
kViT (t)k ≤ V M .
(6.25)
Now, we consider the control law (6.15) with di replaced by d˜i . The inequality (6.25) together
with (6.10), (6.21) and (6.23) implies that under this control law, there exists a constant > 0
such that
ψ̇i (t) < − if ψi (t) > 0
ψ̇i (t) > if ψi (t) < 0.
(6.26)
Therefore, there exists a time τ > 0 such that
ψi (t) = 0 ∀t ≥ τ.
(6.27)
Notice that the closed-loop system (6.1), (6.15) is a system of differential equations with discontinuous right-hand sides; the equation ψi = 0 defines a switching surface of this system, and
a solution satisfying (6.27) is a sliding mode; see, e.g., [178]. Also, (6.5), (6.15) that belongs to
78
the class of switched control laws and the system (6.1) with such controller is a hybrid dynamical system [179–181]. The inequalities (6.26) guarantee that this sliding mode solution of the
closed-loop system looks as it is shown in Fig. 6.2 and satisfies
ψ̇i (t) = 0
∀i ∀t ≥ τ.
(6.28)
From this and (6.21), we obtain that
ωi (t) = −
kViT (t)k sin βi (t)
||d˜i (t)k
(6.29)
for all sliding mode solutions. Therefore, for any initial condition, the sliding mode solution is
unique and well-defined. Furthermore, (6.29), (6.25), (6.10) and (6.23) imply that the constraint
(6.2) holds for any sliding mode solution satisfying (6.27).
Furthermore, the condition (6.27) means that the velocity vector Vi (t) is parallel to the
vector d˜i (t) for all t ≥ τ . Hence, for all t ≥ τ , we have that the robot’s velocity vector is always
pointed at g̃i (t) . Since g̃iy (t) = Ỹ0 + Yi , we obtain that yi (t) → Ỹ0 + Yi . The second of the
conditions (6.7) immediately follows from this. Furthermore, Assumption 6.2.2 implies that
V m ≤ ṽ0 ≤ V M . The fact that the velocity vector Vi (t) is parallel to the vector d˜i (t) for all
t ≥ τ , and the control law (6.15) with di replaced by d˜i imply that
c
˜
di (t) =
0
for all i and all large enough t. The first of the conditions (6.7) immediately follows from this.
We proved the statement of the theorem for the control law (6.15) with di replaced by d˜i . This
and the exponential convergence (6.20) together with the inequality (6.23) imply that the same
statement holds for the original control law (6.15). This completes the proof of Theorem 6.2.1.
Remark 6.2.1 It is evident from the proof of Theorem 6.2.1 that the main idea of the control
law (6.15) can be explained as follows. Each robot i is guided towards a fictitious target Ti that
is always located ahead of the desired robot’s position relative to its neighbours. The reason
we guide the robot towards a fictitious target but not the desired relative robot’s position itself
is clear from (6.29). If we conducted the robot towards the desired relative position, we would
have kdi (t)k → 0; therefore, ωi (t) → ∞ and the constraint (6.2) would be violated. Notice that
our method for guidance towards a fictitious target Ti is a pure pursuit type guidance law (see,
e.g., [182]).
6.3
Formation Building with Anonymous Robots
In the area of robotics, it is common to use the multi-robot task allocation approach to similar
problems. However, most work on multi-robot task allocation has been ad hoc and empirical
especially in the case of an arbitrarily large number of robots; see, e.g., [183,184]. In this section,
we propose a randomized algorithm to handle this problem which leads to a mathematically
rigorous theoretical analysis for any number of robots. In Section 6.2, an algorithm of formation
building for a team of mobile robots was proposed in which positions of all robots are preassigned, i.e., each robot knows a priori its final position in the desired geometric configuration.
79
In this section, we present the algorithm of formation building with anonymous robots meaning
that the robots do not know their final position in the desired geometric configuration at
the beginning but using a randomized algorithm, they eventually reach a consensus on their
positions. In other words, each robot does not know a priori its position in the configuration
C = {X1 , X2 , . . . , Xn , Y1 , Y2 , . . . , Yn }.
Definition 6.3.1 A navigation law is said to be globally stabilizing with anonymous robots
and the configuration C = {X1 , X2 , . . . , Xn , Y1 , Y2 , . . . , Yn } if for any initial conditions
(xi (0), yi (0), θi (0)), there exists a permutation r(i) of the index set {1, 2, . . . , n} such that
for any i = 1, 2, . . . , n, there exist a Cartesian coordinate system and ṽ0 such that the solution
of the closed-loop system (6.1) with the proposed navigation law in this Cartesian coordinate
system satisfies (6.8) and
lim (xi (t) − xj (t)) =Xr(i) − Xr(j)
t→∞
lim (yi (t) − yj (t)) =Yr(i) − Yr(j)
(6.30)
t→∞
for all 1 ≤ i 6= j ≤ n.
Let R > 0 be a given constant. We assume that each robot i has the capacity to detect all
other robots inside the circle of radius R centred at the current position of robot i. Furthermore,
R
let 0 < < be a given constant. For any configuration C = {X1 , X2 , . . . , Xn , Y1 , Y2 , . . . , Yn }
2
introduce a undirected graph P consisting
of n vertices. Vertices i and j of the graph P are
p
connected by an edge if and only if (Xi − Xj )2 + (Yi − Yj )2 ) ≤ R − 2. We will need the
following assumption.
Assumption 6.3.1 The graph P is connected.
We present a randomized algorithm to build an index permutation function r(i) . Let N ≥ 1
be a given integer. Let r(0, i) ∈ {1, 2, . . . , n} be any initial index values where i = 1, 2, . . . , n.
As in the navigation law (6.5), (6.15), for any time t and any robot i, we consider a Cartesian coordinate system with the x-axis in the direction θ̃i (t) (according to the definition (6.9),
θ̃i (t) is piecewise constant). In other words, in this coordinate system θ̃i (t) = 0, xi (t), yi (t) are
now coordinates of robot i in this system. Furthermore, we say that a vertex j of the graph P
is vacant at time kN for robot i if there is no any robot inside the circle of radius centred at
the point
(xi (kN ) + x̃i (kN )) + Xj + kN ṽi (kN )
(yi (kN ) + ỹi (kN )) + Yj
Let S(kN, i) denote the set of vertices of P consisting of r(kN, i) and those of vertices of
P that are connected to r(kN, i) and vacant at time kN for robot i. Let |S(kN, i)| be the
number of elements in S(kN, i). It is clear that 1 ≤ |S(kN, i)| because r(kN, i) ∈ S(kN, i).
Moreover, introduce the Boolean variable bi (kN ) such that bi (kN ) := 1 if there exists another
robot j 6= i that is inside of the circle of radius centred at
(xi (kN ) + x̃i (kN )) + Xi + kN ṽi (kN )
(yi (kN ) + ỹi (kN )) + Yi
80
at time kN , and bi (kN ) := 0 otherwise. We propose the following random algorithm:
r(kN, i) if (bi (kN ) = 0 or (bi (kN ) = 1)
and |S(kN, i)| = 1
r((k + 1)N, i) =
(6.31)
j
if
bi (kN ) = 1
and |S(kN, i)| > 1
Now, we are in a position to present the main result of this section.
Theorem 6.3.1 Consider the autonomous robots described by the equations (6.1) and the constraints (6.2), (6.3). Let C = {X1 , X2 , . . . , Xn , Y1 , Y2 , . . . , Yn } be a given configuration.
Suppose that Assumptions 6.1.1, 6.1.2, 6.2.2 and 6.3.1 hold, and c is a constant satisfying
(6.10). Then, for initial conditions (xi (0), yi (0), θi (0)), i = 1, 2, . . . , n, there exists an integer N0 > 0 such that for any N ≥ N0 , the decentralized control law (6.5), (6.15), (6.31) with
probability 1 is globally stabilizing with these initial conditions and the configuration C.
Proof of Theorem 6.3.1: The algorithm (6.31) defines an absorbing Markov chain which
contains a number of absorbing states that are impossible to leave; in this case, the states when
different robots correspond to different vertices of the desired configuration. It is also obvious
that these absorbing states can be reached from any initial state with a non-zero probability.
It is a well-known theorem of the Markov chain theory that a Markov chain with finite number
of states has, at least, one absorbing state which can be reached from any other states with
non-zero probability. Then, with probability 1, one of the absorbing states will be reached.
This completes the proof of Theorem 6.3.1.
6.4
Obstacle Avoidance
We consider a more challenging problem of navigation of a group of mobile robots for formation
in the existence of obstacles. The map of the environment, information about the obstacles
including their shapes, positions and geometric distribution are not known to the robots a
priori. To detect an obstacle, the robots must be equipped with a range sensor like sonar or
laser. The robots can detect an obstacle when it lies within their range. Then, they obtain
range and angle to the obstacle. The algorithm of obstacle avoidance employs this information
and calculates an appropriate route to avoid collision with the obstacle.
We apply an algorithm for obstacle avoidance that uses angles and distances provided by
range sensors. We assume that the range sensors are located on the robot’s perimeter, in the
forepart with 180◦ field of view; ±90◦ with respect to robot’s heading. Also, we assume that the
maximum range of robots’ range sensors is rs . As shown in Fig. 6.3, a robot moving toward an
obstacle detects the obstacle as soon as the obstacle is placed in the sensing range of the robot.
Then, the robot changes its route to turn the obstacle preserving a distance to it. Suppose
Rt be the turning radius of the robot and d be the distance to the obstacle when the robot’s
heading is parallel to the obstacle surface. Since
Rtmax =
VM
ω min
and
dmin = rs − Rtmax
thus, we need following assumption.
81
d
Figure 6.3: Detecting an obstacle
Figure 6.4: Moving with a constant distance to the obstacle
Assumption 6.4.1 dmin > d0 where d0 is a given constant.
Assume a robot is moving along the circumference of an obstacle (see Fig. 6.4). Also,
suppose that the curvature radius of the obstacle is big enough such that the surface of the
obstacle is assumed flat. As shown in Fig. 6.4, if the robot picks the farthest detectable point
on the obstacle surface using its range sensor as a reference point, there exists an angle between
the robot’s heading and range sensor’s ray is termed as avoiding angle. To have a constant
distance to the obstacle, we need a constant avoiding angle φ0 satisfying d0 = rs sin φ0 .
Now, consider the robot encounters a curved obstacle (see Fig. 6.5). Therefore, the robot
must follow a trajectory preserving the given distance of d0 to the obstacle surface. For instance,
as shown in Fig. 6.5, the robot’s distance to the obstacle is d0 but the range sensor detects
that the avoiding angle φ is greater than φ0 and their difference is ∆φ = |φ − φ0 |. Thus, the
robot must turn in order to remove this gap; by turning equal to ∆φ to the right in this case.
Fig. 6.6 shows the details of obstacle avoidance approach when the obstacle is convex. As
shown in Fig. 6.6, the robot is moving along the surface of the obstacle with avoiding distance
of d0 . As previously mentioned and shown in Fig. 6.4, there is an angle of φ0 between the
82
Figure 6.5: Detecting the curvature of an obstacle
robot’s heading and the range sensor’s ray for a flat surface. However, in order to keep moving
with the avoiding distance of d0 , the robot must turn by ∆φ toward the obstacle. Assume a
fictitious target T , a point with a distance of rs to the robot and angle of ∆φ respect to the
robot’s heading toward the obstacle (see Fig .6.6). As depicted in Fig. 6.6, angles θ0 and θ00
d = CD
d thus
are equal thus the line segments d0 and d00 will be equal. In addition, since AB
]ODB = ]OCD = γ which satisfies that triangles BED and AFC are equal. Therefore, line
segment AF, which is the distance to the obstacle at F, will be equal to BE= d0 . It means that
if point C is selected as the fictitious target, the distance to the obstacle will be maintained to
a given constant.
Fig. 6.7 shows the case that the obstacle is concave. This case is similar to the convex case
except the fictitious target that is away from the obstacle; therefore, the robot must turn by
∆φ away from the obstacle.
As a result, we propose the following control law that enables robots to avoid a collision by
calculating a smooth path around the obstacles.
vi (t) = V M
ωi (t) = ω max sign(ψi (t))
(6.32)
for all i = 1, 2, . . . , n, where
If φ < φ0
1
0
If φ = φ0
ψi (t) =
−1 If φ > φ0
also V M , ω max and sign(.) are given in (6.2), (6.3) and (6.17), respectively.
Now, we are in a position to present the main results of this chapter.
83
(6.33)
E
O
D
F
C
B
A
Figure 6.6: A convex obstacle
Theorem 6.4.1 Consider the autonomous mobile robots described by the equations (6.1) and
the constraints (6.2), (6.3). Let C = {X1 , X2 , . . . , Xn , Y1 , Y2 , . . . , Yn } be a given configuration. Suppose that Assumptions 6.1.1, 6.1.2, 6.2.1 and 6.2.2 hold, and c is a constant satisfying
(6.10). Then, the distributed control law (6.5), (6.15), (6.32) is globally stabilizing with any
initial conditions and the configuration C.
Proof of Theorem 6.4.1: proof of Theorem 6.4.1 is completely similar to the proof
of Theorem 6.2.1. Both control laws, (6.15) for formation building and (6.32) for obstacle
avoidance are the same. The main difference is that the fictitious target T in this case is
variable between (6.12) and what is defined in this section. In other words, whenever a robot
encounters an obstacle, the fictitious target switches from (6.12) to a point with a distance of
rs to the robot and angle of ∆φ respect to the robot’s heading toward the obstacle (point C in
Fig. 6.6 and Fig. 6.7).
6.5
Simulation Results
We present computer simulation results for all algorithms proposed in this chapter: obstacle
avoidance, formation building with obstacle avoidance and anonymous formation building with
obstacle avoidance. We use Mobotsim 1.0 simulator, a powerful 2D simulator of mobile robots
that simulates robots’ motion, environment and range sensors like sonar. Simulation parameters
are given in Table 6.1.
84
F
C
E
O
D
A
B
Figure 6.7: A concave obstacle
Table 6.1: Simulation Paremeters
Parameter
Sampling Intervals
Robot’s Platform Diameter
Distance Between Wheels
Wheels Diameter
Maximum Angular Velocity
Maximum Linear Velocity
Minimum Linear Velocity
Sonars Maximum Range
Number of Ranging Sonars
Sonars’ Radiation Cone
6.5.1
Value
0.1
0.5
0.35
0.2
2
1.5
.2
2
12
15
Comment
s
meter
meter
meter
rad/s
m/s
m/s
meter
degree
Obstacle Avoidance
First, we present the simulation results for the proposed obstacle avoidance rule (6.32). Fig.
6.8 shows the simulation results for obstacle avoidance rule (6.32) with some different obstacles.
As Fig. 6.8 displays, by applying the proposed obstacle avoidance rule, the robots successfully
bypass the obstacles with different shapes and sizes.
85
(a)
(b)
(c)
(d)
Figure 6.8: Applying the obstacle avoidance rule; robots pass the obstacles
6.5.2
Formation Building with Obstacle Avoidance
To simulate the algorithm of formation building with obstacle avoidance, we consider a team
consisting of five robots randomly located on the plane with different headings. The goal is to
build a formation as well as avoiding the obstacles that might obstruct robots’ movement. The
robots are to build the edge ( ’>’) by applying the proposed algorithm of formation building in
Section 6.2 and the obstacle avoidance rule in Section 6.4. First, we assume that there is not
any obstacle; therefore, only the formation building rule of the proposed algorithm is used. As
depicted in Fig. 6.9, the robots build the desired formation (’>’).
Second, we assume the same problem but this time with an obstacle. The simulation results
of applying the proposed algorithm are displayed in Fig. 6.10. As shown in Fig. 6.10(a), the
robots build the desired formation before they encounter the obstacle, and move such that the
formation configuration holds. When the robots detect an obstacle on their direction, they
avoid the obstacle by turning around. Fig. 6.10(b) shows the snapshot of this phase. Passing
the obstacle, the robots restart the formation building phase and as Fig. 6.10(c) shows, the
desired formation is built again. Note that as shown in Fig. 6.10, it is not necessary for the
robots to pass the obstacle all together and then start the formation building, e.g., while one
robot is still in the obstacle avoidance phase, the other robots that have passed the obstacle
begin the formation building phase again.
86
Figure 6.9: Robots form the desired pattern without any obstacles
(a)
(b)
(c)
Figure 6.10: Robots pass the obstacle and build the desired form
To confirm that the proposed algorithm is effective even with any number of obstacles
with different shapes and sizes, more simulations are fulfilled. Fig. 6.11 shows the results of
87
(a)
(b)
(c)
Figure 6.11: Robots form the desired patterns and avoid obstacles
88
these simulations in which more obstacles with different shapes and sizes are used, and the
robots build various formations. In Fig. 6.11(a), the robots build an edge (’>’) while avoiding
the obstacles on their routes. In Fig. 6.11 (b) and (c), they form shapes of a line and an
arc, respectively. The results confirm that the proposed algorithm is effective even with any
number of obstacles with different shapes and sizes. It should be pointed out that the proposed
obstacle avoidance algorithm prevents the collision between robots too; as a robot considers
another robot in its sensing range as an obstacle.
6.5.3
Formation Building with Anonymous Robots and Obstacle
Avoidance
In Section 6.5.2, simulation results for formation building with obstacle avoidance have been
presented. As previously explained, in the algorithm of Section 6.2, positions of all the robots
are pre-assigned, i.e., each robot knows its position in the final formation building. Therefore,
the positions of the robots do not change during the formation process; as it has been shown
in Fig. 6.11 when positions of the robots are invariant before and after passing the obstacles.
On the other hand, applying the algorithm of Section 6.3, result in altering the position of the
robots during the formation process. Simulation results for the algorithm of formation building
with anonymous robots presented in Section 6.3, are given in Fig. 6.12. In Fig. 6.12(a), robots
form an edge (’>’) before encountering the obstacles, and Fig. 6.12(b) shows the positions of
the robots after passing the obstacles. As shown in Fig. 6.12(b), the positions of the robots after
passing the obstacles are not the same as before; robots make the same formation building but
with a different arrangement of the robots. Fig. 6.12(c-f), show something like that for line ’|’
and arc ’(’ formation buildings. Fig. 6.13 shows a comparison between the formation building
algorithms, with and without anonymous robots. As depicted in Fig. 6.13(a), the positions of
the robots in the formation building are the same before and after passing the obstacles while
in Fig. 6.13(b) where the algorithm of anonymous formation building is applied, the positions
of the robots change.
6.6
Summary
The problem of formation building with obstacle avoidance for a team of mobile robots have
been considered. The algorithm of global formation building has been combined with a local
obstacle avoidance algorithm. We have proposed a distributed motion coordination control
algorithm so that the robots collectively move in a desired geometric pattern from any initial
position while avoiding the obstacles on their way. We have considered unicycles with standard
kinematic equations and hard constraints on their linear and angular velocities for the type
of the robots. A consensus variables rule has been used for the formation building phase that
is based on the local information. Also, a novel technique based on the information from
the range sensors have been employed for the obstacle avoidance phase. Furthermore, we
propose a randomized algorithm for the anonymous robots which achieves the convergence to
the desired configuration with probability 1. Mathematically rigorous proofs of the proposed
control algorithms have been given, and the effectiveness of the algorithms have been confirmed
via computer simulations.
89
(a)
(b)
(c)
(d)
(e)
(f)
Figure 6.12: Robots form (a),(b) an edge; (c),(d) a line and (e),(f) an arc
90
(a) Formation building without anonymous robots
(b) Formation building with anonymous robots
Figure 6.13: A comparison between formation building with and without anonymous robots
91
Chapter 7
Conclusions
The main purpose of this report was to design some algorithms for search by multi-robot
systems. We assumed an unknown area including some obstacles to be searched by a team
of autonomous mobile robots either partially for a given number of targets or entirely for
the unknown number of targets. For that purpose, we developed three decentralized control
algorithms to drive a multi-robot team to explore unknown environments. We used a triangular
grid pattern and two-stage algorithms for the control law so that robots move through the
vertices of the grid during the search procedure. In the first stage of the proposed search
algorithms, using a consensus variables rule, the robots are located on the vertices of a triangular
grid. For the second stage of the algorithm, three different scenarios were considered. First, a
pure random grid-based algorithm described in Chapter 3 was presented, by which the robots
randomly move between the vertices of a common triangular grid so that in each step they
only move to the one of the closest neighbouring vertices. Note that there are at most six
closest neighbouring vertices for each vertex. In the second scenario, presented in Chapter 4,
we changed the pure random rule to a semi-random rule. In this case, the robots still randomly
move between the vertices of a common triangular grid so that in each step they move to the
one of the closest neighbouring vertices; but, only to those vertices which have not been visited
by the robots yet. If all the (at most) six neighbouring vertices have been visited already, one
of them will randomly be selected. Finally, a modified algorithm was proposed in Chapter 5.
This algorithm does not confine the robots to move only to the closest neighbouring vertices;
but, they move to the nearest unvisited vertex anywhere in the search area.
It has been shown that a triangular grid pattern is asymptotically optimal in terms of
the minimum number of robots required for the complete coverage of an arbitrary bounded
area. That is we employed a triangular grid pattern for the proposed algorithms, i.e., robots
certainly go through the vertices of a triangular grid during the search operation. Therefore,
using the vertices of a triangular grid coverage guarantees complete search of the whole area as
well as better performance in terms of search time. Furthermore, we presented a new kind of
topological map which robots make and share during the search operation. Unlike many other
hubristic algorithms in this area, we gave mathematically rigorous proofs of convergence with
probability 1 of the proposed algorithms.
The procedures of this study were approved by computer simulation results using a simulator
of real robots and environments. To evaluate the performance of the algorithms, we presented
the experiment results with real Pioneer 3DX mobile robots for one of the algorithms with
detailed descriptions and explanations. The results demonstrated the features of the proposed
algorithms and their performance with real systems. Moreover, we compared the proposed
92
algorithms with each other and also with three other algorithms from other researchers. The
comparison showed the strength of our proposed algorithms over the other existing algorithms.
Also, a further study on networked multi-robot formation building algorithms was presented
in this report. The problem of formation building for a group of mobile robots was considered.
A decentralized formation building with obstacle avoidance algorithm for a group of mobile
robots to move in a defined geometric configuration was proposed. Furthermore, we considered
a more complicated formation problem with a group of anonymous robots where the robots
are not aware of their position in the final configuration and have to reach a consensus during
the formation process while avoiding obstacles. We proposed a randomized algorithm for the
anonymous robots which achieves the convergence to the desired configuration with probability
1. Moreover, we presented a novel obstacle avoidance rule which was employed in the formation
building algorithms. We demonstrated mathematically rigorous proofs of convergence of the
presented algorithms. Also, we confirmed the performance and applicability of the proposed
algorithms by computer simulation results.
Future Work
In terms of directions for future research, further work could be as follows:
• It is an interesting direction for future research to apply the proposed search algorithms
to swarm systems. In that case, it could also be conducted to determine the effectiveness
of limited wireless communication and memory resources [185, 186].
• During the experiments, no visible or severe drift on wheel odometry was observed; but,
it can be significant if the number of vertices of the grid is large. Addressing this problem
is a direction for our future work [187, 188].
• We have assumed static obstacles in the workspace. For a real application, more challenges
may appear with moving obstacles in the environment that would be a fruitful area for
further work [189, 190].
• In the proposed search algorithms, we have assumed static targets. It is recommended
that further research be undertaken with moving targets [191].
• In Chapter 6, we proposed new strategies in formation control with obstacle avoidance of
autonomous robots and presented computer simulation results. It would be interesting to
investigate and verify the effectiveness of the proposed algorithms through experiments
with real robots.
• Regarding the proposed formation building algorithms, the future work can be modifying
the proposed algorithms so that the formation holds while passing the obstacles [192].
• Another possible area of future research would be to consider the problem of environmental extremum seeking by multi-robot teams using the algorithms presented in this report
for search and formation [193–196].
93
Bibliography
[1] H. H. Viet, V.-H. Dang, M. N. U. Laskar, and T. Chung, “Ba*: an online complete
coverage algorithm for cleaning robots,” Applied Intelligence, vol. 39, no. 2, pp. 217–235,
2013.
[2] J. Hess, M. Beinhofer, and W. Burgard, “A probabilistic approach to high-confidence
cleaning guarantees for low-cost cleaning robots,” in Proceedings of the 2014 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2014, pp. 5600–5605.
[3] A. Marjovi and L. Marques, “Multi-robot olfactory search in structured environments,”
Robotics and Autonomous Systems, vol. 59, no. 11, pp. 867–881, 2011.
[4] F. Kendoul, “Survey of advances in guidance, navigation, and control of unmanned rotorcraft systems,” Journal of Field Robotics, vol. 29, no. 2, pp. 315–378, 2012.
[5] J. Linchant, J. Lisein, J. Semeki, P. Lejeune, and C. Vermeulen, “Are unmanned aircraft
systems (uass) the future of wildlife monitoring? a review of accomplishments and
challenges,” Mammal Review, vol. 45, no. 4, pp. 239–252, 2015. [Online]. Available:
http://dx.doi.org/10.1111/mam.12046
[6] C. Zhang and J. M. Kovacs, “The application of small unmanned aerial systems for
precision agriculture: a review,” Precision Agriculture, vol. 13, no. 6, pp. 693–712, 2012.
[7] Q.-V. Dang, I. Nielsen, K. Steger-Jensen, and O. Madsen, “Scheduling a single mobile
robot for part-feeding tasks of production lines,” Journal of Intelligent Manufacturing,
vol. 25, no. 6, pp. 1271–1287, 2014.
[8] E. Guizzo and E. Ackerman, “The rise of the robot worker,” IEEE Spectrum, vol. 49,
no. 10, pp. 34–41, 2012.
[9] H. Chi, K. Zhan, and B. Shi, “Automatic guidance of underground mining vehicles using
laser sensors,” Tunnelling and Underground Space Technology, vol. 27, no. 1, pp. 142–148,
2012.
[10] M. Dunn, D. Reid, and J. Ralston, “Control of automated mining machinery using aided
inertial navigation,” in Machine Vision and Mechatronics in Practice. Springer, 2015,
pp. 1–9.
[11] M. Subhan, A. Bhide, and B. SSGB COE, “Study of unmanned vehicle (robot) for coal
mines,” International Journal of Innovative Research in Advanced Engineering (IJIRAE),
vol. 1, no. 10, pp. 116–120, 2014.
94
[12] U. Reiser, T. Jacobs, G. Arbeiter, C. Parlitz, and K. Dautenhahn, “Care-o-bot R 3–vision
of a robot butler,” Your Virtual Butler, pp. 97–116, 2013.
[13] C. Wang, A. V. Savkin, R. Clout, and H. T. Nguyen, “An intelligent robotic hospital bed
for safe transportation of critical neurosurgery patients along crowded hospital corridors,”
IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 23, no. 5, pp.
744–754, 2015.
[14] S. Satake, K. Hayashi, K. Nakatani, and T. Kanda, “Field trial of an informationproviding robot in a shopping mall,” in Proceedings of the 2015 IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS). IEEE, 2015, pp. 1832–1839.
[15] M. Guarnieri, R. Kurazume, H. Masuda, T. Inoh, K. Takita, P. Debenest, R. Hodoshima,
E. Fukushima, and S. Hirose, “Helios system: A team of tracked robots for special urban search and rescue operations,” in Proceedings of the 2009 IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS). IEEE, 2009, pp. 2795–2800.
[16] G.-J. M. Kruijff, M. Janı́ček, S. Keshavdas, B. Larochelle, H. Zender, N. J. Smets,
T. Mioch, M. A. Neerincx, J. V. Diggelen, F. Colas et al., “Experience in system design for human-robot teaming in urban search and rescue,” in Field and Service Robotics,
2014, pp. 111–125.
[17] A. Marjovi, J. G. Nunes, L. Marques, and A. de Almeida, “Multi-robot fire searching in
unknown environment,” in Field and Service Robotics, 2010, pp. 341–351.
[18] G. Tuna, V. C. Gungor, and K. Gulez, “An autonomous wireless sensor network deployment system using mobile robots for human existence detection in case of disasters,” Ad
Hoc Networks, vol. 13, pp. 54–68, 2014.
[19] B. Liu, O. Dousse, P. Nain, and D. Towsley, “Dynamic coverage of mobile sensor networks,” IEEE Transactions on Parallel and Distributed Systems, vol. 24, no. 2, pp. 301–
311, 2013.
[20] P. Stone and M. Veloso, “Multiagent systems: A survey from a machine learning perspective,” Autonomous Robots, vol. 8, no. 3, pp. 345–383, 2000.
[21] “Mars pathfider,” http://mars.nasa.gov/MPF/index1.html, accessed: 2016-01-17.
[22] “Sony aibo,” http://www.sony-aibo.com, accessed: 2016-02-12.
[23] “Bigdog-the
most
advanced
rough-terrain
robot
on
http://www.bostondynamics.com/robot bigdog.html, accessed: 2016-02-01.
earth,”
[24] G. Dudek, M. R. Jenkin, E. Milios, and D. Wilkes, “A taxonomy for multi-agent robotics,”
Autonomous Robots, vol. 3, no. 4, pp. 375–397, 1996.
[25] V. Zadorozhny and M. Lewis, “Information fusion based on collective intelligence for
multi-robot search and rescue missions,” in Proceedings of the 14th IEEE International
Conference on Mobile Data Management. IEEE, 2013, pp. 275–278.
95
[26] T. Pereira, A. P. Moreira, and M. Veloso, “Coordination for multi-robot exploration
using topological maps,” in Proceedings of the 11th Portuguese Conference on Automatic
Control (CONTROLO). Springer, 2015, pp. 515–524.
[27] F. F. Carvalho, R. C. Cavalcante, M. Vieira, L. Chaimowicz, M. F. Campos et al., “A
multi-robot exploration approach based on distributed graph coloring,” in Proceedings of
the 2013 Latin American Robotics Symposium and Competition (LARS/LARC). IEEE,
2013, pp. 142–147.
[28] S. Sharma, C. Sur, A. Shukla, and R. Tiwari, “Multi-robot area exploration using particle
swarm optimization with the help of cbdf-based robot scattering,” in Computational
Vision and Robotics. Springer, 2015, pp. 113–123.
[29] Z. Yan, N. Jouandeau, and A. A. Cherif, “A survey and analysis of multi-robot coordination,” International Journal of Advanced Robotic Systems, vol. 10, 2013.
[30] S. I. Roumeliotis and G. A. Bekey, “Distributed multirobot localization,” IEEE Transactions on Robotics and Automation, vol. 18, no. 5, pp. 781–795, 2002.
[31] D. Fox, W. Burgard, H. Kruppa, and S. Thrun, “A probabilistic approach to collaborative
multi-robot localization,” Autonomous Robots, vol. 8, no. 3, pp. 325–344, 2000.
[32] J. Wawerla and R. T. Vaughan, “A fast and frugal method for team-task allocation
in a multi-robot transportation system,” in Proceedings of the 2010 IEEE International
Conference on Robotics and Automation (ICRA). IEEE, 2010, pp. 1432–1437.
[33] A. Prorok, A. Bahr, and A. Martinoli, “Low-cost collaborative localization for largescale multi-robot systems,” in Proceedings of the 2012 IEEE International Conference on
Robotics and Automation (ICRA). Ieee, 2012, pp. 4236–4241.
[34] N. Agmon, S. Kraus, and G. A. Kaminka, “Multi-robot perimeter patrol in adversarial
settings,” in Proceedings of the 2008 IEEE International Conference on Robotics and
Automation (ICRA). IEEE, 2008, pp. 2339–2345.
[35] L. Chaimowicz, A. Cowley, D. Gomez-Ibanez, B. Grocholsky, M. Hsieh, H. Hsu, J. Keller,
V. Kumar, R. Swaminathan, and C. Taylor, “Deploying air-ground multi-robot teams in
urban environments,” in Multi-Robot Systems. From Swarms to Intelligent Automata
Volume III. Springer, 2005, pp. 223–234.
[36] K. M. Wurm, C. Stachniss, and W. Burgard, “Coordinated multi-robot exploration using
a segmentation of the environment,” in Proceedings of the 2008 IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS). IEEE, 2008, pp. 1160–1165.
[37] A. Baranzadeh, “A decentralized control algorithm for target search by a multi-robot
team,” in Proceedings of the 2013 Australasian Conference on Robotics and Automation
(ACRA). ARAA, 2013.
[38] V. Nazarzehi and A. Baranzadeh, “A decentralized grid-based random search algorithm
for locating targets in three dimensional environments by a mobile robotic network,” in
Proceedings of the 2015 Australasian Conference on Robotics and Automation (ACRA).
ARAA, 2015.
96
[39] ——, “A distributed bio-inspired algorithm for search of moving targets in three dimensional spaces,” in Proceedings of the 2015 IEEE International Conference on Robotics
and Biomimetics (ROBIO). IEEE, 2015, pp. 2507–2512.
[40] A. Baranzadeh and A. V. Savkin, “A distributed algorithm for grid-based search by
a multi-robot system,” in Proceedings of the 10th Asian Control Conference (ASCC).
IEEE, 2015, pp. 1–6.
[41] ——, “A distributed control algorithm for area search by a multi-robot team,” Robotica,
2016 (Accepted).
[42] H. Sugiyama, T. Tsujioka, and M. Murata, “Integrated operations of multi-robot rescue
system with ad hoc networking,” in Proceedings of the 1st International Conference on
Wireless Communication, Vehicular Technology, Information Theory and Aerospace &
Electronic Systems Technology (Wireless VITAE). IEEE, 2009, pp. 535–539.
[43] D. Portugal and R. Rocha, “A survey on multi-robot patrolling algorithms,” in Technological Innovation for Sustainability. Springer, 2011, pp. 139–146.
[44] A. S. Matveev, H. Teimoori, and A. V. Savkin, “A method for guidance and control of an
autonomous vehicle in problems of border patrolling and obstacle avoidance,” Automatica,
vol. 47, no. 3, pp. 515–524, 2011.
[45] A. Howard, L. E. Parker, and G. S. Sukhatme, “Experiments with a large heterogeneous
mobile robot team: Exploration, mapping, deployment and detection,” The International
Journal of Robotics Research, vol. 25, no. 5-6, pp. 431–447, 2006.
[46] A. Fagiolini, M. Pellinacci, G. Valenti, G. Dini, and A. Bicchi, “Consensus-based distributed intrusion detection for multi-robot systems,” in Proceedings of the 2008 IEEE
International Conference on Robotics and Automation (ICRA). IEEE, 2008, pp. 120–
127.
[47] S. Tadokoro, Rescue robotics: DDT project on robots and systems for urban search and
rescue. Springer Science & Business Media, 2009.
[48] R. R. Murphy, J. Kravitz, S. Stover, and R. Shoureshi, “Mobile robots in mine rescue
and recovery,” IEEE Robotics & Automation Magazine, vol. 16, no. 2, pp. 91–103, 2009.
[49] R. W. Beard and T. W. McLain, “Multiple uav cooperative search under collision avoidance and limited range communication constraints,” in Proceedings of 42nd IEEE Conference on Decision and Control, vol. 1. IEEE, 2003, pp. 25–30.
[50] P. Vincent and I. Rubin, “A framework and analysis for cooperative search using uav
swarms,” in Proceedings of the 2004 ACM Symposium on Applied Computing. ACM,
2004, pp. 79–86.
[51] Y. Yang, M. M. Polycarpou, and A. A. Minai, “Multi-uav cooperative search using an opportunistic learning method,” Journal of Dynamic Systems, Measurement, and Control,
vol. 129, no. 5, pp. 716–728, 2007.
97
[52] J. L. Baxter, E. Burke, J. M. Garibaldi, and M. Norman, “Multi-robot search and rescue:
A potential field based approach,” in Autonomous Robots and Agents. Springer, 2007,
pp. 9–16.
[53] A. Marjovi, J. Nunes, L. Marques, and A. de Almeida, “Multi-robot exploration and fire
searching,” in Proceedings of the 2009 IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS). IEEE, 2009, pp. 1929–1934.
[54] H. Sugiyama, T. Tsujioka, and M. Murata, “Autonomous chain network formation by
multi-robot rescue system with ad hoc networking,” in Proceedings of the 2010 IEEE
International Workshop on Safety Security and Rescue Robotics (SSRR). IEEE, 2010,
pp. 1–6.
[55] C. Luo, A. P. Espinosa, D. Pranantha, and A. De Gloria, “Multi-robot search and rescue
team,” in Proceedings of the 2011 IEEE International Symposium on Safety, Security,
and Rescue Robotics (SSRR). IEEE, 2011, pp. 296–301.
[56] M. Lewis and K. Sycara, “Network-centric control for multirobot teams in urban search
and rescue,” in Proceedings of the 44th Hawaii International Conference on System Sciences (HICSS). IEEE, 2011, pp. 1–10.
[57] A. Macwan, G. Nejat, and B. Benhabib, “Target-motion prediction for robotic search
and rescue in wilderness environments,” IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics, vol. 41, no. 5, pp. 1287–1298, 2011.
[58] B. Mobedi and G. Nejat, “3-d active sensing in time-critical urban search and rescue
missions,” IEEE/ASME Transactions on Mechatronics, vol. 17, no. 6, pp. 1111–1119,
2012.
[59] H. Sugiyama, T. Tsujioka, and M. Murata, “Real-time exploration of a multi-robot rescue
system in disaster areas,” Advanced Robotics, vol. 27, no. 17, pp. 1313–1323, 2013.
[60] Y. Liu, G. Nejat, and J. Vilela, “Learning to cooperate together: A semi-autonomous
control architecture for multi-robot teams in urban search and rescue,” in Proceedings
of the 2013 IEEE International Symposium on Safety, Security, and Rescue Robotics
(SSRR). IEEE, 2013, pp. 1–6.
[61] R. Cipolleschi, M. Giusto, A. Q. Li, and F. Amigoni, “Semantically-informed coordinated
multirobot exploration of relevant areas in search and rescue settings,” in Proceedings of
the 2013 European Conference on Mobile Robots (ECMR). IEEE, 2013, pp. 216–221.
[62] F. Amigoni, N. Basilico, and A. Q. Li, “How much worth is coordination of mobile robots
for exploration in search and rescue?” in RoboCup 2012: Robot Soccer World Cup XVI.
Springer, 2013, pp. 106–117.
[63] S. V. Spires and S. Y. Goldsmith, “Exhaustive geographic search with mobile robots
along space-filling curves,” in Collective Robotics. Springer, 1998, pp. 1–12.
[64] D. Enns, D. Bugajski, and S. Pratt, “Guidance and control for cooperative search,” in
Proceedings of the 2002 American Control Conference (ACC), vol. 3. IEEE, 2002, pp.
1923–1929.
98
[65] B. A. Bash and P. J. Desnoyers, “Exact distributed voronoi cell computation in sensor
networks,” in Proceedings of the 6th International Conference on Information Processing
in Sensor Networks. ACM, 2007, pp. 236–243.
[66] L. Wu, M. Á. Garcı́a Garcı́a, D. Puig Valls, and A. Solé Ribalta, “Voronoi-based space
partitioning for coordinated multi-robot exploration,” vol. 1, pp. 37–44, 2007.
[67] F. Bullo, J. Cortés, and S. Martinez, Distributed control of robotic networks: a mathematical approach to motion coordination algorithms. Princeton University Press, 2009.
[68] A. D. Haumann, K. D. Listmann, and V. Willert, “Discoverage: A new paradigm for
multi-robot exploration,” in Proceedings of the 2010 IEEE International Conference on
Robotics and Automation (ICRA). IEEE, 2010, pp. 929–934.
[69] A. Cowley, C. J. Taylor, and B. Southall, “Rapid multi-robot exploration with topometric maps,” in Proceedings of the 2011 IEEE International Conference on Robotics and
Automation (ICRA). IEEE, 2011, pp. 1044–1049.
[70] S. Bhattacharya, N. Michael, and V. Kumar, “Distributed coverage and exploration in unknown non-convex environments,” in Distributed Autonomous Robotic Systems. Springer,
2013, pp. 61–75.
[71] B. Yang, Y. Ding, Y. Jin, and K. Hao, “Self-organized swarm robot for target search and
trapping inspired by bacterial chemotaxis,” Robotics and Autonomous Systems, vol. 72,
pp. 83–92, 2015.
[72] K. Guruprasad and D. Ghose, “Automated multi-agent search using centroidal voronoi
configuration,” IEEE Transactions on Automation Science and Engineering, vol. 8, no. 2,
pp. 420–423, 2011.
[73] ——, “Performance of a class of multi-robot deploy and search strategies based on centroidal voronoi configurations,” International Journal of Systems Science, vol. 44, no. 4,
pp. 680–699, 2013.
[74] ——, “Heterogeneous locational optimisation using a generalised voronoi partition,” International Journal of Control, vol. 86, no. 6, pp. 977–993, 2013.
[75] R. Zlot, A. Stentz, M. Dias, and S. Thayer, “Multi-robot exploration controlled by a
market economy,” in Proceedings of the 2002 IEEE International Conference on Robotics
and Automation (ICRA), vol. 3. IEEE, 2002, pp. 3016–3023.
[76] J. Fink, M. A. Hsieh, and V. Kumar, “Multi-robot manipulation via caging in environments with obstacles,” in Proceedings of the 2008 IEEE International Conference on
Robotics and Automation (ICRA). IEEE, 2008, pp. 1471–1476.
[77] K. Derr and M. Manic, “Multi-robot, multi-target particle swarm optimization search in
noisy wireless environments,” in Proceedings of the 2nd Conference on Human System
Interactions (HSI’09). IEEE, 2009, pp. 81–86.
[78] Z. Li and Z. Duan, Cooperative Control of Multi-Agent Systems: A Consensus Region
Approach. CRC Press, 2014.
99
[79] A. Okubo, “Dynamical aspects of animal grouping: swarms, schools, flocks, and herds,”
Advances in Biophysics, vol. 22, pp. 1–94, 1986.
[80] C. W. Reynolds, “Flocks, herds and schools: A distributed behavioral model,” vol. 21,
no. 4, pp. 25–34, 1987.
[81] A. Howard, M. J. Matarić, and G. S. Sukhatme, “Mobile sensor network deployment
using potential fields: A distributed, scalable solution to the area coverage problem,”
in Proceedings of the 6th International Symposium on Distributed Autonomous Robotics
Systems. Springer, 2002, pp. 299–308.
[82] F. Sivrikaya and B. Yener, “Time synchronization in sensor networks: a survey,” Network,
IEEE, vol. 18, no. 4, pp. 45–50, 2004.
[83] R. W. Beard, J. Lawton, F. Y. Hadaegh et al., “A coordination architecture for spacecraft
formation control,” IEEE Transactions on Control Systems Technology, vol. 9, no. 6, pp.
777–790, 2001.
[84] M. Aung, A. Ahmed, M. Wette, D. Scharf, J. Tien, G. Purcell, M. Regehr, and B. Landin,
“An overview of formation flying technology development for the terrestrial planet finder
mission,” in Proceedings of the 2004 IEEE Aerospace Conference, vol. 4. IEEE, 2004,
pp. 2667–2679.
[85] C. Candea, H. Hu, L. Iocchi, D. Nardi, and M. Piaggio, “Coordination in multi-agent
robocup teams,” Robotics and Autonomous Systems, vol. 36, no. 2, pp. 67–86, 2001.
[86] M. Mesbahi and M. Egerstedt, Graph theoretic methods in multiagent networks. Princeton University Press, 2010.
[87] F. L. Lewis, H. Zhang, K. Hengster-Movric, and A. Das, Cooperative control of multiagent systems: optimal and adaptive design approaches. Springer Science & Business
Media, 2013.
[88] H. Bai, M. Arcak, and J. Wen, Cooperative control design: a systematic, passivity-based
approach. Springer Science & Business Media, 2011.
[89] Y. Cao, W. Yu, W. Ren, and G. Chen, “An overview of recent progress in the study
of distributed multi-agent coordination,” IEEE Transactions on Industrial Informatics,
vol. 9, no. 1, pp. 427–438, 2013.
[90] G. Antonelli, “Interconnected dynamic systems: An overview on distributed control,”
Control Systems, vol. 33, no. 1, pp. 76–88, 2013.
[91] X. Jia and M. Q.-H. Meng, “A survey and analysis of task allocation algorithms in multirobot systems,” in Proceedings of the 2013 IEEE International Conference on Robotics
and Biomimetics (ROBIO). IEEE, 2013, pp. 2280–2285.
[92] J. P. Hespanha, P. Naghshtabrizi, and Y. Xu, “A survey of recent results in networked
control systems,” IEEE Proceedings, vol. 95, no. 1, p. 138, 2007.
[93] A. S. Matveev and A. V. Savkin, Estimation and control over communication networks.
Springer Science & Business Media, 2009.
100
[94] ——, “Optimal state estimation in networked systems with asynchronous communication
channels and switched sensors,” in Proceedings of the 40th IEEE Conference on Decision
and Control, vol. 1. IEEE, 2001, pp. 825–830.
[95] ——, “The problem of state estimation via asynchronous communication channels with
irregular transmission times,” IEEE Transactions on Automatic Control, vol. 48, no. 4,
pp. 670–676, 2003.
[96] ——, “Comments on” control over noisy channels” and relevant negative results,” IEEE
Transactions on Automatic Control, vol. 50, no. 12, pp. 2105–2110, 2005.
[97] A. V. Savkin, “Analysis and synthesis of networked control systems: Topological entropy,
observability, robustness and optimal control,” Automatica, vol. 42, no. 1, pp. 51–62,
2006.
[98] M. O. F. Sarker, T. S. Dahl, E. Arcaute, and K. Christensen, “Local interactions over
global broadcasts for improved task allocation in self-organized multi-robot systems,”
Robotics and Autonomous Systems, vol. 62, no. 10, pp. 1453–1462, 2014.
[99] R. Luna and K. E. Bekris, “Efficient and complete centralized multi-robot path planning,”
in Proceedings of the 2011 IEEE/RSJ International Conference on Intelligent Robots and
Systems (IROS). IEEE, 2011, pp. 3268–3275.
[100] T. M. Cheng and A. V. Savkin, “Decentralized control for mobile robotic sensor network
self-deployment: barrier and sweep coverage problems,” Robotica, vol. 29, no. 2, pp.
283–294, 2011.
[101] G. Pini, A. Brutschy, M. Frison, A. Roli, M. Dorigo, and M. Birattari, “Task partitioning
in swarms of robots: An adaptive method for strategy selection,” Swarm Intelligence,
vol. 5, no. 3-4, pp. 283–304, 2011.
[102] K. Choi, S. J. Yoo, J. B. Park, and Y. H. Choi, “Adaptive formation control in absence
of leader’s velocity information,” IET Control Theory & Applications, vol. 4, no. 4, pp.
521–528, 2010.
[103] S. P. Hou and C. C. Cheah, “Dynamic compound shape control of robot swarm,” IET
Control Theory & Applications, vol. 6, no. 3, pp. 454–460, 2012.
[104] P. Chand and D. A. Carnegie, “A two-tiered global path planning strategy for limited
memory mobile robots,” Robotics and Autonomous Systems, vol. 60, no. 2, pp. 309–321,
2012.
[105] P. Chand and D. A. Carnegie, “Mapping and exploration in a hierarchical heterogeneous
multi-robot system using limited capability robots,” Robotics and Autonomous Systems,
vol. 61, no. 6, pp. 565–579, 2013.
[106] A. V. Savkin and H. Teimoori, “Decentralized navigation of groups of wheeled mobile
robots with limited communication,” IEEE Transactions on Robotics, vol. 26, no. 6, pp.
1099–1104, 2010.
101
[107] A. V. Savkin, T. M. Cheng, Z. Xi, F. Javed, A. S. Matveev, and H. Nguyen, Decentralized
Coverage Control Problems For Mobile Robotic Sensor and Actuator Networks. Wiley
& IEEE Press, 2015.
[108] T. M. Cheng and A. V. Savkin, “A distributed self-deployment algorithm for the coverage
of mobile wireless sensor networks,” IEEE Communications Letters, vol. 13, no. 11, pp.
877–879, 2009.
[109] T. Cheng and A. Savkin, “Decentralized control of mobile sensor networks for asymptotically optimal blanket coverage between two boundaries,” IEEE Transactions on Industrial
Informatics, vol. 9, no. 1, pp. 365–376, 2013.
[110] T. M. Cheng, A. V. Savkin, and F. Javed, “Decentralized control of a group of mobile
robots for deployment in sweep coverage,” Robotics and Autonomous Systems, vol. 59,
no. 7, pp. 497–507, 2011.
[111] V. Borkar and P. P. Varaiya, “Asymptotic agreement in distributed estimation,” IEEE
Transactions on Automatic Control, vol. 27, no. 3, pp. 650–655, 1982.
[112] J. N. Tsitsiklis, D. P. Bertsekas, and M. Athans, “Distributed asynchronous deterministic
and stochastic gradient optimization algorithms,” in Proceedings of the 1984 American
Control Conference (ACC), 1984, pp. 484–489.
[113] T. Vicsek, A. Czirók, E. Ben-Jacob, I. Cohen, and O. Shochet, “Novel type of phase
transition in a system of self-driven particles,” Physical review letters, vol. 75, no. 6, p.
1226, 1995.
[114] A. V. Savkin, “Coordinated collective motion of groups of autonomous mobile robots:
Analysis of vicsek’s model,” IEEE Transactions on Automatic Control, vol. 49, no. 6, pp.
981–983, 2004.
[115] A. Jadbabaie, J. Lin, and A. S. Morse, “Coordination of groups of mobile autonomous
agents using nearest neighbor rules,” IEEE Transactions on Automatic Control, vol. 48,
no. 6, pp. 988–1001, 2003.
[116] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks of agents with
switching topology and time-delays,” IEEE Transactions on Automatic Control, vol. 49,
no. 9, pp. 1520–1533, 2004.
[117] W. Ren, R. W. Beard, and E. M. Atkins, “A survey of consensus problems in multi-agent
coordination,” in Proceedings of the 2005 American Control Conference (ACC). IEEE,
2005, pp. 1859–1864.
[118] R. Olfati-Saber, A. Fax, and R. M. Murray, “Consensus and cooperation in networked
multi-agent systems,” Proceedings of the IEEE, vol. 95, no. 1, pp. 215–233, 2007.
[119] A. Stanoev and D. Smilkov, “Consensus theory in networked systems,” in Consensus and
Synchronization in Complex Networks. Springer, 2013, pp. 1–22.
[120] A. V. Savkin, C. Wang, A. Baranzadeh, Z. Xi, and H. T. Nguyen, “Distributed formation
building algorithms for groups of wheeled mobile robots,” Robotics and Autonomous
Systems, vol. 75, pp. 463–474, 2016.
102
[121] V. Blondel, J. M. Hendrickx, A. Olshevsky, J. Tsitsiklis et al., “Convergence in multiagent
coordination, consensus, and flocking,” in Proceedings of the 2005 IEEE Conference on
Decision and Control, vol. 44, no. 3. IEEE; 1998, 2005, p. 2996.
[122] A. V. Savkin and F. Javed, “A method for decentralized self-deployment of a mobile sensor
network with given regular geometric patterns,” in Proceedings of the Seventh International Conference on Intelligent Sensors, Sensor Networks and Information Processing
(ISSNIP). IEEE, 2011, pp. 371–376.
[123] V. Nazarzehi, A. V. Savkin, and A. Baranzadeh, “Distributed 3d dynamic search coverage
for mobile wireless sensor networks,” IEEE Communications Letters, vol. 19, no. 4, pp.
633–636, 2015.
[124] L. E. Parker, “Distributed algorithms for multi-robot observation of multiple moving
targets,” Autonomous robots, vol. 12, no. 3, pp. 231–255, 2002.
[125] A. Cunningham, K. M. Wurm, W. Burgard, and F. Dellaert, “Fully distributed scalable
smoothing and mapping with robust multi-robot data association,” in Proceedings of the
2012 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2012,
pp. 1093–1100.
[126] G. Erinc and S. Carpin, “Anytime merging of appearance-based maps,” Autonomous
Robots, vol. 36, no. 3, pp. 241–256, 2014.
[127] T.-D. Vu, J. Burlet, and O. Aycard, “Grid-based localization and local mapping with
moving object detection and tracking,” Information Fusion, vol. 12, no. 1, pp. 58–69,
2011.
[128] D. Marinakis and G. Dudek, “Pure topological mapping in mobile robotics,” IEEE Transactions on Robotics, vol. 26, no. 6, pp. 1051–1064, 2010.
[129] A. Baranzadeh and V. Nazarzehi, “Distributed formation building with obstacle avoidance for a team of wheeled mobile robots,” in Proceedings of the 2015 Australasian Conference on Robotics and Automation (ACRA). ARAA, 2015.
[130] ——, “A decentralized formation building algorithm with obstacle avoidance for multirobot systems,” in Proceedings of the 2015 IEEE International Conference on Robotics
and Biomimetics (ROBIO). IEEE, 2015, pp. 2513–2518.
[131] R. Olfati-Saber and R. M. Murray, “Graph rigidity and distributed formation stabilization
of multi-vehicle systems,” in Proceedings of the 41st IEEE Conference on Decision and
Control, vol. 3. IEEE, 2002, pp. 2965–2971.
[132] M. Farrokhsiar and H. Najjaran, “An unscented model predictive control approach to
the formation control of nonholonomic mobile robots,” in Proceedings of the 2012 IEEE
International Conference on Robotics and Automation (ICRA). IEEE, 2012, pp. 1576–
1582.
[133] M. Turpin, N. Michael, and V. Kumar, “Trajectory design and control for aggressive
formation flight with quadrotors,” Autonomous Robots, vol. 33, no. 1-2, pp. 143–156,
2012.
103
[134] Y. Hong, J. Hu, and L. Gao, “Tracking control for multi-agent consensus with an active
leader and variable topology,” Automatica, vol. 42, no. 7, pp. 1177–1182, 2006.
[135] H. Yu and Y. Wang, “Coordinated collective motion of groups of autonomous mobile
robots with directed interconnected topology,” Journal of Intelligent and Robotic Systems,
vol. 53, no. 1, pp. 87–98, 2008.
[136] T. Liu and Z.-P. Jiang, “Distributed formation control of nonholonomic mobile robots
without global position measurements,” Automatica, vol. 49, no. 2, pp. 592–600, 2013.
[137] W. Dong, “Robust formation control of multiple wheeled mobile robots,” Journal of
Intelligent & Robotic Systems, vol. 62, no. 3-4, pp. 547–565, 2011.
[138] J.-W. Kwon and D. Chwa, “Hierarchical formation control based on a vector field method
for wheeled mobile robots,” IEEE Transactions on Robotics, vol. 28, no. 6, pp. 1335–1345,
2012.
[139] J. Guo, Z. Lin, M. Cao, and G. Yan, “Adaptive leader-follower formation control for
autonomous mobile robots,” in Proceedings of the 2010 American Control Conference
(ACC). IEEE, 2010, pp. 6822–6827.
[140] E. M. Low, I. R. Manchester, and A. V. Savkin, “A biologically inspired method for visionbased docking of wheeled mobile robots,” Robotics and Autonomous Systems, vol. 55,
no. 10, pp. 769–784, 2007.
[141] L. Consolini, F. Morbidi, D. Prattichizzo, and M. Tosques, “On a class of hierarchical
formations of unicycles and their internal dynamics,” IEEE Transactions on Automatic
Control, vol. 57, no. 4, pp. 845–859, 2012.
[142] M. Defoort, T. Floquet, A. Kokosy, and W. Perruquetti, “Sliding-mode formation control
for cooperative autonomous mobile robots,” IEEE Transactions on Industrial Electronics,
vol. 55, no. 11, pp. 3944–3953, 2008.
[143] L. Krick, M. E. Broucke, and B. A. Francis, “Stabilisation of infinitesimally rigid formations of multi-robot networks,” International Journal of Control, vol. 82, no. 3, pp.
423–439, 2009.
[144] Q. Wang and Y.-P. Tian, “Minimally rigid formations control for multiple nonholonomic
mobile agents,” in Proceedings of the 31st Chinese Control Conference (CCC). IEEE,
2012, pp. 6171–6176.
[145] H. Mehrjerdi, J. Ghommam, and M. Saad, “Nonlinear coordination control for a group
of mobile robots using a virtual structure,” Mechatronics, vol. 21, no. 7, pp. 1147–1155,
2011.
[146] M. Hoy, A. S. Matveev, and A. V. Savkin, “Algorithms for collision-free navigation of
mobile robots in complex cluttered environments: a survey,” Robotica, vol. 33, no. 03,
pp. 463–497, 2015.
[147] H. Teimoori and A. V. Savkin, “A biologically inspired method for robot navigation in a
cluttered environment,” Robotica, vol. 28, no. 5, pp. 637–648, 2010.
104
[148] A. S. Matveev, C. Wang, and A. V. Savkin, “Real-time navigation of mobile robots in
problems of border patrolling and avoiding collisions with moving and deforming obstacles,” Robotics and Autonomous systems, vol. 60, no. 6, pp. 769–788, 2012.
[149] A. V. Savkin and C. Wang, “Seeking a path through the crowd: Robot navigation in unknown dynamic environments with moving obstacles based on an integrated environment
representation,” Robotics and Autonomous Systems, vol. 62, no. 10, pp. 1568–1580, 2014.
[150] ——, “A simple biologically inspired algorithm for collision-free navigation of a unicyclelike robot in dynamic environments with moving obstacles,” Robotica, vol. 31, no. 6, pp.
993–1001, 2013.
[151] M. Hoy, A. S. Matveev, and A. V. Savkin, “Collision free cooperative navigation of
multiple wheeled robots in unknown cluttered environments,” Robotics and Autonomous
Systems, vol. 60, no. 10, pp. 1253–1266, 2012.
[152] A. V. Savkin and M. Hoy, “Reactive and the shortest path navigation of a wheeled mobile
robot in cluttered environments,” Robotica, vol. 31, no. 2, pp. 323–330, 2013.
[153] A. V. Savkin, C. Wang, A. Baranzadeh, Z. Xi, and H. T. Nguyen, “A method for decentralized formation building for unicycle-like mobile robots,” in Proceedings of the 9th
Asian Control Conference (ASCC). Istanbul, Turkey: IEEE, 2013, pp. 1–5.
[154] Y. Liang and H.-H. Lee, “Decentralized formation control and obstacle avoidance for
multiple robots with nonholonomic constraints,” in Proceedings of the 2006 American
Control Conference. IEEE, 2006, pp. 6–pp.
[155] C. De La Cruz and R. Carelli, “Dynamic model based formation control and obstacle
avoidance of multi-robot systems,” Robotica, vol. 26, no. 03, pp. 345–356, 2008.
[156] H. Rezaee and F. Abdollahi, “A decentralized cooperative control scheme with obstacle
avoidance for a team of mobile robots,” IEEE Transactions on Industrial Electronics,
vol. 61, no. 1, pp. 347–354, 2014.
[157] H. Choset, “Coverage for robotics–a survey of recent results,” Annals of Mathematics and
Artificial Intelligence, vol. 31, no. 1-4, pp. 113–126, 2001.
[158] S. Kumar, T. H. Lai, and A. Arora, “Barrier coverage with wireless sensors,” in Proceedings of the 11th Annual International Conference on Mobile Computing and Networking.
ACM, 2005, pp. 284–298.
[159] R. Cassinis, G. Bianco, A. Cavagnini, and P. Ransenigo, “Strategies for navigation of
robot swarms to be used in landmines detection,” in Proceedings of the Third European
Workshop on Advanced Mobile Robots (Eurobot’99). IEEE, 1999, pp. 211–218.
[160] A. Jeremic and A. Nehorai, “Design of chemical sensor arrays for monitoring disposal sites
on the ocean floor,” IEEE Journal of Oceanic Engineering, vol. 23, no. 4, pp. 334–343,
1998.
[161] E. Borhaug, A. Pavlov, and K. Y. Pettersen, “Straight line path following for formations
of underactuated underwater vehicles,” in Proceedings of the 46th IEEE Conference on
Decision and Control. IEEE, 2007, pp. 2905–2912.
105
[162] A. V. Savkin, F. Javed, and A. S. Matveev, “Optimal distributed blanket coverage selfdeployment of mobile wireless sensor networks,” IEEE Communications Letters, vol. 16,
no. 6, pp. 949–951, 2012.
[163] R. Kershner, “The number of circles covering a set,” American Journal of Mathematics,
vol. 61, no. 3, pp. 665–671, 1939.
[164] A. V. Savkin and I. R. Petersen, “Model validation for robust control of uncertain systems
with an integral quadratic constraint,” Automatica, vol. 32, no. 4, pp. 603–606, 1996.
[165] S. R. Moheimani, A. V. Savkin, and I. R. Petersen, “Robust filtering, prediction, smoothing, and observability of uncertain systems,” IEEE Transactions on Circuits and Systems
- I: Fundamental Theory and Applications, vol. 45, no. 4, pp. 446–457, 1998.
[166] I. R. Petersen and A. V. Savkin, Robust Kalman filtering for signals and systems with
large uncertainties. Birkhauser, Boston, 1999.
[167] P. N. Pathirana, N. Bulusu, A. V. Savkin, and S. Jha, “Node localization using mobile robots in delay-tolerant sensor networks,” IEEE Transactions on Mobile Computing,
vol. 4, no. 3, pp. 285–296, 2005.
[168] O. Khatib, “Real-time obstacle avoidance for manipulators and mobile robots,” The international journal of robotics research, vol. 5, no. 1, pp. 90–98, 1986.
[169] D. K. Sutantyo, S. Kernbach, P. Levi, V. Nepomnyashchikh et al., “Multi-robot searching
algorithm using lévy flight and artificial potential field,” in Proceedings of the 2010 IEEE
International Workshop on Safety Security and Rescue Robotics (SSRR). IEEE, 2010,
pp. 1–6.
[170] I. R. Manchester and A. V. Savkin, “Circular-navigation-guidance law for precision missile/target engagements,” Journal of Guidance, Control, and Dynamics, vol. 29, no. 2,
pp. 314–320, 2006.
[171] R. Fierro and F. L. Lewis, “Control of a nonholonomic mobile robot using neural networks,” IEEE Transactions on Neural Networks, vol. 9, no. 4, pp. 589–600, 1998.
[172] J. P. Desai, J. P. Ostrowski, and V. Kumar, “Modeling and control of formations of
nonholonomic mobile robots,” IEEE Transactions on Robotics and Automation, vol. 17,
no. 6, pp. 905–908, 2001.
[173] J.-M. Yang and J.-H. Kim, “Sliding mode control for trajectory tracking of nonholonomic
wheeled mobile robots,” IEEE Transactions on Robotics and Automation, vol. 15, no. 3,
pp. 578–587, 1999.
[174] T. D. Barfoot and C. M. Clark, “Motion planning for formations of mobile robots,”
Robotics and Autonomous Systems, vol. 46, no. 2, pp. 65–78, 2004.
[175] S. G. Tzafestas, Introduction to mobile robot control. Elsevier, 2013.
[176] V. Malyavej and A. V. Savkin, “The problem of optimal robust Kalman state estimation
via limited capacity digital communication channels,” Systems & Control Letters, vol. 54,
no. 3, pp. 283–292, 2005.
106
[177] H. Teimoori and A. V. Savkin, “Equiangular navigation and guidance of a wheeled mobile
robot based on range-only measurements,” Robotics and Autonomous Systems, vol. 58,
no. 2, pp. 203–215, 2010.
[178] V. I. Utkin, Sliding Modes in Control and Optimization.
Media, 2013.
Springer Science & Business
[179] A. V. Savkin, I. R. Petersen, E. Skafidas, and R. J. Evans, “Hybrid dynamical systems:
robust control synthesis problems,” Systems & Control Letters, vol. 29, no. 2, pp. 81–90,
1996.
[180] A. S. Matveev and A. V. Savkin, Qualitative theory of hybrid dynamical systems.
Birkhauser, Boston, 2000.
[181] A. V. Savkin and R. J. Evans, Hybrid dynamical systems: controller and sensor switching
problems. Birkhauser, Boston, 2002.
[182] A. V. Savkin and H. Teimoori, “Bearings-only guidance of a unicycle-like vehicle following a moving target with a smaller minimum turning radius,” IEEE Transactions on
Automatic Control, vol. 55, no. 10, pp. 2390–2395, 2010.
[183] B. P. Gerkey and M. J. Matarić, “A formal analysis and taxonomy of task allocation in
multi-robot systems,” The International Journal of Robotics Research, vol. 23, no. 9, pp.
939–954, 2004.
[184] K. Lerman, C. Jones, A. Galstyan, and M. J. Matarić, “Analysis of dynamic task allocation in multi-robot systems,” The International Journal of Robotics Research, vol. 25,
no. 3, pp. 225–241, 2006.
[185] M. Saleem, G. A. Di Caro, and M. Farooq, “Swarm intelligence based routing protocol
for wireless sensor networks: Survey and future directions,” Information Sciences, vol.
181, no. 20, pp. 4597–4624, 2011.
[186] M. Dorigo, D. Floreano, L. M. Gambardella, F. Mondada, S. Nolfi, T. Baaboura, M. Birattari, M. Bonani, M. Brambilla, A. Brutschy et al., “Swarmanoid: a novel concept for
the study of heterogeneous robotic swarms,” IEEE Robotics & Automation Magazine,
vol. 20, no. 4, pp. 60–71, 2013.
[187] A. Censi, A. Franchi, L. Marchionni, and G. Oriolo, “Simultaneous calibration of odometry and sensor parameters for mobile robots,” IEEE Transactions on Robotics, vol. 29,
no. 2, pp. 475–492, 2013.
[188] K. Lee, W. Chung, and K. Yoo, “Kinematic parameter calibration of a car-like mobile
robot to improve odometry accuracy,” Mechatronics, vol. 20, no. 5, pp. 582–595, 2010.
[189] A. S. Matveev, M. C. Hoy, and A. V. Savkin, “A globally converging algorithm for
reactive robot navigation among moving and deforming obstacles,” Automatica, vol. 54,
pp. 292–304, 2015.
[190] A. Matveev, A. V. Savkin, M. Hoy, and C. Wang, Safe Robot Navigation Among Moving
and Steady Obstacles. Elsevier, 2015.
107
[191] G. Hollinger, S. Singh, J. Djugash, and A. Kehagias, “Efficient multi-robot search for
a moving target,” The International Journal of Robotics Research, vol. 28, no. 2, pp.
201–219, 2009.
[192] R. Olfati-Saber, “Flocking for multi-agent dynamic systems: Algorithms and theory,”
IEEE Transactions on Automatic Control, vol. 51, no. 3, pp. 401–420, 2006.
[193] A. S. Matveev, H. Teimoori, and A. V. Savkin, “Navigation of a unicycle-like mobile
robot for environmental extremum seeking,” Automatica, vol. 47, no. 1, pp. 85–91, 2011.
[194] N. Ghods and M. Krstic, “Multiagent deployment over a source,” IEEE Transactions on
Control Systems Technology, vol. 20, no. 1, pp. 277–285, 2012.
[195] J. Cochran, E. Kanso, S. D. Kelly, H. Xiong, and M. Krstic, “Source seeking for two
nonholonomic models of fish locomotion,” IEEE Transactions on Robotics, vol. 25, no. 5,
pp. 1166–1176, 2009.
[196] S.-J. Liu and M. Krstic, “Stochastic source seeking for nonholonomic unicycle,” Automatica, vol. 46, no. 9, pp. 1443–1453, 2010.
108
| 3cs.SY
|
1
Diffusion Based Molecular Communication
with Limited Molecule Production Rate
arXiv:1802.08965v1 [cs.IT] 25 Feb 2018
Hamid G. Bafghi, Amin Gohari, Mahtab Mirmohseni, Masoumeh Nasiri-Kenari
Department of Electrical Engineering, Sharif University of Technology, Tehran, Iran
Abstract—This paper studies the impact of a transmitter’s
molecule generation process on the capacity of a concentration based Molecular Communication (MC) system. Constraints
caused by the molecule generation process affect the availability
of the molecules at the transmitter. The transmitter has a storage
of molecules, and should decide whether to release or save the
currently produced molecules. As a result, the MC system has
conceptual connections with energy harvesting systems. In this
paper, we consider two scenarios on the propagation channel.
The first scenario assumes a channel with no Inter-symbol
Interference (ISI), i.e., a memoryless channel. We derive bounds
on the capacity of the MC system in this scenario. The second
scenario assumes the MC network with ISI, in which the output
of the channel depends on the history of released molecules in
the pervious time-slots. Based on the assumptions that either
the transmitter or the receiver knows the channel statistics, we
compute a lower bound on the channel capacity.1
Index Terms—Molecular communication (MC) network, Channel Capacity, inter-symbol interference (ISI).
I. I NTRODUCTION
Unlike the classical wireless communication, diffusion
based Molecular Communication (MC) utilizes molecules as
the carriers of information between the communicating parties.
Type, concentration, or the release time of molecules can be
used for signaling by a molecular transmitter. As a result, a
mechanism must be set in place for production of molecules
at the transmitters [1]. For instance, this may be realized by
chemical reactions inside the transmitter nodes. The impact
of this molecule production process on the capacity of a
molecular channel is the subject of this paper.
We assume that the transmitter includes a production unit
as well as a storage unit. The production unit adds some
amount of new molecules in each time-slot to the storage
unit. The production rate of molecules may depend on the
amount of molecules already exist in the storage unit, e.g., the
chemical process responsible for molecule production might
have a faster production rate if the storage unit is empty. The
transmitter communicates its message by controlled opening
of an outlet of the storage unit for a short period of time
and thereby releasing a concentration of molecules into the
environment at the beginning of each time-slot. This model
is comparable with an Energy Harvesting (EH) system in the
classical communications, in which the transmitter harvests
energy and wishes to send its message such that its transmitted
signal is amplitude-constrained to the amount of harvested and
1 Authors are listed in the alphabetical order, not according to their contributions.
stored energy at the transmitter [2]–[7]. The transmitter (in the
EH system) may have finite [6] or infinite [8] energy storage
(battery) or it can be assumed with no battery [9].
There are several approaches and models of transmitters
and receivers for diffusion based MC in the literature. We
follow the common approach of choosing one of the models
and adapting our analysis to it. In particular, we adopt the
macro-scale mode of MC and the molecular Poisson model
for our study (see [1], [10] and references therein for a
review of different models and the related results). More
specifically, we assume that the amount of released molecules
is a deterministic concentration (in molar) and the Fick’s law
of diffusion describes the medium. The reception noise is
modeled by a Poisson random variable, i.e., the received signal
has a Poisson distribution whose mean is proportional to the
average concentration of molecules at the receiver.
In this paper, we consider two cases of the Poisson channel,
with or without Inter-Symbol Interference (ISI). The ISI occurs
when a non-negligible portion of transmitted molecules from
the previous time-slots remain in the medium and affect the
communication in the current time-slot. We begin by providing
a number of capacity results when the channel is without ISI.
One should note that though the channel is memoryless in this
case, the problem is still complicated due to the fact that the
transmitter has memory; the number of released molecules
in each transmission is limited by the level of the storage,
which itself depends on the previous transmissions. A similar
phenomenon occurs in the classical energy harvesting systems.
Next, we consider a channel with ISI, i.e., a channel with
memory, and provide a result on its capacity depending on
whether the channel statistics are known at the transmitter or
at the receiver.
A. Related works
In practice, a MC transmitter suffers from constraints on
its molecule production and storage processes [11]–[13]. The
molecule production may be constrained by limitations on
the chemical reactions or the availability of food and energy
for molecule generation at the transmitter [13]. Moreover, in
practical scenarios, the bio-nanomachines store the molecules
that they produce internally or capture externally from the
environment [13]. Thus, the limitation on molecule storage
forces some constraints on the molecule transmitting process [14].
Most of the existing works in the literature of diffusion
based MC networks assume availability of a constant number
2
of molecules at the transmission times [15]. For instance in
the on/off keying, the transmitter releases a fixed amount of
molecules to send bit ‘1’ and stays silent to send bit ‘0’ [15]–
[19]. However, there are a few works that consider the limitations on molecule production at the transmitter [11]–[13]:
the random chemical reactions in the molecule production
process is considered in [11]. This work studies the chemical
description of the transmitters in terms of Langevin equation.
In [12] the total molecule concentration at the transmitter is
assumed to be a function of the number of random chemical
reactions between different molecule types. More precisely,
the inherent randomness in the availability of food and energy
is assumed to affect the molecule generation process which
limits the availability of molecules at the transmitter in the
beginning of each time-slot [13]. Due to lack of sufficient
molecules, the lengths of the symbol intervals may vary in
practical MC systems and the constrained transmitter may not
be able to emit molecules with a fixed release frequency [20];
the authors in [20] suggest using two types of molecules for
communicating and symbol synchronization over the channel.
In [21], it is assumed that the amount of available molecules
in the present time-slot, which is referred to as the “state”
of the transmitter, depends on the state and the inputs in the
previous time-slots.
In another line of works, the transmitter is assumed to
actively capture the needed molecules from the environment.
This process is referred to as “molecule harvesting” [14]. Here
the transmitter has the ability to harvest the arrived molecules,
in addition to the background molecules that randomly hit the
transmitter and the ones captured through chemical synthesis
from other materials. In this model, the harvesting process
is constrained by the size of the molecule storage at the
transmitter. The state of the transmitter (i.e., the number of the
molecules at the transmitter’s storage) is updated according to
a weighted combination of the released, the received and the
harvested molecules in the past time-slot [14].
In this paper, we study the transmitter molecule production
constraint. Our work is novel since previous works do not
consider capacity degradation due to the molecule production
constraint at the transmitter. For our study, we adopt a molecular Poisson channel model, an important molecular channel
adopted in many works in the literature [19], [22]–[27],
The rest of the paper is organized as follows. In Section II,
the MC channel and some preliminary definitions are presented. The main results of the paper on the MC networks
with no ISI are presented in III, including the capacity in the
case of infinite molecule storage, and inner and outer bounds
on the capacity in the case of finite molecule storage. The
main results of the paper on the MC networks with ISI are
presented in IV. The paper is concluded in Section V.
II. S YSTEM M ODEL
A. Notation and Definitions
We use the notation xn to denote the sequence
(x0 , x1 , . . . , xn ). Random variables (r.v.s) are denoted by
uppercase letters, while their realizations are denoted by the
lowercase letters. We say that random variables X, Y, Z form
a Markov chain if pZ|XY (z|x, y) = pZ|Y (z|y) for all x, y, z.
We show this Markov chain relation by X → Y → Z. A
sequence of random variables {Xk } for k = 1, 2, · · · is said
to be Asymptotically Mean Stationary (AMS) if
n
1X
P[Xk ∈ A] = P (A)
n→∞ n
lim
(1)
k=1
exists for all measurable A. Under this condition, P is also a
probability measure and is named the stationary mean of the
AMS sequence [33].
B. Channel Model
In our setting, we have a point-to-point communication
channel in which the information is conveyed by the molecule
concentration released into the environment by the transmitter. A deterministic channel between the transmitter and the
receiver, based on the Fick’s law of diffusion, is assumed. A
reception noise is assumed at the receiver; the concentration
of molecules is detected by a Poisson reception process at
the receiver [28]. A common example of the Poisson reception process is the particle counting noise of a transparent
receiver [29], [30]. A transparent receiver consists of a transparent sphere of a certain volume. It counts the number of
molecules that fall into its sphere. A transparent receiver is
passive in the sense that it does not affect the diffusion medium
by imposing a boundary condition to the differential equation
describing the diffusion process. The precise mathematical
description of a Poisson reception process is as follows:
if the concentration of molecules around the receiver is ρ
moles, the receiver’s measurement is distributed according to
Poisson(κρ) for some constant κ. Since the variance of a
Poisson distribution is proportional to its mean, the larger the
ρ, the noisier the receiver’s reception will be.
We assume a time-slotted transmission. The transmitter
instantaneously releases Xi moles of molecules into the environment at the beginning of each time-slot for i = 0, 1, 2, . . . .
In other words, if the transmitter is located at ~r = 0 and has
a clock with frequency 1/Ts , the transmitter’s channel input
is the impulse train
X
Xk δ(~r = 0)δ(t − kTs ).
k
The Fick’s law of diffusion describes propagation of molecules
in the environment. Here, we assume that the communication
is invariant over time, and the boundary conditions are set
to zero, i.e., there is no molecule production source besides
the transmitter. As a result, the diffusion medium (described
by the Fick’s law of diffusion) becomes a linear time-invariant
(LTI) system and can be characterized by its impulse response.
Thus, the concentration of molecules at the receiver at the
end of
Pithe i-th time-slot can be expressed as the convolution j=0 ζj Xi−j , where ζj represents the channel impulse
response coefficient. The receiver’s measurement
follows a
Pi
Poisson distribution as Yi ∼ Poisson(κ j=0 ζj Xi−j ). Letting
P
pj = κζj , we can write Yi ∼ Poisson( ij=0 pj Xi−j ) for
i = 0, 1, 2, . . . . We refer to {pi } as the channel coefficients.
We assume that the reception noise at the receiver in different
3
time slots are mutually independent, i.e., Yi ’s are conditionally independent given the transmission amounts {Xi }. If
the reception noise is the particle counting noise, to satisfy
independence of the reception noises one should adopt the
common assumption in the literature that the sampling period
Ts is not very small [31, Section IV], [10, Remark 1].
We say that the channel is without ISI if pi is negligible for
i > 0. In this case, Yi ∼ Poisson(p0 Xi ), and
p(y n |xn ) =
n
Y
i=1
p(yi |xi ).
(2)
The channel is said to be with ISI if the output of the channel
in i-th time-slot depends on the past inputs of the channel in
the previous time-slots with weights pj , j ∈ {0, . . . , i} (see
Fig. 1).
C. Transmitter Model
The transmitter opens the outlet of its molecule storage
unit for a short period of time at the beginning of each
time-slot. The channel input xi , at time-slot i, represents the
deterministic released molecule concentration (in molar) into
the environment. We assume that the transmitter has si moles
of molecules in the beginning of i-th time-slot (in its storage
unit). Moreover, the amount of molecules in the storage unit
after recharging for the duration of the i-th time-slot is denoted
by f (si ), where f (·) is a known function. We expect that
f (si ) ≥ si , i.e., molecule production is non-negative in the ith time slot. Also, we expect f (·) to be a non-decreasing
function, i.e., f (s) ≥ f (s′ ) for s ≥ s′ . The intuition here
is that if the transmitter starts off with s mole of molecules, it
will have more molecules after recharging than if it starts off
with s′ < s moles of molecules. In some practical scenarios,
the production rate of new molecules (f (s)−s) decreases in s,
the amount of molecules already existing in the storage unit.
We do not need to make such a restrictive assumption on the
function s 7→ f (s) − s for our results to hold.
If Si is the amount of molecules in the transmitter at i-th
time-slot, we assume that Xi ≤ Si moles of molecules are
released into the environment. Thus, we will have
Si+1 = f (Si − Xi ).
(3)
In other words, Si reduces to Si − Xi because of molecule
release and then recharges to f (Si − Xi ).
We assume that the transmitter starts off empty, i.e., S0 = 0.
If we have no transmission (Xi = 0) and just recharging, we
will have Si = f (Si−1 ) for i = 1, 2, . . . ; and thus, Si = f i (0)
where
f i (s) , (f ◦ f ◦ f · · · ◦ f )(s).
{z
}
|
i times
Since we assumed that f (s) ≥ s for all s, the equation Si =
f (Si−1 ) implies that Si ≥ Si−1 . Therefore, f i (0) is a nondecreasing sequence of non-negative reals. Let
ϕ = lim f i (0),
i→∞
(4)
be the amount of molecules after recharging for infinite time
(when there is no transmission); we set ϕ = ∞ if the sequence
Fig. 1. Modeling the molecular communication networks with constrained
transmitter and ISI effect as a point-to-point molecule-harvesting Poisson
channel with memory.
f i (0) tends to infinity (as discussed later). From (4), it follows
that given any s∗ ∈ [0, ϕ), it is possible to save molecules for
a finite number of time instances n0 and reach the molecular
level s∗ or larger in the transmitter’s reservoir (when there is
no transmission). That is Sn0 ≥ s∗ .
We say that the transmitter has finite molecule capacity
and saturates if ϕ < ∞, and has unbounded molecule
capacity if ϕ = ∞. If ϕ < ∞, from (4) we have f (ϕ) =
limi→∞ f i+1 (0) = ϕ, i.e., ϕ is a fixed point of f . If ϕ = ∞,
from monotonicity of f and (4) we have
lim f (x) = ∞,
x→∞
which we can interpret as f (ϕ) = ϕ for ϕ = ∞ by
symbolically extending the domain of f to include ∞.
With the assumption that the transmitter starts off empty, we
only need to define the function f over the interval [0, ϕ]. The
reason is that for any transmission sequence {Xi } satisfying
Xi ≤ Si , we will have Si ≤ ϕ. This follows from induction
Si+1 = f (Si − Xi ) ≤ f (Si ) ≤ f (ϕ) = ϕ,
(5)
where we use the monotonicity property of f (.). Hence, we
assume that the function f (.) is well-defined over [0, ϕ].
D. Channel Capacity
The transmitter wishes to send the message M which is
uniformly distributed over the set M = {0, 1}k to the receiver
in n channel uses. The communication rate is R = k/n bits
per channel use. The rate R is achievable if for any 0 <
ǫ ≤ 1 there exists a code (respecting the transmitter’s input
constraints) with rate R − ǫ whose average error probability is
less than ǫ. The channel capacity C is the maximum achievable
communication rate.
Capacity calculation is complicated due to the memory
introduced into the problem by the constraints on the transmitter. The transmitter should adapt its molecule release to
the amount of produced molecules in the current time-slot and
the consumed molecules in the previous time-slots. Thus, the
transmitter has memory.
III. MC N ETWORK W ITHOUT ISI
In this section, we present our main results on the capacity
of the MC network without ISI, i.e., Yi ∼ Poisson(p0 Xi ) for
all i.
4
For a given memoryless channel p(y|x) (here a Poisson
channel) on input alphabet X = [0, ∞) and a real c ≥ 0, let
us define:
Rc =
sup
I(X; Y ),
(6)
I(X; Y ).
(7)
pX : E[X]≤c
R̃c =
sup
pX : 0≤X≤c
The difference between Rc and R̃c is that in Rc we take
supremum over all distributions on [0, ∞) with an expected
value be at most c, while in R̃c we take supremum over all
distribution whose support is [0, c].
Theorem 1: Let
∆u = sup f (s) − s.
(8)
s∈[0,ϕ]
Then the channel capacity C satisfies:
R̃∆u ≤ C ≤ R∆u .
(9)
1
n
i=1
Xi ≤ ∆u .
(10)
Observe that (6) is the channel capacity under the average
input cost constraint ∆u given in (10) [32]. As a result,
the upper bound then follows from the known result on the
capacity of memoryless channels with an average input cost
constraint [32]. Because we assume no ISI in this section, our
channel is a memoryless Poisson channel and the result of
[32] can be utilized.
It remains to prove the lower bound on the capacity: C ≥
R̃∆u . Take an arbitrary s ∈ [0, ϕ) and let
c = f (s) − s.
(11)
We prove that R̃c = suppX : X≤c I(X; Y ) is an achievable
rate. This will complete the proof because s has been taken
arbitrary in [0, ϕ), and furthermore
∆u = sup f (s) − s = sup f (s) − s,
s∈[0,ϕ]
Si+1 = f (Si − Xi ) ≥ f (Si − c) ≥ f (s∗ − c) = s∗ . (12)
The above argument shows that it is possible to transmit
any sequence Xi satisfying Xi ≤ c for all i > n0 , as the
molecular reservoir is always at least s∗ and hence never
hits zero. By letting the blocklength n tend to infinity, the
initial finite time instances n0 becomes negligible compared
to transmission length n and results in no rate loss. We obtain
the achievability of R̃c = suppX : X≤c I(X; Y ) via Shannon’s
point to point achievability scheme, i.e., by constructing i.i.d.
codewords from a pX defined on [0, c].
Theorem 2: Assume that ϕ = ∞. Let
∆ℓ = lim inf (f (s) − s) .
s→∞
Proof: To see the upper bound, note that the number of
molecules produced in each recharging period cannot exceed
∆u . Therefore, the average number of consumed molecules
cannot be larger than ∆u :
n
X
We would like to start transmitting in time instances i > n0 .
We claim that if we limit the transmission level to c, then
transmitter’s molecular reservoir will not drop below s∗ . In
other words, if Xi ≤ c for i > n0 we have Si ≥ s∗ for all
i > n0 . This follows from induction. Assuming that Si ≥ s∗
and using the fact that Xi ≤ c, we have
s∈[0,ϕ)
as f (ϕ) − ϕ = 0. To show that R̃c is achievable, it suffices
to show that for any arbitrary distribution pX on [0, c], the
mutual information I(X; Y ) is an achievable rate.
Let s∗ = f (s) = s+c. Then from (11) we have f (s∗ −c) =
∗
s . We claim that s∗ ∈ [0, ϕ]. This is because s∗ = f (s) ≤
f (ϕ) = ϕ where we used the monotonicity property of f .
Since s ∈ [0, ϕ), we can use the save strategy at the beginning
and wait for a finite number of time instances to reach the
molecule level s or larger in the transmitter’s reservoir. If we
wait for one more time step, we reach the molecule level s∗ =
f (s). Hence, there is some finite number n0 such that Sn0 ≥
s∗ assuming that we do not transmit any molecules in the first
n0 time steps.
(13)
Then the channel capacity can be bounded from below as
follows:
C ≥ R∆ℓ .
(14)
Proof: We prove that Rc is an achievable rate for any
c < ∆ℓ . By the definition of ∆ℓ , there exists some s∗ > 0
such that for any s ≥ s∗ we have
f (s) ≥ s + c.
(15)
To prove that Rc = suppX : E[X]≤c I(X; Y ) is achievable,
we need to prove achievability of I(X; Y ) for an arbitrary
distribution pX satisfying E[X] ≤ c. The idea is to mimic the
proof of [5] here and prove this by considering distributions
on a finite support. Take some positive real number K, and
an arbitrary distribution pX of finite support on [0, K]. Here
K may be larger than c, but the distribution of X is required
to satisfy E[X] ≤ c. It is shown in [5] that achievability of
I(X; Y ) for such finite support distributions is enough to show
the achievability of Rc .
As in [5], we wait at the beginning for a finite amount of
time and save molecules so that the amount of saved molecules
exceeds a number w (to be specified later). Because ϕ = ∞,
we can take w as large as we want.
As long as Si ≥ K + s∗ , we have Si − Xi ≥ s∗ and hence
from (15) we have
Si+1 = f (Si − Xi ) ≥ Si − Xi + c.
(16)
In other words, a production of c moles of molecules is
guaranteed as long as the molecular reservoir level stays above
K + s∗ .
In [5], the update rule Si+1 = Si −Xi +c is considered. Here
in the inequality (16), the sequence Si has a larger growth and
dominates the process of [5]. This indicates that the same proof
can be mimicked here since (16) shows that in each step, more
molecules are produced than the update rule Si+1 = Si −Xi +c
of [5]. The number w should be chosen to ensure that with
high probability, starting from w molecules, Si remains above
5
K + s∗ so that (16) remains valid. But this is possible with
the same argument as in [5].
Remark 1: Theorems 1 and 2 give a tight characterization
of capacity C for the case of infinite battery capacity ϕ = ∞ if
∆ℓ = ∆u . For example, this occurs if f (s) = s + v for some
constant v, i.e., when we have a constant production rate.
Theorem 2 is comparable with an EH system with an infinite
battery [5]. In [5], it is shown that the capacity of Gaussian
EH channel with an infinite battery, with the average arrived
energy P , is equal to the capacity of the Gaussian point-topoint channel with average power constraint P .
In the following, we find a lower bound on the capacity
when ϕ < ∞.
Theorem 3: Assume that ϕ < ∞. Let qX|S (x|s) be a
conditional distribution on
is a lower bound on the capacity of the channel. Next, note
that
I(X n ; Y n ) = H(Y n ) − H(Y n |X n )
n
X
H(Yi |Xi )
= H(Y n ) −
i=1
=
≥
=
=
{(x, s) : f (0) ≤ s ≤ ϕ, 0 ≤ x ≤ s}
such that the Markov chain S0 = 0, Si+1 = f (Si − Xi )
with pXi |Si = qX|S asymptotically converges to a stationary
distribution qS , i.e., pSi tends to qS in total variation distance
as i tends to infinity. Then the capacity is lower bounded as
C ≥ I(X; Y |S),
(17)
where qX,S,Y = qX,S pY |X where pY |X is the Poisson channel
Y ∼ Poisson(p0 X).
Proof: Assume that we start with S0 = 0 and choose
Xi from Si according to qX|S . Consider the update rule
Si+1 = f (Si − Xi ). This yields a sequence {(Si , Xi , )} for
i = 0, 1, 2, · · · . Let Yi ∼ Poisson(p0 Xi ) be a memoryless
channel. Then we have
p(y n , xn , sn ) =
n
Y
i=1
p(si |xi−1 , si−1 )q(xi |si )p(yi |xi ). (18)
We claim that {(Si , Xi , Yi )} is an Asymptotically Mean
Stationary (AMS) sequence. As argued in [33], the AMS
condition allows us to conclude that
lim
n→∞
1
I(X n ; Y n )
n
is a lower bound on the capacity of the channel. Next, we
show that {(Si , Xi , Yi )} is an Asymptotically Mean Stationary
(AMS) sequence. Observe that {Si } is a Markov chain (on
a countable state space) starting from S0 = 0. Furthermore,
from the definition of the qS,X in the statement of the theorem,
qS is a limiting stationary distribution for this Markov chain
and p(Si ) converges to qS in total variation distance. Because
pXi ,Yi |Si (xi , yi |si ) = qX|S (xi |si )pY |X (yi |xi ), the distribution
of pSi ,Xi ,Yi also converges to qS,X,Y and {(Si , Xi , Yi )} will
be proven to be an Asymptotically Mean Stationary (AMS)
sequence.
To sum this up, we have that
1
I(X n ; Y n )
n→∞ n
lim
=
n
X
i=1
n
X
i=1
n
X
i=1
n
X
i=1
n
X
i=1
H(Yi |Y1:i−1 ) − H(Yi |Xi )
H(Yi |Y1:i−1 , Si ) − H(Yi |Xi )
H(Yi |Si ) − H(Yi |Xi )
H(Yi |Si ) − H(Yi |Xi , Si )
I(Xi ; Yi |Si )
where we used the fact that Yi → Xi → Si and Yi → Si →
Y1:i−1 form Markov chains, which are implied from (18).
Because pSi ,Xi ,Yi tends to qS,X,Y as i tends to infinity, we
have
n
1
1X
lim I(X n ; Y n ) ≥ lim
I(Xi ; Yi |Si ) = Iq (X; Y |S).
n→∞ n
n→∞ n
i=1
IV. MC N ETWORK W ITH ISI
In this section, we present our main results on the capacity
of the MC network with ISI. As we mentioned
in Section II-B,
P
in this channel Xi and Yi ∼ Poisson( ij=0 pj Xi−j ) represent
the channel input and the channel output, respectively (see
Fig. 1). Let C(p) denote the capacity of a channel with
coefficient sequence p. Calculation of C(p) is complicated by
the fact that the channel has memory.
Assume that the channel coefficient sequence p belong to
a class P. Here, we consider two scenarios: the first scenario
assumes that the transmitter is aware of the actual p ∈ P, but
the receiver only knows the class P and is unaware of the
sequence p ∈ P. A rate is achievable if there is a sequence of
coding strategies for the transmitter and receiver whose error
probability vanishes regardless of the choice of p ∈ P. The
second scenario is the other way around; it assumes that the
receiver is aware of the actual p ∈ P, and the transmitter only
knows the class P. Thus, the two scenarios differ in terms
of whether the “channel state information” is available at the
transmitter or receiver.
To state our results, we begin by a definition:
Definition 1: Given two sequences p = (pj ) and p̃ = (p̃j )
of channel coefficients, we say p̃ p if
p̃j =
j
X
qj−k pk ,
(19)
k=0
for some non-negative sequence
P
∞
j=0 qi ≤ 1.
q0 , q1 , · · ·
satisfying
6
Intuitively speaking, p̃ p can be understood as a channel
with a coefficient sequence p̃ having a more spread out channel
coefficient profile than the channel with a coefficient sequence
p.
Example 1: As an example, given any p̃ we have p̃ p
where p = (pj ) is defined as follows: piP
= 0 for i ≥ 1 and p0
is any number greater than or equal to i p̃i .
Our main results are as follows:
Theorem 4: Assume that the molecule production function
f (·) is a concave function (in addition to the constraints of
Section II-C). Take some channel coefficient sequence p̃ =
(p̃j ) such that p̃ p for all p ∈ P. Then, C(p̃) is an achievable
rate for any channel coefficient profile p ∈ P when p is known
only at the transmitter.
Theorem 5: Take some channel coefficient sequence p̃ =
(p̃j ) such that p̃ p for all p ∈ P. Then, for any
molecule production function f (·), C(p̃) is an achievable rate
for any channel coefficient profile p ∈ P when p is known
only at the receiver.
Remark 2: Theorem 4 imposes more restrictions on the
function f (·) compared to
√ Theorem 5. Functions f (x) =
min(x+ c, ϕ) and f (x) = x + c for c > 0 are two examples
of increasing and concave functions.
Remark 3: From Example 1, we know that given any
channel
coefficient sequence p̃ = (p̃j ), the sequence p =
P
( j p̃j , 0, 0, · · · , 0) satisfies p̃ p. Since p corresponds to
a memoryless channel, the above theorems confirms that ISI
cannot be utilized to increase the channel capacity.
Proof of Theorem 4: In this case, the receiver is unaware
of the p ∈ P. The receiver proceeds with the assumption that
the true channel coefficient is p̃. The idea is to show that
any code for a channel with coefficient p̃ can be “simulated”
by a code for channel p at the transmitter. More specifically,
choose an arbitrary n-code of rate R designed for a channel
with coefficients p̃ consisting of a set of codewords x̃n (m)
for m ∈ {1, 2, . . . , 2nR }. The codeword x̃n (m) satisfies the
transmitter molecule production constraint (i.e., transmission
in each stage is less than the amount of molecules available
in the transmitter reservoir). The receiver assumes that the
actual channel coefficient is p̃ and uses the decoding algorithm
for this channel code. The transmitter gets the true channel
coefficient p and uses the following strategy:
1) From p̃ p, it computes a non-negative sequence
(q0 , q1 , . . .) such that
p̃j =
j
X
qj−k pk .
k=0
2) To transmit message m ∈ {1, 2, . . . , 2nR }, it transmits
xj (m) =
j
X
qj−k x̃k (m), j = 0, 1, . . . , n,
(20)
k=0
on the channel, where x̃n (m) is the codeword of the
codebook for p̃ corresponding to message m.
With this
Pi strategy, the receiver
Poisson( j=0 pi−j xj (m)). Observe that
i
X
pi−j xj (m) =
j=0
pi−j
i
X
i
X
qj−k pi−j
x̃k (m)
i
X
x̃k (m)p̃i−k .
k=0
=
Yi
i
X
j=0
=
j
X
gets
qj−k x̃k (m)
∼
(21)
k=0
j=k
k=0
Pi
Therefore, Yi ∼ Poisson( j=0 p̃i−j x̃j (m)), as if the codeword x̃n (m) was transmitted over a channel with the coefficient sequence p̃. Hence, (20) allows for the simulation of
channel p̃ from channel p.
The crucial and more difficult step is to show that the
sequence of transmission xn (m) defined in (20) satisfies
the transmitter cost constraint. To prove this, we use the
assumption that each codeword x̃n (m) satisfies the transmitter
cost constraint. That is,
x̃i (m) ≤ s̃i (m)
where s̃0 (m) = 0 and s̃i (m) = f (s̃i−1 (m) − x̃i−1 (m)) is
the number of molecules in transmitter’s reservoir at time i if
message m is transmitted. Our claim follows from Lemma 1
in the Appendix.
Proof of Theorem 5: In this case, the transmitter is
unaware of the p ∈ P. The transmitter proceeds with the
assumption that the true channel coefficient sequence is p̃.
The idea is to convert the channel p to the channel p̃ by an
operation at the receiver. Choosing an arbitrary n-code of rate
R designed for a channel with a coefficient sequence p̃ consisting of a set of codewords x̃n (m) for m ∈ {1, 2, . . . , 2nR },
the transmitter sends the codeword x̃n (m), assuming that the
actual channel coefficient sequence is p̃. The receiver gets the
true channel coefficient sequence p and uses the following
strategy: from p̃ p, the receiver computes a non-negative
sequence (q0 , q1 , . . .) such that
p̃j =
j
X
qk pj−k .
k=0
Next, having received the sequence (Y0 , Y1 , Y2 , · · · , Yn ),
the receiver produces the sequence (Ỹ0 , Ỹ1 , Ỹ2 , · · · , Ỹn ) as
follows: for each 0 ≤ k ≤ n, we produce random variables Ŷ0k , Ŷ1k , Ŷ2k , · · · , Ŷkk , Zk from a multinomial distribution withPYk balls and probability sequence
(qk , qk−1 , · · · , q0 , 1 − ki=0 qi ). We assume that the variables
Ŷ0k , Ŷ1k , Ŷ2k , · · · , Ŷkk , Zk for different values of k are produced independently. Then, we let
Ỹj =
j
X
Ŷij .
i=0
Pi
Note that Yi ∼ Poisson( k=0 pi−k x̃k (m)), 0 ≤ i ≤ n
are independent given x̃n (m). By the thinning property of
Poisson distribution, and the fact that sum of independent
Poisson random variables is again a Poisson random variable,
7
Pi
we obtain that Ŷij ∼ Poisson(qj−i k=0 pi−k x̃k (m)) and
furthermore, Ŷij for different values of i, j are mutually
independent. We have
Ỹj =
j
X
i=0
j
i
X
X
qj−i
pi−k x̃k (m))
Ŷij ∼ Poisson(
i=0
k=0
j
j
X
X
qj−i pi−k )
x̃k (m)
= Poisson(
k=0
i=k
j
X
= Poisson(
x̃k (m)p̃j−k ).
k=0
Furthermore, Ỹj for different values of j are mutually independent because they involve summation over disjoint sets of
Ŷij . As a result, Ỹi is exactly what we obtain if x̃n (m) were
passed through the channel p̃. This completes the proof of the
receiver being able to apply a post-processing to convert the
channel p to p̃.
V. C ONCLUSIONS
In this paper, a molecular communication (MC) system
is considered in which the information is encoded in the
concentration of the molecules, and the molecule generation
process causes some constraints on the transmitter. Moreover,
it is assumed that the number of received molecules at the
receiver follows a Poisson distribution of the channel input.
Two scenarios on the MC channel with no Inter-symbol
Interference (ISI) and the MC channel with ISI were studied.
For the case of no-ISI scenario, lower and upper bounds on
the channel capacity were derived. For the case of channels
with ISI, we provided a capacity result for the cases where the
channel coefficient sequence is known either at the transmitter
or at the receiver.
R EFERENCES
[1] N. Farsad, H. B. Yilmaz, A. Eckford, C. B. Chae, and W. Guo, “A
comprehensive survey of recent advancements in molecular communication,” IEEE Commun. Surv. Tutorials, vol. 18, no. 3, pp. 1887–1919,
thirdquarter 2016.
[2] M. L. Ku, W. Li, Y. Chen, and K. J. R. Liu, “Advances in energy
harvesting communications: Past, present, and future challenges,” IEEE
Comm. Surveys Tutorials, vol. 18, no. 2, pp. 1384–1412, Secondquarter
2016.
[3] O. Ozel, J. Yang, and S. Ulukus, “Broadcasting with a battery limited
energy harvesting rechargeable transmitter,” in Int. Symp. on Modeling
and Optimization in Mobile, Ad Hoc and Wireless Networks (WiOpt),
2011, pp. 205–212.
[4] R. Rajesh, V. Sharma, and P. Viswanath, “Information capacity of energy
harvesting sensor nodes,” in Int. Symp. on Inf. Theory (ISIT), July 2011,
pp. 2363–2367.
[5] O. Ozel and S. Ulukus, “Achieving AWGN capacity under stochastic
energy harvesting,” IEEE Trans. on Inf. Theory, vol. 58, no. 10, pp.
6471–6483, Oct. 2012.
[6] K. Tutuncuoglu and A. Yener, “Optimum transmission policies for battery limited energy harvesting systems,” IEEE Trans. Wireless Comm.,
vol. 11, no. 3, pp. 1180–1189, Mar. 2012.
[7] S. Ulukus, A. Yener, E. Erkip, O. Simeone, M. Zorzi, P. Grover, and
K. Huang, “Energy harvesting wireless communications: a review of
recent advances,” IEEE Jour. Selec. Area Comm., vol. 33, no. 3, pp.
360–381, Mar. 2015.
[8] O. Ozel and S. Ulukus, “Information-theoretic analysis of an energy
harvesting communication system,” in Int. Symp. on Personal, Indoor
and Mob. Radio Comm. Workshops (PIMRC Workshops), Sept. 2010,
pp. 330–335.
[9] H. G. Bafghi, B. Seyfe, M. Mirmohseni, and M. R. Aref, “Capacity of
channel with energy harvesting transmitter,” IET Comm., vol. 9, no. 4,
pp. 526–531, 2015.
[10] A. Gohari, M. Mirmohseni, and M. Nasiri-Kenari, “Information theory
of molecular communication: Directions and challenges,” IEEE Trans.
on Molecular, Biological and Multi-Scale Commun., vol. 2, no. 2, pp.
120–142, Dec 2016.
[11] D. T. Gillespie, “The chemical langevin equation,” The Journal of
Chemical Physics, vol. 113, 2000.
[12] M. Pierobon and I. F. Akyildiz, “A statistical–physical model of interference in diffusion-based molecular nanonetworks,” IEEE Trans. on
Commun., vol. 62, 6 2014.
[13] B. Alberts, D. Bray, K. Hopkin, A. Johnson, J. Lewis, M. Raff,
K. Roberts, and P. Walter, Essential Cell Biology, 4th ed. Garland
Science, 2014.
[14] J.-T. Huang and C.-H. Lee, “On capacity bounds of two-way diffusion
channel with molecule harvesting,” in IEEE Int. Conf. on Commun.
(ICC), Paris, France, 2017, 2017.
[15] M. Pierobon and I. F. Akyildiz, Fundamentals of Diffusion–Based
Molecular Communication in Nanonetworks. Now Publisher, 2014.
[16] R. Mosayebi, H. Arjmandi, A. Gohari, M. Nasiri-Kenari, and U. Mitra,
“Receivers for diffusion-based molecular communication: Exploiting
memory and sampling rate,” IEEE Journal on Selected Areas in Commun., vol. 32, no. 12, pp. 2368–2380, Dec 2014.
[17] C. Jiang, Y. Chen, and K. J. Ray Liu, “Nanoscale molecular communication networks: a game-theoretic perspective,” EURASIP Jour. on Adv.
in Signal Proc., vol. 5, 2015.
[18] J. Suzuki, T. Nakano, and M. J. Moore, Modeling, Methodologies and
Tools for Molecular and Nano-scale Communications. Springer Int.
Publishing, 2017.
[19] H. Arjmandi, M. Movahednasab, A. Gohari, M. Mirmohseni, M. NasiriKenari, and F. Fekri, “Isi-avoiding modulation for diffusion-based
molecular communication,” IEEE Trans. on Molecular, Biological and
Multi-Scale Commun., vol. 3, no. 1, pp. 48–59, March 2017.
[20] V. Jamali, A. Ahmadzadeh, and R. Schober, “Symbol synchronization
for diffusive molecular communication systems,” in IEEE Int. Conf. on
Commun. (ICC), Paris, France, 2017.
[21] T. Nakano, T. Suda, Y. Okaie, M. J. Moore, and A. V. Vasilakos,
“Molecular communication among biological nanomachines: A layered
architecture and research issues,” IEEE Trans. on NanoBioscience,
vol. 13, 9 2014.
[22] M. Pierobon and I. F. Akyildiz, “Capacity of a diffusion-based molecular
communication system with channel memory and molecular noise,”
IEEE Trans. on Inf. Theory, vol. 59, 02 2013.
[23] G. Aminian, H. Arjmandi, A. Gohari, M. Nasiri-Kenari, and U. Mitra,
“Capacity of diffusion-based molecular communication networks over
lti-poisson channels,” IEEE Trans. on Molecular, Biological and MultiScale Communications, vol. 1, no. 2, pp. 188–201, June 2015.
[24] S. M. Mustam, S. K. Syed-Yusof, and S. Zubair, “Capacity and delay
spread in multilayer diffusion-based molecular communication (dbmc)
channel,” IEEE Trans. on NanoBioscience, vol. 15, no. 7, pp. 599–612,
Oct 2016.
[25] J. Sun and H. Li, “On the capacity of amplitude modulation based
molecular communication channels,” in 2016 8th Int. Conf. on Wireless
Commun. Signal Proc. (WCSP), Oct 2016, pp. 1–6.
[26] B. Atakan and O. B. Akan, On Channel Capacity and Error Compensation in Molecular Communication. Berlin, Heidelberg: Springer Berlin
Heidelberg, 2008, pp. 59–80.
[27] G. Aminian, M. Farahnak-Ghazani, M. Mirmohseni, M. Nasiri-Kenari,
and F. Fekri, “On the capacity of point-to-point and multiple-access
molecular communications with ligand-receptors,” IEEE Trans. on
Molecular, Biological and Multi-Scale Commun., vol. 1, no. 4, pp. 331–
346, Dec 2015.
[28] H. Mahdavifar and A. Beirami, “Diffusion channel with poisson reception process: capacity results and applications,” in 2015 IEEE Int. Symp.
on Inf. Theory (ISIT), June 2015, pp. 1956–1960.
[29] D. Kilinc and O. B. Akan, “Receiver design for molecular communication,” IEEE Journal on Selected Areas in Communications, vol. 31,
no. 12, pp. 705–714, 2013.
[30] M. Pierobon and I. F. Akyildiz, “Diffusion-based noise analysis for
molecular communication in nanonetworks,” IEEE Transactions on
Signal Processing, vol. 59, no. 6, pp. 2532–2547, 2011.
8
[31] A. Noel, K. C. Cheung, and R. Schober, “Optimal receiver design for
diffusive molecular communication with flow and additive noise,” IEEE
transactions on nanobioscience, vol. 13, no. 3, pp. 350–362, 2014.
[32] A. El Gamal and Y.-H. Kim, Network Information Theory. Cambridge
University Press, 2011.
[33] R. Rajesh, V. Sharma, and P. Viswanath, “Capacity of Gaussian channels
with energy harvesting and processing cost,” IEEE Transactions on
Information Theory, vol. 60, no. 5, pp. 2563–2575, May 2014.
The base of the induction is clear. Since x̃1 ≤ s̃1 = 0, we
have x1 = q0 x̃1 = 0. Therefore, x1 ≤ s1 = 0 and s1 ≥ q0 s̃1
hold. Equation (26) holds because
sT +1 = f (sT − xT )
T
X
qT −k s̃k − xT )
≥ f(
k=0
T
X
= f(
A PPENDIX
k=0
Lemma 1: Let (x̃0 , s̃0 , x̃1 , s̃1 , · · · ) be a sequence satisfying
s̃0 = 0, x̃i ≤ s̃i , and s̃i+1 = f (s̃i − x̃i ) for i ≥ 0. Let
qP
0 , q1 , · · · be a sequence of non-negative numbers satisfying
i qi ≤ 1. Let
xj =
j
X
qj−k x̃k ,
T
X
≥
k=0
T
X
=
qT −k (s̃k − x̃k ))
(29)
qT −k s̃k+1
(30)
k=0
j = 0, 1, . . . .
= qT +1 s̃0 +
X
X
αi f (vi ).
αi vi ) ≥
f(
(22)
P
j
αj ∈ [0, 1]. Then,
qT −k s̃k+1
(31)
i
≥ (1 − κ)f (0) +
≥
X
i
X
where (27) follows from the induction hypothesis and monotonicity of f , (28) follows from the definition of xT , (29)
follows from (22), (30) follows from the definition of s̃k+1
and (31) follows from the fact that s̃0 = 0.
Equation (25) holds because
xT +1 =
αi f (vi )
≤
(23)
i
=
αi f (vi ),
i
where (23) from Jensen’s inequality for the concave function
f.
We prove the Lemma by induction on i. We form the
induction by assuming a stronger induction hypothesis: we
assume that we have defined si for i ≤ T such that xi ≤ si ,
si = f (si−1 − xi−1 ) and additionally
qj−k s̃k ,
j = 1, 2, . . . , T.
(24)
k=0
Then, we prove that sT +1 = f (sT − xT ) will satisfy
xT +1 ≤ sT +1
(25)
sT +1 ≥
(26)
T
+1
X
k=0
qT +1−k s̃k .
qT +1−k s̃k ,
k=0
=
X
X
αi vi )
αi vi ) = f ((1 − κ) · 0 +
f(
T
+1
X
=
i
i
j
X
T
X
k=0
Then, assuming that f is a concave molecule production
function, there is a sequence (s0 , s1 , · · · ) of non-negative
numbers satisfying s0 = 0, xi ≤ si and si+1 = f (si − xi ) for
any i ≥ 0.
Proof of Lemma 1: We begin by first proving the
following fact about the function
P f : given any non-negative
numbers αi and vi such that i αi ≤ 1 we have
sj ≥
(28)
qT −k f (s̃k − x̃k )
k=0
To see this, let κ =
(27)
T
+1
X
k=0
T
+1
X
k=1
T
+1
X
k=1
T
+1
X
qT +1−k x̃k
qT +1−k x̃k
(32)
qT +1−k s̃k
(33)
qT +1−k f (s̃k−1 − x̃k−1 )
(34)
k=1
T
+1
X
≤ f(
qT +1−k (s̃k−1 − x̃k−1 ))
= f(
qT −k (s̃k − x̃k ))
k=1
T
X
=f
k=0
X
T
k=0
qT −k s̃k − xT
≤ f (sT − xT )
= sT +1 ,
(35)
(36)
(37)
where (32) follows from x̃0 = 0, (33) follows from x̃k ≤ s̃k ,
(34) follows from the definition of s̃k , (35) follows from (22),
(36) follows from definition of xT and finally, (37) follows
from induction hypothesis (24) and the fact that f is nondecreasing.
| 7cs.IT
|
Theory and Techniques for Synthesizing a Family of Graph
Algorithms
Srinivas Nedunuri
William R. Cook
Douglas R. Smith
University of Texas at Austin
Kestrel Institute
nedunuri|[email protected]
[email protected]
Although Breadth-First Search (BFS) has several advantages over Depth-First Search (DFS) its prohibitive space requirements have meant that algorithm designers often pass it over in favor of DFS.
To address this shortcoming, we introduce a theory of Efficient BFS (EBFS) along with a simple
recursive program schema for carrying out the search. The theory is based on dominance relations,
a long standing technique from the field of search algorithms. We show how the theory can be used
to systematically derive solutions to two graph algorithms, namely the Single Source Shortest Path
problem and the Minimum Spanning Tree problem. The solutions are found by making small systematic changes to the derivation, revealing the connections between the two problems which are
often obscured in textbook presentations of them.
1
Introduction
Program synthesis is experiencing something of a resurgence [21, 20, 4] [14, 22] following negative
perceptions of its scalability in the early 90s. Many of the current approaches aim for near-automated
synthesis. In contrast, the approach we follow, we call guided program synthesis, also incorporates a
high degree of automation but is more user-guided. The basic idea is to identify interesting classes of
algorithms and capture as much generic algorithm design knowledge as possible in one place.The user
instantiates that knowledge with problem-specific domain information. This step is often carried out with
machine assistance. The approach has been applied to successfully derive scores of efficient algorithms
for a wide range of practical problems including scheduling [18], concurrent garbage collection [13], and
SAT solvers [19].
One significant class of algorithms that has been investigated is search algorithms. Many interesting
problems can be solved by application of search. In such an approach, an initial search space is partitioned into subspaces, a process called splitting, which continues recursively until a feasible solution
is found. A feasible solution is one that satisfies the given problem specification. Viewed as a search
tree, spaces form nodes, and the subspaces after a split form the children of that node. The process
has been formalized by Smith [15, 17]. Problems which can be solved by global search are said to be
in the Global Search (GS) class. The enhancements in GS over standard branch-and-bound include a
number of techniques designed to improve the quality of the search by eliminating unpromising avenues.
One such technique is referred to as dominance relations. Although they do not appear to have been
widely used, the idea of dominance relations goes back to at least the 70s [5]. Essentially, a dominance
relation is a relation between two nodes in the search tree such that if one dominates the other, then the
dominated node is guaranteed to lead to a worse solution than the dominating one, and can therefore be
discarded. Establishing a dominance relation for a given problem is carried out by a user. However this
process is not always obvious. There are also a variety of ways in which to carry out the search, for
example Depth-First (DFS), Breadth-First (BFS), Best-First, etc. Although DFS is the most common,
Doron Peled and Sven Schewe (Eds.):
First Workshop on Synthesis (SYNT2012)
EPTCS 84, 2012, pp. 33–46, doi:10.4204/EPTCS.84.3
c S.Nedunuri, W. R. Cook & D.R. Smith
This work is licensed under the
Creative Commons Attribution License.
Theory and Techniques for Synthesizing a Family of Graph Algorithms
34
BFS actually has several advantages over DFS were it not for its exponential space requirement. The key
to carrying out BFS space-efficiently is to limit the size of the frontier at any level. However, this has not
been investigated in any systematic manner up to now.
This paper has two main contributions:
• We show how to limit the size of the frontier in search using dominance relations, thereby enabling
space-efficient BFS. From this formal characterization, we derive a characteristic recurrence that
serves as the basis of a program schema for implementing Global Search. Additionally, we show
that limiting the size of the undominated frontier to one results in a useful class of greedy algorithms.
• We show how to derive dominance relations and demonstrate they satisfy the greediness conditions
for two graph problems, namely Single Source Shortest Path and Minimum Spanning Tree by a
systematic process, which though not automatic, we believe has the potential to be automated.
2
Background To Guided Program Synthesis
2.1 Process
The basic steps in guided program synthesis are:
1. Start with a logical specification of the problem to be solved. A specification is a quadruple
hD, R, o, ci where D is an input type, R an output or result type, o : D × R is a predicate relating
correct or feasible outputs to inputs, and c : D×R → Int is a cost function on solutions. An example
specification is in Eg. 1 (This specification is explained in more detail below)
2. Pick an algorithm class from a library of algorithm classes (G LOBAL S EARCH , L OCAL S EARCH ,
D IVIDE AND C ONQUER , F IXPOINT I TERATION, etc). An algorithm class comprises a program
schema containing operators to be instantiated and an axiomatic theory of those operators (see [9]
for details). A schema is analogous to a template in Java/C++ , with the difference that both the
template and template arguments are formally constrained.
3. Instantiate the operators of the program schema using information about the problem domain and
in accordance with the axioms of the class theory. To ensure correctness, this step can be carried
out with mechanical assistance. The result is an efficient algorithm for solving the given problem.
4. Apply low-level program transforms such as finite differencing, context-dependent simplification,
and partial evaluation, followed by code generation. Many of these are automatically applied by
Specware [1], a formal program development environment.
The result of Step 4 is an efficient program for solving the problem which is guaranteed correct by construction. The power of the approach stems from the fact that the common structure of many algorithms
is contained in one reusable program schema and associated theory. Of course the program schema needs
to be carefully designed, but that is done once by the library designer. The focus of this paper is the
G LOBAL S EARCH class, and specifically on how to methodically carry out Step 3 for a wide variety of
problems. Details of the other algorithm classes and steps are available elsewhere [7, 15, 13].
Example 1. Specification of the Single Pair Shortest Path (SPSP) problem is shown in Fig. 2.1 (The 7→
reads as “instantiates to”) The input D is a structure with 3 fields, namely a start node, end node and a
set of edges. The result R is a sequence of edges ([. . .] notation). A correct result is one that satisfies the
predicate path? which checks that a path z must be a contiguous path from the start node to the end node
( simple recursive definition not shown). Finally the cost of a solution is the sum of the costs of the edges
in that solution. Note that fields of a structure are accessed using the ’.’ notation.
S.Nedunuri, W. R. Cook & D.R. Smith
35
2.2 Global Search
Before delving into a program schema for Global
Search, it helps to understand the structures over D 7→ hs : Node, e : Node, edges : {Edge}i
which the program schema operates. In [15], a
Edge = h f : Node,t : Node, w : Nati
search space is represented by a descriptor of R 7→ [Edge]
b which is an abstraction of the result
some type R,
o 7→ λ (x, z) · path?(z, x.s, x.e)
type R. The initial or starting space is denoted
path?(p, s, f ) = ...
b
b
⊥. There are also two predicates split: D × R × R,
c 7→ λ (x, z) · ∑edge∈z edge.w
written ⋔, and extract: Rb × R, written χ . Split defines when a space is a subspace of another space, Figure 2.1: Specification of Shortest Path problem
and extract captures when a solution is extractable
from a space. We say a solution z is contained in a space y (written z ∈ y) if it can be extracted after a finite number of splits. A feasible space is one that contains feasible solutions. We often write
⋔ (x, y, y′ ) as y ⋔x y′ for readability, and even drop the subscript when there is no confusion. Global
Search theory (GS-theory) [15] axiomatically characterizes the relation between the predicates ⊥, ⋔
and χ , as well as ensuring that the associated program schema computes a result that satisfies the
b ⊥, ⋔, χ , ⊕ are all assumed to be drawn from GS-theory.
specification. In the sequel, the symbols R,
A theory for a given problem is created by instantiating these terms, as shown in the next example.
Example 2. Instantiating GS-theory for the Single Pair Shortest Path problem. The type of solution
spaces Rb is the same as the result type R 1 . A space is split by adding an edge to the current path - that is
the subspaces are the different paths that result from adding an edge to the parent path. Finally a solution
can be trivially extracted from any space by setting the result z to the space p. This is summarized in Fig.
2.2 ([] denotes the empty list, and ++ denotes list concatenation).
2.3 Dominance Relations
As mentioned in the introduction, a dominance relation provides a way of comparing two subspaces
Rb 7→ R
in order to show that one will always contain at
⊥ 7→ λ x · []
′
least as good a solution as the other. (Goodness
⋔ 7→ λ (x, p, p ) · ∃e ∈ x.edges·
′
in this case is measured by some cost function on
p = p++[e]
solutions). The first space is said to dominate (⊲)
χ 7→ λ (z, p) · p = z
the second, which can then be eliminated from the
Figure 2.2: GS instantiation for Single Pair Shortest search. Letting c∗ denote the cost of an optimal
Path
solution in a space, this can be formalized as (all
free variables are assumed to be universally quantified):
y ⊲ y′ ⇒ c∗ (x, y) ≤ c∗ (x, y′ )
(2.1)
Another way of expressing the consequent of (2.1) is
∀z′ ∈ y′ · o(x, z′ ) ⇒ ∃z ∈ y · o(x, z) ∧ c(x, z) ≤ c(x, z′ )
(2.2)
To derive dominance relations, it is often useful to first derive a semi-congruence relation [15]. A
semi-congruence between two partial solutions y and y′ , written y y′ , ensures that any way of extending
1 there is a covariant relationship between an element of R
b and of R. For example, the initial space, corresponding to all
possible paths, is the empty list.
Theory and Techniques for Synthesizing a Family of Graph Algorithms
36
y′ into a feasible solution can also be used to extend y into a feasible solution. Like ⋔,
is a ternary
b × Rb but as we have done with ⋔ and many other such relations in this work, we drop
relation over D × R
the input argument when there is no confusion and write it as a binary relation for readability. Before
defining semi-congruence, we introduce two concepts. One is the idea of useability of a space. A space y
is is useable, written o∗ (x, y), if ∃z. χ (y, z) ∧ o(x, z), meaning a feasible solution can be extracted from the
space. The second is the notion of incorporating sufficient information into a space to make it useable.
This is defined by an operator ⊕ : Rb × t → Rb that takes a space and some additional information of type
b For example if Rb is the type of lists, then t
t and returns a more defined space. The type t depends on R.
might also be the same type. Now the formal definition of semi-congruence is:
y
y′ ⇒ o∗ (x, y′ ⊕ e) ⇒ o∗ (x, y ⊕ e)
That is, y
y′ is a sufficient condition for ensuring that if y′ can be extended into a feasible solution
than so can y with the same extension. If c is compositional (that is, c(s ⊕ t) = c(s) + c(t)) then it can be
shown [9] that if y y′ and y is cheaper than y′ , then y dominates y′ (written y ⊲ y′ ). Formally:
y
y′ ∧ c(x, y) ≤ c(x, y′ ) ⇒ y ⊲ y′
(2.3)
The axioms given above extend GS-theory [15].
Example 3. Single Pair Shortest Path. If there are two paths p and p′ leading from the start node, if p
and p′ both terminate in the same node then p
p′ . The reason is that any path extension e (of type
t = [Edge]) of p′ that leads to the target node is also a valid path extension for p. Additionally if p is
shorter than p′ then p dominates p′ , which can be discarded. Note that this does not imply that p leads
to the target node, simply that no optimal solutions are lost in discarding p′ . This dominance relation is
formally derived in Eg. 8
Example 4. 0-1 Knapsack
The 0-1 Knapsack problem is, given a set of items each of which has a weight and utility and a
knapsack that has some maximum weight capacity, to pack the knapsack with a subset of items that
maximizes utility and does not exceed the knapsack capacity. Given combinations k, k′ , if k and k′ have
both examined the same set of items and k weighs less than k′ then any additional items e that can be
feasibly added to k′ can also be added to k, and therefore k k′ . Additionally if k has at least as much
utility as k′ then k ⊲ k′ .
The remaining sections cover the original contributions of this paper.
3
A Theory Of Efficient Breadth-First Search (EBFS)
While search can in principle solve for any computable function, it still leaves open the question of how
to carry it out effectively. Various search strategies have been investigated over the years; two of the
most common being Breadth-First Search (BFS) and Depth-First Search (DFS). It is well known that
BFS offers several advantages over DFS. Unlike DFS which can get trapped in infinite paths2 , BFS will
always find a solution if one exists. Secondly, BFS does not require backtracking. Third, for deeper trees,
BFS will generally find a solution at the earliest possible opportunity. However, the major drawback of
BFS is its space requirement which grows exponentially. For this reason, DFS is usually preferred over
BFS.
2 resolvable
in DFS with additional programming effort
S.Nedunuri, W. R. Cook & D.R. Smith
37
Our first contribution in this paper is to refine GS-theory to identify the conditions under which a
BFS algorithm can operate space-efficiently. The key is to show how the size of the undominated frontier
of the search tree can be polynomially bounded. Dominance relations are the basis for this.
In [15], the relation ⋔l for l ≥ 0 is recursively defined as follows:
y ⋔0 y′ = (y = y′ )
y ⋔l+1 y′ = ∃y′′ · y ⋔ y′′ ∧ y′′ ⋔l y′
From this the next step is to define those spaces at a given frontier level that are not dominated. However,
this requires some care because dominance is a pre-order, that is it satisfies the reflexivity and transitivity
axioms as a partial order does, but not the anti-symmetry axiom. That is, it is quite possible for y to
dominate y′ and y′ to dominate y but y and y′ need not be equal. An example in Shortest Path is two paths
of the same length from the start node that end at the same node. Each path dominates the other. To
eliminate such cyclic dominances, define the relation y ≈ y′ as y ⊲ y′ ∧ y′ ⊲ y. It is not difficult to show
that ≈ is an equivalence relation. Now let the quotient frontier at level l be the quotient set frontierl / ≈
. For type consistency, let the representative frontier rfrontierl be the quotient frontier in which each
equivalence class is replaced by some arbitrary member of that class. The representative frontier is the
frontier in which cyclic dominances have been removed. Finally then the undominated frontier undoml
is rfrontierl − {y | ∃y′ ∈ rfrontierl · y′ ⊲ y}.
Now given a problem in the GS class, if it can be shown that kundoml k for any l is polynomially
bounded in the size of the input, a number of benefits accrue: (1) BFS can be used to tractably carry
out the search, as implemented in the raw program schema of Alg. 1, (2) The raw schema of Alg. 1
can be transformed into an efficient tail recursive form, in which the entire frontier is passed down and
(3) If additionally the tree depth can be polynomially bounded (which typically occurs for example in
constraint satisfaction problems or CSPs [3]) then, under some reasonable assumptions about the work
being done at each node, the result is a polynomial-time algorithm for the problem.
3.1 Program Theory
A program theory for EBFS defines a recursive function which given a space y, computes a non-trivial
subset Fx (y) of the optimal solutions contained in y, where
Fx (y) = optc {z | z ∈ y ∧ o(x, z)}
optc is a subset of its argument that is the optimal set of solutions (w.r.t. the cost function c), defined as
follows:
optc S = {z | z ∈ S ∧ (∀z′ ∈ S · c(z) ≤ c(z′ ))}
Also let undom(y) be undoml(y)+1 ∩ {yy | y ⋔ yy} where l(y) is the level of y in the tree. The following
proposition defines a recurrence for computing the feasible solutions in a space:
b o, c, ⊥, ⋔, χ , ⊲, ⊕i be a well-founded GS-Theory w.r.t. the subspace relation
Proposition 5. Let hD, R, R,
⋔ and let Fx (y) = {z | z ∈ y ∧ o(x, z)} be the set of feasible solutions contained in y and Gx (y) = {z |
S
χ (y, z) ∧ o(x, z)} ∪ yy⋔y Gx (yy)} be a recurrence. Then Gx (y) = Fx (y) for any y
Proof. See [15].
Finally he following theorem defines a recurrence that can be used to compute FOx (y):
38
Theory and Techniques for Synthesizing a Family of Graph Algorithms
Theorem 6. Let ⋔ be a well-founded relation of GS-theory and let GOx (y) = optc {z | χ (y, z) ∧ o(x, z)} ∪
yy∈undom(y) GOx (yy)) be a recurrence. Then GOx (y) ⊆ FOx (y)
S
Proof. By generalized induction. The base case is those spaces which do not have subspaces. Then
GOx (y) = optc {z | χ (y, z) ∧ o(x, z)}. By Prop. 5 {z | χ (y, z) ∧ o(x, z)} = {z | z ∈ y ∧ o(x, z)}. The inductive
case is as follows:
FOx (y)
= {defn}
optc {z | z ∈ y ∧ o(x, z)}
= {defn of Fx }
opt(Fx (y))
= {Fx (y) = Gx (y) by Prop.5}
S
opt({z | χ (y, z) ∧ o(x, z)} ∪ y⋔yy Gx (yy))
= {Gx (yy) = Fx (yy) by Prop.5}
S
opt({z | χ (y, z) ∧ o(x, z)} ∪ y⋔yy Fx (yy))
= {distributivity of opt}
S
opt(opt{z | χ (y, z) ∧ o(x, z)} ∪ opt( y⋔yy Fx (yy))) by
= {distributivity and idempotence of opt}
S
opt({z | χ (y, z) ∧ o(x, z)} ∪ y⋔yy opt(Fx (yy)))
= {unfold defn of Fx , fold defn of FOx }
S
opt({z | χ (y, z) ∧ o(x, z)} ∪ y⋔yy FOx (yy))
⊇ {yy ∈ undom(y) ⇒ y ⋔ yy}
S
opt({z | χ (y, z) ∧ o(x, z)} ∪ yy∈undom(y) FOx (yy))
⊇ {induction hypothesis:FOx (yy) ⊇ GOx (yy)}
S
opt({z | χ (y, z) ∧ o(x, z)} ∪ yy∈undom(y) GOx (yy))
= {fold defn of GOx }
GOx (y)
The theorem states that if the feasible solutions immediately extractable from a space y are combined
with the solutions obtained from GOx of each undominated subspace yy, and the optimal ones of those
retained, the result is a subset of FOx (y). The next theorem demonstrate non-triviality3 of the recurrence
by showing that if a feasible solution exists in a space, then one will be found.
Theorem 7. Let ⋔ be a well-founded relation of GS-Theory and GOx be defined as above. Then
FOx (y) 6= 0/ ⇒ GOx (y) 6= 0/
Proof. The proof of Theorem 6 is a series of equalities except for two steps. It is sufficient to show that
both of these steps preserve non-triviality. The proof is again by induction over the subspace relation.
S
S
The first refinement reduces y⋔yy FOx (yy) to yy∈undom(y) FOx (yy). Suppose ∃yy · y ⋔ yy ∧ FOx (yy) 6= 0.
/
If yy∈ undom(y) then we are done. Otherwise if yy is dominated, then there is some yy′ ⊲ yy and by the
S
property of dominance, FOx (yy′ ) 6= 0,
/ so yy∈undom(y) FOx (yy) 6= 0.
/ The second refinement follows again
by induction, using the induction hypothesis FOx (yy) 6= 0/ ⇒ GOx (yy) 6= 0.
/
From the characteristic recurrence we can straightforwardly derive a simple recursive function bfs
to compute a non-trivial subset of Fx for a given y, shown in Alg. 1
3 Non-triviality is similar but not identical to completeness. Completeness requires that every optimal solution is found by
the recurrence, which we do not guarantee.
S.Nedunuri, W. R. Cook & D.R. Smith
39
Algorithm 1 pseudo-Haskell Program Schema for EBFS (schema parameters underlined)
solve :: D -> {R}
solve(x) = bfs x {initial(x)}
bfs :: D -> {RHat}-> {R}
bfs x frontier =
let localsof y = let z = extract x y
in if z!={} && o(x,z) then z else {}
locals = (flatten.map) localsof frontier
allsubs = (flatten.map) (subspaces x) frontier
undom = {yy : yy∈allsubs &&
(yy’∈subs && yy’ ‘dominates‘ yy ⇒ yy==yy’)}
subsolns = bfs x undom
in opt(locals ∪ subsolns)
subspaces :: D -> RHat -> {RHat}
subspaces x y = {yy: split(x,y,yy))
opt :: {R} -> {R}
opt zs = min {c x z | z ∈zs}
The final program schema that is included in the Specware library is the result of incorporating a
number of other features of GS such as necessary filters, bounds tests, and propagation, which are not
shown here. Details of these and other techniques are in [15].
3.2 A class of strictly greedy algorithms (SG)
A greedy algorithm [2] is one which repeatedly makes a locally optimal choice. For some classes of
problems this leads to a globally optimum choice. We can get a characterization of optimally greedy
algorithms within EBFS by restricting the size of undoml for any l to 1. If undoml 6= 0/ then the singleton
member y∗ of undoml is called the greedy choice. In other work [12] we show how to derive greedy
algorithms for a variety of problems including Activity Selection, One machine scheduling, Professor
Midas’ Traveling Problem, Binary Search.
4
Methodology
We strongly believe that every formal approach should be accompanied by a methodology by which it
can be used by a competent developer, without needing great insights. Guided program synthesis already
goes a long way towards meeting this requirement by capturing design knowledge in a reusable form. The
remainder of the work to be done by a developer consists of instantiating the various parameters of the
program schema. In the second half of this paper, we demonstrate how carrying this out systematically
allows us to derive several related graph algorithms, revealing connections that are not always obvious
from textbook descriptions. We wish to reiterate that once the dominance relation and other operators
in the schema have been instantiated, the result is a complete solution to the given problem. We focus
on dominance relations because they are arguably the most challenging of the operators to design. The
remaining parameters can usually be written down by visual inspection.
The simplest form of derivation is to reason backwards from the conclusion of y y′ ⇒ o∗ (x, y′ ⊕
e) ⇒ o∗ (x, y ⊕ e), while assuming o∗ (x, y′ ⊕ e) . The additional assumptions that are made along the way
form the required semi-congruence condition. The following example illustrates the approach.
40
Theory and Techniques for Synthesizing a Family of Graph Algorithms
o∗ (x, y ⊕ e)
= {defn of o∗ }
∃z · χ (y ⊕ e, z) ∧ o(x, z)
= {defn of χ }
o(x, y ⊕ e)
= {defn of o}
path?(y ⊕ e, x.s, x.e)
= {distributive law for path?}
∃n · path?(y, x.s, n) ∧ path?(e, n, x.e)
⇐ {o∗ (x, y′ ⊕ e), ie.∃m · path?(y′ , x.s, m) ∧ path?(e, m, x.e). Let m be witness for n}
path?(y, x.s, m) ∧ path?(e, m, x.e)
= {m = last(y).t, (where last returns the last element of a sequence)}
last(y).t = last(y′ ).t ∧ path?(y, x.s, n)
Figure 4.1: Derivation of semi-congruence relation for Single Pair Shortest Path
Example 8. Derivation of the semi-congruence relation for Single Pair Shortest Path in Eg. 1 is a
straightforward calculation as shown in Fig 4.1. It relies on the specification of Shortest Path given in
Eg. 1 and the GS-theory in Eg. 2.
The calculation shows that a path y is semi-congruent to y′ if y and y′ both end at the same node and
additionally y is itself a valid path from the start node to its last node. Since the cost function is compositional, this immediately produces a dominance relation y ⊲ y′ = last(y) = last(y′ ) ∧ path?(y, x.s, n) ∧
∑edge∈y edge.w ≤ ∑edge′ ∈y′ edge′ .w. Note the use of the distributive law for path? in step 4. Such laws are
usually formulated as part of a domain theory during a domain discovery process, or even as part of the
process of trying to carry out a derivation such as the one just shown. Given an appropriate constructive
prover (such as the one in KIDS [16]) such a derivation could in fact be automated. Other examples that
have been derived using this approach are Activity Selection [11], Integer Linear Programming [15], and
variations on the Maximum Segment Sum problem [10].
While this dominance relation could in principle be used to computer Single Source Shortest Path
using a Best-First search (such as A*) it would not be very efficient as every pair of nodes on the frontier
would need to be compared. In the next section, a more powerful dominance relation is derived which can
be used in a Breadth-First search, and even more importantly, be shown to be in the SG class, resulting in
a very efficient greedy algorithm. The dominance relation just derived is still utilized, but in a subsidiary
role.
5
More complex derivations: A family of related algorithms
5.1 (Single Source) Shortest Path
Previously in Eg. 8 we derived a dominance relation for the (undirected) single pair shortest path problem. To solve the problem of finding all shortest paths from a given start node to every other node in
the graph it is convenient to consider the output as a set of edges that form what is called a path tree, a
subgraph of the input graph which forms a spanning tree rooted at the start node. The desired output is a
path tree in which every path from the root is the shortest. The specification of Single Pair Shortest Path
in Fig. 2.1 is revised as shown in Fig. 5.1 The revised instantiation of Global Search theory is shown
in Fig. 5.2 In what follows, the extends operator ⊕ is shown by simple concatenation. The goal is to
S.Nedunuri, W. R. Cook & D.R. Smith
41
D 7→ hs : Node, edges : {Edge}i
Edge = ha : Node, b : Node, w : Nati
R 7→ {Edge}
o 7→ λ (x, z) · connected(x, z) ∧ acyclic(x, z)
c 7→ λ (x, z) · ∑ p∈pathsFrom(x.s) c′ (p)
c′ (p) = ∑edge∈p edge.w
Figure 5.1: Specification of Shortest Path problem
Rb
⊥
⋔
χ
⊕
7→
7→
7→
7→
7→
R
λ x · {}
λ (x, p, pe) · ∃e ∈ x.edges · pe = p ∪ {e}
λ (z, p) · p = z
∪
Figure 5.2: GS instantiation for Shortest Path
show that there is at most one undominated child following a split of a partial solution α . Let α e and
α e′ be two children following a split of α , that is the graphs α with edge e added and that with e′ added.
Without loss of generality (w.l.o.g.) assume neither e nor e′ are already contained in α and both connect
to α . Let z′ = α e′ ω ′ be a feasible solution derived from α e′ . The task is to construct a feasible solution
z from α e and discover the conditions under which it is cheaper than z′ . We will use the definition of
general dominance (2.2), repeated here for convenience:
∀z′ ∈ y′ · o(x, z′ ) ⇒ ∃z ∈ y · o(x, z) ∧ c(x, z) ≤ c(x, z′ )
Establishing o(α eω ) requires connected(α eω ) and acyclic(α eω ).
In guided program synthesis, it is often useful to write down laws [16] that will be needed during the
derivation. Some of the most useful laws are distributive laws and monotonicity laws. For example the
following distributive law applies to ayclic :
acyclic(αβ ) = acyclic(α ) ∧ acyclic(β ) ∧ ac(α , β )
where ac defines what happens at the “boundary” of α and β :
ac(α , β ) = ∀m, n · ∃p ∈ α ∗ · path?(p, m, n) ⇒ ¬∃q ∈ β ∗ · path?(q, m, n)
requiring that if p is a path in α (α ∗ is a regular expression denoting all possible sequences of edges in
α ) connecting m and n then there should be no path between m and n in β . Examples of monotonicity laws are acyclic(αβ ) ⇒ acyclic(α ) and connected(α ) ∧ connected(β ) ∧ nodes(α ) ∩ nodes(β ) 6=
0/ ⇒ connected(αβ ). By using the laws constructively, automated tools such as KIDS [16] can often
suggest instantiations for terms. For instance, after being told connected(α e′ ω ′ ), the tool could apply
the monotonicity law for connected to suggest connected(α ee′ ω ′ ), that is to try ω = e′ ω ′ . With this
binding for ω we can attempt to establish acyclic(α eω ), by expanding its definition. One of the terms
is . . . ac(e, α e′ ω ′ ) . . . which fails because conn(α e′ ω ′ ) implies path?(α e′ ω ′ , e.a, e.b) so adding edge e
creates a cycle. The witness to the failure is some path π ′ = e j . . . ek where e j .a = e.a ∧ ek .b = e.b.
One possibility is that e j = ek = e, that is ω ′ contains e. If so, let ω ′ = eψ ′ for some ψ ′ . Then
42
Theory and Techniques for Synthesizing a Family of Graph Algorithms
Figure 5.3: Feasible solution α e′ ω ′
z′ = α e′ ω ′ = α e′ eψ ′ = α ee′ ψ ′ . Let ω = e′ ψ ′ and now z = z′ . Otherwise, w.l.o.g assume that e connects with α at e.a, and therefore so does e j , so the case e j .b = e.b is not very interesting. The only
option then is to remove edge ek from ω ′ . Let ω ′ = ek ψ ′ and so ω is e′ ψ ′ . Now the requirement
becomes
acyclic(α e) ∧ acyclic(e′ ψ ′ ) ∧ ac(α e, e′ ψ ′ )
acyclic(e′ ψ ′ ) follows from acyclic(α e′ ψ ′ ) by monotonicity and ac(α e, e′ ψ ′ ) follows from acyclic(α e′ ψ ′ )
and the fact that e was chosen above to remove the cycle in α ee′ ω ′ . This demonstrates o(α eω ) provided
acyclic(α e) where ω is constructed as above.
Finally, to establish general dominance it is necessary to show that z costs no more than z′ . Note that
the cost of a solution is the sum of individual path costs starting from x.s. Let m denote e.a and n denote
e.b (and analogously for e′ ). Now consider a path to a node p in z′ . If the path to p does not contain edge
ek ie. pass through n then the same path holds in z. Otherwise let β ′ e′i γ ′ e”δ be a path to p in z′ where e”
is the edge ek above (see Fig. 5.3). β ′ is a path from x.s in α and e′i is some edge that leads out of α on
the path to p. Then the corresponding path in z is β eδ (see Fig. 5.4).
c(α eω , β eδ ) ≤ c(α e′ ω ′ , β ′ ei γ ′ e”δ )
= {expand defns}
c(z, β e) + c(z, δ ) ≤ c(z′ , β ′ ei ) + c(z′ , e”) + c(z, δ )
= {+ve edge weights, triangle inequality}
c(z, β e) ≤ c(z′ , β ′ ei )
As there were no restrictions on e′ above, let ei be the witness for e′ and this establishes That is, provided
the path β e is shorter than β ′ e′ , there cannot be a shorter path via e′ to n. As cost is the sum of the path
costs, it follows that c(x, α eω ) ≤ c(α e′ ω ′ ). The dominance condition is then α e ⊲ α e′ ⇐ acyclic(α e) ∧
c(z, β e) ≤ c(z′ , β ′ e′ ) . Finally, as it is not known at the time of the split which e′ will lie on the path to e.b,
to be conservative, let e be that edge whose endpoint e.b is the closest to the start node. This is therefore
the greedy choice. Iincorporating finite differencing to incrementally maintain the distances to the nodes,
using the dominance relation derived earlier for Single Pair Shortest Path to eliminate the longer paths to
a node, and data structure refinement results in an algorithm similar to Dijkstra’s algorithm for MSTs.
S.Nedunuri, W. R. Cook & D.R. Smith
43
Figure 5.4: Feasible solution α eω
D 7→ {Edge} | connected
Edge = h f : Node,t : Node, w : Nati
R 7→ {Edge}
o 7→ λ (x, z) · connected(x, z) ∧ acyclic(x, z)
c 7→ λ (x, z) · ∑edge∈z edge.w
Figure 5.5: Specification of Min. Spanning Tree problem
5.2 Minimum Spanning Tree
The specification of MST is very similar to that of Shortest Path, with the difference that there is no
longer a distinguished node s, the input graph must be connected, and the cost of a solution is simply
the sum of the weights of the edges in the tree The instantiation of the simple GS operators is as for SP.
Again, any edge that is added must not create a cycle or it cannot lead to a feasible solution. We will
describe the algorithm construction process informally so as to expose the connection with the Shortest
Path algorithm more clearly. However the derivations shown here can also be similarly formalized.
There are now two ways to satisfy the acyclicity requirement. One is by choosing an edge connecting
a node in α to one outside of α . Another is to choose an edge that connects two nodes within α , being
careful not to create cycles. The two options are examined next,
Option 1: Let z′ = α e′ ω ′ be a feasible solution derived from α e′ . If ω ′ includes e then let ω in a
feasible solution z = α eω simply be ω ′ − {e} ∪ {e′ } and then z = z′ . Otherwise, if ω ’ does not contain
e there must be some other path connecting α with e.t. W.l.o.g. assume that path is via e′ . If α e′ ω ′ is
feasible, then it is a tree, so ω ’ is also a tree. Therefore it is not difficult to show that z = α eω ′ is also a
spanning tree. Now to show dominance, derive conditions under which z is cheaper than z′ :
c(x, α eω ′ ) ≤ c(x, α e′ ω ′ )
= {defn of c}
∑edge∈α eω ′ edge.w ≤ ∑edge∈α e′ ω ′ edge.w
=
e.w ≤ e′ .w
Theory and Techniques for Synthesizing a Family of Graph Algorithms
44
Finally, as it is not known at the time of the split which e′ will lie on the path to e.t, to be conservative,
let e be that edge with the least weight connecting α with an external node . This is therefore the greedy
choice. The result is an algorithm that is similar to Prim’s algorithm for MSTs.
Option 2: The difference with Option 1 is in how e is chosen in order to ensure acyclicity. For a
feasible solution, α must not contain any cycles. Therefore it consists of a collection of acyclic connected
components, ie trees. Any new edge cannot connect nodes within a component without introducing a
cycle. Therefore it must connect two component trees. Connecting two trees by a single edge results in
a new tree. As in Option 1, let z′ = α e′ ω ′ be a feasible solution derived from α e′ . If ω ′ includes e then
let ω in a feasible solution z = α eω simply be ω ′ − {e} ∪ {e′ } and then z = z′ . Otherwise, if ω ’ does
not contain e there must be some other edge used to connect the two trees that e would have connected.
W.l.o.g. assume that edge is e′ . If α e′ ω ′ is feasible, then it is a tree, so ω ’ is also a tree. Therefore it is
not difficult to show that z = α eω ′ is also a spanning tree. The derivation of a cost comparison relation
is identical to Option 1, and once again the greedy choice is the edge e that connects two trees and is of
least weight. The result of this option is an algorithm that is similar to Kruskal’s algorithm.
In conclusion, we observe that far from being completely different algorithms, Dijkstra’s algorithm,
Prim’s algorithm and Kruskal’s algorithm differ only in very small number of (albeit important) ways. In
contrast, many textbook descriptions of the algorithms introduce the algorithms out of the blue, followed
by separate proofs of correctness. We have shown how a systematic procedure can derive different
algorithms, with relatively minor changes to the derivations.
6
Related Work
Gulwani et al. [21, 4] describe a powerful program synthesis approach called template-based synthesis.
A user supplies a template or outline of the intended program structure, and the tool fills in the details. A
number of interesting programs have been synthesized using this approach, including Bresenham’s line
drawing algorithm and various bit vector manipulation routines. A related method is inductive synthesis
[6] in which the tool synthesizes a program from examples. The latter has been used for inferring spreadsheet formulae from examples. All the tools rely on powerful SMT solvers. The Sketching approach of
Solar-Lezama et al [14] also relies on inductive synthesis. A sketch, similar in intent to a template, is
supplied by the user and the tool fills in such aspects as loop bounds and array indexing. Sketching relies
on efficient SAT solvers. To quote Gulwani et al. the benefit of the template approach is that “the programmer only need write the structure of the code and the tool fills out the details” [21].Rather than the
programmer supplying an arbitrary template, though, we suggest the use of a program schema from the
appropriate algorithm class (refer to Step 2 of the process in Sec. 2.1). We believe that the advantage of
such an approach is that, based on a sound theory, much can already be inferred at the abstract level and
this is captured in the theory associated with the algorithm class. Furthermore, knowledge of properties
at the abstract level allows specialization of the program schema with information that would otherwise
have to either be guessed at by the programmer devising a template or inferred automatically by the tool
(e.g. tail recursive implementation or efficient implementation of dominance testing with hashing). We
believe this will allow semi-automated synthesis to scale up to larger problems such as constraint solvers
(SAT, CSP, LP, MIP, etc.), planning and scheduling, and O/S level programs such as garbage collectors
[13].
Program verification is another field that shares common goals with program synthesis - namely a
correct efficient program. The difference lies in approach - we prefer to construct the program in a way
that is guaranteed to be correct, as opposed to verifying its correctness after the fact. Certainly some
S.Nedunuri, W. R. Cook & D.R. Smith
45
recent tools such as Dafny [8] provide very useful feedback in an IDE during program construction.
But even such tools requires significant program annotations in the form of invariants to be able to
automatically verify non-trivial examples such as the Schorr-Waite algorithm [8]. Nevertheless, we do
not see verification and synthesis as being necessarily opposed. For example, ensuring the correctness of
the instantiation of several of the operators in the program schema which is usually done by inspection
is a verification task, as is ensuring correctness of the schema that goes in the class library. We also feel
that recent advances in verification via SMT solvers will also help guided synthesis by increasing the
degree of automation.
Refinement is generally viewed as an alternative to synthesis. A specification is gradually refined
into an efficient executable program. Refinement methods such as Z and B have proved to be very popular. In contrast to refinement, guided program synthesis already has the program structure in place,
and the main body of work consists of instantiating the schema parameters followed by various program
transformations many of which can be mechanically applied. Both refinement and synthesis rely extensively on tool support, particularly in the form of provers. We expect that advances in both synthesis and
refinement will benefit the other field.
7
Summary and Future Work
We have formulated a theory of efficient breadth-first search based on dominance relations. A very useful
specialization of this class occurs when there is at most one undominated child node. This is the class of
Strictly Greedy algorithms. We have also derived a recurrence from which a simple program schema can
be easily constructed. We have shown how to systematically derive dominance relations for a family of
important graph algorithms revealing connections between them that are obscured when each algorithm
is presented in isolation.
Nearly all the derivations shown in this paper have been carried out by hand. However, they are
simple enough to be automated. We plan on building a prover that incorporates the ideas mentioned
in here. We are encouraged by the success of a similar prover that was part of KIDS, a predecessor to
Specware.
References
[1] Specware. Http://www.specware.org.
[2] T. Cormen, C. Leiserson, R. Rivest & C. Stein (2001): Introduction to Algorithms, 2nd edition. MIT Press.
[3] R Dechter (2003): Constraint Processing. Morgan Kauffman.
[4] S. Gulwani, S. Jha, A. Tiwari & R. Venkatesan (2011): Synthesis of loop-free programs. In: PLDI, pp. 62–73,
doi:10.1145/2F1993498.1993506.
[5] T. Ibaraki (1977): The Power of Dominance Relations in Branch-and-Bound Algorithms. J. ACM 24(2), pp.
264–279, doi:10.1145/2F322003.322010.
[6] S. Itzhaky, S. Gulwani, N. Immerman & M. Sagiv (2010): A simple inductive synthesis methodology and its
applications. In: OOPSLA, pp. 36–46, doi:10.1145/2F1869459.1869463.
[7] C. Kreitz (1998): Program Synthesis. In W. Bibel & P. Schmitt, editors: Automated Deduction – A Basis for
Applications, chapter III.2.5, III, Kluwer, pp. 105–134.
[8] K. R. M. Leino (2010): Dafny: an automatic program verifier for functional correctness. In: Proc. 16th intl.
conf. on Logic for Prog., AI, & Reasoning, LPAR, pp. 348–370, doi:10.1007/2F978-3-642-17511-4_20.
46
Theory and Techniques for Synthesizing a Family of Graph Algorithms
[9] S. Nedunuri (2012): Theory and Techniques for Synthesizing Efficient Breadth-First Search Algorithms. Ph.D.
thesis, Univ. of Texas at Austin.
[10] S. Nedunuri & W.R. Cook (2009): Synthesis of Fast Programs for Maximum Segment Sum Problems. In:
Intl. Conf. on Generative Prog. and Component Engineering (GPCE), doi:10.1145/2F1621607.1621626.
[11] S. Nedunuri, D. R. Smith & W. R. Cook (2010): A Class of Greedy Algorithms and Its Relation to Greedoids.
Intl. Colloq. on Theoretical Aspects of Computing (ICTAC), doi:10.1007/2F978-3-642-14808-8_24.
[12] S. Nedunuri, D. R. Smith & W. R. Cook (2012): Theory and Techniques for a Class of Efficient Breadth-First
Search Algorithms. In: Intl. Symp. on Formal Methods (FM).
[13] D. Pavlovic, P. Pepper & D. R. Smith (2010): Formal Derivation of Concurrent Garbage Collectors. In:
Math. of Program Constr. (MPC), doi:10.1007/2F978-3-642-13321-3_20.
[14] Y. Pu, R. Bodík & S. Srivastava (2011): Synthesis of first-order dynamic programming algorithms. In:
OOPSLA, pp. 83–98, doi:10.1145/2F2048066.2048076.
[15] D. R. Smith (1988): Structure and Design of Global Search Algorithms. Tech. Rep. Kes.U.87.12, Kestrel
Institute.
[16] D. R. Smith (1990): KIDS: A Semi-Automatic Program Development System. IEEE Trans. on Soft. Eng.,
Spec. Issue on Formal Methods 16(9), pp. 1024–1043, doi:10.1109/2F32.58788.
[17] D. R. Smith (2010): Global Search Theory Revisited. Unpublished.
[18] D. R. Smith, E. A. Parra & S. J. Westfold (1995): Synthesis of high-performance transportation schedulers.
Technical Report, Kestrel Institute.
[19] D. R. Smith & S. Westfold (2008): Synthesis of Propositional Satisfiability Solvers. Final Proj. Report,
Kestrel Institute.
[20] A. Solar-Lezama, L. Tancau, R. Bodik, S. Seshia & V. Saraswat (2006): Combinatorial sketching for finite
programs. In: Proc. of the 12th intl. conf. on architectural support for prog. lang. and operating systems
(ASPLOS), pp. 404–415, doi:10.1145/2F1168857.1168907.
[21] S. Srivastava, S. Gulwani & J. S. Foster (2010): From program verification to program synthesis. In: POPL,
pp. 313–326, doi:10.1.1.148.395.
[22] M. Vechev & E. Yahav (2008): Deriving linearizable fine-grained concurrent objects. PLDI ’08, pp. 125–
135, doi:10.1145/2F1375581.1375598.
| 6cs.PL
|
1
Video Denoising and Enhancement via
Dynamic Video Layering
arXiv:1710.02213v1 [cs.CV] 5 Oct 2017
Han Guo and Namrata Vaswani
Iowa State University, Ames, IA, USA
Email: {hanguo,namrata}@iastate.edu
Abstract—Video denoising refers to the problem of removing
“noise” from a video sequence. Here the term “noise” is used in a
broad sense to refer to any corruption or outlier or interference
that is not the quantity of interest. In this work, we develop
a novel approach to video denoising that is based on the idea
that many noisy or corrupted videos can be split into three
parts - the “low-rank layer”, the “sparse layer”, and a small
residual (which is small and bounded). We show, using extensive
experiments, that our denoising approach outperforms the stateof-the-art denoising algorithms.
I. I NTRODUCTION
Video denoising refers to the problem of removing “noise”
from a video sequence. Here the term “noise” is used in a
broad sense to refer to any corruption or outlier or interference
that is not the quantity of interest.
In the last few decades there has been a lot of work
on video denoising. Many of the approaches extend image
denoising ideas to 3D by also exploiting dependencies across
the temporal dimension. An important example of this is
“grouping and collaborative filtering” approaches which try
to search for similar image patches both within an image
frame and across nearby frames, followed by collaboratively
filtering the noise from the stack of matched patches [2], [3],
[4], [5], [6]. One of the most effective methods for image
denoising, Block Matching and 3D filtering (BM3D) [3], is
from this category of techniques. In BM3D, similar image
blocks are stacked in a 3D array followed by applying a noise
shrinkage operator in a transform domain. In its video version,
VBM3D [7], the method is generalized to video denoising
by searching for similar blocks across multiple frames. Other
related works [8], [9] apply batch matrix completion or matrix
decomposition on grouped image patches to remove outliers.
Other recent works on video denoising include approaches
that use motion compensation algorithms from the video
compression literature followed by denoising of similar nearby
blocks [10], [11]; and approaches that use wavelet transform
based [12], [13], [14] and discrete cosine transform (DCT)
based [15] denoising solutions. Very recent video denoising
methods include algorithms based on learning a sparsifying
transform [16], [17], [18], [19]. Within the image denoising
literature, the most promising recent approaches are based on
deep learning [20], [21].
This work was partially supported by Rockwell Collins. A portion of this
work was presented at IEEE Workshop on Statistical Signal Processing (SSP)
2016 [1].
Contribution. In this paper, we develop a novel approach
to video denoising, called ReProCS-based Layering Denoising
(ReLD), that is based on the idea that many noisy or corrupted
videos can be split into three parts or layers - the “low-rank
layer”, the “sparse layer” and the “small bounded residual
layer”. Here these terms mean the following. “Low-rank layer”
ℓt : the matrix formed by each vectorized image of this layer
is low-rank. “Sparse layer” st : each vectorized image of
this layer is sparse. “Small bounded residual layer”: each
vectorized image of this layer has l∞ norm that is bounded.
Our proposed algorithm consists of two parts. At each time
instant, it first separates the video into “noisy” versions of
the two layers ℓt and st . This is followed by applying an
existing state-of-the-art denoising algorithm, VBM3D [7] on
each layer. For separating the video, we use a combination of
an existing batch technique for sparse + low-rank matrix decomposition called principal components’ pursuit (PCP) [22]
and an appropriately modified version of our recently proposed
dynamic robust PCA technique called Recursive Projected
Compressive Sensing (ReProCS) [23], [24]. We initialize using
PCP and use ReProCS afterwards to separate the video into
a “sparse layer” and a “low-rank layer”. The video-layering
step is followed by VBM3D on each of the two layers. In
doing this, VBM3D exploits the specific characteristics of
each layer and, hence, is able to find more matched blocks
to filter over, resulting in better denoising performance. The
motivation for picking ReProCS for the layering task is its
superior performance in earlier video experiments involving
videos with large-sized sparse components and/or significantly
changing background images.
The performance of our algorithm is compared with PCP
[22], GRASTA [25] and non-convex rpca (NCRPCA) [26],
which are some of the best solutions from the existing sparse
+ low-rank (S+LR) matrix decomposition literature [22], [25],
[26], [27], [28], [29], [30], [31], [32] followed by VBM3D for
the denoising step; as well as with just using VBM3D directly
on the video. We also compare with one neural network based
image denoising method that uses a Multi Layer Perceptron
(MLP) [33] and with the approach of [9] which performs
standard sparse + low-rank (S+LR) matrix approximation
(SLMA) on grouped image patches. As we show, our approach
is 6-10 times faster than SLMA while also having improved
performance. The reason is we use a novel online algorithm
(after a short batch initialization) for S+LR and then use
VBM3D on each layer.
2
A. Example applications
A large number of videos that require denoising/enhancement
can
be
accurately
modeled
in the above fashion. Some examples are as
follows. All videos referenced below are posted at
http://www.ece.iastate.edu/∼hanguo/denoise.html.
1) In a traditional denoising scenario, consider slowly
changing videos that are corrupted by salt-and-pepper
noise (or other impulsive noise). For these types of
videos, the large magnitude part of the noise forms
the “sparse layer”, while the video-of-interest (slowlychanging in many applications, e.g., waterfall, waving
trees, sea water moving, etc) forms the approximate
“low-rank layer”. The approximation error in the lowrank approximation forms the “small bounded residual”.
See the waterfall-salt-pepper video for an example. The
goal is to denoise or extract out the “low-rank layer”.
2) More generally, consider slow-changing videos corrupted by very large variance white Gaussian noise. As
we explain below, large Gaussian noise can, with high
probability, be split into a very sparse noise component
plus bounded noise. Thus, our approach also works on
this type of videos, and in fact, in this scenario, we show
that it significantly outperforms the existing state-of-theart video denoising approaches. (See Fig. 4a.)
3) In very low-light videos of moving targets/objects (the
moving target is barely visible), the denoising goal is to
“see” the barely visible moving targets (sparse). These
are hard to see because they are corrupted by slowlychanging background images (well modeled as forming
the low-rank layer plus the residual). The dark-room
video is an example of this. The goal is to extract out
the sparse targets or, at least, the regions occupied by
these objects. (See Fig. 3.)
Moreover, in all these examples, it is valid to argue that
the columns of the low-rank matrix lie in a low-dimensional
subspace that is either fixed or slowly changing. This is
true, for example, when the background consists of moving
waters, or the background changes are due to illumination
variations. These also result in global (non-sparse) changes.
In special cases where foreground objects are also present, the
video itself become “low-rank + sparse”. In such a scenario,
the “sparse layer” that is extracted out will consist of the
foreground object and the large magnitude part of the noise.
Some examples are the curtain and lobby videos. The proposed
ReLD algorithm works for these videos if VBM3D applied
to the foreground layer video is able to separate out the
foreground moving object(s) from the noise.
B. Problem formulation
Let mt denote the image at time t arranged as a 1D vector
of length n. We consider denoising for videos in which each
image can be split as
mt = ℓt + st + wt
where st is a sparse vector, ℓt ’s lie in a fixed or slowly
changing low-dimensional subspace of Rn so that the matrix
L := [ℓ1 , ℓ2 , . . . , ℓtmax ] is low-rank, and wt is the residual
noise that satisfies kwt k∞ ≤ bw . We use Tt to denote the
support set of st , i.e., Tt := support(st ).
In the first example given above, the moving targets’ layer is
st , the slowly-changing dark background is ℓt + wt . The layer
of interest is st . In the second example, the slowly changing
video is ℓt + wt , while the salt-and-pepper noise is st . The
layer of interest is ℓt . In the third example, the slowly changing
video is ℓt + w1,t with w1,t being the residual; and, as we
explain next, with high probability (whp), white Gaussian
noise can be split as st + w2,t with w2,t being bounded. In
this case, wt = w1,t + w2,t .
Let n denote a Gaussian noise vector in Rn with zero mean
and covariance σ 2 I. Let β(b) := 2Φ(b) − 1 with Φ(z) being
the cumulative distribution function (CDF) of the standard
Gaussian distribution. Then, it is not hard to see that n can
be split as
n=s+w
where w is bounded noise with kwk∞ ≤ b0 and
s is a sparse
vector with support size |Tt | ≈ 1 − β bσ0 n whp. More
precisely, with probability at least 1 − 2 exp(−2ǫ2 n),
b0
b0
− ǫ n ≤ |Tt | ≤ 1 − β
+ ǫ n.
1−β
σ
σ
In words, whp, s is sparse with support size roughly (1 − β)n
where β = β bσ0 . The above claim is a direct consequence
of Hoeffding’s inequality for a sum of independent Bernoulli
random variables1.
II. R E P RO CS- BASED L AYERING D ENOISING (R E LD)
We summarize the ReProCS-based Layering Denoising
(ReLD) algorithm in Algorithm 1 and detail each step in
Algorithm 3. The approach is explained below.
Algorithm 1 Overall ReLD Algorithm
1) For t < t0 , initialization using PCP [22].
2) For all t > t0 , implement an appropriately modified
ReProCS algorithm
a) Split the video frame mt into layers ℓ̂t and ŝt
b) For every α frames, perform subspace update, i.e.,
update P̂t
3) Denoise using VBM3D
Initialization. Take M0 = [m1 , m2 , · · · , mt0 ] as training
data and use PCP [22] to separate it into a sparse matrix
[ŝ1 , ŝ2 , · · · , ŝt0 ] and a low-rank matrix [ℓ̂1 , ℓ̂2 , · · · , ℓ̂t0 ]. Compute the top b% left singular vectors of [ℓ̂1 , ℓ̂2 , · · · , ℓ̂t0 ] and
denote by P̂0 . Here b% left singular vectors of a matrix M
refer to the left singular vectors of M whose corresponding
1 If
p is the probability of zi = 1, then Hoeffding’s inequality says that:
X
Pr((p − ǫ)n ≤
zi ≤ (p + ǫ)n) ≥ 1 − 2 exp(−2ε2 n)
i
We apply it to the Bernoulli random variables zi ’s with zi defined as zi = 1
if {si 6= 0} and zi = 0 if {si = 0}. Clearly, Pr(zi = 0) = Pr(si = 0) =
Pr(n2i ≤ b20 ) = Φ(b0 /σ) − Φ(−b0 /σ) = 2Φ(b0 /σ) − 1 = β(b0 /σ).
3
Algorithm 2 Details of each step of ReLD
Parameters: We used α = 20, Kmin = 3, Kmax = 10, t0 =
50 in all experiments.
1) Initialization
using
PCP
[22]:
Compute
(L̂0 , Ŝ0 ) ← PCP(M0 ) and compute [P̂0 , Σ̂0 ] ←
approx-basis(L̂0 , 90%). The notation PCP(M ) means
implementing the PCP algorithm on matrix M and P
= approx-basis(M , b%) means that P is the b% left
singular vectors’ matrix for M .
Set r̂ ← rank(P̂0 ), σ̂min ← (Σ̂0 )r̂,r̂ , t̂0 = t0 ,
flag=detect
2) For all t > t0 , implement an appropriately modified
ReProCS algorithm
a) Split mt into layers ℓ̂t and ŝt :
i) Compute yt ← Φt mt with Φt ← I −
′
P̂t−1 P̂t−1
ii) Compute ŝt as the solution of
min kxk1 s.t.kyt − Φt xk2 ≤ ξ
x
with ξ = kΦt ℓ̂t−1 k
p
iii) T̂t ← Thresh(ŝt , ω) with ω = 3 kmt k2 /n.
Here T ← Thresh(x, ω) means that T =
{i : |(x)i | ≥ ω}
ŝt,∗ ← LS(yt , Φt , T̂t ). Here x̂ ← LS(y, A, T )
means that x̂T = (A′T AT )−1 A′T y, which is
least-squared estimate of x on T .
iv) ℓ̂t ← mt − ŝt , ℓ̂t,∗ ← mt − ŝt,∗
b) Perform subspace update, i.e., update P̂t (see details in supplementary material)
3) Denoise using VBM3D:
a) σ̂fg ← Std-est([ŝt , . . . , ŝt0 ])
σ̂bg ← Std-est([ℓ̂t , . . . , ℓ̂t0 ]). Here Std-est(M )
denotes estimating the standard deviation of noise
from M : we first subtract column-wise mean from
M and then compute the standard deviation by
seeing it as a vector.
b) Ŝdenoised ← VBM3D([ŝ1 , . . . , ŝtmax ], σ̂fg )
L̂denoised ← VBM3D([ℓ̂1 , . . . , ℓ̂tmax ], σ̂bg ). Here
VBM3D(M , σ) implements the VBM3D algorithm on matrix M with input standard deviation σ.
Output: Ŝ, Ŝdenoised , L̂denoised or Iˆdenoised = Ŝdenoised + L̂denoised
based on applications
singular values form the smallest set of singular values that
contains at least b% of the total singular values’ energy.
Splitting phase. Let P̂t−1 be the basis matrix (matrix with
orthonormal columns) for the estimated subspace of ℓt−1 . For
t ≥ t0 +1, we split mt into ŝt and ℓ̂t using prac-ReProCS [24].
To do this, we first project mt onto the subspace orthogonal
to range(P̂t−1) to get the projected measurement vector,
′
)mt := Φt mt .
yt := (I − P̂t−1 P̂t−1
(1)
Observe that yt can be expressed as
yt = Φt st + βt where βt := Φt (ℓt + wt ).
(2)
Because of the slow subspace change assumption, the projection nullifies most of the contribution of ℓt and hence βt is
small noise. The problem of recovering st from yt becomes a
traditional noisy sparse recovery/CS problem and one can use
ℓ1 minimization or any of the greedy or iterative thresholding
algorithms to solve it. We denote its solution by ŝt , and obtain
ℓ̂t by simply subtracting ŝt from mt .
Denoising phase.
We perform VBM3D on Ŝ =
[ŝ1 , . . . , ŝtmax ] and L̂ = [ℓ̂1 , . . . , ℓ̂tmax ] and obtain the denoised
data Ŝdenoised and L̂denoised . Based on applications, we output
different results. For example, in the low-light denoising case,
our output is Ŝ since the goal is to extract out the sparse
targets. In traditional denoising scenarios, the output can
be L̂denoised or Iˆdenoised =Ŝdenoised +L̂denoised . This depends on
whether the video contains only background or background
and foreground. In practice, even for videos with only backgrounds, adding Ŝdenoised helps improve PSNR.
Subspace Update phase (Optional). In long videos the
span of the ℓt ’s will change with time. Hence one needs to
update the subspace estimate P̂t every so often. This can be
done efficiently using the projection-PCA algorithm from [24].
III. E XPERIMENTS
Due to limited space, in this paper we only present
a part of the experimental results. The complete presentation of experimental results are in the supplementary material. Video demos and all tables of peak signal to noise ratio (PSNR) comparison are also available
at http://www.ece.iastate.edu/∼hanguo/denoise.html. Code for
ReLD as well as for all the following experiments is also
posted here.
A. Removing Salt & Pepper noise
First we compare the denoising performance on two
dataset – Curtain and Lobby which are available at
http://www.ece.iastate.edu/∼hanguo/denoise.html. The algorithms being compared are ReLD, SLMA [9], VBM3D and a
neural network image denoising method, Multi Layer Perceptron (MLP) [33]. The codes for algorithms being compared are
downloaded from the authors’ webpages. The available MLP
code contains parameters that are trained solely from image
patches that were corrupted with Gaussian noise with σ = 25
and hence the denoising performance is best with σ = 25 and
deteriorates for other noise levels. The noise being added to
the original image frames are Gaussian (σ = 25) plus 8%
salt and pepper noise. In Fig.6 we show a plot of the framewise PSNRs for the lobby video, which shows that ReLD
outperforms all other algorithms – the PSNR is the highest in
all image frames.
B. Removing Gaussian noise
Next, with different levels of Gaussian noise, we compare performance of our proposed denoising framework
with video layering performed using either ReProCS [24]
(our proposed algorithm), or using the other robust PCA
algorithms - PCP [22], NCRPCA [26], and GRASTA
4
σ
36
50
70
σ
50
70
34
32
PSNR
30
28
RELD: mixed noise
VBM3D: mixed noise
MLP: mixed noise
SLMA: mixed noise
26
24
σ
50
70
22
20
18
σ
0
50
100
150
t
200
250
300
Fig. 1: Frame-wise PSNR for Lobby dataset with Gaussian(σ = 25)
plus salt and pepper noise.
50
70
σ
50
70
σ
50
70
original
noisy
SLMA
Dataset: waterfall
ReLD
VBM3D
MLP
PCP-LD
33.08(73.14)
27.99(24.14)
18.87(477.60)
32.93(195.77)
29.25(69.77)
24.42(21.01)
15.03(478.73)
29.17(197.94)
NCRPCA-LD
GRASTA-LD
30.48(128.35)
25.33(58.23)
27.97(133.53)
21.89(55.45)
Dataset: fountain
ReLD
VBM3D
MLP
SLMA
30.53(15.82)
26.55(5.24)
18.53(109.79)
18.55(3.13 × 104 )
27.53(15.03)
22.08(4.69)
14.85(107.52)
16.25(3.19 × 104 )
Dataset: escalator
ReLD
VBM3D
MLP
SLMA
27.84(16.03)
25.10(5.27)
18.83 (109.40)
17.98(3.21 × 104 )
25.15(15.28)
20.20(4.72)
15.20(108.78)
15.90(3.18 × 104 )
Dataset: curtain
ReLD
VBM3D
MLP
SLMA
31.91(17.17)
30.29(4.42)
18.58(188.30)
19.12(7.86 × 104 )
28.10(16.50)
26.15(3.85)
14.73(192.00)
16.68(8.30 × 104 )
Dataset: lobby
ReLD
VBM3D
MLP
SLMA
35.15(58.41)
29.23(19.35)
18.66(403.59 ) 18.21(3.99 × 105 )
29.68(56.51)
24.90(17.00)
14.85(401.29)
16.82(4.09 × 105 )
TABLE I: PSNR (running time in seconds) for different denoising
algorithms. The running time for MLP does not include training time.
C. Denoising in Low-light Environment
ReLD
VBM3D
MLP
Fig. 2: Visual comparison of denoising performance for the Curtain
dataset for very large Gaussian noise (σ = 70)
In this part we show how ReLD can also be used for
video enhancement of low-light videos, i.e., to see the target
signal in the low-light environment. The video was taken in
a dark environment where a barely visible person walked
through the scene. The output we are using here for ReLD
is Ŝ, which is the output of ReProCS. In Fig. 3 we see
ReLD is able to enhance the visual quality – observing the
walking person. Observe that Histogram-Equalization, which
is a standard technique for enhancing low-light images, does
not work for this video.
[25]. We call the respective algorithms ReLD, PCP-LD,
NCRPCA-LD, and GRASTA-LD for short. We test all
these on the Waterfall dataset (downloaded from Youtube
https://www.youtube.com/watch?v=UwSzu 0h7Bg). Besides
these Laying-Denoising algorithms, we also compare with
VBM3D and MLP. Since the video length was too long,
img1
SLMA code failed for this video.
The waterfall video is slow changing and hence is well
modeled as being low-rank. We add i.i.d. Gaussian noise
with different variances to the video. The video consists of img2
650 frames of size 108 × 108. The results are summarized
original
ReLD
Hist-Eq
in Table V. This table contains results on a smaller sized
waterfall video. This is done because SLMA and PCP-LD Fig. 3: Ability of “seeing” in the dark for two sample frames. From
become too slow for the full size video. Comparisons of left to right: orignal dark image, results using ReLD, and using
the other algoithms on full size video are included in the Histogram-Equalization (Hist-Eq).
supplementary material. ReLD has the best performance when
the noise variance is large. We also show the time taken by
each method in parantheses. As can be seen, ReProCS is
IV. C ONCLUSION
slower than VBM3D and GRASTA-LD, but has significantly
In this paper we developed a denoising scheme, called
better performance than both.
ReLD which enhances the denoising performance of the stateIn Table V, we also provide comparisons on four more of-the-art algorithm VBM3D, and is able to achieve denoising
videos - fountain, escalator, curtain and lobby (same videos as in a broad sense. Using ReProCS to split the video first
in the salt and pepper noise experiment) with different levels of results in a clean low-rank layer since the large noise goes
Gaussian noise. In Fig.4a we show sample visual comparisons to the sparse layer. The clean layer improves the “grouping”
for the curtain video with noise standard deviation σ = 70. As accuracy in VBM3D. One drawback of our algorithm is that
can be seen, ReLD is able to recover more details of the images VBM3D needs to be executed twice, once on each layer. As
a result, the running time is at least doubled, but that also
while other algorithms either fail or cause severe blurring.
5
results in significantly improved PSNRs especially for very
large variance Gaussian noise. Moreover ReLD is still 6-10
times faster than SLMA and MLP, while also being better. It
is also much faster than PCP-LD and NCRPCA-LD.
6
R EFERENCES
[1] Han Guo and Namrata Vaswani, “Video denoising via online sparse
and low-rank matrix decomposition,” in Statistical Signal Processing
Workshop (SSP), 2016 IEEE. IEEE, 2016, pp. 1–5.
[2] Antoni Buades, Bartomeu Coll, and Jean-Michel Morel, “Image denoising by non-local averaging,” in Acoustics, Speech, and Signal Processing, 2005. Proceedings.(ICASSP’05). IEEE International Conference on.
IEEE, 2005, vol. 2, pp. 25–28.
[3] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen
Egiazarian, “Image denoising by sparse 3-d transform-domain collaborative filtering,” Image Processing, IEEE Transactions on, vol. 16, no.
8, pp. 2080–2095, 2007.
[4] Michael Elad and Michal Aharon, “Image denoising via sparse and
redundant representations over learned dictionaries,” Image Processing,
IEEE Transactions on, vol. 15, no. 12, pp. 3736–3745, 2006.
[5] Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian, “Pointwise shape-adaptive dct for high-quality denoising and deblocking of
grayscale and color images,” Image Processing, IEEE Transactions on,
vol. 16, no. 5, pp. 1395–1411, 2007.
[6] Julien Mairal, Francis Bach, Jean Ponce, Guillermo Sapiro, and Andrew
Zisserman, “Non-local sparse models for image restoration,” in
Computer Vision, 2009 IEEE 12th International Conference on. IEEE,
2009, pp. 2272–2279.
[7] Kostadin Dabov, Alessandro Foi, and Karen Egiazarian, “Video denoising by sparse 3d transform-domain collaborative filtering,” 2007.
[8] Hui Ji, Chaoqiang Liu, Zuowei Shen, and Yuhong Xu, “Robust video
denoising using low rank matrix completion,” in Computer Vision and
Pattern Recognition (CVPR), 2010 IEEE Conference on, pp. 1791–1798.
[9] Hui Ji, Sibin Huang, Zuowei Shen, and Yuhong Xu, “Robust video
restoration by joint sparse and low rank matrix approximation,” SIAM
Journal on Imaging Sciences, vol. 4, no. 4, pp. 1122–1142, 2011.
[10] Ce Liu and William T Freeman, “A high-quality video denoising
algorithm based on reliable motion estimation,” in Computer Vision–
ECCV 2010, pp. 706–719. Springer, 2010.
[11] Liwei Guo, Oscar C Au, Mengyao Ma, and Zhiqin Liang, “Temporal
video denoising based on multihypothesis motion compensation,” Circuits and Systems for Video Technology, IEEE Transactions on, vol. 17,
no. 10, pp. 1423–1429, 2007.
[12] SM Mahbubur Rahman, M Omair Ahmad, and MNS Swamy, “Video
denoising based on inter-frame statistical modeling of wavelet coefficients,” Circuits and Systems for Video Technology, IEEE Transactions
on, vol. 17, no. 2, pp. 187–198, 2007.
[13] H Rabbani and S Gazor, “Video denoising in three-dimensional complex
wavelet domain using a doubly stochastic modelling,” IET image
processing, vol. 6, no. 9, pp. 1262–1274, 2012.
[14] Shigong Yu, M Omair Ahmad, and MNS Swamy, “Video denoising using motion compensated 3-d wavelet transform with integrated recursive
temporal filtering,” Circuits and Systems for Video Technology, IEEE
Transactions on, vol. 20, no. 6, pp. 780–791, 2010.
[15] Michal Joachimiak, Dmytro Rusanovskyy, Miska M Hannuksela, and
Moncef Gabbouj, “Multiview 3d video denoising in sliding 3d dct domain,” in Signal Processing Conference (EUSIPCO), 2012 Proceedings
of the 20th European. IEEE, 2012, pp. 1109–1113.
[16] Saiprasad Ravishankar and Yoram Bresler, “Learning sparsifying transforms,” Signal Processing, IEEE Transactions on, vol. 61, no. 5, pp.
1072–1086, 2013.
[17] Saiprasad Ravishankar, Bihan Wen, and Yoram Bresler, “Online sparsifying transform learning - part 1: algorithms,” accepted to Signal
Processing, IEEE Transactions on.
[18] Saiprasad Ravishankar and Yoram Bresler, “Online sparsifying transform learning - part 2: convergence analysis,” accepted to Signal
Processing, IEEE Transactions on.
[19] Bihan Wen, Saiprasad Ravishankar, and Yoram Bresler, “Video denoising by online 3d sparsifying transform learning,” in Image Processing
(ICIP), 2015 IEEE International Conference on. IEEE, 2015, pp. 118–
122.
[20] Junyuan Xie, Linli Xu, and Enhong Chen, “Image denoising and inpainting with deep neural networks,” in Advances in Neural Information
Processing Systems, 2012, pp. 341–349.
[21] Forest Agostinelli, Michael R Anderson, and Honglak Lee, “Adaptive
multi-column deep neural networks with application to robust image
denoising,” in Advances in Neural Information Processing Systems,
2013, pp. 1493–1501.
[22] E. J. Candès, X. Li, Y. Ma, and J. Wright, “Robust principal component
analysis?,” Journal of ACM, vol. 58, no. 3, 2011.
[23] C. Qiu, N. Vaswani, B. Lois, and L. Hogben, “Recursive robust pca
or recursive sparse recovery in large but structured noise,” IEEE Trans.
Info. Th., 2014.
[24] Han Guo, Chenlu Qiu, and Namrata Vaswani, “An online algorithm
for separating sparse and low-dimensional signal sequences from their
sum,” Signal Processing, IEEE Transactions on, vol. 62, no. 16, pp.
4284–4297, 2014.
[25] Jun He, Laura Balzano, and Arthur Szlam, “Incremental gradient on
the grassmannian for online foreground and background separation in
subsampled video,” in IEEE Conf. on Comp. Vis. Pat. Rec. (CVPR),
2012.
[26] Praneeth Netrapalli, UN Niranjan, Sujay Sanghavi, Animashree Anandkumar, and Prateek Jain, “Non-convex robust pca,” in Advances in
Neural Information Processing Systems, 2014, pp. 1107–1115.
[27] F. De La Torre and M. J. Black, “A framework for robust subspace
learning,” International Journal of Computer Vision, vol. 54, pp. 117–
142, 2003.
[28] Clemens Hage and Martin Kleinsteuber,
“Robust pca and subspace tracking from incomplete observations using l0-surrogates,”
arXiv:1210.0805 [stat.ML], 2013.
[29] A.E. Abdel-Hakim and M. El-Saban, “Frpca: Fast robust principal
component analysis for online observations,” in Pattern Recognition
(ICPR), 2012 21st International Conference on, 2012, pp. 413–416.
[30] J. Feng, H. Xu, and S. Yan, “Online robust pca via stochastic
optimization,” in Adv. Neural Info. Proc. Sys. (NIPS), 2013.
[31] J. Feng, H. Xu, S. Mannor, and S. Yan, “Online pca for contaminated
data,” in Adv. Neural Info. Proc. Sys. (NIPS), 2013.
[32] Morteza Mardani, Gonzalo Mateos, and G Giannakis, “Dynamic
anomalography: Tracking network anomalies via sparsity and low rank,”
J. Sel. Topics in Sig. Proc., Feb 2013.
[33] Harold C Burger, Christian J Schuler, and Stefan Harmeling, “Image
denoising: Can plain neural networks compete with bm3d?,” in Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference
on. IEEE, 2012, pp. 2392–2399.
7
Supplementary Material
C OMPLETE A LGORITHM
Detailed algorithm of ReLD is summarized in Algorithm 3.
Algorithm 3 Details of each step of ReLD
Parameters: We used α = 20, Kmin = 3, Kmax = 10, t0 = 50 in all experiments.
1) Initialization using PCP: Compute (L̂0 , Ŝ0 ) ← PCP(M0 ) and compute [P̂0 , Σ̂0 ] ← approx-basis(L̂0 , 90%). The notation
PCP(M ) means implementing the PCP algorithm on matrix M and P = approx-basis(M , b%) means that P is the b%
left singular vectors’ matrix for M .
Set r̂ ← rank(P̂0 ), σ̂min ← (Σ̂0 )r̂,r̂ , t̂0 = t0 , flag=detect
2) For all t > t0 , implement an appropriately modified ReProCS algorithm
a) Split mt into layers ℓ̂t and ŝt :
′
i) Compute yt ← Φt mt with Φt ← I − P̂t−1 P̂t−1
ii) Compute ŝt as the solution of
min kxk1 s.t.kyt − Φt xk2 ≤ ξ
x
with ξ = kΦt ℓ̂t−1 k
p
iii) T̂t ← Thresh(ŝt , ω) with ω = 3 kmt k2 /n. Here T ← Thresh(x, ω) means that T = {i : |(x)i | ≥ ω}
ŝt,∗ ← LS(yt , Φt , T̂t ). Here x̂ ← LS(y, A, T ) means that x̂T = (A′T AT )−1 A′T y, which is least-squared
estimate of x on T .
iv) ℓ̂t ← mt − ŝt , ℓ̂t,∗ ← mt − ŝt,∗
b) Perform subspace update, i.e., update P̂t :
i) If flag = detect and mod(t − t̂j + 1, α) = 0,
′
)[ℓ̂t−α+1,∗ , . . . ℓ̂t,∗ ] and check if any singular values are above
A) compute the SVD of √1α (I − P̂(j−1) P̂(j−1)
σ̂min
B) if the above number is more than zero then set flag ← pPCA, increment j ← j + 1, set t̂j ← t − α + 1,
reset k ← 1
Else P̂t ← P̂t−1 .
ii) If flag = pPCA and mod(t − t̂j + 1, α) = 0,
′
)[ℓ̂t−α+1,∗ , . . . ℓ̂t,∗ ],
A) compute the SVD of √1α (I − P̂j−1 P̂(j−1)
B) let P̂j,new,k retain all its left singular vectors with singular values above σ̂min or all α/3 top left singular
vectors whichever is smaller,
C) update P̂t ← [P̂(j−1) P̂j,new,k ], increment k ← k + 1
D) If k ≥ Kmin and
k
Pt
′
′
t−α+1 (P̂j,new,i−1 P̂j,new,i−1 −P̂j,new,i P̂j,new,i )ℓ̂t,∗ k2
Pt
′
k t−α+1 P̂j,new,i−1 P̂j,new,i−1 ℓ̂t,∗ k2
< 0.01 for i = k − 2, k − 1, k; or k = Kmax ,
then K ← k, P̂(j) ← [P̂(j−1) P̂j,new,K ] and reset flag ← detect.
Else P̂t ← P̂t−1 .
3) Denoise using VBM3D:
a) σ̂fg ← Std-est([ŝt , . . . , ŝt0 ])
σ̂bg ← Std-est([ℓ̂t , . . . , ℓ̂t0 ]). Here Std-est(M ) denotes estimating the standard deviation of noise from M : we first
subtract column-wise mean from M and then compute the standard deviation by seeing it as a vector.
b) Ŝdenoised ← VBM3D([ŝ1 , . . . , ŝtmax ], σ̂fg )
L̂denoised ← VBM3D([ℓ̂1 , . . . , ℓ̂tmax ], σ̂bg ). Here VBM3D(M , σ) implements the VBM3D algorithm on matrix M
with input standard deviation σ.
Output: Ŝ, Ŝdenoised , L̂denoised or Iˆdenoised = Ŝdenoised + L̂denoised based on applications
C OMPLETE E XPERIMENTS
First, we compare the performance of our propsed denoising framework with video layering performed using eiher ReProCS
(our proposed algorithm), or using the other robust PCA agorithms - PCP, NCRPCA, and GRASTA. We call the respective
algorithms ReLD, PCP-LD, NCRPCA-LD, and GRASTA-LD for short. We test the algorithms on the Waterfall dataset
(downloaded from Youtube https://www.youtube.com/watch?v=UwSzu 0h7Bg). Besides these Laying-Denoing algorithms, we
also compare with VBM3D and a neural network image denoising method, Multi Layer Perceptron (MLP). The codes for
algorithms being compared are downloaded from the authors’ webpages. The available MLP code contains parameters that are
8
σ
25
30
50
70
35.00,
34.51,
33.08,
29.25,
ReLD
32.78 (73.54)
32.68 (73.33)
32.27 (73.14)
31.79 (69.77)
34.92,
34.42,
32.93,
29.17,
PCP-LD
32.84 (198.87)
32.60 (185.47)
31.65 (195.77)
30.67 (197.94)
NCRPCA-LD
33.34, 31.98 (101.78)
32.53, 31.56 (106.30)
30.48, 30.09 (128.35)
27.97, 29.63 (133.53)
GRASTA-LD
30.45, 28.11 (59.43)
29.40, 26.89 (58.76)
25.33, 23.97 (58.23)
21.89, 21.81 (55.45)
VBM3D
32.02 (24.83)
30.96 (23.96)
27.99 (24.14)
24.42 (21.01)
MLP
28.26 (477.22)
26.96 (474.26)
18.87 (477.60)
15.03 (478.73)
TABLE II: Comparison of denoising performance (PSNR) on the Waterfall dataset (length 650 and the images are of size
108 × 192): data shown in format of PSNR using Iˆdenoised , PSNR using L̂denoised , (and runing time in second).
σ
25
30
50
70
33.84,
33.01,
30.48,
27.39,
ReLD
29.98 (335.97)
29.79 (338.51)
28.86 (333.26)
27.77 (321.71)
33.38,
32.49,
29.79,
26.80,
PCP-LD
29.17 (413.45)
28.72 (406.45)
26.83 (399.09)
25.06 (386.60)
NCRPCA-LD
28.99, 27.54 (507.03)
27.63, 27.63 (510.39)
23.74, 23.29 (544.60)
20.95, 20.88 (589.22)
GRASTA-LD
29.37, 16.12 (435.54)
28.16, 12.53 (405.53)
22.17, 11.27 (417.50)
13.92, 9.78 (411.87)
VBM3D
33.67 (110.89)
32.75 (110.92)
30.18 (110.83)
26.75 (105.80)
31.11
29.18
19.00
15.08
MLP
(1.38 × 103 )
(1.38 × 103 )
(1.39 × 103 )
(1.40 × 103 )
TABLE III: Comparison of denoising performance (PSNR) on the Waterfall dataset (length 100 and the images are of size
540 × 960): data shown in format of PSNR using Iˆdenoised , PSNR using L̂denoised , (and runing time in second).
σ
25
30
50
70
35.13,
34.14,
31.10,
27.71,
ReLD
29.79 (2.83 × 103 )
29.61 (2.34 × 103 )
28.72 (2.31 × 103 )
27.66 (2.38 × 103 )
34.52,
33.45,
30.31,
27.08,
PCP-LD
29.03 (6.31 × 103 )
28.59 (5.50 × 103 )
26.75 (4.89 × 103 )
25.00 (4.82 × 103 )
30.69,
29.22,
25.18,
22.23,
NCRPCA-LD
28.50 (2.88 × 103 )
27.62 (2.38 × 103 )
24.58 (2.45 × 103 )
22.09 (2.48 × 103 )
GRASTA-LD
29.94, 10.72 (1.83 × 103 )
28.46, 10.42 (1.88 × 103 )
22.27, 8.69 (1.98 × 103 )
13.59, 8.33 (1.98 × 103 )
VBM3D
36.04 (533)
35.18 (550)
32.55 (536)
28.45 (608)
33.73
30.94
19.06
15.09
MLP
(5.56 × 103 )
(5.49 × 103 )
(5.50 × 103 )
(5.55 × 103 )
TABLE IV: Comparison of denoising performance (PSNR) on the Waterfall dataset (length 100 and the images are of size
1080 × 1920): data shown in format of PSNR using Iˆdenoised , PSNR using L̂denoised , (and runing time in second).
σ
25
30
50
70
σ
25
30
50
70
ReLD
32.67(16.70)
32.25(15.84)
30.53(15.82)
27.53(15.03)
ReLD
35.47(16.78)
34.58(17.35)
31.91(17.17)
28.10(16.50)
Dataset: fountain
VBM3D
MLP
31.18(5.44)
26.86(105.64)
30.26(5.17)
25.67(107.41)
26.55(5.24)
18.53(109.79)
22.08(4.69)
14.85(107.52)
Dataset: curtain
VBM3D
MLP
34.60(4.15)
31.14(189.14)
33.59(4.37)
28.90(191.14)
30.29(4.42)
18.58(188.30)
26.15(3.85)
14.73(192.00)
SLMA
22.93(3.05 × 104 )
21.85(3.06 × 104 )
18.55(3.13 × 104 )
16.25(3.19 × 104 )
ReLD
31.01(16.64)
30.27(16.45)
27.84(16.03)
25.15(15.28)
SLMA
23.28(7.75 × 104 )
22.74(9.05 × 104 )
19.12(7.86 × 104 )
16.68(8.30 × 104 )
ReLD
39.78(57.96)
38.76(57.99)
35.15(58.41)
29.68(56.51)
Dataset: escalator
VBM3D
MLP
30.32(5.34)
25.53(107.51)
29.29(5.38)
24.54(108.65)
25.10(5.27)
18.83 (109.40)
20.20(4.72)
15.20(108.78)
Dataset: lobby
VBM3D
MLP
35.00(19.57)
29.22(384.11)
33.64(19.09)
27.72(395.67)
29.23(19.35)
18.66(403.59 )
24.90(17.00)
14.85(401.29)
SLMA
21.17(3.09 × 104 )
20.49(3.15 × 104 )
17.98(3.21 × 104 )
15.90(3.18 × 104 )
SLMA
23.43(3.75 × 105 )
21.15(3.82 × 105 )
18.21(3.99 × 105 )
16.82(4.09 × 105 )
TABLE V: PSNR (and running time in second) for different denoising algorithms on datasets of fountain, escalator, curtain
and lobby.
original
noisy
SLMA
original
noisy
SLMA
ReLD
VBM3D
MLP
ReLD
VBM3D
MLP
(a) Curtain
(b) Lobby
Fig. 4: Visual comparison of denoising performance for Curtain and Lobby dataset for very large Gaussian noise (σ = 70)
36
32
38
34
30
36
32
28
34
30
26
32
28
30
RELD: Gaussian (σ=25)
VBM3D: Gaussian (σ=25)
MLP: Gaussian (σ=25)
SLMA: Gaussian (σ=25)
28
26
PSNR
40
PSNR
PSNR
9
RELD: mixed noise
VBM3D: mixed noise
MLP: mixed noise
SLMA: mixed noise
26
24
24
20
22
18
24
20
16
22
18
0
50
100
150
t
200
250
300
0
50
100
RELD: Gaussian (σ=70)
VBM3D: Gaussian (σ=70)
MLP: Gaussian (σ=70)
SLMA: Gaussian (σ=70)
22
150
t
200
250
14
300
0
50
100
150
t
200
250
300
(a)
(b)
(c)
Fig. 5: Frame-wise PSNR for Curtain dataset with different noise level: (a) Gaussian noise with σ = 25, (b) Gaussian noise
(σ = 25) plus salt and pepper noise, (c) Gaussian noise with σ = 70.
45
RELD: Gaussian (σ=25)
VBM3D: Gaussian (σ=25)
MLP: Gaussian (σ=25)
SLMA: Gaussian (σ=25)
40
36
32
34
30
32
28
30
26
28
PSNR
PSNR
PSNR
35
RELD: mixed noise
VBM3D: mixed noise
MLP: mixed noise
SLMA: mixed noise
26
30
24
0
50
100
150
t
200
250
300
18
RELD: Gaussian (σ=70)
VBM3D: Gaussian (σ=70)
MLP: Gaussian (σ=70)
SLMA: Gaussian (σ=70)
18
20
20
22
20
22
25
24
16
0
50
100
150
t
200
250
300
14
0
50
100
150
t
200
250
(a)
(b)
(c)
Fig. 6: Frame-wise PSNR for Lobby dataset with different noise level: (a) Gaussian noise with σ = 25, (b) Gaussian noise
(σ = 25) plus salt and pepper noise, (c) Gaussian noise with σ = 70.
trained solely from image patches that were corrupted with Gaussian noise with σ = 25 and hence the denoising performance
is best with σ = 25 and deteriorates for other noise levels. The video is a background scene without foreground, and hence
has no sparse component. We add i.i.d. Gaussian noise with different variance onto the video. Since there is no foreground in
the video, the splitting phase can generate a sparse layer which basically consists of the large-magnitude part of the Gaussian
noise. The denoising operation followed on such layer does not have the problem of degrading the video quality since this
layer is foreground-free.
The video consists of 650 frames and the images are of size 1080 × 1920. To speed up the algorithms, we first test on
the under-sampled data which has image size of 108 × 192. As can be seen from TABLE II, ReLD has the best denoising
performance. We also compare PSNRs using L̂denoised and Iˆdenoised , and we find that using L̂denoised shows an advantage when
the noise variance is very large. We then test the algorithms on larger image sizes. To avoid out-of-memory in computation,
we only use 100 frames of data. In TABLE III and TABLE IV we summarize the result for images of size 540 × 960 and
1080 × 1920 (original), respectively. We notice that, on very large data set (TABLE IV), using VBM3D without video-layering
algorithms achieves the best denoising performance. This may due to the fact that with larger image size, VBM3D has better
chance to find similar image blocks.
Next we thoroughly compare the denoising performance on two more dataset – curtain and lobby which are available at
http://www.ece.iastate.edu/∼hanguo/denoise.html. The algorithms being compared are ReLD, VBM3D, MLP and SLMA. The
noise being added to the original image frames are Gaussian (σ = 25), Gaussian (σ = 25) plus salt and pepper noisy, and
Gaussian (σ = 70). The input σ for all algorithms is estimated from the noisy data rather then given the true value. We compute
the frame-wise PSNRs (using Iˆdenoised ) for each case in Fig. 5 and Fig. 6 and show sample visual comparisons in Fig. 4a and
Fig. 4b. We can see in Fig. 5 and Fig. 6 that ReLD outperforms all other algorithms in all three noise level – the PSNR is the
highest in almost all image frames. Visually, ReLD is able to recover more details of the images while other algorithms either
fail or cause severe blurring effect. We test the algorithms on two more datasets (the fountain dataset and escalator dataset).
The noise being added to the original video is Gaussian, with standard deviation σ increases from 25 to 70. We present the
PSNRs in Table V.
300
| 1cs.CV
|
Energy Efficient and Throughput
Optimal CSMA Scheme
Ali Maatouk* , Mohamad Assaad* , and Anthony Ephremides†
arXiv:1712.03063v1 [cs.IT] 8 Dec 2017
*
TCL Chair on 5G, Laboratoire des Signaux et Systèmes, CentraleSupélec, Gif-sur-Yvette, France
† ECE Dept., University of Maryland, College Park, MD 20742
Abstract—Carrier Sense Multiple Access (CSMA) is widely
used as a Medium Access Control (MAC) in wireless networks
thanks to its simplicity and distributive nature. This motivated
researchers to find CSMA schemes that achieves throughput
optimality. In 2008, it has been shown that even simple CSMAtype algorithms is able to achieve optimality in terms of
throughput and has been given the name ”adaptive” CSMA.
Lately, new technologies emerged where prolonged battery life
is crucial such as environment and industrial monitoring. This
inspired the foundation of new CSMA based MAC protocols
where nodes are allowed to transition into sleep mode e.g.
IEEE 802.15.4 MAC. However, throughput optimality of these
protocols was not established. This paper therefore aims to
find a new CSMA scheme that combines both throughput
optimality and energy efficiency by adapting to the throughput
and power consumption needs of each node. This is done by
smartly controlling operational parameters such as back-off and
sleeping timers. The main advantages about this scheme is being
asynchronous, completely distributed and its ability to adapt to
different power consumption profiles required by each node.
We present numerical results that show the gain in terms of
energy efficiency compared to the conventional adaptive CSMA
scheme.
Index Terms—CSMA, Throughput Optimality, IoT, Wireless
MAC.
I. I NTRODUCTION
T
optimality for network scheduling, which
is by definition the ability to withstand any arrival
rate inside the capacity region of the network [1], has
been thoroughly investigated in the litterature. Tassiulas
and Ephremides introduced in their seminal paper [2] the
maximal-weight scheduling (MWS), a throughput optimal
scheduling scheme. This scheme is however centralized and
requires finding the maximum weighted independent sets, a
well known NP-Complete problem. In the aim of overcoming
the complexity, a surge of papers have been published. The
efforts were divided in two classes: the first one aimed to
propose low complexity algorithms for certain interference
models (e.g. [3] and the references therein) while the other focused on proposing low complexity algorithms that guarantee
a portion of capacity region [4] or a worst-case performance
[5]. These approaches suffer from message passing, which
can create a lot of congestion especially in the case of high
nodes density.
On the other hand, another class of simple and distributed
algorithms is the Carrier Sense Multiple Access (CSMA),
HROUGHPUT
which is one of the most popular random access protocols.
CSMA is referred to as the class of algorithms in which a
node senses the medium and transmit a packet only if the
medium is sensed idle. The reason for its widespread is due
to the fact that it is simple and completely distributed [5].
Classical CSMA such as the one adopted in IEEE 802.11
suffers however from throughput sub-optimality. This issue
has been solved in a series of work where it was shown that
throughput optimality can be achieved using fully distributed
algorithms (e.g. [1][6]). For instance, it was shown in [1] that
adaptive CSMA maximizes the throughput of the system for
a general interference model.
Both classical and adaptive CSMA suffer however from
high energy consumption. When in back-off stages, continuous sensing of the environment by each node is required to
spot any conflicting transmissions from its neighbors which
would result in an inevitable power consumption [1]. With
long battery life being a strict requirement for emerging
technologies, new MAC protocols were therefore proposed.
In particular, low energy consumption has been the main
focus in the design of MAC protocols in Wireless Sensor
Networks (WSN). These protocols rely on the idea of duty
cycling in which nodes regularly sleep. Sensor-MAC is one of
the most famous low-energy MAC protocols, which is based
on sleep schedules that are managed by virtual clusters [7].
Numerous other protocols were also introduced to further
enhance performance [8][9]. These protocols however suffer
from message passing in order to maintain synchronization.
Another procedure available in the literature is to make the
nodes wake up periodically with an offset between them
while compensating for the clock drift through several proposed methods [10]. Since the throughput optimality of these
proposed methods was not established, we seek to propose
a new distributed CSMA scheme that does not involve any
message passing and achieves throughput optimality while
reducing power consumption.
In the aim of reducing power consumption, our scheme
involves giving each node the freedom to transition between
AWAKE and SLEEP states. Conventional duty cycling proposed by the above methods will however not fit our case due
to a lack of adaptation of the duty cycle period. We argue
that by smartly controlling both the sleeping and back-off
duration of each node, power consumption can be reduced
while still maintaining the ability to withstand any feasible
arrival rate. This of course drastically changes the CSMA
model which requires new analysis as one will see in the
sequel. The result will be a fully distributed algorithm in
which time is not slotted (hence no synchronization required)
and does not suffer from any message passing or clock drift
issues. A new parameter will reveal itself which is assigned
to each node based on a satisfactory power-delay tradeoff.
The Paper is organized as follows: Section II describes the
model used and present the proposed CSMA scheme. The
Markov Chain and optimality analysis is provided in section
III. Section IV provides numerical results that corroborate
the theoretical assumptions while Section V concludes the
paper.
II. S YSTEM M ODEL AND P ROPOSED A LGORITHM
A. Interference Model
We consider a wireless network composed of multiple links
(i.e. multiple transmitters and receivers pairs) using the same
bandwidth. The interference between the links is modeled
by a conflict graph model, which is the most common model
used in the literature and more precisely in the area of random
access and CSMA based modeling. Assuming that we have
K links in the network, we associate a corresponding conflict
graph G={V, E}, where V is the set of vertices of the graph
which denotes the links of the network and E being the set of
edges. Two vertices have an edge between them iff these two
links cannot transmit at the same time. Clearly, not all links
necessarily have the ability to transmit at the same time hence
we define the notion of independent set of this conflict graph.
We denote the independent set of this graph by xi ∈ {0, 1}K
where xik = 1 if link k is active in this set.
In the aim of reducing power consumption, we give the
ability to each node to choose either to be awake or asleep.
In this context, we have 2K possible awake configurations
for the links. For each configuration j, there exist a unique
conflict graph Gj ={Vj , Ej }. We define for each of these
graphs, the set of independent sets Ij ={xi : the network is
in awake configuration j}.
B. Proposed Scheme
The words nodes, links and users are used interchangeably
throughout the paper. In the following, we adopt the standard
CSMA assumptions [1]:
• Sensing is considered instantaneous, there is no sensing
delay
• The problem of Hidden Nodes does not exist
These two assumptions make the model easier to work
with and mathematically tractable. In addition to that, we
assume that the links are always back-logged. Taking these
assumptions into consideration, we can now state the way our
mechanism works. The node can be in one of the following
states:
• SLEEP state: In this case, power consumption is minor
and no sensing of the environment takes place
•
AWAKE state: The adaptive CSMA sensing and back-off
mechanisms take place
The decision to either Wake-up/Sleep is dictated by an
appropriate timer. When a node decides to sleep, it picks
an exponentially distributed wake-up time with mean 1/Wk
after which it wakes up. Once the node is awake, it picks
an exponentially chosen sleep timer with mean 1/Sk after
which it goes back to sleep. The motivation behind choosing
exponentially distributed timers is the memory-less property
that enables us to pursue a MC based analysis.
When the node is awake, the adaptive CSMA scheme takes
place:
•
•
•
•
•
An exponentially distributed back-off timer with mean
1/Rk is picked
Continuous sensing of the channel takes place and
whenever the channel is sensed idle, the back-off timer
runs otherwise it is frozen. In both cases, the sleep timer
keeps running
Once the back-off timer runs out, the channel is captured
by the node and the sleep timer is hence frozen
Packets transmission time is assumed to be exponentially distributed with an average channel holding time
1/Hk
After successful transmission, the sleep-timer is resumed
and another back-off timer is picked for the next transmission
III. MC BASED O PTIMALITY A NALYSIS
A. Proposed Markov Chain
Due to the introduction of a new degree of freedom
for each node, the MC proposed in [1] would not fit our
analysis. The description of our scheme can be modeled by
a 2D CTMC where each state in the chain is made of two
components (aj , xi ):
Configuration State aj : A K-tuplet of binary variables ak
that indicates if the node k is awake (binary value 1) or
asleep (binary value 0).
Transmission State xi : A K-tuplet of variables xk that
indicates if the node k is transmitting or not. We assume that
a node transmit just one packet when it acquires the channel
hence xik is a binary variable. To note that if xik = 1, this
necessarily means that the associated configuration state with
this transmission state must have ajk = 1 since a node cannot
transmit unless it is awake.
Proposition 1. This CTMC is irreducible and time reversible
that is fully characterized by the two parameters rk =
log(Rk /Hk ) and ρk = log(Wk /Sk ) called transmission
aggressiveness and waking-up aggressiveness respectively.
This chain admits the following distribution as stationary:
exp(
j
i
π(a , x ; r, ρ) =
K
P
k=1
ajk ρk )exp(
K
P
k=1
C(r, ρ)
xik rk )
(1)
Fig. 1: Time-line of the proposed CSMA scheme
Where C(r, ρ) is a normalization constant and is equal to:
K
C(r, ρ) =
2
X
exp(
j=1
K
X
ajk ρk )
|Ij |
X
i=1
k=1
exp(
K
X
xik rk )
k=1
Proof: It is sufficient to show that the preceding distribution verifies the detailed balance equations.
Step 1: Consider the following two states (aj , xi ) and
(aj + ek , xi ) where ajk = 0 and ek represents the canonical
vector in RK . It can be verified that:
π(aj + ek , xi ; r, ρ)
= exp(ρk )
π(aj , xi ; r, ρ)
Step 2: Consider the two states (aj , xi ) and (aj , xi + ek )
where ajk = 1 (the node has to be awake to be able to
transmit), xik = 0 and xik0 = 0 ∀k 0 ∈ N eighbor(k). It can
also be verified that:
π(aj , xi + ek ; r, ρ)
= exp(rk )
π(aj , xi ; r, ρ)
To further clarify the model, a simple example of two
conflicting links is taken in Fig. 2. The dashed lines partition
the chain in 4 regions where in each, the states share the same
configuration state. One can see how our chain is much more
general by observing that the partition where aj = (1, . . . , 1)
is nothing but the MC in [4].
The throughput achieved by user k using this scheme is
simply the amount of time the chain is in a state where node
k is both awake and transmitting. This can be formulated in
the following way:
K
sk (r, ρ) =
2
X
j=1
ajk
|Ij |
X
xik π(aj , xi ; r, ρ)
(2)
i=1
Remark 1. One can see that our framework provides more
degrees of freedom as compared to the adaptive CSMA
since new parameters, for instance 1/Wk and 1/Sk , are
introduced. In the adaptive CSMA, the overall optimal transmission aggressiveness (and therefore the optimal back-off
timer mean) is just dictated by the network. This might
lead to relatively large back-off timers which would result
in an inevitable power loss due to continuous sensing of
the environment. The motivation behind the utility of these
new degree of freedom is the fact that nodes that sleep tend
to be more aggressive on the channel once they are awake
to compensate for the time spent in Sleep mode. By doing
so, back-off timers are drastically shrinked which reduces
power consumption. One can see in Fig. 1 an example of the
case of two nodes where Node 1 is highly constrained by its
battery life unlike Node 2. Node 1 therefore eventually end
up in sleeping mode for large duration of time while being
extremely aggressive on the network (short back-off timer)
once it transitions to Awake state. On the other hand, Node
2 barely sleeps and is moderately aggressive on the network.
Our aim is therefore to develop a well designed sleep-wake
up mechanism that would still maintain throughput optimality
while dealing with this power consumption issue and this is
what we are going to tackle in the following sections.
B. Optimality Analysis
We consider that that the arrival rate is normalized to the
capacity hence λk ≤ 1. In this case, λk can be seen as the
portion of time the chain should be in a state where user k
is both awake and transmitting. Hence a rate is said to be
feasible if it can be written in function of a joint probability
distribution over our MC space P where:
K
λk =
|Ij |
2 X
X
pij ajk xik
(3)
j=1 i=1
and P verifies:
K
|Ij |
2 X
X
pij = 1
j=1 i=1
The rate region is then defined by:
K
C = {λ ∈ R
where P verifies:
K
: λk =
|Ij |
2 X
X
j=1 i=1
pij > 0
j|
2K |I
P
P
pij = 1
pij ajk xik }
j=1 i=1
For any arrival rate λ ∈ C, our goal would therefore be
to calibrate the parameters (r, ρ) in a way to make the
MC stationary distribution as close as possible to P . This
is equivalent to reducing the distance between these two
Fig. 2: Two conflicting nodes MC example
distributions. Different measures between distributions exist
but in our paper we consider the Kullback-Leibler divergence.
This approach was firstly introduced by the authors in [1] to
prove the optimality of the adaptive CSMA. For our case, due
to the introduction of new parameters, the subsequent analysis is different and the resulting problem is not necessarily
convex which requires further investigation. Our optimization
problem is the following:
|Ij |
2K X
X
pij
minimize D(pkπ(r, ρ)) =
pij log(
)
j
r,ρ
π(a , xi ; r, ρ)
j=1 i=1
The objective function can be reduced through several steps:
K
2K |Ij |
|Ij |
2 X
X
XX
pij
)=
pij log(pij )
pij log(
j
i
π(a , x ; r, ρ)
j=1 i=1
j=1 i=1
K
−
|Ij |
2 X
X
pij log(π(aj , xi ; r, ρ)) =
|Ij |
2 X
X
K
pij (
j=1 i=1
K
X
xik rk +
k=1
K
X
k=1
pij log(pij )
ajk ρk − log(C(r, ρ)))
The first term being independent of (r, ρ), the problem can
be reformulated as minimizing the following function:
F (r, ρ) = −
|Ij |
K X
2K X
X
k=1 j=1 i=1
2K
X
+ log(
j=1
pij xik rk −
exp(
K
X
k=1
ajk ρk )
|Ij |
K X
2K X
X
pij ajk ρk
k=1 j=1 i=1
|Ij |
X
i=1
pij xik rk
=
|Ij |
2 X
K X
X
pij ajk xik rk =
k=1 j=1 i=1
k=1 j=1 i=1
exp(
K
X
k=1
xik rk ))
K
X
λk rk
k=1
For the second term we will define a new variable
K
|Ij |
K X
2 X
X
K
pij ajk ρk
k=1 j=1 i=1
=
K X
2
X
ajk ρk
k=1 j=1
where:
|Ij |
X
pij =
i=1
fk =
ajk αj
= E(ak ) and
K
X
fk ρk
k=1
K
2
X
αj =
j=1
j=1 i=1
|Ij |
2 X
X
K
K
|Ij |
K X
2 X
X
K
j=1 i=1
−
The first term can be further reduced to an easier form since
xik cannot be equal to 1 unless the node k is actually awake.
So multiplying by the binary variable ajk does not change the
value of this term:
|Ij |
X
pij
i=1
αj represents the portion of time required for the network to
be at configuration state aj . fk can be seen as the portion
of time that the node k is awake as dictated by the arrival
rate λk . At a first glance, our problem as formulated is not
necessarily convex in (r, ρ). However, this difficulty can be
alleviated by introducing a new variable tji in the following
manner:
minimize
r,ρ,T
−
K
X
k=1
subject to tji =
λk rk −
K
X
k=1
K
X
K
fk ρk + log(
j=1 i=1
k=1
ajk ρk +
K
X
|Ij |
2 X
X
xik rk
k=1
j = 1, . . . , 2K i = 1, . . . , |Ij |.
exp(tji ))
The objective function is now made of convex/linear functions in (r, ρ, t) since the log-sum-exponential function is a
well known convex function [11]. On top of that, our equality
constraints are simply linear hence our transformed OP is
indeed convex. We can now formulate the Lagrangian and
state the necessary and sufficient conditions for optimality.
K
L(r, ρ, T , µ) =
−
K
X
k=1
|Ij |
2 X
X
µji (−tji +
j=1 i=1
λk rk −
K
X
ajk ρk +
k=1
K
X
K
X
xik rk )
k=1
|Ij |
2K X
X
fk ρk + log(
exp(tji ))
j=1 i=1
k=1
Where µji represents the dual variable corresponding to the
{ji} constraint. At the optimal point, the following conditions
are met:
exp(t∗ji )
∂L(r ∗ , ρ∗ , T ∗ , µ∗ )
=0
(4)
= −µ∗ji +
∂tji
C(r ∗ , ρ∗ )
K
K
k=1
k=1
X j
X
∂L(r ∗ , ρ∗ , T ∗ , µ∗ )
= −t∗ji +
ak ρ∗k +
xik rk∗ = 0 (5)
∂µji
Combinining the above two conditions, we can conclude that:
exp(
µ∗ji
K
P
k=1
=
ajk ρ∗k )exp(
K
P
k=1
C(r ∗ , ρ∗ )
xik rk∗ )
= π(aj , xi ; r ∗ , ρ∗ )
which is nothing but the stationary distribution of our chain.
Furthermore:
K
|Ij |
2 X
X
∂L(r ∗ , ρ∗ , T ∗ , µ∗ )
= −fk +
µ∗ji ajk = 0
∂ρk
j=1 i=1
(6)
Using our previous conclusion on µ∗ji , (6) will lead to:
K
−fk +
2
X
ajk
j=1
|Ij |
X
π(aj , xi ; r ∗ , ρ∗ ) = 0
i=1
K
−fk +
2
X
ajk αˆj = 0 and αˆj =
j=1
|Ij |
X
π(aj , xi ; r ∗ , ρ∗ )
i=1
αˆj is interpreted as the amount of time the chain is in
configuration state aj . Which means that at the optimal point,
. P2K
fˆk = j=1 ajk αˆj = fk or in other words, the node k is awake
just as needed to be. Lastly:
K
|Ij |
2 X
X
∂L(r ∗ , ρ∗ , T ∗ , µ∗ )
= −λk +
µ∗ji xik = 0
∂rk
j=1 i=1
(7)
As it has been used before, multiplying by the binary variable
ajk does not change this value due to the fact a node cannot
transmit if not awake. Also using the previous conclusions
concerning µ∗ji , (7) becomes:
K
−λk +
2
X
j=1
ajk
|Ij |
X
i=1
xik π(aj , xi ; r ∗ , ρ∗ ) = 0
Hence at the optimal point we have sk (r ∗ , ρ∗ ) = λk which
ensures rate stability of the queues. Consequently, our OP
problem can be solved distributively using a simple gradient
descent method by each user monitoring its own service rate
and awake time.
C. Discussions and Parameter Birth
Now the question to answer is: what does the distribution
P truly dictates concerning the configuration state?
For any feasible λ, it is clear from the expression of λk that
the arrival rate only dictates the portion of time when user k
has to be both awake and transmitting. The other states where
the user is either awake or in a back-off stage are irrelevant
to the arrival rate so the portion of time in which the chain
is in those states presents for us a degree of freedom we can
take advantage of. To see that more clearly, we know that
the minimal value of fk is λk . The only way to achieve this
is if the node wakes up and instantly acquire the channel.
Consequently for this case, the value of rk tends to +∞. In
other words, the user is really aggressive to get the channel
as soon as he wakes up.
The maximal value for fk is 1, in other words the node
is always awake. Consequently in this case, ρk −→ +∞.
The special case when fk −→ 1 ∀k, we will have the
configuration state aj = (1, . . . , 1) almost all the time and
we are back to the adaptive CSMA scheme in [1]. In order
to mitigate these extreme cases on ρk and rk , we impose that
λk < fk < 1 ∀k. In overall we can write fk in the following
manner:
fk = λk + ωk
where 0 < ωk < 1 − λk . This ωk is the degree of freedom
(DoF) previously mentioned. When ωk is really small, we
are forcing the user to wake up just as necessary, so we
expect him to be extremely aggressive on the channel (small
back-off timer) and therefore power consumption is reduced.
However due to the frequent long sleep, delays are to be
anticipated which would manifest in long queues lengths.
On the other extreme, if ωk is close to its maximal value,
the user is almost continuously active so less delays would
be expected but a higher power consumption. In summary,
ωk is a power-delay trade off parameter that is assigned to
each node depending on the desired trade-off. The beauty
of this parameter comes from the fact that IoT applications
present a mixture of both delay-sensitive and delay-tolerant
applications. This can be exploited by simply calibrating the
parameter assignment accordingly, the following table gives
a few examples:
IoT Applications
Category
Delay
Emergencies
Highly Sensitive
E-Health
Med. Sensitivity
Environment Monitors Highly Tolerant
TABLE I: Parameter Assignment
Parameter
High
Medium
Low
This scheme would only be plausible if the optimum is
always attained for a finite (r ∗ , ρ∗ ) and that is what the
following theorem addresses.
Theorem 2. If λ is strictly feasible and if λk < fk < 1 ∀k,
then the optima is attained for a finite (r ∗ , ρ∗ )
To note: A rate λ is said to be strictly feasible if λ ∈
int(C) or in other words if there exist a neighborhood of
λ such as every λI that belongs to this neighborhood is
feasible. The case of two conflicting links can be found in
Figures 3 and 4. Both λ and f that are pointed to verify the
required conditions by this theorem.
Proof: The proof can be found in the Appendix.
frame) has elapsed after which the transmission aggressiveness and wake-up aggressiveness attain their optimal value,
a new constant ωk = gk (Qk ) can be assigned to calibrate
the power-delay trade off based on the link’s queue. It can
be seen as a penalty function for the high queue length,
that should be compensated in the next convergence frame.
However this chosen function verifies certain properties, the
first being that gk is an increasing function in Qk . Other
properties include:
0 < gk (Qk ) < 1 − λk
∀Qk
Update: Since the gradient of the desired objective function
can be calculated in a distributed manner by each node k, the
update will be done using a simple gradient descent method.
However due to the fact that the mixing time is slow [1], the
update should be done using sufficiently large time frames Tm
called update frame in order to ensure that the chain reaches
its stationary distribution. Suppose that the links update their
aggressiveness at time tm , we define Tm = tm − tm−1 as
being the m-th update frame.
rk (m + 1) = rk (m) + 41k (m)(λk − sk (m))
ρk (m + 1) = ρk (m) + 42k (m)(λk + ωk − fˆk (m))
Fig. 3: Feasible capacity region for two conflicting links case
Fig. 4: The allowed region for the awake vector in the case
of two conflicting links
D. Implementation Considerations
The DoF parameter assignment can be either static (kept
constant) or it can be dynamic by choosing an appropriate
function gk (Qk ) where Qk is the queue length of link k.
First case: ωk is assigned to each node and kept constant
throughout transmission without change.
Second case: Once a large time frame Tc (called convergence
where 4 refers to a well chosen step size. Both sk (m) and
fˆk (m) are calculated as follows:
R tm+1
R tm+1
ak (τ )dτ
ak (τ )xk (τ )dτ
tm
fˆk (m) = tm
sk (m) =
Tm+1
Tm+1
If λk is unknown, it can be itself approximated using its
empirical average. A simple note to be taken in consideration,
dummy packets are considered to be sent by node k if its
buffer was empty to ensure the ideal throughput and awake
duration are used during the updates.
As we can see from the expression of the wake-up aggressiveness and Transmission aggressiveness, the chain depends
Rk
k
on the ratio between W
Sk and Hk and not individually on
them. For instance, if we fix Sk to a value and update the
wake-up aggressiveness, it is equivalent to changing just Wk
and vice-versa. This means that the node’s sleep timer is
kept with the same mean but how often the node wakes up
depends on the corresponding wake-up aggressiveness. We
will adopt this approach by fixing both Sk and Hk .
Due to the fact that the updates create a fresh new chain,
residual effects from the previous chain should be mitigated:
•
•
•
If the node was awake and transmitting when the update
time frame has elapsed, the update of the parameters will
not affect it so no further actions
If the node was awake but in a back-off stage, a
new back-off timer according to the new transmission
aggressiveness should be taken.
If the node was asleep, a new wake-up timer according
to the new wake-up aggressiveness should be generated
Power Model
Power Parameter
Value
Pz
5 µW
Pt
27 mW
Ps
1.8 mW
TABLE II: Power model values
For the sake of simplicity, we assume that the channel mean
holding time is 1 ms, and as mentioned before we are going
to assume that the sleeping timer’s mean is kept constant
throughout the simulations and is equal to 1 ms. In this
way, the calibrations of parameters rk an ρk is equivalent
to calibrating the back-off and the wake-up timers mean
respectively.
Scenario 1: The following simulation is a simple proof
of concept and approval of theoretical results in terms of
parameters convergence. For this purpose, we take a simple
example of two conflicting links with medium traffic of
arrival rate λ = (0.35, 0.35). The simulations are run for
100 s with a fixed update time frame Tm = 10 ms and with
the same step size 4 = 0.1 in all updates for both of the
algorithms. For the delay-power tradeoff parameter, it is taken
as ωk = 0.3 for both links which would require both links
to be awake 65% of the time.
For the adaptive CSMA, the minimization of the convex
objective function leads to a theoretical Transmission Aggressiveness vector r ∗ = (0.154, 0.154). For our algorithm, the
OP leads to an optimal Transmission Aggressiveness vector
r ∗ = (0.9273, 0.9273) and Wake-up Aggressiveness vector
ρ∗ = (−0.1541, −0.1541). When adopting our scheme, as it
can be seen from these values, the nodes will sleep but are
more aggressive on the channel when they wake up than the
adaptive CSMA to compensate for the time spent sleeping.
Evolution of the aggressivenesses
1.4
r1
r2
rho1
rho2
CSMA r1
CSMA r2
1.2
1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
0
1
2
3
4
5
Time
6
7
8
9
10
4
x 10
Fig. 5: Evolution of aggressiveness for ωk = 0.3 ∀k
Scenario 2: In this scenario, we take a realistic heterogeneous case in which several groups of nodes each have their
own desired delay-power tradeoff. 3 groups of 4 nodes each
are considered:
-Group 1: This group is made of nodes that are extremely
delay sensitive but can tolerate a higher energy consumption
-Group 2: This group is made of nodes that fall between
the two extremes, they need a moderate power consumption
without introducing a lot of delay
-Group 3: This group is made of nodes that can tolerate long
delays however they are power limited
We consider that the arrival for each node is λk = 0.077 ∀k.
ω1 = 0.8, ω2 = 0.4 and ω3 = 0.1 are assigned for Group
1,2 and 3 respectively.
Average queue
150
100
Average queue
The goal of these simulations is to show the gain experienced from using this method compared to the adaptive
CSMA. Two scenarios are taken in consideration but before
we present them we should state the power model taken. The
power consumption is as follows:
• When the node is in a SLEEP state, the node consumes
Pz
• When the node is transmitting, it consumes Pt
• When the node is sensing the medium: it is essentially
receiving radio signals and after signal processing, it
takes a decision whether the medium is to be considered
busy or not. We neglect the processing power consumption and assume that the power consumption in this case
is simply the power required to receive the radio signal
We will show that even with this assumption that gives
adaptive CSMA an edge since our goal is to minimize power
induced by the medium sensing, we will still achieve a decent
power gain.
As for the numerical values, we will adopt the values in
WiseNET low power radio transceiver used in [12].
The algorithm is completely distributed and by only observing the amount of time the node spends transmitting and
being awake, the parameters converge to their optimal point
as seen in Fig.5.
Aggressiveness
IV. S IMULATIONS
50
0
Group 1
Group 2
Group 3
Fig. 6: Average queue lengths for each group
Power Gain
70
60
Power Gain in %
50
40
30
20
10
0
Group 1
Group 2
Group 3
Fig. 7: Power gain of each group
In order to really demonstrate the difference, we are going to
let the initial parameters to be close to their optimal values
for both our scheme and adaptive CSMA. The results are
shown in both Fig.6 and Fig.7.
As expected, Group 3 experienced the highest queue length
but benefited from a high power gain (recall that Group 3
is delay tolerant). Group 1 on the other extreme had the
complete opposite results while Group 2 laid in between
these two cases. Fig.7 shows the gain in terms of power of
our scheme as compared to the adaptive CSMA. Our results
show clearly that our algorithm can handle multiple services
at the same time by simply adjusting the parameters. For
some IoT applications (with delay tolerance), our scheme
provides a huge power gain.
V. C ONCLUSION
In this paper, we introduced a new scheme that belongs
to the CSMA family. In this scheme, we give each node the
choice to be awake or asleep. We have shown that we can
still maintain throughput optimality even if the node have the
ability to transition into a SLEEP state. Each node controls
two parameters for which the convergence to their optimal
values has been proven to be completely distributed without
any message passing. The theoretical analysis resulted in the
birth of a parameter which had the interpretation as being
a delay-power tradeoff. This parameter is assigned to each
node depending on the application concerned. The simulations conformed with the theoretical results and showed
the advantages in terms of power gain with respect to the
adaptive CSMA.
R EFERENCES
[1] L. Jiang and J. Walrand. A distributed csma algorithm for throughput
and utility maximization in wireless networks. IEEE/ACM Transactions on Networking, 18(3):960–972, June 2010.
[2] L. Tassiulas and A. Ephremides. Stability properties of constrained
queueing systems and scheduling policies for maximum throughput in
multihop radio networks. IEEE Transactions on Automatic Control,
37(12):1936–1948, Dec 1992.
[3] A. Eryilmaz, A. Ozdaglar, and E. Modiano. Polynomial complexity
algorithms for full utilization of multi-hop wireless networks. In IEEE
INFOCOM 2007 - 26th IEEE International Conference on Computer
Communications, pages 499–507, May 2007.
[4] X. Wu, R. Srikant, and J. R. Perkins. Scheduling efficiency of
distributed greedy scheduling algorithms in wireless networks. IEEE
Transactions on Mobile Computing, 6(6):595–605, June 2007.
[5] S. Y. Yun, Y. Yi, J. Shin, and D. Y. Eun. Optimal csma: A survey.
In 2012 IEEE International Conference on Communication Systems
(ICCS), pages 199–204, Nov 2012.
[6] Devavrat Shah and Jinwoo Shin. Randomized scheduling algorithm
for queueing networks. Ann. Appl. Probab., 22(1):128–171, 02 2012.
[7] Wei Ye, J. Heidemann, and D. Estrin. An energy-efficient mac protocol
for wireless sensor networks. In Proceedings.Twenty-First Annual Joint
Conference of the IEEE Computer and Communications Societies,
volume 3, pages 1567–1576 vol.3, 2002.
[8] Tijs van Dam and Koen Langendoen. An adaptive energy-efficient
mac protocol for wireless sensor networks. In Proceedings of the
1st International Conference on Embedded Networked Sensor Systems,
SenSys ’03, pages 171–180, New York, NY, USA, 2003. ACM.
[9] G. Lu, B. Krishnamachari, and C. S. Raghavendra. An adaptive energyefficient and low-latency mac for data gathering in wireless sensor
networks. In 18th International Parallel and Distributed Processing
Symposium, 2004. Proceedings., pages 224–, April 2004.
[10] Bharath Sundararaman, Ugo Buy, and Ajay D. Kshemkalyani. Clock
synchronization for wireless sensor networks: a survey. Ad Hoc
Networks, 3(3):281 – 323, 2005.
[11] Stephen Boyd and Lieven Vandenberghe. Convex Optimization. Cambridge University Press, New York, NY, USA, 2004.
[12] Amre El-Hoiydi and Jean-Dominique Decotignie. Low power downlink mac protocols for infrastructure wireless sensor networks. Mobile
Networks and Applications, 10(5):675–690, 2005.
A PPENDIX
P ROOF OF T HEOREM 2
Since we do not impose any positivity constraints on our
parameters, the use of Slatter’s conditions theorem as in [1]
to prove finiteness of our parameters cannot be applied here
and we have to recourse to other tools.
Lemma 3. pij > 0 ∀i, j if and only if
λk > 0 ∀k
λ is strictly feasible
λk < fk < 1 ∀k
Proof: We divide the proof of the Lemma in two parts
as will be seen in the sequel.
Part 1: We start by proving that if pij > 0 ∀i, j then:
λk > 0 ∀k
λ is strictly feasible
λk < fk < 1 ∀k
- f < 1 : the state S = (0, 0) has a non-null probability
then no node stay awake all the time and therefore we have
that fk < 1 ∀k
- f > λ : states where nodes are awake and not transmitting
have non zero probability i.e. nodes enter in back-off stages
and therefore spend more time awake than transmitting
- λ > 0 : it is straightforward since each node has an arrival
rate of at least pk = (aj = ek , xi = ek ) > 0
- Strict feasibility : we will have to prove that there is exist
a neighborhood of λ such that each arrival rate inside of
it is feasible. Let p0 = P ((1, 0)) > 0 the portion of time
users are all awake and none of them is transmitting and
pk = P ((1, ek )) > 0 where ek refers to the canonical vector,
we define = min{p0 /K, min pk } > 0. We can satisfy any
arrival rate λ0 such as:
k
|λ0k − λk | 6 ∀k
We can do that by simply constructing the following distribution p0 such as:
K
P
0
(λ0k − λk )
p0 = p0 −
k=1
p0k = pk + λ0k − λk ∀k
p0 = p for all other states
To note that fk0 = fk ∀k since we only change the
transmission state and not the configuration state.
Part 2: Throughout this section, a visualization of the proof
in the case of two conflicting nodes is presented to clarify
the explanation. In this part, we suppose that λ and f verify:
λk > 0 ∀k
λ is strictly feasible
λk < fk < 1 ∀k
Fig. 8: Case (b) capacity region
We will try to construct p > 0 such as:
j|
2K |I
P
P
pij ajk xik
λk =
j=1 i=1
j|
2K |I
P
P
pij ajk
fk =
j=1 i=1
We start by choosing an arbitrary probability distribution
pA > 0, then four cases can occur:
(a): the randomly constructed pA verifies the requirements
λA = λ and fA = f then we choose p = pA
(b): pA verifies λA = λ but fA 6= f . Since both fA , f > λ
(for fA we recall part 1 since pA > 0) hence they lay in
the same region and we can find d > 0 as shown in Fig 9
such as:
fB = f + d(f − fA )
Therefore we can find pB > 0 such as λB = λ. We can
then construct in this case p in the following way:
pij = θpAij + (1 − θ)pBij
> 0. We know that pAij > 0 and pBij > 0,
0.
(c): pA verifies the requirement for fA = f but λA 6= λ. It
is tricky here since the region of f depends on λ hence fA
and f do not lay in the same region. This is shown in Fig 11
where the green region represents the allowed region of fA
while the wavy one is for f . However since λ < f , we can
always find a neighborhood of λ such as for each element
in it λI , it verifies λI < f . This neighborhood is marked
as a red disk in Figure 12. Now we take the intersections
d
where θ = 1+d
therefore pij >
Fig. 9: Case (b) awake vector allowed region
between this neighborhood and all the affine combinations
between λ and λA , then we can state that there is exist λB
in this intersection ( hence verifies λB < f ) such as ∃d > 0
in a way that:
λB = λ + d(λ − λA )
Let pB > 0 its corresponding probability distribution with
fB = f which is possible since λB < f . Then we can
conclude that we can write p in the following way:
pij = θpAij + (1 − θ)pBij
d
where θ = 1+d
> 0. We know that pAij > 0 and pBij > 0,
therefore pij > 0.
(d): in this case we have fA 6= f and λA 6= λ. This might
seem like the hardest case but it is merely a combination of
both cases (c) and (d). Since the priority is to coincide the
regions of fA and f , we start by using the same analysis as
case (c). By doing so, we can have:
λB = λ + d1 (λ − λA )
d1
> 0. We know that pAij > 0 and pBij > 0,
where θ1 = 1+d
1
therefore pCij > 0. We have now λC = λ but fC 6= f . We
are back to case (2) and therefore we can find d2 > 0 such
as:
fD = f + d2 (f − fC )
and λD = λ. We can then construct in this case p in the
following way:
pij = θ2 pCij + (1 − θ2 )pDij
d2
where θ2 = 1+d
> 0. We know that pCij > 0 and pDij > 0,
2
therefore pij > 0 which concludes the proof of the Lemma
Fig. 10: Case (c) capacity region
The next step is to use the results of Lemma 3 to prove
that our optimum is attained for a finite (r ∗ , ρ∗ ).
Finite Minimzer: First, we know that D(pkπ(r, ρ)) > 0
hence
Inf D(pkπ(r, ρ)) exists
r,ρ
Now it remains to show that it’s attained by a finite (r ∗ , ρ∗ ).
As proven before, we consider in our case that pij > 0 and
we recall our optimization problem:
K
minimize F (r, ρ) = −
r,ρ
Fig. 11: Case (c) awake vector allowed region
|Ij |
2 X
X
pij log(π(aj , xi ; r, ρ))
j=1 i=1
We have proven that this minimization is indeed a convex
OP after appropriate transformations. Therefore, we either
have one of the following cases:
(a): a finite minimizer (r ∗ , ρ∗ ) exists and it is unique due
to the convexity of our problem
(b): there exist a sequence (r n )n such that
n→+∞
n→+∞
F ((r n )n , ρ∗ ) −→ F ∗ and ||r n || −→ +∞
(c): there exist a sequence (ρn )n such that
n→+∞
n→+∞
F (r ∗ , (ρn )n ) −→ F ∗ and ||ρn || −→ +∞
n
(d): there exist two sequences (r )n and (ρm )m such that
n,m→+∞
F ((r n )n , (ρm )m ) −→ F ∗ while the norm of these two
n,m→+∞
vectors ||r n ||, ||ρm || −→ +∞.
First, we recall a fundamental mathematical theorem on
bounded sequences.
Bolzano–Weierstrass Theorem: Before stating the theorem,
we call to mind the notion of accumulation point. Let (an )
be a sequence of real vectors, the vector L is said to be an
accumulation point of (an ) if there exists a subsequence
(ank ) that converges to L. In other words:
∀ > 0, ∃K ∈ N such that if k > K then ||ank − L|| 6
Fig. 12: Visualization of case (c) analysis
and let pB > 0 its corresponding probability distribution
with fB = f . We construct now the probability distribution
pC in the following way:
pCij = θ1 pAij + (1 − θ1 )pBij
The theorem states that each bounded sequence in RK has
at least one accumulation point or equivalently at least one
convergent subsequence.
Armed with this theorem, we proceed our proof by conn
tradiction. Consider that case (b) occurs, r n = ||r n || ||rrn ||
n
where ||rrn || is a bounded sequence. We denote by r one
n→+∞
of its accumulations point. Since F ((r n )n , ρ∗ ) −→ F ∗
then this is still true for any subsequence extracted from
(r n )n . Consider the subsequence corresponding to the accumulation point r, then for y ≥ 0 which refers to the
modulus, F ∗ = lim F (yr, ρ∗ ). We define the set of all
y→+∞
states Ω = {(aj , xi ), j = 1, . . . , 2K i = 1, . . . , |Ij |}, let
m = max {hxi , ri} and denote Λ = {S ∈ Ω : hxi , ri = m}
S∈Ω
where h·, ·i refers to the scalar product. Then as nk → +∞,
the stationary distribution for a random S = (aj , xi ) ∈ Ω of
our chain becomes:
π(aj , xi ; yr, ρ∗ ) =
exp(haj , ρ∗ i)exp(hxi , yri)
j|
P |I
P
exp(haj , ρ∗ i)exp(hxi , yri)
2K
j=1 i=1
By multiplying both the numerator and denominator by the
same quantity exp(−ym), it leads to:
exp(y(hxi , ri − m + haj , ρ∗ i/y))
y→+∞
−→
K |Ij |
2
P P
∗
j
i
exp(y(hx , ri − m + ha , ρ i/y))
1{S ∈ Λ}
|Λ|
j=1 i=1
y→+∞
Since haj , ρ∗ i/y −→ 0, we are left with the factor
hxi , ri − m 6 0 with y tending to infinity. In order to have a
non zero numerator we need that factor to be null or in other
words S ∈ Λ.
We can distinguish two cases:
- Λ = Ω: In this case, all the states share the same
maximum value m hence the limiting distribution is simply
the uniform distribution over all the state space. In this
case, π(aj , xi ; yr, ρ∗ ) = π(aj , xi , 0, 0) and F ∗ = F (0, 0)
therefore F has a finite minimizer (0, 0) which contradicts
with our assumption
- Λ 6= Ω: In this case, there exist at least a state S 0 =
0
0
0
0
y→+∞
(aj , xi ) ∈
/ Λ and therefore π(aj , xi ; yr, ρ∗ ) −→ 0. But
we assumed that pi0 j 0 > 0 then from the expression of the
objective function we can conclude that F ∗ = +∞ which
is clearly not minimal since F (0, 0) < +∞ hence case (b)
cannot occur.
The proof that case (c) cannot occur is identical to the
preceding one but by taking the sequence (ρn )n into account
instead. The same reasoning can be applied to prove the
impossibility of case (d). The two sequences are taken
simultaneously and we extract from each a subsequence with
their corresponding accumulation points. We will have to
define the following quantities:
- m1 = max {haj , ρi} and denote Λ1 = {S ∈ Ω : haj , ρi =
S∈Ω
m1 }
- m2 = max {hxi , ri} and denote Λ2 = {S ∈ Ω : hxi , ri =
S∈Ω
m2 }
By multiplying the stationary distribution by exp(−m1 y1 −
m2 y2 ), we will end up with:
π(aj , xi ; y2 r, y1 ρ)
y1 ,y2 →+∞
−→
1{S ∈ Λ1 ∩ Λ2 }
|Λ1 ∩ Λ2 |
By following the same analysis as before, we will get to
the conclusion that case (d) cannot occur and we are left with
case (a) which proves the existence of a unique minimizer
(r∗ , ρ∗ ) as long as pij > 0 ∀ S ∈ Ω which brings us back to
our original condition that it is enough to have λ is strictly
feasible and the awake vector f verifying λk < fk < 1 ∀k
to have a unique finite minimizer (r∗ , ρ∗ ).
| 7cs.IT
|
arXiv:1307.2396v1 [math.AC] 9 Jul 2013
DE RAHM COHOMOLOGY OF LOCAL COHOMOLOGY
MODULES-THE GRADED CASE
TONY J. PUTHENPURAKAL
Abstract. Let K be a field of characteristic zero, R = K[X1 , . . . , Xn ]. Let
An (K) = K < X1 , . . . , Xn , ∂1 , . . . , ∂n > be the nth Weyl algebra over K.
We consider the case when R and An (K) is graded by giving deg XiP
= ωi and
deg ∂i = −ωi for i = 1, . . . , n (here ωi are positive integers). Set ω = n
k=1 ωk .
Let I be a graded ideal in R. By a result due to Lyubeznik the local cohomology
modules HIi (R) are holonomic An (K)-modules for each i ≥ 0. In this article we
prove that the De Rahm cohomology modules H ∗ (∂; HI∗ (R)) is concentrated in
degree −ω, i.e., H ∗ (∂; HI∗ (R))j = 0 for j 6= −ω. As an application when A =
1 (R) to H n−1 (∂(f ); A),
R/(f ) is an isolated singularity we relate H n−1 (∂; H(f
)
the (n − 1)th Koszul cohomology of A with respect to ∂1 (f ), . . . , ∂n (f ).
Introduction
Let K be a field of characteristic zero and let R = K[X1 , . . . , Xn ]. We consider
R graded with deg Xi = ωi for i = 1, . . . , n; here ωi are positive integers. Set
m = (X1 , . . . , Xn ). Let I be a graded ideal in R. The local cohomology modules
HI∗ (R) are clearly graded R-modules. Let An (K) = K < X1 , . . . , Xn , ∂1 , . . . , ∂n >
be the nth Weyl algebra over K. By a result due to Lyubeznik, see [2], the local
cohomology modules HIi (R) are holonomic An (K)-modules for each i ≥ 0. We can
consider An (K) graded by giving deg ∂i = −ωi for i = 1, . . . , n.
Let N be a graded left An (K) module. Now ∂ = ∂1 , . . . , ∂n are pairwise commuting K-linear maps. So we can consider the De Rahm complex K(∂; N ). Notice that
the De Rahm cohomology modules H ∗ (∂; N ) are in general only graded K-vector
spaces. They are finite dimensional if N is holonomic; [1, Chapter 1, Theorem 6.1].
In particular H ∗ (∂; HI∗ (R)) are finite dimensional graded K-vector spaces.
Our first result is
Pn
Theorem 1. Let I be a graded ideal in R. Set ω = i=1 ωi . Then the De Rahm
cohomology modules H ∗ (∂1 , . . . , ∂n ; HI∗ (R)) is concentrated in degree −ω, i.e.,
H ∗ (∂1 , . . . , ∂n ; HI∗ (R))j = 0,
for j 6= −ω.
We give an application of Theorem 1. Let f be a homogenous polynomial in
R with A = R/(f ) an isolated singularity, i.e., AP is regular for all homogeneous
prime ideals P 6= m. Let H i (∂(f ); A) be the ith Koszul cohomology of A with
respect to ∂1 (f ), . . . , ∂n (f ). We show
Date: December 11, 2017.
1991 Mathematics Subject Classification. Primary 13D45; Secondary 13N10 .
Key words and phrases. local cohomology, associated primes, D-modules, Koszul homology.
1
2
TONY J. PUTHENPURAKAL
Theorem 2. (with hypotheses as above) There exists a filtration F = {Fν }ν≥0
1
n−1
1
consisting of K-subspaces of H n−1 (∂; H(f
(∂; H(f
) (R)) with Fν = H
) (R)) for
ν ≫ 0, Fν ⊇ Fν−1 and F0 = 0 and injective K-linear maps
Fν
ην :
−→ H n−1 (∂(f ); A)(ν+1) deg f −ω .
Fν−1
1
The techniques in this theorem is generalized in [5] to show that H i (∂; H(f
) (R)) =
0 for i < n − 1. There is no software to compute De-Rahm cohomology of a An (K)module M . As an application of Theorem 2 we prove
2
Example 0.1. Let R = K[X1 , . . . , Xn ] and let f = X12 + X22 + . . . + Xn−1
+ Xnm
with m ≥ 2. Then
1
(1) if m is odd then H n−1 (∂; H(f
) (R)) = 0.
(2) if m is even then
1
(a) If n is odd then H n−1 (∂; H(f
) (R)) = 0.
n−1
1
(b) If n is even then dimK H
(∂; H(f
) (R)) ≤ 1.
We now describe in brief the contents of the paper. In section one we discuss a few
preliminaries that we need. In section two we introduce the concept of generalized
Eulerian modules. In section three we give a proof of Theorem 1. In section four
we give an outline of proof of Theorem 2. In section five we prove Theorem 2. In
section six we give a proof of Example 0.1.
1. Preliminaries
In this section we discuss a few preliminary results that we need.
Remark 1.1. Although all the results are stated for De-Rahm cohomology of
a An (K)-module M , we will actually work with De-Rahm homology. Note that
Hi (∂, M ) = H n−i (∂, M ) for any An (K)-module M . Let S = K[∂1 , . . . , ∂n ]. Consider it as a subring of An (K). Then note that Hi (∂, M ) is the ith Koszul homology
module of M with respect to ∂.
1.2. Let M be a holonomic An (K)-module. Then for i = 0, 1 the De-Rahm homology modules Hi (∂n , M ) are holonomic An−1 (K)-modules, see [1, 1.6.2].
The following result is well-known.
Lemma 1.3. Let ∂ = ∂r , ∂r+1 , . . . , ∂n and ∂ ′ = ∂r+1 , . . . , ∂n . Let M be a left
An (K)-module. For each i ≥ 0 there exist an exact sequence
0 → H0 (∂r ; Hi (∂ ′ ; M )) → Hi (∂; M ) → H1 (∂r ; Hi−1 (∂ ′ ; M )) → 0.
2. generlized Eulerian modules
Consider the Eulerain operator
En = ω1 X1 ∂1 + ω2 X2 ∂2 + · · · + ωn Xn ∂n .
If r ∈ R is homogenous then recall that En r = (deg r) · r. Note that degree of En is
zero.
Let M be a graded An (K) module. If m is homogenous, we set |m| = deg m
We say that M is Eulerian An (K)-module if En m = |m| · m for each homogenous
m ∈ M . This notion was discovered by Ma and Zhang, see the very nice paper [3].
They prove that local cohomology modules HI∗ (R) are Eulerian An (K)-modules,
DE RAHM
3
see [3, 5.3]. Infact they prove this when R is standard graded. The same proof can
be adapted to prove the general case.
It can be easily seen that if M is Eulerian An (K) module then so are each graded
submodule and graded quotient of M . However extensions of Eulerian modules need
not be Eulerian, see [3, 3.5]. To rectify this we introduce the following notion: A
graded An (K)-module M is said to be generalized Eulerian if for a homogenous
element m of M there exists a positive integer a, (here a may depend on m) such
that
(En − |m|)a m = 0.
We now prove that the class of generalized Eulerian modules is closed under extensions.
α
α
2
1
M3 → 0 be a short exact sequence of
M2 −→
Proposition 2.1. Let 0 → M1 −→
graded An (K)-modules. Then the following are equivalent:
(1) M2 is generalized Eulerian.
(2) M1 and M3 is generalized Eulerian.
Proof. The assertion (1) =⇒ (2) is clear. We prove (2) =⇒ (1). Let m ∈ M2 be
homogenous. As M3 is generalized Eulerian we have
(En − |m|)b α2 (m) = 0 for some b ≥ 1.
Set v2 = (En − |m|)b m ∈ M2 . As α2 is An (K)-linear we get α2 (v2 ) = 0. So
v2 = α1 (v1 ) for some v1 ∈ M1 . Note deg v1 = deg v2 = |m|. As M1 is generalized
Eulerian we have that
(En − |m|)a v1 = 0 for some a ≥ 1.
As α1 is An (K)-linear we get (En − |m|)a v2 = 0. It follows that
(En − |m|)a+b m = 0.
If M is a graded An (K)-module then for l ∈ Z the module M (l) denotes the
shift of M by l, i.e., M (l)n = Mn+l for all n ∈ Z. The following result was proved
for Eulerian An (K)-modules in [3, 2.4].
Proposition 2.2. Let M be a non-zero generalized Eulerian An (K)-module. Then
for l 6= 0 the module M (l) is NOT a generalized Eulerian An (K)-module.
Proof. Suppose M (l) is a generalized Eulerian An (K)-module for some l 6= 0. Let
m ∈ M be homogenous of degree r and non-zero. As M is generalized Eulerain
An (K)-module we have that
(En − r)a m = 0 for some a ≥ 1.
We may assume (En − r)a−1 m 6= 0. Now m ∈ M (l)r−l . As M (l) is generalized
Eulerain we get that
(En − r + l)b m = 0 for some b ≥ 1.
Notice
b
0 = (En − r + l) m =
b
X
b b−i
l +
l (En − r)i
i
i=1
b
!
m.
4
TONY J. PUTHENPURAKAL
Multiply on the left by (En − r)a−1 . We obtain
lb (En − r)a−1 m = 0
As l 6= 0 we get (En − r)a−1 m = 0 a contradiction.
3. Proof of Theorem 1
In this section we prove Theorem 1. Notice that HIi (R) are Eulerian An (K)module for all i ≥ 0. Hence Theorem 1 follows from the following more general
result.
Theorem 3.1. Let M be a generalized
Eulerian An (K)-module. Then Hi (∂; M )
P
is concentrated in degree −ω = − nk=1 ωk .
Before proving 3.1 we need to prove a few preliminary results.
Proposition 3.2. Let M be a generalized Eulerian An (K)-module. Then for i =
0, 1, the An−1 (K)-modules Hi (∂n ; M )(−ωn ) are generalized Eulerian.
Proof. Clearly Hi (∂n ; M )(−ωn ) are An−1 (K)-modules for i = 0, 1. We have an
exact sequence of An−1 (K)-modules
∂
n
M → H0 (∂n ; M ) → 0.
0 → H1 (∂n ; M ) → M (ωn ) −→
Note that H1 (∂n ; M )(−ωn ) ⊂ M . Let ξ ∈ H1 (∂n ; M )(−ωn ) be homogeneous. As
M is generalized Eulerian we have that
(En − |ξ|)a ξ = 0 for some a ≥ 1.
Notice En = En−1 + ωn Xn ∂n . Also note that Xn ∂n commutes with En−1 . So
0 = (En−1 − |ξ| + ωn Xn ∂n )a ξ = ((En−1 − |ξ|)a + (∗)Xn ∂n ) ξ.
As ∂n ξ = 0 we get that (En−1 − |ξ|)a ξ = 0. It follows that H1 (∂n ; M )(−ωn ) is a
generalized Eulerian An−1 (K)-module.
Let ξ ∈ H0 (∂n ; M )(−ωn ) be homogenous of degree r. Then ξ = α + ∂n M where
α ∈ Mr−ωn . As M is generalized Eulerian we get that
(En − r + ωn )a α = 0
for some a ≥ 1.
Notice En = En−1 + ωn Xn ∂n = En−1 + ωn ∂n Xn − ωn . So En − r + ωn = En−1 − r +
ωn ∂n Xn . Notice that ∂n Xn commutes with En−1 . Thus
0 = (En−1 − r + ωn ∂n Xn )a α = (En−1 − r)a α + ∂n · ∗α.
Going mod ∂n M we get
(En−1 − r)a ξ = 0.
It follows that H0 (∂n ; M )(−ωn ) is a generalized Eulerian An−1 (K)-module.
Remark 3.3. If M is Eulerian then the same proof shows that Hi (∂n ; M )(−ωn )
is a Eulerian An−1 (K)-module for i = 0, 1. However as the proof of the following
theorem shows that we can only prove that H1 (∂n−1 , ∂n ; M )(−ωn−1 − ωn ) is a
generalized Eulerian An−1 (K)-module.
We now show that
DE RAHM
5
Proposition 3.4. Let M be a generalized Eulerian An (K)-module. Let
∂ = ∂i , ∂i+1 , . . . , ∂n ; here i ≥ 2. Then for each j ≥ 0 the De Rahm homology
module
n
X
Hj (∂; M )(−
ωk )
k=i
is a generalized Eulerian Ai−1 (K)-module.
Proof. We prove this result by descending induction on i. For i = n the result holds byPProposition 3.2. Set ∂ ′ = ∂i+1 , . . . , ∂n . By induction hypothesis
n
Hj (∂ ′ ; M )(− k=i+1 ωk ) is generalized Eulerian Ai (K)-module. By Proposition
3.2 again we get that for l = 0, 1 and for each j ≥ 0
!
n
n
X
X
′
ωk ) (−ωi ) = Hl (∂i ; Hj (∂ ′ ; M )) (−
ωk )
Hl ∂i ; Hj (∂ ; M )(−
k=i+1
k=i
is generalized Eulerian. By 1.3 we have exact sequence
0 → H0 (∂i ; Hj (∂ ′ ; M )) → Hj (∂; M ) → H1 (∂i ; Hj−1 (∂ ′ ; M )) → 0.
The
Pnmodules at the left and right end become generalized Eulerian after shifting by
− k=i ωk . By 2.1 it follows that for each j ≥ 0 the De Rahm homology module
Hj (∂; M )(−
n
X
ωk )
k=i
is a generalized Eulerian Ai−1 (K)-module.
We now consider the case when n = 1.
Proposition 3.5. Let M be a generalized Eulerian A1 (K)-module. Then for l =
0, 1 the modules Hl (∂1 ; M ) is concentrated in degree −ω1 .
Proof. We have an exact sequence of K-vector spaces
∂
1
M → H0 (∂1 ; M ) → 0.
0 → H1 (∂1 ; M ) → M (ω1 ) −→
Let ξ ∈ H1 (∂1 ; M )(−ω1 ) be homogenous and non-zero. As ξ ∈ M we have that
(ω1 X1 ∂1 − |ξ|)a ξ = 0 for some a ≥ 1.
Notice (ω1 X1 ∂1 − |ξ|)a = (∗)∂1 + (−1)a |ξ|a . As ∂1 ξ = 0 we get (−1)a |ξ|a ξ = 0. As
ξ 6= 0 we get that |ξ| = 0. It follows that H1 (∂1 ; M ) is concentrated in degree −ω1 .
Let ξ ∈ H0 (∂1 , M ) be non-zero and homogeneous of degree r. Let ξ = α + ∂1 M
where α ∈ Mr . As M is generalized Eulerian we get that
(ω1 X1 ∂1 − r)a α = 0 for some a ≥ 1.
Notice ω1 X1 ∂1 = ω1 ∂1 X1 − ω1 . So we have
0 = (ω1 ∂1 X1 − (r + ω1 ))a α = (∂1 ∗ +(−1)a (r + ω1 )a ) α.
In M/∂1 M we have (−1)a (r + ω1 )a ξ = 0. As ξ 6= 0 we get that r = −ω1 . It follows
that H0 (∂1 ; M ) is concentrated in degree −ω1 .
We now give
6
TONY J. PUTHENPURAKAL
Pn
Proof of Theorem 3.1. Set ∂ ′ = ∂2 , . . . , ∂n . By 3.4 Nj = Hj (∂ ′ ; M )(− k=2 ωk ) is
generalized Eulerian
P A1 (K)-module, for each j ≥ 0. We use exact sequence in 1.3
and shift it by − nk=2 ωk to obtain an exact sequence
0 → H0 (∂1 , Nj ) → Hj (∂; M )(−
n
X
ωk ) → H1 (∂1 , Nj−1 ) → 0,
k=2
for each j ≥ 0. By Proposition 3.5 the modules on the left and right of the above
exact sequence is concentrated in degree −ω1 . It follows that
Pnfor each j ≥ 0 the
K-vector space Hj (∂; M ) is concentrated in degree −ω = − k=1 ωk .
4. Outline of proof of Theorem 2
The proof of Theorem 2 is a bit long and has a lot of technical details. For the
convenience of the reader we give an outline of the proof.
∼ H1 (∂, H 1 (R)). So it is sufficient to
4.1. By [4, Lemma 2.7] we have H1 (∂, Rf ) =
(f )
work with H1 (∂, Rf ) in order to prove Theorem 2. We consider elements of Rfm
as column-vectors. For x ∈ Rfm we write it as x = (x1 , . . . , xm )′ ; here ′ indicates
transpose.
4.2. Let ξ ∈ Rfm \ Rm . The element (a1 /f i , a2 /f i , . . . , am /f i )′ , with aj ∈ R for all
j, is said to be a normal form of ξ if
(1) ξ = (a1 /f i , a2 /f i , . . . , am /f i )′ .
(2) f does not divide aj for some j.
(3) i ≥ 1.
It can be easily shown that normal form of ξ exists and is unique; see Proposition
5.1. Let (a1 /f i , a2 /f i , . . . , am /f i )′ be the normal form of ξ. Set L(ξ) = i. Notice
L(ξ) ≥ 1.
4.3. Construction of a function θ : Z1 (∂, Rf ) \ Rn → H1 (∂(f ); A).
n
i
i
i ′
Let ξ ∈ ZP
1 (∂, Rf ) \ R . Let (a1 /f , a2 /f , . . . , an /f ) be the normal form of ξ.
n
i
So we have j=1 ∂/∂Xj (aj /f ) = 0. So we have
n
n
∂f
i X
1 X ∂aj
aj
− i+1
= 0.
f i j=1 ∂Xj
f
∂Xj
j=1
It follows that
f divides
n
X
j=1
′
So (a1 , . . . , an ) ∈ Z1 (∂(f ); A). We set
aj
∂f
.
∂Xj
θ(ξ) = [(a1 , . . . , an )′ ] ∈ H1 (∂(f ); A).
Remark 4.4. It can be shown that if ξ ∈ Z1 (∂, Rf )−ω is non-zero then ξ ∈
/ Rn ,
see 5.2. If L(ξ) = i then by 5.3 we have
θ(ξ) ∈ H1 (∂(f ); A)(i+1) deg f −ω .
The next result uses the fact that A is an isolated singularity.
Proposition 4.5. If ξ ∈ B1 (∂, Rf )−ω is non-zero then θ(ξ) = 0.
DE RAHM
7
4.6. Let ξ ∈ Rfm . We define L(f ) as follows.
Case 1: ξ ∈ Rfm \ Rm .
Let (a1 /f i , a2 /f i , . . . , am /f i )′ be the normal form of ξ. Set L(ξ) = i. Notice
L(ξ) ≥ 1 in this case.
Case 2: ξ ∈ Rm \ {0}.
Set L(ξ) = 0.
Case 3: ξ = 0.
Set L(ξ) = −∞.
The following properties of the function L can be easily verified.
Proposition 4.7. (with hypotheses as above) Let ξ, ξ1 , ξ2 ∈ Rfm and α, α1 , α2 ∈ K.
(1)
(2)
(3)
(4)
(5)
(6)
(7)
If L(ξ1 ) < L(ξ2 ) then L(ξ1 + ξ2 ) = L(ξ2 ).
If L(ξ1 ) = L(ξ2 ) then L(ξ1 + ξ2 ) ≤ L(ξ2 ).
L(ξ1 + ξ2 ) ≤ max{L(ξ1 ), L(ξ2 )}.
If α ∈ K ∗ then L(αξ) = L(ξ).
L(αξ) ≤ L(ξ) for all α ∈ K.
L(α1 ξ1 + α2 ξ2 ) ≤ max{L(ξ1 ), L(ξ2 )}.
Let ξ1 , . . . , ξr ∈ Rfm and let α1 , . . . , αr ∈ K. Then
r
X
L
αj ξj ≤ max{L(ξ1 ), L(ξ2 ), . . . , L(ξr )}.
j=1
4.8. We now use the fact that H1 (∂, Rf ) is concentrated in degree −ω = −
So
Z1 (∂, Rf ))−ω
H1 (∂, Rf ) = H1 (∂, Rf )−ω =
.
B1 (∂, Rf )−ω
Let x ∈ H1 (∂, Rf ) be non-zero. Define
Pn
k=1
ωk .
L(x) = min{L(ξ) | x = [ξ], where ξ ∈ Z1 (∂, Rf )−ω .}.
It can be shown that L(x) ≥ 1. If x = 0 then set
L(0) = −∞.
We now define a function
θe: H1 (∂, Rf ) → H1 (∂(f ); A)
(
θ(ξ) if x 6= 0, x = [ξ], and L(x) = L(ξ),
x 7→
0
if x = 0.
e
It can be shown that θ(x)
is independent of choice of ξ; see Proposition 5.6. Also
e
note that if L(x) = i then θ(x)
∈ H1 (∂(f ); A)(i+1) deg f −ω .
4.9. We now construct a filtration F = {Fν }ν≥0 of H1 (∂, Rf ). Set
Fν = {x ∈ H1 (∂, Rf ) | L(x) ≤ ν}.
In the next section we prove
Proposition 4.10. (1) Fν is a K subspace of H1 (∂, Rf ).
(2) Fν ⊇ Fν−1 for all ν ≥ 1.
(3) Fν = H1 (∂, Rf ) for all ν ≫ 0.
(4) F0 = 0.
8
TONY J. PUTHENPURAKAL
Let G =
L
ν≥1
Fν /Fν−1 . For ν ≥ 1 we define
ην :
Fν
→ H1 (∂(f ); A)(ν+1) deg f −ω
Fν−1
(
0,
if ξ = 0;
ξ 7→ e
θ(x), if ξ = x + Fν−1 is non-zero.
It can be shown that ην (ξ) is independent of choice of x; see 5.10. Finally we prove
the following result.
Theorem 4.11. (with Notation as above). For all ν ≥ 1,
(1) ην is K-linear.
(2) ην is injective.
5. Proof of Theorem 2
In this section we give a proof of Theorem 2 with all details. The reader is
advised to read the previous section before reading this section.
We first prove
Proposition 5.1. Let ξ ∈ Rfm \ Rm . Then normal form of ξ exists and is unique.
Proof. Existence: Let ξ ∈ Rfm \ Rm . Let ξ = (b1 /f i1 , b2 /f i2 , · · · , bm /f im )′ with
f ∤ bj if bj 6= 0. Note ij ≤ 0 is possible. Let
ir = max{ij | ij ≥ 1 and bj 6= 0}.
Notice ir ≥ 1. Then
b1 f ir −i1 b2 f ir −i2
bm f ir −im ′
,
,
·
·
·
,
)
f ir
f ir
f ir
Note that f ∤ br . Thus the expression above is a normal form of ξ.
Uniqueness: Let (a1 /f i , · · · , am /f i )′ and (b1 /f r , · · · , bm /f r )′ be two normal
forms of ξ. We first assert that i < r is not possible. For if this holds then note as
aj /f i = bj /f r we get bj = aj f r−i . So f |bj for all j; a contradiction.
A similar argument shows that i > r is not possible. So i = r. Thus aj = bj for
all j. Thus the normal form of ξ is unique.
ξ=(
5.2. Let ξ ∈ Z1 (∂, Rf )−ω be non-zero. Let ξ = (ξ1 , . . . , ξn )′ . Note that
ξ ∈ (Rf (ω1 ) ⊕ Rf (ω2 ) ⊕ · · · ⊕ Rf (ωn ))−ω
It follows that
ξj ∈ (Rf )− Pk6=j ωk .
It follows that ξ ∈ Rfn \ Rn .
5.3. Let (a1 /f i , · · · , an /f i )′ be the normal form of ξ. Then
X
deg aj = i deg f −
ωk .
k6=j
In particular going mod f we get
aj ∈ A (− deg f + ωj )(i+1) deg f −ω .
Notice deg ∂f /∂Xj = deg f − ωj . It follows that
(a1 , · · · , an )′ ∈ Z1 (∂(f ); A)(i+1) deg f −ω .
DE RAHM
9
Thus θ(ξ) ∈ H1 (∂(f ); A)(i+1) deg f −ω .
5.4. Let K = K(∂; Rf ) be the De Rahm complex on Rf written homologically. So
φ1
φ2
φ3
K = · · · → K3 −→ K2 −→ K1 −→ K0 → 0.
L
Here K0 = Rf , K1 = nk=1 Rf (ωk ),
M
M
K2 =
Rf (ωi + ωj ) and K3 =
Rf (ωi + ωj + ωl ).
1≤i<j≤n
1≤i<j<l≤n
Let K′ = K(∂(f ); A) be the Koszul complex on A with respect to
∂f /∂X1, . . . , ∂f /∂Xn. So
ψ3
ψ2
ψ1
K′ = · · · → K′3 −−→ K′2 −−→ K′1 −−→ K′0 → 0.
Ln
Here K′0 = A, K′1 = k=1 A(− deg f + ωk ),
M
M
K′2 =
A(−2 deg f +ωi +ωj ) and K′3 =
A(−3 deg f +ωi +ωj +ωl ).
1≤i<j≤n
1≤i<j<l≤n
We now give
proof of Proposition 4.5. Let u ∈ B1 (∂; Rf )−ω be non-zero. Let ξ ∈ (K2 )−ω be
homogeneous with φ2 (ξ) = u. Let ξ = (ξij | 1 ≤ i < j ≤ n)′ . Notice that
ξij ∈ Rf (ωi + ωj )−ω = (Rf )− Pk6=i,j ωk .
n
(n)
It follows that ξ ∈ Rf 2 \ R( 2 ) . Set
c = min{j | j = L(ξ) where φ2 (ξ) = u and ξ ∈ (K2 )−ω is homogeneous}.
Notice c ≥ 1. Let ξ ∈ (K2 )−ω be such that L(ξ) = c and φ2 (ξ) = u. Let (bij /f c |
1 ≤ i < j ≤ n)′ be the normal form of ξ. Let u = (u1 , . . . , un )′ . Then for
l = 1, . . . , n
X ∂ bil X ∂ blj
−
ul =
∂Xi f c
∂Xj f c
j>l
i<l
So
ul =
Set
f
f c+1
Therefore
X ∂(bil )
i<l
∂Xi
−
X ∂(blj )
j>l
vl = c −
∂Xj
X
i<l
+
c
−
f c+1
X
i<l
bil
∂f
+
∂Xi
X
∂f
∂f
+
blj
bil
∂Xi
∂Xj
ul =
X
j>l
blj
∂f
∂Xj
j>l
f ∗ +vl
.
f c+1
Claim: f ∤ vl for some l.
First assume the claim. Then ((f ∗ +v1 )/f c+1 , . . . , (f ∗ +vn )/f c+1 )′ is normal
form of u. So
θ(u) = [(v1 , . . . , vn )′ ] = [ψ2 (−cb)] = 0.
We now prove our claim. Suppose if possible f |vl for all l. Then
ψ2 (−cb) = (v1 , . . . , vl )′ = 0.
10
TONY J. PUTHENPURAKAL
So −cb ∈ Z2 (∂(f ); A). As H2 (∂(f ); A) = 0 we get −cb ∈ B2 (∂(f ); A). Thus
−cb = ψ3 (γ). Here
γ = (γijl | 1 ≤ i < j < l ≤ n)′ .
So
X
X
X
∂f
∂f
∂f
(5.4.1)
− cbij =
γkij
−
+
+ αij f.
γikj
γijk
∂Xk
∂Xk
∂Xk
k<i<j
i<k<j
i<j<k
We need to compute degree of γijl . Note that ξ ∈ (K2 )−ω . So
bij
∈ (Rf (ωi + ωj ))−ω .
fc
It follows that
(5.4.2)
deg bij = c deg f − ω + ωi + ωj
It can be easily checked that
b ∈ (K′2 )(c+2) deg f −ω .
So
γ ∈ (K′3 )(c+2) deg f −ω .
It follows that
(5.4.3)
deg γijl = (c − 1) deg f − ω + ωi + ωj + ωl .
We first consider the case when c = 1. Then by equation 5.4.1 we have αij = 0.
Also
deg γijl = −ω + ωi + ωj + ωl < 0 if n > 3.
So if n > 3 we get γijl = 0. So b = 0 and so ξ = 0, a contradiction.
We now consider the case when n = 3. Note that γ = γ123 is a constant. Thus
′
∂f
∂f
∂f
, −γ
,γ
.
b= γ
∂X3
∂X2 ∂X3
A direct computation yields u = 0, a contradiction.
We now consider the case when c ≥ 2. Notice by equation 5.4.1 we have
−cbij
1 X
1 X
1 X
αij
∂f
∂f
∂f
= c
− c
+ c
+ c−1 .
γkij
γikj
γijk
c
f
f
∂Xk
f
∂Xk
f
∂Xk
f
k<i<j
Notice
i<k<j
∂
γkij ∂f /∂Xk
=
c
f
∂Xk
Put
γe∗ =
i<j<k
γkij /(1 − c)
f c−1
−
∗
f c−1
1
γ∗ .
c(c − 1)
So we obtain
X ∂ γg
X ∂ γg
X ∂ γg
bij
bf
ij
kij
ikj
ijk
−
+
+ c−1 .
=
c
c−1
c−1
c−1
f
∂Xk f
∂Xk f
∂Xk f
f
k<i<j
Set
δ=
i<k<j
γg
ijl
|1≤i<j<l≤n
f c−1
i<j<k
and ξe =
!
bf
ij
|1≤i<j≤n .
f c−1
DE RAHM
11
Then
e
ξ = φ3 (δ) + ξ.
e This contradicts choice of c.
So we have u = φ2 (ξ) = φ2 (ξ).
5.5. By Theorem 3.1 we have
H1 (∂; Rf ) = H1 (∂; Rf )−ω =
Z1 (∂; Rf )−ω
.
B1 (∂; Rf )−ω
Let x ∈ H1 (∂; Rf ) be non-zero. Define
L(x) = min{L(ξ) | x = [ξ], where ξ ∈ Z1 (∂, Rf )−ω .}.
Let ξ = (ξ1 , . . . , ξn )′ ∈ Z1 (∂, Rf )−ω be such that x = [ξ]. So ξ ∈ (K1 )−ω . Thus
ξi ∈ Rf (+ωi )−ω . So if ξ 6= 0 then ξ ∈ Rfn \ Rn . It follows that L(ξ) ≥ 1. Thus
L(x) ≥ 1.
We now define a function
θe: H1 (∂, Rf ) → H1 (∂(f ); A)
(
θ(ξ) if x 6= 0, x = [ξ], and L(x) = L(ξ),
x 7→
0
if x = 0.
g is independent of choice of ξ.
Proposition 5.6. (with hypotheses as above) θ(x)
Proof. Suppose x = [ξ1 ] = [ξ2 ] is non-zero and L(x) = L(ξ1 ) = L(ξ2 ) = i. Let
(a1 /f i , . . . , an /f i )′ be normal form of ξ1 and let (b1 /f i , . . . , bn /f i )′ be normal form
of ξ2 . It follows that ξ1 = ξ2 + δ where δ ∈ B1 (∂; Rf )−ω . By Proposition 4.7.1 we
get j = L(δ) ≤ i. Let (c1 /f j , . . . , cn /f j )′ be normal form of δ. We consider two
cases.
Case 1. j < i. Then note that ak = bk + f i−j ck for k = 1, . . . , n. It follows that
θ(ξ1 ) = [(a1 , . . . , an )] = [(b1 , . . . , bn )] = θ(ξ2 ).
Case 2. j = i. Then note that ak = bk + ck for k = 1, . . . , n. It follows that
θ(ξ1 ) = θ(ξ2 ) + θ(δ).
g is independent
However by Proposition 4.5, θ(δ) = 0. So θ(ξ1 ) = θ(ξ2 ). Thus θ(x)
of choice of ξ.
5.7. We now construct a filtration F = {Fν }ν≥0 of H1 (∂, Rf ). Set
Fν = {x ∈ H1 (∂, Rf ) | L(x) ≤ ν}
We prove
Proposition 5.8. (1) Fν is a K subspace of H1 (∂; Rf ).
(2) Fν ⊇ Fν−1 for all ν ≥ 1.
(3) Fν = H1 (∂; Rf ) for all ν ≫ 0.
(4) F0 = 0.
Proof. (1) Let x ∈ Fν and let α ∈ K. Then by *
L(αx) ≤ L(x) ≤ ν.
So αx ∈ Fν .
12
TONY J. PUTHENPURAKAL
Let x, x′ ∈ Fν be non-zero. Let ξ, ξ ′ ∈ Z1 (∂; Rf ) be such that x = [ξ], x′ = [ξ ′ ]
and L(x) = L(ξ), L(x′ ) = L(ξ ′ ) Then x + x′ = [ξ + ξ ′ ]. It follows that
L(x + x′ ) ≤ L(ξ + ξ ′ ) ≤ max{L(ξ), L(ξ ′ )} ≤ ν.
Note that the second inequality follows from Proposition 4.7. Thus x + x′ ∈ Fν .
(2) This is clear.
(3) Let B = {x1 , . . . , xm } be a K-basis of H1 (∂; Rf ) = H1 (∂; Rf )−ω . Let
c = max{L(xi ) | i = 1, . . . , m}.
We claim that
Fν = H1 (∂; Rf ) for all ν ≥ c.
Fix ν ≥ c. Let ξi ∈ Z1 (∂; Rf )−ω be such that xi = [ξi ] and L(xi ) = L(ξi ) for
i = 1, . . . , m.
Pm
u ∈ H1 (∂; Rf ). Say u = i=1 αi xi for some α1 , . . . , αm ∈ K. Then u =
PLet
m
[ i=1 αi ξi ]. It follows that
m
X
αi xi ) ≤ max{L(ξi ) | i = 1, . . . , m} = c ≤ ν.
L(u) ≤ L(
i=1
Here the second inequality follows from Proposition 4.7. So u ∈ Fν . Thus Fν =
H1 (∂; Rf ).
(4) If x ∈ H1 (∂; Rf ) is non-zero then L(x) ≥ 1. It follows that F0 = 0.
L
5.9. Let G = ν≥1 Fν /Fν−1 . For ν ≥ 1 we define
ην :
Fν
→ H1 (∂(f ); A)(ν+1) deg f −ω
Fν−1
(
0,
if u = 0;
u 7→ e
θ(x), if u = x + Fν−1 is non-zero.
Proposition 5.10. [with hypotheses as above.] ην (u) is independent of choice of
x.
Proof. Suppose u = x + Fν−1 = x′ + Fν−1 be non-zero. Then x = x′ + y where
y ∈ Fν−1 . As u 6= 0 we have that x, x′ ∈ Fν \ Fν−1 . So L(x) = L(x′ ) = ν. Say
x = [ξ], x′ = [ξ ′ ] and y = [δ] where ξ, ξ ′ , δ ∈ Z1 (∂; Rf ) with L(ξ) = L(ξ ′ ) = ν and
L(δ) = L(y) = k ≤ ν − 1. So we have that ξ = ξ ′ + δ + α where α ∈ B1 (∂; Rf )−ω .
Let L(α) = r. Note r ≤ ν.
Let (a1 /f ν , . . . , an /f ν )′ , (a′1 /f ν , . . . , a′n /f ν )′ , (b1 /f k , . . . , bn /f k )′ and
(c1 /f r , . . . , cn /f r )′ be normal forms of ξ, ξ ′ , δ and α respectively. So we have
aj = a′j + f ν−k bj + f ν−r cj
for j = 1, . . . , n.
Case 1: r < ν. In this case we have that aj = a′j in A for each j = 1, . . . , n. So
e
e ′ ).
θ(ξ) = θ(ξ ′ ). Thus θ(x)
= θ(x
Case 2: r = ν. In this case notice that aj = a′j + cj in A for each j = 1, . . . , n.
So θ(ξ) = θ(ξ ′ ) + θ(α). However θ(α) = 0 as α ∈ B1 (∂; Rf )−ω ; see Proposition 4.5.
e
e ′ ).
Thus θ(x)
= θ(x
Note that neither θ nor θe is linear. However we prove the following:
Proposition 5.11. (with Notation as above). For all ν ≥ 1, ην is K-linear.
DE RAHM
13
Proof. Let u, u′ ∈ Fν /Fν−1 . We first show ην (αu) = αην (u) for all α ∈ K. We
have nothing to show if α = 0 or if u = 0. So assume α 6= 0 and u 6= 0. Say
e
e
u = x + Fν−1 . Then αu = αx + Fν−1 . As θ(αx)
= αθ(x)
we get the result.
Next we show that ην (u + u′ ) = ην (u) + ην (u′ ). We have nothing to show if
u or u′ is zero. Next we consider the case when u + u′ = 0. Then u = −u′ . So
ην (u) = −ην (u′ ). Thus in this case
ην (u + u′ ) = 0 = ην (u) + ην (u′ ).
Now consider the case when u, u′ are non-zero and u + u′ is non-zero. Say u =
x + Fν−1 and u′ = x′ + Fν−1 . Note that as u + u′ is non-zero x + x′ ∈ Fν \ Fν−1 .
Let x = [ξ] and x′ = [ξ ′ ] where ξ, ξ ′ ∈ Z1 (∂; Rf )−ω and L(ξ) = L(ξ ′ ) = ν. Then
x + x′ = [ξ + ξ ′ ]. Note that L(ξ + ξ ′ ) ≤ ν by Proposition 4.7. But L(x + x′ ) = ν.
So L(ξ + ξ ′ ) = ν. Let (a1 /f ν , . . . , an /f ν )′ , (a′1 /f ν , . . . , a′n /f ν )′ be normal forms of
ξ and ξ ′ respectively. Note that ((a1 + a′1 )/f ν , . . . , (an + a′n )/f ν )′ is normal form
e + x′ ) = θ(x)
e + θ(x
e ′ ).
of ξ + ξ ′ . It follows that θ(ξ + ξ ′ ) = θ(ξ) + θ(ξ ′ ). Thus θ(x
Therefore
ην (u + u′ ) = ην (u) + ην (u′ ).
Finally we have the main result of this section
Proof of Theorem 2. Let ν ≥ 1. By Proposition 5.11 we know that ην is a linear
map of K-vector spaces. We now prove that ην is injective.
Suppose if possible ην is not injective. Then there exists non-zero u ∈ Fν /Fν−1
with ην (u) = 0. Say u = x + Fν−1 . Also let x = [ξ] where ξ ∈ Z1 (∂; Rf )−ω and
L(ξ) = L(x) = ν. Let (a1 /f ν , . . . , an /f ν )′ be the normal form of ξ. So we have
e
0 = ην (u) = θ(x)
= θ(ξ) = [(a1 , . . . , an )′ ].
It follows that (a1 , . . . , an )′ = ψ2 (b) where b = (bij | 1 ≤ i < j ≤ n)′ . it follows that
for l = 1, . . . , n
X
X
∂f
∂f
al =
bil
blj
−
.
∂Xi
∂Xj
i<l
l>j
It follows that for l = 1, . . . , n we have the following equation in R:
X
X
∂f
∂f
(5.11.1)
al =
bil
−
+ dl f,
blj
∂Xi
∂Xj
i<l
l>j
for some dl ∈ R. Note that the above equation is of homogeneous elements in R.
So we have the following
P
P
∂f
∂f
dl
al
l>j blj ∂Xj
i<l bil ∂Xi
=
−
+ ν−1 .
(5.11.2)
ν
ν
ν
f
f
f
f
We consider two cases:
Case 1: ν ≥ 2.
Set bf
ij = −bil /(c − 1). Then note that
∂f
bil ∂X
i
fν
∂
=
∂Xi
bf
il
f ν−1
!
−
∗
f ν−1
.
14
TONY J. PUTHENPURAKAL
By 5.11.2 we have for l = 1, . . . , n,
!
X ∂
X ∂
al
bf
il
−
=
ν
ν−1
f
∂Xi f
∂Xj
l<j
i<l
bf
lj
f ν−1
!
+
cl
.
ν−1
f
ν−1
Put ξ ′ = (c1 /f ν−1 , . . . , cn /f ν−1 )′ and δ = (bf
| 1 ≤ i < j ≤ n). Then we
ij /f
have
ξ = φ2 (δ) + ξ ′ .
′
So we have x = [ξ] = [ξ ]. This yields L(x) ≤ L(ξ ′ ) ≤ ν − 1. This is a contradiction.
Case 2: ν = 1.
Note that ξ ∈ (K1 )−ω . Thus for l = 1, . . . , n we have
al
∈ (Rf (ωl ))−ω .
f
It follows that
X
deg al = deg f −
ωk .
k6=l
Also note that deg ∂f /∂Xi = deg f − ωi . By comparing degrees in 5.11.1 we get
al = 0 for all l. Thus ξ = 0. So x = 0. Therefore u = 0 a contradiction.
6. Example 0.1
2
+ Xnm with m ≥ 2. Set
Let R = K[X1 , . . . , Xn ] and let f = X12 + . . . + Xn−1
1
A = R/(f ). In this section we compute H1 (∂; H(f ) (R)).
6.1. We give ωi = deg Xi = m for i = 1, . . . , n − 1 and ωn = deg Xn = 2.
NotePthat f is a homogeneous polynomial in R of degree 2m. Also note that
ω = nk=1 ωk = (n − 1)m + 2.
6.2. First note that the Jacaobian ideal J of f is primary to the unique graded
maximal ideal of R. It follows that A is an isolated singularity. Note J =
(X1 , . . . , Xn−1 , Xnm−1 ). Let Hi (J; A) be the ith -Koszul homology of A with respect to J.
Proposition 6.3. The Hilbert series, P (t), of H1 (J; A) is
P (t) =
m−2
X
t2m+2j .
k=0
Proof. It is easily verified that X1 , . . . , Xm−1 is an A-regular sequence. Set
K[Xn ]
= K ⊕ KXn ⊕ Xn2 ⊕ · · · ⊕ KXnm−1 .
(Xnm )
Note that we have an exact sequence
B = A/(X1 , . . . , Xm−1 )A =
X m−1
0 → H1 (J; A) → B(−2(m − 1)) −−n−−→ B.
It follows that H1 (J; A) = Xn B(−2(m − 1)). The result follows
6.4. By Theorem 2 there exists a filtration F = {Fν }ν≥0 consisting of K-subspaces
1
n−1
1
of H1 (∂; H(f
(∂; H(f
) (R)) with Fν = H
) (R)) for ν ≫ 0, Fν ⊇ Fν−1 and F0 = 0
and injective K-linear maps
Fν
ην :
−→ H1 (∂(f ); A)(ν+1) deg f −ω .
Fν−1
DE RAHM
15
Notice
(ν + 1) deg f − ω = (ν + 1)2m − (n − 1)m − 2 = (2ν − n + 3)m − 2.
If ην 6= 0 then by Proposition 6.3 it follows that
(2ν − n + 3)m − 2 = 2m + 2j
for some j = 0, . . . , m − 2.
So we obtain
(6.4.1)
2νm = (n − 1)m + 2(j + 1)
It follows that m divides 2(j + 1). As 2(j + 1) ≤ 2m − 2 it follows that 2(j + 1) = m.
Thus m is even.
6.5. Say m = 2r. Then by equation (6.4.1) we have
2νr = (n − 1)r + r.
So ν = n/2. It follows that n is even. Furthermore note that ην = 0 for ν 6= n/2
and that if ν = n/2 then by 6.3, dim Fn/2 /Fn/2−1 ≤ 1. It follows that in this case
1
dim H1 (∂; H(f
) (R)) ≤ 1.
6.6. In conclusion we have
1
(1) if m is odd then H n−1 (∂; H(f
) (R)) = 0.
(2) if m is even then
1
(a) If n is odd then H n−1 (∂; H(f
) (R)) = 0.
n−1
1
(b) If n is even then dimK H
(∂; H(f
) (R)) ≤ 1.
This proves Example 0.1.
References
[1] J.-E. Björk, Rings of differential operators. North-Holland Mathematical Library, 21. NorthHolland Publishing Co., Amsterdam-New York, 1979.
[2] G. Lyubeznik, Finiteness properties of local cohomology modules (an application of Dmodules to commutative algebra). Invent. Math. 113 (1993), no. 1, 4155.
[3] L. Ma and W. Zhang, Eulerian graded D-modules, eprint arXiv:1210.8402.
[4] T. J. Puthenpurakal, De Rahm cohomoology of local cohomology modules, Preprint.arXiv:
1302.0116v1.
1 (R), where V (f ) is
[5] T. J. Puthenpurakal and R. B. T. Reddy, De Rahm cohomology of H(f
)
a smooth hypersurface in Pn . In preparation
Department of Mathematics, Indian Institute of Technology Bombay, Powai, Mumbai
400 076
E-mail address: [email protected]
| 0math.AC
|
QuPARA: Query-Driven Large-Scale Portfolio
Aggregate Risk Analysis on MapReduce
Andrew Rau-Chaplin, Blesson Varghese, Duane Wilson, Zhimin Yao, and Norbert Zeh
arXiv:1308.3615v1 [cs.DC] 16 Aug 2013
Risk Analytics Lab, Dalhousie University
Halifax, Nova Scotia, Canada
Email: {arc, varghese, yao, nzeh}@cs.dal.ca
Abstract—Modern insurance and reinsurance companies use
stochastic simulation techniques for portfolio risk analysis. Their
risk portfolios may consist of thousands of reinsurance contracts
covering millions of individually insured locations. To quantify
risk and to help ensure capital adequacy, each portfolio must
be evaluated in up to a million simulation trials, each capturing
a different possible sequence of catastrophic events (e.g., earthquakes, hurricanes, etc.) over the course of a contractual year.
In this paper, we explore the design of a flexible framework
for portfolio risk analysis that facilitates answering a rich variety
of catastrophic risk queries. Rather than aggregating simulation
data in order to produce a small set of high-level risk metrics
efficiently (as is often done in production risk management
systems), the focus here is on allowing the user to pose queries on
unaggregated or partially aggregated data. The goal is to provide
a flexible framework that can be used by analysts to answer a
wide variety of unanticipated but natural ad hoc queries. Such
detailed queries can help actuaries or underwriters to better
understand the multiple dimensions (e.g., spatial correlation,
seasonality, peril features, construction features, financial terms,
etc.) that can impact portfolio risk and thus company solvency.
We implemented a prototype system, called QuPARA (QueryDriven Large-Scale Portfolio Aggregate Risk Analysis), using
Hadoop, which is Apache’s implementation of the MapReduce
paradigm. This allows the user to take advantage of large parallel
compute servers in order to answer ad hoc risk analysis queries
efficiently even on very large data sets typically encountered in
practice. We describe the design and implementation of QuPARA
and present experimental results that demonstrate its feasibility.
A full portfolio risk analysis run consisting of a 1,000,000 trial
simulation, with 1,000 events per trial, and 3,200 risk transfer
contracts can be completed on a 16-node Hadoop cluster in just
over 20 minutes.
Keywords—ad hoc risk analytics; aggregate risk analytics;
portfolio risk; MapReduce; Hadoop
I.
I NTRODUCTION
At the heart of the analytical pipeline of a modern insurance/reinsurance company is a stochastic simulation technique
for portfolio risk analysis and pricing referred to as Aggregate
Analysis [1], [2], [3], [4]. At an industrial scale, a risk portfolio
may consist of thousands of annual reinsurance contracts
covering millions of individually insured locations. To quantify
annual portfolio risk, each portfolio must be evaluated in up
to a million simulation trials, each consisting of a sequence
of possibiliy thousands of catastrophic events, such as earthquakes, hurricanes or floods. Each trial captures one scenario
how globally distributed catastrophic events may unfold in a
year.
Aggregate analysis is computationally intensive as well
as data-intensive. Production analytical pipelines exploit parallelism in aggregate risk analysis and ruthlessly aggregate
results. The results obtained from production pipelines summarize risk in terms of a small set of standard portfolio
metrics that are key to regulatory bodies, rating agencies, and
an organisation’s risk management team, such as Probable
Maximum Loss (PML) [5], [6] and Tail Value-at-Risk (TVaR)
[7], [8]. While production pipelines can efficiently aggregate
terabytes of simulation results into a small set of key portfolio
risk metrics, they are typically very poor at answering the
types of ad hoc queries that can help actuaries or underwriters
to better understand the multiple dimensions of risk that can
impact a portfolio, such as spatial correlation, seasonality, peril
features, construction features, and financial terms.
This paper proposes a framework for aggregate risk analysis that facilitates answering a rich variety of ad hoc queries
in a timely manner. A key characteristic of the proposed
framework is that it is designed to allow users with extensive
mathematical and statistical skills but perhaps limited programming background, such as risk analysts, to pose a rich variety
of complex risk queries. The user formulates their query by
defining SQL-like filters. The framework then answers the
query based on these filters, without requiring the user to
make changes to the core implementation of the framework
or to reorganize the input data of the analysis. The challenges
that arise due to the amounts of data to be processed and
due to the complexity of the required computations are largely
encapsulated within the framework and hidden from the user.
Our prototype implementation of this framework for
Query-Driven Large-Scale Portfolio Aggregate Risk Analysis,
referred to as QuPARA, uses Apache’s Hadoop [9], [10]
implementation of the MapReduce programming model [11],
[12], [13] to exploit parallelism, and Apache Hive [14], [15]
to support ad hoc queries. Even though QuPARA is not as fast
as a production system on the narrow set of standard portfolio
metrics, it can answer a wide variety of ad hoc queries in an
efficient manner. For example, our experiments demonstrate
that an industry-size risk analysis with 1,000,000 simulation
trials, 1,000 events per trial, and on a portfolio consisting of
3,200 risk transfer contracts (layers) with an average of 5 event
loss tables per layer can be carried out on a 16-node Hadoop
cluster in just over 20 minutes.
The remainder of this paper is organized as follows.
Section II gives an overview of reinsurance risk analysis. Section III proposes our new risk analysis framework. Section IV
considers an example query processed by the framework and
various queries for fine-grained aggregate risk analysis. Section V describes implementation details. Section VI presents a
performance evaluation of our framework. Section VII presents
conclusions and discusses future work.
II.
A N OVERVIEW OF R ISK A NALYSIS
A reinsurance company typically holds a portfolio of programs that insure primary insurance companies against largescale losses, like those associated with catastrophic events.
Each program contains data that describes (1) the buildings
to be insured (the exposure), (2) the modelled risk to that
exposure (the event loss tables), and (3) a set of risk transfer
contracts (the layers).
The exposure is represented by a table, one row per
building covered, that lists the building’s location, construction
details, primary insurance coverage, and replacement value.
The modelled risk is represented by an event loss table (ELT).
This table lists for each of a large set of possible catastrophic
events the expected loss that would occur to the exposure
should the event occur. Finally, each layer (risk transfer
contract) is described by a set of financial terms that includes
aggregate deductibles and limits (i.e., deductibles and maximal
payouts to be applied to the sum of losses over the year)
and per-occurrence deductibles and limits (i.e., deductibles and
maximal payouts to be applied to each loss in a year), plus
other financial terms.
Consider, for example, a Japanese earthquake program. The
exposure might list 2 million buildings (e.g., single-family
homes, small commercial buildings, and apartments) and, for
each, its location (e.g., latitude and longitude), constructions
details (e.g., height, material, roof shape, etc.), primary insurance terms (e.g., deductibles and limits), and replacement
value. The event loss table might, for each of 100,000 possible
earthquake events in Japan, give the sum of the losses expected
to the associated exposure should the associated event occur.
Note that ELTs are the output of stochastic region peril models
[17] and typically also include some additional financial terms.
Finally, a risk transfer contract may consist of two layers as
shown in Figure 1. The first layer is a per-occurrence layer that
pays out a 60% share of losses between 160 million and 210
million associated with a single catastrophic event. The second
layer is an aggregate layer covering 30% of losses between
40 million and 90 million that accumulate due to earthquake
activity over the course of a year.
Given a reinsurance company’s portfolio described in terms
of exposure, event loss tables, and layers, the most fundamental
type of analysis query computes an Exceedance Probability
(EP) curve, which represents, for each of a set of user-specified
loss values, the probability that the total claims a reinsurer
will have to pay out exceeds this value. Not surprisingly
there is no computationally feasible closed-form expression
for computing such an EP curve over hundreds of thousands
of events and millions of individual exposures. Consequently
a simulation approach must be taken. The idea is to perform a
stochastic simulation based on a year event table (YET). This
table describes a large number of trials, each representing one
possible sequence of catastrophic events that might occur in
a given year. This YET is generated by an event simulator
that uses the expected occurrence rate of each event plus other
Fig. 1.
An example two-layer reinsurance program.
hazard information like seasonality. The process to generate
the YET is beyond the scope of this paper, and we focus on
the computationally intensive task of computing the expected
loss distribution (i.e., EP curve) for a given portfolio, given
a particular YET. Given the sequence of events in a given
trial, the loss for this particular trial can be computed, and the
overall loss distribution is obtained from the losses computed
for the set of trials.
While computing the EP curve for the company’s entire
portfolio is critical in assessing a company’s solvency, analysts
are often interested in digging deeper into the data and posing a
wide variety of queries with the goal of analysing such things
as cash flow throughout the year, diversity of the portfolio,
financial impact of adding a new contract or contracts to the
portfolio, and many others.
The following is a representative, but far from complete,
set of example queries. Note that while all of them involve
some aspects of the basic aggregate risk analysis algorithm
used to compute Exceedance Probability curves, each puts its
own twist on the computation.
EP Curves with secondary uncertainty: In the basic aggregate risk analysis algorithm, the loss value for each event is
represented by a mean value. This is an oversimplification
because for any given event there are a multitude of possible
loss outcomes. This means that each event has an associated
probability distribution of loss values rather than a single
associate loss value. Secondary uncertainty arises from the
fact that we are not just unsure whether an event will occur
(primary uncertainty) but also about many of the exposure
and hazard parameters and their interactions.
Performing aggregate risk analysis accounting for secondary uncertainty (i.e., working with the loss distributions
associated with events rather than just mean values) is computationally intensive due to the statistical tools employed—
for example, the beta probability distribution is employed in
estimating the loss using the inverse beta cumulative density
function [16]—but is essential in many applications.
Return period losses (RPL) by line of business (LOB),
class of business (COB) or type of participation (TOP):
In the reinsurance industry, a layer defines coverage on
different types of exposures and the type of participation.
Exposures can be classified by class of business (COB) or
line of business (LOB) (e.g., marine, property or engineering coverage). The way in which the contractual coverage
participates when a catastrophic event occurs is defined by
the type of participation (TOP). Decision makers may want
to know the loss distribution of a specific layer type in
their portfolios, which requires the analysis to be restricted
to layers covering a particular LOB, COB or TOP.
Region/peril losses: This type of query calculates the expected losses or a loss distribution for a set of geographic
regions (e.g., Florida or Japan), a set of perils (e.g., hurricane or earthquake) or a combination of region and peril.
This allows the reinsurer to understand both what types
of catastrophes provide the most risk to their portfolio
and in which regions of the globe they are most heavily
exposed to these risks. This type of analysis helps the
reinsurer to diversify or maintain a persistent portfolio in
either dimension or both.
Multi-marginal analysis: Given the current portfolio and a
small set of potential new contracts, a reinsurer will have to
decide which contracts to add to the portfolio. Adding a new
contract means additional cash flow but also increases the
exposure to risk. To help with the decision which contracts
to add, multi-marginal analysis calculates the difference
between the loss distributions for the current portfolio and
for the portfolio with any subset of these new contracts
added. This allows the insurer to choose contracts or to
price the contracts so as to obtain the “right” combination
of added cash flow and added risk.
Stochastic exceedance probability (STEP) analysis:
This analysis is a stochastic approach to the weighted
convolution of multiple loss distributions. After the
occurrence of a natural disaster not in their event
catalogue, catastrophe modelling [17] vendors attempt to
estimate the distribution of possible loss outcomes for that
event. One way of doing this is to find similar events in
existing stochastic event catalogues and propose a weighted
combination of the distributions of several events that
best represents the actual occurrence. A simulation-based
approach allows for the simplest method of producing this
combined distribution. To perform this type of analysis,
a customized Year Event Table must be produced from
the selected events and their weights. In this YET, each
trial contains only one event, chosen with a probability
proportional to its weight. The final result is a loss
distribution of the event, including various statistics such
as mean, variance and quantiles.
Periodic Loss Distribution: Many natural catastrophes have
a seasonal component to them, that is, do not occur
uniformly throughout the year. For example, hurricanes
on the Atlantic coast occur between July and November.
Flooding in equatorial regions occurs in the rain season.
As a result, the reinsurer may be interested in how their
potential losses fluctuate throughout the year, for example
to reduce their exposure through reduced contracts or increased retrocessional coverage during riskier periods. To
aid in these decisions, a periodic loss distribution represents
the loss distribution for different periods of the year, such
as quarters, months, weeks, etc.
III.
Q U PARA F RAMEWORK
In this section, we present our QuRARA framework. Before describing QuPARA, we give an overview of the steps
involved in answering an aggregate query sequentially. This
will be helpful in understanding the parallel evaluation of
aggregate queries using QuPARA.
The loss distribution is computed from the portfolio and
the YET in two phases. The first phase computes a year loss
table (YLT). For each trial in the YET, each event in this trial,
and each ELT that includes this event, the YLT contains a
tuple htrial, event, ELT, lossi recording the loss incurred by
this event, given the layer’s financial terms and the sequence
of events in the trial up to the current event. The second phase
then aggregates the entries in the YLT to compute the final loss
distribution. Algorithm 1 shows the sequential computation of
the YLT.
Algorithm 1: Sequential Aggregate Risk Analysis
Input: Portfolio and YET
Output: YLT
1
2
3
4
5
6
7
8
9
10
11
12
for each trial T in the YET do
for each event X in T do
for each program P in the portfolio do
for each layer L in P do
for each ELT E covered by L do
Lookup X in E to determine the
loss lX associated with X
lL ← lL + lX
Apply per-occurrence and aggregate
financial terms to lL
lP ← lP + lL
lP F ← lP F + lP
lT ← lT + lP F
Populate Y LT
After looking up the losses incurred by a given event in the
ELTs of a given layer and summing these losses, the resulting
layer loss lL is reduced by applying the layer’s per-occurrence
and aggregate financial terms. The remaining loss is added to
the program loss lP for this event, which in turn is added tho
the portfolio loss lP F for this event. Finally, the loss of an
entire trial, lT is computed by summing the portfolio losses
for all events in the trial. Depending on the level of detail of
the final analysis, the YLT is populated with the loss values
at different levels of aggregation. At one extreme, if only the
loss distribution for the entire portfolio is of interest, there is
one loss value per trial. At the other extreme, if the filtering of
losses based on some combination of region, peril, LOB, etc.
is required, the YLT contains one loss value for each htrial,
program, layer, ELT, eventi tuple.
In order to answer ad hoc aggregate queries efficiently,
our QuPARA framework provides a parallel implementation of
such queries using the MapReduce programming model. The
computation of the YLT is the responsibility of the mapper,
while the computation of the final loss distribution(s) is done
by the reducer. Next we describe the components of QuPARA.
Fig. 2.
The Query-Driven Portfolio Aggregate Risk Analysis (QuPARA) Framework
A. Components
Figure 2 visualizes the design of QuPARA. The framework
is split into a front-end offering a query interface to the user,
and a back-end consisting of a distributed file system and a
query engine that executes the query using MapReduce. As
already stated, the user poses a query to the query interface
using an SQL-like syntax and does not need to have any
knowledge of the implementation of the back-end.
The query interface offers a web-based portal through
which the user can issue ad hoc queries in an SQL-like syntax.
The query is then parsed and passed to the query engine.
The distributed file system, implemented using Hadoop’s
HDFS, stores the raw portfolio data, as well as the YET in
tabular form.
The query engine employs the MapReduce framework
to evaluate the query using a single round consisting of a
map/combine step and a reduce step. During the map step,
the engine uses one mapper per trial in the YET, in order to
construct a YLT from the YET, equivalent to the sequential
construction of the YLT from the YET using Algorithm 1.
The combiner and reducer collaborate to aggregate the loss
information in the YLT into the final loss distribution for
the query. There is one combiner per mapper. The combiner
pre-aggregates the loss information produced by this mapper,
in order to reduce the amount of data to be sent across
the network to the reducer(s) during the shuffle step. The
reducer(s) then carry out the final aggregation. In most queries,
which require only a single loss distribution as output, there is
a single reducer. Multi-marginal analysis is an example where
multiple loss distributions are to be computed, one per subset
of the potential contracts to be added to the portfolio. In this
case, we have one reducer for each such subset, and each
reducer produces the loss distribution for its corresponding
subset of contracts.
Each mapper retrieves the set of ELTs required for the
query from the distributed file system using a layer filter and
an ELT filter. Specifically, the query may specify a subset of
the layers in the portfolio to be the subject of the analysis.
The layer filter retrieves the identifiers of the ELTs contained
in these layers from the layer table. If the query specifies, for
example, that the analysis should be restricted to a particular
type of peril, the ELT filter then extracts from this set of
ELTs the subset of ELTs corresponding to the specified type
of peril. Given this set of ELT identifiers, the mapper retrieves
the actual ELTs and, in memory, constructs a combined ELT
associating a loss with each hevent, ELTi pair. It then iterates
over the sequence of events in its trial, looks up the ELTs
recording non-zero losses for each event, and generates the
corresponding htrial, event, ELT, lossi tuple in the YLT, taking
each ELT’s financial terms into account.
The aggregation to be done by the combiner depends on
the query. In the simplest case, a single loss distribution for
the selected set of layers and ELTs is to be computed. In
this case, the combiner sums the loss values in the YLT
output by the mapper and sends this aggregate value to the
reducer. A more complicated example is the computation of
a weekly loss distribution. In this case, the combiner would
aggregate the losses corresponding to the events in each week
and send each aggregate to a different reducer. Each reducer
is then responsible for computing the loss distribution for one
particular week.
Algorithm 2: Mapper in parallel aggregate risk analysis
Input: hT, E := {E1 , E2 , · · · , Em }i, where m is the
number of events in a trial
Output: A list of entries hT, Ei , lP F i of the YLT
1
2
3
4
5
6
7
8
9
The following is a more detailed description of the mapper,
combiner, and reducer used to implement parallel aggregate
risk analysis in MapReduce.
1) Mapper: The mapper, shown in Algorithm 2, takes as
input an entire trial and the list of its events, represented as
a pair hT, E := {E1 , E2 , . . . , Em }i. Algorithm 2 does not
show the construction of the combined ELT (CELT) and layer
list (LLT) performed by the mapper before carrying out the
steps in lines 1–9. The loss estimate of an event in a portfolio
is computed by scanning through every layer L in the LLT,
retrieving and summing the loss estimates for all ELTs covered
by this layer, and finally applying the layer’s financial terms.
2) Combiner: The combiner, shown in Algorithm 3, receives as input the list of triples hT , Ei , lP F i generated by
a single mapper, that is, the list of loss values for the events
in one specific trial. The combiner groups these loss values
according to user-specified grouping criteria and outputs one
aggregate loss value per group.
3) Reducer: The reducer, shown in Algorithm 4, receives
as input the loss values for one specific group and for all trials
in the YET. The reducer then aggregates these loss values into
the loss statistic requested by the user. For example, to generate
an exceedance probability curve, the reducer sorts the received
loss values in increasing order and, for each loss value v in a
user-specified set of loss values, reports the percentage of trials
Apply L’s financial terms to lL
lP F ← lP F + lL
Emit(hT , Ei , lP F i)
Algorithm 3: Combiner in parallel aggregate risk analysis
Input: A list of YLT entries hT , Ei , lP F i for the
events in a given trial T .
Output: A list of aggregate YLT entries hGi , T , lG i
with key Gi for the event groups in trial T
1
2
3
A reducer, finally, receives one loss value per trial. It sorts
these loss values in increasing order and uses this sorted list
to generate a loss distribution.
for each event, Ei in E do
Look up Ei in the CELT and find corresponding
1
2
n
losses, lEi = {lE
, lE
, · · · , lE
}, where ELT1 , ELT2 ,
i
i
i
. . . , ELTn are the ELTS in the CELT
for each layer, L, in the LLT do
for each ELT ELTj covered by L do
j
Lookup lE
in lEi
i
j
lL ← lL + lEi
4
5
Join input tuples with event catalogue to annotate
events with their attributes (region, peril, etc.)
Group events in the input list by event features
according to the user’s query
for each group Gi do
lGi ← sum of the loss values associated with the
events in Gi in trial T
Emit(hT , Gi , LG i)
with a loss value greater than v as the probability of incurring
a loss higher than v.
Algorithm 4: Reducer in parallel aggregate risk analysis
Input: A list of loss tuples hGi , T , lP F i for an event
group Gi .
Output: Loss statistics for event group Gi based on
user’s query
1
Based on user query, generate:
(i) Group loss distribution, or
(ii) Group loss statistics, or
(iii) Group value-at-risk (VaR) and/or Tail value-at-risk
(TVaR), or
(iv) Exceedance probability curves
B. Data Organization
The data used by QuPARA is stored in a number of tables:
• The year event table YET contains tuples htrial_ID,
event_ID, time_Index, z_PEi. trial_ID is a
unique identifier associated with each of the one million
trials in the simulation. event_ID is a unique identifier associated with each event in the event catalogue.
•
•
•
•
•
•
•
•
time_Index determines the position of the occurrence
of the event in the sequence of events in the trial. z_PE
is a random number specific to the program and event occurrence. Each event occurrence across different programs
has a different associated random number.
The layer table LT contains tuples hlayer_ID, cob,
lob, tob, elt_IDsi. layer_ID is a unique identifier associated with each layer in the portfolio. cob is
an industry classification according to perils insured and
the related exposure and groups homogeneous risks. lob
defines a set of one or more related products or services
where a business generates revenue. tob describes how
reinsurance coverage and premium payments are calculated. elt_IDs is a list of event loss table IDs that are
covered by the layer.
The layer list table LLT contains tuples hlayer_ID,
occ_Ret, occ_Lim, agg_Ret, agg_Limi. Each
entry is a simplified representation of the layer identified
by layer_ID. occ_Ret is the occurrence retention or
deductible of the insured for an individual occurrence
loss. occ_Lim is the occurrence limit or coverage the
insurer will pay for occurrence losses in excess of the
occurrence retention. agg_Ret is the aggregate retention
or deductible of the insured for an annual cumulative loss.
agg_Lim is the aggregate limit or coverage the insurer
will pay for annual cumulative losses in excess of the
aggregate retention.
The event loss table pool ELTP contains tuples helt_ID,
region, perili. Each such entry associates a particular type of peril and a particular region with the ELT with
ID elt_ID.
The extended event loss table EELT contains tuples
hevent_ID, z_E, mean_Loss, sigma_I,
sigma_C, max_Lossi. event_ID is the unique
identifier of an event in the event catalogue. z_E is
a random number specific to the event occurrence.
Event occurrences across different programs have the
same random number. mean_Loss denotes the mean
loss incurred if the event occurs. max_Loss is the
maximum expected loss incurred if the event occurs.
sigma_I represents the variance of the loss distribution
for this event. sigma_C represents the error of the
event-occurrence dependencies.
The combined event loss table CELT is not stored on
disk but is constructed by each mapper in memory from
the extended event loss tables corresponding to the user’s
query. It associates with each event ID event_ID a list
of tuples helt_ID, z_E, mean_Loss, sigma_I,
sigma_C, max_Lossi, which is the loss information
for event event_ID stored in the extended ELT elt_ID.
The year event loss table YELT is an intermediate table produced by the mapper for consumption by the
combiner. It contains tuples htrial_ID, event_ID,
time_Index, estimated_Lossi.
The event catalogue ECT contains tuples hevent_ID,
region, perili associating a region and a type of peril
with each event.
The year region peril loss table YRPLT contains tuples htrial_ID, time_Index, region, peril,
estimated_Lossi, listing for each trial the estimated
loss at a given time, in a given region, and due to a
particular type of peril. This is yet another intermediate
table, which is produced by the combiner for consumption
by the reducer.
C. Data Filters
QuPARA incorporates three types of data filters that allow
the user to focus their queries on specific geographic regions,
types of peril, etc. These filters select the appropriate entries
from the data tables they operate on, for further processing my
the mapper, combiner, and reducer.
• The layer filter extracts the set of layers from the layer
table LT and passes this list of layers to the mapper as the
“portfolio” to be analyzed by the query. The list of selected
layers is also passed to the ELT filter for selection of the
relevant ELTs.
• The ELT filter is used to select, from the ELT pool ELTP,
the set of ELTs required by the layer filter and for building
the combined ELT CELT.
• The event filter selects the features of events from the
event catalogue ECT to provide the grouping features to
the combiner.
IV.
A N E XAMPLE Q UERY
An example of an ad hoc request on QuPARA is to generate
a report on seasonal loss value-at-risk (VaR) with confidence
level of 99% due to hurricanes and floods that affects all commercial properties in different locations in Florida. The user
poses the query through the query interface, which translates
the query into a set of five SQL-like queries to be processed
by the query engine:
Q1 : The first part of processing any user query is the query
to be processed by the layer filter. In this case, we are
interested in all layers covering commercial properties,
which translates into the following SQL query:
SELECT * FROM LT
WHERE lob IN commercial
Q2 : The second query is processed by the ELT filter to extract
the ELTs relevant for the analysis. In this case, we are
interested in all ELTs covered by layers returned by
query Q1 and which cover Florida (FL) as the region
and hurricanes (HU) and floods (FLD) as perils:
SELECT elt_ID FROM ELTP
WHERE elt_ID IN Q1
AND region IN FL
AND peril IN HU, FLD
Q3 : This query is provided to the event filter for retrieving
event features required for grouping estimated losses in
the YELT.
SELECT event_ID AND region FROM YELT
Q4 : This query is provided to the combiner for grouping all
events in a trial based on their order of occurrence. For
example, if there are 100 events equally distributed in a
year and need to be grouped based on the four seasons
in a year, then the estimated loss for each season is the
sum of 25 events that occur in that season.
SELECT trial_ID,
SEASON_SUM(estimated_Loss)
FROM YELT
GROUP BY time_Index
Q5 : This query is provided to the reducer to define the final
output of the user request. The seasonal loss Value-at-Risk
(VaR) with 99% confidence level is estimated.
SELECT VaR IN 0.01 FROM YRPLT
V.
I MPLEMENTATION
For our implementation of QuPARA, we used Apache
Hadoop, an open-source software framework that implements
the MapReduce programming model [9], [10], [18]. We chose
Hadoop because other available frameworks [20], [21] require
the use of additional interfaces, commercial or web-based, for
deploying an application.
A number of interfaces provided by Hadoop, such as the
InputFormat and the OutputFormat are implemented as
classes. The Hadoop framework works in the following way
for a MapReduce round. The data files are stored on the
Hadoop Distributed File System (HDFS) [19] and are loaded
by the mapper. Hadoop provides an InputFormat interface,
which specifies how the input data is to be split for processing
by the mapper. HDFS provides an additional functionality,
called distributed cache, for distributing small data files that
are shared by the nodes of the cluster. The distributed cache
provides local access to the shared data. The Mapper interface
receives the partitioned data and emits intermediate key-value
pairs. The Partitioner interface receives the intermediate
key-value pairs and controls the partitioning of these keys for
the Reducer interface. Then the Reducer interface receives
the partitioned intermediate key-value pairs and generates the
final output of this MapReduce round. The output is received
by the OutputFormat interface and provides it back to
HDFS.
The input data for a MapReduce round in QuPARA is
the year event table YET, the event loss table pool ELTP,
the list of layers LT, and the event catalogue ECT, which are
stored on HDFS. The master node executes Algorithm 5 and
requires access to the ELT pool and the portfolio before the
MapReduce round. Firstly, the master node uses the portfolio
to decompose the aggregate analysis job into a set of sub-jobs
{J1 , · · · , Jx } as shown in line 1, each covering 200 layers.
This partition into sub-jobs was chose to ensure a balance
between Input/Output cost and the overhead for (sequentially)
constructing the combined ELT in each mapper; more sub-jobs
means reading the YET more often, whereas fewer and larger
jobs increase the overhead of the sequential combined ELT
construction. The layers and the corresponding ELTs required
by each sub-job are then submitted to the Hadoop job scheduler
in lines 2–6. The MapReduce round is illustrated in Figure 3.
The InputFormat interface splits the YET based on
the number of mappers specified for the MapReduce round.
By default, Hadoop splits large files based on HDFS block
size, but in this paper, the InputFormat is redefined to
split files based on the number of mappers. The mappers are
configured so that they receive all the ELTs covered by the
entire portfolio via the distributed cache. Each mapper then
constructs its own copy of the combined ELT form the ELTs
in the distributed cache. Using a combined ELT speeds up
the lookup of events in the relevant ELTs in the subsequent
processing of the events in the YET. If individual ELTs were
employed, then one lookup would be required for fetching the
Fig. 3.
Apache Hadoop for MapReduce in the QuPARA Framework
Algorithm 5: Algorithm for master node in aggregate
risk analysis
Input: ELT pool, portfolio
1
2
3
4
5
6
Split the portfolio into jobs J1 , · · · , Jx
for each Job Ji do
Distribute its layers to nodes via the distributed
cache
for each layer Lj in job Ji do
Distribute the ELTs covered by Lj to nodes via
the distributed cache
Submit Ji to Hadoop job scheduler
loss value of an event from each ELT. Using a combined ELT,
only a single lookup is required. The mapper now implements
Algorithm 2 to compute the loss information for all events in
its assigned portion of the YET.
The combiner implements Algorithm 3 to group the eventloss pairs received from the mapper and emits the group-loss
pairs to the partitioner. The event catalogue is contained in the
distributed cache of the combiner and is used by the combiner
to implement the grouping of events. The Combiner delivers
the grouped loss pairs to the partitioner. The partitioner ensures
that all the loss pairs with the same group key are delivered
to the same reducer.
The reducer implements Algorithm 4 to collect the sorted
group-loss pairs and produces year region peril loss table
YRPLT. Based on the query, the OutputFormat generates
reports which are then saved to the HDFS.
The layer filter, ELT filter, and event filter, described
earlier, are implemented using Apache Hive [14], which is built
on top of the Hadoop Distributed File System and supports
data summarization and ad hoc queries using an SQL-like
language.
VI.
P ERFORMANCE E VALUATION
In this section, we discuss our experimental setup for
evaluating the performance of QuPARA and the results we
obtained.
Fig. 4. Running time of QuPARA on an increasing number of nodes with a
fixed number of layers per job and one job per node
A. Platform
We evaluated QuPARA on the Glooscap cluster of the
Atlantic Computational Excellence Network (ACEnet). We
used 16 nodes of of the cluster, each of which was an SGI
C2112-4G3 with four quad-core AMD Opteron 8384 (2.7
GHz) processors and 64 GB RAM per node. The nodes were
connected via Gigabit Ethernet. The operating system on each
node was Red Hat Enterprise Linux 4.8. The global storage
capacity of the cluster was 306 TB of Sun SAM-QFS, a
hierarchical storage system using RAID 5. Each node had
500GB of local storage. The Java Virtual Machine (JVM)
version was 1.6. The Apache Hadoop version was 1.0.4. The
Apache Hive version was 0.10.0.
B. Results
Figure 4 shows the total time taken in seconds for performing aggregate risk analysis on our experimental platform.
Up to 16 nodes were used in the experiment. Each node
processed one job with 200 layers, each covering 5 unique
ELTs. Thus, up to 16,000 ELTs were considered. The YET
in our experiments contained 1,000,000 trials, each consisting
of 1,000 events. The graph shows a very slow increase in
the total running time, in spite of the constant amount of
computation to be performed by each node (because every
node processes the same number of layers, ELTs, and YET
entries). The gradual increase in the running time is due to
the increase in the setup time required by the Hadoop job
scheduler and the increase in network traffic (reflected in an
increase in the time taken by the reducer). Nevertheless, this
scheduling and network traffic overhead amounted to only
1.67%–7.13% of the total computation time. Overall, this
experiment demonstrates that, if the hardware scales with the
input size, QuPARA’s processing time of a query remains
nearly constant.
Figure 5 shows the increase in running time when the
number of layers is increased from 200 to 3200 while keeping
the number of nodes fixed. Once again, each layer covered
5 ELTs, the YET contains 1,000,000 trials, each consisting
of 1,000 events. 16 nodes were used in this experiment. As
expected, the running time of QuPARA increases linearly with
the input size as we increase the number of layers. With the
Fig. 5. Running time of QuPARA on an increasing number of layers with a
fixed number of nodes
increase in the number of layers, the time taken for setup,
I/O time, and the time for all numerical computations scale
in a linear fashion. The time taken for building the combined
ELT, by the reducer, and for clean up are a constant. For 200
Layers, only 30% of the time is taken for computations, and
the remaining time accounts for system and I/O overheads. For
3200 layers, on the other hand, more than 80% of the time is
spent on computations.
Figure 6 shows the decrease in running time when the
number of nodes is increased but the input size is kept fixed at
3,200 layers divided into 16 jobs of 200 layers. Figure 7 shows
the relative speed-up achieved in this experiment, that is, the
ratio between the running time achieved on a single node and
the running time achieved on up to 16 nodes. Up to 4 nodes, the
speed-up is almost linear. Beyond 4 nodes, the speed-up starts
to decrease. This is not due to an increase in overhead, but
since the computation times have significantly reduced after 4
nodes, the overheads start to take away a greater percent of
the total time.
In summary, our experiments show that QuPARA is capable of performing ad-hoc aggregate risk analysis queries on
industry-size data sets in a matter of minutes. Thus, it is a
viable tool for analysts to carry out such analyses interactively.
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
Fig. 6. Running time of QuPARA on a fixed input size using between 1 and
16 nodes
[7]
[8]
[9]
[10]
[11]
[12]
[13]
Fig. 7.
Speed-up achieved on the QuPARA framework
VII.
C ONCLUSIONS
Typical production systems performing aggregate risk analysis used in the industry are efficient for generating a small
set of key portfolio metrics such as PML or TVAR required
by rating agencies and regulatory bodies and essential for
decision making. However, production systems can neither
accommodate nor solve ad hoc queries that provide a better
view of the multiple dimensions of risk that can impact a
portfolio.
The research presented in this paper proposes a framework
for portfolio risk analysis capable of expressing and solving a
variety of ad hoc catastrophic risk queries, based on MapReduce, and provided a prototype implementation, QuPARA,
using the Apache Hadoop implementation of the MapReduce
programming model and Apache Hive for expressing ad hoc
queries in an SQL-like language. Our experiments demonstrate
the feasibility of answering ad hoc queries on industry-size
data sets efficiently using QuPARA. As an example, a portfolio
analysis on 3,200 layers and using a YET with 1,000,000 trials
and 1,000 events per trial took less than 20 minutes.
Future work will aim to develop a distributed system by
providing an online interface to QuPARA which can support
multiple user queries.
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
R. R. Anderson and W. Dong, “Pricing Catastrophe Reinsurance with
Reinstatement Provisions using a Catastrophe Model,” Casualty Actuarial Society Forum, Summer 1998, pp. 303-322.
G. G. Meyers, F. L. Klinker and D. A. Lalonde, “The Aggregation
and Correlation of Reinsurance Exposure,” Casualty Actuarial Society
Forum, Spring 2003, pp. 69-152.
W. Dong, H. Shah and F. Wong, “A Rational Approach to Pricing of
Catastrophe Insurance,” Journal of Risk and Uncertainty, Vol. 12, 1996,
pp. 201-218.
R. M. Berens, “Reinsurance Contracts with a Multi-Year Aggregate
Limit,” Casualty Actuarial Society Forum, Spring 1997, pp. 289-308.
G. Woo, “Natural Catastrophe Probable Maximum Loss,” British Actuarial Journal, Vol. 8, 2002.
M. E. Wilkinson, “Estimating Probable Maximum Loss with Order
Statistics, in Casualty Actuarial Society Forum, 1982, pp. 195209.
A. A. Gaivoronski and G. Pflug, “Value-at-Risk in Portfolio Optimization: Properties and Computational Approach,” Journal of Risk, Vol. 7,
No. 2, Winter 2004-05, pp. 131.
P. Glasserman, P. Heidelberger, and P. Shahabuddin, “Portfolio Valueat-Risk with Heavy-Tailed Risk Factors,” Mathematical Finance, Vol.
12, No. 3, 2002, pp. 239269.
T. White, “Hadoop: The Definitive Guide,” 1st Edition, O’Reilly Media,
2009.
Apache Hadoop Project website: http://hadoop.apache.org/ [Last accessed: 25 May, 2013].
J. Dean and S. Ghemawat, “MapReduce: Simplified Data Processing on
Large Clusters,” Communications of the ACM, Vol. 51, Issue 1, 2008,
pp. 107–113.
T. Condie, N. Conway, P. Alvaro, J. M. Hellerstein, K. Elmeleegy and R. Sears, “Mapreduce Online,” EECS Department, University of California, Berkeley, Technical Report No. UCB/EECS-2009136, October 2009. http://www.eecs.berkeley.edu/Pubs/TechRpts/2009/
EECS-2009-136.pdf [Last accessed: 25 May, 2013].
K. -H. Lee, Y. -J. Lee, H. Choi, Y. D. Chung and B. Moon, “Parallel
Data Processing with MapReduce: A Survey,” SIGMOD Record Vol.
40, No. 4, 2011, pp. 11–20.
HiveQL website: http://hive.apache.org/ [Last accessed: 25 May, 2013].
E. Capriolo, D. Wampler and J. Rutherglen, “Programming Hive,”
O’Reilly Media, 1st Edition, 2012.
A. Rau-Chaplin, B. Varghese and Z. Yao, “A MapReduce Framework
for Analysing Portfolios of Catastrophic Risk with Secondary Uncertainty,” accepted for publication in the Workshop Proceedings of the
International Conference on Computational Science, 2013.
P. Grossi and H. Kunreuter, “Catastrophe Modelling: A New Approach
to Managing Risk,” Springer, 2005.
K. Shvachko, K. Hairong, S. Radia, R. Chansler, “The hadoop Distributed File System,” Proceedings of the 26th IEEE Symposium on
Mass Storage Systems and Technologies, 2010, pp. 1–10.
Hadoop Distributed File System website: http://hadoop.apache.org/docs/
r1.0.4/hdfs design.htmlS [Last accessed: 25 May, 2013]
Amazon Elastic MapReduce (Amazon EMR) website: http://aws.
amazon.com/elasticmapreduce/ [Last accessed: 25 May, 2013].
Google MapReduce website: https://developers.google.com/appengine/
docs/python/dataprocessing/overview [Last accessed: 25 May, 2013].
C. Eaton, D. Deroos, T. Deutsch, G. Lapis and P. Zikopoulos, “Understanding Big Data: Analytics for Enterprise Class Hadoop and
Streaming Data,” McGraw Hill, 2012.
| 5cs.CE
|
AD in Fortran
Part 1: Design
arXiv:1203.1448v2 [cs.PL] 8 Mar 2012
Alexey Radul, Barak A. Pearlmutter, and Jeffrey Mark Siskind
Abstract We propose extensions to F ORTRAN which integrate forward and reverse
Automatic Differentiation (AD) directly into the programming model. Irrespective
of implementation technology, embedding AD constructs directly into the language
extends the reach and convenience of AD while allowing abstraction of concepts
of interest to scientific-computing practice, such as root finding, optimization, and
finding equilibria of continuous games. Multiple different subprograms for these
tasks can share common interfaces, regardless of whether and how they use AD internally. A programmer can maximize a function F by calling a library maximizer,
XSTAR=ARGMAX(F, X0), which internally constructs derivatives of F by AD, without having to learn how to use any particular AD tool. We illustrate the utility of
these extensions by example: programs become much more concise and closer to
traditional mathematical notation. A companion paper describes how these extensions can be implemented by a program that generates input to existing F ORTRANbased AD tools.
Key words: Nesting, multiple transformation, forward mode, reverse mode, TAPE NADE , A DIFOR, programming-language design
1 Introduction
The invention of F ORTRAN was a major advance for numeric computing, allowing
Alexey Radul
Hamilton Institute, National University of Ireland Maynooth, Ireland, [email protected]
Barak A. Pearlmutter
Department and Computer Science and Hamilton Institute, National University of Ireland
Maynooth, Ireland, [email protected]
Jeffrey Mark Siskind
Electrical and Computer Engineering, Purdue University, IN, USA, [email protected]
1
2
Alexey Radul, Barak A. Pearlmutter, and Jeffrey Mark Siskind
g(x; α , β ) =
β α α −1 −β x
e
x
Γ (α )
✞to be transcribed into a natural but unambiguous notation
FUNCTION G(X,ALPHA,BETA)
G=BETA**ALPHA/GAMMA(ALPHA)*X**(ALPHA-1)*EXP(-BETA*X)
END
which could be automatically translated into an executable program. However, transcribing
xi+1 = xi − f (xi )/ f ′ (xi )
✞to F ORTRAN in
FUNCTION RAPHSN(F, FPRIME, X0, N)
EXTERNAL F, FPRIME
X = X0
DO 1690 I=1,N
1690 X = X-F(X)/FPRIME(X)
RAPHSN = X
END
requires that the caller provide both F and FPRIME. Manually coding the latter from
the former is, in most cases, a mechanical process, but tedious and error prone.
This problem has traditionally been addressed by arranging for an AD preprocessor to produce FPRIME (Speelpenning, 1980; Wengert, 1964). That breakthrough
technology not only relieves the programmer of the burden of mechanical coding of
derivative-calculation codes, it also allows the derivative code to be updated automatically, ensuring consistency and correctness. However, this caller derives discipline has several practical difficulties. First, the user must learn how to use the AD
preprocessor, which constitutes a surprisingly serious barrier to adoption. Second,
it makes it very difficult to experiment with the use of different sorts of derivatives
(e.g., adding a Hessian-vector product step in an optimization) in such called subprograms, or to experiment with different AD preprocessors. Third, although preprocessors might be able to process code which has already been processed in order to
implement nested derivatives, the maneuvers required by current tools can be somewhat arcane (Siskind and Pearlmutter, 2008a). Fourth, software engineering principles of locality and atomicity are being violated: knowledge of what derivatives are
needed is distributed in a number of locations which must be kept consistent; and
redundant information, which must also be kept consistent, is being passed, often
down a long call chain. We attempt to solve these problems, making the use of AD
more concise, convenient, and intuitive to the scientific programmer, while keeping to the spirit of F ORTRAN. This is done using the Forward And Reverse Fortran
Extension Language or FARFEL, a small set of extensions to F ORTRAN, in concert
with an implementation strategy which leverages existing F ORTRAN compilers and
AD preprocessors (Bischof et al., 1992; Hascoët and Pascual, 2004).
The remainder of the paper is organized as follows: Section 2 describes FARFEL.
Section 3 describes a concrete example FARFEL program to both motivate and illuminate the proposed extensions. Section 4 situates this work in its broader context.
AD in Fortran: Design
3
Section 5 summarizes this work’s contributions. A companion paper (Radul et al.,
2012) describes how FARFEL can be implemented by generating input to existing
AD tools.
2 Language Extensions
FARFEL consists of two principal extensions to F ORTRAN: syntax for AD and for
nested subprograms. We currently support only F ORTRAN77, but there is no barrier,
in principle, to adding FARFEL to more recent dialects.
Extension 1: AD Syntax
Traditional mathematical notation allows one to specify
φ′ =
d 1
1 x − x̄ 2
√
exp −
dσ
2 σ
2πσ 2
✞By analogy, we extend F ORTRAN to encode this as
ADF(TANGENT(SIGMA) = 1)
PHI = 1/SQRT(2*PI*SIGMA**2)*EXP(-0.5*((X-XBAR)/SIGMA)**2)
END ADF(PHIPRM = TANGENT(PHI))
which computes the derivative PHIPRM of PHI with respect to SIGMA by forward AD.
For syntactic details see companion paper (Radul et al., 2012).
An analogous FARFEL construct supports computing the same derivative with
reverse
AD:
✞
ADR(COTANGENT(PHI) = 1)
PHI = 1/SQRT(2*PI*SIGMA**2)*EXP(-0.5*((X-XBAR)/SIGMA)**2)
END ADR(PHIPRM = COTANGENT(SIGMA))
Note that with the ADR construct, the dependent variable appears at the beginning of
the block and the independent variable at the end—the variables and assignments in
the opening and closing statements specify the desired inputs to and outputs from
the reverse phase, whereas the statements inside the block give the forward phase.
These constructs allow not just convenient expression of AD, but also modularity
and encapsulation of code which employs AD. For instance, we can write a general
✞scalar-derivative subprogram DERIV1 at user level
FUNCTION DERIV1(F, X)
EXTERNAL F
ADF(X)
Y = F(X)
END ADF(DERIV1 = TANGENT(Y))
END
✞which could be used in, for example,
FUNCTION PHI(SIGMA)
PHI = 1/SQRT(2*PI*SIGMA**2)*EXP(-0.5*((X-XBAR)/SIGMA)**2)
END
4
Alexey Radul, Barak A. Pearlmutter, and Jeffrey Mark Siskind
PHIPRM = DERIV1(PHI, SIGMA)
✞DERIV1 can be changed to use reverse AD without changing its API:
FUNCTION DERIV1(F, X)
EXTERNAL F
ADR(Y)
Y = F(X)
END ADR(DERIV1 = COTANGENT(X))
END
allowing codes written with DERIV1 to readily switch between using forward and
reverse AD.
To take a more elaborate example, we can write a general gradient calculation
GRAD using repeated forward AD:
✞
SUBROUTINE GRAD(F, X, N, DX)
EXTERNAL F
DO 1492 I=1,N
ADF(TANGENT(X(J)) = 1-MIN0(IABS(I-J),1), J=1,N)
Y = F(X)
1492 END ADF(DX(I) = TANGENT(Y))
END
(Note that the ADF and ADR constructs support implied-DO syntax for arrays.)
✞ This can be modified to instead use reverse AD without changing the API:
SUBROUTINE GRAD(F, X, N, DX)
EXTERNAL F
ADR(Y)
Y = F(X)
END ADR(DX(I) = COTANGENT(X(I)), I=1,N)
END
Although not intended to support checkpoint-reverse AD, our constructs are suffi✞ciently powerful to express a reverse checkpoint:
C
CHECKPOINT REVERSE F->G. BOTH 1ST ARG IN, 2ND ARG OUT
CALL F(X, Y)
ADR(COTANGENT(Z(I)) = . . ., I=1,NZ)
CALL G(Y, Z)
END ADR(DY(I) = COTANGENT(Y(I)), I=1,NY)
ADR(COTANGENT(Y(I)) = DY(I), I=1,NY)
CALL F(X, Y)
END ADR(DX(I) = COTANGENT(X(I)), I=1,NX)
This sort of encapsulation empowers numeric programmers to conveniently experiment with the choice of differentiation method, or with the use of various sorts
of derivatives, including higher-order derivatives, without tedious modification of
lengthy call chains.
Extension 2: Nested Subprograms
We borrow from A LGOL 60 (Backus et al., 1963) and generalize the F ORTRAN
“statement function” construct by allowing subprograms to be defined inside other
subprograms, with lexical scope.
For example, given a univariate maximizer ARGMAX, we can express the idea of
line search as follows:
AD in Fortran: Design
5
✞
C
MAXIMIZE F ALONG THE LINE PARALLEL TO XDIR THROUGH X
SUBROUTINE LINMAX(F, X, XDIR, LENX, N, XOUT)
EXTERNAL F
DIMENSION Y(50)
FUNCTION ALINE(DIST)
DO 2012 I=1,LENX
2012
Y(I) = X(I)+DIST*XDIR(I)
ALINE = F(Y)
END
BESTD = ARGMAX(ALINE, 0.0, N)
DO 2013 I=1,LENX
2013 XOUT(I) = X(I)+BESTD*XDIR(I)
END
Here we are using a library univariate maximizer to maximize the univariate function ALINE, which maps the distance along the given direction to the value of our
multidimensional function of interest F at that point. Note that ALINE refers to variables defined in its enclosing scope, namely F, X, XDIR, LENX, and Y. Note that if
ARGMAX uses derivative information, AD will be performed automatically on ALINE.
3 Concrete Example
We employ a concrete example to show the convenience of the above constructs. We
will also illustrate the implementation on this example in (Radul et al., 2012). Let
two companies, Apple and Banana, be engaged in competition in a common fashion
accessories market. Each chooses a quantity of their respective good to produce, and
sells all produced units at a price determined by consumer demand. Let us model
the goods as being distinct, but partial substitutes, so that availability of products of
A decreases demand for products of B and vice versa (though perhaps not the same
amount). We model both companies as having market power, so the price each gets
will depend on both their own output and their competitor’s. Each company faces
(different) production costs and seeks to maximize its profit, so we can model this
situation as a two player game. Let us further assume that the quantities involved
are large enough that discretization effects can be disregarded.
An equilibrium (a∗ , b∗ ) of a two-player game with continuous scalar strategies a
and b and payoff functions A(a, b) and B(a, b) must satisfy a system of equations:
a∗ = argmax A(a, b∗ )
b∗ = argmax B(a∗ , b)
a
(1)
b
Equilibria can be sought by finding roots of
a∗ = argmax A(a, argmax B(a∗ , b))
a
(2)
b
which is the technique we shall employ.1 Translated into computer code in the most
natural way, solving this equation involves a call to an optimization subprogram
within the function passed to an optimization subprogram, itself within the function
6
Alexey Radul, Barak A. Pearlmutter, and Jeffrey Mark Siskind
passed to a root-finding subprogram. If said optimization and root-finding subprograms need derivative information, this gives rise to deeply nested AD.
Note that in (2), the payoff function B is bivariate but argmax takes a univariate
(in the variable of maximization) objective function. The a∗ variable passed to B is
free in the innermost argmax expression. Free variables occur naturally in mathematical notation, and we support them by allowing nested subprogram definitions.
✞ We can use our extensions to code finding the roots of (2) in a natural style:
C
ASTAR & BSTAR: GUESSES IN, OPTIMIZED VALUES OUT
SUBROUTINE EQLBRM(BIGA, BIGB, ASTAR, BSTAR, N)
EXTERNAL BIGA, BIGB
FUNCTION F(ASTAR)
FUNCTION G(A)
FUNCTION H(B)
H = BIGB(ASTAR, B)
END
BSTAR = ARGMAX(H, BSTAR, N)
G = BIGA(A, BSTAR)
END
F = ARGMAX(G, ASTAR, N)-ASTAR
END
ASTAR = ROOT(F, ASTAR, N)
END
where we implement just the minimal cores of one-dimensional optimization and
✞root finding to illustrate the essential point — root finding by the Rhapson method:
FUNCTION ROOT(F, X0, N)
X = X0
DO 1669 I=1,N
CALL DERIV2(F, X, Y, YPRIME)
1669 X = X-Y/YPRIME
ROOT = X
END
✞
SUBROUTINE DERIV2(F, X, Y, YPRIME)
EXTERNAL F
ADF(X)
Y = F(X)
END ADF(YPRIME = TANGENT(Y))
END
✞and optimization by finding the root of the derivative:
FUNCTION ARGMAX(F, X0, N)
FUNCTION FPRIME(X)
FPRIME = DERIV1(F, X)
END
ARGMAX = ROOT(FPRIME, X0, N)
END
1
The existence or uniqueness of an equilibrium is not in general guaranteed, but our particular A
and B have a unique equilibrium. Coordinate descent (alternating optimization of a∗ and b∗ ) would
require less nesting, but has inferior convergence properties. Although this example involves AD
through iterative processes, we do not address that issue in this work: it is beyond the scope of this
paper, and used here only in a benign fashion, for vividness.
AD in Fortran: Design
7
On our concrete objective functions these converge rapidly, so for clarity we skip
the clutter of convergence detection.
This strategy impels us to compute derivatives nested five deep, in a more complicated pattern than just a fifth-order derivative of a single function. This undertaking is nontrivial with current AD tools (Siskind and Pearlmutter, 2008a), but
becomes straightforward with the proposed extensions—embedded AD syntax and
nested subprograms make it straightforward to code sophisticated methods that require complex patterns of derivative information.
4 Discussion
The FARFEL AD extensions hew to the spirit of F ORTRAN, which tends to prefer
blocks rather than higher-order operators for semantic constructs. (In this, these extensions are syntactically quite similar to a set of AD extensions integrated with
the NAGware F ORTRAN 95 compiler (Naumann and Riehme, 2005), albeit quite
different semantically. Unfortunately those NAGware extensions are no longer publicly available, and the limited available documentation is insufficient to allow a
detailed comparison.) A reasonably straightforward implementation technology involves changing transformed blocks into subprograms that capture their lexical
variable context and closure-converting these into top-level subprograms, rendering them amenable to processing with existing tools (Radul et al., 2012). Since
the machinery for nested subprograms is present, allowing them imposes little additional implementation effort. Moreover, as seen in the example above, that extension
makes code that involves heavy use of higher-order functions, which is encouraged
by the availability of the AD constructs, more straightforward. In this sense AD
blocks and nested subprograms interact synergistically.
These new constructs are quite expressive, but this very expressiveness can tax
many implementations, which might not support some combinations or usages. For
instance, code which makes resolution of the AD at compile time impossible (an
n-th derivative subprogram, say) would be impossible to support without a dynamic
run-time AD mechanism. This would typically not be available. Another common
restriction would be that many tools do not support reverse mode at all and even
those that do typically do not allow nesting over reverse mode, either reverse-overreverse or forward-over-reverse. It is the responsibility of the implementation to
reject such programs with a cogent error.
The FARFEL extensions are implemented by the FARFALLEN preprocessor (Radul
et al., 2012), which generates input for and invokes existing AD tools. This leverages existing AD systems to provide the differentiation functionality in a uniform
and integrated way, extending the reach of AD by making its use easier, more natural, and more widely applicable.
Such a prepreprocessor can target different AD systems (like A DIFOR (Bischof
et al., 1992) and TAPENADE (Hascoët and Pascual, 2004)), allowing easy porting of
code from one AD system to another. It could even mix AD systems, for example
8
Alexey Radul, Barak A. Pearlmutter, and Jeffrey Mark Siskind
using TAPENADE to reverse-transform code generated by using A DIFOR in forward
mode, capturing their respective advantages for the application at hand. The effort of
implementing such retargetings and mixings could then be factored to one developer
(of the prepreprocessor) instead of many end users of AD.
A more important benefit of extending F ORTRAN with AD syntax and nested
subprograms is that a host of notions become reusable abstractions—not just firstorder derivatives, but also their variations, combinations, and uses, e.g., Jacobians,
Hessians, Hessian-vector products, filters, edge detectors, Fourier transforms, convolutions, Hamiltonians, optimizations, integration, differential-equation solvers.
The interfaces to different methods for these tasks can be made much more uniform
because, as our ARGMAX did, they can accept subprograms that just accept the variables of interest (in this case, the argument of maximization) and take any needed
side information from their lexical scope; and subprograms such as ARGMAX can obtain any derivative information they wish from AD without having to demand it be
passed in as arguments. So different maximization methods can be tried out on the
same objective function with ease, regardless of how much derivative information
they require; and at the same time, different objective functions, that carry different
side information, can be maximized by the same maximization subprogram without
having to adjust it to transmit the needed side information. Essentially, derivatives
are requested where they are needed, and the implementation does the necessary
bookkeeping.
These modularity benefits are illustrated by our example program: the FARFEL
input is only 64 lines of code, whereas the amount of code it expands into, which
is comparable to what would need to be written by hand without these extensions,
weighs in at a much more substantial 160 for TAPENADE and 315 for A DIFOR,
including the configuration data needed to run the AD preprocessors to produce
the needed derivatives. Manually performing the 5 nested applications of AD this
example calls for is a tedious, error prone, multi-hour effort, which must be undertaken separately for each preprocessor one wishes to try.2 Existing AD tools do
already save the major labor of manually implementing derivative and gradient subprograms, and keeping them in sync with the subprograms being differentiated. The
further preprocessing step outlined above leverages these tools into being even more
useful. For larger programs, the savings of implementation and maintenance effort
would be considerable.
The present suggestion is not, of course, limited to F ORTRAN. Nested subprograms have gained wide adoption in programming-language designs from A L GOL 60 and beyond, and have yielded proven gains in programmer productivity.
Their advantages for code expressiveness have led to functions with lexical scope
being used as a mathematical formalism for reasoning about computing (Church,
1941), to programming languages organized around the function as the primary
program construct (Jones, 2003), and to compilers that specialize in the efficient
representation and use of functions (Jones et al., 1993; Steele, 1978).
2
A detailed step-by-step discussion of the transformation of this example along with all intermediate code is available at http://www.bcl.hamilton.ie/˜ qobi/fortran/.
AD in Fortran: Design
Fig. 1 Performance comparison. Smaller is faster.
Numeric solution of (2)
with above FARFEL code,
N = 1000 iterations at each
level, FARFALLEN targeting
two F ORTRAN -based AD
tools; for comparison, the
same computation is coded
in VLAD (Pearlmutter and
Siskind, 2008) and compiled
with S TALINGRAD (Siskind
and Pearlmutter, 2008b).
Computer: Intel i7 870 @
2.93GHz, G FORTRAN 4.6.29, 64-bit Debian sid, -Ofast
-fwhole-program, single
precision. See (Radul et al.,
2012) for details.
9
CPU Time (seconds)
10
8.92
6.97
5.83
5
0
TAPENADE
A DIFOR
S TALINGRAD
One can also add ADF- and ADR-like constructs to other languages that have preprocessor implementations of AD, for example, C and ADIC (Bischof et al., 1997).
One would not even need to add nested subprograms in the preprocessor, because
GCC already implements them for GNU C. Doing so would expand the convenience
(and therefore reach) of existing AD technology even further.
Retrofitting AD onto existing languages by preprocessing is not without its limitations, however. Efficient AD preprocessors must construct a call graph in order to
determine which subprograms to transform, along with a variety of other tasks already performed by the compiler. Moreover, optimizing compilers cannot be relied
upon to aggressively optimize intricate machine-generated code, as such code often
exceeds heuristic cutoffs in various optimization transformations. This imposes a
surprisingly serious limitation on AD preprocessors. (Together, these also imply
a significant duplication of effort, while providing room for semantic disagreements between AD preprocessors and compilers which can lead to subtle bugs.)
This leads us to anticipate considerable performance gains from designing an optimizing compiler with integrated AD. Indeed, translating our concrete example into
VLAD (Pearlmutter and Siskind, 2008) and compiling with S TALINGRAD (Siskind
and Pearlmutter, 2008b), our prototype AD-enabled compiler, justifies that suspicion (see Fig. 1). We therefore plan to make a VLAD back-end available in version
2 of FARFALLEN.
5 Conclusion
We have defined and motivated extensions to F ORTRAN for convenient, modular
programming using automatic differentiation. The extensions can be implemented
10
Alexey Radul, Barak A. Pearlmutter, and Jeffrey Mark Siskind
as a prepreprocessor (Radul et al., 2012). This strategy enables modular, flexible
use of AD in the context of an existing legacy language and tool chain, without
sacrificing the desirable performance characteristics of these tools: only about 20%–
50% slower than a dedicated AD-enabled compiler, depending on which F ORTRAN
AD system is used.
Acknowledgements This work was supported, in part, by Science Foundation Ireland grant
09/IN.1/I2637, National Science Foundation grant CCF-0438806, the Naval Research Laboratory
under Contract Number N00173-10-1-G023, and the Army Research Laboratory accomplished
under Cooperative Agreement Number W911NF-10-2-0060. Any views, opinions, findings, conclusions, or recommendations contained or expressed in this document or material are those of the
author(s) and do not necessarily reflect or represent the views or official policies, either expressed
or implied, of SFI, NSF, NRL, the Office of Naval Research, ARL, or the Irish or U.S. Governments. The U.S. Government is authorized to reproduce and distribute reprints for Government
purposes, notwithstanding any copyright notation herein.
References
J. W. Backus, F. L. Bauer, J. Green, C. Katz, J. McCarthy, P. Naur, A. J. Perlis,
H. Rutishauser, K. Samelson, B. Vauquois, J. H. Wegstein, A. van Wijngaarden,
and M. Woodger. Revised report on the algorithmic language ALGOL 60. The
Computer Journal, 5(4):349–367, 1963. doi: 10.1093/comjnl/5.4.349.
Christian H. Bischof, Alan Carle, George F. Corliss, Andreas Griewank, and Paul D.
Hovland. ADIFOR: Generating derivative codes from Fortran programs. Scientific Programming, 1(1):11–29, 1992.
Christian H. Bischof, Lucas Roh, and Andrew Mauer. ADIC — An extensible automatic differentiation tool for ANSI-C. Software–Practice and Experience, 27(12):1427–1456, 1997. doi: 10.1002/(SICI)1097-024X(199712)
27:12h1427::AID-SPE138i3.0.CO;2-Q. URL http://www-fp.mcs.anl.
gov/division/software.
Alonzo Church. The Calculi of Lambda Conversion. Princeton University Press,
Princeton, NJ, 1941.
Laurent Hascoët and Valérie Pascual. TAPENADE 2.1 user’s guide. Rapport technique 300, INRIA, Sophia Antipolis, 2004. URL http://www.inria.fr/
rrrt/rt-0300.html.
Simon Peyton Jones. Haskell 98 language and libraries: the revised report. Journal
of Functional Programming, 13(1), January 2003.
S.L.P. Jones, C. Hall, K. Hammond, W. Partain, and P. Wadler. The Glasgow Haskell
compiler: a technical overview. In Proc. UK Joint Framework for Information
Technology (JFIT) Technical Conference, volume 93, 1993.
Uwe Naumann and Jan Riehme. Computing adjoints with the NAGWare Fortran 95
compiler. In H. M. Bücker, G. Corliss, P. Hovland, U. Naumann, and B. Norris,
editors, Automatic Differentiation: Applications, Theory, and Implementations,
AD in Fortran: Design
11
volume 50 of Lecture Notes in Computational Science and Engineering, pages
159–169. Springer, New York, NY, 2005. doi: 10.1007/3-540-28438-9 14.
Barak A. Pearlmutter and Jeffrey Mark Siskind. Using programming language
theory to make automatic differentiation sound and efficient. In Christian H.
Bischof, H. Martin Bücker, Paul D. Hovland, Uwe Naumann, and J. Utke, editors, Advances in Automatic Differentiation, volume 64 of Lecture Notes in Computational Science and Engineering, pages 79–90. Springer, Berlin, 2008. ISBN
978-3-540-68935-5. doi: 10.1007/978-3-540-68942-3 8.
Alexey Radul, Barak A. Pearlmutter, and Jeffrey Mark Siskind. AD in Fortran,
Part 2: Implementation via prepreprocessor. In Advances in Automatic Differentiation, Lecture Notes in Computational Science and Engineering, Berlin, 2012.
Springer. To appear. Also arXiv technical report http://arxiv.org/abs/
1203.1450.
Jeffrey Mark Siskind and Barak A. Pearlmutter. Putting the automatic back into AD:
Part I, What’s wrong. Technical Report TR-ECE-08-02, School of Electrical and
Computer Engineering, Purdue University, West Lafayette, IN, USA, jan 2008a.
URL http://docs.lib.purdue.edu/ecetr/368.
Jeffrey Mark Siskind and Barak A. Pearlmutter. Using polyvariant union-free flow
analysis to compile a higher-order functional-programming language with a firstclass derivative operator to efficient Fortran-like code. Technical Report TRECE-08-01, School of Electrical and Computer Engineering, Purdue University,
West Lafayette, IN, USA, jan 2008b. URL http://docs.lib.purdue.
edu/ecetr/367.
B. Speelpenning. Compiling Fast Partial Derivatives of Functions Given by Algorithms. PhD thesis, Department of Computer Science, University of Illinois at
Urbana-Champaign, Urbana-Champaign, IL, January 1980.
Guy Lewis Steele, Jr. RABBIT, a compiler for Scheme. Technical Report TR474,
Artificial Intelligence Laboratory, Massachusetts Institute of Technology, Cambridge, MA, USA, May 1978.
R. Wengert. A simple automatic derivative evaluation program. Communications of
the ACM, 7(8):463–464, 1964.
| 6cs.PL
|
A Model Predictive Control Approach for
Low-Complexity Electric Vehicle Charging
Scheduling: Optimality and Scalability
Wanrong Tang Student Member, IEEE and Ying Jun (Angela) Zhang, Senior Member, IEEE
arXiv:1502.01456v3 [math.OC] 1 Apr 2016
Department of Information Engineering, The Chinese University of Hong Kong
Shatin, New Territories, Hong Kong
Abstract—With the increasing adoption of plug-in electric vehicles (PEVs), it is critical to develop efficient charging coordination
mechanisms that minimize the cost and impact of PEV integration to the power grid. In this paper, we consider the optimal PEV
charging scheduling, where the non-causal information about
future PEV arrivals is not known in advance, but its statistical
information can be estimated. This leads to an “online” charging
scheduling problem that is naturally formulated as a finitehorizon dynamic programming with continuous state space and
action space. To avoid the prohibitively high complexity of solving
such a dynamic programming problem, we provide a Model
Predictive Control (MPC) based algorithm with computational
complexity O(T 3 ), where T is the total number of time stages.
We rigorously analyze the performance gap between the nearoptimal solution of the MPC-based approach and the optimal
solution for any distributions of exogenous random variables.
Furthermore, our rigorous analysis shows that when the random
process describing the arrival of charging demands is first-order
periodic, the complexity of proposed algorithm can be reduced
to O(1), which is independent of T . Extensive simulations show
that the proposed online algorithm performs very closely to the
optimal online algorithm. The performance gap is smaller than
0.4% in most cases.
I. I NTRODUCTION
A. Background and Contributions
The massive deployment of PEVs imposes great challenges
to smart power grid, such as voltage deviation, increased power
losses, and higher peak load demands. It is critical to design
PEV charging mechanisms that minimize the cost and impact
of PEV integration. Previously, PEV charging coordination has
been extensively studied to minimize power loss, minimize
load variance, or minimize charging cost, etc [3]–[7]. Ideally,
the load demand can be flattened as much as possible if
the information about future charging demand is known noncausally when calculating the charging schedule. However, in
practice, a PEV charging station only knows the load demand
This work was presented in part as [1]. This work was supported in part
by the National Basic Research Program (973 program Program number
2013CB336701), and three grants from the Research Grants Council of
Hong Kong under General Research Funding (Project number 2150828 and
2150876) and Theme-Based Research Scheme (Project number T23-407/13N).
W. Tang is with the Department of Information Engineering, The Chinese
University of Hong Kong, Shatin, New Territories, Hong Kong (Email:
[email protected]).
Y. J. Zhang is with the Department of Information Engineering, The Chinese
University of Hong Kong. She is also with Shenzhen Research Institute, The
Chinese University of Hong Kong (Email: [email protected]).
of the PEVs that have arrived, but not that of the PEVs
coming in the future. Fortunately, the statistical information
of the future charging demands can often be acquired through
historic data, which benefits the control of the PEV charging
scheduling in practical scenarios.
In this paper, we consider the optimal PEV charging
scheduling, assuming that the future charging demand is not
known a priori, but its statistical information can be estimated.
In particular, we define the cost of PEV charging as a general
strictly convex increasing function of the instantaneous load
demand. Minimizing such a cost leads to a flattened load demand, which is highly desirable for many reasons [3]–[7]. The
online PEV charging scheduling problem is formulated as a
finite-horizon dynamic programming problem with continuous
state space and action space. To avoid the prohibitively high
complexity of solving such a dynamic programming problem,
we provide a Model Predictive Control (MPC) approach to
obtain a near-optimal solution. Instead of adopting the generic
convex optimization algorithms to solve the problem, we
propose an algorithm with computational complexity O(T 3 )
by exploring the load flattening feature of the solution, where
T is the total number of time stages. We rigorously analyze
the performance gap between the near-optimal solution of the
MPC-based approach and the optimal solution, and the result
applies to any distributions of exogenous random variables.
Specially, the performance gap is evaluated by the Value
of the Stochastic Solution (VSS), which represents the gap
between the solution of the approximate approach and that
of dynamic programming problem [8]–[10]. Furthermore, our
analysis shows that when the random process describing
the arrival of charging demands is first-order periodic, the
complexity of proposed algorithm can be reduced to O(1),
which is independent of T . Extensive simulations show that
the proposed algorithm performs very closely to the optimal
solution. The performance gap is smaller than 0.4% in most
cases. As such, the proposed algorithm is very appealing for
practical implementation due to its scalable computational
complexity and close to optimal performance.
The rest of the paper is organized as follows. A review of the
related work on the PEV charging scheduling with uncertain
load demand is presented in Section I-B. We introduce the
problem formulations of both offline and online PEV charging
problem in Section II. In Section III, we propose a MPC based
online algorithm and analyze its performance gap. The O(1)-
complexity algorithm is given when the arrival process is firstorder periodic in Section IV. Simulation results are presented
in Section V. Finally, the paper is concluded in Section VI.
B. Related Work
The works on the PEV charging scheduling with uncertain
PEV load demand include both simulation-based evaluations
[11], [12] and theoretical performance guarantees [13]–[17].
Meanwhile, MPC is one of most commonly approaches for
which has been widely adopted in recent studies [12]–[15].
[12] leverages the MPC based method to design a dynamic
charging and driving cost control scheme. Both [13] and [14]
apply MPC algorithms to minimize the load variation. [15]
proposes a plug and play MPC approach to minimize the
voltage fluctuations by assuming that the load demand is timeperiodic. Compared to [12]–[15], in this paper we analyze
the performance gap between the solution of MPC approach
and the optimal solution regardless of the distribution of the
load demand. Besides, we provide a more scalable algorithm
with O(1)− complexity as well as the optimality analysis
for the case when the load demand is first-order periodic.
Additionally, the objective functions in [12]–[15] are quadratic
forms of load demand. Whereas in this paper, the objective
function is a general strictly convex increasing function which
reflects both the charging cost and the load variance.
As to the amount of information needed, the EV charging
scheduling algorithms in [16] and [17] require the probability
distribution of the random PEV arrival process. In contrast,
the proposed algorithm in this paper only requires the firstorder moment, i.e., the expected values of the random demand
patterns. In practice, it is a lot easier to obtain the expected
values of random process than to obtain the probability distribution of a random process. Convex cost functions are also
considered in [18] and [19]. Both of them devise the online
algorithms for battery charging control problems, where there
is no charging deadline for the battery. The PEV charging
scheduling problem in this paper differs from stationary battery
charging in that each PEV has a demand to be satisfied before
a certain deadline.
II. P ROBLEM F ORMULATION
We consider the PEV charging scheduling problem, where
PEVs arrive at the charging station at random instants with
random charging demands that must be fulfilled before a
random departure time.
A. Optimal Offline PEV Charging Problem
For the ease of understanding, we first introduce an ideal
case, where all the non-causal information of based load and
the PEVs, including the arrival times, departure times and
charging demands are known to the charging station before
the system time. The entire system time is divided into T
equal-length time slots. Let N denote the set of PEVs that
arrive during the system time. Notice that for a given time
slot number T , N is itself a random set due to the random
arrival of PEVs. We denote by I(t) the set of PEVs that
(s)
are in the charging station during slot t. Denote by ti and
(e)
ti the arrival time slot and the departure time slot of PEV
i, respectively. di denotes the charging demand that PEV i
requires. The charging station needs to decide, the charging
rate xit , ∀i ∈ I(t). To satisfy the demand di , xit must satisfy
Pt(e)
i
(s) xit = di . Let st be the total charging rate of time slot
t=ti
t, i.e.,
X
xit , ∀t = 1, 2, · · · , T,
st =
(1)
i∈I(t)
which is also called charging load at time t. The total load
consists of both the charging load and the inelastic base load in
the same location. The base load, denoted by lt , represents the
load of other electricity consumptions at time t except
P for PEV
charging. Then, the total load at time t is given by i∈It xit +
lt . Suppose that the charging cost at time t is a strictly convex
increasing function of the total load, denoted by f (st + lt ).
The convexity and increasing property of f (st + lt ) reflects
the fact that each unit of additional power demand becomes
more expensive to obtain and make available to the consumer.
For example, in the wholesale market, the instantaneous cost
can be modeled as an increasing quadratic function of the
instant load [4]–[6]. On the other hand, the convexity of f (st +
lt ) also captures the intent of reducing the load fluctuation
over time [7]. Then the total cost over time T is computed
PT
(s) (e)
as t=1 f (st + lt ). In the ideal case, assume that lt , ti , ti ,
and di for all t = 1, · · · , T, i ∈ N are known non-causally at
the beginning of the system time. Then, the charging station
can solve (2) and obtain the optimal charging rate, denoted by
x∗it for all time t and the optimal total cost, denoted by Ψ1 .
Such a solution is referred to as an “optimal offline solution”.
T
X
Ψ1 = min
xit
f
X
xit + lt
(2a)
xit = di , ∀i ∈ N ,
(2b)
t=1
i∈I(t)
(e)
ti
X
s.t.
(s)
t=ti
(e)
(s)
xit ≥ 0, ∀t = ti , · · · , ti , ∀i ∈ N .
(2c)
In particular, the optimal
rate, denoted by
P total charging
∗
s∗t , is defined as s∗t =
x
.
Note
that there are in
i∈I(t) it
total O(T |I(t)|) variables in (2), where |I(t)| denotes the
cardinality of the set I(t). This number can be quite large
when the number of cars present at each time slot, |I(t)|, is
large. Next, we propose an equivalent transformation of (2)
that drastically reduces the number of variables. In particular,
the following Theorem 1 shows that as long as we find the
optimal s∗t ∀t, the optimal x∗it ∀i, t can be obtained by earliest
deadline first (EDF) scheduling.
Theorem 1: If a set of st ’s satisfy the following inequality
for all n = 1, · · · , T
n
X
t=1
X
(e)
i∈{i|ti
di ≤
=t}
n
X
t=1
st ≤
n
X
t=1
X
(s)
i∈{i|ti
di ,
(3)
=t}
then there exists at least a set of xit ’s that is feasible to (2).
One such set of xit ’s can be obtained by EDF scheduling,
which charges the PEV i ∈ I(t) with the earliest deadline
at a rate st at each time t. Moreover, when st = s∗t , the set
of xit ’s obtained by EDF scheduling are the optimal solution,
x∗it , to (2).
Proof: Please see the detailed proof in Appendix A. To
see Theorem 1, note that (3) implies that the total energy
charged by any time slot n is no less than the total charging
demand that must be satisfied by time n and no more then
the total charging demand of PEVs which have arrived up
to time n. On the other hand, by EDF scheduling, PEVs
with earlier deadlines must be fully charged before those
with later deadlines can be charged. Thus, (3) guarantees
the fulfillment of the charging demands of each individual
PEV. With Theorem 1, we can transform (2) to the following
equivalent problem with T variables.
Ψ1 = min
st
s.t.
T
X
t=1
n
X
f (st + lt )
st ≥
t=1
n
X
n
X
j=1
st ≤
t=1
n
X
j=1
statistical information of the unknown future demands. The
charging rate sk , once determined, cannot be changed in the
future. Specifically, the remaining charging
Pk−1 demand of PEV
i at time k is given by dˆki = di − t=t(s) xit . Note that,
i
dˆki = di for all PEVs that have not yet arrived by time k − 1.
A close look at (4) suggests that the charging schedule st
only depends on the total charging demand that needs to be
finished before a certain time, but not the demand due to
individual
P PEVs. Thus, for notational simplicity, we define
d˜kt = i∈{i|t(e) =t} dˆki , ∀t = k, · · · , T, as the total unfinished
i
charging demand at time k that must be completed by time t.
With this, we define the state of system at time t as
Dt = [lt , d˜tt , d˜tt+1 , · · · , d˜tT ],
(4a)
(5)
d˜tt0
X
(e)
i∈{i|ti
di , ∀n = 1, · · · , T. (4b)
=j}
X
where lt is the base load at time t,
is the total unfinished
charging demand at time t that must be completed by time
t0 . Let ξ t represent the random arrival events at time t. ξ t is
defined as
t
ξt = [ιt , ηtt , ηt+1
, · · · , ηet t ],
di , ∀n = 1, · · · , T. (4c)
(s)
i∈{i|ti =j}
s∗t
to (4) has an interesting feature: it
The optimal solution
does not change with the cost function f (st + lt ), as long as
f is strictly convex. Moreover, s∗t also minimizes the variance
of total load subjecting to (4b) and (4c), where
the variance
PT
PT
t=1 st +lt 2
of total load is defined as t=1 (st + lt −
) [13],
T
[14]. This is proved in Theorem 2.
Theorem 2: The optimal solution s∗t to (4) does not change
with the cost function f (.), as long as f (.) is strictly convex.
Moreover, s∗t is a load flattening solution that minimizes the
variance of total load.
Proof: Please see the detailed proof in Appendix B.
Remark 1: In practice, a capacity constraint on st + lt is
present for each t due to the hardware limitations and security
concerns. The constraint is omitted in our formulation for
the following reason. Theorem 2 indicates that the optimal
solution s∗t to (4) minimizes the variance of total load. That is,
any other scheduling solution would have a higher peak load,
and therefore is more likely to violate the capacity constraint.
In this sense, the optimal solution s∗t to (4) is “capacity
optimal” in the sense that if the optimal solution to Problem
4 (or equivalently Problem 2) violates the capacity constraint,
then there does not exist any other scheduling solutions that
satisfy the capacity constraint.
B. Online PEV Charging problem
For the online PEV charging problem, the charging schedule only depends on the statistic information of future load
demand, the current based load and the remaining charging
demands and deadlines of the PEVs that have arrived so far.
In contrast to the ideal case in the last subsection, in practice
the charging station only knows the remaining charging demands and departure deadlines of the PEVs that have already
arrived, as well as the current base load. The online charging
scheduling algorithm computes the charging rate sk at each
time slot k based on the known causal information and the
(6)
where ιt is the base load at time t, ηtt0 is the total charging
demand that arrive at time t and must be fulfilled by time t0 ,
et is the latest deadline among the PEVs that arrive at time t.
Then, the state transition, defined as
Dt+1 := g(st , Dt , ξt+1 ),
(7)
is calculated as follows:
lt+1 = ιt+1
(8)
and
= d˜tt0 − st −
d˜t+1
t0
0
tX
−1
+ +
, ∀t0 = t+1, · · · , T. (9)
d˜tj +ηtt+1
0
j=t
+
Here, [x] = max{x, 0}. With the above definitions of system
state and state transition, we are now ready to rewrite (4) into
the following finite-horizon dynamic programming problem.
Qk (Dk ) = min
sk
f (sk + lk ) + Eξk+1 [Qk+1 (g(sk , Dk , ξk+1 ))]
(10a)
s. t.
d˜kk ≤ sk ≤
T
X
d˜kt ,
(10b)
t=k
where Qk (Dk ) is the optimal value of the dynamic programming at time k. The left side of (10b) ensures all charging
demands to be satisfied before their deadlines. The right side
of (10b) implies that the total charging power up to a certain
time cannot exceed the total demands that have arrived up to
that time. By slight abuse of notation, in the rest of the paper
we denote the optimal solutions to both the online and offline
problems as s∗k , when no confusion arises. The actual meaning
of s∗k will be clear from the context. Suppose that s∗k is the
optimal solution to (10) at stage k. Then, the total cost at the
end of system time, denoted by Ψ2 , is provided by
Ψ2 =
T
X
k=1
f (s∗k + lk ).
(11)
Note that (10a) comprises nested expectations with respect
to the random PEV arrivals in the future time stages. Except
for few special cases, it is hard to provide the closed-form
of the optimal solution to (10). On the other hand, (10)
can be solved by the standard numerical methods, such as
backward reduction and the sample average approximation
(SAA) based on Monte Carlo sampling techniques [8]–[10],
[22]. These algorithms typically incur a computational complexity that grows exponentially with both the time span T
and the dimensions of state and decision spaces. Note that (10)
involves continuous state and decision spaces. Discretization
of these spaces leads to a curse of dimensionality, rendering
the computational complexity prohibitively high.
III. MPC- BASED O NLINE C HARGING A LGORITHM
In view of the extremely high complexity of standard
numerical methods, we are motivated to obtain a near-optimal
solution by solving a much simpler problem, which replace
all exogenous random variables by their expected values. This
is referred to as the expected value problem [8]–[10] or the
MPC approach [12]–[15] in the literature. Notice that the firstorder moment, i.e., expectation, of a random process is much
easier to estimate than the other statistics, e.g., variance or the
probability distribution. Thus, the assumption of knowing the
expected values is weaker than the assumptions in other EVcharging algorithms [20], which assume that the probability
distributions of the random processes are known.
Instead of solving the problem using generic convex optimization approaches, we propose a low-complexity online
Expected Load Flattening (ELF) algorithm by exploring the
load flattening feature of the optimal solution to the expected
value problem, as shown in Section III-A. Section III-B provides the theoretical analysis of the performance gap between
the optimal solution to the expected value problem and the
optimal solution to (10).
end of system time, denoted by Ψ3 , is defined as
Ψ3 =
min
sk
s. t.
f (sk + lk ) +
T
X
f (st + νt )
(12a)
t=k+1
j
X
st ≥
j
X
t=k
t=k
j
T
X
X
t=k
st ≤
t=k
d˜kt +
j
X
j
X
µm
n , ∀j
j
d˜kt +
X
em
X
0
d¯tt00 =
µm
n , ∀j = k, · · · , T. (12c)
In each time k, we solve problem (12) and obtain the optimal
charging solution s∗k . Then, problem (12) is resolved with the
updated d˜kt according to the realization of the PEVs arrived in
next time. So on and so forth, we obtain the optimal charging
solution s∗k for time stage k = 2, · · · , T . The total cost at the
(
0
d˜tt00 , for t00 = k, · · · , T, t0 = k,
0
µtt00 , for t00 = t0 , · · · , T, t0 = k + 1, · · · , T .
(14)
The key idea of online ELF algorithm is to balance the
charging load among all time slots k, · · · , T . Specifically, step
3 - 5 is to search the time interval [i∗ , j ∗ ] that has the maximum
load density among current time slots and record the maximum
load density. The time slots with maximum load density are
then deleted, and the process is repeated until the current time
slot k belongs to the maximum-density interval, i.e., i∗ = k.
Algorithm 1: Online ELF Algorithm
1
2
3
4
input : Dk , µt , t = k + 1, · · · , T
output: sk
initialization i = 0, j = 0;
repeat
For all time slot i = k, · · · , T, j = i, · · · , T , compute
Pj
P
0
( jt00 =t0 d¯tt00 + νt0 )
0
i∗ , j ∗ = arg max { t =i
}. (15)
k≤i≤j≤T
j−i+1
Set
y∗ =
5
m=k+1 n=m
(13)
where s∗k is the optimal solution to (12) at time stage k. The
solution to (12) is always feasible to (10) in the sense that
it always guarantees fulfilling the charging demand of the
current parking PEVs before their departures. This is because
the constraints of sk in (10) are included in (12).
Due to the convexity of f (·), the optimal solution is the one
that flattens the total load as much as possible. By exploiting
the load flattening feature of the solution, we present in
Algorithm 1 the online ELF algorithm that solves (12) with
complexity O(T 3 ). The online ELF algorithm have a lower
computational complexity than generic convex optimization
algorithms, such as the interior point method, which has a
complexity O(T 3.5 ) [21]. Notice that similar algorithms have
been proposed in the literature of speed scaling problems [23],
[24] and PEV charging problems [5]. The optimality and the
complexity of the algorithm have been proved therein, and
hence omitted here. The algorithm presented here, however,
paves the way for further complexity reduction to O(1) in
Section IV. For notation brevity, we denote in the online ELF
algorithm
= k, · · · , T, (12b)
m=k+1 n=m
f (s∗k + lk ),
k=1
A. Algorithm Description
Denote the expectation of ξ t as µt = [νt , µtt , · · · , µtT ],
where νt = E[ιt ], µtt0 = E[ηtt0 ], ∀t0 = t, · · · , T. Replacing ξ t
in (10) with µt , we obtain the following deterministic problem:
T
X
6
7
Pj ∗
Pj ∗
¯t0
t00 =t0 dt00
j ∗ − i∗ + 1
t0 =i∗ (
+ νt0 )
.
(16)
Delete time slot i∗ , · · · , j ∗ and relabel the existing time
slot t > j ∗ as t − j ∗ + i∗ − 1.
until i∗ = k;
Set sk = y ∗ − lk .
B. Optimality Analysis
In this subsection, we analyze the optimality of the solution to (12). Notice that MPC approximates the non-causal
random variables by their expected values regardless of their
distribution functions. As a result, such approximation may
lead to unacceptably large performance loss, depending on the
distribution of the random variables. Therefore, the MPC approximation is not always justifiable. A well-accepted metric,
Value of the Stochastic Solution (VSS) is adopted to evaluate
optimality gap between the optimal online solution and the
solution to the expected value problem [8]–[10]. Previous
work, e.g., [9], [10], mainly evaluates VSS using numerical
simulations. Whereas in our analysis, we show that VSS is
always bounded regardless of the distribution of the future
EV charging demands. This provides a strong theoretical
justification of adopting MPC approximation to solve our
problem.
Let Ξ denote a scenario, which is defined as a possible
realization of the sequence of random load demand [22],
Ξ = [ξ2 , ξ3 , · · · , ξT ].
(17)
Here, we treat ξ 1 as deterministic information since the
demand of PEVs arrived at the first stage is known by the
scheduler. Let Φ1 , Φ2 and Φ3 be the expectation of the optimal
value of the offline problem (4), the online problem (10)
and the expected value problem (12), respectively, where the
expectation is taken over the random scenarios. That is,
Φ1 = EΞ [Ψ1 (Ξ)] , Φ2 = EΞ [Ψ2 (Ξ)] , Φ3 = EΞ [Ψ3 (Ξ)] .
(18)
It has been proved previously [8], [9] that
Φ1 ≤ Φ2 ≤ Φ3 .
(19)
To assess the benefit of knowing and using the distributions
of the future outcomes, the VSS is defined as
VSS = Φ3 − Φ2 .
(20)
To show that the online ELF algorithm yields a bounded VSS,
we need to bound Φ3 and Φ2 . Generally, it is hard to calculate
Φ2 or analyze the lower bound of Φ2 directly [9], [10]. Thus,
we choose to analyze the lower bound of Φ1 instead, since
(19) shows that the lower bound of Φ1 is also the bound of
Φ2 . In what follows, we will show the lower bound of Φ1 in
Proposition 1 and the upper bound of Φ3 in Proposition 2.
Proposition 1:
Φ1 ≥ T f
!
Pe1 ˜1 PT Pet j PT
t=1 dt +
t=2
j=t µt +
t=1 νt
T
.
(21)
Proof: Please see the detailed proof in Appendix C.
Let O(t) be the set that O(t) = {(m, n)|em ≥ t, m =
1, · · · , t, n = t, · · · , em }. Then, we show that Φ3 is bounded
by Proposition 2.
Proposition 2: For any distribution of ξ t , t = 1, · · · , T ,
there is
Φ3 ≤ E
T
X
t=1
f
X
ηnm + ιt .
(22)
(m,n)∈O(t)
Proof: Please see the detailed proof in Appendix D.
Now, we are ready to present Theorem 3, which states that
the VSS is bounded for any distribution of random variables.
Theorem 3: For any distribution of random vector ξ t , t =
1, · · · , T, n = t, · · · , T, there is
VSS ≤ E
T
X
t=1
f
X
ηnm + ιt − T f
(m,n)∈O(t)
Γ
T
,
(23)
Pe1 ˜1 PT Pet j PT
dt + t=2 j=t µt + t=1 νt .
where Γ = t=1
Theorem 3 can be easily derived by Proposition 1 and
Proposition 2. In practice, the performance gap between the
online ELF algorithm and the optimal online algorithm is often
much smaller than the bound of VSS. This will be elaborated
in the numerical results in Section V.
IV. O NLINE ELF A LGORITHM UNDER F IRST- ORDER
P ERIODIC R ANDOM P ROCESSES
Notice that the complexity of O(T 3 ) of online ELF algorithm mainly comes from step 3, which exhaustively searches
the maximum-density period [i∗ , j ∗ ] over all subintervals
within [k, T ]. When the random arrival process is first-order
periodic stochastically 1 , we argue that the searching scope
can be reduced to one period from the whole system time
T . Thus, the complexity of step 3 is limited by the length
of a period instead of T . As a result, the complexity of
the algorithm reduces from O(T 3 ) to O(1), implying that
it does not increase with the system time T , and thus the
algorithm is perfectly scalable. In practice, the arrival process
of the charging demands are usually periodic stochastically.
For example, the arrival of charging demands at a particular location is statistically identical at the same time every
day during weekdays (or weekends). the National Household
Travel Survey (NHTS) 2009 gathers information about daily
travel patterns of different types of households in 2009, and
shows that the daily travel statistics (e.g., Average Vehicle
Trip Length, Average Time Spent Driving, Person Trips,
Person Miles of Travel) are very similar for each weekday or
weekend, but different between weekday and weekend [28]. In
Section IV-A, we investigate the case when the random load
demand process is first-order periodic. In Section IV-B, we
provide a closed-form solution to (12) for a special case when
the load demand process is the first-order stationary.
A. First-Order Periodic Process
In this subsection, we consider the case when the arrival
process is first-order periodic. Specifically, the first-order periodic process means that the first-order moment (i.e., mean)
of the random process is periodic. That is, at current time stage
k, for all t = k + 1, · · · , T, we have
µt = E[ξ t ] = µt+p ,
(24)
where ξ t is the random arrival events at time t, µt is the
expectation of ξ t , and p is the length of period. Then, instead
1 The first-order periodic stochastic process is defined as a stochastic process
whose first-order moment, i.e., mean is periodic. That is, the mean of the
random arrival events, i.e., µt is periodic. However, the actual realizations of
the arrival events ξt are uncertain and not periodic.
of considering µt for t = k + 1, · · · , T , we only need to
consider µt for one period, i.e., for t = k + 1, k + p:
k+1
k+1
µk+1 = [νk+1 , µk+1
k+1 , µk+2 , · · · , µk+e1 , 0, · · · , 0],
..
.
(25)
k+p
k+p
µk+p = [νk+p , µk+p
k+p , µk+p+1 , · · · , µk+ep , 0, · · · , 0].
Here, en ≤ T, n = 1, · · · , p is the maximum parking time
for PEVs arriving at time k + n. Specially, we define ê as
ê = max{ek+1 , ek+2 , · · · , ek+p }. We decompose the search
region {i, j|i = k, · · · , T, j = i, · · · , T } into three subregions, defined as Π1 = {i, j|i = k, j = k, · · · , k + ê}, Π2 =
{i, j|i = k, j = k + ê + 1, · · · , T } and Π3 = {i, j|i = k +
1, · · · , T, j = i, · · · , T }, respectively. We denote by X̂, Ŷ , Ẑ
the maximum densities of region Π1 , Π2 , Π3 , respectively.
Indeed, the largest of X̂, Ŷ , Ẑ is the maximum density of the
interval [i∗ , j ∗ ] ⊆ [k, T ] over all possible pairs i, j ∈ {i =
k, · · · , T, j = i, · · · , T }. Let [î1 , ĵ1 ], [î2 , ĵ2 ], [î3 , ĵ3 ] be the
intervals with the maximum density over region Π1 , Π2 , Π3 ,
respectively. By definition, î1 = î2 = k. Similar to the
stationary case, X̂ can be calculated by searching ĵ1 over
{k, · · · , k + ê}. That is,
˜k
n=k (dn
Pt
X̂ =
max
k≤t≤k+ê
P
P
+ νn ) + tn=k tm=n µn
m
.
n−k+1
(26)
Moreover, Lemma 1 shows that Ŷ and Ẑ can be calculated
once the maximum density of interval [k + 1, k + ê] has been
obtained. First, we introduce some definitions which help to
show Lemma 1. Let Π4 be a subset of Π3 , where Π4 is defines
as Π4 = {i, j|i = k + 1, · · · , k + ê, j = i, · · · , k + ê}, and
[ī, j̄] be the interval with
maximum
density of region Π4 , i.e.,
Pk+en n
Pj
m=n µm +νn )
n=i (
ī, j̄ = arg max
.
j−i+1
k+1≤i≤j≤k+ê
Lemma 1: The maximum densities of Π2 and Π3 are calculated by
Pĵ2
Ŷ =
n=k (
Pk+en
m=n
µn
m + νn )
ĵ2 − k + 1
Pĵ3
, Ẑ =
n=î3
(
Pk+en
m=n
µn
m + νn )
ĵ3 − î3 + 1
,
(27)
respectively, where î3 = ī,
ĵ2 =
and
(
max{j̄, k + ê + 1}, if j̄ < ī + p,
j̄ + (r − 1)p,
otherwise.
(
j̄,
if j̄ < ī + p,
ĵ3 =
j̄ + (r − 1)p, otherwise.
(28)
(29)
Proof: Please see the detailed proof in Appendix E.
Based on Lemma 1, we can modified the searching region
of step 3 of online ELF algorithm as follows:
• if j̄ < ī + p, the interval with the maximum density
during time stages [k + 1, T ] is [ī, j̄]. Then, in step 3
of the online ELF algorithm, the search region of i, j
is reduced from {i, j|i = k, · · · , T, j = i, · · · , T } to
{i, j|i = k, · · · , ī, j = i, · · · , ī, j̄}.
• If j̄ ≥ ī+p, the interval with the maximum density during
time stages [k + 1, T ] is [ī, j̄ + (r − 1)p]. Then, in step
3 of the online ELF algorithm, the search region of i, j
can be reduced from {i, j|i = k, · · · , T, j = i, · · · , T }
to {i, j|i = k, · · · , ī, j = i, · · · , ī, j̄ + (r − 1)p}.
As a result, the searching region of the online ELF algorithm
is only related to [k + 1, k + ê] instead of T . Thus, the
computational complexity of the online ELF algorithm is O(1)
instead of O(T 3 ) under first-order periodic process.
B. First-order Stationary Process
In this subsection, we show that the optimal solution to
(12) can be calculated in closed form if the arrival process is
first-order stationary. Due to the page limit, we only provide
the main results here. By first-order stationary, we mean that
the statistical mean of ξ t , i.e., νt and µtt0 , t0 = t, · · · , T only
depends on the relative time difference τ = t0 − t, but not the
absolute value of t. We can then replace νt by ν and replace
µtt0 by µτ , where τ = t0 − t. Then, µt is no longer a function
of t, and can be represented as
µ = [ν, µ1 , µ2 , · · · , µē , 0, · · · , 0],
(30)
where ē is the maximum parking time of a PEV. We denote
by X, Y, Z the maximum densities of region Π1 , Π2 , Π3 ,
respectively. Then, X is calculated by
X=
( Pn
t=k
max
P
d˜kt + n
j=1 (n − k − j + 1)µj + lk − ν
n−k+1
k≤n≤k+ē
)
+ν
,
(31)
and Y, Z are provided in Lemma 2.
Lemma 2: The maximum densities of Π2 and Π3 are
achieved by setting i2 = k, j2 = T, i3 = k + 1, j3 = T ,
and calculated by
Y =
Pk+ē ˜k Pk+ē
t=k dt +
j=1 (T − k − j + 1)µj + lk − ν
T −k+1
(T
−
k
−
j
+
1)µj
j=1
+ ν.
T −k
+ ν, (32a)
Pk+ē
Z=
(32b)
Proof: Please see the detailed proof in Appendix F. The
largest of X, Y, and Z is the maximum density of the
interval [i∗ , j ∗ ] ⊆ [k, T ] over all possible pairs i, j ∈ {i =
k, · · · , T, j = i, · · · , T }. Specially, if X or Y is the largest
one, then k is already contained in the maximum-density
interval, and thus X or Y is the optimal charging rate at time
k. On the other hand, if Z is the largest, then the maximumdensity interval, i.e., [k + 1, T ], does not include k. Following
Algorithm 1, we will delete the maximum-density interval and
repeat the process. Now, time slot k is the only remaining time
slot after deletion. This implies that all charging demands that
have arrived by time slot k should be fulfilled during time slot
k. These arguments are summarized in Proposition 3, which
provides the closed form solution to (12).
Proposition 3: When the random load demand process is
first-order stationary, the optimal charging schedule to (12) is
TABLE I
PARAMETER SETTINGS OF THE PEV TRAFFIC PATTERNS
(33a)
(33b)
(33c)
t=k
V. S IMULATIONS
In this section, we investigate the performance of the
proposed online ELF algorithm through numerical simulations.
All the computations are solved in MATLAB on a computer
with an Intel Core i3-2120 3.30GHz CPU and 8 GB of
memory. For comparison, we also plot the optimal solution to
(10), which is obtained by SAA method [22], and a heuristic
solution by the online AVG algorithm [6], which is obtained by
charging each PEV at a fixed rate, i.e., its charging demand
divided by its parking time. Let the expected cost of AVG
algorithm is denoted by Φ4 . Define the relative performance
loss of ELF and AVG compared with the optimal online
2
2
and Φ4Φ−Φ
, respectively. Similar to [5]
solution as Φ3Φ−Φ
2
2
[6], we adopt an increasing quadratic cost function in the
simulations, i.e., f (st + lt ) = (st + lt )2 . Note that the cost
function is increasing and strictly convex, since the load st +lt
is always non-negative.
A. Average Performance Evaluation
In this subsection, we evaluate the average performance of
the online ELF algorithm under three different traffic patterns,
i.e., light, moderate, and heavy traffics. In particular, the
system time is set to be 24 hours, and each time slot lasts
10 minutes. The PEV arrivals follow a Poisson distribution
and the parking time of each PEV follows an exponential
distribution [25]–[27]. The mean arrival and parking durations
of the three traffic patterns are listed in Table I. The main
difference lies in the arrival rates at the two peak hours, i.e.
12 : 00 to 14 : 00 and 18 : 00 to 20 : 00. The settings of the
peak hour match with the realistic vehicle trips in National
Household Travel Survey (NHTS) 2009 [28]. Specially, the
average number of total PEVs simulated in scenario 1, 2
and 3 are 104, 204 and 304, respectively. We choose the
base load profile of one day in the service area of South
California Edison from [7]. Each PEV’s charging demand is
uniformly chosen from [25, 35]kW h. Each point in Fig. 1,
Fig. 2 and Table II is an average of 105 independent instances
of the scenarios listed in Table I. In Fig. 2, the total loads
s∗k are plotted over time. We notice that the total load of
the online ELF algorithm follows closely to that of optimal
online algorithm, whereas that of the AVG algorithm has a
larger gap from the optimal online algorithm. The average
costs normalized by that of the optimal online algorithm
are plotted in Fig. 1. Moreover, the VSS and the relative
performance loss are listed in Table II. Both the figure and
the table show that ELF performs very close to the optimal
online algorithm. The VSS and the relative performance loss
are no more than 0.1536 and 0.38% respectively. In contrast,
the relative performance loss of AVG algorithm is up to 5.82%,
Arrival Rate (PEVs/hour)
S. 1 S. 2
S. 3
Time of Day
08 : 00 − 10 : 00
10 : 00 − 12 : 00
12 : 00 − 14 : 00
14 : 00 − 18 : 00
18 : 00 − 20 : 00
20 : 00 − 24 : 00
24 : 00 − 08 : 00
7
5
10
5
10
5
0
7
5
35
5
35
5
0
Mean Parking
Time (hour)
7
5
60
5
60
5
0
10
1/2
2
1/2
2
10
0
TABLE II
AVERAGE PERFORMANCE COMPARISON UNDER THREE TRAFFIC PATTERNS
Scenario
VSS
Φ3 −Φ2
Φ2
Φ4 −Φ2
Φ2
1
2
3
0.1178
0.1319
0.1536
0.19%
0.28%
0.38%
3.50%
4.46%
5.82%
which is more than 15 times of that of ELF algorithm. The
relative performance loss of the approximate online algorithm
reflects the percentage of the extra cost compared with the
optimal online algorithm. Obviously, the performance loss is
always the smaller the better. For example, from the report
of Rocky Mountain Institute, the average electricity rate is
11.2cents/kW h, and the average load for a charging station is
100kW [29]. Then, the expected electricity cost of a charging
station for one years is $967680. 6% relative performance loss
means AVG algorithm leads to $58060 extra cost, while the
proposed online algorithm with 0.38% relative performance
loss leads to $3677 extra cost, or an annual saving of $54383
compared with the AVG algorithm.
B. Complexity of The Online Algorithm ELF
In this subsection, we verify the computational complexity
of the online ELF algorithm and also compare the complexity
of online ELF algorithm with that of the optimal online
algorithm and online AVG algorithm, respectively. We still
adopt the SAA method as the optimal online algorithm. Since
the complexity of SAA is too high, truncation is often adopted
in the SAA to reduce the complexity at a cost of performance
loss [8]. As such, we also simulate the complexity of truncated
SAA with a truncation period of 3 hours. Each PEV’s charging
demand is uniformly chosen from [25, 35]kW h, the arrivals
follow a Poisson distribution and the parking time of each
1.07
Optimal online
Online ELF
Online AVG
1.0582
1.06
1.05
1.0446
1.04
Normalized cost
given by the following close-form:
X − lk ,
if X = max{X, Y, Z},
Y − lk ,
if Y = max{X, Y, Z},
s∗k =
k+ē
X
d˜kt ,
otherwise.
1.035
1.03
1.02
1.01
1 1.0019
1
1.0028
1
1.0038
1
0.99
0.98
0.97
Scenario 1
Fig. 1.
Scenario 2
Scenario 3
Normalized costs of three algorithms in three scenarios.
120
Total load of optimal online
Total load of online ELF
Total load of online AVG
65
90
Total load of optimal online
Total load of online ELF
Total load of online AVG
110
100
60
80
Load (kW)
Load (kW)
Load (kW)
90
55
70
50
80
70
60
60
45
Total load of optimal online
Total load of online ELF
Total load of online AVG
50
50
40
40
40
8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 1 2 3 4 5 6 7 8
Time (hour)
8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 1 2 3 4 5 6 7 8
Time (hour)
(a) Scenario 1: light traffic
(b) Scenario 2: moderate traffic
Fig. 2.
(c) Scenario 3: heavy traffic
Base load and total load of three algorithms in three scenarios.
4
10
65
Optimal online without truncation
Optimal online with truncation
Online ELF
Online AVG
3
10
60
Load (kW)
Computional time (s)
8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 1 2 3 4 5 6 7 8
Time (hour)
2
10
55
50
1
10
45
0
10
Total load of optimal online
Total load of online ELF
Total load of online ORCHARD
40
1
Fig. 3.
2
3
4
5
6
Time (hour)
7
8
9
10
CPU computational time over the system time.
PEV follows an exponential distribution, where the arrival rate
and mean parking durations are the same as that in the peak
hours of scenario 2 in Section V.A. We simulate 10 cases in
total, where the system time are set to be 1, 2, · · · , 10 hours.
For each algorithm, we record the CPU computational time at
each time stage and calculate the average CPU computational
time as the sum of CPU computational times of all time
stages divided by the number of time stages. Each point in
Fig. 3 is an average of 100 independent instances. Fig. 3
shows that the CPU computational time of both the online
algorithm ELF and AVG almost grow linearly with the system
time. In contrast, for the SAA based online algorithm with
or without truncation, the average CPU computational times
grow very quickly as system time increases. We notice that
when system time increases to 4 hours, the optimal online
algorithm without truncation consumes more than 2 hours and
the optimal online algorithm with truncation consumes more
than 30 minutes. Meanwhile the proposed online algorithm
ELF only takes several seconds. It is foreseeable that the
computational complexity of optimal online algorithm will
become extremely expensive as we further increase the system
time.
C. Performance Comparison with Online Algorithm ORCHARD [5]
In this section, we compare the proposed online algorithm
ELF with the online algorithm ORCHARD proposed in refer-
8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 1 2 3 4 5 6 7 8
Time (hour)
Fig. 4.
Load comparison of ORCHARD and ELF
ence [5] on the properties of the optimality and computational
complexity. First, we evaluate the average performance of the
proposed online algorithm and online algorithm ORCHARD.
To facilitate the comparison, we also adopt the average performance of optimal online algorithm as a bench mark. For
the system parameters, we use the default settings of scenario
1 in Section V.A. We simulate 105 cases and plot the total
load (the sum of the base load and the PEV charging load)
over time for the three online algorithms in Fig. 4. In addition,
the average performance ratios normalized against the optimal
online solution are shown in Fig. 5 respectively. Fig. 4 shows
that compared with online algorithm ELF, the online algorithm
ORCHARD always produces a larger gap from the optimal
online solution. From Fig. 5, we can see that the proposed
online algorithm ELF achieves a much lower expected cost
than the online algorithm ORCHARD, which indicates that
online algorithm ELF owns a better average performance than
online algorithm ORCHARD.
To compare the computation complexity of online algorithm
ORCHARD and ELF, we adopt the similar case study in
Section V.D of reference [5]. Specifically, we simulate the
CPU computational time of online algorithms by varying the
arrival rates of the PEVs during one day. For the system
parameter, we use the same settings as scenario 1 in Section
V.A except the arrival rates, which are assumed to be the same
during 8 : 00 − 18 : 00 and 0 after 18 : 00. We vary the arrival
Optimal online
2
Online ORCHARD
Online ELF
1.73
1.8
1.6
Normalized cost
1.4
1.2
1.0019
1
1
0.8
0.6
0.4
0.2
problem. Instead of adopting the standard numerical methods with high complexity, we provide a MPC-based online
algorithm with O(T 3 )-complexity. We rigorously analyze the
performance gap between the solution of the MPC-based
approach and the optimal solution for any distribution of
exogenous random variables. Moreover, we show that the
proposed algorithm can be made scalable under the first-order
periodic process of load demand. Besides, our analyses are
validated through extensive simulations.
0
A PPENDIX
Fig. 5.
ELF
Average performance ratios of online algorithm ORCHARD and
A. Proof of Theorem 1:
5
Computional time (s)
10
Optimal online without truncation
Optimal online with truncation
Online ELF
Online ORCHARD
4
10
(s)
(e)
i∈{i|ti =1}
3
10
2
10
100
Fig. 6.
We use the inductive method to show that through EDF
scheduling, all the PEVs can be fulfilled charging before
deadlines. For n = 1, (3) becomes
X
X
di ≥ s1 ≥
di .
(34)
150
200
250
300
350
Number of PEVs
400
450
500
Thus, by EDF scheduling, we can first satisfy the demand of
PEVs whose deadline at time stage 1. That is, for any PEV
(e)
i ∈ {i|ti = 1}, we set
xi1 = di .
CPU computational time over the number of PEVs.
rate in 8 : 00−18 : 00 from 10 to 50 (PEVs/hour) that leads to
the average number of total PEVs during one day varies from
100 to 500. For each specified average number of PEVs, we
simulate the average performance of 108 independent instances
for the online ELF algorithm, the optimal online algorithm and
the online ORCHARD algorithm, respectively, and record the
average CPU computational times of each cases for the three
algorithms, respectively. The results are plotted in Fig. 6. Fig. 6
shows that the average CPU computational time of the optimal
online algorithm grows quickly as the number of total PEVs
increases, while the average CPU computational time of the
online ELF algorithm and the online ORCHARD algorithm
grows slowly as the number of total PEVs increases. When
the number of PEVs is 200, the average CPU computational
time of the optimal online algorithm without truncation is
more than 24 hours. Even for the the optimal online algorithm
with truncation, the average CPU computational time is about
100 minutes. Whereas, the proposed online algorithm ELF
only takes about 4 minutes. Fig. 6 also indicates that the
computational time of online ELF and online ORCHARD as
the number of the PEVs increases.
As a conclusion, the case study shows that the online
algorithm ELF has a better average performance than online
algorithm ORCHARD. It also indicates that the CPU computational time of online ELF and online ORCHARD are similar.
VI. C ONCLUSIONS
In this paper, we formulate the optimal PEV charging
scheduling problem as a finite-horizon dynamic programming
i∈{i|ti =1}
(35)
Assuming that for all time stage m, EDF scheduling can fulfill
charge all the PEVs which depart at or before time stage m,
i.e., there exists at least a set of xit ’s that satisfy
(e)
ti
X
(e)
xit = di , ∀i ∈ {i|ti
≤ m},
(36a)
(s)
t=ti
(s)
(e)
(e)
xit ≥ 0, ∀t = ti , · · · , ti , ∀i ∈ {i|ti
Since
m
X
st ≥
t=1
m
X
t=1
X
≤ m}.
di ,
(36b)
(37)
(e)
i∈{i|ti =t}
Pm
Pm P
then, t=1 st − t=1 i∈{i|t(e) =t} di represents the amount
i
of power which is outputted from the charging station during
time stage 1, · · · , m and charged to the PEVs with deadline
after time stage m. By EDF scheduling, once the PEVs which
depart at time m have been fulfilled charging, we will first
charge the PEVs which depart at time stage m + 1. Thus, if
m
X
st −
t=1
m
X
X
di ≥
t=1 i∈{i|t(e) =t}
i
X
di ,
(38)
(e)
i∈{i|ti =m+1}
we finish charging of PEVs with deadline m + 1, and then go
to charge the PEVs with deadline m + 2. If
m
X
t=1
st −
m
X
X
di <
t=1 i∈{i|t(e) =t}
i
X
di ,
(39)
(e)
i∈{i|ti =m+1}
then the
PEVs with
m + 1 have been charged as
Pm
Pmdeadline
P
power t=1 st − t=1 i∈{i|t(e) =t} di . At time stage m + 1.
i
Since
m+1
X
st ≥
t=1
m+1
X
t=1
X
di ,
(e)
i∈{i|ti
(40)
=t}
1) If x∗it1 = 0 for a particular PEV i at a time slot t1 ∈
(s)
(e)
{ti , · · · , ti }, then, by complementary slackness, we
have ωit1 > 0. From (47a),
f 0 (st1 + lt1 ) = λi + ωit1 .
then,
m
m
X
X
di −
st −
X
sm+1 ≥
(e)
i∈{i|ti
t=1
=m+1}
t=1
X
(e)
i∈{i|ti
di , (41)
=t}
which means all the PEVs with deadline m+1 can be fulfilled
charging. This is because we will charge the PEVs with
deadline m + 1 first by the EDF scheduling. Thus, there exists
at least a set of xit ’s that satisfy
(48)
2) If x∗it2 > 0 for PEV i during a time slot t2 ∈
(s)
(e)
{ti , · · · , ti }, we can infer from (47c) that ωit2 = 0.
Then,
f 0 (st2 + lt2 ) = λi .
(49)
On the other hand, since f (st +lt ) is a strictly convex function
of st + lt , then f 0 (st + lt ) is an increasing function. From the
above discussions, we get the following two conclusions:
(e)
ti
X
(e)
xit = di , ∀i ∈ {i|ti
= m + 1},
(42a)
1) If x∗it1 > 0, x∗it2 > 0, then by (49),
(s)
f 0 (st1 + lt1 ) = f 0 (st2 + lt2 ) = λi .
t=ti
xi,m+1 ≥ 0, ∀i ∈
(e)
{i|ti
= m + 1}.
(42b)
Combining (36) and (42), we get that all the PEVs whose
deadline at or before stage m + 1 can be fulfill charging, i.e.,
there exist at least a set of xit ’s that satisfy
Due to the monotonicity of f 0 (st ), we have s∗t1 + lt1 =
s∗t2 + lt2 .
2) If x∗it1 = 0, x∗it2 > 0, then by (48) and (49), there is
f 0 (st1 + lt1 ) = λi + ωit1 > f 0 (st2 + lt2 ) = λi . (51)
(e)
ti
X
(e)
xit = di , ∀i ∈ {i|ti
≤ m + 1},
(50)
Since f 0 (st ) is a increasing function, we have s∗t1 +lt1 ≥
s∗t2 + lt2 .
(43a)
(s)
t=ti
(s)
(e)
(e)
xit ≥ 0, ∀t = ti , · · · , ti , ∀i ∈ {i|ti
≤ m + 1}.
(43b)
Therefore, we can conclude that by EDF scheduling, there
always exists at least a set of xit ’s that is feasible to (2). This
completes the proof.
Consider two function fˆ(st + lt ) and f¯(st + lt ). Let x̂∗it and
denote the optimal solutions to (2) with f (st +lt ) replaced
by fˆ(st + lt ) and f¯(st + lt ), respectively. Define ŝ∗t , s̄∗t as
X
X
ŝ∗t =
x̂∗it , s̄∗t =
x̄∗it , t = 1, · · · , T,
(52)
x̄∗it
i∈I(t)
B. Proof of Theorem 2:
First, we show that if there exists a PEV parking in the
station at both time t1 and t2 , i.e.,
(s)
(e)
t1 , t2 ∈ {ti , · · · , ti },
(44)
respectively. Suppose that there exists a time slot t1 such that
ŝ∗t1 < s̄∗t1 .
Since
T
X
and
x∗it1 ≥ 0, x∗it2 > 0,
s∗t1 + lt1 ≥ s∗t2 + lt2 .
f 0(
(s)
(e)
xit + lt ) − λi − ωit = 0, i ∈ N , t = ti , · · · , ti ,
s̄∗t =
t=1
X
di ,
(54)
i∈N
ŝ∗t2 > s̄∗t2
(55)
ŝ∗t1 + ŝ∗t2 = s̄∗t1 + s̄∗t2
(56)
Thus, we can find a PEV i ∈ N such that
(47a)
x̂∗it1 < x̄∗t1 , x̂∗it2 > x̄∗it2 .
(57)
x̂∗it2 > 0
(58)
As a result,
(e)
λi (di −
T
X
and
i∈I(t)
ti
X
ŝ∗t =
(53)
there must exist another time slot t2 such that
(46)
The Karush-Kuhn-Tucker (KKT) conditions to the convex
problem (2) are
X
t=1
(45)
then the optimal total loads at time t1 and t2 must satisfy that
i∈I(t)
xit ) = 0, i ∈ N ,
(47b)
(s)
t=ti
ωit xit = 0, i ∈ N , t =
since
(s)
ti , · · ·
(e)
, ti ,
(47c)
where λ, ω are the non-negative optimal Lagrangian multipliers corresponding to (2b) and (2c), respectively. We separate
our analysis into the following two cases:
x̄∗it2
≥ 0. Based on (46), there is
ŝ∗t2 + lt2 ≤ ŝ∗t1 + lt1 .
(59)
Combining (53)(56)(59), we get
s̄∗t2 + lt2 < ŝ∗t2 + lt2 ≤ ŝ∗t1 + lt1 < s̄∗t1 + lt1 .
(60)
Since f¯(st + lt ) is a strictly convex function of st + lt , then,
based on (56) and (60), we have
f¯(s̄∗t1 + lt1 ) + f¯(s̄∗t2 + lt2 ) > f¯(ŝ∗t1 + lt1 ) + f¯(ŝ∗t2 + lt2 ). (61)
PT
t=2
Pet
j=t
µjt . Then, by Jensen’s inequality [30],
Ψ1 (E[Ξ]) ≥
T
X
f
PT
s +l
which indicates that P t=1T t t is a constant. Then, we see
T
PT
s +l
that t=1 (st + lt − t=1T t t )2 is a strictly convex function
of st + lt . This completes the proof.
C. Proof of Proposition 1:
First, we show that Ψ1 (Ξ) is a convex function of Ξ. For
any Ξ, define s∗t (Ξ) as the optimal solution that minimizes
Ψ1 (Ξ) subject to (4b) - (4c). For an arbitrary pair of Ξ0 and
Ξ00 , let Ξ000 = λΞ0 + (1 − λ)Ξ00 ∀λ ∈ [0, 1]. Then, the charging
schedule st (Ξ000 ) such that st (Ξ000 ) = λs∗t (Ξ0 )+(1−λ)s∗t (Ξ00 )
still satisfies (4b) - (4c) with Ξ = Ξ000 due to the linearity of
the constraints. Based on the convexity of f (st + lt ), we have
T
X
t=1
≤λ
T
X
f (s∗t (Ξ0 )
T
X
+ lt ) + (1 − λ)
t=1
(63)
f (s∗t (Ξ00 )
+ lt )
T
X
f (s∗t (Ξ000 ) + lt ) ≤
t=1
T
X
f (st (Ξ000 ) + lt ).
(64)
t=1
Combining (63) and (64), we have
T
X
T
X
n=t
d˜tn ≤
T
X
t=1
n=t−1
T
X
ηnt =
em
X X
ηnm ,
(68)
m∈M n=t
where the inequality holds since st−1 ≥ d˜t−1
t−1 and M =
{m|e
≥
t,
m
=
1,
·
·
·
,
t}
Thus,
we
have st ≤
m
P
m
,
where
O(t)
is
a
bounded
set
for t =
η
(m,n)∈O(t) n
hP
i
P
T
m
1, · · · , T . Therefore, E
f
(
+
ι
)
is an
η
t
t=1
(m,n)∈O(t) n
upper bound of Φ3 . This completes the proof.
E. Proof of Lemma 1:
We provide the proof by discussing the following two cases:
1)If j̄ ≥ ī + p, which means that [ī, j̄] and [ī + p, j̄ + p]
overlaps with each other, then the density of interval [ī, j̄ + p]
is higher than that of [ī, j̄], and the density of [ī, j̄ + 2p] is
higher than that of [ī, j̄ + p]. So on and so forth. Finally, we
see that the interval [ī, j̄ + (r − 1)p] has the maximum density
over region Π3 . Thus, we have î3 = ī, ĵ3 = j̄ + (r − 1)p, and
Pj̄+(r−1)p Pk+en n
( m=n µm + νn )
n=ī
.
j̄ + (r − 1)p − ī + 1
(69)
Likewise, for the region {i, j|i = k, j = k + ê + 1, · · · , T },
we have ĵ2 = j̄ + (r − 1)p, and
Ŷ =
Pj̄+(r−1)p Pk+en n
( m=n µm + νn )
n=k
.
j̄ + (r − 1)p − k + 1
(70)
2)If j̄ < ī + p, then, the density of interval [ī, j̄] is higher
than that of [ī, j̄ + p], and the density of [ī, j̄ + p] is higher
than that of [ī, j̄ + 2p]. So on and so forth. Finally, we see
that the interval [ī, j̄] has the maximum density over region
Π3 . Thus, we have î3 = ī, ĵ3 = j̄, and the corresponding
maximum density
n=ī
P n n
( k+e
m=n µm + νn )
.
j̄ − ī + 1
(71)
For the region Π2 , if j̄ ≤ k + ê + 1, then
f (s∗t (Ξ00 ) + lt ).
Thus, we have established the convexity of Ψ1 (Ξ) over the
set of Ξ. Therefore, we have
E [Ψ1 (Ξ)] ≥ Ψ1 (E[Ξ]).
et
X
n=t
(65)
t=1
On the other hand, based on
1, · · · , e1 , j = t, · · · , et , we
d˜t−1
− d˜t−1
n
t−1 +
Ẑ =
f (s∗t (Ξ0 ) + lt ) + (1 − λ)
D. Proof of Proposition 2:
PT
st ≤ n=t d˜tn holds for all stage t. On the other hand, we
have
Pj̄
f (s∗t (Ξ000 ) + lt )
t=1
≤λ
(67a)
!
Pe1 ˜1 PT Pet j PT
t=1 νt
j=t µt +
t=2
t=1 dt +
.
T
(67b)
This completes the proof.
t=1
for all λ ∈ [0, 1]. On the P
other hand, let s∗t (Ξ000 ) be the optimal
T
solution that minimizes t=1 f (st + lt ) subject to (4b) - (4c)
000
with Ξ = Ξ . Then,
T
X
= Tf
Ẑ =
f (st (Ξ000 ) + lt )
T
t=1
s̄∗t
This contradicts with the fact that the
is the optimal
total charging rate for objective function f¯(st + lt ). Therefore, the optimal charging solution s∗t is the same for any
strictly convex function f (st + lt ). Next, we show that optimal solution s∗t isPa load flattening solution that minimizes
T
PT
t=1 st +lt 2
) subjecting to (4b) and (4c).
t=1 (st + lt −
T
Based on the argument that s∗t is the same for any strictly
convex function fP(st + lt ), then it is equivalent to show that
T
PT
t=1 st +lt 2
) is a strictly convex function of
t=1 (st + lt −
T
st + lt . Since
P
PT
PT
i∈N di +
t=1 lt
t=1 st + lt
=
,
(62)
T
T
!
Pe1 ˜1 PT Pet j PT
j=t µt +
t=1 νt
t=2
t=1 dt +
Pk+ê+1 Pk+en n
n=k (
m=n µm + νn )
Ŷ =
.
ê + 2
If j̄ > k + ê + 1, then
(66)
the definition
of st ,P
d˜it , µjt , i
PT
e1 ˜1
have
t=1 st =
t=1 dt
Pj̄
Ŷ =
=
+
(72)
n=k (
Pk+en
This completes the proof.
µn
m + νn )
.
j̄ − k + 1
m=n
(73)
we have
F. Proof of Lemma 2:
Pj−i+2
µt
(j − i + 2 − t)µt + t=1
+ν
j−i+2
Pj−i+1
(82)
(j − i + 2 − t)µt
t=1
≥
+ν
j−i+1
= ρ(i, j),
Pj−i+1
ρ(i, j + 1) =
Let ρ(i, j) denote the maximum density of [i, j]. For any
i = k, j = k + ē + 1, · · · , T , the density of interval [i, j] is
given by
Pk+ē
Pj−k Pj−2k+1
µt + t=k d˜kt + lk + (j − k)ν
ρ(i, j) = k=1 t=1
j−k+1
Pk+ē
Pj−k
(j − k + 1 − t)µt + t=k d˜kt + lk − ν
+ ν.
= t=1
j−k+1
(74)
To prove that the maximum density is achieved by setting
j = T , we only need to show ρ(i, j) is a non-decreasing
function of j for each given i, i.e.,
ρ(i, j) ≤ ρ(i, j + 1), ∀k + ē + 1 ≤ j ≤ T − 1.
Since
Pj−k
t=1 (j
− k + 1 − t)µt +
j−k
Pk+ē ˜k
t=k dt
≤
j−k+1
X
µt ,
t=1
(75)
(76)
we have
ρ(i, j + 1)
Pj−k
P
˜k Pj−k+1 µt + lk − ν
(j − k + 1 − t)µt + k+ē
t=k dt +
t=1
+ν
= t=1
j−k+1
Pj−k
Pk+ē ˜k
(j − k + 1 − t)µt + t=k dt + lk − ν
+ν
≥ t=1
j−k
=ρ(i, j),
(77)
which implies (75). Hence, Y is the maximum density of
[k, j], j = k + ē + 1, · · · , T . Next, we show that Z is the
maximum density of [k + 1, T ]. For any k + 1 ≤ i ≤ j ≤ T ,
the density of interval [i, j] is given by
Pj−i+1 Pj−i+2−k
µt
t=1
ρ(i, j) = k=1
+ν
j−i+1
(78)
Pj−i+1
t=1 (j − i + 2 − t)µt
=
+ ν.
j−i+1
To prove that the maximum density is achieved by setting
i = k + 1, j = T , we only need to show ρ(i, j) is a nondecreasing function of j for each given i, i.e.,
(79)
and a non-increasing function of i for each given j, i.e.,
ρ(i, j) ≥ ρ(i + 1, j), ∀k + 1 ≤ i + 1 ≤ j ≤ T.
which implies (79). On the other hand, as
Pj−i
j−i+1
X
t=1 (j − i + 1 − t)µt
≤
µt , ∀k + 1 ≤ i ≤ j ≤ T,
j−i
t=1
(83)
then
Pj−i
(j − i + 1 − t)µt
ρ(i + 1, j) = t=1
+ν
j−i
Pj−i
Pj−i+1
µt
t=1 (j − i + 1 − t)µt +
t=1
+ν
≤
j−i+1
= ρ(i, j),
(84)
which implies (80). This completes the proof.
R EFERENCES
k + ē + 1 ≤ j ≤ T − 1,
ρ(i, j) ≤ ρ(i, j + 1), ∀k + 1 ≤ i ≤ j ≤ T − 1,
t=1
(80)
On one hand, since
Pj−i+1
j−i+2
X
t=1 (j − i + 2 − t)µt
≤
µt , ∀k + 1 ≤ i ≤ j ≤ T,
j−i+1
t=1
(81)
[1] W. Tang and Y. J. Zhang, “Online electric vehicle charging control
with multistage stochastic programming,” in IEEE 48th Annu. Conf.
Information Sciences and Systems (CISS), pp. 1-6, Mar. 2014.
[2] J. A. P. Lopes, F. J. Soares, and P. M. R. Almeida, “Integration of electric
vehicles in the electric power system,” Proc. of the IEEE, vol. 99, no. 1,
pp. 168-183, 2011.
[3] E. Sortomme, M. M. Hindi, S. D. J. MacPherson, and S. S. Venkata,
“Coordinated charging of plug-in hybrid electric vehicles to minimize
distribution system losses,” IEEE Trans. on Smart Grid, vol.2, no.1, pp.
198-205, 2011.
[4] Z. Ma, D. Callaway, and I. Hiskens, “Decentralized charging control of
large populations of plug-in electric vehicles,” IEEE Trans. on Control
Systems Technology, vol. 21, no. 1, pp. 67-78, 2013.
[5] W. Tang, S. Bi, and Y. J. Zhang, “Online coordinated charging decision
algorithm for electric vehicles without future information,” IEEE Trans.
on Smart Grid, vol. 5, no. 6, pp. 2810 - 2824, 2014.
[6] Y. He, B. Venkatesh, and L. Guan, “Optimal scheduling for charging and
discharging of electric vehicles,” IEEE Trans. on Smart Grid, vol. 3, no.
3, pp. 1095-1105, 2012.
[7] L. Gan, U. Topcu, and S. H. Low, “Optimal decentralized protocol for
electric vehicle charging,” IEEE Trans. on Power System, vol.28, iss. 2,
pp. 940-951, 2012.
[8] J. R. Birge and F. Louveaux, Introduction to Stochastic Programming,
New York: Springer, 1997.
[9] B. Defourny, D. Ernst, and L. Wehenkel, “Multistage stochastic programming: a scenario tree based approach to planning under uncertainty,” Decision Theory Models for Applications in Artificial Intelligence: Concepts
and Solutions, 2011.
[10] F. Maggioni and S. Wallace, “Analyzing the quality of the expected value
solution in stochastic programming,” Annals of Operations Research, pp.
37-54, 2012.
[11] R. Leou, C. Su, and C. Lu, “Stochastic analyses of electric vehicle
charging impacts on distribution network,” IEEE Trans. on Power System,
vol. 29, no. 3, pp. 1055-1063, May 2014.
[12] L. Rao and J. Yao, “SmartCar: smart charging and driving control
for electric vehicles in the smart grid,” IEEE Global Communications
Conference (GLOBECOM), pp. 2709-2714, Dec. 2014.
[13] N. Chen, L. Gan, S. H. Low, and A. Wierman, “Distributional analysis
for model predictive deferrable load control,” 53rd IEEE Conference on
Decision and Control (CDC), pp. 6433-6438, Dec. 2014.
[14] L. Gan, A. Wierman, U. Topcu, N. Chen, and S. H. Low, “Realtime deferrable load control: handling the uncertainties of renewable generation,”
in Proceedings of the fourth international conference on Future energy
systems (ACM e-Energy) , pp. 113-124, May 2013.
[15] S. Bansal, M. N. Zeilinger, and C. J. Tomlin, “Plug-and-play model
predictive control for electric vehicle charging and voltage control in
smart grids,” 53rd IEEE Conference on Decision and Control (CDC), pp.
5894-5900, 2014.
[16] G. Li and X. Zhang, “Modeling of Plug-in Hybrid Electric Vehicle
Charging Demand in Probabilistic Power Flow Calculations,” IEEE Trans.
on Vehicular Technology, vol. 63, no. 6, pp. 2600-2612, 2014.
[17] T. Zhang, W. Chen, Z. Han, and Z. Cao, “Charging scheduling of electric
vehicles with local renewable energy under uncertain electric vehicle
arrival and grid power price,” IEEE Trans. on Smart Grid, vol. 3, no.
1, pp. 492-499, 2012.
[18] I. Koutsopoulos, V. Hatzi, and L. Tassiulas, “Optimal energy storage
control policies for the smart power grid,” in Proc. of IEEE International
Conference on Smart Grid Communications (SmartGridComm), pp. 475480, 2011.
[19] L.
Huang, J.
Walrand, and K.
Ramchandran, “Optimal demand response with energy storage management,”
http://arxiv.org/pdf/1205.4297.pdf, 2012.
[20] W. Feller, An introduction to probability theory and its applications,
John Wiley & Sons, 2008.
[21] Y. Ye, Interior Point Algorithms: Theory and Analysis, WileyInterscience Press, 1997.
[22] A. Shapiro, D. Dentcheva, and A. Ruszczynski, Lectures on Stochastic
Programming: Modeling and Theory, MPS-SIAM, Philadelphia, 2009.
[23] F. Yao, A. Demers, and S. Shenker, “A scheduling model for reduced
cpu energy,” in Proc. IEEE Symp. Foundations of Computer Science, pp.
374-382, 1995.
[24] N. Bansal, T. Kimbrel, and K. Pruhs, “Speed scaling to manage energy
and temperature,” Journal of the ACM (JACM), vol. 54, no. 1, pp. 1-39,
2007.
[25] M. Alizadeh, A. Scaglione, J. Davies, and K. S. Kurani, “A scalable
stochastic model for the electricity demand of electric and plug-in hybrid
vehicles,” IEEE Trans. on Smart Grid, vol. 5, no. 2, pp. 848-860, 2014.
[26] X. Zhang and S. Grijalva, “An Advanced Data Driven Model for
Residential Electric Vehicle Charging Demand,” technique report, Georgia
Institute of Technology, 2015.
[27] S. Chen and L. Tong, “iEMS for large scale charging of electric vehicles
architecture and optimal online scheduling,” in Proc. IEEE Int. Conf.
Smart Grid Commun. (SmartGridComm), pp. 629-634, Nov. 2012.
[28] A. Santos, A. N. McGuckin, H. Y. Nakamoto, D. Gray, and S. Lis,
Summary of travel trends: 2009 national household travel survey, Federal
Highway Administration, Washington, DC, 2011.
[29] James W. May and Matt Mattila, “Plugging In: A Stakeholder Investment Guide for Public Electric-Vehicle Charging Infrastructure,” Rocky
Mountain Institute, 2009.
[30] S. Boyd and L. Vandenberghe, Convex optimization, Cambridge university press, 2004.
| 3cs.SY
|
arXiv:1802.02074v1 [math.ST] 6 Feb 2018
ORIGINAL ARTICLE
Splitting models for multivariate count data
Pierre Fernique1∗
|
Jean Peyhardi2∗
|
Jean-Baptiste
Durand3
1 UMR AGAP, Cirad Inra, Univ.
Montpellier,
Montpellier, 34000, France
2 IGF, Univ.
Montpellier, CNRS, INSERM,
Montpellier, France.
3 Univ.
Grenoble Alpes, CNRS, Inria,
Grenoble INP? , LJK, 38000 Grenoble,
France ? Institute of Engineering Univ.
Grenoble Alpes
Correspondence
Jean Peyhardi, Université de Montpellier,
34000, France
Email: [email protected]
Funding information
Considering discrete models, the univariate framework has
been studied in depth compared to the multivariate one.
This paper first proposes two criteria to define a sensu stricto
multivariate discrete distribution. It then introduces the
class of splitting distributions that encompasses all usual
multivariate discrete distributions (multinomial, negative
multinomial, multivariate hypergeometric, multivariate negative hypergeometric, etc . . . ) and contains several new.
Many advantages derive from the compound aspect of splitting distributions. It simplifies the study of their characteristics, inferences, interpretations and extensions to regression
models. Moreover, splitting models can be estimated only by
combining existing methods, as illustrated on three datasets
with reproducible studies.
KEYWORDS
singular distribution, convolution distribution, compound
distribution, discrete multivariate regression
1
|
INTRODUCTION
The analysis of multivariate count data is a crucial issue in numerous application settings, particularly in the fields of
biology (Allen and Liu, 2012), ecology (Digby and Kempton, 2012) and econometrics (Winkelmann, 2013). Multivariate
count data are defined as the number of items of different categories issued from sampling within a population, which
individuals are grouped. Denoting by J this number of categories, multivariate count data analysis relies on modeling
the joint distribution of the discrete random vector Y = (Y1 , . . . ,YJ ). In genomics for instance, the data obtained from
sequencing technologies are often summarized by the counts of DNA or RNA fragments within a genomic interval (e.g.,
∗ Equally contributing authors.
1
2
PIERRE FERNIQUE ET AL .
RNA seq data). The most usual models in this situation, are multinomial and Dirichlet multinomial regression to take
account of some environmental covariate effects on these counts. In this way, Xia et al. (2013) and Chen and Li (2013)
studied the microbiome composition (whose output are J bacterial taxa counts) and Zhang et al. (2017) studied the
expression count of J exon sets.
However, the multinomial and Dirichlet multinomial distributions are not appropriate for multivariate count data
n
o
Í
because of their support: the discrete simplex ∆n = y ∈ ÎJ : jJ=1 y j = n . This kind of distributions are said to be
singular and will be denoted by S∆n (θ). The parameter n being related to the support, is intentionally noted as an
index of the distribution, distinguishing it from other parameters θ used to described the probability mass function
(pmf). In this case, any component Yj is deterministic when the J − 1 other components are known. Note that initially,
singular versions of some multivariate distributions have been defined by Patil (1968) and Janardan and Patil (1972).
But these distinctions were unheeded until now, leading to misuse of these distributions (Zhang et al., 2017). Therefore,
a distribution will be considered as a J -multivariate distribution if
1.
the dimension of its support is equal to the number of variables (i.e., dim{Supp(Y )} = J ).
If no confusion could arise, J will be omitted in the former notations. Another problem that occurs when defining
multivariate distributions is the independence relationships between components Y1 , . . . ,YJ . For instance, the multiple
Poisson distribution described by Patil and Bildikar (1967), involves J mutually independent variables. Therefore, a
multivariate distribution will be considered as a sensu stricto multivariate distribution if:
2.
its probabilistic graphical model is connected (i.e., there is a path between every pair of variables).
Additionally, such a distribution is considered as an extension of a given univariate distribution if:
3.
all the univariate marginal distributions belong to the same family (extension),
4.
all the multivariate marginal distributions belong to the same family (natural extension).
Even if a singular distribution is not a sensu stricto J -multivariate distribution, it is very powerful as soon as the
parameter n is considered as a random variable. It then becomes a map between spaces of univariate and multivariate
distributions. Assuming that n follows an univariate distribution L(ψ) (e.g., binomial, negative binomial, Poisson etc . . . ),
the resulting compound distribution, denoted by S∆N (θ) ∧ L(ψ), is called splitting distribution. Under mild hypothesis,
N
splitting distributions can be considered as sensu stricto multivariate distributions. They include all usual multivariate
discrete distributions and several news. Many advantages derive from the compound aspect of splitting distributions.
The interpretation is simply decomposed into two parts: the sum distribution (intensity of the distribution) and the
singular distribution (repartition into the J components). The log-likelihood can also be decomposed according to these
two parts and thus easily computed. All usual characteristics (support, pmf, expectation, covariance and probability
generative function (pgf)) are also easily obtained using this decomposition. Finally, the generalization to regression
models is naturally made by compounding a singular regression by an univariate regression.
This article is organized as follows. In Section 2 notations used all along the paper are introduced. The definition of
singular distributions is used as a building block to introduce splitting distributions. Positive and symmetric singular
distributions are introduced, easing respectively the study of criteria 1-2 and 3-4 for resulting splitting distributions. In
Section 3 the subclass of symmetric convolution distributions is introduced (e.g., the generalized Dirichlet multinomial
is a singular distribution but is not symmetric). Additive and proportional convolution distributions are then introduced
PIERRE FERNIQUE ET AL .
3
Positive
Symmetric
Multinomial
×
×
Dirichlet multinomial
×
×
Multivariate hypergeometric
×
Generalized Dirichlet multinomial
TA B L E 1
×
Properties of four singular distributions.
to simplify respectively the calculation of marginal distributions and the inference procedure. Sections 4 and 5 focus
on splitting distributions obtained with the multinomial and the Dirichlet multinomial distributions since they are
both positive and additive (e.g., the multivariate hypergeometric is additive convolution distribution but not positive).
This leads us to precisely describe fifteen multivariate extensions (among which five are natural extensions) of usual
univariate distributions giving their usual characteristics. In Section 6, the splitting distributions are extended to
regression models. In Section 7 a comparison of these regression models on two benchmark datasets and an application
on a mango tree dataset are proposed.
2
|
SPLITTING DISTRIBUTIONS
Notations
All along the paper focus will be made only on non-negative discrete distributions (and regression models).
ÍJ
j =1 Yj denote the sum of the
For notational convenience, the term discrete will therefore be omitted. Let |Y | =
random vector Y and assume that |Y | ∼ L(ψ). Let y = (y 1 , . . . , y J ) ∈ ÎJ denote an outcome of the random vector
Y . Let PB (A) denotes the conditional probability of A given B. Let E | Y | (Y ) and Cov | Y | (Y ) denote respectively the
conditional expectation and covariance of the random vector Y given the sum |Y |. Let NnJ = y ∈ ÎJ : |y | ≤ n
denote the discrete corner of the hypercube. If no confusion could arise, J will be omitted in the former notations. Let
ÎJ
n
y = n!/(n − |y |)! j =1 y j ! denote the multinomial coefficient defined for y ∈ Nn . This notation replaces the usual
Î
notation yn = n!/ jJ=1 y j ! which is defined only for y ∈ ∆n . Let (a)n = Γ(a + n)/Γ(a) denote the Pochhammer symbol
ÎJ
and B(α) =
j =1 Γ(αj )/Γ( |α|) the multivariate beta function. Let
J
0
0
2 F 2 {(a, a ); b; (c, c ); s}
Õ (a) | y | (a 0 ) | y |
=
y ∈ÎJ
ÎJ
j =1 (b j ) y j
(c) | y | (c 0 ) | y |
y
J s j
Ö
j
j =1
yj !
denote a multivariate hypergeometric function. Remark that a 0 = c 0 lead to 1J F1 (a; b; c; s) the Lauricella’s type D function
(Lauricella, 1893). Moreover, if J = 1 then it turns out to be the usual Gauss hypergeometric function 2 F1 (a; b; c; s) or
the confluent hypergeometric 1 F1 (b; c; s).
Definitions
The random vector Y is said to follow a splitting distribution1 if there exists a singular distribution S∆n (θ)
and an univariate distribution L(ψ) such that Y follows the compound distribution S∆N (θ) ∧ L (ψ). The pmf is then
N
given by P (Y = y) = P (|Y | = |y |)P | Y |=| y | (Y = y) assuming that |Y | follows L(ψ) and Y given |Y | = n follows
S∆n (θ). Moreover, a singular distribution is said to be:
1 It is named splitting distribution since an outcome y
∈ Î of the univariate distribution L(ψ) is split into the J components.
4
PIERRE FERNIQUE ET AL .
•
positive, if its support is the whole simplex,
•
symmetric, if it is invariant under any permutation of its components.
If the singular distribution is symmetric, then it is possible to define a non-singular extension2 having as support a
subset of Nn (if the singular distribution is positive, then the support of the non-singular version is exactly Nn ). Such a
distribution for the random vector Y , denoted by SNn (θ, γ), is such that (Y , n − |Y |) ∼ S∆J +1 (θ, γ). Remark that all
n
univariate distributions bounded by n (denoted by Ln (θ)) are non-singular distributions. The variable Y is said to follow
a damage distribution3 if there exists a bounded distribution Ln (θ) and a distribution L(ψ) such that Y follows the
compound distribution LN (θ) ∧ L (ψ).
N
Examples The multinomial, the Dirichlet multinomial (also known as the multivariate negative hypergeometric), the
multivariate hypergeometric and the generalized Dirichlet multinomial distributions are four examples of singular
distributions (see Table 1). Contrarily to others, the multivariate hypergeometric distribution with parameters n ∈ Î
and k ∈ ÎJ , is not positive since its support is the intersection of the simplex ∆n and the hyper-rectangle k = {y ∈
ÎJ : y 1 ≤ k 1 , . . . , y J ≤ k J }. Contrarily to others, the generalized Dirichlet multinomial distribution is not symmetric.
2.1
|
Support
Splitting distributions as sensu stricto multivariate extensions
Firstly, let us remark that a singular distribution could be viewed as particular splitting distribution if the sum
follows a Dirac distribution, i.e. S∆n (θ) = S∆N (θ) ∧ 1n . Assume that the dimension of a set A ⊆ ÎJ is defined as the
N
dimension of the smaller Ò-vectorial space including A. The dimension of the support of a positive splitting distribution
is depending on the support of the sum distribution as follows:
dim Supp S∆N (θ) ∧ L(ψ)
N
0
=
J −1
J
ifL(ψ) = 10 ,
ifL(ψ) = 1n with n ∈ Î∗ ,
otherwise.
Therefore all positive splitting distributions are considered as multivariate distribution (the criterion 1 holds) when the
sum is not a Dirac distribution (only non-Dirac distributions will therefore be considered all along the paper). For nonpositive splitting distributions, the dimension also depends on the support of the singular distribution. In order to study
the support of splitting distributions in a general way, the singular distribution is thus assumed to be positive (e.g., the
case of the multivariate hypergeometric splitting distributions is omitted). The support of Y can be expressed in terms
of the sum support. If supp ( |Y |) = {a, · · · , b } (with a ∈ Î, b ∈ Î ∪ {∞} and a < b) then supp (Y ) = Nb \ Na−1 . But, for
any j = 1, . . . , J , the marginal support is supp Yj = Nb . It can thus be interesting to consider a shifted sum |Y | = Z + δ,
with Z ∼ L(ψ), supp (Z ) ⊆ {b + 1, . . . , ∞} and δ ∈ {−b, . . . , ∞}. As illustrated on Figure 1, such distributions can be
useful for splitting distributions since they enable their supports to be modified. In particular, for sum distributions
with a = 0 (e.g., binomial, negative binomial and Poisson distributions), δ = 1 enable to remove the null vector from
the joint support while it is kept in marginal supports. Conversely, for sum distributions with a = 1 (e.g., geometric and
logarithmic series distributions), δ = −1 enable to add the null vector to the joint support. If multivariate zeros are from
one structural source, it is possible to use splitting models with shifted sums to extend hurdle models described by
Mullahy (1986).
2 The symmetry ensures that the choice of the last category to complete the vector, has no impact on the distribution.
3 It is named damage distribution since an outcome y
∈ Î of the distribution L(ψ) is damaged into a smaller value.
PIERRE FERNIQUE ET AL .
a+δ
(a)
5
b+δ
a
(b)
a+δ
b
b+δ
(c)
F I G U R E 1 Illustration of the effect of shifting the sum distribution on the support of positive splitting distributions.
The support of the sum distribution is {a, · · · , b }. Vectors of integers that are part of the shifted support are
represented by dark circles. Vectors of integers that were part of the non-shifted support but are not part of the shifted
support are represented by white circles. (a) The splitting distribution support with a negative shift (i.e., δ < 0). (b) The
splitting distribution support without any shift. (c) the splitting distribution support with a positive shift (i.e., δ > 0).
Graphical model A probabilistic graphical model (or graphical model, in short) is defined by a distribution and a graph
such that all independence assertions that are derived from the graph using the global Markov property hold in the
distribution Koller and Friedman (2009). A graphical model is said to be minimal, if any edge removal in the graph
induces an independence assertion that is not held in the distribution. A graphical model is said to be connected if there
exists a path containing all its vertices (i.e., there is no pair of independent variables). This is a necessary condition
(criterion 2) to obtain a sensu stricto multivariate distribution. Peyhardi and Fernique (2017) characterized the graphical
model of multinomial and Dirichlet multinomial splitting distributions according to the sum distribution. But, when
the graphical model is unknown, it is sufficient to show that covariances are strictly positive to obtain a graph with at
least one path between every pair of random variables. Moments can be derived using the law of total expectation
E (Y ) = E E | Y | (Y ) and covariance Cov (Y ) = E Cov | Y | (Y ) + Cov E | Y | (Y ) . This method could be used for
multivariate hypergeometric and generalized Dirichlet multinomial splitting distributions since their graphical models
have not yet been characterized. Similarly, the pgf of splitting distributions can be obtain from the pgf of the singular
distribution since G Y (s) = E Ḡ Y (s) where s = (s 1 , . . . , s J ) and Ḡ Y denotes the pgf of Y given the sum |Y |.
Marginal distributions Splitting distributions that are sensu stricto multivariate distributions (i.e., with criteria 1 and
2) are not necessarily multivariate extensions. To be considered as a multivariate extension of a specific family, marginal
distributions ofYj must belong to this family. The symmetry of the singular distribution is a sufficient condition to obtain
a multivariate extension. In fact, marginals of symmetric splitting distributions follow the same damage distribution but
with different parameters (see Appendix A for details).
6
PIERRE FERNIQUE ET AL .
Distribution
Notation
Parameter Inference
Binomial
Bn (p)
See Blumenthal and Dahiya (1981)
Negative binomial
N B (r , p)
See Dai et al. (2012)
Poisson
P (λ)
See Johnson et al. (1993)
Logarithmic series
L (p)
See Johnson et al. (1993)
Beta binomial
β Bn (a, b)
See Tripathi et al. (1994); Lesnoff et al. (2012) for n known
Beta negative binomial
β N B (r , a, b)
See Irwin (1968)
Beta Poisson
β λ P (a, b)
See Gurland (1958); Vu et al. (2016)
TA B L E 2
|
2.2
References of parameter inference procedures for seven usual univariate discrete distributions.
Inference
If the parameters θ and ψ are unrelated, the log-likelihood of the splitting distribution, denoted by L (θ, ψ; y), can be
decomposed into log-likelihoods for the singular distribution and the sum distribution:
L (θ, ψ; y) = log P | Y |=| y | (Y = y) + log {P ( |Y | = |y |)} ,
L (θ, ψ; y) = L (θ; y) + L (ψ; |y |) .
(1)
Therefore, the maximum likelihood estimator (maximum likelihood estimator (MLE)) of a splitting distribution with
unrelated parameters can be obtained separately using respectively the MLE of the singular distribution and the MLE of
the sum distribution. Hence, with C estimators of singular distributions and L estimators of univariate distributions,
one is able to estimate C × L multivariate distributions, with time complexity in O (C + L). Let us remark that decomposition (1) stays true for decomposable scores such as AIC and BIC. Hence model selection using decomposable scores is
also reduced to two separate model selection problems and has the same linear time and space complexity. To limit the
scope of this paper, focus will be given only to parametrization and inference of singular distributions. Sum distributions
considered in this paper are usual power series distributions and some beta compound distributions (see Appendix C
for definitions and Table 2 for some inference references).
3
|
CONVOLUTION SPLITTING DISTRIBUTIONS
In order to study thoroughly the graphical models and the marginals of splitting distributions, additional assumptions
are necessary concerning the parametric form of the singular distribution. Convolution splitting distributions have been
introduced by Shanbhag (1977) for J = 2 and extended by Rao and Srivastava (1979) for J ≥ 2, but were only used as a
tool for characterizing univariate discrete distributions L(ψ). We here consider convolution splitting distributions as a
general family of multivariate discrete distributions.
PIERRE FERNIQUE ET AL .
7
Definition The random vector Y given |Y | = n is said to follow a (singular) convolution distribution if there exists a
non-negative parametric sequence a := {a θ (y )} y ∈Î such that for all y ∈ ∆n we have
P | Y |=n (Y = y) =
J
1 Ö
a θj (y j ),
c θ (n)
j =1
where c θ denotes the normalizing constant (i.e., the convolution of a θ1 , . . . , a θJ over the simplex ∆n ). This convolution
distribution, characterized by the parametric sequence a = {a θ (y )} y ∈Î , is denoted by C∆n (a; θ) where θ = (θ1 , . . . , θJ ) ∈
ΘJ . If the distribution of |Y | belongs to a family of univariate parametric distribution L(ψ) then, we note this compound
distribution as follows Y ∼ C∆N (a; θ) ∧ L(ψ). Remark that a convolution distribution is positive if and only if a θ (y ) > 0
N
for all θ ∈ Θ and all y ∈ Î. Moreover, a convolution distribution is said to be:
•
additive, if c (θ,θ0 ) (n) = a θ+θ0 (n) for all (θ, θ 0 ) ∈ Θ2 and all n ∈ Î,
•
proportional, if C∆N (a; θ) = C∆N (a; λ · θ) for all θ ∈ ΘJ and all λ ∈ Θ.
Thanks to the symmetry, the non-singular extension denoted by CNn (a; θ, γ) is well defined. Its pmf is given by
J
P (Y = y) =
Ö
1
a γ (n − |y |)
a θj (y j ),
c θ,γ (n)
j =1
for all y ∈ Nn . If the non-singular convolution distribution is univariate then it is denoted by Cn (a; θ, γ). The random
variable Y is said to follow a convolution damage distribution if there exists a latent variable N such that Y given
N = n follows Cn (a; θ, γ) for all n ∈ Î. Moreover if N ∼ L (ψ), we denote this compound distribution as follows:
Y ∼ CN (a; θ, γ) ∧ L (ψ).
N
Examples
Remark that these convolutions distributions are symmetric by construction. The generalized Dirichlet
multinomial distribution thus not belong to this family. The three other singular distributions are additive convolution
distributions; details for the multinomial and Dirichlet multinomial distributions are given in the next two sections.
Additivity is necessary to easily determine the marginal distributions of corresponding splitting distributions. Among
those four distributions, only the multinomial distribution is a proportional convolution distribution (details are given
in the next section); see Table 3 for summary of these properties. For the univariate case, the binomial, beta binomial
(negative hypergeometric) and hypergeometric distributions are three non-singular convolutions distributions.
3.1
|
Properties of convolution splitting distributions
Graphical model
Until recently, only marginal independence have been studied through the well-known Rao-Rubin
condition (Radhakrishna Rao and Rubin, 1964). Peyhardi and Fernique (2017) generalized this condition for conditional
independence and deduced the graphical models for positive convolution distributions under mild hypotheses. In this
case, it has been shown that there exists only one univariate distribution L(ψ) = (p k∗ )k ∈Î such that the minimal graphical
model for Y is empty. This distribution belongs to the family of power series distributions since there exists some
α > 0 such that p k∗ = p 0∗ α k c θ (k )/c θ (0), for all k ∈ Î. In this case and only this case, the resulting convolution splitting
distribution is not a senso stricto multivariate distribution. For all other univariate distributions L(ψ) , (p k∗ )k ∈Î , the
graphical model is complete (the criterion 2 holds) and thus the resulting convolution splitting distribution is a sensu
8
PIERRE FERNIQUE ET AL .
stricto multivariate distribution. This characterization of the graphical model (Peyhardi and Fernique, 2017, Theorem 4)
holds for multinomial and Dirichlet multinomial splitting distributions. But if the convolution is not positive (e.g., the
multivariate hypergeometric distribution), this characterization does not hold anymore and stays an open issue.
Derived distributions Several stability properties hold for additive convolution splitting distributions including
results of Patil (1968); Janardan and Patil (1972); Xekalaki (1986) as particular cases.
Theorem 1 Let Y follow an additive convolution splitting distribution Y ∼ C∆N (a; θ) ∧ L(ψ) then:
N
1.
The marginal sum |YI | follows the damage distribution CN (a; |θI |, |θ−I |) ∧ L(ψ).
2.
The subvector YI given |YI | = n follows the singular convolution distribution C∆n (a; θI ).
3.
The subvector YI follows the convolution splitting damage distribution
N
C∆N (a; θI ) ∧ CN 0 (a; |θI |, |θ−I |) ∧0 L(ψ) .
N
4.
N
The subvector YI given Y−I = y−I follows the convolution splitting truncated and shifted distribution
h
i
C∆N (a; θI ) ∧ T S | y−I | { L(ψ)} .
N
5.
The subvector YI given Y J = y J follows the convolution splitting truncated and shifted damage distribution
C∆N (a; θI ) ∧ T S | y J | CN 0 (a; |θI∪J |, |θ−I∪J |) ∧0 L(ψ) .
N
N
where I ⊂ {1, . . . , J }, −I = {1, . . . , J } \ I, J ⊂ −I and T S δ { L(ψ)} denotes the truncated and shifted distribution L(ψ)
with parameter δ ∈ Î (i.e., X ∼ T S δ { L(ψ)} means that P (X = x ) = P Z ≥δ (Z = δ + x ) with Z ∼ L(ψ)).
Corollary 2 An additive convolution splitting distribution C∆N (a; θ) ∧ L(ψ) is a natural multivariate extension of L(ψ) if
N
the latter is stable under the damage process CN (a; |θI |, |θ−I |) ∧ (·) for any subset I ⊂ {1, . . . , J }. Univariate marginals are
thus following the distribution L(ψj ) for some ψj ∈ Ψ.
N
Corollary 3 The non-singular version of an additive convolution distribution is a specific convolution splitting distribution:
C∆N (a; θ) ∧ Cn (a; |θ | , γ) = CNn (a; θ, γ) .
N
The proof of the three previous result are given in Appendix A. As a consequence from Corollary 3, the univariate distribution L(ψ) = Cn (a; |θ | , γ) is considered as the canonical case for a given convolution distribution. The parameters of
the convolution distribution and the univariate bounded distribution are independent given the sum |θ |. This property can be benefited from into parameter inference procedures. Such dependence disappears when the convolution
distribution is proportional, implying that inference procedure of the non-singular distribution is separable into two independent parts. Note that univariate marginals follow the convolution damage distribution: Yj ∼ CN (a; θj , |θ−j |) ∧ L(ψ).
It is thus easy to highlight multivariate extensions studying these marginal distributions.
N
PIERRE FERNIQUE ET AL .
9
Positive
Additive
Proportional
Multinomial
×
×
×
Dirichlet multinomial
×
×
Multivariate hypergeometric
TA B L E 3
4
|
×
Properties differentiating three convolutions distributions.
MULTINOMIAL SPLITTING DISTRIBUTIONS
In this section the multinomial distribution is introduced as a positive, additive and proportional convolution distribution.
Then, the general case of multinomial splitting distributions (i.e., for any sum distribution L(ψ)) is addressed. For six
specific sum distributions, the usual characteristics of multinomial splitting distributions are described in Tables 4 and 5.
Finally, the canonical case of binomial sum distribution is detailed, with particular emphasize on parameter inference.
Multinomial distribution Let a θ (y ) = θ y /y ! be the parametric sequence that characterizes the multinomial distribution as a convolution distribution. It is positive since θ y /y ! > 0 for all θ ∈ Θ = (0, ∞) and all y ∈ Î. It is additive, as a
Í
consequence from the binomial theorem: (θ + γ)n = ny =0 yn θ y γ n−y . It implies, by induction on n, that the normalizing
constant is c θ (n) = a | θ | (n) = |θ | n /n!. The pmf of the singular multinomial distribution is thus given by
P | Y |=n (Y = y) =
Ö
J
θj y j
n
· 1∆n (y),
|θ |
y
(2)
j =1
and is denoted by M∆n (θ) with θ ∈ (0, ∞)J . This convolution is proportional, implying that the equivalence class of
distributions {M∆n (λ · θ), λ ∈ (0, ∞)} can be summarized by the representative element M∆n (π) where π =
The parameters vector π lies in the continuous simplex c∆ = {π ∈
(0, 1)J
1
|θ |
· θ.
: |π | = 1} and the pmf reduces to its usual
form. The pmf of the non-singular multinomial distribution, denoted by MNn (θ, γ), is given by
P (Y = y) =
n−| y | Ö
yj
J
θj
n
γ
· 1Nn (y).
y |θ | + γ
|θ | + γ
j =1
In the same way there exists a representative element MNn (π ∗ , γ ∗ ) with (π ∗ , γ ∗ ) ∈ (0, 1)J +1 such that |π ∗ | + γ ∗ = 1.
Knowing this constraint, the last parameter γ ∗ = 1 − |π ∗ | could be let aside to ease the notation and obtain MNn (π ∗ )
where the parameters vector π ∗ lies in the continuous corner of the open hypercube cN = {π ∗ ∈ (0, 1)J : |π ∗ | < 1}. As
a particular case of the non-singular multinomial distribution (when J = 1), the binomial distribution is finally denoted
by Bn (p) with p ∈ (0, 1) (which is also the representative element of its class). Even if this new definition of multinomial
distributions based on equivalence classes seems quite artificial, it is necessary to obtained all the properties that hold
for convolution splitting distributions. For instance Corollary 3 becomes
M∆N (π) ∧ Bn (p) = MNn (p · π) ,
N
(3)
with representative element notations (see Appendix A for the proof). A second point that is important to highlight is
the difference between singular and non-singular multinomial distributions. Contrarily to the widely held view that the
10
PIERRE FERNIQUE ET AL .
multinomial distribution is the extension of the binomial distribution (Johnson et al., 1997), only the non-singular one
should be considered as the natural extension. In fact, criterion 4 does not hold for the singular multinomial distribution
(multivariate marginals follow non-singular multinomial distributions). Moreover, when confronted to multivariate
counts, the usual inference of multinomial distributions (Johnson et al., 1997; Zhang et al., 2017) is the inference of
singular multinomial distributions such that [n ∈ Î the random vector Y given |Y | = n follows M∆n (π). Such a point of
view therefore limits the possibility of comparing these distributions to other classical discrete multivariate distributions
such as the negative multinomial distribution (Johnson et al., 1997) or the multivariate Poisson distributions (Karlis and
Meligkotsidou, 2005) used for modeling the joint distribution of Y . The singular multinomial distribution must not be
considered as a J -multivariate distribution since criterion 1 does not hold.
Properties of multinomial splitting distributions Let Y follow a multinomial splitting distribution M∆N (π) ∧ L(ψ).
The criteria 1 and 3 hold, as a consequence from the positivity and the symmetry. The pmf is given by
P (Y = y) = P ( |Y | = |y |)
N
J
|y | Ö y j
πj .
y
(4)
j =1
The expectation and covariance are given by
E (Y ) = µ 1 · π,
Cov (Y ) =
(5)
µ 1 · diag(π) + (µ2 − µ 12 )
· ππ ,
t
(6)
where π t denotes the transposition of the vector π and µ k denotes the factorial moments of the sum distribution for
k = 1, 2. Moreover, using the pgf formula of the singular multinomial distribution (Johnson et al., 1997) we obtain the
pgf of multinomial splitting distributions as
n
o
G Y (s) = E (π t s) | Y | = G | Y | π t s ,
(7)
where G | Y | denote the pgf of the sum distribution. The graphical model is characterized by the following property.
Property 1 (Peyhardi and Fernique, 2017) The minimal graphical model for a multinomial splitting distribution is empty if the
sum follows a Poisson distribution and is complete otherwise.
Therefore, all multinomial splitting distribution are sensu stricto multivariate distributions (the criteria 2 holds) except
when the sum follows a Poisson distribution. As a consequence from additivity, Theorem 1 holds and yields the marginals
distributions in Corollary .
Corollary 4 Let Y follow a multinomial splitting distribution, Y ∼ M∆N (π) ∧ L(ψ). Then, the marginals follow the binomial
damage distribution Yj ∼ BN (πj ) ∧ L(ψ). Moreover, for y ∈ Î
N
N
y
πj (y )
G (1 − πj ),
P Yj = y =
y ! |Y |
(8)
(y )
where G | Y | denotes the y -th derivative of the pgf of the sum distribution L(ψ).
Using equation (8), it is easy to study the action of the binomial distribution among the set of univariate distribution.
Assume that L(ψ) is a power series distribution denoted by P S D {g (α)}. It can be seen by identifiability that the
PIERRE FERNIQUE ET AL .
11
resulting distribution of Y is exactly the multivariate sum-symmetric power series distribution (MSSPSD) introduced
by Patil (1968) with M∆N (π) ∧ P S D {g (α)} = MSSPSD{α · π }. The non-singular multinomial distribution, the negative
N
multinomial distribution and the multivariate logarithmic series distribution are thereby encompassed in multinomial
splitting distributions (see Table 4).
Assume now that L(ψ) is a standard beta compound distribution. We obtaine three new multivariate distributions
which are multivariate extensions of the non-standard beta binomial, non-standard beta negative binomial and beta
Poisson distributions (see Table 5 for details about these three multivariate distributions and Appendix C for definitions
of the non-standard beta binomial and the non-standard beta negative binomial distributions). All the characteristics
of these six multinomial splitting distributions (pmf, expectation, covariance, pgf and marginal distributions) have
been calculated using equations (4), (5), (6), (7), (8) according to the sum distribution L(ψ). The singular multinomial
distribution belongs to the exponential family. For any j ∈ {1, . . . , J } the MLEs πbj are given by the following closed-form
expression:
Í
π
bj = Í
Canonical case of binomial sum distribution.
y ∈N
yj
y ∈N
|y |
.
(9)
The case L(ψ) = Bn (p) is considered as the canonical case since the
binomial distribution is the univariate version of the non-singular multinomial distribution. Usual characteristics of the
multinomial splitting binomial distribution are obtained using equations (4), (5), (6), (7), (8) with L(ψ) = Bn (p). It should
be remarked that the constraint between parameters of the singular distribution and the sum distribution described in
Corollary 3 disappears in this case, as for equation (3). More generally the constraint disappears when the convolution
is proportional. Relation (3) and equation (1) allow parameters of non-singular multinomial distributions to be inferred
using a parameter inference procedure for singular multinomial and binomial distributions. In particular, this enables us
to propose joint or conditional maximum likelihood estimation of parameters n and π of the non-singular multinomial
distribution directly by combining MLE inference for the binomial distribution (Blumenthal and Dahiya, 1981) with
the closed-form expression given in equation (9). Moreover, equation (1) allows us to generalize studies related to
parameter inference in binomial distributions (Deriggi, 1983) to parameter inference in the non-singular multinomial
distributions. If the sum dataset is
•
overdispersed, the likelihood is an increasing function of n and there is no finite MLE of n,
•
underdispersed, the likelihood is either a decreasing function of n, or has an unique maximum and there is a finite
MLE of n.
Note that some previous works considered the joint estimation of parameters n and π for non-singular multinomial
distributions (Sanathanan, 1972; Royle et al., 2007) but only MLEs for specific constraints were derived.
j =1 (pπj )
ÎJ
·π
n
· diag(π) +
r
1−p
1−p π t s
p
1−p
p
1−p
Yj ∼ N B(r , pπj )4
r
r
(1 − p)r
yj
o
p
1−p
· ππ t
yj
j =1 πj
Yj ∼ L (p 0, ω) 5
ln(1−p·π t s)
ln(1−p)
−p
(1−p) ln(1−p)
·π
n
· diag(π) +
| y | ln(1−p)
−p
(1−p) ln(1−p)
y
ÎJ \ (0, . . . , 0)
ÎJ
|y |
−1
| y |+r −1
y
M L(p · π)
ÎJ
L(p)
p {1−ln(1−p)}
(1−p) ln(1−p)
Yj ∼ β πj Bn (a, b)
Marginals
· ππ t
o
y
j
j =1 πj
ÎJ
N
(b)
Yj ∼ β πj N B(r , a, b)
b
a−1
·π6
n
o
(b−1)−b−1
b
r a−1
· diag(π) + a(b+r +1)+r
· ππ t 7
(a−1)(a−2)
a (b)
t
(r +a) 2 F 1 (r , b); r + a + b; π s
r
| y |+r −1 B(r +a,| y |+b)
y,r −1
B(a,b)
ÎJ
β N B(r , a, b)
Y ∼ M∆N (π) ∧ L(ψ)
o
+ |y |, a + b + |y |; −λ)
· ππ t
Yj ∼ β πj λ P(a, b)
yj
πj
j =1 y j !
ÎJ
a
λ a+b
·π
n
o
a
b
λ a+b
· diag(π) + λ (a+b)(a+b+1)
· ππ t
t
1 F 1 a; a + b; λ(π s − 1)
(a) | y | λ | y |
(a+b) | y | 1 F 1 (a
ÎJ
β λ P(a, b)
Usual characteristics of multinomial splitting beta binomial, beta negative binomial and beta Poisson distribution.
2 F 1 {(−n, a); α
TA B L E 5
b(n−1)−a(a+b+1)
(a+b)(a+b+1)
y
j
j =1 πj
ÎJ
+ b; 1 − π t s}
a
n a+b
·π
n
a
n a+b
· diag(π) +
G Y (s)
Cov(Y )
E(Y )
n B(| y |+a,n−| y |+b)
y
B(α,b)
Nn
Supp(Y )
P (Y = y)
β Bn (a, b)
L(ψ)
Distribution
TA B L E 4
yj
N
N M(r , p · π)
N B(r , p)
Y ∼ M∆N (π) ∧ L(ψ)
Usual characteristics of multinomial splitting binomial, negative binomial and logarithmic series distribution.
Yj ∼ Bn (pπj ) 4
Marginals
G Y (s)
Cov(Y )
j =1 (pπj )
ÎJ
np · π
np · diag(π) − p · ππ t
n
1 − p + p πt s
E(Y )
(1 − p)n−| y |
Nn
n
Supp(Y )
y
MNn (p · π)
Re-parametrization
P (Y = y)
Bn (p)
L(ψ)
Distribution
12
PIERRE FERNIQUE ET AL .
PIERRE FERNIQUE ET AL .
5
|
13
DIRICHLET MULTINOMIAL SPLITTING MODELS
In this section the Dirichlet multinomial distribution is introduced as a positive and additive convolution distribution.
Then, the general case of Dirichlet multinomial splitting distributions is studied. For six specific sum distributions, the
usual characteristics of Dirichlet multinomial splitting distributions are described in Tables 6, 7, 8, and 9. Finally, the
canonical case of beta binomial sum distribution is detailed, with particular emphasis on parameter inference.
Dirichlet multinomial distribution
Let a θ (y ) =
y +θ−1
y
be the parametric sequence that characterizes the Dirichlet
multinomial distribution as a convolution distribution. It is positive since y +θ−1
> 0 for all θ ∈ Θ = (0, ∞) and all y ∈ Î.
y
Ín
n−y +γ−1
It is additive, as a consequence from the convolution identity of Hagen and Rothe: n+θ+γ−1
= y =0 y +θ−1
.
n
y
n−y
It implies, by induction on n, that the normalizing constant is c θ (n) = a | θ | (n) = n+| θn |−1 . In order to respect the
usual notation, parameter α will be used instead of θ, and thus the Dirichlet multinomial distribution will be denoted
by D M ∆n (α) with n ∈ Î and α ∈ (0, ∞)J . The non-singular Dirichlet multinomial distribution will be denoted by
D M Nn (α, b) with b ∈ (0, ∞). The beta binomial distribution will be denoted by β Bn (a, b) with (a, b) ∈ (0, ∞)2 . For
similar reasons as in the multinomial case, the non-singular Dirichlet multinomial distribution should be considered as
the natural extension of the beta binomial distribution, rather than the singular one. Note that a Dirichlet multinomial
distribution is exactly a multivariate negative hypergeometric distribution.
Dirichlet multinomial splitting distributions In this paragraph let Y follow a Dirichlet multinomial splitting distribution D M ∆N (α) ∧ L(ψ). Criteria 1 and 3 hold, as a consequence of positivity and symmetry. The pmf is given
N
by
P (Y = y) =
J
P ( |Y | = |y |) Ö y j + αj − 1
.
| y |+| α |−1
yj
|y |
(10)
j =1
The expectation and covariance are given by
µ1
· α,
|α |
|α | + 1 2
1
Cov (Y ) =
· {( |α| + 1)µ1 + µ 2 } · diag(α) + µ 2 −
µ 1 · ααt .
|α |( |α | + 1)
|α|
E (Y ) =
(11)
(12)
The pgf is given by
G Y (s) =
Õ
ÎJ
Γ (|y | + 1) P (|Y | = |y |)
y ∈ÎJ
j =1 (αj ) y j
( |α |) | y |
y
J s j
Ö
j
j =1
yj !
.
(13)
The graphical model is characterized by the following property.
Property 2 (Peyhardi and Fernique, 2017) Let Y follow a Dirichlet multinomial splitting distribution with parameter α ∈
(0, ∞)J . The minimal graphical model for Y is empty if the sum follows a negative binomial distribution with parameters ( |α |, p)
for some p ∈ (0, 1) and is complete otherwise.
Therefore, all Dirichlet multinomial splitting distribution are senso stricto multivariate distributions except when the
sum follows a negative binomial distribution N B(r , p) with the specific constraint r = |α|. Finally, due to additivity,
Property 1 can be applied to describe the marginal distributions.
14
PIERRE FERNIQUE ET AL .
Corollary 5 Let Y follow a Dirichlet multinomial splitting distribution, Y ∼ D M ∆N (α) ∧ L(ψ) with α ∈ (0, ∞)J . Then, the
marginals follow the beta-binomial damage distribution Yj ∼ β BN (αj , |α−j |) ∧ L(ψ).
N
N
Using the Fubini theorem, it can be shown that
Yj ∼ BN (π) ∧ L(ψ) ∧ β (αj , |α−j |),
π
N
(14)
since N and π are independent latent variables. Therefore, results previously obtained for the binomial damage
distributions can be used to describe the beta-binomial damage distributions. Assume that L(ψ) is a standard beta
compound distribution. Four new and two already known multivariate distributions are obtained or recovered (see
Tables 6, 7 and 8). In particular, natural multivariate extensions of three beta compound distributions are described.
The non-singular Dirichlet multinomial is recovered when L(ψ) = β Bn (a, b) with the specific constraint a = |α |. The
multivariate generalized waring distribution, introduced by Xekalaki (1986), is recovered when L(ψ) = β N B(r , a, b)
with the specific constraint r = |α |. Finally, a multivariate extension of the beta Poisson distribution is proposed when
L(ψ) = β λ P(a, b) with the specific constraint a = |α |.
Assume now that L(ψ) is a power series distributions leading to three new multivariate extensions (see Table 9).
Remark that several multivariate extensions of the same univariate distribution could be defined. For instance the
multinomial splitting beta binomial distribution M∆N (π) ∧ β Bn (a, b) and the Dirichlet multinomial splitting binomial
N
distribution D M ∆N (α) ∧ Bn (p) are two multivariate extensions of the non-standard beta binomial distribution (see
N
Tables 5 and 9). Note that the singular Dirichlet multinomial distribution does not belongs to the exponential family.
Either if |α | is known or not, MLEs αbj can be computed using various iterative methods (Minka, 2000; Sklar, 2014).
Canonical case of beta binomial sum distribution
The case L(ψ) = β Bn (a, b) is considered as the canonical case
since the beta binomial distribution is the univariate version of the non-singular Dirichlet multinomial distribution. Usual
characteristics of the Dirichlet multinomial splitting beta binomial distribution are derived from equations (10), (11),
(12), (13) and (14) with L(ψ) = β Bn (a, b). According to Corollary 3, the Dirichlet multinomial splitting beta binomial
distribution with the specific constraint a = |α | is exactly the non-singular Dirichlet multinomial distribution:
D M ∆N (α) ∧ β Bn (|α|, b) = D M Nn (α, b) .
N
(15)
The constraint a = |α | has to be taken into account in the inference procedure, either on the singular distribution or on
the sum distribution. We propose to use the first alternative since the inference procedure of a constrained Dirichlet
multinomial distribution (i.e., with a fixed sum |α |) has already been proposed by Minka (2000). The sum distribution
β Bn (a, b) can then be estimated without constraint on parameters a or b (see Table 2). Note that, if no constraint
between parameters of singular and sum distributions is assumed then the inference procedure is straightforward, since
it can be separated into two independent procedures. The resulting splitting distribution is more general, including the
non-singular Dirichlet multinomial distribution as a special case. As a consequence from equation (14), the marginals
follow beta square binomial distributions β 2 Bn (αj , |α−j |, a, b) and β Bn (αj , |α−j | + b) when the constraint a = |α| is
assumed (see Appendices B and C for definition of beta square distribution and beta square compound distributions).
Identifying these two distributions we obtain a property about the product of two independent beta distributions.
Property 3 For (a, b, c) ∈ (0, ∞)3 , let X ∼ β (a, b) and Y ∼ β (a + b, c) be two independent random variables. Then
XY ∼ β (a, b + c).
This result can be extended by induction for a product of n independent beta distributions.
na
a+b
+ |α | · diag(α)
i
o
b(a+b+n)
(a+b)(a+b+1)
−
Yj ∼
β 2 Bn
αj , |α−j |, a, b
o
− 1 · ααt
− n + 1, |α|); s}
na
| α |(a+b)
+
·α
n
hn
Yj ∼ β Bn
+
rb
a−1
TA B L E 7
o
− 1 · ααt
n |α|
| α |+b
o
i
− 1 · ααt
+
o
+ |α | · diag(α)
8
o
+ |α | · diag(α)
i
n
hn
yj !
(αj ) y j
b
a−1
Yj ∼ β N B αj , a, b
4
+
| α |b
a−1
o
+ |α| · diag(α)
i
o
− 1 · ααt 9
|α| + a + b; s)
−
(| α |+a−1)(a+b−1)
(a−1)(a−2)
(| α |+a−1)(a+b−1)
(a−1)(a−2)
·
(a) | α | J
(a+b) | α | 1 F 1 (b; α;
+
b
(| α |+1)(a−1)
j =1
ÎJ
Usual characteristics of Dirichlet multinomial splitting standard beta negative binomial distribution respectively without constraint and with r = |α|.
+ a + b, |α |); s}
rb
| α |(a−1)
Yj ∼ β 2 N B r , αj , |α−j |, a, b
−
(r +a−1)(a+b−1)
(a−1)(a−2)
Marginals
(r +a−1)(a+b−1)
(a−1)(a−2)
hn
(a)r J
(a+b)r 2 F 2 {(r , b); α; (r
n
·
G Y (s)
+
rb
| α |(| α |+1)(a−1)
b
(| α |+1)(a−1) α
(a) | α |
(b) | y |
(a+b) | α | (| α |+a+b) | y |
E(Y )
yj !
rb
| α |(| α |+1)(a−1) α
j =1
(r ) | y | (b) | y |
(a)r
(a+b)r (r +a+b) | y | (| α |) | y |
P (Y = y)
ÎJ
ÎJ
(αj ) y j
ÎJ
r = |α |
N
Supp(Y )
no constraint
Y ∼ D M ∆N (α) ∧ β N B(r , a, b)
MGWD(b, α, a)
Cov(Y )
n
| α |+b
− n + 1; s)
αj , |α−j | + b 4
−
b(| α |+b+n)
(| α |+b)(| α |+b+1)
b(| α |+b+n)
(| α |+b)(| α |+b+1)
·
(b)n J
(| α |+b)n 1 F 1 (−n; α; −b
+
y j +αj −1
)
yj
(n+|αn|+b−1)
n
(| α |+1)(| α |+b)
n |α|
| α |(| α |+b)
j =1 (
ÎJ
Re-parametrization
Constraint
n
·
b(a+b+n)
(a+b)(a+b+1)
(b)n J
(a+b)n 2 F 2 {(−n, a); α; (−b
+
na
| α |(| α |+1)(a+b)
|−1
(n+|α
)
n
n−| y |+b−1
n−| y |
Usual characteristics of Dirichlet multinomial splitting standard beta binomial distribution respectively without constraint and with a = |α | .
Distribution
TA B L E 6
Marginals
G Y (s)
Cov(Y )
E(Y )
hn
na
| α |(a+b)
·α
n B(a+| y |,b+n−| y |)
|y |
B(a,b)
P (Y = y)
Nn
y j +αj −1
)
j =1 ( y j
ÎJ
Supp(Y )
Nn
N
a = |α |
D M Nn (α, b)
no constraint
Y ∼ D M ∆N (α) ∧ β Bn (a, b)
Re-parametrization
Constraint
Distribution
PIERRE FERNIQUE ET AL .
15
(a) | y |
(a+b) | y | (| α |) | y |
λa
| α |(a+b)
P (Y = y)
E(Y )
np
|α|
E(Y )
yj !
Î
j ∈J
y
(λs j ) j
yj !
n
·
hn
Yj ∼ β λ P αj , |α−j | + b
4
+ |y |); a + b; (λ · s, −λ)}
o
λ| α |
+ |α
|+b + |α | + 1 · diag(α)
o
i
λ
− a+b
· ααt
+ |y |; a + b + |y |; −λ)
λb
(| α |+b)(| α |+b+1)
y j ! 1 F 1 (a
(αj ) y j
λb
(| α |+b)(| α |+b+1)
J +1 F {(α, a
0 1
+
λ
(| α |+1)(a+b)
·α
j =1
ÎJ
·α
(r ) | y |
(| α |) | y |
p |y |
j =1
ÎJ
yj !
(αj ) y j
N
Yj ∼ β p N B(r , αj , |α−j |)
(1 − p)r 1J F1 (r ; α; |α |; p · s)
(r −| α |)p+| α |+1
rp
1−p
| α |(| α |+1)(1−p) ·
(| α |−r )p
+ | α |(1−p) · ααt
rp
| α |(1−p)
(1 − p)r
ÎJ
N B(r , p)10
· diag(α)
Y ∼ D M ∆N (α) ∧ L(ψ)
·α
yj !
(αj ) y j
n
· (λ + |α | + 1) · diag(α) −
j =1
ÎJ
Yj ∼ β λ P(αj , |α−j |)
e −λ 0J F1 (α; |α |; λ · s)
λ
| α |(| α |+1)
λ
|α|
|y |
|y |
λ
e −λ (| α
|)
ÎJ
P(λ)
Usual characteristics of Dirichlet multinomial splitting binomial, negative binomial and Poisson distribution.
Yj ∼ β p Bn (αj , |α−j |)
p
Marginals
· ααt
(1 − p)n 1J F1 (−n; α; |α |; − 1−p · s)
np
| α |(| α |+1)
p(n+| α |)
− |α|
j =1
ÎJ
(αj ) y j
G Y (s)
TA B L E 9
i
+ |α| + 1 · diag(α)
· {(n − 1)p + |α | + 1} · diag(α)
Γ(n+1)p | y | (1−p)n−| y |
Γ(n−| y |+1)(| α |) | y |
P (Y = y)
Nn
Supp(Y )
·α
Bn (p)
L(ψ)
Distribution
Cov(Y )
k!
+
λa
a+b
λ
| α |+b
λ |y |
(a+b) | y |
ÎJ
Usual characteristics of Dirichlet multinomial splitting beta Poisson distribution respectively without constraint and with a = |α |.
Yj ∼ β λ2 P αj , |α−j |, a, b
TA B L E 8
o
k ∈Î (a+b) | y |+k (| α |) | y |
Marginals
Í
−
λb
(a+b)(a+b+1)
o
N
a = |α |
Y ∼ D M ∆N (α) ∧ β λ P(a, b)
+ |y |; a + b + |y |; −λ)
λa
t
| α |(a+b) · αα
ÎJ
(a) | y |+k j =1 (αj ) y j (−λ)k
λb
(a+b)(a+b+1)
y ∈ÎJ
n
·
hn
y j ! 1 F 1 (a
Í
+
λa
| α |(| α |+1)(a+b)
·α
j =1
ÎJ
(αj ) y j
G Y (s)
Cov(Y )
ÎJ
Supp(Y )
λ |y |
no constraint
Constraint
Distribution
λ
|α|
· ααt
o
16
PIERRE FERNIQUE ET AL .
PIERRE FERNIQUE ET AL .
17
(a)
Regression
Link function
Multinomial
πj =
Dirichlet multinomial
exp(xt βj )
,
1+exp(xt βj )
j = 1, . . . , J − 1
αj = exp(xt βj ), j = 1, . . . , J
Generalized Dirichlet multinomial
aj =
exp(xt β1,j ),
j = 1, . . . , J − 1
bj =
exp(xt β
j = 1, . . . , J − 1
2,j ),
Parameter inference
See Zhang et al. (2017)
see Zhang et al. (2017)
see Zhang et al. (2017)
(b)
Regression
Link function
Parameter inference
Poisson
λ = exp(xt β)
See McCullagh and Nelder (1989)
Binomial
Negative binomial
p=
exp(xt β )
1+exp(xt β )
See McCullagh and Nelder (1989) for n known
p = exp(xt β)
Beta Poisson
a
a+b
Beta binomial
a
a+b
Beta negative binomial
a
a+b
See Hilbe (2011)
=
exp(xt β )
1+exp(xt β )
See Vu et al. (2016)
=
exp(xt β )
1+exp(xt β )
See Forcina and Franconi (1988)
=
exp(xt β )
1+exp(xt β )
and Lesnoff et al. (2012) for n known
See Rodríguez-Avi et al. (2009)
and Saez-Castillo et al. (2017)
TA B L E 1 0
6
|
References of inference procedures for (a) singular regressions and (b) univariate regressions.
SPLITTING REGRESSION MODELS
Let us consider the regression framework, with the discrete multivariate response variable Y and the vector of Q
explanatory variables X = X 1 , . . . , XQ . The random vector Y is said to follow a splitting regression if there exists
ψ : Supp(X) → Ψ and θ : Supp(X) → Θ such that:
•
the random vector Y given |Y | = n and X = x follows the singular regression S∆n {θ (x)} for all n ∈ Î.
•
the sum |Y | given X = x follow the univariate regression L {ψ (x)}.
Such a compound regression model will be denoted by Y | X = x ∼ S∆N {θ (x)} ∧ L {ψ (x)}. The decomposition of
N
log-likelihood (1) still holds when considering covariates if parametrizations of the singular distribution and the sum
distribution are unrelated. Table 10 gives some references for parameter inference and variable selection adapted to
three singular regressions and six univariate regressions. The choice of the link function for the singular regression is
related to the symmetry of the resulting splitting regression. Using the singular multinomial regression for instance,
only the canonical link function implies the symmetry of the splitting regression (see Peyhardi et al. (2015) for details
about invariance properties of categorical regression models). Note that all generalized Dirichlet multinomial splitting
regressions are not symmetric since the singular distribution is not.
18
7
PIERRE FERNIQUE ET AL .
|
EMPIRICAL STUDIES
All studies presented in this section are reproducible. Packages used are installable using the conda package management system and each study is available as a Jupyter notebook (see Appendix D).
7.1
|
A comparison of multivariate models for count data
In order to illustrate the variety of splitting models, we considered two datasets used in the literature to illustrate
models for count data. The first one consists in outcomes of football games (Karlis and Ntzoufras, 2005) and the second
one consists in simulated data mimicking data obtained from sequencing techonologies such as RNA-seq data (Zhang
et al., 2017). The goal being to compare distributions and regressions models, comparisons were performed when
considering all covariates or none of the covariates (see Table 11). Remark that variable selection (Zhang et al., 2017,
e.g., using regularization methods) is possible, but is out of the scope of this paper.
Let us first remark that the inference methodology for multinomial, Dirichlet multinomial and generalized Dirichlet
multinomial regressions presented by Zhang et al. (2017) and implemented by Zhang and Zhou (2017) is only valid for
singular versions. Their comparisons of these models against the negative multinomial is therefore invalid since the first
three models focus on Y given |Y | and the latter focuses on Y . Hence, we only compared our results to their unique
J -multivariate model that is the negative multinomial model and the multivariate Poisson model defined by Karlis and
Meligkotsidou (2005). By limiting the number of sum models to 7 and the number of singular models to 6, we were able
to propose 42 splitting models. Among those 42 models, only 4 models were not sensu stricto multivariate models since
multinomial splitting Poisson models induce independent response variables.
For the first dataset, the best splitting model is a singular multinomial regression compounded by a Poisson
distribution with a BIC of 508.14 + 1, 130.64 = 1, 638.78. This score is inferior to the one of the best multivariate Poisson
model (i.e., 1, 710.05) and the one of the best negative multinomial model (i.e., 1, 705.93). This indicates that there is no
relationship between football team goals. For the second dataset, the best splitting model is a singular generalized
Dirichlet multinomial regression compounded by a negative binomial regression with a BIC of 8843.48 + 2514.3 =
11, 357.78. This score is also inferior to the one of the best negative multinomial model (i.e., 17, 657.63).
7.2
|
An application to mango patchiness analysis
Recently, a statistical methodology has been proposed to characterize plant patchiness at the plant scale (Fernique et al.,
2016). However, little is known about patchiness at the whole population scale. To characterize patchiness at the plant
scale, a segmentation/clustering of tree-indexed data method has been proposed in order to split an heterogeneous
tree into multiple homogeneous subtrees. After the clustering, the tree can be summarized into a multivariate count
denoting the number of subtrees in each cluster (i.e., patch type). Mixture of multinomial splitting distributions can
therefore be considered to recover the different types of tree patchiness that can be found in the plant population. Such
a mixture model is of high interest since it enables to discriminate the types of tree patchiness according to the :
•
number of patches present on trees, by fitting different sum distributions within components of the mixture model,
•
distribution of these patches among types, by fitting different singular distributions within components of the
mixture model.
We here consider results presented by Fernique et al. (2016) to conduct our post-hoc analysis. Three different types of
PIERRE FERNIQUE ET AL .
19
(a)
(b)
Y given |Y | = n and X = x
BIC0
BIC1
|Y | given X = x
BIC0
BIC1
M∆n (π)
574.18
38, 767.91
P (λ)
1, 130.64
13, 074.12
D M ∆n (α)
579.49
9, 969.121
Bn (p)
1, 165.6
26, 474.38
G D M ∆n (α, β)
579.49
9, 735.45
N B (r , p)
1, 131.85
2, 678.55
M∆n {π (x)}
508.14
15, 145.24
L (p)
1, 370.84
3, 513.92
D M ∆n {α (x)}
836.4
8, 932.83
P {λ (x)}
1, 258.65
6, 353.13
G D M ∆n {α (x) , β (x)}
836.4
8, 843.479
Bn {p (x)}
1, 272.7
12, 999.47
N B {r , p (x)}
1, 264.38
2, 514.30
(c)
(d)
Y given X = x
BIC0
BIC1
Y given X = x
BIC0
BIC1
M P (λ)
1, 710.05
M N (r , π)
1, 705.93
41, 384.52
M P {λ (x)}
1, 956.10
M N {r , π (x)}
2, 176.3
17, 657.63
TA B L E 1 1 Bayesian Information Criteria (BIC) obtained for the first dataset (Karlis and Ntzoufras, 2005, BIC0 ) and
the second one (Zhang et al., 2017, BIC1 ) for (a) singular models, (b) sum models , (c) Poisson and (d) negative
multinomial models. Multivariate Poisson models could not be fit to the second dataset since, to our knowledge, there is
no implementation available in R for more than 2 response variables (Karlis and Ntzoufras, 2005).
patches have been identified for mango trees: vegetative patches which contain almost only vegetative growth units
(GU, plant elementary component), reproductive patches which contain almost only GUs that flowered or fructified and
quiescent patches which contain GUs that did not burst, flowered nor fructified. Multinomial splitting distributions
of mixture components were therefore of dimension 3, where N 0 (resp. N 1 and N 2 ) denotes the number of vegetative
(resp. reproductive and quiescent) patches observed within a tree. Since there is at least one patch in a mango tree (i.e.,
the tree itself), shifted singular multinomial splitting distributions were considered with a shift equal to 1 for binomial,
negative binomial and Poisson sum distributions but without shift for geometric and logarithmic distributions. Within
each component the parametric form of the sum distribution was selected using the BIC.
The mixture model selected using BIC has two components (see Figure 2) with weights P (L = 1) = 0.44 and
P (L = 2) = 0.56. In the two components i = 1, 2, the number of patches followed a multinomial splitting shifted negative
binomial distribution Y | L = i ∼ M∆N (πi ) ∧ N B (r i , p i ; δi ) with estimations π̂1 = (0.21, 0.00, 0.79), rˆ1 = 0.16, p̂ 1 = 0.76,
N
δ̂ 1 = 1 for the first component and π̂2 = (0.54, 0.17, 0.28), rˆ2 = 3.96, p̂ 2 = 0.40, δ̂ 2 = 1 for the second component. This
mixture of two components indicates that the population of mango trees can be separated into two types of trees (see
Figure 3):
•
mango trees with a relatively low number of patches that can be either vegetative or quiescent but not reproductive
(component 1),
•
mango trees with a relatively high number of patches that can be of any type and in particular reproductives
(component 2).
20
PIERRE FERNIQUE ET AL .
BIC
-400
-600
-800
1
FIGURE 2
2
3
Number of mixture components
4
5
BIC according to the number of multinomial splitting components of mixtures.
Frequency
80
60
40
20
0
2
4
6
n
8
10
12
F I G U R E 3 Representation of the mixture of sum distributions estimated (with a solid line) confronted to data
frequencies (gray bars). Note that the sum distribution of the first (resp. second) component is represented with a
dotted (resp. dashed) line.
These types of trees are almost equally represented in the period considered (52% for the first component against 48%).
This result tends to imply that the reproductive period of mango trees leads to an increase in patch number increase
while the vegetative period leads to a decrease in patch number.
8
| DISCUSSION
Convolutions splitting distributions that are positive and additive, have been studied in depth in this paper since their
graphical models and their marginal distributions are easily obtained. The characterization of the graphical model of
hypergeometric splitting distributions stay an open issue because of the non-positivity. But thanks to the additivity,
Theorem 1 still holds. It would be interesting to instantiate some univariate distributions L(ψ) and precisely describe
the resulting splitting distributions. More generally, the multivariate Polya distribution with parameters n ∈ Î, θ ∈ Θ
and c ∈ Ò encompasses the multivariate hypergeometric (c = −1), the multinomial (c = 0) and the Dirichlet multinomial
(c = 1) distributions (Janardan and Patil, 1970). It would therefore be interesting to study the properties of multivariate
Polya splitting distributions according to the c value. Otherwise, non-symmetric convolution distributions could be
defined (including the generalized Dirichlet multinomial distribution as a special case) to ease the study of corresponding
PIERRE FERNIQUE ET AL .
21
splitting distributions.
Another alternative to define new singular distributions is to consider their mixture. Finite mixture can be inferred
using classical expectation-maximization algorithm for multivariate distributions. Moreover, in an application context
the principle of mixture models is quite interesting for splitting models. If we consider the mango tree application,
we inferred mixture of splitting distributions in order to characterize plant patchiness at the plant scale. This relied
on the assumption that tree patchiness is both expressed in terms of number of patches and their type distribution.
One the one hand, if tree patchiness is only a phenomenon expressed in term of number of patches, a mixture of sum
distributions could be considered to distinguish trees. On the other hand, if tree patchiness is only a phenomenon
expressed in term of patch type distribution, singular distributions constructed using mixture of singular distributions
could be of most interest.
Finally, this work could be used for learning graphical models with discrete variables, which is an open issue.
Although the graphical models for convolution splitting distributions are basic (complete or empty), they could be
used as building blocks for partially directed acyclic graphical models. Therefore, the procedure of learning partially
directed acyclic graphical models described by Fernique et al. (2014) could be used for learning graphical models based
on convolution splitting distributions and regressions. It could be used for instance to infer gene co-expression network
from RNA seq dataset.
REFERENCES
Allen, G. I. and Liu, Z. (2012) A log-linear graphical model for inferring genetic networks from high-throughput sequencing data.
In IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 1–6. IEEE.
Blumenthal, S. and Dahiya, R. C. (1981) Estimating the binomial parameter n. Journal of the American Statistical Association, 76,
903–909.
Chen, J. and Li, H. (2013) Variable selection for sparse dirichlet-multinomial regression with an application to microbiome data
analysis. The Annals of Applied Statistics, 7, 418–442.
Dai, H., Bao, Y. and Bao, M. (2012) Maximum likelihood estimate for the dispersion parameter of the negative binomial distribution. Statistics & Probability Letters, 83, 21–27.
Deriggi, D. F. (1983) Unimodality of likelihood functions for the binomial distribution. Journal of the American Statistical Association, 78, 181–183.
Digby, P. and Kempton, R. A. (2012) Multivariate analysis of ecological communities, vol. 5. Springer Science & Business Media.
Dunkl, C. F. (2013) Products of Beta distributed random variables. ArXiv e-prints.
Fernique, P., Dambreville, A., Durand, J.-B., Pradal, C., Lauri, P.-E., Normand, F. and Guédon, Y. (2016) Characterization of
mango tree patchiness using a tree-segmentation/clustering approach. In 2016 IEEE International Conference on FunctionalStructural Plant Growth Modeling, Simulation, Visualization and Applications (FSPMA), 68–74. IEEE.
Fernique, P., Durand, J.-B. and Guédon, Y. (2014) Estimation of discrete partially directed acyclic graphical models in multitype
branching processes. In COMPSTAT 2014, 21st International Conference on Computational Statistics.
Forcina, A. and Franconi, L. (1988) Regression analysis with the beta-binomial distribution. Rivista di Statistica Applicata, 21.
Gurland, J. (1958) A generalized class of contagious distributions. Biometrics, 14, 229–249.
Hilbe, J. M. (2011) Negative binomial regression. Cambridge University Press.
22
PIERRE FERNIQUE ET AL .
Irwin, J. O. (1968) The generalized waring distribution applied to accident theory. Journal of the Royal Statistical Society. Series A
(General), 205–225.
Janardan, K. G. and Patil, G. P. (1970) On the multivariate polya distribution: a model of contagion for data with multiple counts.
Random Count in Scientific Work, 3, 143–161.
— (1972) A unified approach for a class of multivariate hypergeometric models. Sankhyā: The Indian Journal of Statistics, Series
A (1961-2002), 34, 363–376. URL: http://www.jstor.org/stable/25049821.
Johnson, N., Kemp, A. and Kotz, S. (1993) Univariate discrete distributions. Wiley-Interscience.
Johnson, N., Kotz, S. and Balakrishnan, N. (1997) Discrete multivariate distributions. Wiley New York.
Karlis, D. and Meligkotsidou, L. (2005) Multivariate Poisson regression with covariance structure. Statistics and Computing, 15,
255–265.
Karlis, D. and Ntzoufras, I. (2005) Bivariate poisson and diagonal inflated bivariate Poisson regression models in R. Journal of
Statistical Software, 14, 1–36.
Koller, D. and Friedman, N. (2009) Probabilistic Graphical Models. The MIT Press.
Lauricella, G. (1893) Sulle funzioni ipergeometriche a piu variabili. Rendiconti del Circolo Matematico di Palermo, 7, 111–158.
URL: https://doi.org/10.1007/BF03012437.
Lesnoff, M., Lancelot and R. (2012) aod: Analysis of Overdispersed Data. URL: http://cran.r-project.org/package=aod. R
package version 1.3.
McCullagh, P. and Nelder, J. (1989) Generalized linear models. Monographs on Statistics and Applied Probability 37. Chapman &
Hall, London.
Merkel, D. (2014) Docker: Lightweight Linux Containers for Consistent Development and Deployment. Linux Journal, 2014, 2.
Minka, T. (2000) Estimating a dirichlet distribution.
Mullahy, J. (1986) Specification and testing of some modified count data models. Journal of econometrics, 33, 341–365.
Patil, G. (1968) On sampling with replacement from populations with multiple characters. The Indian Journal of Statistics, Series
B, 30, 355–366.
Patil, G. P. and Bildikar, S. (1967) Multivariate logarithmic series distribution as a probability model in population and community ecology and some of its statistical properties. Journal of the American Statistical Association, 62, 655–674.
Peyhardi, J. and Fernique, P. (2017) Characterization of convolution splitting graphical models. Statistics & Probability Letters,
126, 59–64.
Peyhardi, J., Trottier, C. and Guédon, Y. (2015) A new specification of generalized linear models for categorical responses.
Biometrika, 102, 889–906.
Radhakrishna Rao, C. and Rubin, H. (1964) On a characterization of the Poisson distribution. Sankhyā Ser. A, 26, 295–298.
Rao, C. R. and Srivastava, R. (1979) Some characterizations based on a multivariate splitting model. Sankhyā: The Indian Journal
of Statistics, Series A, 124–128.
Rodríguez-Avi, J., Conde-Sánchez, A., Sáez-Castillo, A., Olmo-Jiménez, M. and Martínez-Rodríguez, A. M. (2009) A generalized
waring regression model for count data. Computational Statistics & Data Analysis, 53, 3717–3725.
Royle, J. A., Dorazio, R. M. and Link, W. A. (2007) Analysis of multinomial models with unknown index using data augmentation.
Journal of Computational and Graphical Statistics, 16, 67–85. URL: http://dx.doi.org/10.1198/106186007X181425.
PIERRE FERNIQUE ET AL .
23
Saez-Castillo, A. J., Vilchez-Lopez, S., Olmo-Jimenez, M. J., Rodriguez-Avi, J., Conde-Sanchez, A. and Martinez-Rodriguez, A. M.
(2017) GWRM: Generalized Waring Regression Model for Count Data. URL: https://cran.r-project.org/package=GWRM.
GWRM R package version 2.1.0.3.
Sanathanan, L. (1972) Estimating the size of a multinomial population. The Annals of Mathematical Statistics, 43, 142–152. URL:
http://www.jstor.org/stable/2239906.
Shanbhag, D. N. (1977) An extension of the Rao-Rubin characterization of the Poisson distribution. J. Appl. Probability, 14,
640–646.
Sklar, M. (2014) Fast mle computation for the dirichlet multinomial. arXiv preprint arXiv:1405.0099.
Tripathi, R. C., Gupta, R. C. and Gurland, J. (1994) Estimation of parameters in the beta binomial model. Annals of the Institute
of Statistical Mathematics, 46, 317–331.
Vu, T. N., Wills, Q. F., Kalari, K. R., Niu, N., Wang, L., Rantalainen, M. and Pawitan, Y. (2016) Beta-poisson model for single-cell
rna-seq data analyses. Bioinformatics, 32, 2128–2135.
Whitby, O. (1972) Estimation of parameters in the generalized beta distribution.
Winkelmann, R. (2013) Econometric analysis of count data. Springer Science & Business Media.
Xekalaki, E. (1981) Chance mechanisms for the univariate generalized waring distribution and related characterizations. In
Statistical distributions in scientific work, 157–171. Springer.
— (1986) The multivariate generalized waring distribution. Communications in Statistics - Theory and Methods, 15, 1047–1064.
Xia, F., Chen, J., Fung, W. K. and Li, H. (2013) A logistic normal multinomial regression model for microbiome compositional
data analysis. Biometrics, 69, 1053–1063.
Zhang, Y. and Zhou, H. (2017) MGLM: Multivariate Response Generalized Linear Models. URL: https://CRAN.R-project.org/
package=MGLM. R package version 0.0.8.
Zhang, Y., Zhou, H., Zhou, J. and Sun, W. (2017) Regression models for multivariate count data. Journal of Computational and
Graphical Statistics, 26, 1–13.
A
|
PROOFS
Details about marginal of symmetric splitting distributions
For any j ∈ {1, . . . , J } and y j ∈ Î we have
P (Yj = y j ) =
Õ
P (Y = y),
y−j
=
Õ
n ≥y j
P (Yj = y j ) =
Õ
P ( |Y | = n)
Õ
P | Y |=n (Y = y),
y−j
P ( |Y | = n)P | Y |=n (Yj = y j ),
n ≥y j
The marginal distribution of the singular distribution, i.e., the distribution of Yj given |Y | = n, is a distribution bounded
by n. Its parametrization has the same form fj (θ) for all marginals Yj given |Y | = n, since the singular distribution is
symmetric. It implies that all marginals Yj follow the damage distribution LN {fj (θ)} ∧ L(ψ).
N
24
PIERRE FERNIQUE ET AL .
Proof of Theorem 1
Let I denote a subset of {1, . . . , J } withI , {1, . . . , J }. Let −I denote the subset {1, . . . , J } \ I and yI (respectively
y−I ) denote the corresponding sub-vectors.
Proof of 1
P ( |YI | = n) =
Õ
P (YI = yI )
yI ∈∆n
=
Õ Õ
P (Y = y)
yI ∈∆n y−I
=
Õ
P | Y |=k (Y = y)
Õ P ( |Y | = k ) Õ Ö
a θj (y j )
c θ (k )
yI ∈∆n j ∈I
k ≥n
P ( |YI | = n) =
Õ
yI ∈∆n y−I ∈∆k −n
k ≥n
=
Õ
P ( |Y | = k )
Õ c θ (n)c θ (k − n)
−I
I
c θ (k )
k ≥n
Õ
Ö
a θj (y j )
y−I ∈∆k −n j ∈−I
P ( |Y | = k )
where c θI (n) denotes the convolution of (a θj )j ∈I over the simplex ∆n . Since the convolution distribution is assumed to
be additive, we obtain by recursion on j ∈ I (resp. j ∈ −I and j ∈ {1, . . . , J }) that
P ( |YI | = n) =
Õ a | θ | (n)a | θ | (k − n)
I
−I
a | θ | (k )
k ≥n
Moreover we obtain the convolution identity
Ík
n=0
P ( |Y | = k )
(16)
a | θI | (n)a | θ−I | (k − n) = a | θ | (k ) and thus the last equation well defines
the desired convolution damage distribution ∼ CN (a; |θI |, |θ−I |) ∧ L(ψ).
N
Proof of 2 and 3
P (YI = yI , |YI | = n) = P (YI = yI )1∆n (yI )
Õ
= 1∆n (yI )
P ( |Y | = k )
Ö
a θj (y j )1∆n (yI )
j ∈I
P (YI = yI , |YI | = n) =
Ö
P | Y |=k (Y = y)
y−I ∈∆k −n
k ≥n
=
Õ
Õ P ( |Y | = k )
c θ (k )
k ≥n
a θj (y j )1∆n (yI )
j ∈I
Õ
y−I ∈∆k −n j ∈−I
Õ a | θ | (k − n)
−I
k ≥n
a | θ | (k )
P ( |Y | = k )
Using equation 16 we obtain the conditional probability
P | YI |=n (YI = yI ) =
1
Ö
a | θI | (n)
j ∈I
Ö
a θj (y j )1∆n (yI ),
and thus 2 holds. Remark that 1 and 2 imply 3 by definition of a splitting distribution.
a θj (y j )
PIERRE FERNIQUE ET AL .
25
Proof of 4 and 5
P Y−I =y−I (YI = yI ) = PY−I =y−I , | YI |=| yI | (YI = yI )PY−I =y−I ( |YI | = |yI |)
Since the sum |Y | is independent of the vector Y−I given its sum |Y−I | it can be shown that
PY−I =y−I (YI = yI ) = P | YI |=| yI | (YI = yI )P | Y−I |=| y−I | ( |YI | = |yI |)
Thanks to the result 2, the left part of this product is given by the singular convolution distribution. Remarking that
P | Y−I |=| y−I | ( |YI | = |yI |) = P | Y |≥a ( |Y | = a + |yI |) with a = |y−I | the left part is given by the truncated and shifted
distribution T S a { L(ψ)} and thus 4 holds. Remark that 3 and 4 imply 5.
Proof of corollary 2
Assume that L(ψ) is stable under the damage process CN (a; |θI |, |θ−I |) ∧ (·) for any subset I ⊂ {1, . . . , J }. Thanks
N
to the additivity of the convolution distribution, Theorem 1 can be applied. Using the property 3, it is easily seen that
multivariate marginals are stable. The criterion 4 holds and the convolution splitting distribution is considered as a
natural multivariate extension of L(ψ). In particular, L(ψ) is stable under CN (a; |θj |, |θ−j |) ∧ (·) and thus the univariate
N
marginal follow L(ψj ) for some ψj ∈ Ψ.
Proof of corollary 3
Let Y follow the non-singular version of an additive convolution distribution: Y ∼ CNn (a θ ; θ, θ). It means that the
completed vector (Y , n − |Y |) follow the additive convolution C∆J +1 (a θ ; θ, θ). Otherwise this singular distribution can
n
seen as a particular splitting Dirac distribution, i.e., C∆J +1 (a θ ; θ, θ) = C∆J +1 (a θ ; θ, θ) ∧ 1n . Thanks to the additivity, the
n
N
N
Theorem 1 can be applied on the completed vector (Y , n − |Y |) to describe the distribution of Y (property 3):
Y ∼ C∆N (a θ ; θ) ∧ CN 0 (a θ ; |θ |, θ) ∧0 1n ,
N
N
⇔ Y ∼ C∆N (a θ ; θ) ∧ Cn (a θ ; |θ |, θ).
N
Proof of equality (3)
Using the Corollary 3 with a θ (y ) = θ y /y ! we obtain for θ ∈ (0, ∞)J and γ ∈ (0, ∞)
M∆N (θ) ∧ Bn (|θ |, γ) = MNn (θ, γ) .
N
Denoting by π =
1
|θ |
· θ, p =
|θ |
| θ |+γ
and π ∗ =
1
| θ |+γ
· θ and using the proportionality we obtain equivalently
M∆N (π) ∧ Bn (p, 1 − p) = MNn π ∗ , 1 − |π ∗ | .
N
The notation of the binomial and the non-singular multinomial are then simplified by letting aside the last parameter
without loss of generality, i.e. we have M∆N (π) ∧ Bn (p) = MNn (π ∗ ). Finally remarking that π ∗ = p · π we obtain the
desired result.
N
26
B
PIERRE FERNIQUE ET AL .
| REMARKABLE CONTINUOUS UNIVARIATE DISTRIBUTION
Let us recall the definition of the (generalized) beta distribution with positive real parameters c, α and b, denoted by
β c (a, b). Its probability density function described by Whitby (1972) is given by
f (x ) =
1
x a−1 (c − x )b−1
· 1(0,c) (x ).
B (a, b)
c a+b−1
Note that Z = dX with d ∈ (0, ∞) and X ∼ β c (a, b) implies that Z ∼ β cd (a, b). The parameter c of the beta distribution
can thus be interpreted as a rescaling parameter of the standard beta distribution. By convention the standard beta
distribution (i.e., defined with c = 1) will be denoted by β (a, b).
Let us introduce the definition of the (generalized) beta square distribution with parameters (a 1 , b 1 , a 2 , b 2 ) ∈ (0, ∞)4
and c ∈ (0, ∞), denoted by β c2 (a 1 , b 1 , a 2 , b 2 ), as the product of the two independent beta distributions β (a 1 , b 1 ) and
β (a 2 , b 2 ) normalized on (0, c); see Dunkl (2013) for details. It is named the standard beta square distribution when c = 1
and denoted by β 2 (a 1 , b 1 , a 2 , b 2 ). More generally the product of m beta distributions could be defined.
C
C.1
| REMARKABLE DISCRETE UNIVARIATE DISTRIBUTIONS
| Power series distributions
Let (b y ) y ∈Î be a non-negative real sequence such that the series
Í
y ≥0 b y θ
y
converges toward g (θ) for all θ ∈ D = (0, R ),
where R is the radius of convergence. The discrete random variable Z is said to follow a power series distribution if for
all y ∈ Î
P (Y = y ) =
by θy
,
g (θ)
and is denoted by Y ∼ P S D {g (θ)}. Several usual discrete distributions fall into the family of power series distributions:
1.
2.
3.
The Poisson distribution P(λ) with b y = 1/y !, θ = λ, g (θ) = e θ and D = (0, ∞).
The binomial distribution Bn (p) with b y = yn 1 y ≤n , θ = p/(1 − p), g (θ) = (1 + θ)n and D = (0, ∞).
The negative binomial distribution N B(r , p) with b y = r +yy −1 , θ = p, g (θ) = (1 − θ)−r and D = (0, 1).
4.
The geometric distribution G(p) with b y = 1 y ≥1 , θ = 1 − p, g (θ) = θ/(1 − θ) and D = (0, 1).
5.
The logarithmic series distribution L(p) with b y = 1 y ≥1 1/y , θ = p, g (θ) = − ln(1 − θ) and D = (0, 1).
When the support is a subset of Î, the b y values can be weighted by an indicator function as for binomial, geometric and
logarithmic distributions. The b y must be independent of θ but they may depend on other parameters as for binomial
and negative binomial distributions.
Zero modified logarithmic series Johnson et al. (1997) The discrete variable Y is said to follow a zero modified
logarithmic series distribution with parameter ω ∈ [0, 1) and p ∈ (0, 1) if the its probabilities are given by
P (Y = 0) = ω,
P (Y = y ) =
(1 − ω)p y
, y ≥ 1.
−y ln(1 − p)
PIERRE FERNIQUE ET AL .
27
This distribution will be denoted by L(p, ω). Note that if ω = 0 this is the logarithmic series distribution: L(p, 0) = L(p).
C.2
| Beta compound distributions
Usual characteristics of the standard beta binomial (Tripathi et al., 1994), standard beta negative binomial - also described by Xekalaki (1981) as the univariate generalized waring distribution (UGWD) - and the beta Poisson distributions
(Gurland, 1958) are first recalled in Table 12. Then we introduce these beta compound distributions in a general way, i.e.
using the generalized beta distribution (see Appendix B). For the Poisson case we obtain the same distribution since
P(λp) ∧ β (a, b) = P(θ) ∧ β λ (a, b). The two other case lead us to new distributions (13 for usual characteristics). Remark
p
θ
that if π = 1 then, the generalized beta binomial (resp. generalized beta negative binomial) turns out to be the standard
beta binomial (resp. standard negative binomial distribution). In opposition, if π < 1, the non-standard beta binomial
distribution (respectively non-standard beta negative binomial distribution) is obtained.
Generalized beta binomial distribution
Let n ∈ Î, a ∈ (0, ∞), b ∈ (0, ∞) and π ∈ (0, 1) and consider the compound distribution Bn (p) ∧ β π (a, b) denoted by
p
β π Bn (a, b). Considering π as a rescaling parameter, we have β π Bn (a, b) = Bn (πp) ∧ β (a, b). Moreover, using the pgf of
p
the binomial distribution in equation (8), it can be shown that Bn (πp) = BN (π) ∧ Bn (p). Finally, using the Fubini theorem
N
we obtain
β π Bn (a, b) = BN (π) ∧ Bn (p) ∧ β (a, b),
p
N
= BN (π) ∧ Bn (p) ∧ β (a, b) ,
N
p
β π Bn (a, b) = BN (π) ∧ β Bn (a, b) .
N
This is a binomial damage distribution whose the latent variable N follows a standard beta binomial distribution. The
equation (8) can thus be used to compute the probability mass function. The y t h derivative of the pgf of the standard
beta binomial distribution is thus needed
(y )
G N (s) =
(−n) y (a) y
(b)n
2 F 1 {(−n + y , a + y ); −b − n + 1 + y ; s },
(a + b)n (−b − n + 1) y
obtained by induction on y ∈ Î. The moments are obtained with the total law of expectation and variance given the
latent variable N of the binomial damage distribution. In the same way, we obtain the pgf as GY (s) = G N (1 − π + πs). A
similar proof holds for the generalized beta negative binomial case.
Generalized beta square compound distributions
It is also possible to define the (generalized) beta square distribution, as the product of two independent beta distributions (Dunkl, 2013), and then define the (generalized) beta square compound distributions.
•
The standard beta square binomial distribution is defined as Bn (p)∧β 2 (a 1 , b 1 , a 2 , b 2 ) and denoted by β 2 Bn (a 1 , b 1 , a 2 , b 2 ).
•
The standard beta square negative binomial distribution is defined as N B(r , p) ∧ β 2 (a 1 , b 1 , a 2 , b 2 ) and denoted by
p
p
β 2 N B(r , a 1 , b 1 , a 2 , b 2 ).
•
The generalized beta square binomial distribution is defined as Bn (p)∧β π2 (a 1 , b 1 , a 2 , b 2 ) and denoted by β π2 Bn (a 1 , b 1 , a 2 , b 2 )
p
28
•
PIERRE FERNIQUE ET AL .
The generalized beta square negative binomial distribution is defined as N B(r , p) ∧ β π2 (a 1 , b 1 , a 2 , b 2 ) and denoted
p
by β π2 N B(r , a 1 , b 1 , a 2 , b 2 )
•
The beta square Poisson distribution is defined as P(θ) ∧ β λ2 (a 1 , b 1 , a 2 , b 2 ) and denoted by β λ2 P(a 1 , b 1 , a 2 , b 2 ).
θ
D |
REPRODUCIBILITY
The source code (written in C++ and Python) used for the inference of splitting distributions is available on GitHub
(https://github.com/StatisKit) and binaries can be installed using the Conda package management system ((http:
//conda.pydata.org)). Refers to the documentation for more information ((http://statiskit.rtfd.io)).
Our analyses performed with the Python interface or R packages is available in the Jupyter notebook format
as supplementary materials and can be reproduced using the Docker (Merkel, 2014) image statiskit/FPD18 (see
https://hub.docker.com/r/statiskit/FPD18 for image and the documentation for more information).
β Bn (a, b)
Notation
p
+ a + b; s }
1 F 1 {a; a
(b)n (−n) y (a) y π y
(a+b)n (−b−n+1) y y ! 2 F 1 {(−n
a
nπ a+b
b(a+b+n)
a
nπ a+b
π (a+b)(a+b+1)
+1−π
(b)n
(a+b)n 2 F 1 {(−n, a); −b
P (Y = y )
E(Y )
V(Y )
GY (s)
− n + 1; 1 + π(s − 1)}
14
+ y , b + y ); r + a + b + y ; 1 − π }
(a)r
(a+b)r 2 F 1 {(r , b); r
+ a + b; 1 + π(s − 1)}
n
o
−1)(a+b−1)
b
r π a−1
π (a+r
+ 1 − π 15
(a−2)(a−1)
b
r π a−1
β π N B (r , a, b)
+ y ; a + b + y ; −λ)
Usual characteristics of the generalized beta binomial and the generalized beta negative binomial distributions.
n
o
(a)r (r ) y (b) y π y
(a+b)r (r +a+b) y y ! 2 F 1 {(r
{0, 1, . . . , n }
Supp(Y )
+ y , a + y ); −b − n + 1 + y ; 1 − π }
Î
β π Bn (a, b)
p
N B (r , p) ∧ β π (a, b)
Notation
p
Bn (p) ∧ β π (a, b)
Definition
Generalized beta negative binomial
Generalized beta binomial
Name
TA B L E 1 3
+ b; λ(s − 1)}
Usual characteristics of the standard beta compound binomial, negative binomial and Poisson distributions.
GY (s)
TA B L E 1 2
(a)r
(a+b)r 2 F 1 {(r , b); r
(b)n
(a+b)n 2 F 1 {(−n, a); −b
V(Y )
o
n
a
b
λ a+b
1 + λ (a+b)(a+b+1)
n ab(a+b+n)
(a+b)2 (a+b+1)
b(a+r −1)(a+b−1) 13
(a−1)2 (a−2)
r
n
E(Y )
r
(a) y λ y
(a+b) y y ! 1 F 1 (a
a
λ a+b
P (Y = y )
b 12
a−1
P (Y = y ) =
(a)b (r ) y (b) y 1 11
(a+r )b (r +a+b) y y !
(b)n (−n) y (a) y 1
(a+b)n (−b−n+1) y y !
a
a+b
Î
Î
{0, 1, . . . , n }
Supp(Y )
β Pλ (a, b)
P (λp) ∧ β (a, b)
beta Poisson
UGWD(r , b, a)
β N B (r , a, b)
p
N B (r , p) ∧ β (a, b)
Standard beta negative binomial
Re-parametrization
− n + 1; s }
Bn (p) ∧ β (a, b)
Definition
p
Standard beta binomial
Name
PIERRE FERNIQUE ET AL .
29
| 10math.ST
|
Estimation in the convolution structure density model.
Part I: oracle inequalities.
arXiv:1704.04418v1 [math.ST] 14 Apr 2017
O.V. Lepski
∗
T. Willer
∗
Institut de Mathématique de Marseille
Aix-Marseille Université
39, rue F. Joliot-Curie
13453 Marseille, France
e-mail: [email protected]
e-mail: [email protected]
Abstract:
We study the problem of nonparametric estimation under Lp -loss, p ∈ [1, ∞), in the framework of the convolution structure density model on Rd . This observation scheme is a generalization of two classical statistical models, namely density estimation under direct and indirect
observations. In Part I the original pointwise selection rule from a family of ”kernel-type”
estimators is proposed. For the selected estimator, we prove an Lp -norm oracle inequality and
several of its consequences. In Part II the problem of adaptive minimax estimation under
Lp –loss over the scale of anisotropic Nikol’skii classes is addressed. We fully characterize the
behavior of the minimax risk for different relationships between regularity parameters and
norm indexes in the definitions of the functional class and of the risk. We prove that the
selection rule proposed in Part I leads to the construction of an optimally or nearly optimally
(up to logarithmic factor) adaptive estimator.
AMS 2000 subject classifications: 62G05, 62G20.
Keywords and phrases: deconvolution model, density estimation, oracle inequality, adaptive estimation, kernel estimators, Lp –risk, anisotropic Nikol’skii class.
1. Introduction
In the present paper we will investigate the following observation scheme introduced in Lepski and Willer
(2017). Suppose that we observe i.i.d. vectors Zi ∈ Rd , i = 1, . . . , n, with a common probability
density p satisfying the following structural assumption
p = (1 − α)f + α[f ⋆ g], f ∈ Fg (R), α ∈ [0, 1],
(1.1)
where α ∈ [0, 1] and g : Rd → R are supposed to be known and f : Rd → R is the function to be
estimated. We will call the observation scheme (1.1)
convolution structure density model.
Here and later, for two functions f, g ∈ L1 Rd
Z
f (x − z)g(z)νd (dz), x ∈ Rd ,
f ⋆ g (x) =
Rd
and for any α ∈ [0, 1], g ∈ L1 Rd and R > 1,
n
o
Fg (R) = f ∈ B1,d (R) : (1 − α)f + α[f ⋆ g] ∈ P Rd .
Here P
Rd denotes the set of probability densities on Rd , Bs,d (R) is the ball of radius R > 0 in
Ls Rd := Ls Rd , νd , 1 ≤ s ≤ ∞ and νd is the Lebesgue measure on Rd .
∗
This work has been carried out in the framework of the Labex Archimède (ANR-11-LABX-0033) and of the
A*MIDEX project (ANR-11-IDEX-0001-02), funded by the ”Investissements d’Avenir” French Government program
managed by the French National Research Agency (ANR).
1
imsart-generic ver. 2014/10/16 file: DSDM-oracle-arXiv.tex date: April 17, 2017
We remark that if one assumes additionally that f, g ∈ P Rd , this model can be interpreted as
follows. The observations Zi ∈ Rd , i = 1, . . . , n, can be written as a sum of two independent random
vectors, that is,
Zi = Xi + ǫi Yi , i = 1, . . . , n,
(1.2)
where Xi , i = 1, . . . , n, are i.i.d. d-dimensional random vectors with a common density f , to be
estimated. The noise variables Yi , i = 1, . . . , n, are i.i.d. d-dimensional random vectors with a
known common density g. At last εi ∈ {0, 1}, i = 1, . . . , n, are i.i.d. Bernoulli random variables
with P(ε1 = 1) = α, where α ∈ [0, 1] is supposed to be known. The sequences {Xi , i = 1, . . . , n},
{Yi , i = 1, . . . , n} and {ǫi , i = 1, . . . , n} are supposed to be mutually independent.
The observation scheme (1.2) can be viewed as the generalization of two classical statistical
models. Indeed, the case α = 1 corresponds to the standard deconvolution model Zi = Xi +
Yi , i = 1, . . . , n. Another ”extreme” case α = 0 corresponds to the direct observation scheme
Zi = Xi , i = 1, . . . , n. The ”intermediate” case α ∈ (0, 1), considered for the first time in Hesse
(1995), can be treated as the mathematical modeling of the following situation. One part of the
data, namely (1−α)n, is observed without noise, while the other part is contaminated by additional
noise. If the indexes corresponding to that first part were known, the density f could be estimated
using only this part of the data, with the accuracy corresponding to the direct case. The question
we address now is: can one obtain the same accuracy if the latter information is not available? We
will see that the answer to the aforementioned question is positive, but the construction of optimal
estimation procedures is based upon ideas corresponding to the ”pure” deconvolution model.
The convolution structure density model (1.1) will be studied for an arbitrary g ∈ L1 Rd and
f ∈ Fg (R). Then, except in the case α = 0, the function f is not necessarily a probability density.
We want to estimate f using the observations
Z (n) = (Z1 , . . . , Zn ). By estimator, we mean any
Z (n) -measurable map fˆ : Rn → Lp Rd . The accuracy of an estimator fˆ is measured by the Lp –risk
1/p
, p ∈ [1, ∞),
Rn(p) [fˆ, f ] := Ef kfˆ − f kpp
where Ef denotes the expectation with respect to the probability measure Pf of the observations
Z (n) = (Z1 , . . . , Zn ). Also, k · kp , p ∈ [1, ∞), is the Lp -norm on Rd and without further mentioning
we will assume that f ∈ Lp Rd . The objective is to construct an estimator of f with a small
Lp –risk.
1.1. Oracle approach via local selection. Objectives of Part I
Let F = fˆt , t ∈ T be a family of estimators built from the observation Z (n) . The goal is to
propose a data-driven (based on Z (n) ) selection procedure from the collection F and to establish
for it an Lp -norm oracle inequality. More precisely, we want to construct a Z (n) -measurable random
map t̂ : Rd → T and prove that for any p ∈ [1, ∞) and n ≥ 1
Rn(p) fˆt̂(·) ; f ≤ C1 inf An (f, t, ·)
t∈T
1
p
+ C2 n − 2 ,
∀f ∈ Lp Rd .
(1.3)
Here C1 and C2 are numerical constants which may depend on d, p and T only.
We call (1.3) an Lp -norm oracle inequality obtained by local selection, and in Part I we provide
with an explicit expression of the functional An (·, ·, x), x ∈ Rd in the case where F = F Hd is
the family of ”kernel-type” estimators parameterized by a collection of multi-bandwidths Hd . The
selection from the latter family is done pointwisely, i.e. for any x ∈ Rd , which allows to take into
account the ”local structure” of the function to be estimated. The Lp -norm oracle inequality is
2
then obtained by the integration of the pointwise risk of the proposed estimator, which is a kernel
estimator with the bandwidth being a multivariate random function. This, in its turn, allows us
to derive different minimax adaptive results presented in Part II of the paper. They are obtained
thanks to an unique Lp -norm oracle inequality.
Our selection rule presented in Section 2.1 can be viewed as a generalization and modification of
some statistical procedures proposed in Kerkyacharian et al. (2001) and Goldenshluger and Lepski
(2014). As we mentioned above, establishing (1.3) is the main objective of Part I. We will see
however that although An (·, ·, x), x ∈ Rd will be presented explicitly, its computation in particular
problems is not a simple task. The main difficulty here is mostly related to the fact that (1.3) is
proved without any assumption (except for the model requirements) imposed on the underlying
function f . It turns out that under some nonrestrictive assumptions imposed on f , the obtained
bounds can be considerably simplified, see Section 2.3. Moreover these new inequalities allow to
better understand the methodology for obtaining minimax adaptive results by the use of the oracle
approach.
1.2. Adaptive estimation. Objectives of Part II
(p)
Let F be a given subset of Lp Rd . For any estimator f˜n , define its maximal risk by Rn f˜n ; F =
(p)
supf ∈F Rn f˜n ; f and its minimax risk on F is given by
φn (F) := inf Rn(p) f˜n ; F .
(1.4)
f˜n
Here, the infimum is taken over all possible estimators. An estimator whose maximal risk is
bounded,
up to some constant factor, by φn (F), is called
minimax on F.
d
Let Fϑ , ϑ ∈ Θ be a collection of subsets of Lp R , νd , where ϑ is a nuisance parameter which
may have a very complicated structure.
The problem of adaptive estimation can be formulated as follows: is it possible to construct a
single estimator fˆn which would be simultaneously minimax on each class Fϑ , ϑ ∈ Θ, i.e.
(p) ˆ
lim sup φ−1
n (Fϑ )Rn fn ; Fϑ < ∞, ∀ϑ ∈ Θ?
n→∞
We refer to this question as the problem of minimax adaptive estimation over the scale of {Fϑ , ϑ ∈
Θ}. If such an estimator exists, we will call it optimally adaptive.
From oracle approach to adaptation. Let the oracle inequality (1.3) be established. Define
1
Rn Fϑ = sup inf An (f, t, ·) + n− 2 , ϑ ∈ Θ.
f ∈Fϑ
t∈T
p
We immediately deduce from (1.3) that for any ϑ ∈ Θ
lim sup Rn−1 Fϑ Rn(p) fˆt̂(·) ; Fϑ < ∞.
n→∞
Hence, the minimax adaptive optimality of the estimator fˆt̂(·) is reduced to the comparison of the
normalization Rn Fϑ with the minimax risk φn (Fϑ ). Indeed, if one proves that for any ϑ ∈ Θ
lim inf Rn Fϑ φ−1
n (Fϑ ) < ∞,
n→∞
then the estimator fˆt̂(·) is optimally adaptive over the scale Fϑ , ϑ ∈ Θ .
Objectives. In the framework of the convolution structure density model, we will be interested in
adaptive estimation over the scale
3
~ L
~ ~r, L,
~ ∩ Fg (R), ϑ = β,
~ R ,
Fϑ = N~r,d β,
~ L
~ is the anisotropic Nikolskii class (its exact definition will be presented in Part II).
where N~r,d β,
~ L
~ the coordinate βi of the vector β~ = (β1 , . . . , βd ) ∈
Here we only mention that for any f ∈ N~r,d β,
(0, ∞)d represents the smoothness of f in the direction i and the coordinate ri of the vector ~r =
~ L
~
(r1 , . . . , rd ) ∈ [1, ∞]d represents the index of the norm in which βi is measured. Moreover, N~r,d β,
~ ∈ (0, ∞)d represents the
is the intersection of balls in some semi-metric space and the vector L
radii of these balls.
The aforementioned dependence on the direction is usually referred to anisotropy of the underlying function and the corresponding functional class. The use of the integral norm in the definition
of the smoothness is referred to inhomogeneity of the underlying function. The latter means that
the function f can be sufficiently smooth on some part of the observation
domain
and rather
~ L
~ ~r, L
~ , β,
~ ∈
irregular on another part. Thus, the adaptive estimation over the scale N~r,d β,
(0, ∞)d × [1, ∞]d × (0, ∞)d can be viewed as the adaptation to anisotropy and inhomogeneity of
the function to be estimated.
~ L
~ ~r, L,
~ ∩ Fg (R) ∩ B∞,d (Q), ϑ = β,
~ R, Q . It will
Additionally, we will consider Fϑ = N~r,d β,
allow us to understand how the boundedness of the underlying function may affect the accuracy of
estimation.
The minimax adaptive estimation is a very active area of mathematical statistics, and the
theory of adaptation was developed considerably over the past three decades. Several estimation procedures were proposed in various statistical models, such that Efroimovich-Pinsker method,
Efroimovich and Pinsker (1984), Efroimovich (1986), Lepski method, Lepskii (1991) and its generalizations, Kerkyacharian et al. (2001), Goldenshluger and Lepski (2009), unbiased risk minimization, Golubev (1992), wavelet thresholding, Donoho et al. (1996), model selection, Barron et al.
(1999), Birgé and Massart (2001), blockwise Stein method, Cai (1999), aggregation of estimators,
Nemirovski (2000), Wegkamp (2003), Tsybakov (2003), Goldenshluger (2009), exponential weights,
Leung and Barron (2006), Dalalyan and Tsybakov (2008), risk hull method, Cavalier and Golubev
(2006), among many others. The interested reader can find a very detailed overview as well as
several open problems in adaptive estimation in the recent paper, Lepski (2015).
As already said, the convolution structure density model includes itself the density estimation
under direct and indirect observations. In Part II we compare in detail our minimax adaptive results
to those already existing in both statistical models. Here we only mention that more developed
results can be found in Goldenshluger and Lepski (2011), Goldenshluger and Lepski (2014) (density
model) and in Comte and Lacour (2013), Rebelles (2016) (density deconvolution).
1.3. Assumption on the function g
Pd
R
Later on for any U ∈ L1 Rd , let Ǔ denote its Fourier transform, defined as Ǔ (t) := Rd U (x)e−i j=1 xj tj νd (dx), t ∈
Rd . The selection rule from the family of kernel estimators, the Lp -norm oracle inequality as well
as the adaptive results presented in Part II are established under the following condition.
Assumption 1. (1) if α 6= 1 then there exists ε > 0 such that
1 − α + αǧ(t) ≥ ε,
∀t ∈ Rd ;
(2) if α = 1 then there exists µ
~ = (µ1 , . . . , µd ) ∈ (0, ∞)d and Υ0 > 0 such that
4
|ǧ(t)| ≥ Υ0
d
Y
(1 + t2j )−
µj
2
,
j=1
∀t = (t1 , . . . , td ) ∈ Rd .
Remind that the following assumption is well-known in the literature:
Υ0
d
Y
(1 + t2j )−
µj
2
j=1
≤ |ǧ(t)| ≤ Υ
d
Y
(1 + t2j )−
µj
2
,
j=1
∀t ∈ Rd .
It is referred to as a moderately ill-posed statistical problem. In particular, the assumption is satisfied
for the centered multivariate Laplace law.
Note that Assumption 1 (1) is very weak and it is verified for many distributions, including
centered multivariate Laplace and Gaussian ones. Note also that this assumption always holds with
ε = 1 − 2α if α < 1/2. Additionally, it holds with ε = 1 − α if ǧ is a real positive function. The
latter is true, in particular, for any probability law obtained by an even number of convolutions of
a symmetric distribution with itself.
2. Pointwise selection rule and Lp -norm oracle inequality
~ (α) = µ
~ (α) = (0, . . . , 0), α ∈ [0, 1).
To present our results in an unified way, let us define µ
~, α
R = 1, µ
Let K : Rd → R be a continuous function belonging to L1 Rd , R K = 1, and such that its Fourier
transform Ǩ satisfies the following condition.
Assumption 2. There exist k1 > 0 and k2 > 0 such that
Z
Rd
Ǩ(t)
d
Y
(1 +
t2j )
j=1
µj (α)
2
dt ≤ k1 ,
Z
Rd
Ǩ(t)
2
d
Y
(1 + t2j )µj (α) dt ≤ k22 .
(2.1)
j=1
Set H = ek , k ∈ Z and let Hd = ~h = (h1 , . . . , hd ) : hj ∈ H, j = 1, . . . , d . Define for any
~h = (h1 , . . . , hd ) ∈ Hd
d
Y
hj .
K~h (t) = V~−1 K t1 /h1 , . . . , td /hd , t ∈ Rd , V~h =
h
j=1
Rd
Later on for any u, v ∈
the operations and relations u/v, uv, u ∨ v,u ∧ v, u ≥ v, au, a ∈ R, are
understood in coordinate-wise sense. In particular u ≥ v means that uj ≥ vj for any j = 1, . . . , d.
2.1. Pointwise selection rule from the family of kernel estimators
For any ~h ∈ (0, ∞)d let M ·, ~h satisfy the operator equation
Z
~
g(t − y)M t, ~h dt, y ∈ Rd .
(2.2)
K~h (y) = (1 − α)M y, h + α
Rd
P
For any ~h ∈ Hd and x ∈ Rd introduce the estimator fb~h (x) = n−1 ni=1 M Zi − x, ~h .
Our first goal is to propose
x ∈ Rd a data-driven selection rule from the family of
for any given
d
d
b
~
kernel estimators F H = f~h (x), h ∈ H . Define for any ~h ∈ Hd
5
bn x, ~h =
U
s
b2 x, ~h
2λn ~h σ
4M∞ λn ~h
+
,
Q
n
3n dj=1 hj (hj ∧ 1)µj (α)
λn ~h = 4 ln(M∞ ) + 6 ln (n) + (8p + 26)
−d
M∞ = (2π)
Pointwise selection rule
introduce
−1
ε
Ǩ
1
1 α6=1
d
X
j=1
n
1X 2
M Zi − x, ~h ;
σ
b2 x, ~h =
n
i=1
1 + µj (α) ln(hj ) ;
+ Υ−1
0 k1 1α=1
∨ 1.
Let H be an arbitrary subset of Hd . For any ~h ∈ H and x ∈ Rd
h
i
bn x, ~η
b~ (x) = sup fb~ (x) − fb~η (x) − 4U
bn x, ~h ∨ ~η − 4U
;
R
h
h∨~
η
and define
+
η ∈H
~
b ∗ x, ~h =
U
n
sup
η ∈H: ~
~
η≥~h
bn x, ~η ,
U
h
i
~
b ∗ x, ~h .
b~ (x) + 8U
h(x) = arg inf R
n
h
(2.3)
~h∈H
Our final estimator is fb~h(x) (x), x ∈ Rd and we will call (2.3) the pointwise selection rule.
d since
~
b
Note that the estimator fbh(·)
~ (·) does not necessarily belong to the collection f~h (·), h ∈ H
the multi-bandwidth ~
h(·) is a d-variate function, which is not necessarily constant on Rd . The latter
fact allows to take into account the ”local structure” of the function to be estimated. Moreover,
~h(·) is chosen with respect to the observations, and therefore it is a random vector-function.
2.2. Lp -norm oracle inequality
Introduce for any x ∈ Rd and ~h ∈ Hd
Un∗ x, ~h =
sup
η ∈Hd : ~
~
η≥~h
Un x, ~η ,
S~h (x, f ) =
where we have put
s
4M∞ λn ~η
η
2λn ~η σ 2 x, ~
Un x, ~η =
+
,
Q
n
3n dj=1 ηj (ηj ∧ 1)µj (α)
For any H ⊆ Hd , ~h ∈ H and x ∈ Rd introduce also
B~h∗ (x, f ) = sup S~h∨~η (x, f ) − S~η (x, f ) ,
η ∈H
~
Z
Rd
K~h (t − x)f (t)νd (dt);
σ 2 x, ~η =
Z
Rd
M 2 t − x, ~η p(t)νd (dt).
B~h (x, f ) = S~h (x, f ) − f (x) .
(2.4)
Theorem 1. Let Assumptions 1 and 2 be fulfilled. Then for any H ⊆ Hd , n ≥ 3 and p ∈ [1, ∞),
n
o
1
∀f ∈ Fg (R), Rn(p) [fb~h(·) , f ] ≤ inf 2B~h∗ (·, f ) + B~h (·, f ) + 49Un∗ ·, ~h
+ Cp n− 2 .
~h∈H
p
6
The explicit expression for the constant Cp can be found in the proof of the theorem.
Later on we will pay attention to a special choice for the collection of multi-bandwidths, namely
d
Hisotr
:= ~h ∈ Hd : ~h = (h, . . . , h), h ∈ H .
More precisely, in Part II, the selection from the corresponding family of kernel estimators will be
used for the adaptive estimation over the collection of isotropic Nikolskii classes. Note also that if
d
d
H = Hisotr
then obviously for any ~h = (h, . . . , h) ∈ Hisotr
B~h∗ (·, f ) ≤ 2
sup
η ∈Hdisotr : η≤h
~
B~η (·, f )
and we come to the following corollary of Theorem 1.
Corollary 1. Let Assumptions 1 and 2 be fulfilled. Then for any n ≥ 3 and p ∈ [1, ∞)
1
(p) b
∗
~
Rn [f~h(·) , f ] ≤
inf
5
sup
B~η (·, f ) + 49Un ·, h
+ Cp n− 2 , ∀f ∈ Fg (R).
~h∈Hd
isotr
η ∈Hdisotr : η≤h
~
p
The oracle inequality proved in Theorem 1 is particularly useful since it does not require any
assumption on the underlying function f (except for the restrictions ensuring the existence of the
model and of the risk). However, the quantity appearing in the right hand side of this inequality,
namely
n
o
inf 2B~h∗ (·, f ) + B~h (·, f ) + 49Un∗ ·, ~h
~h∈H
p
is not easy to analyze. In particular, in order to use the result of Theorem 1 for adaptive estimation,
one has to be able to compute
n
o
sup inf 2B~h∗ (·, f ) + B~h (·, f ) + 49Un∗ ·, ~h
f ∈F
~h∈H
p
d
. It turns out that under
for a given class F ⊂ Lp Rd ∩ Fg (R) with either H = Hd or H = Hisotr
some nonrestrictive assumptions imposed on f , the obtained bounds can be considerably simplified.
Moreover, the new inequality obtained below will allow us to better understand the way for proving
adaptive results.
2.3. Some consequences of Theorem 1
Thus, furthermore we will assume that f ∈ Fg,u (R, D) ∩ Bq,d (D), q, u ∈ [1, ∞], D > 0, where
o
n
(∞)
Fg,u (R, D) := f ∈ Fg (R) : (1 − α)f + α[f ⋆ g] ∈ Bu,d (D) ,
(∞)
and Bu,d (D) denotes the ball of radius D in the weak-type space Lu,∞ Rd , i.e.
(∞)
Bu,d (D) = λ : Rd → R : kλku,∞ < D , kλku,∞ = inf C : νd x : |T (x)| > z ≤ C u z−u , ∀z > 0 .
(∞)
(∞)
As usual B∞,d (D) = B∞,d (D) and obviously Bu,d (D) ⊃ Bu,d (D). Note also that Fg,1 (R, D) = Fg (R)
for any D ≥ 1. It is worth noting that the assumption f ∈ Fg,u (R, D) simply means that the common
(∞)
density of the observations p belongs to Bu,d (D).
Remark 1. It is easily seen that Fg,∞ R, Rkgk∞ = Fg (R) if α = 1 and kgk∞ < ∞. Note also
that Fg,∞ (R, Qkgk1 ) ⊃ Fg (R) ∩ B∞,d (Q) for any α ∈ [0, 1] and Q > 0.
7
2.3.1. Oracle inequality over Fg,u (R, D) ∩ Bq,d (D)
For any ~h ∈ Hd and any v > 0, let
A(~h, f, v) = x ∈ Rd : B~h (x, f ) ≥ 2−1 v ,
B~h (·, f ) = 2B~h∗ (·, f ) + B~h (·, f ),
q
P
ln n + dj=1 | ln hj |
,
Fn ~h = √ Q
1
d
µj (α)
2
n j=1 hj (hj ∧ 1)
Gn ~h =
P
ln n + dj=1 | ln hj |
.
Q
n dj=1 hj (hj ∧ 1)µj (α)
d
Furthermore let H be either Hd or Hisotr
and for any v, z > 0 define
H(v) = ~h ∈ H : Gn ~h ≤ av ,
H(v, z) = ~h ∈ H(v) : Fn ~h ≤ avz −1/2 .
(2.5)
Here a > 0 is a numerical constant whose explicit expression is given in the beginning of Section
3.2. Introduce for any v > 0 and f ∈ Fg,u (R, D)
h
i
Λ(v, f ) =
inf νd A(~h, f, v) + v −2 Fn2 ~h ;
~h∈H(v)
Λ(v, f, u) = inf
inf
z≥2 ~h∈H(v,z)
Λp (v, f, u) = inf
inf
z≥2 ~h∈H(v,z)
h
i
νd A(~h, f, v) + z −u ;
Z
p
A(~h,f,v)
p −u
B~h (x, f ) νd (dx) + v z
.
Remark 2. Note that H(v) 6= ∅ and H(v, z) 6= ∅ whatever the values of v > 0 and z ≥ 2. Indeed,
for any v > 0 and z > 2 one can find b > 1 such that
ln n + d ln b (nbd )−1 ≤ a2 v 2 z −1 ∧ av.
The latter means that ~b = (b, . . . , b) ∈ H(v, z) ∩ H(v). Thus, we conclude that the quantities Λ(v, f ),
Λ(v, f, u) and Λp (v, f, u) are well-defined for all v > 0.
Also, It is easily seen that for any v > 0 and f ∈ Fg,∞ (R, D)
Λ(v, f, ∞) =
inf
~h∈H(v,2)
νd A(~h, f, v) ,
Λp (v, f, ∞) =
inf
~h∈H(v,2)
Z
p
A(~h,f,v)
B~h (x, f ) νd (dx).
(2.6)
Put at last for any v > 0, lH (v) = v p−1 (1 + | ln (v)|)t(H) , where t(H) = d − 1 if H = Hd and t(H) = 0
d
if H = Hisotr
.
Theorem 2. Let the assumptions of Theorem 1 be fulfilled and let K be a compactly supported
function. Then for any n ≥ 3, p > 1, q > 1, R > 1, D > 0, 0 < v ≤ v < ∞, u ∈ (p/2, ∞], u ≥ q and
any f ∈ Fg,u (R, D) ∩ Bq,d (D)
Rn(p) [fb~h(·) , f ]
≤C
(1)
lH (v) +
Z
v
v
v
p−1
1
p
1
{Λ(v, f ) ∧ Λ(v, f, u)}dv + Λp (v, f, u) + Cp n− 2 .
Here C (1) is a universal constant independent of f and n. Its explicit expression can be found in
the proof of the theorem. We remark also that only this constant depends on q.
8
The result announced in Theorem 2 suggests a way for establishing minimax and minimax
adaptive properties of the pointwise selection rule given in (2.3). For a given F ⊂ Fg,u (R, D) ∩
Bq,d (D) it mostly consists in finding a careful estimate for
S ~h, z := sup νd x ∈ Rd : B~h (x, f ) ≥ z , ∀~h ∈ H, ∀z > 0.
f ∈F
The choice of v, v > 0 is a delicate problem and it depends on S(·, ·).
In the next section we present several results concerning some useful upper estimates for the
quantities
sup{Λ(v, f ) ∧ Λ(v, f, u)}, sup Λp (v, f, u), v > 0.
f ∈F
f ∈F
We would like to underline that these bounds will be established for an arbitrary F and, therefore,
they can be applied to the adaptation over different scales of functional classes. In particular, the
results obtained below form the basis for our consideration in Part II.
2.3.2. Application to the minimax adaptive estimation
(p)
Our objective now is to bound from above supf ∈F Rn [fb~h(·) , f ] for any F ⊂ Fg,u (R, D) ∩ Bq,d (D).
All the results in this section will be proved under an additional condition imposed on the kernel
K.
R
Assumption 3. Let K : R → R be a compactly supported, bounded function and K = 1. Then
d
Y
K(xj ), ∀x ∈ Rd .
K(x) =
j=1
Without loss of generality we will assume that kKk∞ ≥ 1 and supp(K) ⊂ [−cK , cK ] with cK ≥ 1.
Introduce the following notations. Set for any h ∈ H, x ∈ Rd and j = 1, . . . , d
Z
∗
bh,f,j (x) =
K(u)f x + uhej ν1 (du) − f (x) , bh,f,j (x) = sup b∗η,f,j (x)
η∈H: η≤h
R
where (e1 , . . . , ed ) denotes the canonical basis of Rd . For any s ∈ [1, ∞] introduce
X
B∗j,s,F(h) = sup
b∗h,f,j s , Bj,s,F (h) = sup bh,f,j s , j = 1, . . . , d.
f ∈F h∈H: h≤h
f ∈F
Set for any ~h ∈ Hd , v > 0 and j = 1, . . . , d,
J ~h, v = j ∈ {1, . . . , d} : hj ∈ Vj (v) , Vj (v) = h ∈ H : Bj,∞,F(h) ≤ cv ,
−d
where c = (20d)−1 max(2cK kKk∞ , kKk1 ) . As usual the complement of J ~h, v will be denoted
by J¯ ~h, v . Furthermore, the summation over the empty set is supposed to be zero.
For any ~s = (s1 , . . . , sd ) ∈ [1, ∞)d , u ≥ 1 and v > 0 introduce
X
sj
−sj
−u
v
Bj,sj ,F hj
Λ~s (v, F, u) = inf
inf
+z
;
(2.7)
z≥2 ~h∈H(v,z)
Λ~s v, F = inf
~h∈H(v)
X
¯ ~h,v)
j∈J(
v
−sj
¯ ~h,v)
j∈J(
9
Bj,sj ,F hj
sj
+v
−2
Fn2
~h .
(2.8)
Theorem 3. Let assumptions of Theorem 2 be fulfilled and suppose additionally that K satisfies
Assumption 3. Then for any n ≥ 3, p > 1, q > 1, R > 1, D > 0, 0 < v ≤ v < ∞, u ∈ (p/2, ∞], u ≥
q, ~s ∈ (1, ∞)d , ~q ∈ [p, ∞)d and any F ⊂ Bq,d (D) ∩ Fg,u (R, D)
sup Rn(p) [fb~h(·) , f ]
f ∈F
≤C
(2)
lH (v) +
Z
v
v
p−1
v
If additionally q ∈ (p, ∞) one has also
sup Rn(p) [fb~h(·) , f ]
f ∈F
≤C
(2)
lH (v) +
≤C
(2)
v
v
1
p
1
p
Λ~s (v, F, u) ∧ Λ~s (v, F) dv + v Λ~q (v, F, u) + Cp n− 2 .
p−1
v
Moreover, if q = ∞ one has
sup Rn(p) [fb~h(·) , f ]
f ∈F
Z
lH (v) +
Z
v
v
p−1
v
Λ~s (v, F, u) ∧ Λ~s (v, F) dv + v p−q
1
p
1
+ Cp n− 2 .
1
p
1
Λ~s (v, F, u) ∧ Λ~s (v, F) dv + Λ~s (v, F, u) + Cp n− 2 .
d
Finally, if H = Hisotr
all the assertions above remain true for any ~s ∈ [1, ∞)d if one replaces in
(2.7)–(2.8) Bj,sj ,F (·) by B∗j,sj ,F (·).
It is important to emphasize that C (2) depends only on ~s, ~q, g, K, d, R, D, u and q. Note also that
the assertions of the theorem remain true if we minimize right hand sides of obtained inequalities
w.r.t ~s, ~q since their left hand sides are independent of ~s and ~q. In this context it is important to
realize that C (2) = C (2) (~s, · · · ) is bounded for any ~s ∈ (1, ∞)d but C (2) (~s, · · · ) = ∞ if there exists
d
j = 1, . . . , d such that sj = 1. Contrary to that C (2) (~s, · · · ) < ∞ for any ~s ∈ [1, ∞)d if H = Hisotr
and it explains in particular the fourth assertion of the theorem.
Note also that D, R, u, q are not involved in the construction of our pointwise selection rule.
That means that one and the same estimator can be actually applied on any
[
F⊂
Bq,d (D) ∩ Fg,u (R, D).
R,D,u,q
Moreover, the assertion of the theorem has a non-asymptotical nature; we do not suppose that the
number of observations n is large.
Discussion As we see, the application of our results to some functional class is mainly reduced to
the computation of the functions B∗j,s,F(·) j = 1, . . . , d, for some properly chosen s. Note however
that this task is not necessary for many functional classes used in nonparametric statistics, at least
for the classes defined by the help of kernel approximation. Indeed, a typical description of F can
be summarized as follows. Let λj : R+ → R+ , be such that λj (0) = 0, λj ↑ for any j = 1, . . . , d.
Then, the functional class, say FK ~λ(·), ~r can be defined as a collection of functions satisfying
bh,f,j
rj
≤ λj (h),
∀h ∈ H,
(2.9)
for some ~r ∈ [1, ∞]. It yields obviously
Bj,rj ,F (·) ≤ λj (·),
j = 1, . . . , d,
and the result of Theorem 3 remains valid if we replace formally Bj,rj ,F (·) by λj (·) in all the
expressions appearing in this theorem. In Part II we show that for some particular kernel K ∗ , the
10
~ L
~ is included into the class defined by (2.9) with λj (h) = Lj hβj ,
anisotropic Nikol’skii class N~r,d β,
~ L
~ and ~r.
whatever the values of β,
Denote ϑ = (~λ(·), ~r) and remark that in many cases FK [ϑ] ⊂ Bq,d (D) for any ϑ ∈ Θ for some
class parameter Θ and q ≥ p, D > 0. Then, replacing Bj,r
in (2.7) and (2.8) and
j ,F (·) by λj (·)
choosing ~q = (q, . . . , q) we come to the quantities Λ v, u, ϑ and Λq v, ϑ , completely determined
by the functions λj (·), j = 1, . . . , d, the vector ~r and the number q. Therefore, putting
1
Z v
p
1
p−1
p
p−q
lH (v) +
v
Λ(v, u, θ) ∧ Λ(v, θ) dv + v Λq v, u, ϑ ∧ v
ψn ϑ =
+ n− 2
inf
0<v≤v<∞
v
we deduce from the first and the second assertions of Theorem 3 for any ~λ(·) and ~r and n ≥ 3
sup Rn(p) [fb~h(·) , f ] ≤ C (3) ψn (ϑ).
(2.10)
f ∈FK [ϑ]
Since the estimator fb~h(·) is completely data-driven and, therefore, is independent of ~λ(·) and ~r, the
bound (2.10) holds for the scale of functional classes FK [ϑ] ϑ .
If φn FK [ϑ] is the minimax risk defined in (1.4) and
lim sup ψn (ϑ)φ−1
∀ϑ ∈ Θ,
(2.11)
n FK [ϑ] < ∞,
n→∞
we can assert that our estimator is optimally adaptive over the considered scale FK [ϑ], ϑ ∈ Θ .
To illustrate the powerfulness of our approach, let us consider a particular scale of functional
classes defined by (2.9).
~ ∈ (0, ∞)d be given vectors.
Classes of Hölderian type Let β~ ∈ (0, ∞)d and L
~ L
~ , where K satisfies Assumption
Definition 1. We say that a function f belongs to the class FK β,
3, if f ∈ B∞,d maxj=1,...,d Lj and for any j = 1, . . . , d
bh,f,j
∞
≤ Lj v β j ,
∀h ∈ H.
We remark that this class is a particular case of the one defined in (2.9), since it corresponds to
λj (h) = Lj hj and rj = ∞ for any j = . . . , d. Moreover let us introduce the following notations
2µj (α)+1
d
d
1
X
Y
2µj (α) + 1
1
β
2+1/β(α)
−1
Lj j .
=
, L(α) =
ϕn = δn
, δn = L(α)n ln (n),
β(α)
βj
j=1
j=1
Then the following result is a direct consequence of Theorem 3. Its simple and short proof is
postponed to Section 3.4.
Assertion 1. Let the assumptions of Theorem 3 be fulfilled. Then for any n ≥ 3, p > 1, β~ ∈ (0, ∞)d ,
~ ∈ [L0 , L∞ ]d there exists C > 0 independent of L
~ such that
0 < L0 ≤ L∞ < ∞ and L
~ L
~
lim sup ψn−1 β,
sup Rn(p) [fb~h(·) , f ] ≤ C,
n→∞
~L
~)
f ∈FK (β,
where we have denoted
t(H) (1−1/p)β(α)
ln(n) p δn β(α)+1 ,
2 + 1/β(α) > p;
1∨t(H) (1−1/p)β(α)
~ L
~ =
ψn β,
(2.12)
ln(n) p δn β(α)+1 ,
2 + 1/β(α) = p;
β(α)
2 + 1/β(α) < p.
δ 2β(α)+1 ,
n
11
It is interesting to note that the obtained bound, being a very particular case of our consideration
in Part II, is completely new if α 6= 0. As we already mentioned, for some particular choice of the
~ L
~ is included in the class FK ∗ ~λ(·), ~r with
kernel K ∗ , the anisotropic Nikol’skii class N~r,d β,
~ L
~ and ~r. Therefore, the aforementioned result holds on an
λj (v) = Lj vβj , whatever the values of β,
~ L
~ . Comparing the result of Assertion 1 with the lower bound for
arbitrary Hölder class N∞,d
β,
~
the minimax risk obtained in Lepski and Willer (2017), we can state that it differs only by some
logarithmic factor. Using the modern statistical language, we say that the estimator fb~h(·) is nearly
optimally-adaptive over the scale of Hölder classes.
3. Proofs
3.1. Proof of Theorem 1
The main ingredients of the proof of the theorem are given in Proposition 1. Their proofs are
postponed to Section 3.1.2. Introduce for any ~h ∈ Hd
n
1 X
M Zi − x, ~h − Ef M Zi − x, ~h ,
n
ξn x, ~h =
i=1
x ∈ Rd .
Proposition 1. Let Assumptions 1 and 2 be fulfilled. Then for any n ≥ 3 and any p > 1
Z
n
p o
p
Ef sup ξn x, ~h − Un x, ~h + νd (dx) ≤ Cp n− 2 ;
(i)
Rd
(ii)
Z
Rd
(iii)
Z
~h∈Hd
Ef
Rd
n
Ef
o
bn x, ~h − 3Un x, ~h p νd (dx) ≤ Cp′ n− p2 ;
sup U
+
~h∈Hd
n
o
bn x, ~h p νd (dx) ≤ C ′ n− p2 .
sup Un x, ~h − 4U
p
+
~h∈Hd
The explicit expression of constant Cp and Cp′ can be found in the proof.
3.1.1. Proof of the theorem
We start by proving the so-called pointwise oracle inequality.
Pointwise oracle inequality. Let ~h ∈ H and x ∈ Rd be fixed. We have in view of the triangle inequality
fb~h(x) (x) − f (x) ≤ fb~h(x)∨~h (x) − fb~h(x) (x) + fb~h(x)∨~h (x) − fb~h (x) + fb~h (x) − f (x) .
(3.1)
10 . First, note that obviously fb~h(x)∨~h (x) = fb~h∨~h(x) (x) and, therefore,
bn x, ~h(x) .
bn x, ~h(x) ∨ ~h + 4U
b~ (x) + 4U
fb~h(x)∨~h (x) − fb~h(x) (x) = fb~h∨~h(x) (x) − fb~h(x) (x) ≤ R
h
b ∗ x, ~η for any ~η ∈ Hd .
bn x, ~
η ≤U
Moreover by definition, U
n
bn x, ~h ∨ ~η ≤ U
b ∗ x, ~h ∧ U
b ∗ x, ~η . Thus, we obtain
Next, for any ~h, ~
η ∈ Hd we have obviously U
n
n
b ∗ x, ~h(x) .
b~ (x) + 8U
fb~h(x)∨~h (x) − fb~h(x) (x) ≤ R
n
h
12
(3.2)
Similarly we have
b ~ (x) + 8U
b ∗ x, ~h .
fb~h(x)∨~h (x) − fb~h (x) ≤ R
n
h(x)
(3.3)
The definition of ~h(x) implies that for any ~h ∈ H
b ~ (x) + 8U
b ∗ x, ~h(x) + R
b~ (x) + 8U
b ∗ x, ~h ≤ 2R
b~ (x) + 16U
b ∗ x, ~h
R
n
n
n
h(x)
h
h
and we get from (3.1), (3.2) and (3.3) for any ~h ∈ H
b~ (x) + 16U
bn∗ x, ~h + fb~ (x) − f (x) .
fb~h(x) (x) − f (x) ≤ 2R
h
h
(3.4)
20 . We obviously have for any ~h, ~
η ∈ Hd
fb~h∨~η (x) − fb~η (x) ≤ Ef M Z1 − x, ~h ∨ ~η − Ef M Z1 − x, ~η + ξn x, ~h ∨ ~η + ξn x, ~η .
Note that for any h ∈ Hd
Ef M Z1 − x, ~h
:=
=
Z
M t − x, ~h p(t)νd (dt)
Rd
Z
Z
M t − x, ~h f (t)νd (dt) + α
(1 − α)
Rd
Rd
M t − x, ~h f ⋆ g (t)νd (dt),
in view of the structural assumption (1.1) imposed on the density p. Note that
Z
Z
~
M t − x, ~h f ⋆ g (t)νd (dt)
M t − x, h f (t)νd (dt) + α
(1 − α)
d
Rd
Z R
Z
M u, ~h g(u − [z − x])νd (du) νd (dz)
f (z) (1 − α)M z − x, ~h + α
=
Rd
Rd
and, therefore, in view of the definition of M ·, ~h , c.f. (2.2), we obtain for any h ∈ Hd
Z
~
K~h (z − x)f (z)νd (dz) =: S~h (x, f ).
Ef M Z1 − x, h =
(3.5)
Rd
We deduce from (3.5) that
Ef M Z1 − x, ~h ∨ ~η − Ef M Z1 − x, ~η = S~h∨~η (x, f ) − S~η (x, f )
and, therefore, for any ~h, ~
η ∈ Hd
fb~h∨~η (x) − fb~η (x) ≤ S~h∨~η (x, f ) − S~η (x, f ) + ξn x, ~h ∨ ~η
30 . Set for any ~h ∈ Hd and any x ∈ Rd
υ(x) = sup
η
~ ∈Hd
bn x, ~h
̟1 (x) = sup Un x, ~h − 4U
~h∈Hd
ξn x, ~η
,
+
13
− Un x, ~η
+ ξn x, ~η .
+
bn x, ~h − 3Un x, ~h
̟2 (x) = sup U
+
~h∈Hd
(3.6)
We obtain in view of (3.6) that for any ~h ∈ H (since obviously ~h ∨ ~η ∈ Hd for any ~h, ~η ∈ Hd )
b~ (x) ≤ B ∗ (x, f ) + 2υ(x) + 2̟1 (x).
R
~h
h
(3.7)
Note also that in view of the obvious inequality (supα Fα − supα Gα )+ ≤ supα (Fα − Gα )+
∗
b x, ~h − 3U ∗ x, ~h
bn x, ~η − 3Un x, ~η
U
=: ̟2 (x)
≤
sup
U
n
n
+
+
(3.8)
η ∈Hd
~
We get from (3.4), (3.7) and (3.8)
fb~h(x) (x) − f (x) ≤ 2B~h∗ (x, f ) + 4υ(x) + 4̟1 (x) + 48Un∗ x, ~h + 16̟2 (x) + fb~h (x) − f (x) .
It remains to note that
fb~h (x) − f (x) ≤ B~h (x, f ) + ξn x, ~h
and we obtain for any ~h ∈ H and x ∈ Rd
≤ B~h (x, f ) + Un x, ~h + υ(x),
fb~h(x) (x) − f (x) ≤ 2B~h∗ (x, f ) + B~h (x, f ) + 5υ(x) + 4̟1 (x) + 49Un∗ x, ~h + 16̟2 (x).
Noting that the left hand side of the latter inequality is independent of ~h we obtain for any x ∈ Rd
n
o
fb~h(x) (x) − f (x) ≤ inf 2B~h∗ (x, f ) + B~h (x, f ) + 49Un∗ x, ~h + 5υ(x) + 4̟1 (x) + 16̟2 (x). (3.9)
~h∈H
This is the pointwise oracle inequality.
Application of Proposition 1. Set for any x ∈ Rd
n
o
Rn (x) = inf 2B~h∗ (x, f ) + B~h (x, f ) + 49Un∗ x, ~h
~h∈H
Applying Proposition 1 we obtain in view of (3.9) and the triangle inequality
Rn(p) [fb~h(·) , f ]
≤
Rn
≤
Rn
Z
+5
p
p
1
Ef υ(x)
Rd
− 21
+ Cp n
,
p
1
p
Z
+4
Rd
Ef ̟1 (x)
p
1
p
Z
+ 16
Rd
Ef ̟2 (x)
1
where Cp = 5(Cp ) p + 20(Cp′ ) p . The theorem is proved.
3.1.2. Proof of Proposition 1
Since the proof of the proposition is quite long and technical, we divide it into several steps.
14
p
1
p
Preliminaries 10 . We start the proof with the following simple remark. Let M̌ t, ~h , t ∈ Rd ,
denote the Fourier transform of M ·, ~h . Then, we obtain in view of the definition of M ·, ~h
−1
M̌ t, ~h = Ǩ t~h (1 − α) + αǧ(−t) , t ∈ Rd .
(3.10)
Note that Assumptions 1 and 2 guarantee that M̌ ·, ~h ∈ L1 Rd ∩ L2 Rd for any ~h ∈ Hd and,
therefore,
M ·, ~h ∞ ≤ (2π)−d M̌ ·, ~h 1 ,
M ·, ~h 2 = (2π)−d M̌ ·, ~h 2 .
Thus, putting
d
Y
−µj (α)
M∞ (~h = M∞
h−1
,
j (hj ∧ 1)
j=1
we obtain in view of Assumptions 1 and 2 for any ~h ∈ Hd
M ·, ~h
∞
≤ M∞ (~h ,
where M2 = (2π)−d ε−1 Ǩ
M ·, ~h
4
4
1
2 α6=1
≤ M ·, ~h
2
∞
M ·, ~h
2
+ Υ−1
0 k2 1α=1
M ·, ~h
2
2
≤
≤ M2
d
Y
j=1
−1
hj 2 (hj ∧ 1)−µj (α) ,
(3.11)
∨ 1. Additionally we deduce from (3.11)
2
M22 M∞
d
Y
j=1
∀~h ∈ Hd .(3.12)
−4µj (α)
h−3
,
j (hj ∧ 1)
Let L ·, ~h be either M ·, ~h or M 2 ·, ~h and let L∞ ~h denote either M∞ ~h or M2∞ ~h .
We have in view of (3.11)
P
2 2 dj=1 (1+µj (α))| ln(hj )|
~
~
≤ M∞
e
, ∀~h ∈ Hd .
L−1
∞ h ∨ L∞ h
(3.13)
Additionally, we get from (3.11) and (3.12)
L ·, ~h
Set
σL
2
2
2 4
≤ M22 M∞
e
Pd
j=1 (1+µj (α))| ln(hj )|
,
∀~h ∈ Hd .
(3.14)
qR
d
2
~
~
~
x, h =
Rd L t − x, h p(t)νd (dt) and note that in view of (3.14) for any h ∈ H
Z
Rd
σ L x, ~h
2
νd (dx) = L ·, ~h
Next, we have in view of (3.13)
σ L ·, ~h
∞
2
2
2 4
≤ M22 M∞
e
Pd
j=1 (1+µj (α))| ln(hj )|
P
2 2 dj=1 (1+µj (α))| ln(hj )|
≤ L∞ ~h ≤ M∞
e
.
20 . Define for any x ∈ Rd and ~h ∈ Hd
15
.
(3.15)
(3.16)
n
X
ξ L x, ~h = n−1
L Zi − x, ~h − EL Zi − x, ~h ;
i=1
d
X
~
.
1 + µj (α) ln(hj ) + 2 ln σ L x, ~h ∨ n−3/2 L∞ ~h
zn x, h = 3 ln(n) + (8p + 22)
j=1
s
4zn x, ~h L∞ ~h
2zn x, ~h
V
+
;
n
3n
s
~h
~h L∞ ~h
2λ
4λ
n
n
L
L
+
,
U x, ~h = σ x, ~h
n
3n
P
where remind λn ~h = 4 ln(M∞ ) + 6 ln (n) + (8p + 26) dj=1 1 + µj (α) ln(hj ) .
L
x, ~h = σ L x, ~h
Noting that supz∈[a,b] | ln z| ≤ | ln a| ∨ | ln b| for any 0 < a < b < ∞ we deduce from (3.16)
zn x, ~h ≤ λn ~h for any x ∈ Rd and, therefore, for any ~h ∈ Hd
V L x, ~h ≤ U L x, ~h .
(3.17)
First step Let x ∈ Rd and ~h ∈ Hd be fixed and put b = 8p + 22.
We obtain for any z ≥ 1 and q ≥ 1 by the integration of the Bernstein inequality
√
√ L ~
4zL∞ ~h q
4L∞ ~h q
2zσ L x, ~h
2σ x, h
L
~
√
√
Ef ζ x, h −
≤ 2Γ(q + 1)
exp {−z},
−
+
3n
3n
n
n
+
where Γ is the Gamma-function.
10 . Choose z = zn x, ~h . Noting that for any n ∈ N∗ and x ∈ Rd
√ L
1
2σ x, ~h
4L∞ ~h
√
+
≤ 3L∞ ~h n− 2
n
3n
and taking into account that exp {−| ln(y)|} ≤ y for any y > 0, we get
q
Ef ζ L x, ~h − V L x, ~h +
Pd
2
q
≤ 2 × 3q Γ(q + 1)n− 2 −3 Lq∞ ~h eb j=1 (1+µj (α))| ln(hj )| σ L x, ~h ∨ n−3/2 L∞ ~h
q
≤ Cq(1) n− 2 −3 e(2q−b)
Pd
j=1 (1+µj (α))| ln(hj )|
2
L
.
σ x, ~h ∨ n−3/2 L∞ ~h
(1)
(3.18)
2q q
Here to get the second inequality we have used (3.13) and put Cq = 2M∞
3 Γ(q + 1).
Set X ~h = x ∈ Rd : σ L x, ~h ≥ n−3/2 L∞ ~h , X̄ ~h = Rd \ X ~h and later on the integration
over the empty set is supposed to be zero.
We have in view of (3.17), (3.15) and (3.18) applied with q = p that for any ~h ∈ Hd
Z
P
L
p
(2) − p (2p+4−b) dj=1 (1+µj (α))| ln(hj )|
L
.
(3.19)
Ef ζ x, ~h − U x, ~h + νd (dx) ≤ Cp n 2 e
X ~h
(2)
(1)
2 .
where Cp = Cp M22 M∞
16
20 . Introduce the following notations. For any i = 1, . . . , n set
n
o
Ψi x, ~h = 1 L Zi − x, ~h − EL Zi − x, ~h ≥ n−1 L∞ ~h ,
nP
o
n
~h ≥ 2 . As usual, the complimentary
and introduce the random event D x, ~h =
x,
Ψ
i
i=1
event will be denoted by D̄ x, ~h . Set finally π x, ~h = Pf Ψ1 x, ~h = 1 .
We obviously have
ζ
x, ~h 1
L
and, therefore,
D̄
~h
3L
∞
≤
< U L ~h
~
x,h
n
ζ L x, ~h − U L x, ~h
1
D̄ x,~h
p
+
= 0.
Applying Cauchy-Schwartz inequality, we deduce from (3.20) that
h
p
Ef ζ L x, ~h − U L x, ~h + ≤ Ef ζ L x, ~h − U L x, ~h
Using (3.18) with q = 2p and (3.13) we obtain for any x ∈ X̄ ~h
Ef
ζ L x, ~h
− U L x, ~h
p
+
p
≤ Cp(3) n− 2 −3 e(2p+2−b/2)
(3.20)
2p
+
Pd
i 21
.
Pf D x, ~h
j=1 (1+µj (α))| ln(hj )|
h
i 12
Pf D x, ~h
, (3.21)
(1) 1 2
(3)
where we have put Cp = C2p 2 M∞
.
For any λ > 0 we have in view of the exponential Markov inequality
n
o
nX
n
= Pf
Ψi x, ~h ≥ 2 ≤ e−2λ eλ π x, ~h + 1 − π x, ~h
Pf D x, ~h
i=1
λ
n
(e − 1)π x, ~h + 1 ≤ exp{−2λ + n(eλ − 1)π x, ~h }.
L
~
~ 2 . It yields
We get applying the Tchebychev inequality π x, ~h ≤ n2 L−2
∞ h σ x, h
L
~
~ 2 (eλ − 1) , ∀~h ∈ Hd .
Pf D x, ~h ≤ exp − 2λ + n3 L−2
∞ h σ x, h
~h σ L x, ~h 2 < 1 for any x ∈ X̄ ~h . Hence,
Note that the definition of X̄ ~h implies n3 L−2
∞
~ L ~
choosing λ = ln 2 − 2 ln n3/2 L−1
we have
∞ h σ x, h
L
~ 4 , ∀x ∈ X̄ ~h .
Pf D x, ~h ≤ (e2 /4)n6 L−4
∞ σ x, h
−2λ
= e
It yields, together with (3.13), (3.15) and (3.21) and for any ~h ∈ Hd
Z
P
L
p
(4) − p (2p+10−b/2) dj=1 (1+µj (α))| ln(hj )|
L
(3.22)
Ef ζ x, ~h − U x, ~h + νd (dx) ≤ Cp n 2 e
X¯ ~h
(4)
(3)
(5)
(2)
(4)
6 M 2 . Putting C
where Cp = Cp (e/2)M∞
p = Cp + Cp and noting that 2p + 10 − b/2 < 0 we
2
~
obtain from (3.19) and (3.22) for any h ∈ Hd
Z
Pd
p
p
Ef ζ L x, ~h − U L x, ~h + νd (dx) ≤ Cp(5) n− 2 e(2p+10−b/2) j=1 (1+µj (α))| ln(hj )| . (3.23)
Rd
30 . Choosing L = M and L∞ = M∞ we get from (3.23) and the definition of b
Z
Pd
p
p
Ef ξn x, ~h − Un x, ~h + νd (dx) ≤ Cp(5) n− 2 e− j=1 (1+µj (α))| ln(hj )| , ∀~h ∈ Hd .(3.24)
Rd
(5) P
The first assertion of the proposition follows from (3.24) with Cp = Cp
17
k∈Zd
e−
Pd
j=1
|kj |
.
Second step
2
b x, ~h − σ 2 x, ~h − Un x, ~h + , where
Denoting χ x, ~h = σ
s
~h M2 ~h
~h
4λ
2λ
2
n
n
∞
M
+
,
x, ~h
Un x, ~h = σ
n
3n
and choosing L = M 2 and L∞ = M2∞ , we get from (3.23)
Z
Pd
p
Ef χp x, ~h νd (dx) ≤ Cp(5) n− 2 e(2p+10−b/2) j=1 (1+µj (α))| ln(hj )| ,
∀~h ∈ Hd . (3.25)
Rd
2
Note that σ M x, ~h ≤ M∞ ~h σ x, ~h and, therefore, for any x ∈ Rd and any ~h ∈ Hd
Un x, ~h ≤ M∞ ~h Un x, ~h .
(3.26)
This implies,
2
b x, ~h
2λn ~h σ
2λn ~h σ 2 x, ~h
2λn ~h M∞ ~h Un x, ~h
2λn ~h M∞ ~h χ∗ (x, ~h
≤
+
+
,
n
n
n
n
~
~
where we have denoted χ∗ (x, ~h = M−1
∞ h χ(x, h . Hence
s
2λn ~h M∞ ~h Un x, ~h + χ∗ (x, ~h
~
~
b
Un x, h ≤ Un x, h +
.
(3.27)
n
By the same reason
s
~h M∞ ~h Un x, ~h + χ∗ x, ~h
2λ
n
bn x, ~h +
.
Un x, ~h ≤ U
n
bn x, ~h and Un x, ~h implies that
Note that the definition of U
(3.28)
2λn ~h M∞ ~h
bn x, ~h , Un x, ~h .
(3.29)
≤ (3/2) min U
n
p
Using the inequality |ab| ≤ 2−1 (|ay| + |b/y|), y > 0 we get from (3.27), (3.28) and (3.29)
p
bn x, ~h ≤ 1 + 3/2 + (3/4)y Un x, ~h + (2y)−1 χ∗ x, ~h ;
U
bn x, ~h + (2y)−1 Un x, ~h + (2y)−1 χ∗ x, ~h .
Un x, ~h ≤ 1 + (3/4)y U
Choosing y = 1/2 in the first inequality and y = 1 in the second we get for any x ∈ Rd and ~h ∈ Hd
∗
~h ;
bn x, ~h − 3Un x, ~h
≤
χ
(x,
(3.30)
U
+
∗
bn x, ~h
≤ χ (x, ~h .
(3.31)
Un x, ~h − 4U
+
Remembering that b = 8p + 22 we obtain from (3.30), (3.31), (3.25) and (3.13) for any ~h ∈ Hd
Z
P
2p (5) − p2 − dj=1 (1+µj (α))| ln(hj )|
bn x, ~h − 3Un x, ~h p νd (dx) ≤ M∞
C
n
Ef U
e
;
(3.32)
p
+
d
ZR
P
2p (5) − p2 − dj=1 (1+µj (α))| ln(hj )|
bn x, ~h p νd (dx) ≤ M∞
C
n
e
.
(3.33)
Ef Un x, ~h − 4U
p
+
Rd
(5)
2p
Cp .
The second and third assertions follow from (3.32) and (3.33) with Cp′ = M∞
18
3.2. Proof of Theorem 2
Let f ∈ Fg,u (R, D). Introduce the following notations:
√
where c1 = M2 2D, c2 =
3.2.1. Preliminaries
4M∞
3
n √
o−1
a = 196 (c1 c3 ) ∨ (c2 c3 )
,
and c3 = 2 max 4 ln(M∞ ), (8p + 26) maxj=1,...,d [1 + µj (α)] .
Recall that for any locally integrable function λ : Rd → R its strong maximal function is defined as
Z
1
M[λ](x) := sup
λ(t)dt, x ∈ Rd ,
(3.34)
H νd (H) H
where the supremum is taken over all possible rectangles H in Rd with sides parallel to the coordinate axes, containing point x.
It is well known that the strong maximal operator λ 7→ M[λ] is of the strong (t, t)–type for all
1 < t ≤ ∞, i.e., if λ ∈ Lt (Rd ) then M[λ] ∈ Lt (Rd ) and there exists a constant Ct depending on t
only such that
(3.35)
M[λ] t ≤ Ct kλkt , t ∈ (1, ∞].
Let m[λ] be defined by (3.34), where, instead of rectangles, the supremum is taken over all possible
cubes H in Rd with sides parallel to the coordinate axes, containing point x. Then, it is known
that λ 7→ m[λ] is of the weak (1, 1)-type, i.e. there exists C1 depending on d only such that for any
λ ∈ L1 (Rd )
n
o
νd x : m[λ](x) ≥ z ≤ C1 z−1 kλk1 , ∀z > 0.
(3.36)
The results presented below deal with the weak property of the strong maximal function. The
following inequality can be found in Guzman (1975). There exists a constant C > 0 depending on
d only such that
Z
n
o
νd x : M[λ](x) ≥ z ≤ C
Rd
|λ(x)|
|λ(x)| d−1
1 + ln+
dx,
z
z
z > 0,
where for all z ∈ R, ln+ (z) := max{ln(z), 0}.
Lemma 1. For any given d ≥ 1, R > 0, Q > 0 and q ∈ (1, ∞] there exists C(d, q, R, Q) such that
for any λ ∈ B1,d (R) ∩ Bq,d (Q)
n
o
d−1
νd x : M[λ](x) ≥ z ≤ C(d, q, R, Q)z−1 1 + | ln(z)|
, ∀z > 0.
The proof of the lemma is an elementary consequence of the aforementioned result and can be
omitted.
Recall also the particular case of the Young inequality for weak-type spaces, see Grafakos (2008),
Theorem 1.2.13.
For any u ∈ (1, ∞] there exists Cu > 0 such that for any λ1 ∈ L1 Rd and
λ2 ∈ Lu,∞ Rd one has
kλ1 ⋆ λ2 ku,∞ ≤ Cu kλ1 k1 kλ2 ku,∞ .
(3.37)
19
Auxiliary results
Let us prove several simple facts. First note that for any n ≥ 3 for any ~h ∈ Hd
d
i
h
X
ln(ηj ) .
λn ~h ≤ c3 ln (n) +
(3.38)
j=1
Second it is easy to see that for any any n ≥ 3,
r
~
η ≤ Fn h l V~η /V~h , Gn ~η ≤ Gn ~h l V~η /V~h ,
Fn ~
∀~η, ~h ∈ (0, ∞)d : ~η ≥ ~h,
where l(v) = v −1 (1 + ln v). Since η~ ≥ ~h implies V~η ≥ V~h and l(v) ≤ 1 if v ≥ 1, we have
Fn ~η ≤ Fn ~h , Gn ~η ≤ Gn ~h , ∀~η, ~h ∈ (0, ∞)d : η~ ≥ ~h.
Then by (3.38) and the second inequality in (3.39), we have:
4M∞ λn ~η
~h .
sup
≤
c
c
G
Qd
2
3
n
µj (α)
η ∈Hd : ~
~
η ≥~h 3n
j=1 ηj (ηj ∧ 1)
(3.39)
(3.40)
Now let us establish two bounds for kUn∗ ·, ~h k∞ .
10 a. Let u = ∞. We have in view of the second inequality in (3.11) for any ~η ∈ Hd
√
η
σ x, ~
η ≤ D M ·, ~
2
d
√ Y
−1
≤ M2 D
ηj 2 (ηj ∧ 1)−µj (α) ,
j=1
∀x ∈ Rd .
It yields for any x ∈ Rd in view of the first inequality in (3.39)
s
η σ 2 x, ~
η
2λn ~
√
√
≤ sup c1 c3 Fn ~η ≤ c1 c3 Fn ~h .
sup
n
η ∈Hd : ~
~
η≥~h
η ∈Hd : ~
~
η ≥~h
(3.41)
Then gathering (3.40), (3.41) and by definition of a, we have
(3.42)
kUn∗ ·, ~h k∞ ≤ (196a)−1 Fn ~h + Gn ~h .
10 b. Another bound for kUn∗ ·, ~h k∞ is available regardless of the value of u. Indeed for any
~η ∈ Hd in view of the first inequality in (3.11)
σ x, ~
η ≤ M ·, ~
η
∞
≤ M∞
d
Y
j=1
ηj−1 (ηj ∧ 1)−µj (α) ,
∀x ∈ Rd .
It yields for any x ∈ Rd and any n ≥ 3
q
s
P
2
ln (n) + dj=1 ln(ηj )
√
√
η
2λn ~η σ x, ~
~h .
sup
2c3 M∞ √ Qd
2c
nM
G
≤ sup
≤
3
∞
n
n
n j=1 ηj (ηj ∧ 1)µj (α)
η ∈Hd : ~
~
η ≥~h
η∈Hd : ~
~
η ≥~h
√
Then gathering with (3.40) again, we have kUn∗ ·, ~h k∞ ≤ ( 2c3 nM∞ ) ∨ (c2 c3 ) Gn ~h for any
~h ∈ Hd and, therefore,
inf kUn∗ ·, ~h k∞ = 0.
(3.43)
~h∈H
20
To get this it suffices to choose ~h = (b, . . . , b) and to make b tend to infinity.
20 . Let now u < ∞. Let us prove that for any z > 0, s ∈ {1, u} and any f ∈ Fg,u (R, D)
−2 2 s
d
e
νd x ∈ R :
sup
Un x, ~η , f ≥ z ≤ c5 Dz
Fn ~h ,
(3.44)
η
~∈Hd : ~
η ≥~h
e = 1 if s = 1 and D
e = D if s = u.
where we have put Un2 ·, ~
η , f = 2n−1 λn ~η σ 2 ·, ~η and D
Indeed, if s = 1, applying the Markov inequality, we obtain in view of the second inequality in
η ∈ Hd
(3.11) for any ~
Z
2 −1
d
σ 2 x, ~η νd (dx)
(3.45)
η , f ≥ z ≤ 2(nz ) λn ~η
νd x ∈ R : Un x, ~
Rd
2
λn ~η
−2 2
= 2(nz2 )−1 λn ~η M ·, ~
η 2 ≤ 2M22 (nz2 )−1 Qd
≤
c
z
F
~
η
.
6
n
2µj (α)
η
(η
∧
1)
j
j
j=1
Here we have put c6 = 2M22 c21 c3 and to get the last inequality
we 2haveused (3.38).
2
To get
the similar result if s = u we remark that σ ·, ~η = M ·, ~η ⋆ p(·) and that M 2 ·, ~
η ∈
L1 Rd in view of the second inequality in (3.11). It remains to note that f ∈ Fg,u (R, D) implies
(∞)
p ∈ Bu,d (D) and to apply the inequality (3.37).
It yields together with the second inequality in (3.11) for any ~η ∈ Hd
u
νd x ∈ Rd : Un x, ~η , f ≥ z ≤ c6 Cu Dz−2 Fn2 ~η .
e = 1 if s = 1 and C
e = Cu if s = u, we get from (3.45) and (3.46)
Thus, denoting C
X
e Dz
e −2 s
Fn2s η~ .
νd x ∈ R d :
sup
Un x, ~η, f ≥ z ≤ c6 C
η ∈Hd : η
~
~≥~h
(3.46)
(3.47)
η ∈Hd : ~
~
η ≥~h
It remains to note that since ~
η , ~h ∈ Hd and ~η ≥ ~h we can write ηj = emj hj with mj ≥ 0 for any
j = 1, . . . , d. It yields together with the first inequality in (3.39)
X
η∈Hd : ~
~
η ≥~h
Fn2s ~η ≤ Fn2s ~h
X
(m1 ,...,md )∈Nd
1+
d
X
j=1
mj
s
e−s
Pd
j=1
mj
=: c7 Fn2s ~h .
e s follows from (3.47).
Hence, (3.44) with c5 = c7 [c6 C]
30 . Let cK ≥ 1 be such that supp(K) ⊂ [−cK , cK ]d . We have
Z
K~h (t − x)f (t)νd (dt) ≤ (2cK )d kKkd∞ M[|f |](x),
S~h (x, f ) =
Rd
∀~h ∈ (0, ∞)d .
If ~h = (h, . . . , h), h ∈ (0, ∞), the latter inequality holds with m[|f |] instead of M[|f |]. Thus,
sup B~h (x, f ) ≤ 3(2cK )d kKk∞ MH [|f |](x) + |f (x)|,
~h∈H
∀x ∈ Rd ,
(3.48)
d
where we have denoted MH = M if H = Hd and MH = m if H = Hisotr
.
Moreover, we deduce from (3.48) and (3.43) putting T~h (x, f ) = B~h (x, f ) + 49Un∗ ·, ~h that
inf T~h (x, f ) ≤ 3(2cK )d kKk∞ MH [|f |](x) + |f (x)|.
~h∈H
21
(3.49)
3.2.2. Proof of the theorem
For any v > 0 set Cv (f ) = x ∈ Rd : T(x, f ) ≥ v , where we have put T(x, f ) = inf~h∈H |T~h (x, f )|.
For any given v > 0 one obviously has
Z
Z v
p−1
p
v νd Cv (f ) dv +
|T(x, f )|p νd (dx)
(3.50)
kT(·, f )kp ≤ p
Cv (f )
0
Denoting Wv (~h, f ) = {x ∈ Rd : 49Un∗ x, ~h ≥ 2−1 v} we obviously have for any ~h ∈ H and v > 0
νd Cv (f ) ≤ νd A(~h, f, v) + νd Wv (~h, f ) ;
(3.51)
p
p
|T(x, f )|p 1Cv (f ) (x) ≤ 2p B~h (x, f ) 1A(~h,f,v) + 98p Un∗ x, ~h 1Wv (~h,f ) (x);
νd Cv (f ) ≤ νd x ∈ Rd : 3(2cK )d kKk∞ MH [|f |](x) + |f (x)| > v .
The last inequality follows from (3.49). Set Un∗ x, ~h, f = sup~η∈Hd : ~η≥~h Un x, ~η, f .
10 . Noting that Un∗ x, ~h ≤ Un∗ x, ~h, f + (196a)−1 Gn ~h in view of (3.40), we get
n
o
fv (~h, f ), ∀~h ∈ H(v).
Wv (~h, f ) ⊆ x ∈ Rd : 49Un∗ x, ~h ≥ 4−1 v := W
Applying (3.44) with s = 1 we deduce from (3.51) that
νd Cv (f ) ≤ νd A(~h, f, v) + 1962 c5 v −2 Fn2 ~h ,
(3.52)
(3.53)
(3.54)
∀~h ∈ H(v).
Noting that the left hand side of the latter inequality is independent of ~h we get
νd Cv (f ) ≤ max[1, 1962 c5 ]Λ(v, f ).
(3.55)
and for any u ∈ (p/2, ∞],
Z
|T(x, f )|p νd (dx) ≤ max[2p , 98p c9 ]Λp v, f, u ,
(3.57)
20 . Let us establish the following bounds, where c9 is given in the paragraph 20 b. below.
For any u ∈ [1, ∞],
(3.56)
νd Cv (f ) ≤ max[1, c5 1962 , c5 1962u D u a2u ]{Λ(v, f ) ∧ Λ(v, f, u)}, ∀v > 0.
Cv (f )
∀v > 0.
20 a. Let u = ∞. We remark that Wv (~h, f ) = ∅ for any ~h ∈ H(v, 2) in view of (3.42). Thus, we
deduce from (3.51), (3.52) and (2.6), taking into account that the left hand sides of both inequalities
are independent of ~h
Z
|T(x, f )|p νd (dx) ≤ Λp (v, f, ∞).
(3.58)
νd Cv (f ) ≤ Λ(v, f, ∞),
Cv (f )
This inequality and (3.55) ensure that (3.56) and (3.57) hold if u = ∞.
22
20 b. Let u < ∞. Applying (3.44) with s = u, we obtain in view of (3.54)
νd Wv (~h, f ) ≤ c5 1962u D u v −2u Fn2u ~h ≤ c5 1962u D u a2u z −u , ∀~h ∈ H(v, z)
It yields together with (3.51)
νd Cv (f ) ≤ max[1, c5 1962u D u a2u ]Λ(v, f, u).
(3.59)
This inequality and (3.55) ensure that (3.56) holds if u < ∞.
What is more, we have in view of (3.40) and (3.54) for any ~h ∈ H(v)
p
p
Un∗ x, ~h 1Wv (~h,f ) ≤ 2p Un∗ x, ~h, f 1W
fv (~h,f )
Moreover, applying (3.44) with s = u, we have for any y > 0 and ~h ∈ H(v, z)
fy (~h, f ) ≤ c5 1962u D u y −2u F 2u ~h ≤ c5 1962u D u y −2u (av)2u z −u .
νd W
n
Hence, if additionally u > p/2, we have for any ~h ∈ H(v, z)
Z ∞
Z
p
∗
p
~
fy (~h, f ) dy
y p−1 νd W
Un x, h νd (dx) ≤ 2 p
v
Wv (~h,f )
Z ∞
= c5 1962u D u a2u 2p pv 2u z −u
y p−1−2u dy =: c9 v p z −u .
v
This yields together with (3.52)
Z
|T(x, f )|p νd (dx) ≤ max[2p , 98p c9 ]Λp v, f, u .
(3.60)
Cv (f )
This inequality ensures that (3.57) holds if u < ∞.
30 . Recall that f ∈ Fg (R) implies that f ∈ B1,d (R). Since additionally f ∈ Bq,d (D), q > 1,
Lemma 1 as well as (3.36) is applicable and we obtain in view of (3.53)
νd Cv (f ) ≤ c10 v −1 (1 + | ln v|)t(H) , ∀v > 0.
It yields for any v > 0 and p > 1
Z
Z v
p−1
v νd Cv (f ) dv ≤ c10 p
p
0
v
0
v p−2 (1 + | ln v|)t(H) dv ≤ c11 v p−1 (1 + | ln v|)t(H) . (3.61)
In the case of t(H) = 0 the last inequality is obvious and if t(H) = d − 1 it follows by integration
by parts. The assertion of the theorem follows now from (3.50), where the bound (3.61) is used for
any v < v, the estimate (3.56) for any v ∈ [v, v] and the bound (3.57) with v = v.
3.3. Proof of Theorem 3
The proof of the theorem is based essentially on some auxiliary statements formulated in Section
3.3.1 below.
Some properties related to the kernel approximation of the underlying function f are summarized
in Lemma 2 and in formulae (3.62). The results presented in Lemma 1 and in formulae (3.63) deal
with the properties of the strong maximal function. In the subsequent proof c1 , c2 , . . ., stand for
constants depending only on ~s, ~
q , g, K, d, R, D, u and q.
23
3.3.1. Auxiliary results
Let J denote the set of all the subsets of {1, . . . d} endowed with the empty set ∅. For any J ∈ J and
y ∈ Rd set yJ = {yj , j ∈ J} ∈ R|J| and we will write y = yJ , yJ¯ , where as usual J¯ = {1, . . . d} \ J.
For any j = 1, . . . , d introduce the d × d matrix P
Ej = (0, . . . , ej , . . . , 0) where, recall, (e1 , . . . , ed )
denotes the canonical basis of Rd . Set also E[J] = j∈J Ej . Later on E0 = E[∅] denotes the matrix
with zero entries.
To any J ∈ J and any λ : Rd → R associate the function
λJ yJ , zJ¯ = λ z + E[J](y − z) , y, z ∈ Rd ,
with the obvious agreement λJ ≡ λ if J = {1, . . . d}, Q
which is always the
case if d = 1.
d
For any ~h ∈ Hd and J ⊆ {1, . . . d} set K~h,J (uJ ) = j∈J h−1
K
u
/h
j
j and define for any y ∈ R
j
Z
K~h,J¯(uJ¯ − yJ¯)λ yJ , uJ¯ ν|J|
K~h ◦ λ J (y) =
¯ duJ¯ ,
R|J¯|
¯
|J|
where ν|J|
h, ~η ∈ Hd set
¯ is the Lebesgue measure on R . For any ~
B~h,~η (x, f ) = |S~h∨~η (x, f ) − S~η (x, f )|.
Lemma 2. Let Assumption 3 hold. One can find k ∈ {1, . . . d} and a collection of indexes j1 <
j2 < · · · < jk ∈ {1, . . . , d} such that for any x ∈ Rd and any f : Rd → R
B~h,~η (x, f ) ≤
B~h (x, f ) ≤
k h
X
l=1
k h
X
l=1
K~h∨~η ◦ bhjl ,f,jl
K~h ◦ bhjl ,f,jl
i
Jl
i
Jl
(x),
(x) +
h
K~η ◦ bhjl ,f,jl
i
Jl
(x) ;
Jl = {j1 , . . . , jl }.
The proof of the lemma can be found in Lepski (2015), Lemma 2.
Also, let us mention the following bound which is a trivial consequence of the Young inequality
and the Fubini theorem. If λ ∈ Lt (Rd ) then for any t ∈ [1, ∞]
(3.62)
sup K~h ◦ λ J t ≤ kKkd1 kλkt , ∀~h ∈ Hd .
J∈J
To any J ∈ J and any locally integrable function λ : Rd → R+ we associate the operator
Z
1
(3.63)
λ t + E[J][x − t] ν|J|
MJ [λ](x) = sup
¯ (dtJ¯)
¯ (H|J|
¯ ) H|J|
H|J¯| ν|J|
¯
¯
¯ and with
where the supremum is taken over all hyper-rectangles in R|J| containing xJ¯ = (xj , j ∈ J)
sides parallel to the axis.
As we see MJ [λ] is the strong maximal operator applied to the function obtained from λ by fixing
the coordinates whose indices belong to J. It is obvious that M∅ [λ] ≡ M[λ] and M{1,...,d} [λ] ≡ λ.
The following result is a direct consequence of (3.35) and of the Fubini theorem. For any t ∈ (1, ∞]
there exists Ct such that for any λ ∈ Lt Rd )
sup MJ [λ]
J∈J
t
≤ Ct kλkt .
Obviously this inequality holds if t = ∞ with C∞ = 1.
24
(3.64)
3.3.2. Proof of the theorem
10 . We start with the following obvious observation. For any λ : Rd → R+ , ~u ∈ Rd and J ∈ J
[K~u ◦ λ]J (x) ≤ (2cK kKk∞ )d MJ [λ](x),
∀x ∈ Rd .
(3.65)
Putting C1 = (2cK kKk∞ )d we get for any ~h, ~η ∈ Hd and x ∈ Rd in view of (3.65) and assertions of
Lemma 2 that
B~h,~η (x, f ) ≤ 2C1
d
X
sup MJ bhj ,f,j (x),
j=1 J∈J
B~h (x, f ) ≤ C1
d
X
sup MJ bhj ,f,j (x).
j=1 J∈J
Thus noting that the right hand side of the first inequality above is independent of ~η , we obtain
B~h (x, f ) ≤ 5C1
d
X
sup MJ bhj ,f,j (x),
j=1 J∈J
∀x ∈ Rd , ∀~h ∈ Hd .
(3.66)
Applying (3.64) with t = ∞, we have for any v > 0 in view of the definition of J(~h, v)
X
X
sup MJ bhj ,f,j (x) + 5C1
sup MJ bhj ,f,j
B~h (x, f ) ≤ 5C1
¯ ~h,v)
j∈J(
≤ 5C1
≤ 5C1
X
¯ ~h,v)
j∈J(
X
¯ ~h,v)
j∈J(
J∈J
sup MJ bhj ,f,j (x) + 5C1
J∈J
j∈J(~h,v)
X
∞
J∈J
Bj,∞,F hj
j∈J(~h,v)
sup MJ bhj ,f,j (x) + 4−1 v,
J∈J
∀f ∈ F.
(3.67)
We obtain for any f ∈ F, v > 0 and ~s = (s1 , . . . , sd ) ∈ (1, ∞)d , applying consecutively the Markov
inequality and (3.64) with t = sj ,
o
n
n
o
−1
≤ νd ∪J∈J ∪j∈J(
νd A ~h, f, v
¯ ~h,v) x : 5C1 MJ bhj ,f,j (x) ≥ (4d) v
h
X
X
is j
s
v −sj Bj,sj ,F hj
.
(3.68)
v −sj bhj ,f,j sj ≤ c1
≤ c1
j
¯ ~h,v)
j∈J(
¯ ~h,v)
j∈J(
Noting that the right hand side of the latter inequality is independent of f and the left hand side
is independent of ~s, we get
c−1
s (v, F),
1 sup{Λ(v, f ) ∧ Λ(v, f, u)} ≤ Λ~s (v, F, u) ∧ Λ~
f ∈F
∀v > 0, ~s ∈ (1, ∞)d .
(3.69)
20 . Note also that in view of (3.67), we have for any v > 0
Z
p
B~h (x, f ) νd (dx)
A ~h,f,v
≤ c2
≤ c4
Z
A ~h,f,v
"
¯ ~h,v)
j∈J(
X Z
¯ ~h,v)
j∈J(
X
A ~h,f,v
p
n
o
sup MJ bhj ,f,j (x) νd (dx) + c3 v p νd A ~h, f, v
J∈J
p
n
p
sup MJ bhj ,f,j (x) νd (dx) + v νd A ~h, f, v
J∈J
25
o
#
.
(3.70)
For any v > 0 and j = 1, . . . , d, introduce
o
n
Aj (v) = x ∈ Rd : sup MJ bhj ,f,j (x) ≥ (40C1 )−1 v ,
J∈J
Aj (v) = A ~h, f, v ∩ Āj (v).
¯ ~h, v)
Noting that in view of (3.67) for any v > 0 and any j ∈ J(
(
)
X
d
sup MJ bhj ,f,k (x) ≥ v/8
Aj (v) ⊆
x ∈ R : 5C1
¯ ~h,v), k6=j
k∈J(
(
x ∈ Rd : 5C1
⊆
X
¯ ~h,v)
k∈J(
J∈J
sup MJ bhj ,f,k (x) ≥ v/8
J∈J
q ∈ [p, ∞)d
we deduce from (3.70) that for any ~
Z
X Z
p
B~h (x, f ) νd (dx) ≤ c4
A ~h,f,v
≤ c6
X
Aj (v)
¯ ~h,v)
j∈J(
v p−qj
=: A∗ ~h, f, v ,
p
sup MJ bhj ,f,j (x) νd (dx)
J∈J
n
h n
o i
o
+c5 v p νd A∗ ~h, f, v + νd A ~h, f, v
n
h n
o i
o
qj
. (3.71)
+ c5 v p νd A∗ ~h, f, v + νd A ~h, f, v
sup MJ bhj ,f,j
qj
J∈J
¯ ~h,v)
j∈J(
)
It remains to note that similarly (3.68) for any ~s ∈ (1, ∞)d
n
X
o
v −sj bhj ,f,j
νd A∗ ~h, f, v ≤ c7
sj
sj
¯ ~h,v)
j∈J(
and to apply (3.64) with t = qj to the each term in the sum appeared in (3.71). All of this together
with (3.68), applied with ~s = ~
q yields for any v > 0 and ~q ∈ [p, ∞)d
Z
h
X
X
iqj
p
qj
p−qj
p−qj
.
h
B
v
b
B
(x,
f
)
ν
(dx)
≤
c
v
≤
c
~h
j,qj ,F j
9
9
hj ,f,j qj
d
A ~h,f,v
¯ ~h,v)
j∈J(
¯ ~h,v)
j∈J(
Noting that the right hand side of the latter inequality is independent of f and the left hand side
is independent of ~
q , the we get
sup Λp (v, f, u) ≤ c9 v p Λ~q (v, F, u),
f ∈F
∀v > 0, ~q ∈ [p, ∞)d .
(3.72)
The first assertion of the theorem follows from (3.69), (3.72) and Theorem 2.
30 . Remark that in view of (3.48) and (3.35) f ∈ Bq,d (D) implies
B~h (·, f )
q
≤ 3(2cK )d kKkd∞ Cq + 1 D,
∀~h ∈ (0, ∞)d ,
where Cq is the constant which appeared in (3.35). Hence for any v > 0 and q ∈ [p, ∞)
Z
p
q
q−p p−q
v
B~h (·, f ) q ≤ c10 v p−q .
B~h (x, f ) νd (dx) ≤ 2
A ~h,f,v
26
(3.73)
(3.74)
Remind that H(v) 6= ∅, H(v, z) 6= ∅ whatever v > 0 and z ≥ 2, see Remark 2. Hence, in view of
(3.74) for any f
Λp (v, f, u) ≤ inf c10 v p−q + z −u = c10 v p−q .
z≥2
It remains to note that the right hand side of the obtained inequality is independent of f and the
second assertion of the theorem follows from this inequality, (3.69) and Theorem 2.
40 . Since C∞ = 1 we obtain in view of (3.73) for all f ∈ B∞,d (D)
B~h (·, f )
∞
≤ 3(2cK )d kKkd∞ + 1 D,
∀~h ∈ (0, ∞)d .
It yields for any ~s ∈ (1, ∞) in view of (3.68) if q = ∞
Z
n
o
p
~
≤ c12
ν
(dx)
≤
c
ν
A
h,
f,
v
(x,
f
)
B
~h
11 d
d
A ~h,f,v
X
¯ ~h,v)
j∈J(
h
isj
.
v −sj Bj,s,F hj
Since the left hand side of the obtained inequality is independent of f and the left hand side is
independent of ~s we conclude that
sup Λp (v, f, u) ≤ c12 Λ~s (v, F, u),
f ∈F
∀v > 0, ~s ∈ (1, ∞)d .
(3.75)
The third assertion of the theorem follows now from (3.69), (3.75) and Theorem 2.
50 . We have already seen (Corollary 1), that B~∗ (·, f ) ≤ 2 supη∈H:η≤h B~η (·, f ) if ~h = (h, . . . , h) ∈
h
d
Hisotr
. Therefore by definition of B~h (·, f ):
B~h (·, f ) ≤ 5
sup
η∈H:η≤h
B~η (·, f ) ≤ 5
sup
d
X
sup K~η ◦ b∗η,f,j J (x).
(3.76)
η∈H:η≤h j=1 J∈J
d
where, remind ~
η = (η, . . . , η) ∈ Hisotr
. We remark that (3.76) is similar to (3.66) but the maximal
operator is not involved in this bound. This, in its turn, allows to consider ~s ∈ [1, ∞)d .
Indeed, similarly to (3.67) we have for any v > 0, applying (3.62) with t = ∞
X
(3.77)
sup K~η ◦ b∗η,f,j J (x) + 4−1 v, ∀f ∈ F.
B~h (x, f ) ≤ 5 sup
η∈H:η≤h
¯ ~h,v)
j∈J(
J∈J
We obtain for any f ∈ F, v > 0 and ~s = (s1 , . . . , sd ) ∈ [1, ∞)d applying consecutively the Markov
inequality and (3.62) with t = sj
h
X
X
X
is j
s
v −sj
.
v −sj B∗j,sj ,F h
b∗η,f,j sj ≤ c14
νd A ~h, f, v ≤ c13
j
¯ ~h,v)
j∈J(
η∈H:η≤h
¯
j∈J(h,y)
We note that the obtained inequality coincides with (3.68) if one replaces Bj,sj ,F (·) by B∗j,sj ,F (·). It
remains to remark that Bj,sj ,F (·) ≤ B∗j,sj ,F (·). Indeed,
bv,f,j (x) = lim
sup
k→∞ h∈H: e−k ≤h≤v
27
b∗h,f,j (x).
Therefore, by the monotone convergence theorem and the triangle inequality for any s ∈ [1, ∞)
Bj,s,F (h) := sup kbv,f,j ks = sup lim
f ∈F k→∞
f ∈F
≤
X
sup lim
f ∈F k→∞
sup
h∈H: e−k ≤h≤h
b∗h,f,j
h∈H: e−k ≤h≤h
s
b∗h,f,j
= sup
s
X
f ∈F h∈H: h≤h
b∗h,f,j
s
=: B∗j,s,F(v).
The fourth statement of the theorem follows now from (3.69), (3.72), (3.74) and Theorem 2.
3.4. Proof of Assertion 1
~ L
~ ⊂ B∞,d (L∞ ). Thus, we can choose D = L∞ and q = ∞, which implies u = ∞.
Obviously FK β,
~
For any v > 0 let h(v)
= h1 (v), . . . , hd (v) , where
1/βj
hj (v) = max h ∈ H : h ≤ LL0 L−1
, j = 1, . . . , d,
j v
and L ∈ (0, 1) is chosen to satisfy LL0 ≤ c. This in its turn implies LL0 < 1.
~
~ L
~ implies that
This choice of h(v)
together with the definition of the class FK β,
~
J h(v),
v = {1, . . . , d}, ∀v > 0;
~
h(v)
∈ (0, 1]d , ∀v ∈ (0, L−1 ].
(3.78)
(3.79)
~ such that
Moreover, there exists T1 := T1 β~ < ∞ independent of L
−1
lim sup (ln n)
n→∞
d
Then set T2 = e 2 +
Pm
j=1
sup
d
X
v∈Vn j=1
µj (α) √
i
h β(α)
Vn = δn1+β(α) , 1 .
ln hj (v) ≤ T1 ,
1
− 2β(α)
T1 + 2(LL0 )
.
We have in view of (3.79) and (3.80) for all n large enough and any v ∈ Vn
p
p − 1
(T1 + 2) ln n
Fn ~
h(v) ≤ √ Q
≤ T2 δn v 2β(α)
1
+µ (α)
n dj=1 hj (v) 2 j
h(v) ≤
Gn ~
n
(3.80)
(T1 + 2) ln n
(T1 + 2) ln n
− 1
2
1+µj (α) ≤ Qd
1+2µj (α) ≤ T2 δn v β(α) .
n j=1 hj (v)
j=1 hj (v)
Qd
(3.81)
(3.82)
2β(α)
β(α)
√
Setting T3 = ( 2a−1 T2 ) 2β(α)+1 and T4 = (a−1 T22 ) β(α)+1 we obtain in view of (3.81) and (3.82) for
all n large enough
β(α)
i
h
~
h(v)
∈ H(v), ∀v ∈ T4 δn1+β(α) , 1 ,
~h(v) ∈ H(v, 2), ∀v ∈ T3 ϕn , 1 .
β(α)
It is worth noting that T2 > 1, which implies T4 > 1, and δn1+β(α) ϕ−1
n → 0, n → ∞. Choose
β(α)
v = T4 δn1+β(α) , v = T3 ϕn .
28
(3.83)
~
Since h(v)
∈ H(v) for any v ∈ [v, v] in view of (3.83), we deduce from (3.78) and (3.81) for any ~s
~ L
~ ≤ v −2 F 2 ~h(v) ≤ T 2 δn v −2−1/β(α) , ∀v ∈ [v, v].
Λ~s v, FK β,
n
2
This, in its turn, yields for any ~s
Z v
v
~ L
~ dv ≤ T5 δn Z(v, v),
v p−1 Λ~s v, FK β,
(3.84)
where we have denoted T5 = T22 1 ∨ |p − 2 − 1/β(α)|−1 and
Z(v, v) = v p−2−1/β(α) 1{p>2+1/β(α)} + v p−2−1/β(α) 1{p<2+1/β(α)} + ln (v/v)1{p=2+1/β(α)} .
~
∈ H(v, 2) in view of (3.83), we deduce from (3.78) that for any ~s
Moreover, since h(v)
~ L
~ , ∞ = 0.
Λ~s v, FK β,
(3.85)
At last, putting T6 = T4p−1 (1 + ln T4 )t(H) , we obtain
(p−1)β(α)
lH (v) := v p−1 (1 + | ln (v)|)t(H) ≤ T6 δn 1+β(α) (1 + ln n)t(H) .
(3.86)
Applying the third assertion of Theorem 3, we deduce from (3.84), (3.85) and (3.86) that
Rn(p) [fb~h(·) , f ]
≤C
(1)
(p−1)β(α)
1+β(α)
T6 δn
t(H)
(1 + ln n)
1
p
1
+ T5 δn Z(v, v) + Cp n− 2 .
After elementary computations we come to the statement of Assertion 1.
References
Barron, A., Birgé, L. and Massart, P. (1999). Risk bounds for model selection via penalization.
Probab. Theory Related Fields 113, 301–413.
Birgé, L. and Massart, P. (2001). Gaussian model selection. J. Eur. Math. Soc. (JEMS) 3, 3,
203-268.
Cai, T. T. (1999). Adaptive wavelet estimation: a block thresholding and oracle inequality approach. Ann. Statist. 27, 3, 898–924.
Cavalier, L. and Golubev, G.K. (2006). Risk hull method and regularization by projections of
ill-posed inverse problems. Ann. Statist. 34, 1653–1677.
Comte, F. and Lacour, C. (2013). Anisotropic adaptive kernel deconvolution. Ann. Inst. H.
Poincaré Probab. Statist. 49, 2, 569–609.
Dalalyan, A. and Tsybakov, A.B. (2008). Aggregation by exponential weighting, sharp PACBayesian bounds and sparsity. Machine Learning 72, 39–61.
Donoho, D. L., Johnstone, I. M., Kerkyacharian, G. and Picard, D. (1996). Density
estimation by wavelet thresholding. Ann. Statist. 24, 508–539.
Efroimovich, S. Yu. and Pinsker M.S. (1984). An adaptive algorithm of nonparametric filtering. Automat.Remote Control 45, 58–65.
Efroimovich, S. Yu. (1986). Non-parametric estimation of the density with unknown smoothness.
Theory Probab. Appl. 30, 557–568.
29
Goldenshluger, A. (2009). A universal procedure for aggregating estimators. Ann. Statist. 37,
1, 542-568.
Goldenshluger, A. and Lepski, O.V. (2009). Structural adaptation via Lp -norm oracle inequalities. Probab. Theory Related Fields 143, 41-71.
Goldenshluger, A. and Lepski, O.V. (2011). Bandwidth selection in kernel density estimation:
oracle inequalities and adaptive minimax optimality. Ann. Statist. 39, 1608–1632.
Goldenshluger, A. and Lepski, O.V. (2014). On adaptive minimax density estimation on Rd .
Probab. Theory Related Fields, 159, 479–543.
Golubev, G.K. (1992). Non-parametric estimation of smooth probability densities. Probl. Inform.
Transm. 1, 52–62.
Grafakos, L. (2008). Classical Fourier Analysis. Graduate texts in Mathematics, second edition,
Vol.249. Springer scence+Business media.
de Guzman, M. (1975). Differentiation of Integrals in Rn . With appendices by Antonio Crdoba,
and Robert Fefferman, and two by Roberto Moriyn. Lecture Notes in Mathematics, Vol. 481.
Springer-Verlag, Berlin-New York.
Hesse, C. H. (1995). Deconvolving a density from partially contaminated observations. Journal
of Multivariate Analysis 55, 246–260.
Kerkyacharian, G., Lepski, O. and Picard, D. (2001). Nonlinear estimation in anisotropic
multi–index denoising. Probab. Theory Related Fields 121, 137–170.
Lepskii, O. V. (1991). Asymptotically minimax adaptive estimation. I. Upper bounds. Optimally
adaptive estimates. Theory Probab. Appl. 36, 682–697.
Lepski, O.V. (2015). Adaptive estimation over anisotropic functional classes via oracle approach.
Ann. Statist. 43, 3, 1178–1242.
Lepski, O.V., and Willer T. (2017). Lower bounds in the convolution structure density model
Bernoulli, 23, 2, 884-926.
Leung, G. and Barron, A. R. (2006). Information theory and mixing least-squares regressions.
IEEE Trans. Inform. Theory 52, 8, 3396-3410.
Nemirovski, A. S. (2000). Topics in non-parametric statistics. In Lectures on probability theory
and statistics (Saint-Flour, 1998) Lecture Notes in Math. Springer, Berlin 1738, 85-277.
Rebelles, G. (2016) Structural adaptive deconvolution under Lp -losses. Math. Methods Statist. 25,
1, 26-53.
Tsybakov, A. (2003). Optimal rate of aggregation. Proc. COLT. Lecture Notes in Artificial Intelligence 2777, 303–313.
Wegkamp, M.H. (2003). Model selection in nonparametric regression. Ann. Statist. 31, 252-273.
30
| 10math.ST
|
On Behaviors of Maximal Dimension
arXiv:1801.08327v2 [math.GR] 26 Jan 2018
Wenjun Niu∗
Fudan University
Shanghai, China, 200433
January 29, 2018
Abstract
In this paper, we investigate behaviors of Maximal Dimension, a group
invariant involving certain configuration of maximal subgroups, which we denote by MaxDim. We prove that in some special cases, MaxDim(G × H) =
MaxDim(G) + MaxDim(H). We also prove a conjecture stated by Ellie Thieu
which shows that groups with m < MaxDim can be constructed from groups
with m < i.
1
Introduction and background
For a group G, a subset s ⊆ G of elements in G is a generating set if hg, g ∈ si = G.
s is said to be irredundant if hg, g ∈ s \ {h}i =
6 hg, g ∈ si for each h ∈ s. Thus
for a finite group G, any generating set s contains an irredundant generating set,
simply by removing the redundant elements one at a time from s. Note that this is
generally not true for infinite groups, as has been mentioned in [7] by R. Fernando. In
what follows, we consider exclusively only finite groups, which will not be mentioned
again in the statements. With these in mind, we introduce two different notions of
“dimension” for a group, whose associations has been studied for long. Denote by
m(G) the maximal size of an irredundant generating set of G, and i(G) the maximal
size of an irredundant set of G. By definition one clearly has m(G) ≤ i(G). It is not
hard to see that these functions are not always equal, for example, if G = Zp ≀ Zp ,
then m(G) = 2 but i(G) = p.
There are several reasons these functions are called “dimension”. First, one may
observe that these functions generalizes dimension of a vector space, on which these
conceptions coincide. Second, these functions clearly measure how “large” a group
may be. Finally, thanks to the work of several mathematicians, it has been shown
that these functions behave nicely on groups, e.g., m(G × H) = m(G) + m(H) and
i(G × H) = i(G) + i(H)1 .
There is another counterpart of dimension for finite groups, which comes somehow
naturally from generating sequences, and will be our main gradient. In order to define
this function, let us introduce the following definition:
∗ This material is based on work at 2017 Cornell University Math REU Program. The author want
to express his acknowledgement to Professor R. Keith Dennis and graduate mentor Ravi Fernando
for their guidance and encouragement.
1 The conclusion for i is more straightforward from Whiston’s Lemma. For m, see [9] or [4].
1
Definition 1.1. Let G be a group, and {Hi | 1 ≤ i ≤ n} a set of subgroups of
G.
| 1 ≤ i ≤ n} is said to be in general position if for any 1 ≤ j ≤ n,
T {Hi T
Hi ) Hi .
i6=j
i
In other words, a set of subgroups of G is in general position if the intersection of
subgroups from any proper subset strictly contains the intersection of the subgroups
from the whole set. Now we define:
Definition 1.2. Let G be a group, the maximal dimension of G, denoted by
MaxDim(G), is defined by:
MaxDim(G) = max |S|
S
(1)
where S ranges over collections of maximal subgroups of G that are in general position.
One may recognize the counterpart of this definition in linear algebra. This
function is related to generation of finite groups by the following proposition:
Proposition 1.3. Let G be a group, then m(G) ≤ MaxDim(G) ≤ i(G).
Proof. First, we prove that m(G) ≤ MaxDim(G). Let s = {g1 , . . . , gn } be an irredundant generating set of G with n = m(G). For any 1 ≤ i ≤ n, choose Mi maximal such
that s \ {gi } ⊆ Mi , which is possible because s \ {gi } generates a proper subgroup.
One observe that gi ∈
/ Mi otherwise Mi would contain a generating set. We claim that
{Mi , 1 ≤ i ≤ n} are in general position. For any j, we indeed have ∩i6=j Mi ) ∩i Mi
because the first subgroup contains gj while the second doesn’t. Hence the claim
holds and m(G) = n ≤ MaxDim(G).
Next, we prove that MaxDim(G) ≤ i(G). Let S = {M1 , . . . , Mk } be maximal
subgroups of G with k = MaxDim(G). By definition, for any i, ∩j6=i Mj \ Mi is
not empty, hence one can choose gi in this set. We claim that s = {g1 , . . . , gm } is
irredundant. Indeed, the subgroup generated by s\{gi } is contained in Mi which does
not contain gi , hence itself cannot contain gi , so the claim is true. Thus MaxDim(G) =
k ≤ i(G).
Remark 1.4. We say s in the above proof certifies S.
In light of Proposition 1.3, R. Keith Dennis suggested a way of seeking generating
sets by looking for maximal subgroups in general position attaining maximal length,
and then use the above machinery. But unfortunately, the behavior of this function
has not been studied very profoundly and adapting this machinery can be risky. R.
Fernando([7]), E. Detomi and A. Lucchini([5]) exhibit groups with m ≪ MaxDim,
and proved that m = MaxDim for a large family of groups. Computation by R. Keith
Dennis reveals that this function may have nice behavior on product of groups. We
will present a partial result in this direction. But before going to next section, let us
introduce the following definition:
Definition 1.5. Let G be a group. The Frattini Subgroup of G, denoted by Φ(G), is
defined as the intersection of all maximal subgroups of G.
The importance of Φ(G) for us is that it is negligible in computing m and
MaxDim. We say G is Frattini Free if Φ(G) is trivial. It is not restrictive to always
assume that G is Frattini Free when computing these functions because G/Φ(G) is
always Frattini Free. We will do so when necessary.
2
2
On additivity of MaxDim
In this section, we will prove that MaxDim(H × K) = MaxDim(H) + MaxDim(K)
for appropriate H and K, and will present the obstacle to generality. First, we will
introduce the following:
Lemma 2.1 (Goursat). Let H and K be groups. Then maximal subgroups of H × K
come from one of the following two types:
(1). M × K for M maximal in H or H × M ′ for M ′ maximal in K.
(2). There exists a tuple (S, N, N ′ , α), where S is a simple group, N and N ′
are normal subgroups of H and K respectively, and α : H/N → K/N ′ ∼
= S is an
isomorphism, such that ∆ ⊆ H × K is defined by ∆ = {(h, k) | α(h̄) = k̄}.
Remark 2.2. Subgroups in (1) will be called standard subgroups, and subgroups in
(2) will be called pullback subgroups2 .
Lemma 2.1 partially explains the reason why MaxDim is difficult to understand:
one has to deal with pullback subgroups, which involves common simple quotients of H
and K. One easy observation will be that if pullback subgroups do not exist, in other
words, if H and K are coprime, then MaxDim(H ×K) = MaxDim(H)+MaxDim(K).
This in fact can be generalized slightly to the following:
Theorem 2.3. Let H and K be groups admitting no common nonabelian simple
quotient. The following identity holds:
MaxDim(H × K) = MaxDim(H) + MaxDim(K)
with which a corollary:
Corollary 2.4. Let H and K be groups such that one of them is solvable, then
MaxDim(H × K) = MaxDim(H) + MaxDim(K).
Thus with the help of Theorem 2.3, the additivity of MaxDim on product of finite
solvable groups has been settled. We proceed to the proof of Theorem 2.3, beginning
with a lemma:
Lemma 2.5. Let H be a group and N be a normal subgroup of H. Let M1 , . . . , Mk
be maximal subgroups of H in general position, then it is possible to rearrange them,
so that M1 ∩ N, . . . , Ml ∩ N are in general position and for any j > l, Mj ∩ N contains
∩1≤i≤l Mi ∩N . Moreover, if we let π : H → H/N be the projection and R = ∩1≤i≤l Mi ,
then π(R ∩ Mj ) for j > l are in general position in H/N .
Proof. Without loss of generality we assume that Mi ∩ N are in general position for
1 ≤ i ≤ l and Mj ∩ N contains the intersection of them for any j > l. So it remains
to show that π(R ∩ Mj ) for j > l are in general position. For this purpose, one need
only show that for any l < j ≤ k, π(R ∩ Mj ) does not contain
\
π(R ∩ Mt )
l<t≤k
t6=j
Without loss of generality assume j = k. Since clearly
\
\
π(R ∩ Mt ) ⊇ π(
Mt ∩ R)
l<t≤k−1
2 This
l<t≤k−1
terminology comes from the fact that such groups are pullbacks in the category of groups.
3
we need only show that π(R ∩ Mk ) does not contain the latter group, which is equivalent to that
\
(R ∩ Mk )N +
Mt ∩ R
l<t≤k−1
Denote by T the group on the right. If (R ∩ Mk )N ⊇ T then (R ∩ Mk )(R ∩ N ) ⊇ T
and by assumption R ∩ N ⊆ Mk ∩ N which implies that Mk ⊇ T , violating the fact
that M1 , . . . , Mk are in general position. Hence π(R ∩ Mj ) for j > l are in general
position.
This lemma is very similar to Whiston’s argument3 , but here we are working on
a set of subgroups, hence readers may notice that the direction changes. Now we can
prove Theorem 2.3:
Proof of Theorem 2.3. The proof is by the following strategy: given any set of maximal subgroups in general position, we replace pullback subgroups by standard subgroups in a way that does not change the total number of subgroups in question
, so that the resulting set of maximal subgroups is still in general position. Let
S = {M1 , . . . , Mk } be a collection of maximal subgroups of H × K in general position, if no maximal subgroups are pullbacks then k ≤ MaxDim(H) + MaxDim(K)
and we are done. So suppose among S there is a pullback subgroup, say Mk of the
form ∆(N,N ′ ,α) induced by N ✁ H and N ′ ✁ K and α : H/N ∼
= Zp (By our
= K/N ′ ∼
assumption, pullback maximal subgroups must be of this form). Abbreviate it by ∆,
we know that necessarily ∆ ∩ (N × N ′ ) = N × N ′ and must contain the intersection
of Mj ∩ (N × N ′ ) for j < k. This, together with Lemma 2.5, and the fact that Z2p
is flat with i = 2, indicates that we may assume without loss of generality, either
Mj ∩ (N × N ′ ) for j < k are in general position, or Mj ∩ (N × N ′ ) for j < k − 1 are
in general position and Mk−1 contains the intersection of them. Let us now consider
separately these two cases.
Case 1. Suppose Mj ∩ (N × N ′ ) for j < k are in general position. Let
\
R=
Mj
1≤j<k
and π : H × K → H × K/(N × N ′ ) be the projection, then π(R) must be a nontrivial
subgroup of H × K/(N × N ′ ) ∼
= Z2p , otherwise R ⊆ N × N ′ ⊆ ∆, violating the fact
that S is in general position. Because π(R) is nontrivial, but H/N ∩ K/N ′ is trivial
in the quotient, π(R) is not contained in at least one of them, say H/N . We claim
in this case that S ′ = {M1 , . . . , Mk−1 , H × N ′ } is in general position. Indeed, since
Mi ∩ (H × N ′ ) for 1 ≤ i < k are in general position, and H × N ′ does not contain R
which is the intersection of Mi for 1 ≤ i < k, we see immediately that S ′ is in general
position.
Case 2. Suppose Mj ∩ (N × N ′ ) for j < k − 1 are in general position and Mk−1
contains the intersection of them. Let
\
R=
Mj
1≤j<k−1
and π as in the first case, we claim that π(R) = H/N × K/N ′ . Indeed, by Lemma 2.5,
π(R ∩ Mk−1 ) and π(R ∩ ∆) are in general position, which implies that i(π(R)) = 2.
3 Interested
readers may refer to [10] or [2] for more information
4
But the only subgroup of Z2p having i = 2 is the whole group, hence the claim holds.
Now it is not hard to show, by the same argument as in the first case, that the set
S ′ = {M1 , . . . , Mk−2 , H × N ′ , N × K} is in general position.
This shows that in any case we can replace pullback subgroups by standard
subgroups, and a repeated application of this gives k ≤ MaxDim(H) + MaxDim(K).
This is true for any collection of maximal subgroups of the direct product, and we
conclude that MaxDim(H × K) ≤ MaxDim(H) + MaxDim(K). On the other hand, it
is trivial that MaxDim(H × K) ≥ MaxDim(H) + MaxDim(K), hence MaxDim(H ×
K) = MaxDim(H) + MaxDim(K). This completes the proof.
Lemma 2.5 also yield the following:
Proposition 2.6. Suppose H and K are groups and MaxDim(K) = i(K), then:
MaxDim(H × K) = MaxDim(H) + MaxDim(K)
Proof. Let G = H × K and let M1 , . . . , Mk be maximal subgroups that are in general
position. Now after rearranging we may assume that M1 ∩ H, . . . , Ml ∩ H are in
general position and Mj contains their intersection for j > l. This together with
Lemma 1.3 gives the bound k − l ≤ i(G/H) = i(K) = MaxDim(K). Now for any
maximal subgroup Mi , i ≤ l, since {Mi ∩ H|1 ≤ i ≤ l} is in general position, Mi
is either a standard subgroup induced by some maximal M < H, or a pullback.
In both cases, H ∩ Mi will be an intersection of maximal subgroups. Indeed, if
Mi = ∆(N,N ′ ,α) , then Mi ∩H = N , which is the intersection of all maximal subgroups
containing N because H/N is simple therefore Frattini Free; but if Mi = M × K then
Mi ∩ H = M . We show that l ≤ MaxDim(H). Let Hi = Mi ∩ H for 1 ≤ i ≤ l. Now
we claim that we can choose a maximal subgroup M ⊇ H1 such that M, H2 , . . . , Hl
are in general position. This is enough to show that l ≤ MaxDim(H). Suppose for
any such M , the sequence fail to be in general position, then the only possibility
is that M ⊇ ∩1<i≤l Hi , since if H1 ∩ Hi are in general position for 1 < i ≤ l,
so are M ∩ Hi for any M ⊇ H1 . But this implies that H1 ⊇ ∩1<i≤l Hi since H1
equals the intersection of maximal subgroups containing it. This contradicts the fact
that Hi are in general position, so the claim holds. These arguments together give
k ≤ MaxDim(H) + MaxDim(K). This is true for any set of maximal subgroups in
general position, hence MaxDim(G) ≤ MaxDim(H) + MaxDim(K). Since the other
inequality is trivial, the proof is complete.
Theorem 2.3 and Proposition 2.6 cover a wide range of groups, e.g., all solvable
groups, Mathieu groups, Janko groups J1 , J2 and all symmetric and alternating groups
satisfy assumptions of Theorem 2.3 or Proposition 2.64 , hence MaxDim(H × K) =
MaxDim(H) + MaxDim(K) if K is one of the above groups. These all provide strong
evidence that in general, MaxDim is additive on product of groups.
It is not known yet in general, but in proving it or disproving it, one might need
to consider S × S for S a simple group. There are some reasons for this: first of all,
4 The
results for Mathieu groups M11 and M12 is proved by T. Brooks([1]). The results for
Mathieu group M22 and Janko groups are joint work of Sophie Le, Tianyue Liu and me(to appear).
The result for M23 is from the work of Sophie Le(to appear), and the result for M24 is from a
remarkable theoretical proof of Tianyue Liu and R. Keith Dennis([8]). For alternating and symmetric
groups, see [10] by J. Whiston.
5
MaxDim is additive automatically on S ×S ′ if S and S ′ are simple but not isomorphic;
second, for any group G, one has the conception of rumpf, denoted by R(G), which
is defined as the intersection of maximal normal subgroups, and it is easily seen that
pullback subgroups are determined by G/R(G), which is a product of simple groups.
To this direction, the biggest family of non-flat simple groups is that of linear groups,
and we deal with a part of this family in the following:
Proposition 2.7. Let PSL(2, p) be the projective special linear group over the field
with p ≥ 5 elements, where p is a prime. Then the following identity holds:
MaxDim(PSL(2, p) × PSL(2, p)) = 2MaxDim(PSL(2, p))
Before coming to the proof, it is worthwhile to introduce some important results
about PSL(2, p) that will be used in the proof. The first one is Dickson’s complete
classification of maximal subgroups of PSL(2, p).
Theorem 2.8 (Dickson). Maximal subgroups of PSL(2, p) comes from one of the
following types:
(1). Point stabilizers Ga for a ∈ P1 (Fp ), isomorphic to Zp ⋊ Z p−1 .
2
(2). Stabilizers of a pair G{a,b} for a, b ∈ P1 (Fp ), isomorphic to Dp−1 , dihedral
group of order p − 1.
(3). Stabilizers of a pair G{a,b} for a, b ∈ P1 (Fp2 ), isomorphic to Dp+1 , dihedral
group of order p + 1.
(4). Subgroups isomorphic to A4 , S4 or A5 .
Readers may refer to [6] for more details. In [11], J. Whiston and J. Saxl proved
that 3 ≤ m(PSL(2, p)) ≤ 4. This is in fact also true for MaxDim. In fact, in [3], D.
Collins computed explicitly the intersection of maximal subgroups of the first three
types. The result can be summarized as the following:
Proposition 2.9 (D. Collins). Intersections of maximal subgroups of PSL(2, p) of
first three types are:
(1). Ga ∩ Gb ∼
= Z(p−1)/2 .
(2). Ga ∩ Gb ∩ Gc is trivial.
(3). Ga ∩ G{b,c} is either trivial or isomorphic to Z2 .
(4). G{a,b} ∩ G{c,d} is isomorphic to Z2 × Z2 , the Klein 4-group.
In the following proof, we view PSL(2, p) both as matrices and as mobius functions on the space Fp ∪ {∞} interchangebly. In any case, Ga will be point stabiliser,
and G{a,b} will be the elements that stabilize the set {a, b}.
Proof of Proposition 2.7. Denote by G the group PSL(2, p). Let us first note that
MaxDim(G) ≤ 4. Indeed, if MaxDim(G) > 4, then because i(A5 ) = i(S4 ) = 3, the
set of subgroups attaining maximal length would all come from first three types in
Theorem 2.8. But by Proposition 2.9, the intersection of any three maximal subgroups
from first three types is a subgroup of Z2 if they are in general position. This is a
contradiction. Hence MaxDim(G) ≤ 4.
We fix some notations LM = M × G and RM ′ = G × M ′ , and ∆σ = {(g, σ(g) |
g ∈ G)}. These are the only maximal subgroups of G × G. Since the conclusion
is trivial if no pullback subgroup is included, in what follows we assume that there
is a pullback subgroup. Let LM1 , · · · , LMk and RM1′ , · · · , RMs′ be standard maximal
subgroups induced by Mi , Mj′ < G, and ∆id , ∆σ1 , · · · , ∆σn be pullback subgroups
6
induced by id, σi , such that these subgroups are in general position(note that we can
always assume that ∆id is included by applying appropriate automorphism), we wish
to show that k + s + n + 1 ≤ 2MaxDim(G). First note that {∆id ∩ LMi , ∆id ∩ RMj′ } is
in general position, then {Mi , Mj′ } is in general position in PSL(2, p). On the other
hand, by Proposition 2.9, it is easily seen that if four maximal subgroups are in general
position, then their intersection must be trivial. This indicates that if k + s = 4, then
n = 0 and the cardinality of the whole set is less than 2MaxDim(G).
So we may assume that k + s does not attain 4. But before working directly with
all maximal subgroups, it seems convenient to first deal with ∆σi ’s. In the following,
we prove that in order that ∩i ∆σi ∩ ∆id be nontrivial, n must be less than or equal
to 4.
We begin with pullback subgroups that are induced by inner homomorphism.
Now let ∆id , ∆h1 , ∆h2 , · · · , ∆hn be maximal subgroups that are in general position,
where ∆hi is induced by the conjugation of hi , such that their intersection is not
trivial. We show that n ≤ 3. It is easily seen that their intersection is {(g, g) |
g commutes with h1 , · · · , hk }. Also note that {h1 , · · · , hk } is irredundant. Let C(S)
be the subgroup consists of elements commuting with S. Let H be the subgroup
generated by all hi . If H = G then since G is simple, C(G) is trivial and hence
the intersection is trivial. This shows that in order that the subgroups above have
nontrivial intersection, H must be proper. We assume first that H is contained in
some maximal subgroup of the first three kind. But before going on, we will list here
some observations that will be used throughout the proof:
(a). Suppose h = diag{x, x−1 }, where x2 6= 1, then C(h) consists of only diagonal
matrices iff x4 6= 1. If x4 = 1, then C(h) consists of all diagonal matrices as well as
matrices of the form
0
y
−y −1 0
which we will denote by Ad{y, −y −1}.
(b). Suppose h = Ad{y, −y −1 } then C(h) consists of matrices of the form
x −zy 2
z
x
as well as
x
z
zy 2
−x
where each matrix must have determinant one.
If H < Ga for some a, then consider C(h1 ), it is a subgroup commuting with h1 .
Since h1 fixes a, it also fixes ga for g ∈ C(h1 ), hence in order that h1 be nontrivial,
all elements in C(h1 ) can only possibly fix a or move a to another b(This is true by
(2) of Proposition 2.9). If for all hi , all elements in C(hi ) fixes a, we show that n ≤ 2.
2
Under appropriate conjugation, we may assume that a = ∞ and Ga ∼
= Zp ⋊ (Z×
p)
where the second factor acts by multiplying on the left(viewed as subgroup of the
unit elements in Zp ). Let h1 = (x1 , y12 ) and h2 = (x2 , y22 ). Let (x, y 2 ) be an element
that commute with both h1 and h2 , then we have the formula:
(
x1 + y12 x = x + y 2 x1
x2 + y22 x = x + y 2 x2
7
and more precisely the following matrix equation:
x
1 − y12 x1
x1
=
y2
1 − y22 x2
x2
If the matrix on the left has determinant zero, then the two rows are linearly
dependent, in which case C(h1 ) = C(h2 ), which cannot happen since the subgroups
are in general position. Hence the only possibility is that the matrix is invertible,
which means that there is only one possible solution for the above equation, i.e.,
(0, 1). Hence C(h1 ) ∩ C(h2 ) = 1 and n ≤ 2.
If C(h1 ) consists of elements fixing a and switching {a, b}, then h1 ∈ Ga ∩ Gb ,
and under appropriate conjugation, we may assume that h1 is diagonal. If h2 is
not in this intersection, then h2 maps b to some c, in which case any element g in
C(h1 ) ∩ C(h2 ) will need to fix {a, b} and satisfy gc = gh2 b = h2 gb which equals either
c or a. But since g fixes {a, b}, g must also fix c, in which case C(h1 ) ∩ C(h2 ) ⊆ Z2
and nothing else can be added otherwise the intersection will be trivial. Now consider
the other case, where all hi lies in Ga ∩ Gb , and hence all hi are diagonalisable. In this
case, by the observation (1), we conclude that C(h1 ) = C(h2 ) or C(h1 ) ⊇ C(h2 ) or
C(h1 ) ⊆ C(h2 ). This contradicts the fact that the subgroups are in general position.
Up to now the first case is all done and the conclusion is that whenever H < Ga ,
n ≤ 2, or to say, for any three of the C(hi ), if they are in general position, their
intersection must be trivial.
Now consider the case where H < G{a,b} . Here we may well work at PSL(2, Fp2)
which unifies type (2) and (3) in Theorem 2.8, because the above observations are still
true here. Without loss of generality, we may assume that H consists of matrix of the
form diag{x, x−1 } and Ad{y, −y −1}. By same argument as in the above case, we know
that there will at most be one hi of diagonal form. Suppose h1 is of diagonal form
−1
diag{x1 , x−1
1 } and h2 is of the form Ad{y2 , −y2 }. Direct computation by observation
(2) shows that C(h1 ) ∩ C(h2 ) either contains only diag{x, x−1 } with x4 = 1(if C(h1 )
contains only diagonal matrices) or contains these together with those Ad{y, −y −1}
such that y 4 = y24 . In the first case, C(h1 ) ∩ C(h2 ) = Z2 and nothing more can be
added to the sequence. In the second case, if we choose h3 = Ad{y3 , −y3−1 } with
y34 = y24 , then C(h2 ) ∩ C(h1 ) = C(h3 ) ∩ C(h1 ) which contradicts the fact that they
are in general position. Hence we must choose h3 such that y34 6= y24 , which means
that C(h1 ) ∩ C(h2 ) ∩ C(h3 ) contains only elements of the form diag{x, x−1 } with
x4 = 1, and so C(h1 ) ∩ C(h2 ) ∩ C(h3 ) = Z2 . And we cannot add any more hi in
otherwise the intersection will be trivial. Finally, if we only choose element of antidiagonal form, and assume that we have chosen hi = Ad{yi , −yi−1 } for 1 ≤ i ≤ 3, and
yi4 are not all equal, then it is possible using same argument as above to show that
C(h1 ) ∩ C(h2 ) ∩ C(h3 ) ⊆ Z2 . This finishes the case where H < G{a,b} .
If H < S4 or A5 , and if n = 3, then the hi ’s form an irredundant generating
sequence for respective groups. Because in either case the group is strongly flat and
has trivial center, the intersection will have to be trivial. Hence in order that the
intersection being nontrivial, n ≤ 2. The same is true if H < A4 .
The conclusion here is that if ∆id , ∆h1 , ∆h2 , ∆h3 are in general position, there
intersection is a subset of Z2 . We now apply an argument similar to Whiston’s
to pass the information to Aut(G)5 . Suppose we have ∆σ1 , · · · , ∆σ4 together with
∆id are in general position, and one of the σi , say σ1 is not an inner-morphism.
5 Here
we use the fact that Out(G) = Z2 .
8
Then we can multiply σ1 appropriately to σj ’s to make them inner, say σ1si σj is
inner, and we have Rad(∆σ1 , . . . , ∆σ4 , ∆id )6 = Rad(∆σ1 , ∆σ1s2 σ2 , · · · , ∆σ1s4 σ4 , ∆id ).
Consider ∆id ∩ ∆σsj σj for j > 1. We claim that they are in general position. If not,
1
say ∆id ∩ ∆σ1s1 σ2 ∩ ∆σ1s3 σ3 < ∆id ∩ ∆σ1s4 σ4 , then Rad(∆σ1 , ∆σ1s2 σ2 , · · · , ∆σ1s4 σ4 , ∆id ) =
Rad(∆σ1 , ∆σ1s2 σ2 , ∆σ1s3 σ3 , ∆id ), which implies ∩i ∆σi ∩ ∆id = ∩i<4 ∆σi ∩ ∆id . This
contradicts the fact that the five subgroups are in general position. But if ∆id ∩
∆σsj σj for j > 1 are in general position, then their intersection lies inside Z2 as has
1
been shown above. Hence if ∆σi for 1 ≤ i ≤ 4 and ∆id are in general position, their
intersection is a subset of Z2 .
Now we can finish the proof of the proposition. Suppose we are given some
′
collections of maximal subgroups LM1 , · · · , LMk and RM1′ , · · · , RM
with k + s ≤ 3,
s
∆id , ∆σ1 , ∆σ2 , ∆σ3 , ∆σ4 where σi ’s are nontrivial automorphisms. We are done
if MaxDim(G) = 4 since the argument above indicates that any time if pullback
subgroups are included, the total number of maximal subgroups has to be no more
than 8. If MaxDim(G) = 3, then in order that the equality fails, all but one of
the maximal subgroups of the above need to be included. We show that this is
impossible. Suppose we use all the ∆σi , then the above argument shows that the
intersection of these subgroups with ∆id is either Z2 or trivial, indicating that we
can only possibly add one more in the set. So no more than three ∆σi ’s other than
∆id can be included, and in order to attain 7, we need to use all standard maximal
subgroups above, that is LMi and RMj′ . If the Mi , Mj′ are all from first three types,
then by easy computation, their intersection will be either trivial or isomorphic to
Z2 , in which case only 5 maximal subgroups can be included. Hence we must have
some A5 or S4 in Mi , Mj′ . But in this case, since i(A5 ) = i(S4 ) = 3 we conclude that
we can possibly have 3 more choices of other types other than ∆id . In either case,
k + s + n + 1 ≤ 6 ≤ 2MaxDim(G). This is true for any set of maximal subgroups in
which there is one pullback subgroup, and because it is trivially true if no pullback
subgroup is used, we conclude that MaxDim(G × G) ≤ 2MaxDim(G). Since the other
inequality is trivial, the proof is complete.
We conclude this section by some comments. First of all, it seems to the author
that in proving or disproving MaxDim(S 2 ) = 2MaxDim(S) for simple S one will
inevitably use classification theorem. Second, it is not hard to show that for a simple
S, if MaxDim(S 2 ) = 2MaxDim(S), then MaxDim(S n ) = nMaxDim(S). Third, even
if the identity holds for simple groups, there is no obvious evidence that it should
be true in general. So before going into the proof for all simple groups, I think it
is worthwhile to build the bridge from simple groups to general groups. R. Keith
Dennis, on the other hand, suggested that one should first focus on simple groups
and look for stronger properties than that in Proposition 2.7, among which there
might be illustrations toward general situations. This is possibly doable because in
Proposition 2.7, it seems very likely that once a pullback subgroup is included, the
number of subgroups in general position is no greater than MaxDim(S) + 1. It will
be great if such conclusion does hold, but by now we only know this holds for a small
number of groups, i.e., flat simple groups. This is the following proposition proved by
R. Fernando in Cornell 2017 summer REU program:
Proposition 2.10 (R. Fernando). Let S be a simple group. If M1 , . . . , Mk are max6 Rad
means the radical, i.e., taking intersection.
9
imal subgroups of S × S that are in general position, then we have:
if Mi are all standard
2MaxDim(S)
k≤
i(S) + 1
if some Mi is a pullback
Proof. If all Mi are of standard form, it is trivial that one has k ≤ 2MaxDim(S). If
some Mi , say M1 is a pullback, then M1 ∼
= S and because Mi ∩ M1 are in general
position for i > 1, we have k − 1 ≤ i(S), hence k ≤ i(S) + 1. This completes the
proof.
3
Relative versions of MaxDim
This section serves as a transition between Section 2 and Section 4, some definitions
and results will be introduced and will be used in next section. In order to further
understand Maximal Dimension, inspired by [4], we come to the definition of relative
versions of this function. The first one is the following:
Definition 3.1. Let G be a group and H < G, then the maximal dimension of G
relative to H, which we denote by MaxDim(G, H), is defined by:
MaxDim(G, H) = max |S|
S
where the maximum is taken over all S = {M1 , . . . , Mk }, collections of maximal
subgroups of G such that {M1 ∩ H, . . . , Mk ∩ H} is in general position.
The naturality of this definition can be seen in Lemma 2.5. There is another
version, which comes naturally when studying group extensions:
Definition 3.2. Let G be a group and H another group acting on G. Define the
maximal dimension of G under H, which is denoted by MaxDimH (G), by:
MaxDimH (G) = max |S|
S
where the maximum is taken over all S = {M1 , . . . , Mk }, collections of maximal Hinvariant subgroups of G that are in general position.
Our goal in this and next section is to use these two conceptions to study behaviors of maximal dimension under group extensions, especially extensions by abelian
groups. To this vein, we prove the following:
Proposition 3.3. Let N ✁ H and π : H → H/N . Let M1 , . . . , Mk be maximal
subgroups of H that are in general position, such that M1 ∩N, . . . , Ml ∩N are in general
position and Mj contains ∩1≤i≤l Mi ∩N for any j > l. Let R = ∩1≤i≤l Mi and π : H →
H/N be the projection. Then l ≤ MaxDim(H) − MaxDim(H/N, π(R)). Moreover, if
N/Φ(N ) is abelian, l ≤ MaxDim(H) − MaxDim(H/N ), and l ≤ MaxDimH (N ).
Proof. For the first statement, suppose M1′ , . . . , Ms′ are maximal subgroups of H
containing N such that π(M1′ ) ∩ π(R), . . . , π(Ms′ ) ∩ π(R) are in general position, it
10
is not hard to show that {M1 , . . . , Ml , M1′ , . . . , Ms′ } are in general position, hence
l ≤ MaxDim(H) − s. Taking s to be maximal, we get
l ≤ MaxDim(H) − MaxDim(H/N, π(R))
For the second statement, since Φ(N ) < Φ(H), we assume that H is Frattini
free, which implies by our assumption that N is abelian. Note that if M is a maximal
subgroup of H not containing N , we claim that M ∩ N is a maximal H-invariant
subgroup of N . It is indeed H-invariant because M N = H, so we only need to show
that M ∩ N is maximal. If not, say M ∩ N ( N ′ ( N for some N ′ ✁ H. Since M is
maximal, M N ′ = H and H/N = M/(M ∩ N ) = M N ′ /N ′ = H/N ′ , a contradiction.
Hence M ∩ N is maximal H-invariant. This gives the bound l ≤ MaxDimH (N ).
Now we proceed to prove that π(R) = H/N , which will complete the proof of the
Proposition. Suppose R′ is a subgroup of H such that π(R′ ) = H/N , and M a
maximal subgroup of H not containing N such that M ∩ N + R′ ∩ N . We claim that
π(M ∩ R′ ) = H/N . First of all, since M ∩ N is maximal invariant, (M ∩ N )(R′ ∩ N ) =
N . Now for any h ∈ H/N , there exists m ∈ M and r ∈ R′ such that π(m) = π(r) = h.
This indicates that mr−1 ∈ N = (M ∩ N )(R′ ∩ N ), hence there exists n1 ∈ M ∩ N
′
and n2 ∈ R′ ∩ N such that mr−1 = n−1
1 n2 , which indicates that n1 m = n2 r ∈ R ∩ M .
This element has image h. This shows that every element in H/N has a pre-image in
R′ ∩ M , hence π(R′ ∩ M ) = H/N , and the proof is complete.
Proposition 3.3 gives several bounds on l, a value coming from Lemma 2.5 depending on the collection of maximal subgroups. There is a corollary to this which is
also related to the previous section:
Corollary 3.4. Suppose H and K are groups with R(H) and R(K) abelian, and both
H/R(H) and K/R(K) are product of flat simple groups. Then MaxDim(H × K) =
MaxDim(H) + MaxDim(K).
Proof. Let M1 , . . . , Mn be maximal subgroups of H × K that are in general position,
and consider N = R(H) × R(K). Suppose M1 ∩ N, . . . , Mk ∩ N are in general
position, and Mj contains the intersection of them. These Mi must be type (1)
subgroups either not containing R(H) × K(if they are induced by M < H), or not
containing H × R(K)(if they are induced by M ′ < K). Hence by Proposition 3.5, k ≤
MaxDim(H)+MaxDim(K)−MaxDim(H/R(H))−MaxDim(K/R(K)). On the other
hand, by Lemma 1.3, n − k ≤ i(H/R(H) × K/R(K)) = i(H/R(H)) + i(K/R(K)),
which is equal to MaxDim(H/R(H)) + MaxDim(K/R(K)) by flatness. Combining
these inequalities we get the desired result.
Proposition 3.3 also yield the following corollary, which will be used in next
section.
Corollary 3.5. Suppose H = N ⋊ T where N is abelian and T is flat. Then
MaxDim(H) = MaxDimT (N ) + MaxDim(T )
Proof. Note that Proposition 3.5 gives the bound MaxDim(H) ≤ MaxDimT (N ) +
MaxDim(T ). But since the sequence splits, it is very easy to demonstrate a sequence
of that length, which consist of N1 ⋊ T, . . . , Nk ⋊ T and N ⋊ M1 , . . . , N ⋊ Mt where
k = MaxDimT (N ) and t = MaxDim(T ).
11
4
Groups with MaxDim = m and groups with
MaxDim > m
In this section, we will give another proof, using the results and techniques in Section
3, the theorem that has been proved by E. Detomi and A. Lucchini in [5], stating
that if G′ is nilpotent, then m(G) = MaxDim(G). Also, we will prove a proposition
conjectured by Ellie Thieu in 2017 Cornell summer SPUR forum, which not only
exhibits a connection between groups with m < MaxDim and groups with m < i,
but also explains the reason that the group R = smallgroup(720, 774) found by R.
Fernando in [7] is the smallest solvable group with m < MaxDim. In fact, the proof
of this proposition is based on computations on GAP for R.
To begin with, we introduce the following version of Goursat lemma:
Lemma 4.1 (Goursat Lemma for G-groups). Let G be a group acting on H and K.
Let π1 : H × K → H and π2 : H × K → K. Suppose ∆ is a proper G-invariant
subgroup of H × K such that π1 (∆) = H and π2 (∆) = K, then there exists invariant
normal subgroups N ✁ H and N ′ ✁ K, and a G-isomorphism α : H/N → K/N ′ such
that ∆ = {(h, k)|α(h̄) = k̄}.
Proof. Let N = ∆∩H ×1. This is a G-invariant subgroup, and by definition a normal
subgroup of ∆, being the kernel of the projection π2 . Since ∆ projects on the first
factor, N ✁ H. The same reason indicates that N ′ = ∆ ∩ 1 × K is normal. The rest
of the statements will be routine check and will be omitted.
Another lemma important to us is a structure theorem for groups with nilpotent
commutator subgroups:
Lemma 4.2. Let G be a finite group with G′ nilpotent, and suppose Φ(G) = 1. Then
there are abelian subgroups F, H < G such that F ✁ G and G = F ⋊ H. Moreover,
F can be decomposed into minimal H-groups.
Proof. See [5].
Theorem 4.3. Let G be a group with G′ nilpotent, then MaxDim(G) = m(G).
Proof. Write G = F ⋊ H as in Lemma 4.2 and assume that G is Frattini Free, which
implies that F can be decomposed into minimal H vector spaces. Since H is abelian,
Corollary 3.5 is applied and we get that MaxDim(G) = MaxDimH (F ) + m(H). Since
m(G) = m(H)+number of irreducible H components of F 7 , we are only left to prove
that MaxDimH (F ) is exactly the number of complemented chief factors contained in
F (i.e., the number of irreducible components of F ). For this purpose, we may assume
without loss of generality, that F = V n for some minimal H module V and some
n > 0.
Let U be a maximal H-invariant subgroup of F , then by Lemma 4.1, U either
discard one of the factors V totally(i.e., U = V n−1 ), or is determined by diagonally
linking two copies of them(by some automorphism of V ). We show by induction that
the number of maximal invariant subgroups used should always be less than or equal
to the number of irreducible component. Now take any maximal subgroup U that is
used, we deviate two cases. First, suppose U = V n−1 , i.e., U is of normal type. For
7 See,
for example, [9], or [4, p. 52] Proposition 3.4.1
12
any other W , since W 6= U , W either discard another copy of V , or is determined by
diagonally linking two copies of V . In the first case, W ∩ U is clearly maximal in U .
In the second case, if W links two copies of V that are both inside of U , then W ∩ U is
a maximal invariant subspace of U determined by diagonally linking two copies of V .
Otherwise W links two copies of V one of which is the complement of U . In this case,
W ∩ U is the direct sum of V n−1 and the subspace of V consisting of fixed points of
a nontrivial H-automorphism. Since V is minimal this latter space is trivial. Hence
W ∩ U is maximal again. In either case W ∩ U is maximal and inductive hypothesis
can be applied to give the bound. Second, suppose U is a diagonal subspace. In
this case we also have that U ∼
= V n−1 . For any other W , if W is a normal form
or is determined by diagonally linking a pair different from that of U , then W ∩ U
is maximal. If W is determined by diagonally linking the same pair as U , but in a
different way, then W ∩ U is again the direct sum of V n−2 with a diagonal subgroup
determined by the fixed point space of a nontrivial H-automorphism of V , which is
trivial by the same reason as above. Hence W ∩ U is maximal, and the inductive
hypothesis can be applied. This proves that MaxDimH (F ) is indeed the number of
irreducible components of F , and the proof is complete.
In [4], D. Collins used relative versions of m to deal with m(G) − m(G/N ), by
which he proved that m(G × H) = m(G) + m(H). In [9], A. Lucchini investigated
more closely the quantity m(G) − m(G/N ) for minimal normal N , and showed that
this quantity only depends on the action of G on N . The above results come from the
same idea, i.e., to look at MaxDim(G) − MaxDim(G/N ). But A. Lucchini proved(in
private communication) that this value can be arbitrarily large even for minimal
normal abelian N , so we are still not close to fully understanding this value. Next
result is based on computations for R, which also gives a way to construct infinite
class of groups with m < MaxDim.
Proposition 4.4. Let S be a non flat group. There exists a semi-simple S-group V
such that MaxDim(V ⋊ S) = MaxDimS (V ) + i(S).
Clearly because m(V ⋊ S) = MaxDimS (V ) + m(S), if S is flat, this indeed gives
an example of groups with m < MaxDim, and somehow they are as many as groups
with m < i. We need a lemma to prove the result:
Lemma 4.5. Let H < S be a subgroup, and p a prime not dividing the cardinality
of S. Then there exists a finite S-space V over Fp and a vector v ∈ V such that
Stab(v) = H.
Proof. Let V0 = Fp , viewed as the trivial H module. Let us consider the induced
module V = indSH (V0 ). An element f ∈ V is defined as a map f : S → V0 such that
f (gh) = h−1 (f (g)) = f (g) since V0 is a trivial H module, which is equivalent to that
f takes constant value along the left cosets of H. Now let us choose such a function
f ∈ V , taking 0 on H and 1 on all other left cosets. This is a nontrivial element in
V . We claim that Stab(f ) = H. Pick s ∈ Stab(f ), then by definition sf (g) = f (g)
for all g ∈ S. Taking g = 1 we see that sf (1) = f (s−1 ) = f (1), hence f (s−1 ) = 0 and
so s ∈ H. On the other hand, for any h ∈ H, hf (g) = f (h−1 g) = f (h−1 gh) = f (g)
as can be easily seen. Hence indeed Stab(f ) = H and the proof is complete.
13
Proof. (of Proposition 4.4) Fix a prime p not dividing the cardinality of S, and let
H1 , . . . , Hn be subgroups of S that are in general position with n = i(S). For each
i, there exists Vi over Fp and vi ∈ Vi such that Stab(vi ) = Hi . For each i, we can
decompose Vi into irreducible subspaces and the elements that stabilize vi stabilize
all factors in the decomposition. Meanwhile, if ∩1≤i≤t Mi , N2 , . . . , Nk are in general
position, there must be some i such that Mi , N2 , . . . , Nk are also in general position.
Hence we may assume that each vi spans an irreducible component of Vi under the
action of S. Let V = V1 ⊕ V2 ⊕ · · · ⊕ Vn , and consider G = V ⋊ S. First of all, it
is easily seen using Lemma 1.3 that MaxDim(G) ≤ MaxDimS (V ) + i(S), so we must
prove the other direction. We do this by exhibiting a sequence of this length.
Fix a decomposition of V , say W1 ⊕W2 ⊕· · ·⊕Wm such that the first n factors are
just Wi = spanS (vi ), the irreducible subspaces generated by each vi . Define maximal
subgroups of G by
M
Mi = (
Wj ) ⋊ S
j6=i
for 1 ≤ i ≤ m and
M
Tk = (
Wj )vk Svk−1
j6=k
for 1 ≤ k ≤ n. We claim that Mi , Tk are in general position. Let π : G → S be
the projection. First, any Tk does not contain the intersection of the other, since
π(∩1≤i≤n Mi ∩ Tk ) = π({vk svk−1 |s ∈ Hk }) = Hk and we know that all Hk are in
general position. Next, Mj for j > n does not contain the intersection of the other,
since the intersection of all except Mj contains Wj but Mj doesn’t. Finally, Mi
for each 1 ≤ i ≤ n does not contain the intersection of the other. Without loss of
generality, consider M1 , and let R = ∩j>1 Mj ∩ T1 , by calculation this is just v1 Sv1−1 ,
and we see that π(R ∩ M1 ) = H1 . What is π(R ∩ Tk ) for k > 1? For instance,
for any v1 sv1−1 ∈ R, if it is in Tk , then (v1 sv1−1 s−1 )s = (wvk svk−1 s−1 )s for some
w ∈ ⊕j6=k Wj , and so v1 sv1−1 s−1 − vk svk−1 s−1 = w for some w ∈ ⊕j6=k Wj . But since
the k-th component of w is trivial by definition, vk svk−1 s−1 = 0 and so s ∈ Hk . Now it
is easily seen that π(R ∩ Tk ) = Hk for k > 1, and because Hk are in general position,
M1 ∩ R does not contain the intersection of R ∩ Tk for k > 1. This indicates that the
subgroups are in general position. Counting the number of subgroups in question we
see that MaxDim(V ⋊ S) ≥ MaxDimS (V ) + i(S), and the result follows from these
two inequalities.
In fact, the pattern of subgroups of R is exactly of this form, and the reason
m(R) < MaxDim(R) is because R has a non flat top S(smallest non flat group, in
fact, by computation in GAP) and a bottom V which S acts on, such that stabilizers
of elements in V play a role in making MaxDim one larger. This also shows that in
Lemma 2.5, the subgroups π(R∩Mj ) are not predictable in the quotient, in particular
not necessarily maximal.
14
5
Acknowledgement
The author wants to take this chance to express his gratitude for many people, especially for Professor R. Keith Dennis and graduate mentor Ravi Fernando, for their
encouragements and advise, as well as all Cornell 2017 REU participants, for many
insightful comments and fruitful discussions. The author also wants to thank Professor Andrea Lucchini for his private communication with the author, giving the author
many new perspectives on this topic. The author is also very thankful for Michael
Vincent Crea for being a friend to the author and providing much support during this
summer.
15
References
[1] Thomas Brooks. Generating sets of mathieu groups. 2013.
[2] Peter J Cameron and Philippe Cara. Independent generating sets and geometries
for symmetric groups. Journal of Algebra, 258(2):641–650, 2002.
[3] Dan Collins. Finite Groups. 2017.
[4] Dan Collins. Generating Sequences of Finite Groups. 2017.
[5] Eloisa Detomi and Andrea Lucchini. Maximal subgroups of finite soluble
groups in general position. Annali di Matematica Pura ed Applicata (1923-),
195(4):1177–1183, 2016.
[6] Leonard Eugene Dickson. Linear groups: With an exposition of the Galois field
theory. Courier Corporation, 2003.
[7] Ravi Fernando. On an inequality of dimension-like invariants for finite groups.
arXiv preprint arXiv:1502.00360, 2015.
[8] Tianyue Liu and R Keith Dennis. Maxdim of some simple groups. arXiv preprint
arXiv:1712.04553, 2017.
[9] Andrea Lucchini. The largest size of a minimal generating set of a finite group.
Archiv der Mathematik, 101(1):1–8, 2013.
[10] Julius Whiston. Maximal independent generating sets of the symmetric group.
Journal of Algebra, 232(1):255–268, 2000.
[11] Julius Whiston and Jan Saxl. On the maximal size of independent generating
sets of psl2 (q). Journal of Algebra, 258(2):651–657, 2002.
16
| 4math.GR
|
1
Backlash Identification in Two-Mass Systems by
Delayed Relay Feedback
arXiv:1802.06246v1 [cs.SY] 17 Feb 2018
Michael Ruderman, Shota Yamada, Hiroshi Fujimoto
Abstract—Backlash, also known as mechanical play, is a piecewise differentiable nonlinearity which exists in several actuated
systems, comprising, e.g., rack-and-pinion drives, shaft couplings,
toothed gears, and other elements. Generally, the backlash is
nested between the moving elements of a complex dynamic
system, which handicaps its proper detection and identification. A
classical example is the two-mass system which can approximate
numerous mechanisms connected by a shaft (or link) with
relatively high stiffness and backlash in series. Information about
the presence and extent of the backlash is seldom exactly known
and is rather conditional upon factors such as wear, fatigue and
incipient failures in components. This paper proposes a novel
backlash identification method using one-side sensing of a twomass system. The method is based on the delayed relay operator
in feedback that allows stable and controllable limit cycles to be
induced, operating within the unknown backlash gap. The system
model, with structural transformations required for the one-side
backlash measurements, is given, along with the analysis of the
delayed relay in velocity feedback. Experimental evaluations are
shown for a two-inertia motor bench with gear coupling, with a
low backlash gap of about one degree.
Index Terms—backlash, two-mass system, relay, identification,
play operator, nonlinearities, limit cycles, online estimation
I. I NTRODUCTION
Mechanical backlash, i.e., the phenomenon of play between
adjacent movable parts, is well known in control systems
and causes rather disturbing side-effects such as lost motion,
undesired limit cycles in a closed control loop, reduction
of the apparent natural frequencies of systems and other
effects. For a review of this phenomenon we refer to [1]. One
possibility is that progressive wear and fatigue-related cracks
in mechanical structures can develop over the operation time of
an actuated system with backlash. This may occur in addition
to the appearance of parasitic noise; see, e.g., [2]. Moreover,
increasing backlash leads to more strongly pronounced chaotic
behavior [3] that, in general, mitigates against accurate motion
control of the system.
Two-mass systems with backlash, as schematically shown
in Fig. 1, constitute a rather large class of motion systems with
related issues. Here, a driving member, which is generally a
motor or actuator, is coupled to the driven member (load)
by various construction elements, i.e., gears, couplings or
kinematic pairs. The latter contain a finite gap, here denoted
by 2β , usually of a small size. Within this gap, both moving
M. Ruderman is with University of Agder, Norway
correspondence to: [email protected]
S. Yamada and H. Fujimoto are with the University of Tokyo, Japan
This work has received funding from the European Unions Horizon 2020
research and innovation programme (H2020-MSCA-RISE-2016) under the
Marie Sklodowska-Curie grant agreement No 734832.
masses become decoupled from each other. It is important to
note that this structure has a backlash feedback system which
differs significantly from feedthrough systems where a static
backlash element appears either in the input or output channel.
These types of systems with backlash have also attracted
considerable attention in systems analysis and control, see,
e.g., [4], [5]. However, they should not be confused with
the types of backlash feedback systems depicted in Fig. 1
and addressed in the following. Some previous studies which
consider analysis and control of backlash feedback systems
can be found in, e.g., [6]–[8].
x
!
Fig. 1. Two-mass system with backlash.
Despite the fact that the fundamental understanding of
backlash mechanisms appears to be something of a solved
problem, case-specific modeling, and above all, reliable detection and identification, still remain relevant and open-ended
problems in systems and control engineering. Applications
can be found in industrial robotic manipulators, e.g., [9],
flexible medical robots, e.g., [10], servo drive systems, e.g.,
[11], [12], automotive power trains, e.g., [13] and other areas.
Information about the presence and extent of backlash, which
is a rather undesirable structural element, is seldom exactly
known, and usually requires identification under normal operational conditions. When both the motor-side and the loadside of a two-mass (two-inertia) system are equipped with
high-resolution position sensors (encoders) (see for instance
[14]), the identification of backlash becomes a trivial task
which can be accomplished under various quasi-static, lowexcitation conditions. A frequently encountered case, however,
is when load-side sensing is not available, or is associated with
special auxiliary measures which can differ markedly from the
normal operational conditions. A typical situation occurs when
only the driving actuator is equipped with an encoder or other
position-giving sensor, while the total drive train, presumably
with backlash, has no access to the additional measurements.
In the last two decades, various strategies for estimation,
and hence identification, of the backlash between two moving
members, have been developed. A very early study and analysis of the backlash phenomenon in geared mechanisms was
2
made by [15]. A specially designed impulse excitation and
motor current filtering method was proposed in [11] for highprecision servo systems. While straightforward in realization,
the method relies on the assumption that the torque impulses
are sufficient to move the motor and, at the same time, to
keep the load immobile. For systems with high and uncertain
damping and additional elasticity modes, such excitation can
be challenging with regard to the realization and manifestation
of the backlash. The nonlinear observation of backlash states
was proposed and evaluated in [16], though both motor-side
and load-side measurements were required. An identification approach based on the ridges and skeletons of wavelet
transforms was demonstrated by [17]. The method requires
a relatively broadband frequency excitation and a combined
time-frequency analysis that can be challenging in a practical
application. One of the most established strategies of backlash
identification, using motor-side sensing only, was reported in
[12]. It should be noted, however, that the underlying ideas
were previously provided in [15]. This approach will also
be taken as a reference method for experimental evaluation
in the present work. Another, more frequency-domain-related
approach for analyzing backlash by approximating with a
describing function, can be found in [18]. This relies on the socalled ”exact backlash mode” introduced in [19]. Describing
the function analysis of systems with impact and backlash
can be found in [20]. An experimental comparison of several
backlash identification methods, mainly based on the previous
works of [11]–[13], can be found in [21]. The authors concluded that the method of integration of speed [12] was the
most accurate of those under evaluation.
The objective of this paper is to introduce a new strategy
for identifying the backlash in two-mass systems when only
motor-side sensing is available. In addition, this method does
not require a large excitation of the motion dynamics, in contrast to the case of [12]. This can be advantageous for various
machines and mechanisms where the load cannot be driven
at higher velocities and accelerations. The proposed method
relies on the appearance of stable and controllable limit cycles,
while using a delayed relay in the velocity feedback loop.
We recall that relay feedback systems have been intensively
studied since the earlier pioneering works [22], [23] and
successfully applied in, e.g., auto-tuning of controllers [24]
and other purposes such as mass identification [25]. In the
present study, the uncertainties and parasitic effects of the
system dynamics are not explicitly taken into account, but
the relay in feedback provides the necessary robustification,
which is also confirmed by experiments. Further, we note that
attempts to use relay feedback for identifying backlash in twomass systems were also made in [26]. However, the relay
feedback was solely used for inducing a large-amplitude periodic motion, while the underlying identification strategy relied
on the motor speed pattern analysis, as in [12]. Furthermore,
exact detection of the zero-crossing and extremum instants is
required, along with the ratio of the masses (inertias).
The rest of the paper is organized as follows. In section II we
describe the underlying modeling approach for two-mass systems with backlash, and the required structural transformations
which allow consideration of the two-mass system as a plant
with motor-side sensing only. Section III contains a detailed
analysis of the feedback relay system and induced limit cycles,
including controllable drifting, which enables motor operation
within and beyond the backlash gap. An experimental case
study with the minimal necessary identification of system
parameters and relay-based backlash identification is provided
in section IV, together with a comparison with the method
reported in [12]. The study is summarized, and the conclusions
drawn in section V.
II. T WO -M ASS S YSTEM WITH BACKLASH
The block diagram of a generalized two-mass system with
backlash is shown in Fig. 2. Note that this encompasses the
principal structure introduced in section I, (compare with Fig.
1). The motor-side and load-side dynamics, described by the
relative coordinates xm and xL respectively, are given by
mẍm + d ẋm + f sign(ẋm ) = u − τ ,
(1)
M ẍL + DẋL + Fsign(ẋL ) = τ .
(2)
Both are coupled in the forward and feedback manner by
the overall transmitted force τ , which is often referred to
as a link (or joint) force. Note that in the following we
will use the single terms force, displacement and velocity
while keeping in mind the generalized forces corresponding
to generalized motion variables. Thus, the rotary coordinates
and corresponding rotary degrees of freedom will also be
considered without changing the notation of the introduced
variables, for example, in the case study of the two-inertia
system presented in section IV. Clearly, the parameters m,
M, d, D, f and F are the motor and load masses, damping
and Coulomb friction coefficients respectively. We will also
denote the known invertible mapping (u − τ ) 7→ xm by Gm
and τ 7→ xL by GL . Here, we explicitly avoid any notations
of the transfer function since the dynamics of (1) and (2)
include also the nonlinear terms of Coulomb friction. Using
the nonlinear function g(·), where parameters are unknown
but the structure may be assumed, we will elucidate the force
transmission characteristics of the link.
"
"
"
!
u
"
"
"
Gm
xm
!
g #!
"
GL
"
xL
"
"
Fig. 2. Block diagram of two-mass system with backlash in the link.
Commonly, the link is excited by the relative displacement
δ = xm − xL
(3)
between the motor and load. We assume that the link transfer
characteristics offer a relatively high but finite stiffness, and incorporate a backlash connected in series. Under high stiffness
conditions we detect and identify the first resonance peak of
the measurable frequency range which is sufficiently beyond
the input frequencies and control bandwidth of the system.
We assume that the backlash itself is sufficiently damped
so that the relative displacement between the ”impact pair” is
3
not subject to any long-term oscillations when the mechanics
engage. Here, we recall that the simplest modeling approach
for the backlash in two-mass systems, used for instance in [12],
[14], applies the dead-zone operator only to δ , while the deadzone output is subsequently gained from the stiffness coefficient of the link. This arrangement results in proper backlash
behavior at lower frequencies but tends to produce spurious δ oscillations at higher harmonics. More detailed consideration
of the backlash in two-mass systems incorporates a nonlinear
damping at impact [8], for which advanced impact models can
be found, e.g., in [27], [28]. Note also that in [19], an inelastic
impact was considered when the backlash gap was closed, and
the switching case distinction for gap and contact modes was
made (compare further with eq. (4)).
The above assumption of sufficient backlash damping allows
a simple structural transformation of the block diagram shown
in Fig. 2 to be made. Assuming rigid coupling at the backlash
contacts and loss-of-force transmission at the backlash gap,
an equivalent model can be obtained, as shown in Fig. 3.
Obviously, the consideration of the backlash element in Fig.
3 is purely kinematic, so that an ideal play-type hysteresis
[29], [30] is assumed between the motor and load displacements. The kinematic backlash, which is the Prandtl-Ishlinskii
operator of the play type, can be written in differential form
ẋm if xL = xm − β , ẋm > 0 ,
ẋ
m if xL = xm + β , ẋm < 0 ,
ẋL =
(4)
0
if xm − β < xL < xm + β ,
0
otherwise ,
as in [9], [31], where the total gap is 2β . Note that the backlash
acts as a structure-switching nonlinearity xL = P[xm ], so that
two operational modes are distinguished. The first mode is
within the backlash gap, i.e., when |δ | < β , implies zero
feedback of the G−1
L dynamics. The second mode, is during
mechanics engagement, i.e., |δ | = β provides the feedback
coupling by G−1
L , so that the system performs as a single
mass with lumped parameters (m+M) and (d +D). During the
switching between modes the play operator (4) becomes nondifferentiable. That is, the first time-derivative of the P-output
to an inherently C 2 -smooth motor position input contains step
discontinuities. Correspondingly, the second time-derivative
constitutes the weighted delta impulses, in terms of the distribution theory. These impulses can be interpreted as an instantaneous impact force τ which excites the motor dynamics when
the backlash switches between gap and engagement, and vice
versa. We should stress that this switching mechanism does
not explicitly account for the link stiffness and damping at
impact, as previously mentioned. However, this is not critical
at lower frequencies, since the link stiffness is assumed to
be sufficiently high, and the motor and load dynamics are
subject to viscous and Coulomb friction damping; see (1) and
(2). Also we note that the hybrid approach, with a switching
structure for the gap and engagement, was also used in [32] for
modeling and control of mechanical systems with backlash.
A significant feature of the equivalent model shown in Fig.
3 is that it allows the two-mass system with backlash to be
considered as a single closed loop, while assuming that only
GL !
u
!
xm
xL
Gm
Fig. 3. Equivalent block diagram for modeling the two-mass system with
backlash described by the play-type hysteresis operator.
the motor-side measurement is available. This is the decisive
property we will make use of when identifying backlash using
one-side sensing of the two-mass system. It should be recalled
that the loop is closed during the engagement mode and
open during the gap mode of the backlash. This results in
an impulsive behavior during mode switching. Therefore, the
motor dynamics are disturbed periodically when a controllable
stable limit cycle occurs. We analyze this situation further in
section III-C.
III. D ELAYED R ELAY F EEDBACK
The proposed backlash identification is based on the delayed
relay in the feedback of the motor velocity. Consider the
relay feedback system as shown in Fig. 4. Note that the
structure represents our system, as introduced in section II,
when operating in the backlash gap mode, i.e., y = 0. From
the point of view of a control loop with relay, the switching
feedback force (see Fig. 3) can be seen as an exogenous signal
y. The non-ideal relay, also known as a hysteron [29], switches
between two output values ±h, while the switching input is
”delayed” by the predefined threshold values ±e. Therefore 2e
is the relay width, corresponding to the size of the hysteron,
and h is the amplification gain of the relay operator. The
hysteron with input z is defined, according to [33], as:
h
i
H(t) = h min sign(z + e), max H(t− ), sign(z − e) , (5)
while its initial state at t0 is given by
(
h sign z(t0 ) , if z(t0 ) ∈ (−∞, −e] ∨ [e, ∞),
H(t0 ) =
{−h, +h},
otherwise.
(6)
It can be seen that the hysteron has a memory of the previous
state at t − and keeps its value as long as z ∈ (−e, e).
In the following, we will first assume that the relay feedback
system from Fig. 4 is self-sustaining, i.e., without a disturbing
factor, i.e., y = 0. Subsequently, in sections III-B and III-C we
will allow for y 6= 0 and analyze the resulting behavior while
approaching the proposed strategy of backlash identification.
Note that y 6= 0 corresponds to the engagement mode of the
backlash, with reference to our original system from Fig. 3.
A. Symmetric Unimodal Stable Limit Cycle
It is well known that relay feedback systems can possess a
stable limit cycle [34]. For a single-input-single-output (SISO)
4
are matched operators of one and the same argument ẋm . Both
are discrete switching operators, so that
y
xm
Gm
h> f
u
is required for the relay to overcome the Coulomb friction. It
is essential that the velocity sign changes while the (delayed)
relay simultaneously keeps its control value. For the boundary
case f < h → f , the relay feedback system with, e.g., H < 0
becomes mẍm = −d ẋm + 2 f after the velocity sign changes
from positive to negative. Based on this, the condition for the
relay threshold results in
linear time invariant (LTI) system we write
=
Ax + Bu,
z = Cx,
(7)
(8)
assuming the n × n system matrix A is a Hurwitz matrix, and
assuming that the input and output coupling vectors B and C
satisfy CB > 0. When a relay u = −H[z] closes the feedback
loop, the state space obtains two parallel switching surfaces
which partition the state space into the corresponding cells.
For H > 0, the system dynamics is given by ẋ = Ax − Bh, and
for H < 0 by ẋ = Ax + Bh, while the total amplitude of the
relay switching is 2h, see (5). If the relay in (5) is non-ideal,
i.e., e > 0, then the existence of a solution of (7) and (8) with
relay (5) in feedback is guaranteed [35], and the state trajectory
will reach one of the switching surfaces for an arbitrary initial
state. The necessary condition for globally stable limit cycles
of the relay feedback system with a relay gain h, is given by
CA−1 Bh + e < 0,
(9)
(compare with [35]). Otherwise, a trajectory starting at A−1 BH
would not converge to the limit cycle. In the following, we
describe the most significant characteristics of the stable limit
cycles of the relay feedback system as shown in Fig. 4, based
on the above assumptions and conditions from the literature.
For more details on the existence and analysis of limit cycles
in relay feedback systems we refer to [34]–[36]. The limit
cycle existence condition in control systems with backlash and
friction has also been studied in [37].
Considering the motor velocity dynamics (1), first without
Coulomb friction, and the relay (5) in negative feedback, the
condition for stable limit cycles (9) becomes
e<
h
.
d
2f
,
(12)
d
while following the same line of development as that for
deriving (10) from (9). Note that all three conditions (10)(12) should hold, in order to guarantee the existence of a stable
limit cycle once the Coulomb friction is incorporated.
For determining the position amplitude of the limit cycles
whose velocity amplitude is 2e by definition, consider the state
trajectories of the system mẍm + d ẋm + f sign(ẋm ) = −h. The
corresponding initial velocity is e, immediately after the relay
switches down to −h. It can easily be shown that the associated
phase portrait of state trajectories is given by
e<
xm = −
−0.5mẋ2m
.
d ẋm + h + f sign(ẋm )
(13)
Solving (13) for three characteristic velocities ẋm = {e, 0, −e}
the start, maximal and minimal positions of a limit cycle
are obtained –x3, x4, and x2 respectively –as shown in Fig.
5. Note that in the case of d, f = 0, the points x2 and
e
m
Fig. 4. Relay feedback system with motor dynamics in the loop.
dx /dt
t
ẋ
(11)
0
−e
x1
x2
x
x3
x4
m
Fig. 5. Symmetric unimodal limit cycle.
(10)
In fact, the velocity threshold |e| should be first reachable for
a given system damping and excitation force provided during
the last switching of the relay. Therefore, (10) specifies the
boundaries for parameterization of the relay, in order to ensure
that the trajectory reaches one of the switching surfaces ±e
independently of the initial state. In cases where the motor
damping is not well known, its upper bound can be assumed,
thus capturing the ”worst case” of an overdamped system.
Once the Coulomb friction contributes equally in (1), the
condition in (10) becomes a necessary but not sufficient condition. Obviously, sufficient conditions should relate the motor
Coulomb friction coefficient f to the relay, as in (5), since both
x3 coincide with each other. Since the symmetric relay (5)
induces a symmetric unimodal limit cycle, the second halfcycle can be obtained by mirroring the first one, derived
using (13); once across the axis connecting both switching
points at ±e and once across the orthogonal axis going
through its center (see Fig. 5). The intersection of the axes
constitutes the origin of the limit cycles, which is the manifold
{xm , | ẋm = 0}. Obviously, the xm coordinate depends on the
initial conditions and the system excitation. We also recall
that the limit cycle is symmetric if ξ (t + t ∗ ) = −ξ (t) where
ξ (t) is a nontrivial periodic solution of (5)-(8) with period
2t ∗ . The limit cycle is also called unimodal when it switches
only twice per period; see [35] for details. The symmetry and
5
unimodality of the limit cycle, as depicted in Fig. 5, implies
that x4 − x3 = x2 − x1. Therefore the total position amplitude,
denoted by Xξ , becomes 2(x4 − x3) + (x3 − x2). Evaluating x2,
x3, x4 points, computed by (13), yields
Xξ = −
e2 h m
.
( f + h + d e)( f − h + d e)
xm
xe xs
e
xm
(14)
For determining the period of the limit cycle, the same
differential equation mẍm + d ẋm + f sign(ẋm ) = −h is solved
with respect to time, together with those for the initial and final
values ẋm (0) = e and ẋm (t ∗ ) = −e respectively. This yields
f + h i
de
mh
+ ln
.
(15)
t ∗ = − ln 1 +
d
f −h
f +h+de
B. Drifting Limit Cycle
In the previous subsection, the conditions for a unimodal
limit cycle of the system as shown in Fig. 4 were derived,
and the characteristic features in terms of the displacement
amplitude and period, were given by (14) and (15). We
note that the limit cycle can appear within the gap mode of
backlash, (see section II), provided Xξ < 2β , so that the motor
dynamics and load remain decoupled (compare with Fig. 3).
In order to realize a drifting limit cycle, consider a modified
relay from (5) and (6) so that the amplitude becomes
α+ h0 , if sign(H) > 0,
h=
(16)
α− h0 , if sign(H) < 0.
Here, h0 is the amplitude of the underlying symmetric relay
(5) and (6), and α+ , α− ≥ 1 with α+ 6= α− scales this, so
as to provide an unbalanced control effort when switching.
This leads to differing acceleration and deceleration phases
of the limit cycle, so that the trajectory does not close after
one period (compare with Fig. 5), and becomes continuously
drifting as in the example shown in Fig. 6. Note that the case
where α− > α+ is illustrated here. As in section III-A, one
can solve the trajectories between two consecutive switches
separately for α− h0 and α+ h0 , and obtain the relative shift
XC = xe − xs of the limit cycle per period as follows:
−e2 h20 m(α−2 − α+2 ) ( f + d e)
.
(−α−2 h20 + d 2 e2 + 2de + f 2)(−α+2 h20 + d 2e2 + 2de + f 2)
(17)
Following the same line of developments as in section III-A,
the period of the drifting limit cycle, i.e., the time of arrival
at xe can be obtained, according to:
f + de − α−h0
m
f + α− h0
+ ln
+
TC = −
ln
d
f + de + α−h0
f − α− h0
f + α+ h0
f + de − α+h0
+ ln
+ ln
(18)
f + de + α+h0
f − α+ h0
XC =
Obviously, the average drifting velocity of the limit cycle can
be computed from (17) and (18) as ẊC ≈ XC /TC . Note that (17)
and (18) characterize the drifting limit cycle which proceeds
within the backlash gap without impact with the load side. The
impact behavior of drifting limit cycles is addressed below,
from the associated relative load displacement point of view.
e
XC
Fig. 6. Drifting limit cycle of amplitude-asymmetric relay (α− > α+ )
C. Impact Behavior at Limit Cycles
In order to analyze the impact behavior at limit cycles, i.e.,
the situation where the load side is permanently shifted, as
it is periodically excited by the drifting limit cycle, consider
first the impact phenomenon of two colliding masses m and
M. Note that when considering the impact and the engagement
mode of the backlash, the conditions (10)-(12) are not explicitly re-evaluated due to nontrivial (and also possibly chaotic)
explicit solutions. However, making some smooth assumptions
below, we evaluate the impact behavior at the limit cycles and
show their persistence within numerical simulations and later
in experiments.
Introducing the coefficient of restitution 0 ≤ ε ≤ 1 and
denoting the velocities immediately after impact by the superscript ”+ ” one can show that
ẋL (1 − ε )M + ẋm (m − ε M)
,
(19)
M+m
ẋL (M − ε m) + ẋm (1 + ε )m
ẋ+
=
.
(20)
L
M+m
The above velocity jumps directly follow from Newton’s law
and the conservation of momentum. See, e.g., [20] for details.
Recall that a zero coefficient of restitution means a fully plastic
collision, while ε = 1 constitutes the ideal elastic case. For the
relatively low velocities at impact (since small backlash gaps
are usually assumed) and stiff (metallic) backlash structures,
we take ε ≈ 1 (compare with [27]). Further, we assume the
load velocity to be zero and the motor velocity amplitude to
be (in the worst case) maximal, i.e., max |ẋm | = e, immediately
before the impact. Due to above assumptions, the load velocity
(20) after impact can be determined as an upper bound:
2me
.
(21)
ẋ¯+
L =
M+m
Note that (21) constitutes an ideal case without any structural and frictional damping at impact, whereas a real load
velocity after impact will generally be lower in amplitude than
(21). Nevertheless, (21) provides a reasonable measure of the
relative load motion immediately after the backlash engages.
Solving the unidirectional motion M ẍL + DẋL + F = 0, which,
for zero final velocity and an initial motion given by (21),
yields the relative displacement of the load in the idle state as
2M e m
F(M + m)
FM
+
.
(22)
XL = 2 ln
D
F(M + m) + 2D e m
D(M + m)
ẋ+
m
=
6
nor stiction of the load is taken into account. Thus, in a real
two-mass system with backlash, a nonuniform motion during
engagement mode is expected to differ markedly from that
within the backlash gap, (see the experiments in section IV).
−3
x 10
gap mode
engagement mode
0
xm, xL (−)
Intuitively, assuming well-damped (through F and D) load
dynamics, the load displacement XL due to a single impulse at
the impact should be relatively low. At the same time, it can
be seen that the logarithmic contribution with negative sign in
the first summand of (22) is balanced by the linear increase
of the second summand, depending on the parameter e. In the
sum, the quadratic shape of XL as a function of e occurs at
lower values of e, and it approaches a linear slope once e
grows (see Fig. 7). Hence, the e parameter should be kept as
low-valued as possible, given the system parameters m, M, D
and F. This will ensure a displacement of the load which is
as low as possible, i.e., XL → 0, induced by the single impulse
at impact. Furthermore, low values of e are required firstly to
fulfill the necessary and sufficient conditions (10) and (12) of
the limit cycle, and secondly to keep the amplitude as low as
possible and therefore within the backlash gap (see (14)).
−2
−4
−6
0
motor
load
0.1
0.2
time (s)
0.3
Fig. 8. Motor and load displacement during the gap and engagement modes.
X
L
IV. E XPERIMENTAL C ASE S TUDY
0
0
e
Fig. 7. Relative displacement of the load after the backlash impact as a
function of the relay parameter e according to (22); qualitative example.
Obviously, the maximal load displacement (22) will be
induced periodically by the drifting limit cycle, (see section
III-B), each time the impact occurs. In fact, the motor side,
which is moving in a drifting limit cycle, produces a sequence
of impulses that continuously ”push” the load side once the
instantaneous impulse can overcome the system stiction. The
resulting load motion, though periodic due to the drifting limit
cycle, can be rather chaotic, due to the amplitude and phase
shifts produced by the series of impulses. Due to the inherent uncertainties of the Coulomb friction, viscous damping
[38] and non-deterministic system stiction, an exact (analytic)
computation of the resulting load motion appears to be only
marginally feasible, with unsystematic errors that reveal the
predicted trajectory as less credible. At the same time, it can
be shown that for the h amplitude selection to be close to
f + F, a continuous propulsion of the load should occur in the
direction of ẊC . An example of a numerical simulation of the
system (1) and (2) with backlash and an asymmetric relay (5),
(6) and (16) in feedback, is shown in Fig. 8. It can be seen that
the motor displacement exhibits a uniform drifting limit cycle
during the gap mode, until it reaches the backlash boundary
and begins to interact with the load side. It is evident that the
average drifting velocity of the limit cycle ẊC , corresponding
to the slope in Fig. 8, differs between the gap and engagement
modes. This makes the backlash boundary easily detectable
based on the motor displacement trajectory only. Note that in
the numerical simulation shown, neither damping uncertainties
This section is devoted to an experimental case study on
detecting and identifying backlash in a two-inertia system,
using only motor-side sensing. The parameters of linear system
dynamics, i.e., inertia and damping, are first identified from
the frequency response function. The total Coulomb friction
level required for the analysis of stable limit cycles (see
section III), is estimated using simple unidirectional motion
experiments. In addition, the steady limit cycles previously
analyzed are induced and confirmed with experiments for
unconstrained operation within the backlash gap. The nominal
backlash size was measured by means of both motor-side and
load-side encoders and used as a reference for identification.
The proposed backlash identification approach, as described
in section III, was followed and evaluated in the laboratory
setup. In addition, the reference method [12] was evaluated
experimentally for two different excitation conditions and
compared with the proposed method.
A. Laboratory Setup
An experimental laboratory setup (see picture in Fig. 9),
consisting of two identical motors, each with a 20-bit highresolution encoder, was used in this study. Note that only
Fig. 9. Experimental setup with a gear coupling.
the motor-side encoder is utilized for the required system
identification and for evaluation of the proposed backlash
7
estimation method. The load-side encoder is used for reference
measurements only. In this setup, the backlash can be added or
removed by replacing the rigid coupling with the gear coupling
shown in the picture. A standard proportional-integral (PI)
current controller is implemented on-board with the motor
amplifier, with a bandwidth of 1.2 kHz. In the experiments,
the sampling frequency is set to 2.5 kHz and the control parts
are discretized using the Tustin method.
B. System Identification
1) Two-Mass System Parameters: The parameters of the
two-mass system, i.e., m, d, M and D, can be identified by
measuring the frequency response function (FRF) on the motor
side. It is worth emphasizing that the identified motor inertia
and damping are sufficient for analyzing and applying the
proposed backlash identification approach. Thus, if the above
parameters are available or otherwise previously identified,
the FRF-based identification discussed below will not be
required. The measured frequency characteristics shown in
Fig. 10, show the antiresonance and resonance behavior. The
0
-50
-100
-150
101
102
103
102
103
0
-200
-400
-600
101
The measured FRF shows that the setup can be modeled
as a two-mass system which has an antiresonance frequency
of 409 Hz and a resonance frequency of 583 Hz. The fitted
model response is indicated in Fig. 10 by the blue solid line,
while the measurement results are indicated by the red dashed
line. A visible discrepancy in the phase response is due to an
inherent time delay in the digital control system. However, this
becomes significant in a higher frequency range. The identified
parameters are listed in Table I.
2) Coulomb Friction: The combined motor-side and loadside Coulomb friction f + F is identified by measuring the
motor torque when the motor-side velocity is controlled for
constant reference values. Figure 11 shows the obtained
torque-velocity measurements. Blue circles indicate the measured data, captured from multiple constant velocity drive
experiments, and the red line indicates the linear slope fitted by
the least-squares method. Here, the Coulomb friction torque of
0.1 Nm is determined from the intercept of the fitted line with
the torque axis. In the same way, the friction-velocity curve
is determined for the negative velocity range, not shown here
due to its similarity. The determined Coulomb friction value
is −0.0993 Nm which demonstrates that the friction behavior
of the total drive is sufficiently symmetrical around zero. An
average value of f + F = 0.0999 Nm is further assumed, while
a very rough estimation of the motor-side Coulomb friction
is half of the total value, i.e., f ≈ 0.05 Nm. Note that an
exact knowledge of f is not required to satisfy the conditions
derived in section III. Knowledge of the total Coulomb friction
coefficient f + F alone is sufficient for applying the relay
feedback system.
0.135
0.13
Fig. 10. Frequency characteristics measurements of the setup from the motor
torque to the motor-side angle versus the fitted two-mass system model.
0.125
0.12
0.115
TABLE I
I DENTIFIED PARAMETERS OF TWO - MASS SYSTEM .
2
Motor-side inertia m
8.78e-4
kgm
Motor-side damping coefficient d
6.20e-2
Nms/rad
Load-side inertia M
8.78e-4
kgm2
Load-side damping coefficient D
3.60e-2
Nms/rad
0.11
0.105
2
parameters of the two-mass (two-inertia) system are identified
by fitting both peak regions. Here, the frequency characteristics
from the motor torque to the motor-side angle are determined
for the setup equipped with rigid coupling (without backlash).
This corresponds to the case of a nominal plant where no
backlash disturbance develops during the operation. Note that
for a sufficient system excitation, similar FRF characteristics
can also be obtained in the presence of a small backlash.
Otherwise, the basic parameters, such as lumped mass and
damping of the motor and load sides, are assumed to be
available from the manufacturer’s data sheets and additional
design or manufacturing data.
3
4
5
6
7
8
9
10
11
12
Fig. 11. Friction-velocity identification measurements.
3) Nominal Backlash Size: For evaluation of the proposed
method, the nominal backlash size is first identified using
both the motor- and load-side encoders. Figure 12 shows
the measured (xm , xL ) map when the motor-side position is
open-loop controlled by a low-amplitude and low-frequency
sinusoidal wave. The detected backlash is 2β = 19.05 mrad,
which can be seen as a low value, i.e., below one degree.
It should be emphasized that when the motor is moving
forwards after a negative direction reversal, the load side is at
first moving together with the motor side, even though both are
within the backlash gap (see the lower left-hand range in Fig.
12). This spurious side-effect can be explained by adhesion
between the motor and load sides and the specific structure
8
10
5
0
-5
-10
-10
-5
0
5
10
15
20
25
Fig. 12. Nominal backlash identified using both encoders.
of the gear coupling. Since the gear coupling used in the
setup consists of an internal tooth ring and two external tooth
gears, it exhibits, so to speak, a double backlash, which cannot
be exactly captured by the standard modeling assumptions
previously made for a two-mass system with backlash. It
is evident that the forward transitions, as depicted in Fig.
12 and further in Fig. 16 and Fig. 17, exhibit some nondeterministic creep-like behavior that cannot be attributed
to backlash nonlinearity as assumed. Therefore, in order to
be able to consider the backlash effect without adhesive
disturbances, only the backward-moving phases are regarded
as suitable for evaluation.
delay of the control system so that the actual switchings do
not occur exactly at ±e = ±0.1 rad/sec as required by the relay
parameterization. Rather, they appear at ±0.2 rad/sec, which is
double the set relay parameter value e. Nevertheless, the limit
cycle remains stable, almost without drifting, and displays the
expected characteristic shape as shown in Fig. 5. Here, we note
that the gear coupling with backlash was installed, so that the
steady limit cycle as shown in Fig. 13 clearly occurs within
the backlash gap, without impact with the load side. The halfperiod t ∗ , determined from the measurements, corresponds to
about six or seven samplings, i.e., 2.4 to 2.8 msec, and is in
accordance with that computed by (15). Recall that knowledge
of t ∗ is significant for deciding the sufficiency of the sampling
rate with regard to the relay control parameters.
D. Motor-Side Backlash Identification
The proposed backlash identification method is implemented and evaluated experimentally. The results are evaluated
for two different sets of relay parameters. The two cases are
defined as follows. Case 1: h0 =0.12, e=0.1, α = 2, and Case 2:
h0 =0.1, e=0.1, α = 2.5. The α values are selected by trial and
error when operating the feedback relay system and observing
sufficient movement (drift) of the resulting limit cycle (see
Fig. 8). Note that a periodic sequence for which the relay
asymmetry coefficients α+ and α− alternate, for example,
[α+ , α− ] = [2, 1] for the first five seconds followed by [α+ , α− ]
= [1, 2] for the next five seconds, and so on, has been applied.
This allows for changing the drift direction of the limit cycles,
C. Steady Limit Cycle
As described in Section III-A, the delayed relay feedback
produces steady limit cycles. The experimentally measured
cycle is shown in Fig. 13. The assumed relay parameters
15
10
5
0.3
0
0.2
-5
0.1
-10
0
-15
0
2
4
6
8
10
12
14
16
18
20
-0.1
-0.2
Fig. 14. Backlash identification using the proposed method, Case 1.
-0.3
-0.2
-0.15
-0.1
-0.05
0
0.05
0.1
0.15
Fig. 13. Steady limit cycle caused by delayed relay feedback.
h = 0.1 and e = 0.1 satisfy (10)-(12). The total amplitude Xξ
of the limit cycle is required to be small compared to the
backlash size. In addition, the period of the limit cycle 2t ∗
is required to be large enough, with respect to the sampling
rate, so as to ensure a sufficient number of samples per period
of the limit cycle. The values computed according to (14) and
(15) are Xξ = 0.128 mrad and t ∗ = 2.45 msec respectively. The
average Xξ value, determined from the example measurements
shown in Fig. 13, is about 0.28 mrad. This is not surprising,
since the measured steady limit cycle is subject to the time
and therefore exploring the backlash gap in both directions as
well as beyond the gap, i.e., during the engagement mode. As
stated in section IV-B, only the backward-moving phases are
evaluated in order to estimate the backlash without adhesion
side-effects.
The motor position responses in Case 1 and Case 2 are
shown in Fig. 14 and Fig. 15 respectively. The full backlash
width 2β can be determined by the sudden change of the
motor position’s slope. The slope of the motor position is
steep and uniform within the backlash gap, since the load
is decoupled from the motor side. When the impact between
them occurs, the slope suddenly changes and becomes less
steep and also irregular. The determined 2β values in Case
1 and Case 2 are 19.30 mrad and 23.62 mrad respectively,
9
limit cycles generated by the delayed relay with asymmetric
amplitudes are drifting controllably along all transitions of
the backlash. Only the forward-moving phases suffer from
adhesion side-effects, while the backwards phases coincide
exactly with the backlash shape (see Fig. 12).
25
20
15
10
5
0
E. Comparison with Reference Method
-5
-10
-15
0
2
4
6
8
10
12
14
16
18
20
Fig. 15. Backlash identification using the proposed method, Case 2.
while the nominal backlash width identified using the loadside encoder, as in section IV-B, is 19.05 mrad. This confirms
that the proposed method can identify the backlash width
using the motor-side information only. Note that the recorded
motor position pattern is subject to various deviations between
the slopes over multiple periods which mitigates against an
exact read-off of the backlash size value (see Fig. 14 and
Fig. 15). For accuracy enhancement and correspondingly better
generalization, the 2β values read off over multiple periods
can be averaged. Figures 16 and 17 show the corresponding
2
0
-2
-4
In order to assess the performance of the proposed method,
a well-established backlash identification strategy [12] has
been taken as reference method. This method also uses the
motor-side information only when identifying the backlash in
a two-inertia system. The basics of the reference method are
described below. For more details we refer to [12], [15].
The reference method applies the triangular-wave reference
velocity to the PI motor velocity controller. Considering the
controlled plant as a one-inertia system, i.e., with the lumped
parameters m + M and d + D, the PI velocity controller is
designed by the pole placement such that its control bandwidth
is set to 5 Hz. After the sign of the motor-side acceleration
changes, due to the triangular velocity reference, the motor
side moves back from one end to the opposite end of the
backlash gap. At the same time, the load side continues to
(freely) move in the initial direction until the backlash impact.
The instant when the motor and load sides are decoupled is
defined as t1 and the instant when the motor and load sides are
in contact again is defined as t2 (see Fig. 18 and Fig. 19). The
reference method identifies the backlash width by integrating
the relative motor velocity between these two instants. If the
load-side viscosity is small enough, and t2 −t1 is short enough,
the decrease of the load-side velocity during t2 − t1 can be
neglected. Therefore, one can assume ẋL (t) = ẋˆL = ẋm (t1 ) for
t1 ≤ t ≤ t2 . Then, the total backlash width is estimated as
-6
2β =
-8
-15
-10
-5
0
5
10
15
Fig. 16. Measured xL -xm map in Case 1.
10
5
0
-5
-10
-15
-10
-5
0
5
10
15
20
25
Fig. 17. Measured xL -xm map in Case 2.
(xm , xL ) maps, as a reference measurement, for both experimental cases. Recall that the load-side encoder signal is not
used for backlash identification. Both figures indicate that the
Zt2
t1
ẋL (t) − ẋm (t) dt ≃ Ts
k2
∑
k1
ẋˆL − ẋm (k) ,
(23)
where Ts is the sampling time and k1t = t1 and k2t = t2 .
The results from two different triangular waves are evaluated. Here, two cases are defined as Case 3 –for which the
triangular wave slope is 1400 and the period is 0.2 s, and
Case 4 –for which the triangular wave slope is 500 and the
period is 0.4 s. Obviously, Case 3 is more ”aggressive” in
terms of the system excitation, as regards the higher reference
acceleration/deceleration of the total drive. Fig. 18 and Fig.
19 show the measured velocity responses of Case 3 and Case
4 respectively. Since the acceleration in Case 3 is larger than
that in Case 4, the decrease in ẋL (t) for t1 ≤ t ≤ t2 is smaller,
which reduces the estimation error. According to (23), the 2β
values are calculated as 31.0 mrad and 42.6 mrad for Case
3 and Case 4 respectively. Inherently, the reference method
provides an overestimated backlash width, as is apparent from
Fig. 18 and Fig. 19. In order to reduce the estimation error,
which is mainly driven by the assumption ẋL (t) = ẋˆL = ẋm (t1 ),
the reference method requires larger accelerations, to make
the decoupling between the motor-side and load-side faster
once the sign of the motor acceleration changes. Therefore,
the reference method inherently requires more ”aggressive”
10
70
65
60
55
50
45
40
0.07
0.08
0.09
0.1
0.11
0.12
0.13
Fig. 18. Backlash identification using the reference method, Case 3.
55
for inducing the stable limit cycles of amplitudes significantly
lower than the backlash gap. The limit cycles can then be
operated as drifting while being controlled by an asymmetric
relay amplitude. We provide a detailed analysis of steady
and drifting limit cycles for two-mass systems with backlash,
and derive the conditions for the minimal set of system
parameters which can be easily identified. The experimental
evaluation of the method is performed on a test bench with
two coupled motors, where the rigid coupling is replaceable
by one with a small backlash gap of about one degree. The
known identification method reported in [12] is taken as a
reference for comparison. Several advantages of the proposed
method, in terms of less aggressive system excitations and no
need for high-speed motions of the overall two-mass system,
are shown and discussed along with the experimental results.
54
R EFERENCES
53
52
51
50
49
48
47
46
0.15
0.16
0.17
0.18
0.19
0.2
0.21
0.22
0.23
Fig. 19. Backlash identification using the reference method, Case 4.
system excitations and generally larger load inertias to allow
for a free, correspondingly decoupled, load motion phase. At
the same time, higher and uncertain damping and Coulomb
friction values will lead to higher errors when estimating 2β
according to (23).
The backlash identification results for all four cases, i.e.,
Case 1 and Case 2 for the proposed method, and Case 3 and
Case 4 for the reference method, are summarized in Table II,
while the nominal backlash size was measured as 19.05 mrad.
TABLE II
BACKLASH IDENTIFICATION RESULTS SUMMARY
Case 1
Case 2
Case 3
Case 4
19.3 mrad
23.6 mrad
31.0 mrad
42.6 mrad
V. C ONCLUSIONS
A new approach for identifying the backlash in two-mass
systems is proposed, with the principal advantage of using
the motor-side sensing only, which is a common practice for
various machines and mechanisms. The method is based on the
delayed relay in feedback of the motor velocity, which allows
[1] M. Nordin and P.-O. Gutman, “Controlling mechanical systems with
backlash a survey,” Automatica, vol. 38, no. 10, pp. 1633–1649, 2002.
[2] J.-L. Dion, S. Le Moyne, G. Chevallier, and H. Sebbah, “Gear impacts
and idle gear noise: Experimental study and non-linear dynamic model,”
Mech. Syst. and Sign. Process., vol. 23, no. 8, pp. 2608–2628, 2009.
[3] T. Tjahjowidodo, F. Al-Bender, and H. Van Brussel, “Quantifying
chaotic responses of mechanical systems with backlash component,”
Mech. Syst. and Sign. Process., vol. 21, no. 2, pp. 973–993, 2007.
[4] G. Tao and P. V. Kokotovic, “Continuous-time adaptive control of
systems with unknown backlash,” IEEE Transactions on Automatic
Control, vol. 40, no. 6, pp. 1083–1087, 1995.
[5] T. Hägglund, “Automatic on-line estimation of backlash in control
loops,” Journal of Process Control, vol. 17, no. 6, pp. 489–499, 2007.
[6] G. Brandenburg, H. Unger, and A. Wagenpfeil, “Stability problems of a
speed controlled drive in an elastic system with backlash and corrective
measures by a load observer,” in International conference on electrical
machines, 1986, pp. 523–527.
[7] Y. Hori, H. Iseki, and K. Sugiura, “Basic consideration of vibration suppression and disturbance rejection control of multi-inertia system using
sflac (state feedback and load acceleration control),” IEEE Transactions
on Industry Applications, vol. 30, no. 4, pp. 889–896, 1994.
[8] J. C. Gerdes and V. Kumar, “An impact model of mechanical backlash
for control system analysis,” in American Control Conference, vol. 5,
1995, pp. 3311–3315.
[9] M. Ruderman, F. Hoffmann, and T. Bertram, “Modeling and identification of elastic robot joints with hysteresis and backlash,” IEEE Trans.
on Industrial Electronics, vol. 56, no. 10, pp. 3840–3847, 2009.
[10] T. K. Morimoto, E. W. Hawkes, and A. M. Okamura, “Design of a
compact actuation and control system for flexible medical robots,” IEEE
Robotics and Automation Letters, vol. 2, no. 3, pp. 1579–1585, 2017.
[11] D. Gebler and J. Holtz, “Identification and compensation of gear backlash without output position sensor in high-precision servo systems,”
in IEEE 24th Annual Conference of the Industrial Electronics Society
(IECON’98), 1998, pp. 662–666.
[12] S. Villwock and M. Pacas, “Time-domain identification method for
detecting mechanical backlash in electrical drives,” IEEE Trans. on
Industrial Electronics, vol. 56, no. 2, pp. 568–573, 2009.
[13] A. Lagerberg and B. Egardt, “Backlash estimation with application
to automotive powertrains,” IEEE Transactions on Control Systems
Technology, vol. 15, no. 3, pp. 483–493, 2007.
[14] S. Yamada and H. Fujimoto, “Proposal of high backdrivable control
using load-side encoder and backlash,” in IEEE 42nd Annual Conference
of Industrial Electronics Society (IECON2016), 2016, pp. 6429–6434.
[15] A. Tustin, “The effects of backlash and of speed-dependent friction on
the stability of closed-cycle control systems,” Journal of the Institution
of Electrical Engineers - Part IIA: Automatic Regulators and Servo
Mechanisms, vol. 94, pp. 143–151, 1947.
[16] R. Merzouki, J. Davila, L. Fridman, and J. Cadiou, “Backlash phenomenon observation and identification in electromechanical system,”
Control Engineering Practice, vol. 15, no. 4, pp. 447–457, 2007.
[17] T. Tjahjowidodo, F. Al-Bender, and H. Van Brussel, “Experimental
dynamic identification of backlash using skeleton methods,” Mech. Syst.
and Sign. Process., vol. 21, no. 2, pp. 959–972, 2007.
11
[18] A. Lichtsinder and P.-O. Gutman, “Closed-form sinusoidal-input describing function for the exact backlash model,” IFAC-PapersOnLine,
vol. 49, no. 18, pp. 422–427, 2016.
[19] M. Nordin, J. Galic’, and P.-O. Gutman, “New models for backlash
and gear play,” International journal of adaptive control and signal
processing, vol. 11, no. 1, pp. 49–63, 1997.
[20] R. S. Barbosa and J. T. Machado, “Describing function analysis of
systems with impacts and backlash,” Nonlinear Dynamics, vol. 29, no. 1,
pp. 235–250, 2002.
[21] M. Yang, S. Tang, J. Tan, and D. Xu, “Study of on-line backlash identification for pmsm servo system,” in IECON 38th Annual Conference
on IEEE Industrial Electronics Society (IECON), 2012, pp. 2036–2042.
[22] A. A. Andronov and S. Khajkin, Theory of oscillations. Princeton Univ
Press, 1949.
[23] J. F. Mullin and I. E. Jury, “A phase-plane approach to relay sampleddata feedback systems,” Transactions of the American Institute of
Electrical Engineers, Part II: Applications and Industry, vol. 77, no. 6,
pp. 517–524, 1959.
[24] C. Hang, K. Astrom, and Q. Wang, “Relay feedback auto-tuning of
process controllers a tutorial review,” Journal of process control, vol. 12,
no. 1, pp. 143–162, 2002.
[25] T. Mizuno, T. Adachi, M. Takasaki, and Y. Ishino, “Mass measurement
system using relay feedback with hysteresis,” Journal of System Design
and Dynamics, vol. 2, no. 1, pp. 188–196, 2008.
[26] Y. Han, C. Liu, and J. Wu, “Backlash identification for pmsm servo
system based on relay feedback,” Nonlinear Dynamics, vol. 84, no. 4,
pp. 2363–2375, 2016.
[27] K. Hunt and F. Crossley, “Coefficient of restitution interpreted as
damping in vibroimpact,” Journal of Applied Mechanics, Trans. ASME,
vol. 42, no. 2, pp. 440–445, 1975.
[28] H. M. Lankarani and P. E. Nikravesh, “Continuous contact force models
for impact analysis in multibody systems,” Nonlinear Dynamics, vol. 5,
no. 2, pp. 193–207, 1994.
[29] A. Visintin, Differential models of hysteresis. Springer, 1994.
[30] P. Krejci, Hysteresis, Convexity and Dissipation in Hyperbolic Equations. Tokyo: Gattötoscho, 1996.
[31] M. Ruderman and D. Rachinskii, “Use of Prandtl-Ishlinskii hysteresis
operators for Coulomb friction modeling with presliding,” in Journal of
Physics: Conference Series, vol. 811, no. 1, 2017, p. 012013.
[32] P. Rostalski, T. Besselmann, M. Barić, F. V. Belzen, and M. Morari, “A
hybrid approach to modelling, control and state estimation of mechanical
systems with backlash,” International Journal of Control, vol. 80, no. 11,
pp. 1729–1740, 2007.
[33] M. Ruderman and M. Iwasaki, “Observer of nonlinear friction dynamics for motion control,” IEEE Transactions on Industrial Electronics,
vol. 62, no. 9, pp. 5941–5949, 2015.
[34] K. J. Åström, “Oscillations in systems with relay feedback,” in Adaptive
Control, Filtering, and Signal Processing, 1995, pp. 1–25.
[35] J. M. Gonçalves, A. Megretski, and M. A. Dahleh, “Global stability
of relay feedback systems,” IEEE Transactions on Automatic Control,
vol. 46, no. 4, pp. 550–562, 2001.
[36] K. H. Johansson, A. Rantzer, and K. J. Åström, “Fast switches in relay
feedback systems,” Automatica, vol. 35, no. 4, pp. 539–552, 1999.
[37] A. Lichtsinder and P.-O. Gutman, “Limit cycle existence condition in
control systems with backlash and friction,” IFAC Proceedings Volumes,
vol. 43, no. 14, pp. 469–474, 2010.
[38] M. Ruderman, “Computationally efficient formulation of relay operator
for Preisach hysteresis modeling,” IEEE Transactions on Magnetics,
vol. 51, no. 12, pp. 1–4, 2015.
| 3cs.SY
|
Sequence-Length Requirement of Distance-Based Phylogeny
Reconstruction: Breaking the Polynomial Barrier∗
Sébastien Roch†
arXiv:0908.2061v1 [math.PR] 14 Aug 2009
February 22, 2018
Abstract
We introduce a new distance-based phylogeny reconstruction technique which provably achieves, at
sufficiently short branch lengths, a polylogarithmic sequence-length requirement—improving significantly
over previous polynomial bounds for distance-based methods. The technique is based on an averaging
procedure that implicitly reconstructs ancestral sequences.
In the same token, we extend previous results on phase transitions in phylogeny reconstruction to general
time-reversible models. More precisely, we show that in the so-called Kesten-Stigum zone (roughly, a region
of the parameter space where ancestral sequences are well approximated by “linear combinations” of the
observed sequences) sequences of length poly(log n) suffice for reconstruction when branch lengths are
discretized. Here n is the number of extant species.
Our results challenge, to some extent, the conventional wisdom that estimates of evolutionary distances
alone carry significantly less information about phylogenies than full sequence datasets.
Keywords: Phylogenetics, distance-based methods, phase transitions, reconstruction problem.
∗
The current manuscript is the full version with proofs of [Roc08]. In subsequent work [Roc09] the results stated here were improved
to logarithmic sequence length, thereby matching the best results for general methods.
†
Department of Mathematics, UCLA.
1
Introduction
The evolutionary history of a group of organisms is generally represented by a phylogenetic tree or phylogeny [Fel04, SS03]. The leaves of the tree represent the current species. Each branching indicates a speciation
event. Many of the most popular techniques for reconstructing phylogenies from molecular data, e.g. UPGMA,
Neighbor-Joining, and BIO-NJ [SS63, SN87, Gas97], are examples of what are known as distance-matrix methods. The main advantage of these methods is their speed, which stems from a straightforward approach: 1) the
estimation of a distance matrix from observed molecular sequences; and 2) the repeated agglomeration of the
closest clusters of species. Each entry of the distance matrix is an estimate of the evolutionary distance between
the corresponding pair of species, that is, roughly the time elapsed since their most recent common ancestor.
This estimate is typically obtained by comparing aligned homologous DNA sequences extracted from the extant
species—the basic insight being, the closer the species, the more similar their sequences. Most distance methods run in time polynomial in n, the number of leaves, and in k, the sequence length. This performance compares very favorably to that of the other two main classes of reconstruction methods, likelihood and parsimony
methods, which are known to be computationally intractable [GF82, DS86, Day87, MV05, CT06, Roc06].
The question we address in this paper is the following: Is there a price to pay for this speed and simplicity?
There are strong combinatorial [SHP88] and statistical [Fel04] reasons to believe that distance methods are
not as accurate as more elaborate reconstruction techniques, notably maximum likelihood estimation (MLE).
Indeed, in a typical instance of the phylogenetic reconstruction problem, we are given aligned DNA sequences
{(ξli )ki=1 }l∈L , one sequence for each leaf l ∈ L, from which we seek to infer the phylogeny on L. Generally, all
sites (ξli )l∈L , for i = 1, . . . , k, are assumed to be independent and identically distributed according to a Markov
model on a tree (see Section 1.1). For a subset W ⊆ L, we denote by µW the distribution of (ξli )l∈W under
this model. Through their use of the distance matrix, distance methods reduce the data to pairwise sequence
correlations, that is, they only use estimates of µ2 = {µW : W ⊆ L, |W | = 2}. In doing so, they seemingly
fail to take into account more subtle patterns in the data involving three or more species at a time. In contrast,
MLE for example outputs a model that maximizes the joint probability of all observed sequences. We call
methods that explicitly use the full dataset, such as MLE, holistic methods.
It is important to note that the issue is not one of consistency: when the sequence length tends to infinity,
the estimate provided by distance methods—just like MLE—typically converges to the correct phylogeny. In
particular, under mild assumptions, it suffices to know the pairwise site distributions µ2 to recover the topology
of the phylogeny [CH91, Cha96]. Rather the question is: how fast is this convergence? Or more precisely, how
should k scale as a function of n to guarantee a correct reconstruction with high probability? And are distance
methods significantly slower to converge than holistic methods? Although we do not give a complete answer
to these questions of practical interest here, we do provide strong evidence that some of the suspicions against
distance methods are based on a simplistic view of the distance matrix. In particular, we open up the surprising
possibility that distance methods actually exhibit optimal convergence rates.
Context. It is well-known that some of the most popular distance-matrix methods actually suffer from a
prohibitive sequence-length requirement [Att99, LC06]. Nevertheless, over the past decade, much progress
has been made in the design of fast-converging distance-matrix techniques, starting with the seminal work of
Erdös et al. [ESSW99a]. The key insight behind the algorithm in [ESSW99a], often dubbed the Short Quartet
Method (SQM), is that it discards long evolutionary distances, which are known to be statistically unreliable.
The algorithm works by first building subtrees of small diameter and, in a second stage, putting the pieces back
together. The SQM algorithm runs in polynomial time and guarantees the correct reconstruction with high
probability of any phylogeny (modulo reasonable assumptions) when k = poly(n). This is currently the best
known convergence rate for distance methods. (See also [DMR06, DHJ+ 06, Mos07, GMS08, DMR09b] for
faster-converging algorithms involving partial reconstruction of the phylogeny.)
2
Although little is known about the sequence-length requirement of MLE [SS99, SS02], recent results of
Mossel [Mos04], Daskalakis et al. [DMR06, DMR09a], and Mihaescu et al. [MHR09] on a conjecture of
Steel [Ste01] indicate that convergence rates as low as k = O(log n) can be achieved when the branch lengths
are sufficiently short and discretized, using insights from statistical physics. We briefly describe these results.
As mentioned above, the classical model of DNA sequence evolution is a Markov model on a tree that
is closely related to stochastic models used to study particle systems [Lig85, Geo88]. This type of model
undergoes a phase transition that has been extensively studied in probability theory and statistical physics: at
short branch lengths (in the binary symmetric case, up to 15% divergence per edge), in what is called the
reconstruction phase, good estimates of the ancestral sequences can be obtained from the observed sequences;
on the other hand, outside the reconstruction phase, very little information about ancestral states diffuses to
the leaves. See e.g. [EKPS00] and references therein. The new algorithms in [Mos04, DMR06, DMR09a,
MHR09] exploit this phenomenon by alternately 1) reconstructing a few levels of the tree using distancematrix techniques and 2) estimating distances between internal nodes by reconstructing ancestral sequences at
the newly uncovered nodes. The overall algorithm is not distance-based, however, as the ancestral sequence
reconstruction is performed using a complex function of the observed sequences named recursive majority.
The rate k = O(log n) achieved by these algorithms is known to be necessary in general. Moreover, the slower
rate k = poly(n) is in fact necessary for all methods—distance-based or holistic—outside the reconstruction
phase [Mos03]. In particular, note that distance methods are in some sense “optimal” outside the reconstruction
phase by the results of [ESSW99a].
Beyond the oracle view of the distance matrix. It is an outstanding open problem to determine whether distance methods can achieve k = O(log n) in the reconstruction phase1 . From previous work on fast-converging
distance methods, it is tempting to conjecture that k = poly(n) is the best one can hope for. Indeed, all previous
algorithms use the following “oracle view” of the distance matrix, as formalized by King et al. [KZZ03] and
Mossel [Mos07]. As mentioned above, the reliability of distance estimates depends on the true evolutionary
distances. From standard concentration inequalities, it follows that if leaves a and b are at distance τ (a, b), then
the usual distance estimate τ̂ (a, b) (see Section 1.1) satisfies:
if τ (a, b) < D + ε or τ̂ (a, b) < D + ε then |τ (a, b) − τ̂ (a, b)| < ε,
(1)
for ε, D such that k ∝ (1−e−ε )−2 e2D . Fix ε > 0 small and k poly(n). Let T be a complete binary tree with
log2 n levels. Imagine that the distance matrix is given by the following oracle: on input a pair of leaves (a, b)
the oracle returns an estimate τ̂ (a, b) which satisfies (1). Now, notice that for any tree T 0 which is identical to
T on the first log2 n/2 levels above the leaves, the oracle is allowed to return the same distance estimate as for
T . That is, we cannot distinguish T and T 0 in this model unless k = poly(n). (This argument can be made
more formal along the lines of [KZZ03].)
What the oracle model ignores is that, under the assumption that the sequences are generated by a Markov
model of evolution, the distance estimates (τ̂ (a, b))a,b∈[n] are in fact correlated random variables. More concretely, for leaves a, b, c, d, note that the joint distribution of (τ̂ (a, b), τ̂ (c, d)) depends in a nontrivial way
on the joint site distribution µW at W = {a, b, c, d}. In other words, even though the distance matrix is—
seemingly—only an estimate of the pairwise correlations µ2 , it actually contains some information about all
joint distributions. Note however that it is not immediately clear how to exploit this extra information or even
how useful it could be.
As it turns out, the correlation structure of the distance matrix is in fact very informative at short branch
lengths. More precisely, we introduce in this paper a new distance-based method with a convergence rate of
k = poly(log n) in the reconstruction phase (to be more accurate, in the so-called Kesten-Stigum phase; see
1
Mike Steel offers a 100$ reward for the solution of this problem.
3
below)—improving significantly over previous poly(n) results. Note that the oracle model allows only the
reconstruction of a o(1) fraction of the levels in that case. Our new algorithm involves a distance averaging
procedure that implicitly reconstructs ancestral sequences, thereby taking advantage of the phase transition
discussed above. We also obtain the first results on Steel’s conjecture beyond the simple symmetric models
studied by Daskalakis et al. [DMR06, DMR09a, MHR09] (the so-called CFN and Jukes-Cantor models). In
the next subsections, we introduce general definitions and state our results more formally. We also give an
overview of the proof.
Further related work. For further related work on efficient phylogenetic tree reconstruction, see [ESSW99b,
HNW99, CK01, Csu02].
1.1
Definitions
Phylogenies. We define phylogenies and evolutionary distances more formally.
Definition 1 (Phylogeny) A phylogeny is a rooted, edge-weighted, leaf-labeled tree T = (V, E, [n], ρ; τ )
where: V is the set of vertices; E is the set of edges; L = [n] = {0, . . . , n − 1} is the set of leaves; ρ is
the root; τ : E → (0, +∞) is a positive edge weight function. We further assume that all internal nodes in T
have degree 3 except for the root ρ which has degree 2. We let Yn be the set of all such phylogenies on n leaves
and we denote Y = {Yn }n≥1 .
Definition 2 (Tree Metric) For two leaves a, b ∈ [n], we denote by Path(a, b) the set of edges on the unique
path between a and b. A tree metric on a set [n] is a positive function d : [n] × [n] → (0, +∞) such that
there exists a tree T = (V, E) with leaf set [n] and an edge weight function w : E → (0, +∞) satisfying the
following: for all leaves a, b ∈ [n]
X
d(a, b) =
we .
e∈Path(a,b)
For convenience, we denote by (τ (a, b))a,b∈[n] the tree metric corresponding to the phylogeny T = (V, E, [n], ρ; τ ).
We extend τ (u, v) to all vertices u, v ∈ V in the obvious way.
Example 1 (Homogeneous Tree) For an integer h ≥ 0, we denote by T (h) = (V (h) , E (h) , L(h) , ρ(h) ; τ ) a
rooted phylogeny where T (h) is the h-level complete binary tree with arbitrary edge weight function τ and
(h)
L(h) = [2h ]. For 0 ≤ h0 ≤ h, we let Lh0 be the vertices on level h − h0 (from the root). In particular,
(h)
(h)
L0 = L(h) and Lh = {ρ(h) }. We let HY = {HYn }n≥1 be the set of all phylogenies with homogeneous
underlying trees.
Model of molecular sequence evolution. Phylogenies are reconstructed from molecular sequences extracted
from the observed species. The standard model of evolution for such sequences is a Markov model on a tree
(MMT).
Definition 3 (Markov Model on a Tree) Let Φ be a finite set of character states with ϕ = |Φ|. Typically
Φ = {+1, −1} or Φ = {A, G, C, T}. Let n ≥ 1 and let T = (V, E, [n], ρ) be a rooted tree with leaves
labeled in [n]. For each edge e ∈ E, we are given a ϕ × ϕ stochastic matrix M e = (Mije )i,j∈Φ , with fixed
stationary distribution π = (πi )i∈Φ . An MMT ({M e }e∈E , T ) associates a state σv in Φ to each vertex v in V
as follows: pick a state for the root ρ according to π; moving away from the root, choose a state for each vertex
v independently according to the distribution (Mσeu ,j )j∈Φ , with e = (u, v) where u is the parent of v.
4
The most common MMT used in phylogenetics is the so-called general time-reversible (GTR) model.
Definition 4 (GTR Model) Let Φ be a set of character states with ϕ = |Φ| and π be a distribution on Φ
satisfying πi > 0 for all i ∈ Φ. For n ≥ 1, let T = (V, E, [n], ρ; τ ) be a phylogeny. Let Q be a ϕ × ϕ rate
matrix, that is, Qij > 0 for all i 6= j and
X
Qij = 0,
j∈Φ
for all i ∈ Φ. Assume Q is reversible with respect to π, that is,
πi Qij = πj Qji ,
for all i, j ∈ Φ. The GTR model on T with rate matrix Q is an MMT on T = (V, E, [n], ρ) with transition
matrices M e = eτe Q , for all e ∈ E. By the reversibility assumption, Q has ϕ real eigenvalues
0 = Λ1 > Λ2 ≥ · · · ≥ Λϕ .
We normalize Q by fixing Λ2 = −1. We denote by Qϕ the set of all such rate matrices. We let Gn,ϕ = Yn ⊗ Qϕ
be the set of all ϕ-state GTR models on n leaves. We denote Gϕ = {Gn,ϕ }n≥1 . We denote by ξW the vector of
states on the vertices W ⊆ V . In particular, ξ[n] are the states at the leaves. We denote by LT ,Q the distribution
of ξ[n] .
GTR models include as special cases many popular models such as the CFN model.
Example 2 (CFN Model) The CFN model is the GTR model with ϕ = 2, π = (1/2, 1/2), and
−1/2 1/2
CFN
.
Q=Q
≡
1/2 −1/2
Example 3 (Binary Asymmetric Channel) More generally, letting Φ = {+, −} and π = (π+ , π− ), with
π+ , π− > 0, we can take
−π− π−
Q=
.
π+ −π+
Phylogenetic reconstruction. A standard assumption in molecular evolution is that each site in a sequence
(DNA, protein, etc.) evolves independently according to a Markov model on a tree, such as the GTR model
above. Because of the reversibility assumption, the root of the phylogeny cannot be identified and we reconstruct phylogenies up to their root.
e = {Y
e n }n≥1 be a subset of phylogenies and Q
eϕ
Definition 5 (Phylogenetic Reconstruction Problem) Let Y
e
be a subset of rate matrices on ϕ states. Let T = (V, E, [n], ρ; τ ) ∈ Y. If T = (V, E, [n], ρ) is the rooted
tree underlying T , we denote by T− [T ] the tree T where the root is removed: that is, we replace the two
edges adjacent to the root by a single edge. We denote by Tn the set of all leaf-labeled trees on n leaves
with internal degrees 3 and we let T = {Tn }n≥1 . A phylogenetic reconstruction algorithm is a collection of
i )k
[n] k
maps A = {An,k }n,k≥1 from sequences (ξ[n]
i=1 ∈ (Φ ) to leaf-labeled trees T ∈ Tn . We only consider
algorithms A computable in time polynomial in n and k. Let k(n) be an increasing function of n. We say that
e ⊗Q
e ϕ with sequence length k = k(n) if for all δ > 0,
A solves the phylogenetic reconstruction problem on Y
e n, Q ∈ Q
e ϕ,
there is n0 ≥ 1 such that for all n ≥ n0 , T ∈ Y
h
i
i k(n)
P An,k(n) (ξ[n]
)i=1 = T− [T ] ≥ 1 − δ,
k(n)
i )
where (ξ[n]
i=1 are i.i.d. samples from LT ,Q .
5
An important result of this kind was given by Erdos et al. [ESSW99a].
Theorem 1 (Polynomial Reconstruction [ESSW99a]) Let 0 < f ≤ g < +∞ and denote by Yf,g the set
of all phylogenies T = (V, E, [n], ρ; τ ) satisfying f ≤ τe ≤ g, ∀e ∈ E. Then, for all ϕ ≥ 2 and all
0 < f ≤ g < +∞, the phylogenetic reconstruction problem on Yf,g ⊗ Qϕ can be solved with k = poly(n).
This result was recently improved by Daskalakis et al. [DMR06,
DMR09a] (see also [MHR09]) in the so-called
√
Kesten-Stigum reconstruction phase, that is, when g < ln 2.
Definition 6 (∆-Branch Model) Let 0 < ∆ ≤ f ≤ g < +∞ and denote by Yf,g
∆ the set of all phylogenies
T = (V, E, [n], ρ; τ ) satisfying f ≤ τe ≤ g where τe is an integer multiple of ∆, for all e ∈ E. For ϕ ≥ 2 and
Q ∈ Qϕ , we call Yf,g
∆ ⊗ {Q} the ∆-Branch Model (∆-BM).
√
Theorem 2 (Logarithmic Reconstruction [DMR06, DMR09a]. See also [MHR09].) Let g ∗ = ln 2. Then,
CFN } can be solved with
for all 0 < ∆ ≤ f ≤ g < g ∗ , the phylogenetic reconstruction problem on Yf,g
∆ ⊗ {Q
k = O(log n)2 .
Distance methods. The proof of Theorem 1 uses distance methods, which we now define formally.
Definition 7 (Correlation Matrix) Let Φ be a finite set with ϕ ≥ 2. Let (ξai )ki=1 , (ξbi )ki=1 ∈ Φk be the sequences at a, b ∈ [n]. For υ1 , υ2 ∈ Φ, we define the correlation matrix between a and b by
k
1X
ab
b
Fυ1 υ2 =
1{ξai = υ1 , ξbi = υ2 },
k
i=1
and Fbab = (Fbυab1 υ2 )υ1 ,υ2 ∈Φ .
Definition 8 (Distance Method) A phylogenetic reconstruction algorithm A = {An,k }n,k≥1 is said to be
i )k
[n] k
bab
distance-based if A depends on the data (ξ[n]
i=1 ∈ (Φ ) only through the correlation matrices {F }a,b∈[n] .
The previous definition takes a very general view of distance-based methods: any method that uses only pairwise sequence comparisons. In practice, most distance-based approaches actually use a specific distance estimator, that is, a function of Fbab that converges to τ (a, b) in probability as n → +∞. We give two classical
examples below.
Example 4 (CFN Metric) In the CFN case with state space Φ = {+, −}, a standard distance estimator (up
to a constant) is
D(Fb) = − ln 1 − 2(Fb+− + Fb−+ ) .
Example 5 (Log-Det Distance [BH87, Lak94, LSHP94, Ste94]) More generally, a common distance estimator (up to scaling) is the so-called log-det distance
D(Fb) = − ln | det Fb|.
Loosely speaking, the log-det distance can be thought as a generalization of the CFN metric. We will use a
different generalization of the CFN metric. See section 1.3.
2
The correct statement of this result appears in [DMR09a]. Because of different conventions, our edge weights are scaled by a factor
of 2 compared to those in [DMR09a]. The dependence of k in ∆ is ∆−2 .
6
1.2
Results
In our main result, we prove that phylogenies under GTR models of mutation can be inferred using a distancebased method from k = poly(log n).
Theorem 3 (Main Result) For all ϕ ≥ 2, 0 < ∆ ≤ f ≤ g < g ∗ and Q ∈ Qϕ , there is a distance-based
3
method solving the phylogenetic reconstruction problem on Yf,g
∆ ⊗ {Q} with k = poly(log n).
Note that this result is a substantial improvement over Theorem 1—at least, in a certain range of parameters—
and that it almost matches the bound obtained in Theorem 2. The result is also novel in two ways over Theorem 2: only the distance matrix is used; the result applies to a larger class of mutation matrices. A slightly
weaker version of the result stated here appeared without proof as [Roc08]. Note that in [Roc08] the result was
stated without the discretization assumption which is in fact needed for the final step of the proof. This is further
explained in Section 7.3 of [DMR09a]. In subsequent work [Roc09], the result stated here was improved to
logarithmic sequence length, thereby matching Theorem 2. This new result follows a similar high-level proof
but involves stronger concentration arguments [PR09], as well as a simplified algorithm.
In an attempt to keep the paper as self-contained as possible we first give a proof in the special case of
homogeneous trees. This allows to keep the algorithmic details to a minimum. The proof appears in Section 3.
We extend the result to general trees in Appendix A. The more general result relies on a combinatorial algorithm
of [DMR09a].
1.3
Proof Overview
Distance averaging. The basic insight behind Steel’s conjecture is that the accurate reconstruction of ancestral sequences in the reconstruction phase can be harnessed to perform a better reconstruction of the phylogeny
itself. For now, consider the CFN model with character space {+1, −1} and assume that our phylogeny is homogeneous with uniform branch lengths ω. Generate k i.i.d. samples (σVi )ki=1 . Let a, b be two internal vertices
on level h − h0 < h (from the root). Suppose we seek to estimate the distance between a and b. This estimation
cannot be performed directly because the sequences at a and b are not known. However, we can try to estimate
these internal sequences. Denote by A, B the leaf set below a and b respectively. An estimate of the sequence
at a is the (properly normalized) “site-wise average” of the sequences at A
σ̄ai =
1 X σai 0
,
|A| 0 e−ωh0
(2)
a ∈A
for i = 1, . . . , k, and similarly for b. It is not immediately clear how such a site-wise procedure involving simultaneously a large number of leaves can be performed using the more aggregated information in the correlation
matrices {Fbuv }u,v∈[n] . Nevertheless, note that the quantity we are ultimately interested in computing is the
following estimate of the CFN metric between a and b
!
k
1X i i
σ̄a σ̄b .
τ̄ (a, b) = − ln
k
i=1
3
As in Theorem 2, the dependence of k in ∆ is ∆−2 .
7
Our results are based on the following observation:
τ̄ (a, b) = − ln
k
1X
k
i=1
= − ln
1 X σai 0
|A| 0 e−ωh0
a ∈A
X X
1
0
|A||B|e−2ωh 0
0
a ∈A b ∈B
= − ln
!
1 X σbi 0
|B| 0 e−ωh0
b ∈B
!!
k
1X i i
σa0 σb0
k
i=1
!
X X
1
0 0
e−τ̂ (a ,b )
0
−2ωh
|A||B|e
0
0
!!
,
a ∈A b ∈B
where note that the last line depends only on distance estimates τ̂ (a0 , b0 ) between leaves a0 , b0 in A, B. In other
words, through this procedure, which we call exponential averaging, we perform an implicit ancestral sequence
reconstruction using only distance estimates. One can also think of this as a variance reduction technique. When
the branch lengths are not uniform, one needs to use a weighted version of (2). This requires the estimation of
path lengths.
GTR models. In the case of GTR models, the standard log-det estimator does not lend itself well to the
exponential averaging procedure described above. Instead, we use an estimator involving the right eigenvector
ν corresponding to the second eigenvalue Λ2 of Q. For a, b ∈ [n], we consider the estimator
τ̂ (a, b) = − ln ν > Fbab ν .
(3)
This choice is justified by a generalization of (2) introduced in [MP03]. Note that ν may need to be estimated.
Concentration. There is a further complication in that to obtain results with high probability, one needs to
show that τ̄ (a, b) is highly concentrated. However, one cannot directly apply standard concentration inequalities
because σ̄a is not bounded. Classical results on the reconstruction problem imply that the variance of σ̄a
is finite—which is not quite enough. To amplify the accuracy, we “go down” log log n levels and compute
O(log n) distance estimates with conditionally independent biases. By performing a majority procedure, we
finally obtain a concentrated estimate.
1.4
Organization
In Section 2, we provide a detailed account of the connection between ancestral sequence reconstruction and
distance averaging. We then give a proof of our main result in the case of homogeneous trees in Section 3.
Finally, in Appendix A, we give a sketch of the proof in the general case.
All proofs are relegated to Appendix B.
2
Ancestral Reconstruction and Distance Averaging
√
Let ϕ ≥ 2, 0 < ∆ ≤ f ≤ g < g ∗ = ln 2, and Q ∈ Qϕ with corresponding stationary distribution
π > 0. In this section we restrict ourselves to the homogeneous case T = T (h) = (V, E, [n], ρ; τ ) where we
take h = log2 n and f ≤ τe ≤ g and τe is an integer multiple of ∆, ∀e ∈ E. (See Examples 1 and 2 and
Theorem 2.)4
4
Note that, without loss of generality, we can consider performing ancestral state reconstruction on a homogeneous tree as it is
always possible to “complete” a general tree with zero-length edges. We come back to this point in Appendix A.
8
Throughout this section, we use a sequence length k > logκ n where κ > 1 is a constant to be determined
later. We generate k i.i.d. samples (ξVi )ki=1 from the GTR model (T , Q) with state space Φ.
All proofs are relegated to Appendix B.
2.1
Distance Estimator
The standard log-det estimator does not lend itself well to the averaging procedure discussed above. For reconstruction purposes, we instead use an estimator involving the right eigenvector ν corresponding to the second
eigenvalue Λ2 of Q. For a, b ∈ [n], consider the estimator
τ̂ (a, b) = − ln ν > Fbab ν ,
(4)
where the correlation matrix Fbab was introduced in Definition 7. We first give a proof that this is indeed a
legitimate distance estimator. For more on connections between eigenvalues of the rate matrix and distance
estimation, see e.g. [GL96, GL98, GMY09].
Lemma 1 (Distance Estimator) Let τ̂ be as above. For all a, b ∈ [n], we have
E[e−τ̂ (a,b) ] = e−τ (a,b) .
For a ∈ [n] and i = 1, . . . , k, let
σai = νξai .
Then (4) is equivalent to
k
1X i i
σa σb
k
τ̂ (a, b) = − ln
!
.
(5)
i=1
Note that in the CFN case, we have simply ν =
of the CFN metric.
Let
(1, −1)>
and hence (5) can be interpreted as a generalization
π̄ = min πι ,
ι
and
1
ν̄ ≡ max |νi | ≤ √ .
i
π̄
The following lemmas show that the distance estimate above with sequence length poly(log n) is concentrated
for path lengths of order O(log log n).
Lemma 2 (Distorted Metric: Short Distances) Let δ > 0, ε > 0, and γ > 0. There exists κ > 1, such that if
the following conditions hold for u, v ∈ V :
• [Small Diameter] τ (u, v) < δ log log(n),
• [Sequence Length] k > logκ (n),
then
|τ (u, v) − τ̂ (u, v)| < ε,
with probability at least 1 − n−γ .
9
Lemma 3 (Distorted Metric: Diameter Test) Let D > 0, W > 5, δ > 0, and γ > 0. Let u, v ∈ V not
descendants of each other. Let u0 , v0 ∈ V be descendants of u, v respectively. there exists κ > 1, such that if
the following conditions hold:
• [Large Diameter] τ (u0 , v0 ) > D + ln W ,
• [Close Descendants] τ (u0 , u), τ (v0 , v) < δ log log(n),
• [Sequence Length] k > logκ (n),
then
W
,
2
with probability at least 1 − n−γ . On the other hand, if the first condition above is replaced by
τ̂ (u, v) − τ (u0 , u) − τ (v0 , v) > D + ln
• [Small Diameter] τ (u0 , v0 ) < D + ln W
5 ,
then
τ̂ (u, v) − τ (u0 , u) − τ (v0 , v) ≤ D + ln
W
,
4
with probability at least 1 − n−γ .
2.2
Ancestral Sequence Reconstruction
Let e = (x, y) ∈ E and assume that x is closest to ρ (in topological distance). We define Path(ρ, e) =
Path(ρ, y), |e|ρ = |Path(v, e)|, and
Rρ (e) = 1 − θe2 Θ−1
ρ,y ,
where Θρ,y = e−τ (ρ,y) and θe = e−τ (e) .
Proposition 1 below is a variant of Lemma 5.3 in [MP03]. For completeness, we give a proof.
Proposition 1 (Weighted Majority: GTR Version) Let ξ[n] be a sample from LT ,Q (see Definition 4) with
corresponding σ[n] . For a unit flow Ψ from ρ to [n], consider the estimator
S=
X Ψ(x)σx
.
Θρ,x
x∈[n]
Then, we have
E[S] = 0,
E[S | σρ ] = σρ ,
and
Var[S] = 1 + KΨ ,
where
KΨ =
X
Rρ (e)Ψ(e)2 .
e∈E
10
Let Ψ be a unit flow from ρ to [n]. We will use the following multiplicative decomposition of Ψ: If
Ψ(x) > 0, we let
Ψ(y)
ψ(e) =
,
Ψ(x)
and, if instead Ψ(x) = 0, we let ψ(y) = 0. Denoting x↑ the immediate ancestor of x ∈ V and letting
−τ
θx = e (x↑ ,x) , it will be useful to re-write
KΨ =
h−1 X
X
Y
(1 − θx2 )
h0 =0 x∈L(h)
e∈Path(ρ,x)
h0
ψ(e)2
,
θe2
(6)
and to define the following recursion from the leaves. For x ∈ [n],
Kx,Ψ = 0.
Then, let u ∈ V − [n] with children v1 , v2 with corresponding edges e1 , e2 and define
X
ψ(eα )2
.
Ku,Ψ =
((1 − θv2α ) + Kvα ,Ψ )
θe2α
α=1,2
Note that, from (6), we have Kρ,Ψ = KΨ .
Lemma 4 (Uniform Bound on KΨ ) Let Ψ be the homogeneous flow from ρ to [n]. Then, we have
KΨ ≤
2.3
1
< +∞.
1 − e−2(g∗ −g)
Distance Averaging
The input to our tree reconstruction algorithm is the matrix of all estimated distances between pairs of leaves
{τ̂ (a, b)}a,b,∈[n] . For short sequences, these estimated distances are known to be accurate for leaves that are
close enough. We now show how to compute distances between internal nodes in a way that involves only
{τ̂ (a, b)}a,b,∈[n] (and previously computed internal weights) using Proposition 1. However, the second-moment
guarantee in Proposition 1 is not enough to obtain good estimates with high probability. To remedy this situation, we perform a large number of conditionally independent distance computations, as we now describe.
Let α > 1 and assume h0 > bα log2 log2 nc. Let 0 ≤ h00 < h0 such that ∆h ≡ h0 − h00 = bα log2 log2 nc.
(h)
(h)
Let a0 , b0 ∈ Lh0 . For x ∈ {a, b}, denote by x1 , . . . , x2∆h , the vertices in Lh00 that are below x0 and, for
j = 1, . . . , 2∆h , let Xj be the leaves of T (h) below xj . See Figure 1. Assume that we are given Θa0 ,· , Θb0 ,· .
For 1 ≤ j ≤ 2∆h , we estimate τ (a0 , b0 ) as follows
X
X
1
−1 −τ̂ (a0 ,b0 )
τ̄ (aj , bj ) ≡ − ln
Θ−1
.
a0 ,a0 Θb0 ,b0 e
|Aj ||Bj | 0
0
a ∈Aj b ∈Bj
This choice of estimator is suggested by the following observation
X X
00
−1 −τ̂ (a0 ,b0 )
e−τ̄ (aj ,bj ) ≡
2−2h Θ−1
a0 ,a0 Θb0 ,b0 e
a0 ∈Aj b0 ∈Bj
00 i
00 i
k
−h
−h
X
X
X
2
σa0
2
σb0
1
= e−τ (a0 ,aj )−τ (b0 ,bj )
.
k
Θaj ,a0
Θbj ,b0
0
0
i=1
11
a ∈Aj
b ∈Bj
bα log2 log2 nc
aj
a1
···
A1
b0
a0
···
Aj
bj
b1
a2∆h
···
B1
A2∆h
b2∆h
···
Bj
B2∆h
Figure 1: Accuracy amplification.
τ (a0 , b0 )
rj∗
j
j 0 r∗0
j
Figure 2: Examples of dense balls with six points. The estimate in this case would be one of the central points.
12
Note that the first line depends only on estimates (τ̂ (u, v))u,v∈[n] and {Θv,· }v∈Va0 ∪Vb0 . The last line is the
empirical distance between the reconstructed states at a and b when the flow is chosen to be uniform in Proposition 1.
For d ∈ R and r > 0, let Br (d) be the ball of radius r around d. We define the dense ball around j to be
∆h
the smallest ball around τ̄ (aj , bj ) containing at least 2/3 of J = {τ̄ (aj 0 , bj 0 )}j20 =1 (as a multiset). The radius
of the dense ball around j is
2 ∆h
∗
rj = inf r : |Br (τ̄ (aj , bj )) ∩ J | ≥ (2 ) ,
3
for j = 1, . . . , 2∆h . We define our estimate of τ (a0 , b0 ) to be τ̄ 0 (a0 , b0 ) = τ̄ (aj ∗ , bj ∗ ), where j ∗ = arg minj rj∗ .
See Figure 2. For D > 0, W > 5, we define
W
1
−∆h
j : τ̄ (aj , bj ) ≤ D + ln
>
.
SD(a0 , b0 ) = 1 2
3
2
We extend Lemmas 2 and 3 to τ̄ 0 (a0 , b0 ) and SD(a0 , b0 ).
Proposition 2 (Deep Distance Computation: Small Diameter) Let α > 1, D > 0, γ > 0, and ε > 0. Let
(h)
a0 , b0 ∈ Lh0 as above. There exist κ > 1 such that if the following conditions hold:
• [Small Diameter] τ (a0 , b0 ) < D,
• [Sequence Length] k > logκ (n),
then
|τ̄ 0 (a0 , b0 ) − τ (a0 , b0 )| < ε,
with probability at least 1 − O(n−γ ).
Proposition 3 (Deep Distance Computation: Diameter Test) Let α > 1, D > 0, W > 5, and γ > 0. Let
(h)
a0 , b0 ∈ Lh0 as above. There exists κ > 1 such that if the following conditions hold:
• [Large Diameter] τ (a0 , b0 ) > D + ln W ,
• [Sequence Length] k > logκ (n),
then
SD(a0 , b0 ) = 0,
with probability at least 1 − n−γ . On the other hand, if the first condition above is replaced by
• [Small Diameter] τ (a0 , b0 ) < D + ln W
5 ,
then
SD(a0 , b0 ) = 1,
with probability at least 1 − n−γ .
13
3
Reconstructing Homogeneous Trees
In this section, we prove our main result in the case of homogeneous trees. More precisely, we prove the
following.
Theorem 4 (Main Result: Homogeneous Case) Let 0 < ∆ ≤ f ≤ g < +∞ and denote by HYf,g
∆ the set of
all homogeneous phylogenies
T
=
(V,
E,
[n],
ρ;
τ
)
satisfying
f
≤
τ
≤
g
and
τ
is
an
integer
multiple
of ∆,
e
e
√
∀e ∈ E. Let g ∗ = ln 2. Then, for all ϕ ≥ 2, 0 < ∆ ≤ f ≤ g < g ∗ and Q ∈ Qϕ , there is a distance-based
method solving the phylogenetic reconstruction problem on HYf,g
∆ ⊗ {Q} with k = poly(log n).
All proofs are relegated to Appendix B.
In the homogeneous case, we can build the tree level by level using simple “four-point” techniques [Bun71].
See e.g. [SS03, Fel04] for background and details. See also Section 3.2 below. The underlying combinatorial
algorithm we use here is essentially identical to the one used by Mossel in [Mos04]. From Propositions 2 and 3,
we get that the “local metric” on each level is accurate as long as we compute adequate weights. We summarize
this fact in the next proposition. For ∆ > 0 and z ∈ R+ , we let [z]∆ be the closest multiple of ∆ to z (breaking
ties arbitrarily). We define
0
[τ̄ (a0 , b0 )]∆ , if SD(a0 , b0 ) = 1,
d(a0 , b0 ) =
+∞,
o.w.
Proposition 4 (Deep Distorted Metric) Let D > 0, W > 5, and γ > 0. Let T = (V, E, [n], ρ; τ ) ∈ HYf,g
∆
(h)
with g < g ∗ . Let a0 , b0 ∈ Lh0 for 0 ≤ h0 < h. Assume we are given, for x = a, b, θe for all e ∈ Vx . There
exists κ > 0, such that if the following condition holds:
• [Sequence Length] The sequence length is k > logκ (n),
then we have, with probability at least 1 − O(n−γ ),
d(a0 , b0 ) = τ (a0 , b0 )
under either of the following two conditions:
1. [Small Diameter] τ (a0 , b0 ) < D, or
2. [Finite Estimate] d(a0 , b0 ) < +∞.
It remains to show how to compute the weights, which is the purpose of the next section.
3.1
Estimating Averaging Weights
Proposition 4 relies on the prior computation of the weights θe for all e ∈ Vx , for x = a, b. In this section, we
show how this estimation is performed.
(h)
Let a0 , b0 , c0 ∈ Lh0 . Denote by z the meeting point of the paths joining a0 , b0 , c0 . We define the “threepoint” estimate
1
θ̂z,a0 = O(a0 ; b0 , c0 ) ≡ exp − [d(a0 , b0 ) + d(a0 , c0 ) − d(b0 , c0 )] .
2
Note that the expression in parenthesis is an estimate of the distance between a0 and z.
14
(h)
Proposition 5 (Averaging Weight Estimation) Let a0 , b0 , c0 ∈ Lh0 as above. Assume that the assumptions
of Propositions 2, 3, 4 hold. Assume further that the following condition hold:
• [Small Diameter] τ (a0 , b0 ), τ (a0 , c0 ), τ (b0 , c0 ) < D + ln W ,
then
θ̂z,a0 = θz,a0 ,
with probability at least 1 − O(n−γ ) where θ̂z,a0 = O(a0 ; b0 , c0 ).
3.2
Putting it All Together
(h)
Let 0 ≤ h0 < h and Q = {a0 , b0 , c0 , d0 } ⊆ Lh0 . The topology of T (h) restricted to Q is completely
characterized by a bi-partition or quartet split q of the form: a0 b0 |c0 d0 , a0 c0 |b0 d0 or a0 d0 |b0 c0 . The most basic
operation in quartet-based reconstruction algorithms is the inference of such quartet splits. In distance-based
methods in particular, this is usually done by performing the so-called four-point test: letting
1
F(a0 b0 |c0 d0 ) = [τ (a0 , c0 ) + τ (b0 , d0 ) − τ (a0 , b0 ) − τ (c0 , d0 )],
2
we have
a0 b0 |c0 d0 if F(a0 , b0 |c0 , d0 ) > 0
a0 c0 |b0 d0 if F(a0 , b0 |c0 , d0 ) < 0
q=
a0 d0 |b0 c0 o.w.
Of course, we cannot compute F(a0 , b0 |c0 , d0 ) directly unless h0 = 0. Instead we use Proposition 4.
Deep Four-Point Test. Assume we have previously computed weights θe for all e ∈ Vx , for x = a, b, c, d.
We let
1
F(a0 b0 |c0 d0 ) = [d(a0 , c0 ) + d(b0 , d0 ) − d(a0 , b0 ) − d(c0 , d0 )],
(7)
2
and we define the deep four-point test
FP(a0 , b0 |c0 , d0 ) = 1{F(a0 b0 |c0 d0 ) > f /2},
with FP(a0 , b0 |c0 , d0 ) = 0 if any of the distances in (7) is infinite. Also, we extend the diameter test SD to
arbitrary subsets by letting SD(S) = 1 if and only if SD(x, y) = 1 for all pairs x, y ∈ S.
Algorithm. Fix α > 1, D > 4g, W > 5, γ > 3. Choose κ so as to satisfy Propositions 4 and 5. Let Z0 be
the set of leaves. The algorithm—a standard cherry picking algorithm—is detailed in Figure 3.
Acknowledgments
This work was triggered by a discussion with Elchanan Mossel on lower bounds for distance methods, following
a talk of Joseph Felsenstein. In particular, Elchanan pointed out that the distance matrix has a potentially useful
correlation structure.
15
Algorithm
Input: Distance estimates {τ̂ (a, b)}a,b∈[n] ;
Output: Tree;
• For h0 = 1, . . . , h − 1,
1. Four-Point Test. Let
Rh0 = {q = ab|cd : ∀a, b, c, d ∈ Zh0 distinct such that FP(q) = 1}.
2. Cherry Picking. Identify the cherries in Rh0 , that is, those pairs of vertices that only appear on the same
side of the quartet splits in Rh0 . Let
(h0 +1)
Zh0 +1 = {a1
(h0 +1)
, . . . , a2h−(h0 +1) },
be the parents of the cherries in Zh0
3. Weight Estimation. For all z ∈ Zh0 +1 ,
(a) Let x, y be the children of z. Choose w to be any other vertex in Zh0 with SD({x, y, w}) = 1.
(b) Compute
θ̂z,x = O(x; y, w).
(c) Repeat the previous step interchanging the role of x and y.
Figure 3: Algorithm.
References
[Att99]
K. Atteson. The performance of neighbor-joining methods of phylogenetic reconstruction. Algorithmica, 25(2-3):251–278, 1999.
[BH87]
Daniel Barry and J. A. Hartigan. Statistical analysis of hominoid molecular evolution. Statist.
Sci., 2(2):191–210, 1987. With comments by Stephen Portnoy and Joseph Felsenstein and a reply
by the authors.
[Bun71]
P. Buneman. The recovery of trees from measures of dissimilarity. In Mathematics in the Archaelogical and Historical Sciences, pages 187–395. Edinburgh University Press, Edinburgh, 1971.
[CH91]
Joseph T. Chang and John A. Hartigan. Reconstruction of evolutionary trees from pairwise distributions on current species, 1991.
[Cha96]
Joseph T. Chang. Full reconstruction of Markov models on evolutionary trees: identifiability and
consistency. Math. Biosci., 137(1):51–73, 1996.
[CK01]
Miklós Csurös and Ming-Yang Kao. Provably fast and accurate recovery of evolutionary trees
through harmonic greedy triplets. SIAM Journal on Computing, 31(1):306–322, 2001.
[Csu02]
M. Csurös. Fast recovery of evolutionary trees with thousands of nodes. J. Comput. Biol.,
9(2):277–97, 2002.
[CT06]
Benny Chor and Tamir Tuller. Finding a maximum likelihood tree is hard. J. ACM, 53(5):722–744,
2006.
[Day87]
William H. E. Day. Computational complexity of inferring phylogenies from dissimilarity matrices. Bull. Math. Biol., 49(4):461–467, 1987.
16
[DHJ+ 06]
Constantinos Daskalakis, Cameron Hill, Alexander Jaffe, Radu Mihaescu, Elchanan Mossel, and
Satish Rao. Maximal accurate forests from distance matrices. In RECOMB, pages 281–295, 2006.
[DMR06]
Constantinos Daskalakis, Elchanan Mossel, and Sébastien Roch. Optimal phylogenetic reconstruction. In STOC’06: Proceedings of the 38th Annual ACM Symposium on Theory of Computing, pages 159–168, New York, 2006. ACM.
[DMR09a] Constantinos Daskalakis, Elchanan Mossel, and Sébastien Roch. Evolutionaty trees and the Ising
model on the Bethe lattice: a proof of Steel’s conjecture. Preprint, 2009.
[DMR09b] Constantinos Daskalakis, Elchanan Mossel, and Sébastien Roch. Phylogenies without branch
bounds: Contracting the short, pruning the deep. In RECOMB, pages 451–465, 2009.
[DS86]
William H. E. Day and David Sankoff. Computational complexity of inferring phylogenies by
compatibility. Syst. Zool., 35(2):224–229, 1986.
[EKPS00]
W. S. Evans, C. Kenyon, Y. Peres, and L. J. Schulman. Broadcasting on trees and the Ising model.
Ann. Appl. Probab., 10(2):410–433, 2000.
[ESSW99a] P. L. Erdös, M. A. Steel, L. A. Székely, and T. A. Warnow. A few logs suffice to build (almost) all
trees (part 1). Random Struct. Algor., 14(2):153–184, 1999.
[ESSW99b] P. L. Erdös, M. A. Steel, L. A. Székely, and T. A. Warnow. A few logs suffice to build (almost) all
trees (part 2). Theor. Comput. Sci., 221:77–118, 1999.
[Fel04]
J. Felsenstein. Inferring Phylogenies. Sinauer, Sunderland, MA, 2004.
[Gas97]
O. Gascuel. BIO-NJ: An improved version of the NJ algorithm based on a simple model of
sequence data. Mol. Biol. Evol., 14(7):685–695, 1997.
[Geo88]
H. O. Georgii. Gibbs measures and phase transitions, volume 9 of de Gruyter Studies in Mathematics. Walter de Gruyter & Co., Berlin, 1988.
[GF82]
R. L. Graham. and L. R. Foulds. Unlikelihood that minimal phylogenies for a realistic biological
study can be constructed in reasonable computational time. Math. Biosci., 60:133–142, 1982.
[GL96]
X Gu and W H Li. A general additive distance with time-reversibility and rate variation among
nucleotide sites. Proceedings of the National Academy of Sciences of the United States of America,
93(10):4671–4676, 1996.
[GL98]
Xun Gu and Wen-Hsiung Li. Estimation of evolutionary distances under stationary and nonstationary models of nucleotide substitution. Proceedings of the National Academy of Sciences of
the United States of America, 95(11):5899–5905, 1998.
[GMS08]
Ilan Gronau, Shlomo Moran, and Sagi Snir. Fast and reliable reconstruction of phylogenetic trees
with very short edges. In SODA ’08: Proceedings of the nineteenth annual ACM-SIAM symposium
on Discrete algorithms, pages 379–388, Philadelphia, PA, USA, 2008. Society for Industrial and
Applied Mathematics.
[GMY09]
Ilan Gronau, Shlomo Moran, and Irad Yavneh. Towards optimal distance functions for stochastic
substitutions models. Preprint, 2009.
[HNW99]
D. H. Huson, S. H. Nettles, and T. J. Warnow. Disk-covering, a fast-converging method for
phylogenetic tree reconstruction. J. Comput. Biol., 6(3–4), 1999.
17
[KZZ03]
Valerie King, Li Zhang, and Yunhong Zhou. On the complexity of distance-based evolutionary
tree reconstruction. In SODA ’03: Proceedings of the fourteenth annual ACM-SIAM symposium
on Discrete algorithms, pages 444–453, Philadelphia, PA, USA, 2003. Society for Industrial and
Applied Mathematics.
[Lak94]
JA Lake. Reconstructing Evolutionary Trees from DNA and Protein Sequences: Paralinear Distances. Proceedings of the National Academy of Sciences, 91(4):1455–1459, 1994.
[LC06]
Michelle R. Lacey and Joseph T. Chang. A signal-to-noise analysis of phylogeny estimation by
neighbor-joining: insufficiency of polynomial length sequences. Math. Biosci., 199(2):188–215,
2006.
[Lig85]
Thomas M. Liggett. Interacting particle systems, volume 276 of Grundlehren der Mathematischen
Wissenschaften [Fundamental Principles of Mathematical Sciences]. Springer-Verlag, New York,
1985.
[LSHP94]
PJ Lockhart, MA Steel, MD Hendy, and D Penny. Recovering Evolutionary Trees under a More
Realistic Model of Sequence. Mol Biol Evol, 11(4):605–612, 1994.
[MHR09]
R. Mihaescu, C. Hill, and S. Rao. Fast phylogeny reconstruction through learning of ancestral
sequences. Preprint, 2009.
[Mos03]
E. Mossel. On the impossibility of reconstructing ancestral data and phylogenies. J. Comput.
Biol., 10(5):669–678, 2003.
[Mos04]
E. Mossel. Phase transitions in phylogeny. Trans. Amer. Math. Soc., 356(6):2379–2404, 2004.
[Mos07]
E. Mossel. Distorted metrics on trees and phylogenetic forests. IEEE/ACM Trans. Comput. Bio.
Bioinform., 4(1):108–116, 2007.
[MP03]
E. Mossel and Y. Peres. Information flow on trees. Ann. Appl. Probab., 13(3):817–844, 2003.
[MV05]
Elchanan Mossel and Eric Vigoda. Phylogenetic MCMC Algorithms Are Misleading on Mixtures
of Trees. Science, 309(5744):2207–2209, 2005.
[PR09]
Y. Peres and S. Roch. Reconstruction on trees: Exponential moment bounds for linear estimators.
Preprint, 2009.
[Roc06]
Sébastien Roch. A short proof that phylogenetic tree reconstruction by maximum likelihood is
hard. IEEE/ACM Trans. Comput. Biology Bioinform., 3(1):92–94, 2006.
[Roc08]
Sébastien Roch. Sequence-length requirement for distance-based phylogeny reconstruction:
Breaking the polynomial barrier. In FOCS, pages 729–738, 2008.
[Roc09]
Sébastien Roch. Phase transition in distance-based phylogeny reconstruction. Preprint, 2009.
[SHP88]
M. A. Steel, M. D. Hendy, and D. Penny. Loss of information in genetic distances. Nature,
336(6195):118, 1988.
[SN87]
N. Saitou and M. Nei. The neighbor-joining method: A new method for reconstructing phylogenetic trees. Mol. Biol. Evol., 4(4):406–425, 1987.
[SS63]
R. Sokal and P. Sneath. Principles of Numerical Taxonomy. W. H. Freeman and Co., San Francisco, Calif., 1963.
18
[SS99]
Michael A. Steel and László A. Székely. Inverting random functions. Ann. Comb., 3(1):103–113,
1999. Combinatorics and biology (Los Alamos, NM, 1998).
[SS02]
M. A. Steel and L. A. Székely. Inverting random functions. II. Explicit bounds for discrete maximum likelihood estimation, with applications. SIAM J. Discrete Math., 15(4):562–575 (electronic), 2002.
[SS03]
C. Semple and M. Steel. Phylogenetics, volume 22 of Mathematics and its Applications series.
Oxford University Press, 2003.
[Ste94]
M. Steel. Recovering a tree from the leaf colourations it generates under a Markov model. Appl.
Math. Lett., 7(2):19–23, 1994.
[Ste01]
M. Steel. My Favourite Conjecture. Preprint, 2001.
19
A
Extending to General Trees
It is possible to generalize the previous arguments to general trees, using a combinatorial algorithm of [DMR09a],
thereby giving a proof of Theorem 3. To apply the algorithm of [DMR09a] we need to obtain a generalization
of Proposition 4 for disjoint subtrees in “general position.” This is somewhat straightforward and we give a
quick sketch in this section.
A.1
Basic Definitions
The algorithm in [DMR09a] is called Blindfolded Cherry Picking. We refer the reader to [DMR09a] for a full
description of the algorithm, which is somewhat involved. It is very similar in spirit to the algorithm introduced in Section 3.2, except for complications due to the non-homogeneity of the tree. The proof in [DMR09a]
is modular and relies on two main components: a distance-based combinatorial argument which remains unchanged in our setting; and a statistical argument which we now adapt. The key to the latter is [DMR09a,
Proposition 4]. Note that [DMR09a, Proposition 4] is not distance-based as it relies on a complex ancestral
reconstruction function—recursive majority. Our main contribution in this section is to show how this result
can be obtained using the techniques of the previous sections—leading to a fully distance-based reconstruction
algorithm.
In order to explain the complications due to the non-homogeneity of the tree and state our main result, we
first need to borrow a few definitions from [DMR09a].
Basic Definitions. Fix 0 < ∆ ≤ f ≤ g < g ∗ as in Theorem 3. Let T = (V, E, [n], ρ; τ ) ∈ Yf,g
∆ be a
phylogeny with underlying tree T = (V, E). In this section, we sometimes refer to the edge set, vertex set and
leaf set of a tree T 0 as E(T 0 ), V(T 0 ), and L(T 0 ) respectively.
Definition 9 (Restricted Subtree) Let V 0 ⊆ V be a subset of the vertices of T . The subtree of T restricted to
V 0 is the tree T 0 obtained by 1) keeping only nodes and edges on paths between vertices in V 0 and 2) by then
contracting all paths composed of vertices of degree 2, except the nodes in V 0 . We sometimes use the notation
T 0 = T |V 0 . See Figure 4 for an example.
Definition 10 (Edge Disjointness) Denote by PathT (x, y) the path (sequence of edges) connecting x to y in
T . We say that two restricted subtrees T1 , T2 of T are edge disjoint if
PathT (x1 , y1 ) ∩ PathT (x2 , y2 ) = ∅,
for all x1 , y1 ∈ L(T1 ) and x2 , y2 ∈ L(T2 ). We say that T1 , T2 are edge sharing if they are not edge disjoint.
See Figure 5 for an example.
Definition 11 (Legal Subforest) We say that a tree is a rooted full binary tree if all its internal nodes have
degree 3 except the root which has degree 2. A restricted subtree T1 of T is a legal subtree of T if it is also a
rooted full binary tree. We say that a forest
F = {T1 , T2 , . . .},
is legal subforest of T if the Tι ’s are edge-disjoint legal subtrees of T . We denote by ρ(F) the set of roots of F.
Definition 12 (Dangling Subtrees) We say that two edge-disjoint legal subtrees T1 , T2 of T are dangling if
there is a choice of root for T not in T1 or T2 that is consistent with the rooting of both T1 and T2 . See Figure 6
below for an example where two legal, edge-disjoint subtrees are not dangling.
20
Figure 4: Restricting the top tree to its white nodes.
u3
u2
u4
u1
u5
u7
u8
u6
Figure 5: The subtrees T |{u1 ,u2 ,u3 ,u8 } and T |{u4 ,u5 ,u6 ,u7 } are edge-disjoint. The subtrees T |{u1 ,u5 ,u6 ,u8 } and
T |{u2 ,u3 ,u4 ,u7 } are edge-sharing.
21
x2
z2
y2
x1
y1
z1
u2
w2
v2
u1
w1
v1
T2
T1
Figure 6: Basic Disjoint Setup (General). The rooted subtrees T1 , T2 are edge-disjoint but are not assumed to
be dangling. The white nodes may not be in the restricted subtrees T1 , T2 . The case w1 = x1 and/or w2 = x2
is possible. Note that if we root the tree at any node along the dashed path, the subtrees rooted at y1 and y2 are
edge-disjoint and dangling (unlike T1 and T2 ).
Definition 13 (Basic Disjoint Setup (General)) Let T1 = Tx1 and T2 = Tx2 be two restricted subtrees of T
rooted at x1 and x2 respectively. Assume further that T1 and T2 are edge-disjoint, but not necessarily dangling.
Denote by yι , zι the children of xι in Tι , ι = 1, 2. Let wι be the node in T where the path between T1 and T2
meets Tι , ι = 1, 2. Note that wι may not be in Tι since Tι is restricted, ι = 1, 2. If wι 6= xι , assume without loss
of generality that wι is in the subtree of T rooted at zι , ι = 1, 2. We call this configuration the Basic Disjoint
Setup (General). See Figure 6. Let τ (T1 , T2 ) be the length of the path between w1 and w2 in the metric τ .
A.2
Deep Distorted Metric
Our reconstruction algorithm for homogeneous trees (see Section 3) builds the tree level by level and only
encounters situations where one has to compute the distance between two dangling subtrees (that is, the path
connecting the subtrees “goes above them”). However, when reconstructing general trees by growing a subforest from the leaves, more general situations such as the one depicted in Figure 6 cannot be avoided and have to
be dealt with carefully.
Hence, our goal in this subsection is to compute the distance between the internal nodes x1 and x2 in the
Basic Disjoint Setup (General). We have already shown how to perform this computation when T1 and T2 are
dangling, as this case is handled easily by Proposition 4 (after a slight modification of the distance estimate;
see below). However, in the general case depicted in Figure 6, there is a complication. When T1 and T2 are not
dangling, the reconstructed sequences at x1 and x2 are not conditionally independent. But it can be shown that
for the algorithm Blindfolded Cherry Picking to work properly, we need: 1) to compute the distance between x1
and x2 correctly when the two subtrees are close and dangling; 2) detect when the two subtrees are far apart (but
an accurate distance estimate is not required in that case). This turns out to be enough because the algorithm
Blindfolded Cherry Picking ensures roughly that close reconstructed subtrees are always dangling. We refer
the reader to [DMR09a] for details.
The key point is the following: if one computes the distance between y1 and y2 rather than the distance
between x1 and x2 , then the dangling assumption is satisfied (re-root the tree at any node along the path
22
connecting w1 and w2 ). However, when the algorithm has only reconstructed T1 and T2 , we cannot tell which
pair in {y1 , z1 } × {y2 , z2 } is the right one to use for the distance estimation. Instead, we compute the distance
for all pairs in {y1 , z1 } × {y2 , z2 } and the following then holds: in the dangling case, all these distances will
agree (after subtracting the length of the edges between x1 , x2 and {y1 , z1 , y2 , z2 }); in the general case, at least
one is correct. This is the basic observation behind the routine D ISTORTED M ETRIC in Figure 7 and the proof
of Proposition 6 below. We slightly modify the definitions of Section 3.
Using the notation of Definition 13, fix (a0 , b0 ) ∈ {y1 , z1 } × {y2 , z2 }. For v ∈ V and ` ∈ L, denote
by |`|v be the graph distance (that is, the number of edges) between v and leaf `. Assume that we are given
θe for all e ∈ E(Ta0 ) ∪ E(Tb0 ). Let α > 1 and ∆h = bα log2 log2 nc. Imagine (minimally) completing the
subtrees below a0 and b0 with 0-length edges so that the leaves below a0 and b0 are at distance at least ∆h. For
x ∈ {a, b}, denote by x1 , . . . , x2∆h , the vertices below x0 at distance ∆h from x0 and, for j = 1, . . . , 2∆h , let
Xj be the leaves of T below xj . For 1 ≤ j ≤ 2∆h , we estimate τ (a0 , b0 ) as follows
X
X
1
−1 −τ̂ (a0 ,b0 )
τ̄ (aj , bj ) ≡ − ln
Θ−1
.
0 Θb ,b0 e
a
,a
0
0
|Aj ||Bj | 0
0
a ∈Aj b ∈Bj
Note that, because the tree is binary, it holds that
X −|a0 | X −|b0 |
X X −|a0 | −|b0 |
aj
bj
aj
bj
2
=
2
2
= 1,
a0 ∈Aj b0 ∈Bj
a0 ∈Aj
b0 ∈Bj
and we can think of the weights on Aj (similarly for Bj ) as resulting from a homogeneous flow Ψaj from aj to
Aj . Then, the bound on the variance of
X −|a0 |
aj
0
Saj ≡
2
Θ−1
aj ,a0 σa ,
a0 ∈Aj
in Proposition 1 still holds with
Kaj ,Ψaj =
X
Raj (e)Ψ(e)2 .
e∈E(Taj )
Moreover Kaj ,Ψaj is uniformly bounded following an argument identical to (8) in the proof of Lemma 4. For
d ∈ R and r > 0, let Br (d) be the ball of radius r around d. We define the dense ball around j to be the smallest
∆h
ball around τ̄ (aj , bj ) containing at least 2/3 of J = {τ̄ (aj 0 , bj 0 )}2j 0 =1 (as a multiset). The radius of the dense
ball around j is
2 ∆h
∗
rj = inf r : |Br (τ̄ (aj , bj )) ∩ J | ≥ (2 ) ,
3
for j = 1, . . . , 2∆h . We define our estimate of τ (a0 , b0 ) to be τ̄ 0 (a0 , b0 ) = τ̄ (aj ∗ , bj ∗ ), where j ∗ = arg minj rj∗ .
See Figure 2. For D > 0, W > 5, we define
W
1
−∆h
SD(a0 , b0 ) = 1 2
j : τ̄ (aj , bj ) ≤ D + ln
>
,
3
2
and we let
d(a0 , b0 ) =
[τ̄ 0 (a0 , b0 )]∆ , if SD(a0 , b0 ) = 1,
+∞,
o.w.
Proposition 6 (Accuracy of D ISTORTED M ETRIC) Let α > 1, D > 0, W > 5, γ > 0 and g < g 0 < g ∗ .
Consider the Basic Disjoint Setup (General) with F = {T1 , T2 } and Q = {y1 , z1 , y2 , z2 }. Assume we are given
θe for all e ∈ E(T1 ) ∪ E(T2 ). Let Υ denote the output of D ISTORTED M ETRIC in Figure 7. There exists κ > 0,
such that if the following condition holds:
23
Algorithm D ISTORTED M ETRIC
Input: Rooted forest F = {T1 , T2 } rooted at vertices x1 , x2 ; weights τe , for all e ∈ E(T1 ) ∪ E(T2 );
Output: Distance Υ;
• [Children] Let yι , zι be the children of xι in F for ι = 1, 2 (if xι is a leaf, set zι = yι = xι );
• [Distance Computations] For all pairs (a, b) ∈ {y1 , z1 } × {y2 , z2 }, compute
D(a, b) := d(a, b) − τ (a, x1 ) − τ (b, x2 );
• [Multiple Test] If
(1)
(1)
(2)
(2)
max{ D(r1 , r2 ) − D(r1 , r2 ) :
(ι)
(ι)
(r1 , r2 ) ∈ {y1 , z1 } × {y2 , z2 }, ι = 1, 2} = 0,
return Υ := D(z1 , z2 ), otherwise return Υ := +∞ (return Υ := +∞ if any of the distances above is +∞).
Figure 7: Routine D ISTORTED M ETRIC.
• [Edge Length] It holds that τ (e) ≤ g 0 , ∀e ∈ E(Tx ), x ∈ Q5 ;
• [Sequence Length] The sequence length is k > logκ (n),
then we have, with probability at least 1 − O(n−γ ),
Υ = τ (x1 , x2 )
under either of the following two conditions:
1. [Dangling Case] T1 and T2 are dangling and τ (T1 , T2 ) < D, or
2. [Finite Estimate] Υ < +∞.
The proof, which is a simple combination of the proof of Proposition 4 and the remarks above the statement of
Proposition 6, is left out.
Full Algorithm. The rest of the Blindfolded Cherry Picking algorithm is unchanged except for an additional
step to compute averaging weights as in the algorithm of Section 3. This concludes our sketch of the proof of
Theorem 3.
5
For technical reasons explained in [DMR09a], we allow edges slightly longer than the upper bound g.
24
B
Proofs
Proof of Lemma 1: Note that E[Fbijab ] = πi e−τ (a,b)Q
ij
. Then
i
h
X X
νi
πi e−τ (a,b)Q νj
E ν > Fbab ν =
i∈Φ
=
X
ij
j∈Φ
νi (πi e−τ (a,b) νi )
i∈Φ
= e−τ (a,b)
X
πi νi2
i∈Φ
= e
−τ (a,b)
.
Proof of Lemma 2: By Azuma’s inequality we get
P [τ̂ (u, v) > τ (u, v) + ε]
" k
#
1X i i
=P
σu σv < e−τ (u,v)−ε
k
i=1
#
" k
1X i i
σu σv < e−τ (u,v) − (1 − e−ε )e−τ (u,v)
=P
k
i=1
" k
#
1X i i
1 1
−ε −δ log log(n)
≤P
σu σv < E[σu σv ] − (1 − e )e
k
i=1
2 !
(1 − e−ε )e−δ log log(n)
≤ exp −
2k(2ν̄/k)2
≤ n−γ ,
for κ large enough depending on δ, ε, γ. Above, we used that
τ (u, v) = − ln E[σu1 σv1 ].
A similar inequality holds for the other direction.
25
Proof of Lemma 3: Assume the first three conditions hold. By Azuma’s inequality we get
W
P τ̂ (u, v) − τ (u0 , u) − τ (v0 , v) ≤ D + ln
2
" k
#
1X i i
=P
σu σv ≥ 2W −1 e−D e−τ (u0 ,u)−τ (v0 ,v)
k
i=1
#
" k
1X i i
−τ (u,v)
−1 −D −τ (u0 ,u)−τ (v0 ,v)
σu σv ≥ e
+W e e
≤P
k
i=1
#
" k
X
1
σui σvi ≥ E[σu1 σv1 ] + W −1 e−D e−τ (u0 ,u)−τ (v0 ,v)
=P
k
i=1
2 !
W −1 e−D e−τ (u0 ,u)−τ (v0 ,v)
≤ exp −
2k(2ν̄/k)2
≤ n−γ ,
for κ large enough. A similar argument gives the second claim.
Proof of Proposition 1: We follow the proofs of [EKPS00, MP03]. Let ēi be the unit vector in direction i. Let
x ∈ [n], then
> τ (ρ,x)Q
.
E[ē>
ξx | ξρ ] = ēξρ e
Therefore,
τ (ρ,x)Q
ν = σρ e−τ (ρ,x) ,
E[σx | σρ ] = ē>
ξρ e
and
E[S | σρ ] =
X Ψ(x)σρ e−τ (ρ,x)
X
= σρ
Ψ(x) = σρ .
Θρ,x
x∈[n]
x∈[n]
In particular,
X
E[S] =
πi νi = 0.
ι∈Φ
For x, y ∈ [n], let x ∧ y be the meeting point of the paths between ρ, x, y. We have
X
E[σx σy ] =
P[ξx∧y = ι]E[σx σy | ξx∧y = ι]
ι∈Φ
=
X
πι E[σx | ξx∧y = ι]E[σy | ξx∧y = ι]
ι∈Φ
=
X
πι e−τ (x∧y,x) νι e−τ (x∧y,y) νι
ι∈Φ
= e−τ (x,y)
X
πι νι2
ι∈Φ
−τ (x,y)
= e
.
26
Then
Var[S] = E[S 2 ]
X Ψ(x)Ψ(y)
=
E[σx σy ]
Θρ,x Θρ,y
x,y∈[n]
X
Ψ(x)Ψ(y)e2τ (ρ,x∧y) .
=
x,y∈[n]
For e ∈ E, let e = (e↑ , e↓ ) where e↑ is the vertex closest to ρ. Then, by a telescoping sum, for u ∈ V
X
X
X
e2τ (ρ,e↑ )
e2τ (ρ,e↓ ) −
Rρ (e) =
e∈Path(ρ,u)
e∈Path(ρ,u)
e∈Path(ρ,u)
2τ (ρ,u)
= e
− 1,
and therefore
E[S 2 ] =
X
Ψ(x)Ψ(y)e2τ (v,x∧y)
x,y∈[n]
=
X
Ψ(x)Ψ(y) 1 +
x,y∈[n]
= 1+
= 1+
X
X
Rρ (e)
e∈Path(ρ,x∧y)
Rρ (e)
1{e ∈ Path(ρ, x ∧ y)}Ψ(x)Ψ(y)
X
e∈E
x,y∈[n]
X
Rρ (e)Ψ(e)2 .
e∈E
Proof of Lemma 4: From (6), we have
Kρ,Ψ ≤
h−1
X
(1 − e−2g )2h−i
i=0
≤
h
X
e2jg e−(2 ln
√
e2(h−i)g
22(h−i)
2)j
j=1
=
h
X
e2j(g−g
∗)
j=1
≤
+∞
X
(e−2(g
∗ −g)
)j
j=0
=
1
< +∞,
1 − e−2(g∗ −g)
√
where recall that g ∗ = ln 2.
Proof of Proposition 2: Let
∆h
∆h
Z = {aj }2j=1 ∪ {bj }2j=1 ,
27
(8)
and
i k
E = {(σZ
)i=1 }.
For j = 1, . . . , 2∆h , let
e
−τ̂ (aj ,bj )
k
1X i i
=
σaj σbj .
k
i=1
Note that
E[e−τ̂ (aj ,bj ) ] = e−τ (aj ,bj ) ,
by Lemma 1. For i = 1, . . . , k, j = 1, . . . , 2∆h , and x ∈ {a, b}, let
σ̄xi j =
X 2−h00 σ i 0
x
.
Θ
xj ,x0
0
x ∈Xj
By the Markov property, it follows that, conditioned on E,
{σ̄xi j : i = 1, . . . , k, j = 1, . . . , 2∆h , x ∈ {a, b}},
are mutually independent. Moreover, by Proposition 1, we have
E[σ̄xi j | E] = σxi j ,
and
E[(σ̄xi j )2 | E] ≤ 2π̄ −1
1
1−
e−2(g∗ −g)
.
Therefore, for any ζ > 0 there exists κ > 1 such that
E[e
−τ̄ (aj ,bj )
τ (a0 ,aj )+τ (b0 ,bj )
| E] = e
= eτ (a0 ,aj )+τ (b0 ,bj )
!
k
1X
i
i
E[σ̄aj σ̄bj | E]
k
1
k
i=1
k
X
!
E[σ̄ai j | E]E[σ̄bi j | E]
i=1
−(τ̂ (aj ,bj )−τ (aj ,a0 )−τ (bj ,b0 ))
= e
,
and
−τ̄ (aj ,bj )
Var[e
!
k
1 X
i
i
Var[σ̄aj σ̄bj | E]
k2
2τ (a0 ,aj )+2τ (b0 ,bj )
| E] = e
≤ e2τ (a0 ,aj )+2τ (b0 ,bj )
≤
e2τ (a0 ,aj )+2τ (b0 ,bj )
k
e4gbα log2 log2 nc
k
≤ ζ.
≤
28
1
k2
i=1
k
X
!
E[(σ̄ai j )2 | E]E[(σ̄bi j )2 | E]
i=1
2
1 − e−2(g∗ −g)
2
2
1 − e−2(g∗ −g)
2
Take κ large enough such that Lemma 2 holds for diameter 2gbα log2 log2 nc+D, precision ε/6 and failure
probability O(n−(γ+1) ). Then, we have
|τ̂ (aj , bj ) − τ (aj , bj )| < ε/6,
∀j = 1, . . . , 2∆h ,
(9)
with probability 1 − O(n−γ ). Let E 0 be the event that (9) holds. Note that
E[e−τ̄ (aj ,bj ) | E ∩ E 0 ] = e−(τ̂ (aj ,bj )−τ (aj ,a0 )−τ (bj ,b0 )) .
Let
ε0 = min{(eε/3 − eε/6 )e−D , (e−ε/6 − e−ε/3 )e−D }.
By Chebyshev’s inequality, we have that
1
0
P τ̄ (aj , bj ) < τ (a0 , b0 ) − ε | E ∩ E
3
h
i
≤ P e−τ̄ (aj ,bj ) > e−τ (a0 ,b0 )+ε/3 | E ∩ E 0
i
h
≤ P e−τ̄ (aj ,bj ) − E[e−τ̄ (aj ,bj ) | E ∩ E 0 ] > [e−τ (a0 ,b0 )+ε/3 − e−(τ̂ (aj ,bj )−τ (aj ,a0 )−τ (bj ,b0 )) ] | E ∩ E 0
i
h
≤ P e−τ̄ (aj ,bj ) − E[e−τ̄ (aj ,bj ) | E ∩ E 0 ] > e−τ (a0 ,b0 ) [eε/3 − eτ (aj ,bj )−τ̂ (aj ,bj ) ] | E ∩ E 0
i
h
≤ P e−τ̄ (aj ,bj ) − E[e−τ̄ (aj ,bj ) | E ∩ E 0 ] > e−D [eε/3 − eε/6 ] | E ∩ E 0
h
i
≤ P e−τ̄ (aj ,bj ) − E[e−τ̄ (aj ,bj ) | E ∩ E 0 ] > ε0 | E ∩ E 0
≤ 1/12,
for ζ small enough. A similar argument holds for
1
0
P τ̄ (aj , bj ) > τ (a0 , b0 ) + ε | E ∩ E ≤ 1/12.
3
Conditioned on E ∩ E 0 ,
−∆h
L=2
2∆h
X
1
1 |τ̄ (aj , bj ) − τ (a0 , b0 )| < ε
3
j=1
is a sum of independent {0, 1}-variables with average at least 5/6. By Azuma’s inequality, we have
P[L ≤ 2/3 | E ∩ E 0 ] ≤ P[L − E[L | E ∩ E 0 ] < −1/6 | E ∩ E 0 ]
(1/6)2
≤ exp − −∆h 2 ∆h
2(2
) 2
α
≤ exp (−O(log2 n))
≤ O(n−γ ),
where we used that α > 1.
This implies that with (unconditional) probability at least 1 − O(n−γ )
2
1
|τ̄ (aj , bj ) − τ (a0 , b0 )| < ε > (2∆h ).
3
3
29
In particular, there exists j such that
2
2
0
j : |τ̄ (aj 0 , bj 0 ) − τ̄ (aj , bj )| < ε > (2∆h ),
3
3
and for all j such that
|τ̄ (aj , bj ) − τ (a0 , b0 )| > ε,
we have that
2
j : |τ̄ (aj 0 , bj 0 ) − τ̄ (aj , bj )| < ε
3
0
1
≤ (2∆h ).
3
Finally, we get that
|τ̄ 0 (a0 , b0 ) − τ (a0 , b0 )| ≤ ε.
Proof of Proposition 3: The proof is similar to the proof of Lemma 3 and Proposition 2.
Proof of Proposition 4: We let ε < ∆/2.
The first part of the proposition follows immediately from Proposition 2 and the second part of Proposition 3. Note that Propositions 2 and 3 are still valid when h0 < bα log2 log2 nc if we take instead ∆h =
min{h0 , bα log2 log2 nc}. Indeed, in that case there is no need to perform an implicit ancestral sequence reconstruction and the proofs of the lemmas follow immediately from Lemmas 2 and 3.
For the second part, choose κ so as to satisfy the conditions of Proposition 2 with diameter D + ln W and
apply the first part of Proposition 3.
Proof of Proposition 5: The proof follows immediately from Proposition 4 and the remark above the statement
of Proposition 5.
Proof of Theorem 4: The proof of Theorem 4 follows from Propositions 4 and 5. Indeed, at each level h0 , we
are guaranteed by the above to compute a distorted metric with a radius large enough to detect all cherries on
the next level using four-point tests. The proof follows by induction.
30
| 5cs.CE
|
Operational Semantics of Process Monitors
arXiv:1705.05137v1 [cs.PL] 15 May 2017
Jun Inoue and Yoriyuki Yamagata
National Institute of Advanced Industrial Science and Technology (AIST), 1-8-31
Midorigaoka, Ikeda, Osaka 563-8577 Japan {jun.inoue,
yoriyuki.yamagata}@aist.go.jp
Abstract. CSPE is a specification language for runtime monitors that
can directly express concurrency in a bottom-up manner that composes
the system from simpler, interacting components. It includes constructs
to explicitly flag failures to the monitor, which unlike deadlocks and livelocks in conventional process algebras, propagate globally and aborts the
whole system’s execution. Although CSPE has a trace semantics along
with an implementation demonstrating acceptable performance, it lacks
an operational semantics. An operational semantics is not only more
accessible than trace semantics but also indispensable for ensuring the
correctness of the implementation. Furthermore, a process algebra like
CSPE admits multiple denotational semantics appropriate for different
purposes, and an operational semantics is the basis for justifying such semantics’ integrity and relevance. In this paper, we develop an SOS-style
operational semantics for CSPE , which properly accounts for explicit
failures and will serve as a basis for further study of its properties, its
optimization, and its use in runtime verification.
Keywords: Operational Semantics, Concurrency, Runtime Monitoring, Communicating Sequential Processes
1
Introduction
Specification-based runtime monitoring [5] checks a program’s execution trace
against a formal specification. Often more rigorous than testing due to the
presence of a formal specification, this technique is also computationally much
cheaper than formal verification methods like model checking, as it only needs
to look at concrete program runs with instrumentation. CSPE [10] is a language
based on Hoare’s Communicating Sequential Processes [6] for developing the formal specification. Unlike many other languages in this niche, CSPE can directly
express concurrency. Moreover, it builds up the specification in a bottom-up
manner by composing smaller, interacting components, helping to model complex behavior.
CSPE ’s main appeal as a specification language, compared to plain CSP, is a
FAIL construct that signals a global failure, aborting all processes in the model
at once. This construct can be used like assert(false) in C or Java, allowing to
code invariants that mark some states as (should-be) unreachable. By contrast,
Event
Event Variable
Term
Event Set
Event Set Param
e∈Σ
x∈X
P, Q ∈ Terms ::= STOP | FAIL | ?x : E → P | P Q | P ||E Q
E ::= f (y1 , . . . , yn ) where f : Σ n → 2Σ is computable
y ::= x | e
Fig. 1. Syntax of CSPE .
deadlocks and livelocks, the conventional notions of failure in CSP, affect only
the deadlocked or livelocked process(es). These failures are thus very difficult to
propagate into a failure for the entire system, as desired for assertion failures.
However, the semantics of how FAIL propagates throughout the model requires
special treatment. Because the propagation preempts all other activities, normal
execution rules must apply only when FAIL is not currently propagating. This
is a negative constraint, which is generally problematic [4].
While earlier work [10] demonstrated a trace semantics and a reasonably efficient implementation for CSPE , an operational semantics has been lacking. Developing an operational semantics is highly desirable for several reasons. Firstly,
though a trace semantics more naturally defines the set of behaviors (i.e. traces)
that comply with a CSPE specification, an operational semantics more directly
defines the implementation. Secondly, process algebras admit multiple denotational semantics capturing different aspects of operationally defined behavior [3].
Investigating the full spectrum of such semantics requires an operational semantics. Finally, an operational semantics provides a more accessible presentation
of the semantics than denotational semantics.
1.1
Contributions
In this paper, after reviewing the syntax and trace semantics of CSPE (Section 2), we present the following contributions.
– We define an operational semantics in SOS format [8], which properly captures the propagation of FAIL while avoiding the complexities of rules with
negative premises (Section 3).
– We prove that the operational semantics induces the previously published
trace semantics (Section 4).
2
Syntax and Trace Semantics of CSPE
This section reviews the syntax and trace semantics of CSPE . Figure 1 presents
the syntax. A CSPE term represents a process, which is an entity that successively emits events drawn from an alphabet Σ. Terms are built from the following
constructs, with the indicated meanings. For a thorougher explanation, see [10].
– The stuck term STOP does not emit anything.
– The failing term FAIL aborts all processes.
Trace
Trace Set
Trace Set Operations
s, t ∈ Σ ∗
T ∈ TraceSets ::= prefix-closed subsets of Σ ∗
eT := {ε} ∪ {et | t ∈ T }
T (e) := {t | et ∈ T }
∅ ||E T := T ||E ∅ := ∅
[
T1 ||E T2 :=
e(T1 (e) ||E T2 (e)) ∪
e∈E
[
(1)
(e(T1 (e) ||E T2 ) ∪ e(T1 ||E T2 (e)))
(2)
e∈Σ−E
Trace Semantics
JSTOPK := {ε}
JP QK := JP K ∪ JQK
JFAILK := ∅
JP ||E QK := JP K ||E JQK
J?x : E → P K := {ε} ∪
[
eJ[e/x]P K
e∈E
Fig. 2. Trace semantics of CSPE . Equation (1) takes precedence over eq. (2), so the
latter applies only if the former does not.
– Prefix ?x : E → P chooses and emits an event e ∈ E, then executes [e/x]P .
– Choice P Q executes P or Q, whichever manages to emit something first.
– Parallel composition P ||E Q executes P and Q in parallel. Their events are
interleaved arbitrarily, except events in E are synchronized.
An event set E can be specified by any computable function parametrized by
the x’s bound by surrounding prefix operators.
In this short paper, we omit recursion and the terminating action X in the
interest of conciseness. This paper’s focus is on analyzing FAIL, and X complicates the presentation substantially without adding anything of conceptual
significance. Recursion seems to be similar, though it is still under investigation.
Figure 2 presents the trace semantics. A trace is a (possibly empty) sequence
of events, and Σ ∗ is the set of all traces. The concatenation of traces s and t
is written st. A trace set is any prefix-closed set of traces, which can be empty,
unlike in conventional process algebras. The trace semantics of CSPE assigns to
each term P a trace set JP K, which is intuitively the set of traces P can emit.
The semantic map uses some operations on trace sets. If T is a trace set, eT
prepends e to all members of T and adjoins ε, while T (e) discards all traces in
T that do not start with e and drops the leading e from all remaining traces.
The ||E operator is defined by eqs. (1) and (2). Though significantly simplified,
these equations are equivalent to the ones found in [10] modulo the absence
of X. In [10], this operator was defined “coinductively”, which was correct but
misleading. Formally, by the Knaster-Tarski Theorem, the defining equations (1)
and (2) have a greatest solution in the complete lattice of total binary functions
on TraceSets ordered by point-wise inclusion, which was taken to be ||E . However,
′′
′
if ||E and ||E are any two solutions of these equations, then for any T1 and T2 ,
′′
′
every trace in T1 ||E T2 is also in T1 ||E T2 , by straightforward induction on
the trace’s length. Thus, the solution is unique, and ||E is this unique solution.
a ::= e | τ
D ∈ Doomed ::= FAIL | D D | D ||E P | P ||E D
P̂ , Q̂ ∈ Terms − Doomed
Action
Doomed Term
Viable Term
Operational Semantics
e∈E
e
e
e
e
e
P̂ 7→ P ′
e
e
P̂ ||E Q̂ 7→ P ′ ||E Q′
τ
FAIL FAIL 7→ FAIL
P Q 7→ P Q′
a
Q 7→ Q′
a 6∈ E
τ
D1 7→ P1
τ
D2 7→ P2
τ
τ
τ
FAIL ||E P 7→ FAIL
s
as
a 6∈ E
a
P̂ ||E Q 7→ P̂ ||E Q′
D1 ||E D2 7→ P1 ||E D2
a
ε
τ
a
P 7→ P ′ Z⇒ P ′′
P Z⇒ P
τ
P ||E Q̂ 7→ P ′ ||E Q̂
e∈E
e
τ
P 7→ P ′
P Q 7→ Q′
Q̂ 7→ Q′
Q 7→ Q′
a
Q 7→ Q′
P Q 7→ P ′
τ
P Q 7→ P ′ Q
(?x : E → P ) 7→ [e/x]P
P 7→ P ′
P 7→ P ′
P Z⇒ P ′′
D1 ||E D2 7→ D1 ||E P2
τ
P ||E FAIL 7→ FAIL
τ
s
P 7→ P ′ Z⇒ P ′′
s
P Z⇒ P ′′
Fig. 3. Operational semantics of CSPE .
S
S
S
Lemma 1. ||E is continuous, i.e. ( S1 ) ||E ( S2 ) = T1 ∈S1 ,T2 ∈S2 T1 ||E T2 .
Proof. The defining equations (1) and (2) preserve continuity, so in fact the
Knaster-Tarski construction can be carried out in the space of continuous binary
operators, which is also a complete lattice under point-wise inclusion.
3
Operational Semantics
This section presents the operational semantics. The semantics is given in Figa
ure 3, which defines internal transitions P 7→ Q between terms. Some transitions do not emit events but instead emit the silent action τ . A visible transition
s
P Z⇒ Q happens when P internally transitions to Q in zero or more steps, and
the non-τ actions it emits along the way forms s.
The main challenge in this semantics is capturing the propagation of FAIL.
For example, in P ||∅ FAIL, the P must not be allowed to keep emitting events,
for then P could do so indefinitely, withholding the propagation of FAIL. Instead,
FAIL should kill all processes including P , transitioning the whole term to FAIL.
To achieve this effect, the usual rule that allows the left operand to transition
must apply only when the right operand is not failing. This constraint is tricky
to capture because it is a negative constraint.
In our semantics, the constraint is captured by the viability annotation P̂ .
This annotation restricts the range of the metavariable P̂ to exclude doomed
terms, i.e. terms for which transitioning to FAIL has become inevitable and are
now propagating FAIL within themselves. These annotations are placed so that
when a term is doomed, rules that propagate FAIL become the only applicable
ones, thus forcing the propagation to take place.
Proposition 2. A doomed process always transitions to FAIL while emitting
nothing but τ .
a
Proof. D 7→ P implies a = τ ∧ P ∈ Doomed ∧ |D| > |P |, where |P | denotes
term size, by induction on D. Thus, a doomed term can only τ -transition, and
only finitely many times, while staying doomed. Another induction shows ∀D 6=
τ
FAIL. ∃D′ . D 7→ D′ , so a doomed term keeps transitioning until it reaches FAIL.
4
Correspondence Between the Semantics
This section establishes a correspondence between the two semantics: a process’
denotation is precisely the set of traces it can emit, up to but not including any
transitions that doom the process. This means that the monitor comparing a
system to P can declare a failure as soon as the system’s trace strays out of JP K.
s
Theorem 3. JP K = {s | ∃M. P Z⇒ M 6∈ Doomed}.
A special case of this theorem is particularly illuminating: the doomed set is
precisely the set of terms with empty trace sets, corresponding to the fact that
doomed terms silently transition to FAIL.
Proposition 4. P ∈ Doomed ⇐⇒ JP K = ∅.
Proof. Induction on P .
Furthermore, trace sets faithfully follow non-silent transitions, in that the traces
which follow an event e in JP K are precisely the traces of terms Q that follow P
after a sequence of transitions that emit e.
S
e
Lemma 5. JP K(e) = P Z⇒
JQK.
Q
Proof. Induction on the size of P , where event sets do not count toward size, e.g.
|?x : E → P ′ | := |P ′ |+1. This way, |[e/x]P ′ | = |P ′ |, so when P = (?x : E → P ′ ),
the inductive hypothesis applies to [e/x]P ′ , despite it not being a subterm. Several
lemmas are needed along the way, two of which are of particular note. Take for
example P = PS
JP2 K 6= ∅ and e ∈ E. Inductive hypotheses
1 ||E P2 with JP1 K,S
e
e
give JP K(e) = ( P1 Z⇒
JQ
K)
||
(
JQ2 K). Then, continuity (Lemma 1)
E
P2 Z⇒Q2
SQ1 1
S
e
e
lets us commute the
and ||E , equating this to P1 Z⇒
(JQ1 K ||E JQ2 K).
Q1 ,P2 Z⇒Q2
S
e
e
Then, a lemma characterizing those Q with P Z⇒ Q equates this to P Z⇒
JQK.
Q
Theorem 3 is a straightforward consequence of these facts.
s
Proof (of Theorem 3). We show s ∈ JP K ⇐⇒ ∃Q. P Z⇒ Q 6∈ Doomed by induction on s. For the base case, ε ∈ JP K ⇐⇒ P 6∈ Doomed by Proposition 4. If
ε
P 6∈ Doomed, then P Z⇒ P 6∈ Doomed, and if P ∈ Doomed, then P can only transition inside Doomed as noted in the proof of Proposition 2. For the inductive
step, s breaks down as s = es′ , and s ∈ JP K ⇐⇒ s′ ∈ JP K(e). By Lemma 5, this
e
is equivalent to having s′ ∈ JP ′ K and P Z⇒ P ′ for some P ′ , which by inductive
e
s′
hypothesis is equivalent to ∃P ′ , Q. P Z⇒ P ′ Z⇒ Q 6∈ Doomed.
5
Related Works
The main issue with CSPE semantics is the propagation of FAIL, which entails the negative constraint that normal computation rules apply only if FAILa
propagation rules do not. Negative premises of the form P 7→
6 come quite naturally as a means for codifying such constraints, but negative premises are generally quite problematic. A transition relation satisfying negative rules may be
not-existent, or non-unique, with no obvious guiding principle (such as minimality) in choosing the “right” one. Some formats do guarantee well-definedness,
such as GSOS with the witnessing constraint [2] and ntyft/ntyxt [4]. But even
then, negative rules tend to betray desirable properties such as compositionality
of some forms of bisimulation [1].
Our approach exploits the fact that we only have a very specific negative
constraint – the absence of doomed subprocesses – and encodes it with a restriction on the range of metavariables in transition rules. With trick, we manage to
avoid negative premises altogether, essentially turning the system into a positive
one. This approach is very commonly employed, e.g. in reduction rules for the
call-by-value λ calculus [7], where the argument in a function application should
be evaluated only if the function expression cannot be evaluated any further.
We identify FAIL-induced failures by transitions into FAIL, but an alternative approach would be to have FAIL emit a special event ̥, just as termination
is signalled by X. Though we have not pursued this idea in detail, the central
concern there will be to give ̥ higher priority than all other events. Prioritized
transition also involves a negative constraint but is known to be quite wellbehaved, being translatable to plain CSP [9]. At the moment, it is not clear if
FAIL propagation can be translated to the prioritized-transition primitive in [9].
6
Conclusion
We gave an operational semantics for CSPE that adequately captures the behavior of FAIL, the global failure operator, with positive operational rules. This
semantics induces the previously defined trace semantics. As noted in the introduction, this development enables studies of other types of denotational semantics, while informing the implementation. An interesting direction of future
work is to see if FAIL can be specified by priorities, and if that approach yields
better-behaved semantics.
Acknowledgment
The authors would like to thank Yoshinao Isobe for comments on an earlier draft
of this paper and stimulating discussions.
References
1. Bloom, B.: Structural operational semantics for weak bisimulations. Theoretical
Computer Science 146(1), 25–68 (1995)
2. Bloom, B., Istrail, S., Meyer, A.R.: Bisimulation can’t be traced. Journal of the
ACM 42(1), 232–268 (1995)
3. van Glabbeek, R.J.: The linear time - branching time spectrum I. The semantics
of concrete, sequential processes, chap. 1, pp. 3–100. Elsevier (2001)
4. Groote, J.F.: Transition system specifications with negative premises. Theoretical
Computer Science 118(2), 263 – 299 (1993)
5. Havelund, K., Reger, G.: Specification of parametric monitors, pp. 151–189.
Springer Fachmedien Wiesbaden, Wiesbaden (2015)
6. Hoare, C.A.R.: Communicating Sequential Processes. Prentice-Hall, Inc., Upper
Saddle River, NJ, USA (1985)
7. Mitchell, J.C.: Foundations for Programming Languages. MIT Press (1996)
8. Plotkin, G.D.: A structural approach to operational semantics. The Journal of
Logic and Algebraic Programming 60, 17–139 (2004)
9. Roscoe, A.: The expressiveness of CSP with priority. Electronic Notes in Theoretical Computer Science 319, 387 – 401 (2015)
10. Yamagata, Y., Artho, C., Hagiya, M., Inoue, J., Ma, L., Tanabe, Y., Yamamoto, M.:
Runtime Monitoring for Concurrent Systems, pp. 386–403. Springer International
Publishing, Cham (2016)
| 6cs.PL
|
arXiv:1512.07526v2 [math.GR] 30 Jun 2016
On the acylindrical hyperbolicity of the tame
automorphism group of SL2(C)
Alexandre Martin
Abstract
We introduce the notion of über-contracting element, a strengthening of the notion
of strongly contracting element which yields a particularly tractable criterion to show
the acylindrical hyperbolicity, and thus a strong form of non-simplicity, of groups
acting on non locally compact spaces of arbitrary dimension. We also give a simple
local criterion to construct über-contracting elements for groups acting on complexes
with unbounded links of vertices.
As an application, we show the acylindrical hyperbolicity of the tame automorphism
group of SL2 (C), a subgroup of the 3-dimensional Cremona group Bir(P3 (C)), through
its action on a CAT(0) square complex recently introduced by Bisi–Furter–Lamy.
Cremona groups are groups of birationnal transformations of projective spaces over arbitrary fields, and as such are central objects in birational geometry. These groups have a
long history, going back to work of Castelnouvo, Cremona, Enriques and Noether among
others. A lot of work has been devoted to Cremona groups in dimension 2, and a rather
clear picture is now available regarding the structure of such groups: Many results, such
as the Tits alternative [6], the computation of their automorphism groups [11], the Hopf
property [12], as well as their algebraic non-simplicity [7,16], have been proved. While classical approaches to these groups involve methods from birational geometry and dynamical
systems, methods from geometric group theory have proved very fruitful in recent years to
unveil more of the structure of these groups. Indeed, the group Bir(P2 (C)) acts by isometries on a hyperbolic space of infinite dimension, a subspace of the Picard–Manin space,
and such an action can be used to derive many more properties of the group. For instance,
Cantat–Lamy used methods from hyperbolic geometry (more precisely, ideas reminiscent
of small cancellation theory) to show the non-simplicity of the plane Cremona group over
an arbitrary closed field [7], a result recently extended to arbitrary fields by Lonjou [16].
In another direction, Minasyan–Osin used the action of the group Aut(C2 ) of polynomial
automorphism group of C2 on the Bass–Serre tree associated to its decomposition as an
amalgamated product (a decomposition due to Jung and van der Kilk [14, 24]) to obtain,
among other things, a strong form of non-simplicity for this group [19].
By contrast, the situation is much more mysterious in higher dimension, and very few
results are known for Cremona groups of dimension at least 3. A first step would be to
1
A. Martin
On the acylindrical hyperbolicity of Tame(SL2 (C))
understand subgroups of higher Cremona groups. An interesting subgroup of Bir(Pn (C)) is
the automorphism groups of Cn , or more generally the automorphism group of a space birationally equivalent to Cn . An even smaller subgroup is the group of tame automorphisms
of Cn , that is, the subgroup generated by the affine group of Cn and transformations of the
form (x1 , . . . , xn ) 7→ (x1 + P (x2 , . . . , xn ), . . . , xn ) for some polynomial P in n − 1 variables.
This definition of tame automorphisms was recently extended to a general affine quadric
threefold by Lamy–Vénéreau [15]. Further methods from geometric group theory have been
used recently by Bisi–Furter–Lamy to study the structure of the group Tame(SL2 (C)) of
tame automorphisms of SL2 (C), a subgroup of Bir(P3 (C)), through its action on a CAT(0)
square complex [3]. Such complexes have an extremely rich combinatorial geometry, and
this action was used to obtain for instance the Tits alternative for the group, as well as the
linearisability of its finite subgroups.
In this article, we explain how further methods from geometric group theory allow us
to get a better understanding of the geometry and structure of Tame(SL2 (C)). The aim of
this article is thus twofold. On the birational geometric side, we wish to convince the reader
of the general interest of the wider use of tools coming from geometric group theory in the
study of Cremona groups and their subgroups. On the geometric group theoretical side,
we wish to convince the reader of the interest of studying a group through its non-proper
actions on non locally finite complexes of arbitrary dimension. Indeed, while a lot of work
has been done to study groups either through their proper actions on metric spaces, or
through their actions on simplicial trees, few general results and techniques are available to
study actions in a more general setting. Allowing non-proper actions raises serious geometric obstacles, as this most often implies working with non locally finite spaces of arbitrary
dimension. We show however that, under mild geometric assumptions - assumptions that
are satisfied by large classes of complexes with a reasonable combinatorial geometry, it is
possible to obtain simple and useful tools to study such general actions.
In this article, we shall focus on the hyperbolic-like features of Tame(SL2 (C)), and more
precisely on the notion of acylindrical hyperbolicity. It is a theme which has come to the
forefront of geometric group theory in recent years: Indeed, it is a notion sufficiently general
to encompass large classes of groups (mapping class groups [4], Out(Fn ) [2], many CAT(0)
groups [22], etc.), unifying many previously known results, and yet has strong consequences:
acylindrically hyperbolic groups are SQ-universal (that is, every countable group embeds
in a suitable quotient of the group), the associated reduced C ∗ -algebra is simple in the case
of a countable torsion-free group, etc. [10]. We refer the reader to [20] for more details. It
is thus non surprising to witness a wealth of tools being developed to prove the acylindrical
hyperbolicity of ever larger classes of groups.
Acylindrical hyperbolicity is defined in terms of an acylindrical action of a group on a
hyperbolic space, a dynamical condition which is generally cumbersome to check, particularly for actions on non locally compact spaces: Indeed, while these conditions for actions
2
A. Martin
On the acylindrical hyperbolicity of Tame(SL2 (C))
on trees can be reformulated in terms of pointwise stabilisers of pairs of points (this was
the original definition of Sela [21]), they involve coarse stabilisers of pairs of points for
more general actions, that is, group elements moving a pair of point by a given amount.
Controlling such coarse stabilisers generally requires understanding the set of geodesics between two non-compact metric balls, a substantial geometric obstacle in absence of local
compactness. What is more, few actions naturally associated to a group turn out to be
acylindrical (let us mention nonetheless the action of mapping class groups on their associated curve complexes [4], as well as the action of Higman groups on n ≥ 5 generators
on their associated CAT(-1) polygonal complexes [17]). Instead, when given an action of a
group presumed to be acylindrically hyperbolic on a geodesic metric space, one generally
tries to show that a sufficient criterion is satisfied, namely the existence of a WPD element
with a strongly contracting orbit (see [1, Theorem H]). This approach was for instance followed in [2,8,13,19]. Here again, checking the WPD condition for a given element turns out
to be highly non-trivial for actions on non-locally compact spaces of dimension at least two.
As often in geometric group theory, the situation is much easier to handle in the case of
actions on simplicial trees, that is, for groups admitting a splitting. In this case, Minasyan–
Osin obtained a very simple and useful criterion [19, Theorem 4.17], which was used to
show the acylindrical hyperbolicity of large classes of groups: some one-relator groups, the
affine Cremona group in dimension 2, the Higman group, many 3-manifold groups, etc. For
simplicity, we only state it for amalgamated products: Consider an amalgamated product
G of the form A ∗C B, where the edge group C is strictly contained in both A and B and is
weakly malnormal in G, meaning that there exists an element g ∈ G such that C ∩ gCg −1
is finite. Then G is either virtually cyclic or acylindically hyperbolic.
In order to get an idea on how one could generalise such a criterion to more general
actions, it is useful to outline its proof: Considering two distinct edges e, e0 of the associated
Bass–Serre tree with finite common stabiliser, one can extend the minimal geodesic segment
P containing e and e0 into a geodesic line L which is the axis of some element g ∈ G acting
hyperbolically. Such an element turns out to satisfy the WPD identity. Indeed, given two
balls of radius r whose projection on L are sufficiently far apart with respect to r, any
geodesic between two points in these balls must contain a translate of P , and one concludes
using weak malnormality. At the heart of this proof is the very particular geometry of
trees, and in particular the existence of cut-points, which “force” geodesics to go through
a prescribed set of vertices. Such a behaviour cannot a priori be expected from actions on
higher-dimensional spaces.
In this article, we to show that it is possible to obtain a generalisation of the aforementioned criterion for groups acting on higher dimensional complexes, under a mild geometric
assumption on the complex acted upon. Following what happens for trees, we want conditions that force large portions of a geodesic to be prescribed by a coarse information
about their endpoints. In particular, we want to force geodesics to go through certain fi3
A. Martin
On the acylindrical hyperbolicity of Tame(SL2 (C))
nite subcomplexes. We will be interested in the following strengthening of the notion of
strongly contracting element: We consider a hyperbolic element such that one of its axes
comes equipped with a set of checkpoints, a collection of uniformly finite subcomplexes
whose union is coarsely equivalent to the axis and such that for every two points of the
spaces whose projections on the axis are far enough, every geodesic between them must
meet sufficiently many checkpoints between their respective projections (this will be made
precise in Definition 1.3). Such elements will be called über-contractions.
The advantage of such a notion of an über-contraction is that it allows for a more
tractable criterion to prove the acylindrical hyperbolicity of a group. Recall that the standard criterion of Bestvina–Bromberg–Fujiwara [1, Theorem H] is to find a strongly contracting element satisfying the so-called WPD condition (see Theorem 1.1). If ones considers
only über-contractions, it is enough to check a weaker dynamical condition, which is much
easier to check, as it is formulated purely in terms of stabilisers of pairs of points. Our main
criterion is the following:
Theorem A. Let G be a group acting by isometries on a geodesic metric space X. Let g
be an infinite order element which has quasi-isometrically embedded orbits, and assume that
the following holds:
• g is an über-contraction with respect to a set S of checkpoints,
• g satisfies the following weakening of the WPD condition: There exists a constant m0
such that for every point s ∈ S and every m ≥ m0 , only finitely many elements of G
fix pointwise s and g m s.
Then G is either virtually cyclic or acylindrically hyperbolic.
With such a theorem at hand, it is now important to understand how to construct übercontractions. Forcing geodesics to go through given complexes is reasonably manageable
in a CAT(0) space as geodesics can be understood locally: In a CAT(0) space, if two
geodesics meet along a vertex v and make a very large angle at v, then the concatenation
of these two geodesics is again a geodesic, and what is more, any geodesic between points
close enough to the endpoints of this concatenation will also have to go through v. For
spaces which do not have such a rich geometry, we also provide a way to construct übercontractions, by mimicking what happens in a space with a CAT(0) geometry: We want
to construct a hyperbolic element with an axis such that the “angle” made at some special
vertices of this axis is so large that it will force geodesics between two arbitrary points
having sufficiently far apart projections on this axis to go through these special vertices.
As it turns out, a quite mild geometric condition ensures that such a “strong concatenation
of geodesics” phenomenon occurs: We say that a complex has a bounded angle of view if,
roughly speaking, there is a uniform bound on the angle between two arbitrary vertices v and
v 0 , seen from any point that does not lie on a geodesic between v and v 0 (see Definition 2.6).
Having a bounded angle of view holds for CAT(0) metric spaces and other complexes with a
4
A. Martin
On the acylindrical hyperbolicity of Tame(SL2 (C))
more combinatorial notion of non-positive curvature (C 0 (1/6)-polygonal complexes, systolic
complexes). It implies in particular a Strong Concatenation Property of combinatorial
geodesics (see Definition 2.2), a property also satisfied by hyperbolic complexes satisfying
a very weak form of isoperimetric inequality (see Proposition 2.10).
Under such a mild assumption, we have a very simple way of constructing über-contraction
(see Proposition 2.4). This is turn allows us to obtain a second, more local, criterion for
acylindrical hyperbolicity which generalises the aforementioned Minasyan–Osin criterion to
actions on very general metric spaces:
Theorem B (“Link Criterion” for acylindrical hyperbolicity). Let X be a simply connected
hyperbolic polyhedral complex satisfying the Strong Concatenation Property, together with
an action by polyhedral isomorphisms of a group G. Assume that there exists a vertex v of
X and a group element g such that:
1) the action of Gv on the 1-skeleton of the link of v has unbounded orbits (for the
simplicial metric on the 1-skeleton lk(v) where every edge has length 1),
2) Gv is weakly malnormal, that is, the intersection Gv ∩ gGv g −1 is finite.
Then G is either virtually cyclic or acylindrically hyperbolic.
It is this criterion that we use to prove the acylindrical hyperbolicity of Tame(SL2 (C)),
using its aforementioned action on a CAT(0) square complex.
Theorem C. The group Tame(SL2 (C)) is acylindrically hyperbolic. In particular, it is
SQ-universal and admits free normal subgroups.
The article is organised as follows. In Section 1, after recalling standard definitions and
results about acylindrical hyperbolicity, we introduce über-contractions and prove Theorem
A. In Section 2, we introduce the Strong Concatenation Property and prove Theorem B.
Finally, Section 3 deals with the acylindrical hyperbolicity of Tame(SL2 ) by means of its
action on the CAT(0) square complex introduced by Bisi–Furter–Lamy.
Acknowledgement. We gratefully thank S. Lamy for remarks on a first version of this
article, as well as I. Chatterji for many discussions and suggestions about this article. This
work was partially supported by the European Research Council (ERC) grant no. 259527
of G. Arzhantseva and by the Austrian Science Fund (FWF) grant M 1810-N25.
1
A criterion for acylindrical hyperbolicity via über-contractions
In this section, we give a tractable criterion implying acylindrical hyperbolicity for groups
acting on (not necessarily locally compact) geodesic metric spaces.
5
A. Martin
1.1
On the acylindrical hyperbolicity of Tame(SL2 (C))
Contracting properties of quasi-lines
As many others, our criterion relies on the existence of of a group element whose orbits
possesses hyperbolic-like features. We start by recalling various “contracting” poperties of
a quasi-line in a metric space.
Definition 1.1. Let X be a metric space and Λ a quasi-line of X, i.e. the image by a
quasi-isometric embedding of the real line. For a closed subset Y of X, we denote by πΛ (Y )
the set of points of Λ realising the distance to Y , called the closest-point projection of Y
on Λ.
The quasi-line Λ is Morse if for every K, L ≥ 0 there exists a constant C(K, L) such
that any (K, L) quasi-geodesic with endpoints in Λ stays in the C(K, L)-neighbourhood
of Λ. We say that an isometry of X is Morse if it is a hyperbolic isometry, i.e. its has
quasi-isometrically embedded orbits, and if one (hence every) of its orbits is Morse.
The quasi-line Λ is strongly contracting if there exists a constant C such that every ball
of X disjoint from Λ has a closest-point projection on Λ of diameter at most C. We say
that an isometry of X is strongly contracting if it is a hyperbolic isometry and if one (hence
every) of its orbits is strongly contracting.
Remark 1.2. A strongly contracting quasi-geodesic is Morse.
We now introduce a strengthening of the notion of strongly contracting isometry, which
is central in this article.
Definition 1.3 (system of checkpoints, über-contracting isometry). Let X be a geodesic
metric space, let h be an isometry of X with quasi-isometrically embedded orbits. A system
of checkpoints for h is the
S data of a finite subset S of X, an error constant L ≥ 0, and a
quasi-isometry f : Λ := i∈Z hi S → R such that the following holds:
Let x, y be points of X and x0 , y 0 be projections on Λ of x, y respectively. For every
checkpoint Si := hi S, i ∈ Z such that:
• Si coarsely separates x0 and y 0 , that is, f (x0 ) and f (y 0 ) lie in different unbounded
connected components of R \ f (Si ),
• Si is at distance at least L from x0 and y 0 ,
then every geodesic between x and y meets Si .
A hyperbolic isometry h of X is über-contracting, or is an über-contraction, if it admits
a system of checkpoints.
Example 1.4. If X is a simplicial tree and h is a hyperbolic isometry, the h-translates of
any vertex on the axis of h yields a system of checkpoints.
Example 1.5. By standard arguments of hyperbolic geometry, if X is a δ-hyperbolic
geodesic metric space and h is a hyperbolic isometry, the h-translates of any ball of radius
2δ yields a system of checkpoints.
6
A. Martin
On the acylindrical hyperbolicity of Tame(SL2 (C))
We mention a couple of immediate properties:
Remark 1.6. If h is an über-contraction and Λ the hhi-orbit of some finite subset, then
there is coarsely well-defined closest-point projection on Λ, as the diameter of the set of
projections of a given point is uniformly bounded above.
Remark 1.7. An über-contracting isometry is strongly contracting. In particular, it is
Morse by Remark 1.2.
1.2
Acylindrical hyperbolicity in presence of über-contractions
We start by recalling some standard definitions.
Definition 1.8 (acylindricity, acylindrically hyperbolic group). Let G be a group acting
on a geodesic metric space X. We say that the action is acylindrical if for every r ≥ 0 there
exist constants L(r), N (r) ≥ 0 such that for every points x, y of X at distance at least L(r),
there are at most N (r) elements h of G such that d(x, hx), d(y, hy) ≤ r.
A non virtually cyclic group is acylindrically hyperbolic if it admits an acylindrical action
with unbounded orbits on a hyperbolic metric space.
Our goal is to obtain a criterion for acylindrical hyperbolicity for groups admitting übercontractions under additional assumptions. We start by recalling some standard criterion.
Definition 1.9. Let G be a group acting on a geodesic metric space X. Let g be an element
of G of infinite order with quasi-isometrically embedded orbits. We say that g satisfies the
WPD condition if for every r ≥ 0 and every point x of X, there exists an integer m0 such
that there exists only finitely many elements h of G such that d(x, hx), d(g m0 x, hg m0 x) ≤ r.
Remark 1.10. If g is a Morse element, then the WPD condition is equivalent to the
following strengthening, by [22, Lemma 2.7]: For every r ≥ 0 and every point x of X, there
exists an integer m0 such that for every m ≥ m0 , there exists only finitely many elements
h of G such that d(x, hx), d(g m x, hg m x) ≤ r.
Let us now recall a useful criterion of Bestvina–Bromberg–Fujiwara to prove the acylindrical hyperbolicity of a group:
Theorem 1.1 ( [1, Theorem H]). Let G be a group acting by isometries on a geodesic
metric space X and let g be an infinite order element with quasi-isometrically embedded
orbits. Assume that the following holds:
• g is a strongly contracting element.
• g satisfies the WPD condition.
Then G is either virtually cyclic or acylindrically hyperbolic.
7
A. Martin
On the acylindrical hyperbolicity of Tame(SL2 (C))
We are now ready to state our main criterion:
Theorem 1.2 (Criterion for acylindrical hyperbolicity). Let G be a group acting by isometries on a geodesic metric space X. Let g be an infinite order element with quasi-isometrically
embedded orbits. Assume that the following holds:
• g is über-contracting (with respect to a system of checkpoints (hi S)i∈Z ),
• g satisfies the following weakening of the WPD condition: There exists a constant m0
such that for every point s ∈ S and every m ≥ m0 , only finitely many elements of G
fix pointwise s and g m s.
Then G is either virtually cyclic or acylindrically hyperbolic.
Before starting the proof, let us recall an elementary property of coarse projections on
strongly contracting quasi-lines:
Lemma 1.11 (coarsely Lipschitz projection). Let Λ be a strongly contracting quasi-line,
and C a constant such that balls disjoint from γ project on Λ to subsets of diameter at most
C. Let x, y two points of X and let π(x), π(y) be two closest-point projections on Λ. Then
d(π(x), π(y)) ≤ max(C, 4d(x, y)).
Proof. Let x, y be two points of X and consider the ball of radius d(x, y) around x. Then
either this ball is disjoint from Λ, in which we case the strongly contracting assumption
immediately implies that d(π(x), π(y)) ≤ C, or it contains a point of Λ, in which case the
distance from x (respectively y) to any of its projection on Λ is at most d(x, y) (respectively
2d(x, y)), and thus d(π(x), π(y)) ≤ 4d(x, y).
Before proving Proposition 1.2, we present a key lemma, which reduces the proof of the
WPD condition to the WPD condition for points of the checkpoints.
Lemma 1.12. Let g be an über-contracting element with system of checkpoints (Si )i∈Z and
error constant L ≥ 0. Assume that there exists a constant m0 such that for every point
s ∈ S and every m ≥ m0 , only finitely many elements of G fix pointwise s and g m s. Then
g satisfies the WPD condition.
Proof. Fix r > 0 and x ∈ X. We want to show that there exists an integer m ≥ 1 such
that the coarse stabiliser Stabr (x, g m x), that is, the set of group elements gi such that
d(x, gi x), d(g m x, gi g m x) ≤ r, is finite.
S
Let C be a constant such that balls disjoint from Λ := i Si project on Λ to subsets of
diameter at most C.
Let m := 8r + 2C + m0 (|S| + 1) + 2L, and consider group elements gi in Stabr (x, g m x).
Let Qx be a geodesic path between x and g m x. The constant m has been chosen so that
closest-point projections on Λ of gi x and gi g m x are at distance at least m0 |g|(|S|+1)+2L by
8
A. Martin
On the acylindrical hyperbolicity of Tame(SL2 (C))
Lemma 1.11. Let S1 , . . . , Sk be the checkpoints of Λ coarsely separating πΛ (x) and πΛ (g m x)
and which are at distance at least L from πΛ (x) and πΛ (g m x). Note that k ≥ m0 (|S|+1) by
construction. In particular, for each gi ∈ Stabr (x, g m x) there exist two distinct checkpoints
of Λ, say S, S 0 ∈ {S1 , . . . , Sk }, and two points s ∈ S and s0 ∈ S 0 such that s0 = g m s with
m ≥ m0 , such that gi Qx contains s and s0 .
This allows us to define a map φ (using the same notation as in the previous paragraph):
Stabr (x, g m x) → ∪1≤j≤k Sj × ∪1≤j≤k Sj × Qx × Qx
0
gi 7→ (s, s
, gi−1 s, gi−1 s0 ).
(1)
(2)
Notice that the target is finite. Let F be the preimage of an element in the image of
φ. Choose an element f0 ∈ F and consider the set F f0−1 of elements of G of the form
f f0−1 , f ∈ F . Then elements of F f0−1 fix both s and s0 by construction. As s0 = g m s with
m ≥ m0 , it follows that F f0−1 , and hence F , is finite by the weak WPD condition. It now
follows that Stabr (x, g m x) is finite.
Proof of Proposition 1.2. Let g be an element of G as in the statement of the Theorem and
let (Si )i∈Z and L ≥ 0 be a system of checkpoints as in Definition 1.3. By [1, Theorem H], it
is enough to show that g is strongly contracting and satisfies the WPD condition. The first
part follows directly from Remark 1.7. Since there are only finitely many elements in ∪Si
modulo the action of hgi, we choose m0 ≥ 1 such that for every point s ∈ ∪Si such that for
every m > m0 , there exists only finitely many elements fixing both s and g m s. The second
part follows directly from Lemma 1.12
2
A local criterion
In this section, we give a method for constructing über-contractions for groups acting on
polyhedral complexes with some vertices having unbounded links. This allows us to give a
second, more local, criterion for proving the acylindrical hyperbolicity of such groups.
2.1
The Strong Concantenation Property
Definition 2.1 (angle). Let X be a polyhedral complex. The angle at a vertex v between
two edges of X containing v is their (possibly infinite) distance in the 1-skeleton of the link
of v (where each edge of that graph is given length 1).
For two geodesic paths γ, γ 0 starting at a vertex v, the angle at v between γ and γ 0 ,
denoted ∠v (γ, γ 0 ), is the angle at v between the associated edges of γ and γ 0 .
Finally, for vertices v, w, w0 of X, we define the angle ∠v (w, w0 ) as the minimum of the
angle ∠v (γ, γ 0 ) where γ (respectively γ 0 ) ranges over the combinatorial geodesics between
v and w (respectively w0 ).
The following definition mimicks the strong properties of geodesics in CAT(0) spaces.
9
A. Martin
On the acylindrical hyperbolicity of Tame(SL2 (C))
Definition 2.2. We say that a complex satisfies the Strong Concatenation Property with
constants (A, R) if the following two conditions hold:
• Let γ1 , γ2 be to geodesics of X meeting at a vertex v. If ∠v (γ1 , γ2 ) > A, then γ1 ∪ γ2
is a geodesic of X.
• Let γ be a geodesic segment of X, v a vertex of γ. Let x, y be two vertices of X,
π(x), π(y) be projections of x, y respectively on γ such that π(x) and π(y) are at
distance strictly more than R from v. If γ makes an angle greater than A at v, then
every geodesic between x and y contains v.
Note that this property has the following immediate consequence, which allows for the
construction of many über-contractions:
Lemma 2.3 (Local criterion for über-contractions). Let X be a simply connected complex
with the Strong Concatenation Property with constants (A, R). Then there exists a constant
C such that the following holds:
Let h be a hyperbolic isometry of X with axis γ and assume that for some vertex v of
γ, the angle made by γ at v is at least C. Then h is über-contracting with respect to the
system of checkpoints (hi v)i∈Z .
This allows to give a local criterion to show the acylindrical hyperbolicity of a group:
Proposition 2.4 (Link criterion for acylindrical hyperbolicity). Let X be a simply connected hyperbolic complex with the Strong Concatenation Property, together with an action
by isometries of a group G. Assume that there exists a vertex v of X and a group element
g such that:
1) the action of Gv on the link of v has unbounded orbits (for the simplicial metric on
lk(v) where every edge has length 1),
2) the intersection Gv ∩ gGv g −1 is finite.
Then G is either virtually cyclic or acylindrically hyperbolic.
Remark 2.5. In the first item, one could require that the action of Gv on the link of
v has large orbits, where large is defined in terms of constants appearing in the Strong
Concatenation Property.
Proof of Proposition 2.4. Choose a vertex v and a group element g satisfying 1) and 2).
Choose an integer N greater than 8δ, where δ is the hyperbolicity constant of X. Let
P be a geodesic between v and gv. By condition 1), choose an element h ∈ Gv such
that P and ghP make an angle at least A. Then the element gh is an über-contraction,
with ((gh)n v)n∈Z as a system of gates.
for every i, (gh)i P and (gh)i+1 P make an
S Indeed,
angle of at least A. In particular, i∈Z (gh)i P is a geodesic by the Strong Concatenation
10
A. Martin
On the acylindrical hyperbolicity of Tame(SL2 (C))
Property and thus gh is a hyperbolic isometry with axis γ. Such an axis turns out to be
über-contracting by Lemma 2.3.
Moreover, gh satisfies the so-called WPD condition by condition 2) and Lemma 1.12.
Hence the result follows from Proposition 1.2.
2.2
Complexes with the Strong Concatenation Property
We now give two simple properties that imply the Strong Concatenation Property. The
first one is reminiscent of features of CAT(0) spaces.
Definition 2.6 (bounded angle of view). We say that a simply connected complex X has
an angle of view of at most A ≥ 0 if there exists a constant A such that for every vertices
x, y of X and every vertex z of X which does not lie on a geodesic between x and y, we
have ∠z (x, y) ≤ A. We say that X has a bounded angle of view if there exists a constant
A ≥ 0 such that X has an angle of view of at most A.
Example 2.7. CAT(0) spaces have an angle of view of at most π.
This weak condition seems to be satisfied by many natural examples of complexes that
are non-positively curved in a broad sense. For instance, it follows from the classification
of geodesic triangles by Strebel [23] that simply connected C 0 (1/6) complexes in the sense
of McCammond–Wise [18] have a bounded angle of view. The result also holds for systolic
complexes.
Lemma 2.8. Simply connected complexes with an angle of view at most A have the Strong
Concatenation Property with constants (3A, 0).
Proof. From the bounded angle of view condition, it is immediate that the concatenation of
two geodesics making an angle greater than A is again a geodesic. Let us now consider the
case of a geodesic segment γ of X, and let v be a vertex of γ. Let x, y be two vertices of X,
π(x), π(y) be projections of x, y respectively on γ such that π(x) and π(y) are distinct from
v. If there exists a geodesic γx,y between x and y not containing v, we choose geodesics
γv,x , γv,y between v and x (respectively between v and y), geodesics γx , γy between x and
a point x0 of πγ (x) (respectively between y and a pointy 0 of πγ (y)), and let γv,x0 , γv,y0 be
the sub-segments of γ between v and x0 (respectively between v and y 0 ). By assumption, v
does not belong to γx ∪ γx,y ∪ γy , so we get
∠v (γ) ≤ ∠v (γv,x0 , γv,x ) + ∠v (γv,x , γv,y ) + ∠v (γv,y , γv,y0 ) ≤ 3A,
which concludes the proof.
Remark 2.9. If X has a bounded angle of view, it is not necessary to require the
S space to
be hyperbolic in the statement of Proposition 2.4, as one can check directly that i∈Z (gh)i P
is a geodesic in the proof of Proposition 2.4.
11
A. Martin
On the acylindrical hyperbolicity of Tame(SL2 (C))
The second property, more algorithmic in nature, is particularly suitable for hyperbolic
complexes which are not known to have such a rich combinatorial geometry.
Definition 2.10. We say that a simply connected polyhedral complex satisfies an isoperimetric inequality if there exists a function ϕ : N → N such that every embedded loop of
length n can be filled by a combinatorial disc of area at most ϕ(n).
Lemma 2.11. Simply connected hyperbolic complexes satisfying an isoperimetric inequality
have the Strong Concatenation Property.
Proof. Let δ be the hyperbolic constant of X, and ϕ its isoperimetric function.
We start by the first condition. Let γ1 , γ2 be two geodesics of X meeting at a vertex v,
and γ a geodesic between the two endpoints of the path γ1 ∪ γ2 . For i ∈ {1, 2}, if |γi | > 2δ,
choose the point xi ∈ γi at distance 2δ from v. By the hyperbolicity condition, xi is at
distance at most δ from the other two sides of the geodesic triangle γ1 ∪ γ2 ∪ γ, so choose
a point ui ∈ γ1 ∪ γ2 ∪ γ at distance at most δ from vi , and which is not on γi . Choose a
geodesic path τui ,vi between ui and vi . If |γi | ≤ 2δ, set vi = ui to be the endpoint of γi
distinct from v.
If γ does not contain v, then one could extract from the loop (γ1 )v,v1 ∪ τv1 ,u1 ∪ (γ)u1 ,u2 ∪
τu2 ,v2 ∪ (γ2 )v2 ,v an embedded loop of length at most 12δ containing a neighbourhood of v
in γ1 ∪ γ2 . In particular, this implies that ∠v (γ1 , γ2 ) ≤ ϕ(12δ).
Let us now show the second condition. Let γ be a geodesic segment of X, v a vertex of
γ. Let x, y be two vertices of X, π(x), π(y) be projections of x, y respectively on γ, such
that π(x) and π(y) are at distance at least 100δ from v. Let us assume that there exists a
geodesic γx,y between x and y which does not contain v.
Let γ 0 be the sub-segment of γ centred at v and of length 100δ. Using the theorem
of approximation by trees in a hyperbolic space [9, Thm. 1 Ch. 8] , it follows that there
0
0
exists a sub-segment γx,y
of γx,y such that γ 0 and γx,y
are at Hausdorff distance at most 8δ.
0
0
Projecting the endpoints of γ to γx,y , it is thus possible to construct an embedded loop of
X, of length at most 300δ, which contains γ 0 . By filling this loop, it follows that the angle
of γ at v is bounded above by ϕ(300δ).
Thus X satisfies the Strong Concatenation Property with constants (ϕ(300δ), 100δ).
Remark 2.12. Note that non locally finite hyperbolic complexes do not necessarily satisfy
any isoperimetric inequality, let alone a linear one. For instance, consider the suspension of
the simplicial real line, with its triangular complex structure. Given any number n, there
exists a geodesic bigon of length 4 between the two apices that require at least n triangles
to be filled.
12
A. Martin
3
On the acylindrical hyperbolicity of Tame(SL2 (C))
Application: The tame automorphism group of SL2 (C)
Here we use the Criterion 2.4 to study the subgroup Tame(SL2 ) ⊂ Bir(P3 (C)). We prove
the following:
Theorem 3.1. The group Tame(SL2 ) is acylindrically hyperbolic.
Recall that Tame(SL2 ) can be defined (see [3, Proposition 4.19]) as the subgroup
Tameq (C4 ) of Aut(C4 ) generated by the orthogonal group O(q) associated to the quadratic
form q(x1 , x2 , x3 , x4 ) = x1 x4 − x2 x3 and the subgroup consisting of automorphisms of the
form
(x1 , x2 , x3 , x4 ) 7→ (x1 , x2 + x1 P (x1 , x3 ), x3 , x4 + x3 P (x1 , x3 )), P ∈ C[x1 , x3 ].
This group was studied in [3] from the point of view of its action on a CAT(0) square
complex X, which we now describe.
Vertices of X. To each element (f1 , f2 , f3 , f4 ) of Tame(SL2 (C)), one associates:
• a vertex [f1 ], said of type 1, corresponding to the orbit C∗ · f1 ,
• a vertex [f1 , f2 ], said of type 2, corresponding to the orbit GL2 (C) · (f1 , f2 ),
• a vertex [f1 , f2 , f3 , f4 ], said of type 3, corresponding to the orbit O(q) · (f1 , f2 , f3 , f4 ).
Edges of X. To each element (f1 , f2 , f3 , f4 ) of Tame(SL2 (C)), one associates:
• an edge joining the type 1 vertex [f1 ] and the type 2 vertex [f1 , f2 ].
• an edge joining the type 2 vertex [f1 , f2 ] and the type 3 vertex [f1 , f2 , f3 , f4 ].
Squares of X. To each element (f1 , f2 , f3 , f4 ) of Tame(SL2 (C)), one associates a square
with vertex set being, in cyclic order, [f1 ], [f1 , f2 ], [f1 , f2 , f3 , f4 ], [f1 , f3 ].
Action of Tame(SL2 (C)). The group Tame(SL2 (C)) acts by isometries on X as follows.
For every element (f1 , f2 , f3 , f4 ) of Tame(SL2 (C)) and every g of Tame(SL2 (C)), we set:
• g · [f1 ] := [f1 ◦ g −1 ],
• g · [f1 , f2 ] := [f1 ◦ g −1 , f2 ◦ g −1 ],
• g · [f1 , f2 , f3 , f4 ] := [f1 ◦ g −1 , f2 ◦ g −1 , f3 ◦ g −1 , f4 ◦ g −1 ].
A central result of [3] is the following:
Theorem 3.2. The square complex X is CAT(0) and hyperbolic.
13
A. Martin
On the acylindrical hyperbolicity of Tame(SL2 (C))
We recall further properties of the action that we will need.
Proposition 3.1 ( [3, Lemma 2.7] ). The action is transitive on the squares of X, and the
pointwise stabiliser of a given square is conjugate to the subgroup
{(x1 , x2 , x3 , x4 ) 7→ (ax1 , b(x2 +cx1 ), b−1 (x3 +dx1 ), a−1 (x4 +cx3 +dx2 )), a, b, c, d ∈ C, ab 6= 0}
of Tame(SL2 ).
Proposition 3.2 ( [3, Propositions 3.6, 3.7 and 4.1] ). The link of a vertex of type 1 has
infinite diameter. Moreover, the action of the stabiliser of such a vertex on its link has
unbounded orbits.
We want to construct a super-contraction for this action. To that end, we will use the
following hyperbolic isometry considered in [3]:
Lemma 3.3 ( [3, Example 6.2] ). The element g ∈ Tame(SL2 ) defined by
g(x1 , x2 , x3 , x4 ) =:= (x4 + x3 x21 + x2 x21 + x51 , x2 + x31 , x3 + x31 , x1 )
acts hyperbolically on X. More precisely, there exists a 4 × 4 grid isometrically embedded
in X and a vertex v of type 1, such that the vertices v, gv, g 2 v are as in Figure 1.
v
C
gv
gC
g2v
Figure 1: The 4 × 4 grid K isometrically embedded in X, together with vertices v, gv and
g 2 v (black dots), and squares C and gC (shaded).
14
A. Martin
On the acylindrical hyperbolicity of Tame(SL2 (C))
Proof of Theorem 3.1. Let v, K be the vertex and 4 × 4 grid of X mentioned in Lemma 3.3.
We show that we can apply the Criterion 2.4 to g 2 and v. Note that Item 1) of Criterion
2.4 follows from Proposition 3.2.
We now show that v and g 2 v have a finite common stabiliser, that is, that their stabilisers
intersect along a finite subgroup. Let C be the top-left square of K, as indicated in Figure
1. We start by showing that C and gC have a finite common stabiliser. Indeed, Stab(C) is
conjugated to the subgroup defined by elements of the form
f (x1 , x2 , x3 , x4 ) = (ax1 , b(x2 +cx1 ), b−1 (x3 +dx1 ), a−1 (x4 +cx3 +dx2 )), a, b, c, d ∈ C, ab 6= 0,
by Proposition 3.1. Now an equation of the form gf = f 0 g, with f, f 0 of the previous form,
with coefficients a, b, c, d and a0 , b0 , c0 , d0 respectively, yields the following equation, when
isolating the first coordinate:
a−1 (x4 +cx3 +dx2 )+a2 b−1 x21 (x3 +dx2 )+a2 bx21 (x2 +cx1 )+a5 x5 = a0 (x4 +x3 x21 +x2 x21 +x51 ).
Isolating the various monomials, we successively get a6 = 1, b6 = 1 and c = d = 0
(and analogous equalities for a0 , b0 , c0 , d0 ), hence Stab(C) ∩ gStab(C)g −1 is finite, and thus
C and gC have a finite common stabiliser. This in turn implies that v and g 2 v have a finite
common stabiliser. Indeed, the combinatorial interval between v and g 2 v is exactly K, as
combinatorial intervals embed isometrically in R2 with its square structure by [5, Theorem
1.16]. Thus, up taking a finite index subgroup, elements fixing v and g 2 v will fix pointwise
K, and in particular C and gC.
Since CAT(0) spaces a have bounded angle of view by Example 2.7, one can thus apply
Criterion 2.4 to conclude.
REFERENCES
[1] M. Bestvina, K. Bromberg, and K. Fujiwara. Constructing group actions on quasi-trees and applications to mapping class groups. Publ. Math. Inst. Hautes Études Sci., in press, 2014.
[2] M. Bestvina and M. Feighn. A hyperbolic Out(Fn )-complex. Groups Geom. Dyn., 4(1):31–58, 2010.
[3] C. Bisi, J.-P. Furter, and S. Lamy. The tame automorphism group of an affine quadric threefold acting
on a square complex. Journal de l’Ecole Polytechnique, 1:161–223, 2014.
[4] B. H. Bowditch. Tight geodesics in the curve complex. Invent. Math., 171(2):281–300, 2008,
doi:10.1007/s00222-007-0081-y.
[5] J. Brodzki, S. J. Campbell, E. Guentner, G. A. Niblo, and N. J. Wright. Property A and CAT(0) cube
complexes. J. Funct. Anal., 256(5):1408–1431, 2009.
[6] S. Cantat. Sur les groupes de transformations birationnelles des surfaces. Ann. of Math. (2),
174(1):299–340, 2011.
[7] S. Cantat and S. Lamy. Normal subgroups in the Cremona group. Acta Math., 210(1):31–94, 2013.
With an appendix by Yves de Cornulier.
[8] P.-E. Caprace and D. Hume. Orthogonal forms of Kac-Moody groups are acylindrically hyperbolic.
Ann. Inst. Fourier, in press, 2015.
15
A. Martin
On the acylindrical hyperbolicity of Tame(SL2 (C))
[9] M. Coornaert, T. Delzant, and A. Papadopoulos. Géométrie et théorie des groupes, volume 1441
of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1990. Les groupes hyperboliques de
Gromov. [Gromov hyperbolic groups], With an English summary.
[10] F. Dahmani, V. Guirardel, and D. Osin. Hyperbolically embedded subgroups and rotating families in
groups acting on hyperbolic spaces. Mem. Amer. Math. Soc., in press, 2012.
[11] J. Déserti. Sur les automorphismes du groupe de Cremona. Compos. Math., 142(6):1459–1478, 2006.
[12] J. Déserti. Le groupe de Cremona est hopfien. C. R. Math. Acad. Sci. Paris, 344(3):153–156, 2007.
[13] D. Gruber and A. Sisto. Infinitely presented graphical small cancellation groups are acylindrically
hyperbolic. arXiv:1408.4488v2, 2014.
[14] H. W. E. Jung. Über ganze birationale Transformationen der Ebene. J. Reine Angew. Math., 184:161–
174, 1942.
[15] S. Lamy and S. Vénéreau. The tame and the wild automorphisms of an affine quadric threefold. J.
Math. Soc. Japan, 65(1):299–320, 2013.
[16] A. Lonjou. Non simplicité du groupe de Cremona sur tout corps. arXiv:1503.03731v2, 2015.
[17] A. Martin. Acylindrical actions on CAT(0) square complexes. arXiv: 1509.03131, 2015.
[18] J. P. McCammond and D. T. Wise. Fans and ladders in small cancellation theory. Proc. London Math.
Soc. (3), 84(3):599–644, 2002, doi:10.1112/S0024611502013424.
[19] A. Minasyan and D. Osin. Acylindrically hyperbolic groups acting on trees. Math. Ann., in press,
2015.
[20] D. Osin. Acylindrically hyperbolic groups. Trans. Amer. Math. Soc., in press, 2015.
[21] Z. Sela.
Acylindrical accessibility for groups.
Invent. Math., 129(3):527–565, 1997,
doi:10.1007/s002220050172.
[22] A. Sisto. Contracting elements and random walks. to be published in Crelle, arXiv:1112.2666v2, 2013.
[23] R. Strebel. Appendix. Small cancellation groups. In Sur les groupes hyperboliques d’après Mikhael
Gromov (Bern, 1988), volume 83 of Progr. Math., pages 227–273. Birkhäuser Boston, Boston,
MA, 1990.
[24] W. van der Kulk. On polynomial rings in two variables. Nieuw Arch. Wiskunde (3), 1:33–41, 1953.
16
| 4math.GR
|
ROOT DATA WITH GROUP ACTIONS
arXiv:1707.01935v1 [math.RT] 6 Jul 2017
JEFFREY D. ADLER AND JOSHUA M. LANSKY
Abstract. Suppose k is a field, G is a connected reductive algebraic
k-group, T is a maximal k-torus in G, and Γ is a finite group that acts on
(G, T ). From the above, one obtains a root datum Ψ on which Gal(k)×Γ
acts. Provided that Γ preserves a positive system in Ψ, not necessarily
invariant under Gal(k), we construct an inverse to this process. That is,
given a root datum on which Gal(k) × Γ acts, we show how to construct
a pair (G, T ), on which Γ acts as above.
Although the pair (G, T ) and the action of Γ are canonical only up
to an equivalence relation, we construct a particular pair for which G
is k-quasisplit and Γ fixes a Gal(k)-stable pinning of G. Using these
choices, we can define a notion of taking “Γ-fixed points” at the level
of equivalence classes, and this process is compatible with a general
“restriction” process for root data with Γ-action.
1. Introduction
Let k be a field with separable closure ksep . Let Γ be a finite group.
Suppose Ψ is a (reduced) based root datum on which the absolute Galois
group Gal(k) acts. Then it is well known ([7, Theorem 6.2.7]) that there
exists a connected, reductive, k-quasisplit k-group G, uniquely determined
up to k-isomorphism, such that the root datum of G (with respect to a
maximal k-torus contained in a Borel k-subgroup) is isomorphic to Ψ and
carries the same action of Gal(k). We generalize this result in two directions.
(A) Suppose G is a connected reductive k-group, and T is an arbitrary
maximal torus. Then the root datum Ψ(G, T ) carries an action of
Gal(k). We show that one can reverse this process. That is, given a
root datum Ψ with an action of Gal(k), one can obtain a pair (G, T )
that gives rise to Ψ. In general, the pair (G, T ) need not be uniquely
determined up to k-isomorphism. However, we can always choose G to
be k-quasisplit, and all possibilities for G must be k-inner forms of each
other.
(B) Now suppose that Γ acts on a pair (G, T ) as above via k-automorphisms.
Then Γ acts on the root datum Ψ(G, T ), and the actions of Γ and
Gal(k) commute. We show that one can reverse this process under mild
Date: February 22, 2018.
2010 Mathematics Subject Classification. Primary 20G15, 20G40. Secondary 20C33.
Key words and phrases. Reductive algebraic group, root datum, quasi-semisimple
automorphisms.
1
2
JEFFREY D. ADLER AND JOSHUA M. LANSKY
conditions. That is, suppose that Ψ is a root datum with an action of
Gal(k) × Γ. Assume that Γ (but not necessarily Gal(k)) preserves a
base. Then one can obtain a pair (G, T ) as above, carrying an action
of Γ. That is, under appropriate conditions, we can lift an action of Γ
from a root datum to a reductive group. Moreover, one can choose G
to be k-quasisplit, and can choose the action of Γ to preserve a pinning.
The above are all contained in our main result, Theorem 1. In order to
state it more precisely, let us consider the collection of abstract root data Ψ
that carry an action of Gal(k) × Γ such that the action of Γ stabilizes a base
for Ψ. We consider two data Ψ and Ψ′ with such actions to be equivalent
if there is a Gal(k) × Γ-equivariant isomorphism Ψ −→ Ψ′ . Let RΓ denote
the set of equivalence classes of reduced data with such actions.
Let G be a connected reductive k-group and T ⊆ G a maximal k-torus.
Suppose there exists some Borel subgroup B ⊆ G (not necessarily defined over k) containing T , and a homomorphism ϕ from Γ to the group
Autk (G, B, T ) of k-automorphisms of G stabilizing T and B. Suppose G′ , T ′ ,
and ϕ′ are defined similarly. We say that the triples (G, T, ϕ) and (G′ , T ′ , ϕ′ )
are equivalent if there exists an isomorphism ν : G −→ G′ whose restriction
gives a Γ-equivariant k-isomorphism T −→ T ′ . (In this situation, ν must be
an inner twisting by [5, §3.2].) Let TΓ be the set of equivalence classes of
such triples (G, T, ϕ).
A triple (G, T, ϕ) as above naturally determines a root datum with appropriate actions of Gal(k) and Γ, hence an element of RΓ . It is easily seen
that if (G′ , T ′ , ϕ′ ) and (G, T, ϕ) are equivalent, then they determine the same
class in RΓ . Hence we have a natural map rΓ : TΓ −→ RΓ .
Our main result is the following:
Theorem 1. The map rΓ : TΓ −→ RΓ is a bijection.
We also prove a generalization of a second well-known result. Suppose
char k = 0, and Γ is a finite, cyclic group acting algebraically on a connected
reductive group G. Then Γ must fix a Borel-torus pair (B, T ) in G. If Γ
fixes a pinning, then the root system of the connected part Ḡ := (GΓ )◦ of
the group of fixed points is obtained as follows. The set of restrictions of
roots of G from T to T̄ := (T Γ )◦ is a root system, not necessarily reduced,
but there is a preferred way to choose a maximal reduced subsystem.
We generalize the above result in several directions.
(C) Instead of assuming char k = 0, we impose the weaker condition that Γ
fix a Borel-torus pair.
(D) The group Γ need not be cyclic.
(E) We describe the root datum, not just the root system, of Ḡ with respect
to T̄ .
The above are all contained in Theorem 2. To state this result more
precisely, suppose that the triple (G, T, ϕ) represents an element of TΓ .
Then we know [1, Proposition 3.5] that Ḡ is a connected reductive k-group,
ROOT DATA WITH GROUP ACTIONS
3
and T̄ is a maximal k-torus in Ḡ. Thus, if we let “1” represent the map
from the trivial group 1 to Aut(G), then the triple (Ḡ, T̄ , 1) represents an
element of T1 . The equivalence class of (G, T, ϕ) does not determine that
of (Ḡ, T̄ , 1), or even the ksep -isomorphism class of Ḡ. Nonetheless, we can
obtain a well-defined map TΓ −→ T1 as follows: From Remark 25, we will
see that every class in TΓ contains a triple (G, T, ϕ) such G is k-quasisplit
and ϕ fixes a Gal(k)-invariant pinning. Use this choice of triple to define Ḡ
and T̄ , and it is straightforward to show that our choices determine Ḡ and
T̄ up to k-isomorphism.
Suppose that the root datum Ψ represents an element of RΓ . We will see
in §2 that the action of Γ on Ψ allows us to construct a “restricted” root
datum Ψ̄ that has a preferred choice of reduced subdatum. We thus obtain
a map RΓ −→ R1 .
Theorem 2. Our maps TΓ −→ T1 and RΓ −→ R1 above are compatible with the maps of Theorem 1, in the sense that the following diagram
commutes:
rΓ
// RΓ
TΓ
T1
r1
// R1
We prove both theorems in §4.
2. Restrictions of root data
Let Ψ = (X ∗ , Φ, X∗ , Φ∨ ) be a root datum. (We do not assume that Φ is
reduced.) Let Γ denote a finite group of automorphisms of Ψ. We assume
that there exists a Γ-stable set ∆ of simple roots in Φ. Let V ∗ = X ∗ ⊗ Q and
V∗ = X∗ ⊗ Q. Let i∗ denote the quotient map from V ∗ to its space V̄ ∗ := VΓ∗
of Γ-coinvariants. From [1, §2], there is an embedding ι : V̄ ∗ −→ V ∗ with
image V ∗ Γ given by
1 X
ι(v̄) =
γv,
|Γ|
∗
γ∈Γ
V̄ ∗ . Let
where v is any preimage in V of v̄ ∈
X̄ ∗ and Φ̄ denote the images
∗
∗
∗
of X and Φ under i . Then X̄ is the torsion-free part of the module XΓ∗
of Γ-coinvariants of X ∗ . It is straightforward to see that X̄ ∗ and X̄∗ := X∗Γ ,
and thus V̄ ∗ and V̄∗ := V∗Γ , are in duality via the pairing given by hx̄, λ̄i :=
hιx̄, i∗ λ̄i, where i∗ : X̄∗ −→ X∗ is the inclusion map. With respect to these
pairings, i∗ is the transpose of i∗ .
For each β ∈ Φ, let wβ denote the automorphism of X ∗ defined by
wβ (x) = x − hx, β ∨ iβ.
Let W denote the Weyl group of Ψ, i.e., the (finite) subgroup of Aut(X ∗ )
generated by the wβ . Then Γ acts naturally on W , and W acts on X ∗ . The
group W Γ of Γ-fixed elements of W acts on on both V̄ ∗ and X̄ ∗ via the rule
4
JEFFREY D. ADLER AND JOSHUA M. LANSKY
w(i∗ x) := i∗ (w(x)) for w ∈ W Γ and x ∈ X ∗ . Equivalently, for x̄ ∈ X̄ ∗ and
w ∈ W Γ , we have ι(w(x̄)) = w(ιx̄).
Lemma 3 (cf. [8, §1.32(a)]). The natural action of W Γ on X̄ ∗ is faithful.
Proof. Let w be a nontrivial element of W Γ . Then there exists a positive
root β ∈ Φ such that w(β) is negative. Since Γ stabilizes ∆, it follows that
w(γ ·β) = γ ·(wβ) is also negative for every γ ∈ Γ. Thus ι(w(i∗ β)) is a linear
combination of roots in ∆ in which all of the coefficients are nonpositive, so
w(i∗ β) 6= i∗ β.
Notation 4. For each root β ∈ Φ, define a Γ-orbit Ξβ in Φ as in [1, §5].
That is, let Ξβ = Γ · β if this is an orthogonal set. Otherwise, for each
θ ∈ Γ · β, there exists a unique root θ ′ 6= θ in Γ · β such that θ and θ ′ are not
orthogonal. Moreover, θ + θ ′ is a root in Φ and does not belong to Γ · β. In
this case, let Ξβ = {θ + θ ′ | θ ∈ Γ · β}.
Remark 5. Thus, in all cases, Ξβ is an orthogonal Γ-orbit of roots.
Lemma 6. If α ∈ Φ̄, then i∗−1 (α) is a Γ-orbit of roots in Φ.
Proof. This argument is similar to but more general than that given in the
proof of [6, Lemma 10.3.2(ii)]. Suppose β ∈ Φ and i∗ (β) = α. Then clearly
i∗ θ = α for any θ ∈ Γ · β.
Now suppose β ′ ∈ Φ, β ′ 6= β, and i∗ β ′ = α. Since ι(i∗ (β ′ − β)) = 0 and
since Γ preserves ∆, when β ′ − β is written as a linear combination of simple
roots, the coefficients must sum to 0. In particular, β ′ − β ∈
/ Φ. Since β ′
′
∨
cannot be multiple of β, we have that hβ , β i ≤ 0 by standard results about
root systems. Similarly, hβ ′ , θ ∨ i ≤ 0 for all θ 6= β ′ in Γ · β. Therefore,
D
E D
E
X
X
X
E D
β ′ − β,
θ ∨ = i∗ (β ′ − β),
θ ∨ = β ′ − β, i∗
θ∨ ,
θ∈Γ·β
θ∈Γ·β
i∗ (β ′
θ∈Γ·β
′ ∨
and since
− β) = 0, this pairing vanishes. Thus
θ∈Γ·β hβ , θ i =
P
∨
θ∈Γ·β hβ, θ i = 2 or 1, depending on whether or not Γ · β is orthogonal.
(This follows from the properties of root orbits discussed in [1, §5].) Since
hβ ′ , θ ∨ i ≤ 0 for all θ 6= β ′ in Γ · β, it follows that β ′ ∈ Γ · β.
P
For each α ∈ Φ̄, define
α∨ =
|Γ · β| X ∨
ξ ∈ X̄∗ ,
|Ξβ |
ξ∈Ξβ
where β is any element of Φ such that i∗ β = α. The element of X̄∗ defined by
the above formula is independent of the particular choice of β by Lemma 6.
Note that α∨ does indeed lie in X̄∗ since |Γ · β|/|Ξβ | = 1 or 2. Let Φ̄∨ =
{α∨ | α ∈ Φ̄}.
Theorem 7. With the above notation, Ψ̄ := (X̄ ∗ , Φ̄, X̄∗ , Φ̄∨ ) is a root datum.
ROOT DATA WITH GROUP ACTIONS
5
Remark 8. If Ψ comes equipped with an action of Gal(k), and the action of
Γ commutes with that of Gal(k), then it is clear that the action of Gal(k)
preserves Ψ̄.
Proof of Theorem 7. According to [5, §1.1], it suffices to show that
• X̄ ∗ and X̄∗ are in duality (which we have already observed),
• hα, α∨ i = 2 for all α ∈ Φ̄, and
• The automorphisms wα of X̄ ∗ of the form
wα (x̄) = x̄ − hx̄, α∨ iα
(for α ∈ Φ̄)
stabilize Φ̄ and generate a finite subgroup of Aut(X̄ ∗ ).
Let α ∈ Φ̄. Choose β ∈ Φ such that i∗ β = α, and choose ξ0 ∈ Ξβ . Then
we have
hα, α∨ i = hια, i∗ α∨ i
D 1
X
|Γ · β| X ∨ E
ξ
θ,
=
|Γ · β|
|Ξβ |
θ∈Γ·β
ξ∈Ξβ
E
D
X
X
1
=
ξ∨
θ,
|Ξβ |
θ∈Γ·β
ξ∈Ξβ
1 D X ′ X ∨E
=
ξ,
ξ
|Ξβ | ′
ξ ∈Ξβ
(by the definition of Ξβ )
θ∈Ξβ
hξ0 , ξ0∨ i
=
= 2,
(by Remark 5)
as desired.
Now let x̄ ∈ X̄ ∗ , and choose x ∈ X ∗ such that i∗ x = x̄. Then
(9)
hx̄, α∨ i = hx, i∗ α∨ i
D |Γ · β| X E
(by Remark 5)
ξ∨
= x,
|Ξβ |
ξ∈Ξβ
|Γ · β| X
hx, ξ ∨ i.
=
|Ξβ |
ξ∈Ξβ
6
JEFFREY D. ADLER AND JOSHUA M. LANSKY
It follows that
wα (x̄) = x̄ − hx̄, α∨ iα
= i∗ x − hx̄, α∨ ii∗ β
hx̄, α∨ i ∗ X
i
θ
= i∗ x −
|Γ · β|
(by Lemma 6)
θ∈Γ·β
hx̄, α∨ i ∗ X ′
ξ
i
|Γ · β|
ξ ′ ∈Ξβ
X
1 X
= i∗ x −
hx, ξ ∨ ii∗
ξ′
|Ξβ |
ξ∈Ξβ
ξ ′ ∈Ξβ
1 X
X
ξ′ .
hx, ξ ∨ ii∗
= i∗ x −
|Ξβ | ′
= i∗ x −
ξ∈Ξβ
(by the definition of Ξβ )
(by (9))
ξ ∈Ξβ
But by Remark 5, for any ξ ∈ Ξβ ,
1 X
i∗
ξ ′ = i∗ ξ,
|Ξβ | ′
ξ ∈Ξβ
so we have
(10)
X
wα (x̄) = i∗ x −
x, ξ ∨ ξ .
ξ∈Ξβ
Also by Remark 5, the reflections wξ ∈ W for ξ ∈ Ξβ all commute with one
another. If w denotes their product, then
X
hx, ξ ∨ iξ = w(x),
x−
ξ∈Ξβ
so by (10), we have
(11)
wα (x̄) = i∗ (w(x)).
In particular, if α′ ∈ Φ̄, and β ′ ∈ Φ satisfies i∗ β ′ = α′ , then
wα (α′ ) = i∗ (w(β ′ )) ∈ i∗ (Φ) = Φ̄,
so Φ̄ is stable under the action of wα , as desired.
It remains to show that the group W̄ := hwα | α ∈ Φ̄i ⊂ Aut(X̄ ∗ ) is finite.
To accomplish this, we show that W̄ embeds naturally in the finite group
W Γ . By Lemma 3, there is a natural injection
W Γ −→ Aut(X̄ ∗ ).
To construct an embedding W̄ −→ W Γ , it is therefore enough to show that
the image of this injection contains W̄ . Thus, given w̄ ∈ W̄ , we will show
that there exists w ∈ W Γ whose action on X̄ ∗ coincides with that of w̄. It
suffices to prove the existence of w only in the caseQin which w̄ is a reflection
wα through a root α ∈ Φ̄. In this case, let w = ξ∈Ξβ wξ , where β ∈ Φ is
ROOT DATA WITH GROUP ACTIONS
7
such that i∗ β = α. It follows from Remark 5 that w ∈ W Γ , and it follows
from (11) that for any x ∈ X ∗ ,
wα (i∗ x) = i∗ (w(x)) = w(i∗ x).
This establishes the existence of the desired embedding.
Remark 12. If Φ is reduced, then so is the root system Φ̄ constructed above,
unless Φ has an irreducible factor of type A2n whose stabilizer in Γ acts
upon it nontrivially. To see this, it is easy to reduce to the case where Φ is
irreducible and Γ is cyclic (see [1, Proposition 3.5]). The result then follows
from [3, §1.3].
Remark 13. There is a way to choose a maximal reduced subsystem that
we will later see is preferred. Specficially, take only the nondivisible (resp.
nonmultipliable) roots of Φ̄ according as char k is not two (resp. two).
Lemma 14. The map i∗ induces a bijection between the set of Γ-invariant
positive systems in Φ and the set of positive systems in Φ̄.
Proof. Let Π ⊆ Φ be a Γ-invariant positive system. Let Π̄ = i∗ (Π) ⊆ Φ̄.
Then there is some vector v ∈ V∗ such that for every root β ∈ Φ, we have
that hβ, vi =
6 0, and P
hβ, vi > 0 if and only if β ∈ Π. Since Π is Γ-invariant,
we may replace v by γ∈Γ γv, and thus assume that v is Γ-invariant, and so
lies in V̄∗ . Suppose α ∈ Φ̄. Then α = i∗ β for some β ∈ Φ, so hα, vi = hβ, vi.
Thus, hα, vi =
6 0, and hα, vi > 0 if and only if α ∈ Π̄. This shows that Π̄ is
a positive system in Φ̄.
Conversely, suppose that Π̄ ⊆ Φ̄ is a positive system, and let Π = i∗ −1 Π̄.
Then there is some vector v̄ ∈ V̄∗ such that for every root α ∈ Φ̄, we have
that hα, v̄i =
6 0, and hα, v̄i > 0 if and only if α ∈ Π̄. For every root β ∈ Φ,
we have hβ, i∗ vi = hi∗ β, vi, which is never zero, and is positive if and only
if β ∈ Π. Thus, Π ⊂ Φ is a positive system. Since i∗ v is Γ-invariant, so is
Π.
Corollary 15. Let W̄ be the Weyl group of Ψ̄. Then the embedding of W̄
into W Γ in the proof of Theorem 7 is an isomorphism.
Proof. Since W̄ acts simply transitively on the set of positive systems in Φ̄,
and W Γ acts simply transitively on the set of Γ-invariant positive systems
in Φ, the result follows from Lemma 14.
3. From automorphisms of root data to automorphisms of
reductive groups
Let Ψ = (X ∗ , Φ, X∗ , Φ∨ ) be a root datum on which a group Λ acts via
automorphisms. Choosing a root basis ∆ of Φ, we obtain a corresponding
based root datum Ψ̇. Then Ψ̇ also carries an action of Λ. Namely, for σ ∈ Λ,
there exists a unique element c(σ) in the Weyl group W (Ψ) of Ψ such that
σ(∆) = c(σ)(∆). If we define σ ⋆ to be the automorphism of Ψ̇ given by
(16)
σ ⋆ χ = c(σ)−1 (σχ)
8
JEFFREY D. ADLER AND JOSHUA M. LANSKY
for χ ∈ X ∗ , then the action of Λ on Ψ̇ is given by σ 7→ σ ⋆ .
Since Λ acts on Ψ and Ψ̇, it acts naturally on Aut(Ψ) and Aut(Ψ̇), as
well as on the Weyl groups W (Ψ) ⊂ Aut(Ψ) and W (Ψ̇) ⊂ Aut(Ψ̇). Just as
the actions of Λ on Ψ̇ and on Ψ differ, so the actions of Λ on W (Ψ̇) and on
W (Ψ) differ, even though the Weyl groups themselves are equal. For σ ∈ Λ
and w ∈ W (Ψ̇), let (σ, w) 7→ σ ⋆ (w) denote the action of Λ on W (Ψ̇). Then
we have
σw = c(σ) (σ ⋆ w) c(σ)−1 .
One can check readily that map c : Λ −→ W (Ψ̇) is a cocycle in Z 1 (k, W (Ψ̇)).
We now turn our attention to based root data arising from reductive
algebraic groups. Let G be a connected reductive k-group, B a Borel subgroup of G, and T ⊆ B a maximal torus of G. Let Ψ̇(G, B, T ) denote the
corresponding based root datum.
Any map ϑ ∈ Aut(G) determines an obvious isomorphism
ϑ∗ : Ψ̇(G, B, T ) −→ Ψ̇(G, ϑ(B), ϑ(T )).
There is a natural homomorphism
π : Aut(G) −→ Aut(Ψ̇(G, B, T ))
defined as follows. For ϑ ∈ Aut(G), choose gϑ ∈ G(ksep ) such that Int(gϑ )
takes ϑ(B) to B, and ϑ(T ) to T . Then Int(gϑ ) ◦ ϑ stabilizes B and T , and
we let π(ϑ) be the automorphism (Int(gϑ ) ◦ ϑ)∗ of Ψ̇(G, B, T ) (which is, in
fact, independent of the choice of gϑ ).
Now suppose that T is defined over k. Then an element σ ∈ Gal(k)
naturally determines an automorphism of Ψ(G, T ) hence an automorphism
σ ⋆ of Ψ̇(G, B, T ) as defined in (16). We thus obtain an action of Gal(k)
on Ψ̇(G, B, T ), hence one on Aut(Ψ̇(G, B, T )) as above. These actions are
independent of the particular choice of B and T in the sense that if g ∈
G(ksep ) and σ ∈ Gal(k), then we have
(17)
σ ⋆ ◦ Int(g)∗ = Int(g)∗ ◦ σ ⋆ ,
where we use the notation σ ⋆ to denote both the action of σ on Ψ̇(G, B, T )
and on Ψ̇(G, g B, g T ).
There is a well-known exact sequence
(18)
π
1 −→ Inn(G) −→ Aut(G) −→ Aut(Ψ̇(G, B, T )) −→ 1.
We note that the homomorphisms in (18) are Gal(k)-equivariant.
Remark 19. Let ∆ be the set of simple roots for (G, B, T ). Let {Xα }α∈∆ ⊂
Lie(G)(ksep ) be a pinning. It is well known [5, Cor. 2.14] that {Xα } determines a unique splitting ψ of (18). Namely, if f ∈ Aut(Ψ̇(G, B, T )), define
ψ(f ) to be the automorphism of G such that
• ψ(f ) stabilizes B and T ,
• the restriction of ψ(f ) to T is determined by the automorphism of
X ∗ (T ) given by f , and
ROOT DATA WITH GROUP ACTIONS
9
• ψ(f )(Xα ) = Xf (α) .
Thus im ψ lies in the subgroup Aut(G, B, T, {Xα }) of Aut(G) consisting of
automorphisms that stabilize B, T , and the pinning {Xα }. If B and T are
defined over k, and {Xα } is Gal(k)-stable, it follows from [2, §3.10] that ψ
is Gal(k)-equivariant.
Lemma 20. Retain the notation of the previous remark, and assume that B
and T are defined over k, and {Xα } is Gal(k)-stable. Suppose a group Γ acts
on G via k-automorphisms, preserving B, T , and {Xα }. Then Ḡ = (GΓ )◦
is a reductive k-group, B̄ = (B Γ )◦ is a Borel k-subgroup of G, T̄ = (T Γ )◦ is
a maximal k-torus in B, and W (G, T )Γ = W (Ḡ, T̄ ).
We prove this result by reducing to the well-known case where Γ is cyclic.
Proof. The statements about Ḡ and T̄ follow from [1, Proposition 3.5]. The
lemma follows for G if it holds for a central quotient of G. Therefore, we
may assume that (over ksep ) G is a direct product of almost simple groups.
We can also reduce to the case where Γ acts transitively on the factors of
G. As in the proof loc. cit., we may identify the factors of G with each
other, and replace Γ by a group S × Γ1 such that Ḡ = (GS×Γ1 )◦ , where S
acts by permuting the factors in our product decomposition of G, and Γ1
preserves each factor and acts in the same way on each. It is clear from the
construction that S × Γ1 preserves {Xα }.
Working in stages, we may assume that Γ is simple. Thus, either Γ acts
by permutation of the factors of G, or G is simple. In the former case,
our result is trivial, so assume that G is simple. Then G has a connected
Dynkin diagram, whose automorphism group is solvable. Since Γ embeds in
this automorphism group, it must be cyclic.
We let Ψ = (X ∗ (T ), Φ(G, T ), X∗ (T ), Φ∨ (G, T )) and will freely use the
notation of §2. We may identify X̄ ∗ with X ∗ (T̄ ). Under this identification,
the restriction βres of a root β ∈ Φ(G, T ) to T̄ corresponds to i∗ β. It follows
from [8, §8.2(2)] that since Γ fixes a pinning (i.e., cβ = 1 for each β ∈ ∆,
in the terminology loc. cit.), then for each β ∈ Φ(G, T ), there exists a root
α ∈ Φ(Ḡ, T̄ ) proportional to βres . Meanwhile, it follows from [1, Proposition
3.5(iv)] that every root in Φ(Ḡ, T̄ ) is the restriction of a root in Φ(G, T ).
It follows that the Weyl group W̄ of Ψ̄ is equal to W (Ḡ, T̄ ). But W̄ is
canonically isomorphic to W (G, T )Γ by Corollary 15.
4. Proofs of Theorems
Proof of Theorem 1. Consider an abstract root datum Ψ = (X ∗ , Φ, X∗ , Φ∨ )
with an action of Gal(k) × Γ. Suppose that ∆ is a Γ-stable base for Ψ. Let
Ψ̇ be the corresponding based root datum. As discussed in §3, the action of
Gal(k) × Γ on Ψ determines one of Gal(k) × Γ on Ψ̇. Since ∆ is Γ-stable, the
actions of Γ on Ψ and Ψ̇ coincide. In the notation of (16) with Λ = Gal(k),
the elements c(σ) ∈ W (Ψ̇) that arise from the action of Gal(k) on Ψ̇ must
lie in W (Ψ̇)Γ since this action commutes with that of Γ. Therefore, the
10
JEFFREY D. ADLER AND JOSHUA M. LANSKY
map c : Gal(k) −→ W (Ψ̇)Γ is a cocycle in Z 1 (k, W (Ψ̇)Γ ). We note that the
Gal(k) × Γ-isomorphism class of Ψ̇ depends only on that of Ψ.
By [7, Theorem 6.2.7], there exists a triple (G, B0 , T0 ), unique up to kisomorphism, consisting of a k-quasisplit connected reductive group G, a
Borel k-subgroup B0 of G, and a maximal k-torus T0 of B0 , such that the
associated based root datum Ψ̇(G, B0 , T0 ) is Gal(k)-isomorphic to Ψ̇. We
will identify Ψ̇ and Ψ̇(G, B0 , T0 ) via such an isomorphism.
Let {Xα } be a Gal(k)-stable pinning for G relative to B0 and T0 . The
action of Γ on Ψ̇ determines a homomorphism φ : Γ −→ Aut(Ψ̇). Let ϕ be
the composition
φ
ψ
ϕ : Γ −→ Aut(Ψ̇) = Aut(Ψ̇(G, B0 , T0 )) −→ Aut(G, B0 , T0 , {Xα }),
where ψ : Aut(Ψ̇(G, B0 , T0 )) −→ Aut(G, B0 , T0 , {Xα }) is the homomorphism
from Remark 19.
◦
ϕ(Γ) ◦
Let Ḡ = Gϕ(Γ) and T̄0 = T0
. By Lemma 20, Ḡ is a k-quasisplit
reductive group, T̄0 a maximal k-torus of Ḡ, and
W (Ψ̇)Γ = W (G, T0 )ϕ(Γ) = W (Ḡ, T̄0 ).
Thus we may view c as a cocycle in Z 1 (k, W (Ḡ, T̄0 )).
By [4, Theorem 1.1], there is some g ∈ Ḡ(ksep ) such that for all σ ∈
Gal(k), g−1 σ(g) lies in the normalizer NḠ (T̄0 )(ksep ), and the image of g −1 σ(g)
in W (Ḡ, T̄0 ) is equal to c(σ). Let T = g T0 and B = g B0 . Since g is ϕ(Γ)fixed, T is a ϕ(Γ)-stable maximal k-torus of G, and B is a ϕ(Γ)-stable Borel
subgroup of G containing T . We have therefore associated to Ψ a triple
(G, T, ϕ) of the required type.
Suppose we vary the arbitrary choices made in the above construction of
(G, T, ϕ). That is, suppose we choose
• another based root datum Ψ̇′ with underlying datum Ψ, and hence
a cocycle c′ in Z 1 (k, W (Ψ̇′ )Γ );
• another triple of k-groups (G′ , B0′ , T0′ ) k-isomorphic to (G, B0 , T0 )
and an identification of Ψ̇(G′ , B0′ , T0′ ) with Ψ̇′ ; and
• a Gal(k)-stable pinning {Xα′ } of G′ relative to B0′ and T0′ , along with
the associated map ψ ′ : Aut(Ψ̇(G′ , B0′ , T0′ )) −→ Aut(G′ , B0′ , T0′ , {Xα′ })
from Remark 19.
We will show that these choices lead to a triple (G′ , T ′ , ϕ′ ) that is equivalent
to (G, T, ϕ). We note that replacing Ψ by another datum in its Gal(k) × Γisomorphism class has no additional effect on the triple arising from the
construction.
Use a particular k-isomorphism between (G′ , B0′ , T0′ ) and (G, B0 , T0 ) to
identify these triples. Following the above construction, we obtain a homomorphism
ϕ′ : Γ −→ Aut(G, B0 , T0 , {Xα′ }), as well as an element g ′ ∈
′
′ (Γ) ◦
ϕ
and a k-torus T ′ = g T0 , analogous to g and T , respectively.
G
There is a unique element w ∈ W (Ψ) mapping Ψ̇ to Ψ̇′ , and by uniqueness,
w must in fact lie in W (Ψ)Γ , and the mapping it induces is equivariant with
ROOT DATA WITH GROUP ACTIONS
11
respect to the actions of Gal(k) on these based root data. Via conjugation,
the element w induces Γ-equivariant isomorphisms W (Ψ̇) −→ W (Ψ̇′ ) and
τ : Z 1 (k, W (Ψ̇)) −→ Z 1 (k, W (Ψ̇′ )).
We have a unique element κ of Autk (Ψ̇(G, B0 , T0 )) that produces a commutative square
(21)
Ψ̇
// Ψ̇(G, B0 , T0 )
κ
w
Ψ̇′
// Ψ̇(G, B0 , T0 )
Here the horizontal arrows are the identifications chosen in the respective
constuctions of ϕ and ϕ′ . (That κ is Gal(k)-equivariant follows from the
equivariance of the other three maps in the square.) We therefore obtain a
diagram
(22)
♥66
♥♥♥
♥
♥
♥♥♥
♥♥♥
Γ
// Aut(Ψ̇) ∩ Aut(Ψ̇′ )
PPP
PPP
PPP
PPP
((
Aut(Ψ̇)
// Aut(Ψ̇(G, B0 , T0 ))
// Aut(Ψ̇(G, B0 , T0 ))
Aut(Ψ̇′ )
in which the square on the right is induced by (21) (and hence commutes),
the vertical maps are given respectively by conjugation by w and κ, the
diagonal maps are given by inclusion, and the map out of Γ is given by the
action of Γ on Ψ.
The map τ (c) : σ 7→ wc(σ)w−1 is a cocycle in Z 1 (k, W (Ψ̇′ )Γ ), cohomologous to c′ ; more precisely, for σ ∈ Gal(k),
c′ (σ) = w−1 (wc(σ)w−1 )σ ⋆′ (w) = w−1 (τ (c)(σ))σ ⋆′ (w),
where σ ⋆′ denotes the result of the action of σ on w, viewed as an element of
W (Ψ̇′ ). Identifying c and c′ respectively with cocycles in Z 1 (k, W (G, T0 )ϕ(Γ) )
′
and Z 1 (k, W (G, T0 )ϕ (Γ) ) as in the above construction, it follows from (22)
that
(23)
c′ (σ) = w−1 (κ ◦ c(σ) ◦ κ−1 )σ(w),
where σ(w) here denotes the result of σ ∈ Gal(k) acting on the element
w ∈ W (Ψ̇′ ) via the identification of this group with the concrete Weyl group
W (G, T0 ) in (22).
Let n ∈ NḠ (T̄0 )(ksep ) be a representative for w and set µ = ψ(κ) ∈
Autk (G, B0 , T0 ). Then by (23), g ′ −1 σ(g′ ) and n−1 µ(g−1 σ(g))σ(n) have the
same image in W (G, T0 ). Rearranging terms and letting h = g′ n−1 µ(g)−1 ,
12
JEFFREY D. ADLER AND JOSHUA M. LANSKY
we obtain that σ(h) and h have the same image modulo
(24)
σ(µ(g)n)
T0 = σ(µ(g)n T0 ) = σ(µ(g) T0 ) = σ(µ(g T0 )) = σ(µ(T )) = µ(T ).
Let ν = Int(h) ◦ µ. Since
−1
ν(T ) = Int(h)(µ(T )) = Int(g ′ n−1 µ(g)−1 )(µ(T )) = Int(g ′ n−1 )(µ(g T ))
′
= Int(g′ n−1 )(µ(T0 )) = Int(g ′ n−1 )(T0 ) = g T0 = T ′ ,
it follows from (24) that ν gives a k-isomorphism T −→ T ′ .
To show that (G′ , T ′ , ϕ′ ) is equivalent to (G, T, ϕ), it remains to show that
ν is Γ-equivariant. It follows from the construction of ϕ that π ◦ ϕ is equal
to the composition Γ −→ Aut(Ψ̇) −→ Aut(Ψ̇(G, B0 , T0 )) appearing in (22).
Similarly, π ◦ ϕ′ is equal to the analogous composition Γ −→ Aut(Ψ̇′ ) −→
Aut(Ψ̇(G, B0 , T0 )). Thus for any γ ∈ Γ,
π(ϕ′ (γ)) = κ ◦ π(ϕ(γ)) ◦ κ−1 .
Applying ψ to this equality and noting that ψ ◦ π ◦ ϕ = ϕ by construction,
we obtain
ψ(π(ϕ′ (γ))) = µ ◦ ϕ(γ) ◦ µ−1 .
Note that by definition, ψ(f ) and ψ ′ (f ) agree on T0 for any f ∈ Aut(Ψ̇(G, B0 , T0 )).
Therefore, as automorphisms of T0 , we have
ϕ′ (γ) = ψ ′ (π(ϕ′ (γ))) = ψ(π(ϕ′ (γ))) = µ ◦ ϕ(γ) ◦ µ−1 .
It follows that, as maps on T ,
ϕ′ (γ) ◦ ν
= ϕ′ (γ) ◦ Int(h) ◦ µ
= ϕ′ (γ) ◦ Int(g ′ n−1 µ(g)−1 ) ◦ µ
= Int(g ′ ) ◦ ϕ′ (γ) ◦ Int(µ(g)n)−1 ◦ µ
= Int(g ′ ) ◦ µ ◦ ϕ(γ) ◦ µ−1 ◦ Int(µ(g)n)−1 ◦ µ
= Int(g ′ ) ◦ µ ◦ ϕ(γ) ◦ Int(gµ−1 (n))−1
= Int(g ′ ) ◦ µ ◦ Int(gµ−1 (n))−1 ◦ ϕ(γ),
where the last equality above comes from the fact that g ∈ Ḡ(ksep ) and
Int(µ−1 (n)) ∈ W (G, T0 )ϕ(Γ) . Thus ϕ′ (γ) ◦ ν is equal to
Int(g ′ n−1 µ(g)−1 ) ◦ µ ◦ ϕ(γ) = ν ◦ ϕ(γ),
showing that ν is Γ-equivariant. Therefore, (G′ , T ′ , ϕ′ ) is equivalent to
(G, T, ϕ), and our construction induces a well-defined map sΓ : RΓ −→ TΓ .
We now show that rΓ ◦sΓ is the identity map on RΓ . Let Ψ be a root datum
representing some class in RΓ , and let (G, T, ϕ) be a triple representing the
image of the class of Ψ under sΓ . We need to show that Ψ(G, T ) is Gal(k)×Γisomorphic to Ψ. We will make free use of the notation developed in the
construction of sΓ .
ROOT DATA WITH GROUP ACTIONS
13
The Gal(k)-equivariant isomorphism of based root data Ψ̇ −→ Ψ̇(G, B0 , T0 )
chosen in the definition of sΓ is Γ-equivariant by construction (where the action of Γ on Ψ̇(G, B0 , T0 ) is induced by ϕ). We may therefore identify Ψ̇ and
Ψ̇(G, B0 , T0 ) as based root data with Gal(k)×Γ-action via this isomorphism.
This allows us to identify Ψ and Ψ(G, T0 ) as root data with Γ-action (but
not necessarily with Gal(k)-action since the actions of Gal(k) on Ψ̇ and Ψ
differ in general).
Recall the element g ∈ Ḡ(ksep ) chosen in the definition of sΓ . The map
Int(g)∗ : Ψ = Ψ(G, T0 ) −→ Ψ(G, T ) is Γ-equivariant since g is ϕ(Γ)-fixed.
Furthermore, Int(g)∗ is Gal(k)-equivariant since for σ ∈ Gal(k) and χ ∈
X ∗ (T0 ),
Int(g)∗ (σχ) = Int(g)∗ c(σ)(σ ⋆ χ)
= gg
−1 σ(g)
(σ ⋆ χ)
= σ(g) (σ ⋆ χ)
= σ(g χ)
= σ(Int(g)∗ (χ)).
Thus Ψ(G, T ) is Gal(k) × Γ-isomorphic to Ψ, as desired.
Finally, we show that sΓ ◦ rΓ is the identity map on TΓ . Let (G, T, ϕ)
represent a class in TΓ , and let (G′ , T ′ , ϕ′ ) represent the image of this class
under sΓ ◦ rΓ . Since rΓ ◦ (sΓ ◦ rΓ ) = (rΓ ◦ sΓ ) ◦ rΓ = rΓ , it follows that there
is a Gal(k) × Γ isomorphism Ψ(G, T ) −→ Ψ(G′ , T ′ ). By [5, Theorem 2.9],
this isomorphism is induced by an isomorphism ν : G −→ G′ that restricts
to a Γ-equivariant k-isomorphism T −→ T ′ . Thus (G, T, ϕ) and (G′ , T ′ , ϕ′ )
are equivalent.
Remark 25. Observe that in the definition of the map sΓ above, the triple
(G, T, ϕ) is constructed in such a way that G is k-quasisplit and ϕ fixes a
Gal(k)-invariant pinning of G. Thus, since sΓ ◦ πΓ is the identity map on
TΓ , we see that every equivalence class in TΓ contains such a triple.
Moreover, suppose that (G, T, ϕ) is a triple of this kind. Applying the
construction of sΓ ◦ rΓ to this triple, we see that the triple we obtain is
precisely (G, T, ϕ), provided that we make appropriate choices.
Remark 26. Recall that in the proof, it is shown that if (G, T, ϕ) and
(G′ , T ′ , ϕ′ ) are two triples that arise by applying the sΓ construction to
a root datum Ψ, then (G, T, ϕ) and (G′ , T ′ , ϕ′ ) are equivalent. We note that
the equivalence ν constructed in this case is of a special kind. Namely, ν is
of the form Int(h) ◦ µ, where h ∈ G(ksep ) and µ ∈ Autk (G, B0 , T0 ).
Now suppose that (G, T, ϕ) and (G′ , T ′ , ϕ′ ) are arbitrary equivalent triples
with the properties that G and G′ are k-quasisplit and ϕ and ϕ′ fix Gal(k)invariant pinnings for G and G′ , respectively. Then combining the first part
of this remark with Remark 25, it follows that there is an equivalence ν
between (G, T, ϕ) and (G′ , T ′ , ϕ′ ) of the above special form.
14
JEFFREY D. ADLER AND JOSHUA M. LANSKY
Remark 27. Suppose that G′ is k-quasisplit and T ′ is a maximal k-torus
of G′ . Suppose that the finite group Γ acts via Gal(k)-equivariant automorphisms on Ψ(G′ , T ′ ) preserving a base. Then the equivalence class of
Ψ(G′ , T ′ ) lies in RΓ . Applying the construction in the definition of sΓ
to Ψ(G′ , T ′ ), we obtain a triple (G, T, ϕ) where G is k-quasisplit. Since
Ψ(G′ , T ′ ) and Ψ(G, T ) are Gal(k)-isomorphic, G′ can be taken to equal G.
Moreover, if g ∈ G(ksep ) is chosen such that T ′ = g T0 , then the cocylcle c used to define T can be taken to be the image of σ 7→ g−1 σ(g) in
Z 1 (k, W (G, T0 )Γ ). In particular, it follows from [1, Proposition 6.1] that T ′
is stably conjugate to T .
Proof of Theorem 2. Consider a class in TΓ . From Remark 25, we can represent this class by a triple (G, T, ϕ), where G is k-quasisplit and the action ϕ of Γ on G fixes a Gal(k)-invariant pinning. Let Ḡ = (Gϕ(Γ) )◦ and
T̄ = (T ϕ(Γ) )◦ . Then Ψ = Ψ(G, T ) comes equipped with an action of Γ.
Consider the restricted root datum Ψ̄ of Theorem 7. Construct a new root
datum Ψ̄′ by replacing the root system Φ̄ of Ψ̄ by a maximal reduced subsystem Φ̄′ as in Remark 13, and do likewise with the coroot system. It is clear
from the constructions that the root data Ψ̄′ and Ψ(Ḡ, T̄ ) are equivalent,
provided that their root systems are equivalent.
As in the proof of Lemma 20, one may reduce to the case where Γ is
cyclic. From the proof of [8, §8.2(2′′′′ )], the root system Φ̄′ must contain the
root system Φ(Ḡ, T̄ ), with equality since Γ fixes a pinning.
References
[1] Jeffrey D. Adler and Joshua M. Lansky, Lifting representations of finite reductive
groups I: Semisimple conjugacy classes, Canad. J. Math. 66 (2014), 1201–1224, DOI
10.4153/CJM-2014-013-6, available at arXiv:1106.0706.
[2] M. Demazure, Automorphismes des groupes réductifs, Schémas en Groupes (Sém.
Géométrie Algébrique, Inst. Hautes Études Sci., 1963/64), Inst. Hautes Études Sci.,
Paris, 1965, pp. 87 (French). MR0228503 (37 #4083)
[3] Robert E. Kottwitz and Diana Shelstad, Foundations of twisted endoscopy, Astérisque
255 (1999), vi+190 (English, with English and French summaries). MR1687096
(2000k:22024)
[4] M. S. Raghunathan, Tori in quasi-split groups, J. Ramanujan Math. Soc. 19 (2004),
no. 4, 281–287. MR2125504 (2005m:20114)
[5] Tonny A. Springer, Reductive groups, Automorphic forms, representations, and Lfunctions. Part 1 (Armand Borel and W. Casselman, eds.), Proceedings of Symposia in
Pure Mathematics, XXXIII, American Mathematical Society, Providence, R.I., 1979,
pp. 3–27. MR546587 (80h:20062)
, Linear algebraic groups, Progress in Mathematics, vol. 9, Birkhäuser Boston
[6]
Inc., Boston, MA, 1998. MR1642713 (99h:20075)
, Linear algebraic groups, Algebraic geometry IV, Encyclopedia of Mathemati[7]
cal Sciences, SpringerVerlag, 1994, pp. 1–121. MR1100484 (92g:20061)
[8] Robert Steinberg, Endomorphisms of linear algebraic groups, Memoirs of the American
Mathematical Society, No. 80, American Mathematical Society, Providence, R.I., 1968.
MR0230728 (37 #6288)
ROOT DATA WITH GROUP ACTIONS
15
Department of Mathematics and Statistics, American University, Washington, DC 20016-8050
E-mail address, Adler: [email protected]
E-mail address, Lansky: [email protected]
| 4math.GR
|
An Efficient Primal-Dual Algorithm for Fair
Combinatorial Optimization Problems
Viet Hung Nguyen1 and Paul Weng2
arXiv:1801.07544v1 [cs.DS] 23 Jan 2018
1
Sorbonne Universités, UPMC Univ Paris 06, UMR 7606, LIP6, Paris, France
[email protected]
2
SYSU-CMU Joint Institute of Engineering, Guangzhou, China
SYSU-CMU Joint Research Institute, Shunde, China
School of Electronics and Information Technology, SYSU
[email protected]
Abstract. We consider a general class of combinatorial optimization
problems including among others allocation, multiple knapsack, matching or travelling salesman problems. The standard version of those problems is the maximum weight optimization problem where a sum of values
is optimized. However, the sum is not a good aggregation function when
the fairness of the distribution of those values (corresponding for example to different agents’ utilities or criteria) is important. In this paper,
using the generalized Gini index (GGI), a well-known inequality measure, instead of the sum to model fairness, we formulate a new general
problem, that we call fair combinatorial optimization. Although GGI
is a non-linear aggregating function, a 0, 1-linear program (IP) can be
formulated for finding a GGI-optimal solution by exploiting a linearization of GGI proposed by Ogryczak and Sliwinski [21]. However, the time
spent by commercial solvers (e.g., CPLEX, Gurobi...) for solving (IP)
increases very quickly with instances’ size and can reach hours even for
relatively small-sized ones. As a faster alternative, we propose a heuristic for solving (IP) based on a primal-dual approach using Lagrangian
decomposition. We demonstrate the efficiency of our method by evaluating it against the exact solution of (IP) by CPLEX on several fair
optimization problems related to matching. The numerical results show
that our method outputs in a very short time efficient solutions giving
lower bounds that CPLEX may take several orders of magnitude longer
to obtain. Moreover, for instances for which we know the optimal value,
these solutions are quasi-optimal with optimality gap less than 0.3%.
Keywords: Fair Optimization; Generalized Gini Index; Ordered Weighted
Averaging; Matching; Subgradient Method.
1
Introduction
The solution of a weighted combinatorial optimization problem can be seen
as the selection of n values in a combinatorial set X ⊂ Rn . The maximum
weight version
Pnof such a problem consists in maximizing the sum of these n
values (e.g., i=1 ui ). For instance, in a matching problem on a graph, the sum
of weights that is optimized corresponds to the sum of weights of the edges
2
Viet Hung Nguyen and Paul Weng
selected in a matching. In practice, the vector of weights (u1 , u2 , . . . , un ) could
receive different interpretations depending on the actual problem. In a multiagent setting, each value ui represents the utility of an agent i, as in a bi-partite
matching problem where n objects have to be assigned to n agents. In a multicriteria context, those n values can be viewed as different dimensions to optimize.
For example, in the travelling salesman problem (TSP) with n cities, a feasible
solution (i.e., Hamiltonian cycle) is valued by an n-dimensional vector where
each component represents the sum of the distances to reach and leave a city.
In both interpretations, it is desirable that the vector of values (u1 , u2 , . . . , un )
be both Pareto-optimal (i.e., not improvable on all components at the same time)
and balanced (or fair). We call optimization with such concerns fair optimization
by adopting the terminology from multi-agent systems. In this paper, we focus
on the fair optimization version of a class of combinatorial problems (including
allocation, general matching, TSP...). Note that optimizing the sum of the values
(i.e., maximum weight problem) yields a Pareto-optimal solution, but does not
provide any guarantee on how balanced the vector solution would be.
Various approaches have been proposed in the literature to provide such a
guarantee with different models for fairness or ”balancedness” (see Section 2 for
an overview). In this paper, our approach is based on an inequality measure
called Generalized Gini Index (GGI) [29], which is well-known and well-studied
in economics and can be used to control for both Pareto-efficiency and fairness.
Indeed, fairness has naturally been investigated in economics [17]. In this literature, two important requirements have been identified as essential for fairness:
equal treatment of equals and efficiency. The first notion implies that two agents
with the same characteristics (notably the same preferences) have to be treated
the same way, while the second entails that a fair solution should be Paretooptimal. GGI satisfies both requirements, as it is symmetric in its arguments
and increasing with Pareto dominance. The notion of fairness that GGI encodes
is based on the Pigou-Dalton transfer principle, which states that a small transfer
of resource from a richer agent to a poorer one yields a fairer distribution.
To the best of our knowledge, fair optimization in such general combinatorial
problems has not been considered so far, although the GGI criterion has been
investigated before in some specific problems (allocation [12], capital budgeting
[11], Markov decision process [19,20]...). The difficulty of this combinatorial optimization problem lies in the fact that the objective function is non-linear. The
contribution of this paper is fourfold: (1) we introduce a new general combinatorial problem (e.g., fair matching in general graph or fair TSP have not been
studied so far); (2) we provide an optimality condition and an approximation
ratio; (3) we propose a fast general heuristic method based on a primal-dual approach and on Lagrangian decomposition; (4) we evaluate this method on several
problems related to matching to understand its efficiency. Although our general
combinatorial formulation covers problems whose maximum weight version is
NP-hard, we leave for a follow-up work the integration of our fast heuristic with
approximation algorithms to solve those NP-hard problems.
The paper is organized as follows. Section 2 gives an overview of related work.
Section 3 provides a formal definition of our problem, which can be solved by a
An Efficient Primal-Dual Algorithm for Fair Combinatorial Optimization
3
0, 1-linear program. As a faster alternative, we present a heuristic primal-dual
solving method based on Lagrangian decomposition in Section 4 and evaluate it
experimentally in Section 5. Finally, we conclude in Section 6.
2
Related Work
Fair optimization is an active and quite recent research area [18,14] in multiobjective optimization. Fairness can be modeled in different ways. One simple
approach is based on maxmin, so called Egalitarian approach, where one aims
at maximizing the worse-off component (i.e., objective, agent...). Due to the
drowning effects of the min operator, vectors with the same minimum cannot
be discriminated. A better approach [24] is based on the lexicographic maxmin,
which consists in considering the minimum first when comparing two vectors,
then in case of a tie, focusing on the second smallest values and so on. However,
due to the noncompensatory nature of the min operator, vector (1, 1, . . . , 1)
would be preferred to (0, 100, . . . , 100), which may be debatable. To take into
account this observation, one can resort to use a strictly increasing and strictly
Schur-concave (see Section 3.2 for definition) aggregation function f (see [18] for
examples) that evaluates each vector such that higher values are preferred.
In this paper, we focus on the Generalized Gini Index (GGI) proposed in
the economics literature [29], because it satisfies natural properties for encoding
fairness. GGI is a particular case of a more general family of operators known as
Ordered Weighted Averaging (OWA) [31]. Much work in fair optimization has
applied the OWA operator and GGI in multiobjective (continuous and combinatorial) optimization problems. To cite a few, it was used in network dimensioning
problems [22], capital budgeting [11], allocation problems [12], flow optimization
in wireless mesh networks [10] and multiobjective sequential decision-making under uncertainty [19,20]. One common solving technique is based on a linearization trick of the nonlinear objective function based on GGI [21]. Recently, [9]
considered a similar setting to ours, but tries to solve its continuous relaxation.
In multicriteria decision-making, fair optimization is related to compromise
optimization, which generally consists in minimizing a distance to an ideal point
[27]. More generally, the ideal point can be replaced by any reference point
that a decision maker chooses, as in the reference point method [30]. In this
context, a judiciously chosen reference point can help generate a solution with
a balanced profile on all criteria. One main approach is based on minimizing
the augmented weighted Tchebycheff distance. This method has been applied in
many multicriteria problems, for instance, in process planning [25], in sequential decision-making under uncertainty [23], in discrete bicriteria optimization
problems [6], in multiobjective multidimensional knapsack problems [15].
Note that our combinatorial optimization problem should not be confused
with the multicriteria version of those problems where each scalar weight becomes vectorial and the value of a solution is obtained by aggregating the selected
weight vectors with a componentwise sum. For instance, Anand [1] investigated
a multicriteria version of the matching problem and proved that the egalitarian
approach for vector-valued matching leads to NP-hard problems. In our problem,
4
Viet Hung Nguyen and Paul Weng
the weights are scalar and the value of a solution is not obtained by summing
its scalar weights, but by aggregating them with GGI.
3
Model
In this section, we formally describe the general class of combinatorial problems
considered in this paper and provide some concrete illustrative examples in this
class. Then we recall the generalized Gini index as a measure of fairness and
define the fair combinatorial optimization problems tackled in this paper. We
start with some notations. For any integer n, [n] denotes the set {1, 2, . . . n}. For
any vector x, its component is denoted xi or xij depending on its dimension.
3.1
General Model
We consider a combinatorial optimization problem (e.g., allocation, multiple
knapsack, matching, travelling salesman problem...), whose feasible solutions
X ⊆ {0, 1}n×m can be expressed as follows:
Az ≤ b
z ∈ {0, 1}n×m
where A ∈ Zp×(nm) , b ∈ Zp , n, m and p are three positive integers, and z is
viewed as a one-dimensional vector (z11 , . . . , z1m , z21 , . . . , z2m , . . ., zn1 , . . . , znm )⊺ .
Let uij ∈ N be the utility of setting zij to 1. The maximum weight problem
defined on combinatorial set X can be written as a 0, 1-linear program (0, 1-LP):
max.
X X
uij zij
i∈[n] j∈[m]
s.t. z ∈ X
Because this general problem includes the travelling salesman problem (TSP), it
is NP-hard in general. As mentioned before, this objective function provides no
control on the fairness of the obtained solution. Although possibly insufficient,
one simple approach to fairness consisting in focusing on the worse-off component
is the maxmin problem defined on set X , which can also be written as a 0, 1-LP:
max. v
s.t. v ≤
X
uij zij
∀i ∈ [n]
j∈[m]
z∈X
Even for some polynomial problems like allocation, this version is NP-hard in
general [4]. To avoid any confusion, in this paper, allocation refers to matching
on a bi-partite graph and matching generally implies a complete graph.
For illustration, we now present several instantiations of our general model
on allocation and matching problems, some of which will be used for the experimental evaluation of our proposed methods in Section 5.
Example 1 (Allocation) Let G = (V1 ∪ V2 , E, u) be a valued bipartite graph
where V1 and V2 are respectively an n-vertex set and an m-vertex set with V1 ∩
V2 = ∅, E ⊆ {x, y} | (x, y) ∈ V1 × V2 is a set of non-directed edges and
An Efficient Primal-Dual Algorithm for Fair Combinatorial Optimization
5
u : E → R defines the nonnegative utility (i.e., value to be maximized) of an
edge. As there is no risk of confusion, we identify V1 to the set [n] and V2 to
the set [m]. An allocation of G is a subset of E such that each vertex i in V1 is
connected to αi to βi vertices in V2 and each vertex in V2 is connected to α′j to
βj′ vertices in V1 where (α, β) ∈ Nn×n and (α′ , β ′ ) ∈ Nm×m .
The assignment problem where n tasks need to be assigned to n agents is a
special case where n = m and αi = βi = α′j = βj′ = 1 for i ∈ [n] and j ∈ [n].
The conference paper assignment problem where m papers needs to be reviewed
by n reviewers such that each paper is reviewed by 3 reviewers and each reviewer
receives at most 6 papers can be represented with αi = 0, βi = 6, α′j = 3 and
βj′ = 3 for i ∈ [n] and j ∈ [m]. The Santa Claus problem [3] where m toys needs
to be assigned to n children with n ≤ m is also a particular case with αi = 0,
βi = m, α′j = βj′ = 1 for i ∈ [n] and j ∈ [m].
The maximum weight problem can be solved with the following 0, 1-LP:
X X
max.
uij zij
i∈[n] j∈[m]
s.t. αi ≤
X
zij ≤ βi ∀i ∈ [n]
(3a)
X
zij ≤ βj′ ∀j ∈ [m]
(3b)
j∈[m]
α′j ≤
i∈[n]
z ∈ {0, 1}n×m
Interestingly, its solution can be efficiently obtained by solving its continuous
relaxation because the matrix defining its constraints (3a)–(3b) is totally unimodular [26]. However, the maxmin version is NP-complete [4].
Example 2 (Matching) Let G = (V,E, u) be a valued graph where V is a
2n-vertex set (with n ∈ N\{0}), E ⊆ {x, y} | (x, y) ∈ V 2 , x 6= y is a set of
non-directed edges and u : E → R defines the nonnegative utility of an edge. A
matching M of G is a subset of E such that no pair of edges of M are adjacent,
i.e., they do not share a common vertex: ∀(e, e′ ) ∈ E 2 , e 6= e′ ⇒ e ∩ e′ = ∅. A
perfect matching M is a matching where every vertex of G is incident to an edge
of M . Thus, a perfect matching contains n edges. Without loss of generality, we
identify V to the set [2n] and denote ∀e = {i, j} ∈ E, uij = u(e) when convenient.
The standard maximum weight perfect matching problem aims at finding a
perfect matching for which the sum of the utilities of its edges is maximum. Let
δ(i) = {{i, j} ∈ E | j ∈ V \{i}} be the set of edges that are incident on vertex i.
It is known [13] that this problem can be formalized as a 0, 1-LP (where zij ’s for
i > j are unnecessary and can be set to 0):
max.
(PM )
s.t.
X
X
(4a)
uij zij
i∈[2n] j∈[2n],j>i
X
zij = 1
∀k ∈ [2n]
(4b)
∀i ∈ [2n], j = i + 1, . . . , 2n
(4c)
{i,j}∈δ(k),i<j
zij ∈ {0, 1}
where (4b) states that in a matching only one edge is incident on any vertex.
6
Viet Hung Nguyen and Paul Weng
This problem can be solved as an LP by considering the continuous relaxation
of PM and adding the well-known blossom constraints (5b) in order to remove
the fractional solutions introduced by the relaxation:
max.
s.t.
(RP M )
X
X
uij zij
i∈[2n] j∈[2n],j>i
X
zij = 1
∀k ∈ [2n]
(5a)
z(δ(S)) ≥ 1
∀S ⊂ V, |S| odd, |S| ≥ 3
(5b)
0 ≤ zij ≤ 1
∀i ∈ [2n], j = i + 1, . . . , 2n
(5c)
{i,j}∈δ(k),i<j
P
where z(δ(S)) =
{i,j}∈δ(S),i<j zij and δ(S) = {{i, j} ∈ E | i ∈ S and j ∈
V \S}. Constraints (5a)–(5c) define the so-called perfect matching polytope. In
practice, this problem can be efficiently solved with the Blossom algorithm proposed by Edmonds [8]. To the best of our knowledge, the maxmin version of the
matching problem (on complete graph) has not been investigated so far.
In this paper we focus on a variant of those P
combinatorial problems: search
for a solution z whose distribution of values
j∈[m] uij zij i∈[n] is fair to its
components (e.g., different agents’ utilities or criteria). To model fairness we use
a special case of the ordered weighted averaging operator that we recall next.
3.2
Ordered Weighted Average and Generalized Gini Index
The Ordered Weighted Average (OWA) [31] of v ∈ Rn is defined by:
X
wk vk↑
OW Aw (v) =
k∈[n]
where w = (w1 , . . . , wn ) ∈ [0, 1] is the OWA weight vector and v ↑ = (v1↑ , . . . , vn↑ )
is the vector obtained from v by rearranging its components in an increasing order. OWA defines a very general family of operators, e.g., the sum (for wk = 1,
∀k ∈ [n]), the average, the minimum (for w1 = 1 and wk = 0, ∀k > 1), the maximum (for wn = 1 and wk = 0, ∀k < n), the leximin when differences between
OWA weights tends to infinity or the augmented weighted Tchebycheff distance
[20].
Let the Lorenz components [2] of v be denoted by (L1 (v), . . ., Ln (v)) and be
P
defined by ∀k ∈ [n], Lk (v) = i∈[k] vi↑ . Interestingly, OWA can be rewritten as:
X
wk′ Lk (v)
(6)
OW Aw (v) =
n
k∈[n]
where ∀k ∈ [n], wk′ = wk − wk+1 and wn+1 = 0. With this rewriting, one can see
that OWA is simply a weighted sum in the space of Lorenz components.
The notion of fairness that we use in this paper is based on the Pigou-Dalton
principle [16]. It states that, all other things being equal, we prefer more “balanced” vectors, which implies that any transfer (called Pigou-Dalton transfer)
from a richer component to a poorer one without reversing their relative positions yields a preferred vector. Formally, for any v ∈ Rn where vi < vj and for
An Efficient Primal-Dual Algorithm for Fair Combinatorial Optimization
7
cumulative value
100
80
60
40
20
0
0
20
40
60
80
100
cumulative population (in %)
Fig. 1. Lorenz curves
any ǫ ∈ (0, vj −vi ) we prefer v +ǫ1i −ǫ1j to v where 1i (resp. 1j ) is the canonical
vector, null everywhere except in component i (resp. j) where it is equal to 1.
When the OWA weights are strictly decreasing and positive [29], OWA is
called the Generalized Gini Index (GGI) [29] and denoted Gw . It encodes both:
efficiency: Gw is increasing with respect to Pareto-dominance (i.e., if v ∈ Rn
Pareto-dominates3 v ′ ∈ Rn , then Gw (v) > Gw (v ′ )); and
fairness: Gw is strictly Schur-concave, i.e., it is strictly increasing with PigouDalton transfers (∀v ∈ Rn , vi < vj , ∀ǫ ∈ (0, vj −vi ), Gw (v+ǫ1i −ǫ1j ) > Gw (v)).
The classic Gini index, which is a special case of GGI with wi = (2(n − i) +
1)/n2 for all i ∈ [n], enjoys a nice graphical interpretation (see Figure 1). For a
given distribution
v ∈ Rn+ , let v̄ denote the average of the components of v, i.e.,
1 Pn
v̄ = n i=1 vi . Distribution v can be represented by the curve going through
the points (0, 0) and ( nk , Lk (v)) for k ∈ [n]. The most equitable distribution with
the same total sum as that of v (i.e., nv̄) can be represented by the straight line
going through the points (0, 0) and ( nk , kv̄) for k ∈ [n]. The value 1 − Gw (v)/v̄
is equal to twice the area between the two curves.
Interestingly, the Lorenz components of a vector can be computed by LP
[21]. Indeed, the k-th Lorenz component Lk (v) of a vector v can be found as the
solution of a knapsack problem, which is obtained by solving the following LP:
min.
(LP k )
s.t.
X
aik xi
X
aik = k
i∈[n]
i∈[n]
0 ≤ aik ≤ 1
∀i ∈ [n]
Equivalently, this can be solved by its dual:
(DLk )
3
max.
s.t.
krk −
X
dik
i∈[n]
rk − dik ≤ vi
∀i ∈ [n]
dik ≥ 0
∀i ∈ [n]
Vector v Pareto-dominates vector v ′ if ∀i ∈ [n], vi ≥ vi′ and ∃j ∈ [n], vj > vj′ .
8
Viet Hung Nguyen and Paul Weng
The dual formulation is particularly useful. Contrary to the primal, it can be integrated in an LP where the vi ’s are also variables [21]. We will use this technique
to formulate a 0, 1-LP to solve our general combinatorial optimization problem.
3.3
Fair Combinatorial Optimization
The problem tackled in this paper is defined by using GGI as objective function:
X
Az ≤ b
s.t.
uij zij i∈[n]
max. Gw
z ∈ {0, 1}n×m
j∈[m]
Following Ogryczak and Sliwinski [21], we can combine the rewriting of OWA
based on Lorenz components (6) and LPs (DLk ) for k ∈ [2n] to transform the
previous non-linear optimization program into a 0, 1-LP:
X
max.
wk′ (krk −
X
dik )
(9a)
i∈[n]
k∈[n]
s.t. Az ≤ b
(9b)
z ∈ {0, 1}n×m
X
uij zij
rk − dik ≤
(9c)
∀i ∈ [n], ∀k ∈ [n]
(9d)
∀i ∈ [n], ∀k ∈ [n]
(9e)
j∈[m]
dik ≥ 0
Due to the introduction of new constraints (9d)–(9e) from LPs (DLk ), the relaxation of this 0, 1-LP may yield fractional solutions. The naive approach to solve
it would be to give it to a 0, 1-LP solver (e.g., Cplex, Gurobi...). Our goal in
this paper is to propose an adapted solving method for it, which would be much
faster than the naive approach by exploiting the structure of this problem.
4
Alternating Optimization Algorithm
Before presenting our approach, which is a heuristic method based on a primaldual technique using a Lagrangian decomposition, we first make an interesting
and useful observation. The dual of the continuous relaxation of the previous
0, 1-LP (9) is given by:
min. b⊺ v +
X X
(10a)
tij
i∈[n] j∈[m]
s.t. (v ⊺ A)ij + tij −
X
∀i ∈ [n], ∀j ∈ [m]
(10b)
∀k ∈ [n]
(10c)
0 ≤ yik ≤ wk′
∀i ∈ [n], ∀k ∈ [n]
(10d)
vj ≥ 0
∀j ∈ [p]
(10e)
tij ≥ 0
∀i ∈ [n], ∀j ∈ [m]
(10f)
uij yik ≥ 0
k∈[n]
n
X
yik = kwk′
i=1
An Efficient Primal-Dual Algorithm for Fair Combinatorial Optimization
9
Interestingly, with fixed yik ’s, the dual of the previous program can be written
in the following form, which is simply the continuous relaxation of the original
program with modified weights:
max.
X
i∈[n]
X
yik
k∈[n]
X
uij zij
(11a)
j∈[m]
s.t. Az ≤ b
(11b)
n×m
z ∈ [0, 1]
(11c)
Therefore, solving this program with discrete z yields a feasible solution of the
original problem. We denote (Py ) the 0, 1-LP (11) defined with y = (yik )i∈[n],k∈[n] .
4.1
Optimality Condition and Approximation Ratio
Next we express an optimality condition so that an integer solution z ∗ computed
from a dual feasible solution y ∗ of (10) is optimal for program (9). First, note
that any extreme solution (v, t, y) of program (10) is such that either yik = 0 or
yik = wk′ for all i ∈ [n] and k ∈ [n].
Theorem 1. Let (v, t, y ∗ ) be an extreme solution
of (10) and let z ∗ be the
P
∗
∗
∗
for all i ∈ [n] and
optimal solution of program (Py ). Let Ti = j∈[m] uij zij
assume without loss of generality that T1∗ ≥ T2∗ ≥ . . . ≥ Tn∗ .
∗
∗
= 0 for all i ∈ [n − k]
= wk′ for all i ≥ n + 1 − k and yik
If for all k ∈ [n], yik
then z ∗ is an optimal solution of program (9).
Proof. Let (v ∗ , t∗ ) be the dual optimal solution associated with z ∗ when solving (Py∗ ). Composing them with y ∗ , we obtain a feasible solution (v ∗ , t∗ , y ∗ ) of
(10). By duality P
theory P
of linear programming, the objective value of this solu′
)Ti∗ . Let us now build a feasible solution
tion is equal to i∈[n] ( j∈[i] wn+1−j
∗
∗
∗
∗
(r , d , z ) of (9) based on z as follows. For all k ∈ [n],
∗
and
• rk∗ = Tn+1−k
∗
rk − Ti∗ if i ≥ n + 1 − k
• d∗ik =
for all i ∈ [n] .
0
otherwise
We now show that (r ∗ , d∗ ) satisfy constraints (9d). For any i ∈ [n] and k ∈ [n],
∗
≤ Ti∗ and d∗ik = 0, we have
if i ≤ n + 1 − k then as rk∗ = Tn+1−k
X
∗
uij zij
.
rk∗ − d∗ik ≤ Ti∗ =
j∈[m]
P
∗
. Hence
If i ≥ n + 1 − k then as d∗ik = rk∗ − Ti∗ , rk∗ − d∗ik = Ti∗ = j∈[m] uij zij
P
∗
∗
∗
∗
(r , d , z ) is a feasible solution of (9). For any k ∈ [n], krk − i∈[n] d∗ik =
P
P
P
∗
krk∗ − ni=n+1−k d∗ik = krk∗ − (krk∗ − ni=n+1−k Ti∗ ) = ni=n+1−k
PTi ). Then it is
easy to see that the objective value of this solution, which is k∈[n] wk′ (krk∗ −
P
P
P
d∗ik ) is equal to k∈[n] wk′ ni=n−k+1 Ti∗ . This sum is just a rewriting of
Pi∈[n] P
′
∗
i∈[n] (
j∈[i] wn+1−j )Ti . Thus, by duality of linear programming, the solution
(r ∗ , d∗ , z ∗ ) is optimal for program (9).
⊓
⊔
10
Viet Hung Nguyen and Paul Weng
Theorem 1 provides an optimality condition for any feasible solution z ∗ , but
does not indicate how to find ”good” solutions. Yet, one may be interested in
the quality of some special solutions, e.g., the optimal solution of the maximum
weight version. The following theorem establishes an approximation ratio for the
latter, which also applies to our method as discussed later.
Theorem
P 2. Let z̄ be an optimal solution of the maximum weight version. Let
T̄i =
j∈[m] uij z̄ij for all i ∈ [n] and assume without loss of generality that
′
T̄1 ≥ T̄2 ≥ . . . ≥ T̄n . Let wmax
= maxk∈[n] wk′ . Then the GGI value of z̄ is at
2w ′
worst max( (n+1)wn ′
max
, (P nT̄n T̄i )) ) of the optimal objective value of program (9).
i∈[n]
Proof. Let vector ȳ ∈ Rn×n be defined as ȳik = nk wk′ for i, k ∈ [n], which is
feasible for program (10). The objective function of (Pȳ ) satisfies:
X
i∈[n]
X
k∈[n]
ȳik
X
uij zij =
X X X k
(
w′ )uij zij
n k
i∈[n] j∈[m] k∈[n]
j∈[m]
≤
X X X k
w′ )uij zij
(
n max
(12)
i∈[n] j∈[m] k∈[n]
Program (11) with objective (12) corresponds
to the maximum
weight version
P
P
′
/n)× i∈[n] (T̄i ) for solution
scaled by a constant. It is equal to k∈[n] (kwmax
z̄, which is an upperbound of the objective value associated with ȳ of (10) and
hence an upperbound for the optimal value of (9).
Proceeding as for Theorem 1, we define a feasible solution of (9) based on z̄:
• r̄k = T̄n+1−k for all k ∈ [n], and
r̄k − T̄i if i ≥ n + 1 − k
¯
• dik =
for all i ∈ [n], for all k ∈ [n].
0
otherwise
P
P
′
)T̄i (see proof of TheThe objective value of this solution i∈[n] ( j∈[i] wn+1−j
P
P
′
/n)T̄i .
orem 1) is to be compared with upperbound i∈[n] ( k∈[n] kwmax
By comparing term by term w.r.t. T̄i for i ∈ [n], we can see that
the
worst case
P
′
/n).
happens to the term associated with T̄1 with the ratio wn′ /( k∈[n] kwmax
2w ′
Therefore, we obtain the ratio (n+1)wn ′ . This ratio is consistent since when
max
n = 1, the optimal solution of the maximum weight version coincides with the
optimum solution of (9).
By comparing term by term with respect to wk′ for k ∈ [n], we can seePthat the
n
worst case happens to the term associated with w1′ with the ratio nT̄n /( i=1 T̄i ),
which can be interpreted as the ratio of the smallest utility over the average
utility in the optimal solution of the maximum weight version. This ratio is consistent since in the case of equal utilities in the optimal solution of the maximum
weight version, the latter coincides with the optimum solution of (9).
⊓
⊔
An Efficient Primal-Dual Algorithm for Fair Combinatorial Optimization
4.2
11
Iterative Algorithm
The previous discussion motivates us to design an alternating optimization algorithm that starts with a feasible y for (10), computes the associated z and
uses the latter to iteratively improve y. Formally, it can be sketched as follows:
1: t ← 0
2: compute y (0)
3: repeat
4:
t←t+1
5:
solve 0, 1-LP (Pyt−1 ) to obtain feasible solution z (t)
6:
update y (t) based on y (t−1) and z (t)
(t)
7: until max iteration has been reached or change on yik is small
8: return z (t) with highest GGI
Interestingly, lines 2 and 6 can be performed in different ways. For line 2, an
initial y (0) can be obtained by solving the dual LP (10). Another approach is
to solve the maximum weight version of our combinatorial problem and get the
dual solution variables for y (0) . Note that Theorem 2 then provides a guarantee
on the final solution, as it is at least as good as that of the maximum weight
problem. For line 6, one approach is to solve (9) with z fixed to z (t) in order to
get dual solution variables y (t) . A better approach as observed in the experiments
and explained next is based on Lagrangian relaxation.
The Lagrangian relaxation of (9) with respect to constraint (9d) can be
written as follows with Lagrangian multipliers λ = (λik )i∈[n],k∈[n] :
L(λ) =
max.
X
(wk′ k −
X
λik )rk −
X
+
i∈[n]
X
(wk′ − λik )dik
(13a)
k∈[n] i∈[n]
i∈[n]
k∈[n]
X X
λik
X
(13b)
uij zij
j∈[m]
k∈[n]
s.t. Az ≤ b
(13c)
z ∈ {0, 1}n×m
(13d)
dik ≥ 0
(13e)
∀i ∈ [n], ∀k ∈ [n]
The Lagrangian dual of (13) is then given by:
min. L(λ) s.t. λik ≥ 0
∀i ∈ [n], ∀k ∈ [n]
(14)
n×n
For an optimal solution z ∗ , r ∗ , d∗ of the 0, 1-LP (9), we have for any λ ∈ R+
:
X
k∈[n]
wk′ (krk∗ −
X
i∈[n]
d∗ik ) ≤
X
(wk′ k −
λik )rk −
X
i∈[n]
X
k∈[n]
λik
X X
(wk′ − λik )dik
k∈[n] i∈[n]
i∈[n]
k∈[n]
+
X
X
uij zij ≤ L(λ)
j∈[m]
The first inequality holds because of the nonnegativity of λ and the feasibility
of z ∗ , r ∗ , d∗ . The second is true because of the maximization in (13). Therefore
12
Viet Hung Nguyen and Paul Weng
the best upperbound is provided by the solution of the Lagrangian dual (14),
though this problem is not easy to solve due to the integrality condition over z.
An inspection of program (13) leads to two observations: (i) it can be decomposed into two maximization problems, one over z and the other over r
and d; (ii) for program (13) to yield a useful upperbound, λ should satisfy two
constraints (otherwise L(λ) = ∞):
X
λik = kwk′ ∀k ∈ [n] and λik ≤ wk′ ∀i ∈ [n], ∀k ∈ [n]
i∈[n]
Interestingly, in the above decomposition, the maximization problem over z corresponds to (Pλ ) and therefore λ can be identified to the dual variable y.
Based on those observations, line 6 can be performed as follows. Given λ
(or y), the upperbound L(λ) can be improved by updating λ so as to decrease
(13a), which can be simply done by a projected sub-gradient step:
X
uij zij )
∀i ∈ [n], k ∈ [n]
(15)
λ′ik ← λik − γ(rk − dik −
j∈[m]
′
λ ← arg min ||λ − λ||
(16)
λ∈L
where γ is the sub-gradient
of λ′ on
P step and (16)′ is the Euclidean projection
n×n
′
L = {λ ∈ R+ | ∀k ∈ [n], i∈[n] λik = kwk , ∀i ∈ [n], λik ≤ wk }.
Projection (16) can be performed efficiently by exploiting the structure of L:
X X
(λ′ik − λik )2
arg min ||λ′ − λ|| = arg min ||λ′ − λ||2 = arg min
λ∈L
= arg min
λk ∈Lk
λ∈L
X
i∈[n]
(λ′ik
λ∈L
i∈[n] k∈[n]
X λ′
λik
( ik′ − ′ )2 k∈[n]
− λik )2 k∈[n] = arg min
wk
wk
λk ∈Lk
(17)
i∈[n]
P
where Lk = {λk ∈ Rn+ | i∈[n] λik /wk′ = k, ∀i ∈ [n], λik /wk′ ≤ 1}. Equation (17)
states that projection (16) can be efficiently performed by n projections on
capped simplices [28]. The complexity of this step would be in O(n3 ), which is
much faster than solving the quadratic problem (16). Besides, the n projections
can be easily computed in a parallel way.
We can provide a simple interpretation to the variable λ (or y). Considering
programs (10) and (11), we can observe that y corresponds to an allocation of
weights wk′ ’s over the different component i’s. Indeed, an optimal solution of (10)
would yield an extreme point of L (for a given k ∈ [n], exactly k terms among
(y1k , . . . , ynk ) are equal to wk′ and the other ones are null). The projected subgradient method allows to search for an optimal solution of our fair combinatorial
problem by moving inside the convex hull of those extreme points.
5
Experimental Results
We evaluated our method on two different problems: assignment and matching.
The LPs and 0, 1-LPs were solved using CPLEX 12.7 on a PC (Intel Core i7-6700
An Efficient Primal-Dual Algorithm for Fair Combinatorial Optimization
Instance
CPLEX
CPU1 CPU2
1.02
1.02
3.14
3.14
64.95 14.26
1054.14 100.23
0.89
0.89
8.83
8.83
590.66 45.93
1.55
1.55
342.78 342.78
AlterOpt
CPU Gap
0.23 0%
0.26 0%
0.45 0.28%
0.65 0.26%
0.2 0%
0.3 0.015%
0.48 0.13%
0.18 0%
0.94 0%
13
Instance
CPLEX
AlterOpt
CPU1 CPU2 CPU Gap
v50-20
v50-30 0.86
0.86 0.79 0%
v50-30
v50-40 2.43
2.43 1.42 0%
v50-40
v50-50 5.14
5.14 2.67 0%
v50-50
v50-60 148.5
25.45 13.43 0.01%
v30-20
v50-70 2406.02 1282.8 17.71 0.005%
v30-30
v30-30 1.15
1.15 0.78 0%
v30-40
v30-40 7.13
7.13 1.44 0%
v10-20
v30-50 81.75
75.5 2.45 0.01%
v10-30
v30-60 1003.69 615.16 12.8 0.036%
v10-30 5.33
5.33 0.76 0%
v10-40 1325.7 806.8 1.4 0.06%
v10-50 29617.78 3370.7 2.48 0.053%
Table 1. Numerical results for (left) assignment and (right) general matching problems
3.40GHz) with 4 cores and 8 threads and 32 GB of RAM. Default parameters of
CPLEX were used with 8 threads. The sub-gradient step γt is computed followt
ing the scheme: γt := (val(zt )−bestvalue)ρ
where val(zt ) is the objective value of
sqn
the program (11) with solution zt , bestvalue is the best known objective value
of the program (9) so far and sqn is the square of the Euclidean norm of the
subgradient vector. The parameter ρt is divided by two every 3 consecutive iterations in which the upperbound L(λ) has not been improved. The GGI weights
were defined as follows: wk = 1/k 2 for k ∈ [n] so that they decrease fast in order
to enforce more balanced solutions.
Assignment To demonstrate the efficiency of our heuristic method, we generate
hard random instances for the assignment problem. A random instance of this
problem corresponds to a random generation of the uij ’s, which are generated
as follows. For all i ∈ [n], ui1 follows a uniform distribution over [100] and for
all j ∈ [n], uij = ui1 + ǫ where ǫ is a random variable following a uniform distribution over integers between −d and d (with d a positive integer parameter).
With such a generation scheme, agents’ preferences over objects are positively
correlated and the solution of the fair optimization problem is harder due to the
difficulty of finding a feasible solution that satisfies everyone.
Matching We use the lemon library [7] for solving the maximum weight matching problem. For the generation of the matching problem (in a complete graph
with 2n nodes), we follow a similar idea to the assignment problem. Recall we
only need uij (and zij ) for i < j. For all i ∈ [n], for all j ∈ [n] with i < j,
uij = −1000. For all i ∈ [n], ui,n+1 follows a uniform distribution over [100] and
for all j ≥ max(i + 1, n + 2), uij = ui,n+1 + ǫ where ǫ is defined as above.
Explanations The name of the instances is of the form ”vd-x” where d denotes
the deviation parameter mentioned above and x the number of the vertices of the
graphs (i.e., n = x/2). Column “CPLEX” regroups CPLEX’s results. Subcol-
14
Viet Hung Nguyen and Paul Weng
umn “CPU1” reports the time (in seconds) that CPLEX spent to solve program
(9) to optimal. Subcolumn “CPU2” reports the times needed by the primal
heuristic of CPLEX to obtain a feasible integer solution that is better than or
equal to the solution given by our algorithm. Column “AlterOpt” reports our
algorithm’s results. Subcolumn “CPU” is the time spent by our algorithm. Subcolumn “Gap” reports the gap in percentage between Sol and Opt, which is
equal to (Opt − Sol) × 100/Opt% where Opt is the optimal value and Sol is the
value of the solution given by our algorithm. The times and the gaps reported
are averaged over 10 executions corresponding to 10 random instances.
Table 1 shows that the CPU time spent by CPLEX (subcolumn CPU1) for
solving program (9) increases exponentially with n and can quickly reach up to
around 10 hours. Moreover, the smaller the deviation x, the more difficult the
problem. For example, for x = 50, we cannot solve instances with more than
50 and more than 70 vertices for respectively the fair assignment and general
matching problems within 10 hours of CPU time. For x = 10, this limit is
respectively 30 and 50 vertices. In contrast, the CPU time spent by our algorithm
(subcolumn CPU) seems to increase linearly with n and remains within tens or
so seconds. The quality of the solutions output by our algorithm is very good
as the gap is at maximum around 0.3% for fair assignment. This is even better
for fair general matching, in all cases the gap is smaller than 0.1%. Moreover,
the CPU time that CPLEX needs to find a feasible integer solution of similar
quality by primal heuristic is much longer than the CPU time of our algorithm
(up to hundreds times longer). It is interesting to notice that the fair assignment
seems to be more difficult in our experiments than the fair general matching.
This contrasts with the classical maximum weight version where the assignment
problem is generally easier than the general maximum matching.
6
Conclusion
We formulated the fair optimization with the Generalized Gini Index for a large
class of combinatorial problem for which we proposed a primal-dual algorithm
based on a Lagrangian decomposition. We demonstrated its efficiency on several
problems. We also provided some theoretical bounds on its performance. As
future work, we plan to improve those bounds and investigate other updates
for the Lagrangian multipliers. Another interesting direction is to consider other
linearization techniques such as the one proposed by Chassein and Goerigk [5].
Finally, we will also apply our method to problems whose maximum weight
version is NP-hard.
References
1. Anand, S.: The multi-criteria bipartite matching problem (2006)
2. Arnold, B.: Majorization and the Lorenz Order. Springer (1987)
3. Bansal, N., Sviridenko, M.: The Santa Claus problem. In: STOC. pp. 31–40 (2006)
An Efficient Primal-Dual Algorithm for Fair Combinatorial Optimization
15
4. Bezakova, I., Dani, V.: Allocating indivisible goods. ACM SIGecom Exchanges
5(3), 11–18 (2005)
5. Chassein, A., Goerigk, M.: Alternative formulations for the ordered weighted averaging objective. Information Processing Letters pp. 604–608 (2015)
6. Dachert, K., Gorski, J., Klamroth, K.: An augmented weighted Tchebycheff
method with adaptively chosen parameters for discrete bicriteria optimization
problems. Computers and Operations Research 39(12), 2929–2943 (2012)
7. Dezs, B., Juttner, A., Kovacs, P.: LEMON - an open source C++ graph template
library. Electronic notes in theoretical computer science 264(5), 23–45 (2011)
8. Edmonds, J.: Maximum matching and a polyhedron with 0, 1-vertices. Journal
Res Nat Bureau Standards 69B, 125–130 (1965)
9. Gilbert, H., Spanjaard O.: A Game-Theoretic View of Randomized Fair MultiAgent Optimization. IJCAI Algorithmic Game Theory Workshop (2017)
10. Hurkala, J., Sliwinski, T.: Fair flow optimization with advanced aggregation operators in wireless mesh networks. In: Federated Conference on Computer Science
and Information Systems. pp. 415–421 (2012)
11. Kostreva, M., Ogryczak, W., Wierzbicki, A.: Equitable aggregations and multiple
criteria analysis. Eur. J. Operational Research 158, 362–367 (2004)
12. Lesca, J., Perny, P.: Lp solvable models for multiagent fair allocation problems. In:
ECAI (2011)
13. Lovász, L., Plummer, M.: Matching Theory. North Holland (1986)
14. Luss, H.: Equitable Resource Allocation. Wiley (2012)
15. Lust, T., Teghem, J.: The multiobjective multidimensional knapsack problem: a
survey and a new approach. Intl. Trans. in Op. Res. pp. 1–26 (2012)
16. Moulin, H.: Axioms of cooperative decision making. Cambridge Univ. Press (1988)
17. Moulin, H.: Fair Division and Collective Welfare. MIT Press (2004)
18. Ogryczak, W., Luss, H., Pióro, M., Nace, D., Tomaszewski, A.: Fair optimization
and networks: A survey. Journal of Applied Mathematics 2014 (2014)
19. Ogryczak, W., Perny, P., Weng, P.: On minimizing ordered weighted regrets in
multiobjective Markov decision processes. In: ADT. Lecture Notes in Artificial
Intelligence, vol. 6992, pp. 190–204. Springer (2011)
20. Ogryczak, W., Perny, P., Weng, P.: A compromise programming approach to multiobjective Markov decision processes. IJITDM 12, 1021–1053 (2013)
21. Ogryczak, W., Sliwinski, T.: On solving linear programs with the ordered weighted
averaging objective. Eur. J. Operational Research 148, 80–91 (2003)
22. Ogryczak, W., Sliwinski, T., Wierzbicki, A.: Fair resource allocation schemes and
network dimensioning problems. J. of Telecom. & Info. Tech. 2003(3), 34–42 (2003)
23. Perny, P., Weng, P.: On finding compromise solutions in multiobjective Markov
decision processes. In: ECAI (short paper) (2010)
24. Rawls, J.: The Theory of Justice. Havard university press (1971)
25. Rodera, H., Bagajewicz, M.J., Trafalis, T.B.: Mixed-integer multiobjective process
planning under uncertainty. Ind. Eng. Chem. Res. 41(16), 4075–4084 (2002)
26. Schrijver, A.: Theory of Linear and Integer Programming. John Wiley (1998)
27. Steuer, R.: Multiple criteria optimization. John Wiley (1986)
28. Wang, W., Lu, C.: Projection onto the capped simplex (2015), arXiv:1503.01002
29. Weymark, J.: Generalized Gini inequality indices. Math. Social Sciences 1 (1981)
30. Wierzbicki, A.: A mathematical basis for satisficing decision making. Mathematical
Modelling 3, 391–405 (1982)
31. Yager, R.: On ordered weighted averaging aggregation operators in multi-criteria
decision making. IEEE Trans. on Syst., Man and Cyb. 18, 183–190 (1988)
| 8cs.DS
|
ON GEOMETRICALLY TRANSITIVE HOPF ALGEBROIDS
arXiv:1508.05761v3 [math.AC] 29 Dec 2017
LAIACHI EL KAOUTIT
Abstract. This paper contributes to the characterization of a certain class of commutative Hopf algebroids. It
is shown that a commutative flat Hopf algebroid with a non zero base ring and a nonempty character groupoid
is geometrically transitive if and only if any base change morphism is a weak equivalence (in particular, if any
extension of the base ring is Landweber exact), if and only if any trivial bundle is a principal bi-bundle, and if
and only if any two objects are fpqc locally isomorphic. As a consequence, any two isotropy Hopf algebras of a
geometrically transitive Hopf algebroid (as above) are weakly equivalent. Furthermore, the character groupoid
is transitive and any two isotropy Hopf algebras are conjugated. Several other characterizations of these Hopf
algebroids in relation to transitive groupoids are also given.
Contents
1. Introduction
1.1. Motivation and overview
1.2. Description of the main results
2. Abstract groupoids: General notions and basic properties
2.1. Notations, basic notions and examples
2.2. Groupoids actions, equivariant maps and the orbits sets
2.3. Principal groupoid-bisets and the two sided translation groupoid
2.4. Principal groupoids-biset versus weak equivalences
2.5. Transitive groupoids are characterized by weak equivalences
2.6. Correspondence between transitive groupoids and principal group-sets
3. Hopf algebroids: comodules algebras, principal bundles, and weak equivalences
3.1. Preliminaries and basic notations
3.2. The 2-category of Hopf algebroids
3.3. Comodules, bicomodules (algebras), and presheaves of orbit sets
3.4. Weak equivalences and principal bundles between Hopf algebroids
3.5. Dualizable right comodules
3.6. Dualizable comodule whose endomorphism ring is a principal bundle
4. Geometrically transitive Hopf algebroid: Definition, basic properties and the result
4.1. Definition and basic properties
4.2. Characterization by means of weak equivalences
5. More properties of geometrically transitive Hopf algebroids
5.1. The isotropy Hopf algebras are weakly equivalent
5.2. The transitivity of the character groupoid
5.3. GT Hopf algebroids and principal bundles over Hopf algebras
5.4. GT Hopf algebroids with empty character groupoid
5.5. Dualizable comodules over GT Hopf algebroids are locally free of constant rank
References
1
1
3
5
5
7
7
9
10
11
11
12
12
15
16
17
18
18
18
21
22
23
25
26
27
29
30
1. Introduction
1.1. Motivation and overview. A commutative Hopf algebroid can be thought as an affine groupoid
scheme, that is, a groupoid scheme [10, Définition page 299] in which the schemes defining objects and
morphisms are affine schemes. In other words, this is a representable presheaf of groupoids in the category
of affine schemes, or a prestack of groupoids whose ”stackification” leads to a stack in the fpqc (fidèlement
plate et quasi-compacte) topology. For instance, an action of an affine group scheme on an affine scheme
Date: January 3, 2018.
2010 Mathematics Subject Classification. Primary 16D90, 16T15, 18B40, 18D35, 18D10; Secondary 14M17, 20L05, 58H05.
Key words and phrases. Transitive groupoids; groupoid bisets; weak equivalences; Landweber exactness; geometrically transitive
Hopf algebroids; principal bi-bundles; isotropy Hopf algebras; character groupoid.
Research supported by the Spanish Ministerio de Economı́a y Competitividad and FEDER, grants MTM2016-77033-P and
MTM2013-41992-P.
1
2
LAIACHI EL KAOUTIT
leads to an affine groupoid scheme which gives rise (by passage to the coordinate rings) to a commutative
Hopf algebroid, known as split Hopf algebroid ([26, Appendix A.1], see also [19]). More examples of
commutative Hopf algebroids can be performed using Set-theoretically constructions in groupoids.
Hopf algebroids in relation with groupoids are fundamental objects in both algebraic topology and
algebraic geometry. They appear in the study of stable homotopy theory [26, 17, 24, 25] (see also the
references therein), and prove to be very useful in studying quotients of preschemes, prestacks of groupoids
over affine schemes as well as (commutative) Tannakian categories [14, 9, 4, 1, 20].
As in the case of affine group schemes [10], several constructions and results on groupoids have a certain
geometric meaning in presheaves of groupoids and then a possible algebraic interpretation at the level of
Hopf algebroids. In this way, Hopf algebroids are better understood when looking to classical results in
groupoids, or by mimicking well-known results on classical geometric groupoids, e.g. topological or Lie
groupoids. The main motivation of this paper fits into these lines of research. Specifically, we study a
class of commutative Hopf algebroids called geometrically transitive (see below), by means of transitive
groupoids and their properties, obtaining in this way several new characterizations of these Hopf algebroids.
Besides, much of the properties of transitive groupoids hereby developed and used in the study of Hopf
algebroids can be also seen as a contribution to the theory of groupoids.
The notion of transitivity varies depending on the context. In groupoid theory, a (small) groupoid is said
to be transitive when the cartesian product of the source and the target is a surjective map. A Lie groupoid
is called locally trivial (or a transitive Lie groupoid), when this map is a surjective submersion [21, 8].
For groupoid schemes, the meaning of the abstract notion of transitivity was introduced by Deligne in [9,
page 114]. More precisely, a groupoid scheme is transitive in the fpqc topology sense if the morphism
constituted by the fibred product of the source and the target is a cover in this topology. In [4, Définition
page 5850], Bruguières introduced a class of Hopf algebroids referred to as geometrically transitive, where
he showed ([4, Théorème 8.2 page 5858]) that in the commutative case (the case which we are interested
in) these are Hopf algebroids whose associated affine groupoid schemes are transitive in the fpqc sense. It
is also implicitly shown in [4] that a commutative Hopf algebroid is geometrically transitive if and only
if the unit map (i.e., the tensor product of the source and the target) is a faithfully flat extension. This, in
fact, can be thought of as a proper definition of geometrically transitive (GT for short) commutative Hopf
algebroids. Nevertheless, we will use here the original definition of [4] (see Definition 4.2 below) and show
using elementary methods that the faithful flatness of the unit characterizes in fact GT Hopf algebroids with
nonempty character groupoids (see Theorem A below).
Transitive groupoids are also characterized by the fact that any two objects are isomorphic, or equivalently: a groupoid with only one connected component, or connected groupoid in the terminology of [16].
From the geometric point of view, that is, for presheaves of groupoids, this means that any two objects
are locally isomorphic in the fpqc topology, see [9, Proposition 3.3]. At the level of Hopf algebroids, this
property can be directly expressed in terms of faithfully flat extensions (see Definition 3.6 below), which
in turn characterizes GT Hopf algebroids with nonempty character groupoids, as we will see in the sequel
by using elementary (algebraic) arguments. Our methods, in fact, lead us also to recover other results on
equivalences of categories stated in [9, §3.5] (see the paragraph after Theorem A).
From their very definition one can then see that GT Hopf algebroids can be understood as a natural
algebro-geometric substitute of transitive groupoids. Under this point of view, it is reasonable to expect
that most of the properties or characterizations of transitive groupoids could have an analogous counterpart
at the level of GT Hopf algebroids. However, apart from the basic definition, there are still several characterizations of transitive groupoids which, up to our knowledge, are not known for GT Hopf algebroids. In
the following, we describe the two most interesting of these characterizations.
/
s
ι
A perhaps well-known result (see Proposition 2.15 for details) says that a groupoid G : G1 o
/ G0
t
ς
is transitive if and only if for any map ς : X → G0 the induced morphism of groupoids G → G is a weak
equivalence (i.e., an essentially surjective and fully faithful functor), where G ς is the induced groupoid
whose set of objects is X and its set of arrows is the fibred product X ς × t G1 s × ς X, that is, G ς is the pullback groupoid of G along ς (see [21, §2.3] and [16] for a dual construction). Another more interesting
and perhaps not yet known characterization of the transitivity is by means of principal groupoid-bisets;
for the precise definition see Definitions 2.6, 2.7, and 2.8. This notion is in fact an abstract formulation
of the notion of principal bi-bundles in the context of topological and Lie groupoids [22, 18], or that of
ON GEOMETRICALLY TRANSITIVE HOPF ALGEBROIDS
3
bi-torsors in sheaf theory [10, 15], which is of course based on the natural generalization of the notion
of group-bisets [7] to the context of groupoids. The aforementioned characterization can be expressed as
follows: a groupoid G is transitive if and only if for any map ς : X → G0 the pull-back groupoid-biset
G1 s × ς X is a principal (G , G ς )-biset, see again Proposition 2.15.
The main aim of this paper is to investigate GT Hopf algebroids by means of transitive groupoids. Our
aim is in part to see how the previous characterizations of transitive groupoids can be transferred, by means
of weak equivalences and principal groupoids-bisets, to the commutative Hopf algebroid framework.
1.2. Description of the main results. Let k be the ground field. The term algebra in the following stands
for a commutative k-algebra, as usual the unadorned tensor product ⊗ denotes the tensor product over k.
Our main result is summarized in the following theorem, which includes Theorem 4.8 below:
Theorem A. Let (A, H) be a commutative flat Hopf algebroid over k with A , 0 and denote by H its
associated presheaf of groupoids. Assume that H0 (k) , ∅. Then the following are equivalent:
(i) η = s ⊗ t : A ⊗ A → H is a faithfully flat extension;
(ii) Any two objects of H are fpqc locally isomorphic (see Definition 3.6);
(iii) For any extension φ : A → B, the extension α : A → Ht ⊗A φ B, a 7→ s(a) ⊗A 1B is faithfully flat;
(iv) (A, H) is geometrically transitive (see Definition 4.2);
(v) For any extension φ : A → B, the canonical morphism φ : (A, H) → (B, Hφ) of Hopf algebroids
is a weak equivalence, that is, the induced functor φ∗ : ComodH → ComodHφ is an equivalence of
symmetric monoidal k-linear categories of comodules, where Hφ = B ⊗A H ⊗A B;
(vi) For any extension φ : A → B, the trivial principal left (H, Hφ )-bundle H ⊗A B is a principal bi-bundle
(see subsection 3.4).
The flatness is to ensure that the categories of comodules of the involved Hopf algebroids are Grothendieck
ones with exact forgetful functor to the category of modules over the base algebra. As for the hypothesis
A , 0, one of the conditions in Definition 4.2 below says that the endomorphisms ring of A viewed as a
comodule should coincides with k, so it is reasonable to ask A to be a non zero object.
The examples of Hopf algebroids we have in mind and satisfy the assumptions of Theorem A, are the
ones which can be obtained from Tannaka’s reconstruction process, by using k-linear representations of a
(small) groupoid G and taking A = Map(G0 , k), the set of all maps from G0 to k, as the base algebra.
The proof of Theorem A is done by showing the implications (i) ⇒ (ii) ⇒ (iii) ⇒ (iv) ⇒ (i), and using
the equivalences (iii) ⇔ (v) ⇔ (vi) from [11, Proposition 5.1]. The assumption H0 (k) , ∅ which ensures
that the character groupoid H (k) is nonempty, is needed to prove the implication (iii) ⇒ (iv). Although,
the implication (ii) ⇒ (iii) can be shown under the weaker assumption H0 (L) , ∅ for some field extension
L of k. Moreover, the equivalent conditions of Theorem A are stable under change of scalars for the class of
Hopf algebroids with nonempty character groupoids (i.e., with H0 (k) , ∅). Indeed, if L is a field extension
of k and (A, H) is a Hopf algebroid which verifies the assumptions of Theorem A and satisfies one of these
conditions say (i), then the Hopf algebroid (AL , HL ) = (A ⊗k L, H ⊗k L) satisfies this condition as well.
However, if H0 (k) = ∅ and there is a field extension L such that H0 (L) , ∅, then it is not clear to us how to
show the implication (iii) ⇒ (iv), see Remarks 3.7 and 4.10 for more comments on the change of scalars.
As we have mention before, saying that (A, H) is GT Hopf algebroid is equivalent to say that the kgroupoid H1 acts transitively on H0 in the fpqc sense. Thus, by Theorem A, this can be now easily
deduced by comparing condition (i) with the Definition of [9, §1.6]. On the other hand, there is also a
notion of transitivity for Hopf algebroids [4, Définition page 5850] and as was shown in [4, Proposition
7.3 page 5851], a Hopf algebroid (A, H) is geometrically transitive if and only if (AL , HL ) is transitive for
any field extension L of k. Perhaps this justifies the terminology, although it is not clear, at least to us, how
to express this transitivity at the level of the associated presheaves of groupoids with respect to a certain
topology, see Remark 4.5 for more details.
Condition (v) in Theorem A implies in particular that any algebra extension B is Landweber exact over
A in the sense of [17, Definition 2.1] and shows that certain GT Hopf algebroids do not have a non trivial
hereditary torsion theory in the sense of [17, Theorem A]. On the other hand, following the notation of
[9, §3.5], we know that the category of comodules ComodH is canonically equivalent, as a symmetric
monoidal k-linear category, to the category of linear representations Rep(H0 : H1 ) of the associated affine
presheaf of groupoids. So, up to these canonical equivalences, condition (v) gives the ”affine version” of
4
LAIACHI EL KAOUTIT
the equivalence of categories stated in [9, (3.5.1) page 130]. Furthermore, Theorem A shows that for affine
k-schemes with the induced fpqc topology, the equivalence of categories stated in op.cit., is not only a
necessary condition to have a transitive action (for the class of affine k-groupoids H with H0 (k) , ∅), but
also a sufficient one. Thus we obtain a new characterization of these transitive affine k-groupoids.
The fact that transitive groupoids are characterized by the conjugacy of their isotropy groups, and the
analogue of this characterization in the Hopf algebroid context also attracted our attention. More precisely,
given a Hopf algebroid (A, H) and denote by H (C) the fiber of H at a commutative algebra C, that is,
the groupoid with set of objects H0 (C) = Algk (A, C) and set of arrows H1 (C) = Algk (H, C) (see equation
(13) below). Assume as above that the character groupoid H (k) in nonempty (i.e., H0 (k) , ∅), then for
any object x : A → k, there is a presheaf of sets which assigns to each algebra C, the isotropy group of the
object x∗ (1C ) ∈ H0 (C), where 1C : k → C denotes the unit of C. It turns out that this is an affine group
scheme represented by the Hopf algebra (kx , Hx ) which is the base change of (A, H) by the algebra map x
(here kx denotes k viewed as an A-algebra via x, and Hx = kx ⊗A H ⊗A kx ). The pair (kx , Hx ) is refereed to
as the isotropy Hopf algebra of (A, H) at (the point) x. Now, recall from [17] that two flat Hopf algebroids
(A, H) and (B, K) are weakly equivalent if there is a diagram of weak equivalences
(A, H)
(C, J) i❙
❙❙❙❙
❦❦5
❙
❦❦❦❦
(B, K),
see Subsection 3.4 for more details. The fact that two isotropy groups of a transitive groupoid are isomorphic is translated to the fact that two isotropy Hopf algebras of a GT Hopf algebroid are weakly equivalent;
of course, Hopf algebras are considered here as Hopf algebroids where source and target coincide.
This result is part of the subsequent corollary of Theorem A which contains both Proposition 5.3 and
Corollary 5.17; the last statement in part (2) below follows from Propositions 4.1(GT13) and 4.6.
Corollary A. Let (A, H) be a flat Hopf algebroid as in Theorem A with H0 (k) , ∅. Assume that (A, H) is
geometrically transitive. Then
(1) Any two isotropy Hopf algebras are weakly equivalent.
(2) Any dualizable (right) H-comodule is locally free of constant rank. Moreover, any right H-comodule
is an inductive limit of dualizable right H-comodules.
The notion of the conjugation relation between two isotropy Hopf algebras is not automatic. This relation can be formulated by using 2-isomorphisms in the 2-category of flat Hopf algebroids. More specifically, using the notations and the assumptions of Theorem A, for a given Hopf algebroid (A, H), two
isotropy Hopf algebras (kx , Hx ) and (ky , Hy ), at the points x, y ∈ H0 (k) are said to be conjugated provided
there is an isomorphism g : (kx , Hx ) → (ky , Hy ) of Hopf algebras such that the following diagram
(kx , Hx ) i❙
❙❙❙❙
❙
x
g
(A, H)
/ (k , H )
❦5 y y
❦
❦
❦
❦❦ y
commutes up to a 2-isomorphism, where x and y are the canonical morphisms attached, respectively, to
x and y. The transitivity of the conjugation relation characterizes in fact the transitivity of the character
groupoid. This result is also a corollary of Theorem A and stated as Proposition 5.9 below:
Corollary B. Let (A, H) be a flat Hopf algebroid as in Theorem A with H0 (k) , ∅. Assume that (A, H) is
geometrically transitive. Then the following are equivalent
(i) the character groupoid H (k) is transitive;
(ii) for any two objects x, y in H0 (k), the left H-comodule algebras H ⊗A kx and H ⊗A ky are isomorphic;
(iii) any two isotropy Hopf algebras are conjugated.
Furthermore, under the stated assumptions, we have that condition (i) is always fulfilled.
Let (A, H) be a flat Hopf algebroid over k with A , 0 and H0 (k) = ∅. If (A, H) satisfies condition (i) of
Theorem A and there exists a field extension L of k such that H0 (L) , ∅, then Corollary B can be applied
to (AL , HL ) and implies that H (L) is a transitive groupoid, see Proposition 5.14.
Transitive groupoids are related to principal group-bisets, in the sense that there is a (non canonical)
correspondence between these two notions, see Subsection 2.6. This in fact is an abstract formulation
ON GEOMETRICALLY TRANSITIVE HOPF ALGEBROIDS
5
of Ehresmann’s well-known result dealing with the correspondence between transitive Lie groupoids and
principal fibre bundles, as was expounded in [8]. The analogous correspondence at the level of Hopf
algebroids is not always possible and some technical assumptions are required. The formulation of this
result is given as follows.
For any object x ∈ H0 (k), consider the presheaf of sets which associates to each algebra C the set
Px (C) := t−1 {x∗ (1C )} , where t is the target of the groupoid H (C). In the terminology of [16], this is
the left star set of the object x∗ (1C ). Denote by αx : A → Px := H ⊗A kx the algebra map which sends
a 7→ s(a) ⊗ 1. It turns out that the presheaf of sets Px is affine, and is up to a natural isomorphism
represented by the (H, Hx )-bicomodule algebra Px . The subsequent is a corollary of Theorem A, and
formulates the desired result. It is a combination of Lemma 5.10 and Proposition 5.11 below.
Corollary C. Let (A, H) be a flat Hopf algebroid as in Theorem A with H0 (k) , ∅. If (A, H) is a GT Hopf
algebroid, then for any object x ∈ H0 (k), the comodule algebra (Px , αx ) is a right principal Hx -bundle (i.e.,
a Hopf-Galois extension).
Conversely, let (P, α) be a right principal B-bundle over a Hopf k-algebra B with extension α : A → P.
Denote by υ : H := (P ⊗ P)coinvB → P ⊗ P the canonical injection, where P ⊗ P is a right B-comodule
algebra via the diagonal coaction (here RcoinvB denotes the subalgebra of coinvariant elements of a right
B-comodule algebra R). Assume that υ is a faithfully flat extension and that
coinvB
H ⊗A H = (P ⊗ P) ⊗A (P ⊗ P)
,
where (P ⊗ P) ⊗A (P ⊗ P) is endowed with a canonical right B-comodule algebra structure. Then the pair
of algebras (A, H) admits a unique structure of a GT Hopf algebroid such that (α, υ) : (A, H) → (P, P ⊗ P)
is a morphism of GT Hopf algebroids.
2. Abstract groupoids: General notions and basic properties
This section contains the results about groupoids, which we want to transfer to the context of Hopf
algebroids in the forthcoming sections. For sake of completeness we include some of their proofs.
2.1. Notations, basic notions and examples. A groupoid (or abstract groupoid) is a small category where
/
s
ι
each morphism is an isomorphism. That is, a pair of two sets G := (G1 , G0 ) with diagram G1 o
/ G0 ,
t
where s and t are resp. the source and the target of a given arrow, and ι assigns to each object its identity
arrow; together with an associative and unital multiplication G2 := G1 s × t G1 → G1 as well as a map
G1 → G1 which associates to each arrow its inverse.
Given a groupoid G , consider an object x ∈ G0 , the isotropy group of G at x, is the group of loops:
n
o
G x := g ∈ G1 | s(g) = t(g) = x .
(1)
U
x
Notice that the disjoint union x ∈ G0 G of all isotropy groups form the set of arrows of a subgroupoid of
U
G whose source equal to its target, namely, the projection x ∈ G0 G x → G0 .
A morphism of groupoids φ : H → G is a functor between the underlying categories. That is, φ =
(φ0 , φ1) ), where φ0 : H0 → G0 and φ1 : H1 → G1 satisfying the pertinent compatibility conditions:
φ1 ◦ ι = ι ◦ φ0 , φ0 ◦ s = s ◦ φ1 , φ0 ◦ t = t ◦ φ1 , φ1 ( f g) = φ1 ( f )φ1 (g),
whenever the multiplication f g in H1 is permitted.
Obviously any such a morphism induces morphisms between the isotropy groups: φu : H u → G φ0 (u) , for
every u ∈ H0 . Naturally, groupoids, morphism of groupoids, and natural transformations form a 2-category
Grpds. Next we describe some typical examples of groupoids and their morphisms.
Example 2.1. Let G be a group and fix a set M. Denote by BG M the category whose objects are (left) Gtorsors of the form (P, G, M) and morphisms are G-morphisms. In the terminology of Definition 2.8 below,
an object (P, G, M) in BG M is a principal left G-set P. Precisely, this is a left G-set P with projection π :
P → M to the set of orbits M and where the canonical map G × P → P π × π P, (g, p) 7→ (gp, p) is bijective.
It is clear that any morphism in this category is an isomorphism, thus, BG M is a groupoid (probably not
small). The groupoid BG pt plays a crucial role in the representation theory of the group G. Furthermore,
when M varies in the category Sets of sets, we obtain the presheaf of groupoids BG : Setsop −→ Grpds,
which is known as the classifying stack of the group G.
6
LAIACHI EL KAOUTIT
Example 2.2. Assume that R ⊆ X × X is an equivalence relation on a set X. One can construct a groupoid
/
pr
R o prι12 / X, with structure maps as follows. The source and the target are s = pr2 and t = pr1 , the
second and the first projections, and the map of identity arrows ι is the diagonal one. The multiplication
and the inverse maps use, respectively, the transitivity and reflexivity of R and are given by
(x, x′ ) (x′ , x′′ ) = (x, x′′ ),
and (x, x′ )−1 = (x′ , x).
This is an important class of groupoids known as the groupoid of equivalence relation, see [10, Exemple
1.4, page 301]. A particular situation is when R = X × X, that is, the obtained groupoid is the so called the
groupoid of pairs (called fine groupoid in [3] and simplicial groupoid in [16]); or when R is defined by a
certain fibred product X ν × ν X for a map ν : X → Y.
Example 2.3. Any group G can be considered as a groupoid by taking G1 = G and G0 = {∗} (a set with
one element). Now if X is a right G-set with action ρ : X × G → X, then one can define the so called
the action groupoid: G1 = X × G and G0 = X, the source and the target are s = ρ and t = pr1 , the
identity map sends x 7→ (e, x) = ιx , where e is the identity element of G. The multiplication is given by
(x, g)(x′ , g′ ) = (x, gg′ ), whenever xg = x′ , and the inverse is defined by (x, g)−1 = (xg, g−1 ). Clearly the
pair of maps (pr2 , ∗) : (G1 , G0 ) → (G, {∗}) defines a morphism of groupoids.
Example 2.4. Let G = (G1 , G0 ) be a groupoid and ς : X → G0 a map. Consider the following pair of sets:
n
o
Gς 1 := X ς × t G1 s × ς X = (x, g, x′ ) ∈ X × G1 × X| ς(x) = t(g), ς(x′ ) = s(g) , Gς 0 := X.
Then G ς = (Gς 1 , Gς 0 ) is a groupoid, with structure maps: s = pr3 , t = pr1 , ιx = (ς(x), ις(x) , ς(x)), x ∈ X.
The multiplication is defined by (x, g, y)(x′, g′ , y′ ) = (x, gg′ , y′ ), whenever y = x′ , and the inverse is given
by (x, g, y)−1 = (y, g−1 , x). The groupoid G ς is known as the induced groupoid of G by the map ς, (or
the pull-back groupoid of G along ς, see [16] for dual notion). Clearly, there is a canonical morphism
φς := (pr2 , ς) : G ς → G of groupoids.
Any morphism φ : H → G of groupoids factors through the canonical morphism G φ0 → G , that is we
have the following (strict) commutative diagram
φ
H
(
φ′
of groupoids, where φ′0 = idH0 and
φ′1 : H1 −→ Gφ0 1 ,
G φ0
/7 G
♦♦♦
♦
♦
♦
h 7−→ t(h), φ1(h), s(h) .
A particular and important example of an induced groupoid is the case when G is a groupoid with
one object, that is, a group. In this case, to any group G and a set X, one can associated the groupoid
(X × G × X, X) as the induced groupoid of (G, {∗}) by the map ∗ : X → {∗}.
Recall that a groupoid G = (G1 , G0 ) is said to be transitive if the map (s, t) : G1 → G0 × G0 is surjective.
Example 2.5. The groupoid of pairs is clearly transitive, as well as any induced groupoid of the form
(X × G × X, X). On other hand, if a group G acts transitively on a set X, then the associated action groupoid
is by construction transitive.
Let G be a transitive groupoid. Then if x, y ∈ G0 , there is a non-canonical isomorphism of groups
G x G y given by conjugation: Let g ∈ G1 with x = s(g) and t(g) = y, then
G x −→ G y , h 7−→ ghg−1
is an isomorphism of groups. This fact is essential in showing that any transitive groupoid is isomorphic, in
a non-canonical way, to an induced groupoid of the form (X ×G × X, X). Indeed, given a transitive groupoid
G , fix an object x ∈ G0 with isotropy group G x and chose a family of arrows { fy }y ∈ G0 such that fy ∈ t−1 ({x})
and s( fy ) = y, for y , x while fx = ι(x), for y = x. In this way the morphism
h
i
φx : G −→ (G0 × G x × G0 , G0 ),
(g, z) 7−→ s(g), ft(g) g fs−1
, t(g) , z
(g)
establishes an isomorphism of groupoids.
ON GEOMETRICALLY TRANSITIVE HOPF ALGEBROIDS
7
2.2. Groupoids actions, equivariant maps and the orbits sets. The following definition is a natural
generalization to the context of groupoids, of the usual notion of group-set. It is an abstract formulation
of that given in [21, Definition 1.6.1] for Lie groupoids, and essentially the same definition based on the
Sets-bundles notion given in [27, Definition 1.11].
Definition 2.6. Given a groupoid G , a right G -set is a triple (X, ς, ρ) where X is a set and ς : X → G0 and
ρ : X ς × t G1 → X (shortly written as ρ(x, g) := xg) are the structure and action maps respectively. These
maps obey the following conditions
(1) s(g) = ς(xg), for any x ∈ X and g ∈ G1 with ς(x) = t(g).
(2) xις(x) = x, for every x ∈ X.
(3) (xg)h = x(gh), for every x ∈ X, g, h ∈ G1 with ς(x) = t(g) and t(h) = s(g).
In order to simplify the notation we denote a right G -set by a pair (X, ς), omitting the action ρ. A left
groupoid action is analogously defined by interchanging the source with the target and similar notations
might be adopted. Obviously, any groupoid G acts over itself on both sides by using the regular action, i.e.,
the multiplication G1 s × t G1 → G1 . Thus, we have that (G1 , s) is a right G -set and (G1 ,t) is a left G -set.
Let (X, ς) be a right G -set, and consider the pair of sets X Y G := X ς × t G1 , X as a groupoid with
structure maps s = ρ, t = pr1 , ιx = (x, ις(x) ). The multiplication and the inverse maps are defined by
(x, g)(x′ , g′ ) = (x, gg′ ) and (x, g)−1 = (xg, g−1 ). The groupoid X Y G is known as the right translation
groupoid of X by G .
For sake of completeness let us recall the notion of equivariant maps. A morphism of right G -sets (or
G -equivariant map) F : (X, ς) → (X ′ , ς′ ) is a map F : X → X ′ such that the diagrams
X
ς ♦♦
♦
♦
♦
w♦
F
G0 g❖❖
❖❖❖
❖
′
ς
X′
X ς × t G1
/X
F
F × id
X ′ ς′ × t G 1
/ X′
(2)
commute. Clearly any such a G -equivariant map induces a morphism of groupoids F : X Y G → X ′ Y G .
Next we recall the notion of the orbit set attached to a right groupoid-set. This notion is a generalization
of the orbit set in the context of group-sets. Here we use the (right) translation groupoid to introduce this
set. First we recall the notion of the orbit set of a given groupoid. The orbit set of a groupoid G is the
quotient set of G0 by the following equivalence relation:
x ∼ y ⇐⇒ ∃ g ∈ G1 such that s(g) = x and t(g) = y.
In others words, this is the set of all connected components of G .
Given a right G -set (X, ς), the orbit set X/G of (X, ς) is the orbit set of the translation groupoid X Y G .
If G = (X × G, X) is an action groupoid as in Example 2.3, then obviously the orbit set of this groupoid
coincides with the classical set X/G of orbits.
2.3. Principal groupoid-bisets and the two sided translation groupoid. We give in this subsection an
exhaustive survey on principal groupoids bisets and the formal constructions of their bicategories.
Let G and H be two groupoids and (X, ς, ϑ) a triple consisting of a set X and two maps ς : X → G0 ,
ϑ : X → H0 . The following definitions are abstract formulations of those given in [18, 22] for topological
and Lie groupoids.
Definition 2.7. An (H , G )-biset is a triple (X, ς, ϑ) where (X, ς) is endowed with a right G -action ρ :
X ς × t G1 → X and (X, ϑ) with a left H -action λ : H1 s × ϑ X → X such that
(1) For any x ∈ X, h ∈ H1 , g ∈ G1 with ϑ(x) = s(h) and ς(x) = t(g), we have
ϑ(xg) = ϑ(x) and ς(hx) = ς(x).
(2) For any x ∈ X, h ∈ H1 and g ∈ G1 with ς(x) = t(g), ϑ(x) = s(h), we have h(xg) = (hx)g.
The two sided translation groupoid associated to a given (H , G )-biset (X, ς, ϑ) is defined to be the
groupoid H X X Y G whose set of objects is X and set of arrows is
n
o
H1 s × ϑ X ς × s G1 = (h, x, g) ∈ H1 × X × G1 | s(h) = ϑ(x), s(g) = ς(x) .
8
LAIACHI EL KAOUTIT
The structure maps are:
s(h, x, g) = x,
t(h, x, g) = hxg−1 and ι x = (ιϑ(x) , x, ις(x) ).
The multiplication and the inverse are given by:
(h, x, g)(h′, x′ , g′ ) = (hh′ , x′ , gg′),
(h, x, g)−1 = (h−1 , hxg−1 , g−1 ).
Associated to a given (H , G )-biset (X, ς, ϑ), there are two canonical morphisms of groupoids:
Σ : H X X Y G −→ G ,
(h, x, g), y 7−→ g, ς(y) ,
Θ : H X X Y G −→ H ,
(h, x, g), y 7−→ h, ϑ(y) .
(3)
(4)
Definition 2.8. Let (X, ς, ϑ) be an (H , G )-biset. We say that (X, ς, ϑ) is a left principal (H , G )-biset if it
satisfies the following conditions:
(P-1) ς : X → G0 is surjective;
(P-2) the canonical map
∇ : H1 s × ϑ X −→ X ς × ς X,
(h, x) 7−→ (hx, x)
(5)
is bijective.
By condition (P-2) we consider the map δ := pr1 ◦ ∇−1 : X ς × ς X → H1 . This map clearly satisfies:
s δ(x, y) = ϑ(y)
(6)
δ(x, y)y =
x,
for any x, y ∈ X with ς(x) = ς(y);
(7)
δ(hx, x) =
h,
for h ∈ H1 , x ∈ X with s(h) = ϑ(x).
(8)
Equation (8), shows that the action is in fact free, that is, hx = x only when h = ιϑ(x) . The subsequent lemma
is also immediate from this definition.
Lemma 2.9. Let (X, ς, ϑ) be a left principal (H , G )-biset. Then the map ς induces a bijection between the
orbit set X/H and the set of objects G0 .
Analogously one defines right principal (H , G )-biset. A principal (H , G )-biset is both left and right
principal biset. For instance, (G1 , t, s) is a left and right principal (G , G )-biset, known as the unit principal
biset, which we denote by U (G ). More examples of left principal bisets can be performed, as in the
geometric case, by pulling back other left principal bisets. Precisely, assume we are given (X, ς, ϑ) a left
principal (H , G )-biset, and let ψ : K → G be a morphism of groupoids. Consider the set Y := X ϑ × ψ0 K0
together with maps pr2 : Y → K0 and e
ς := ς ◦ pr1 : Y → H0 . Then the triple (Y, e
ς, pr2 ) is an (H , K )-biset
with actions
λ : H1 s × ς Y −→ Y,
h, (x, u) 7−→ hx, u
(9)
ρ : Y eς × t K1 −→ Y,
(x, u), f 7−→ xψ1 ( f ), s( f ) ,
(10)
which is actually a left principal (H , K )-biset, and known as the pull-back principal biset of (X, ς, ϑ); we
denote it by ψ∗ (X, ς, ϑ) . A left principal biset is called a trivial left principal biset if it is the pull-back of
the unit left principal biset, that is, of the form ψ∗ (U (G )) for some morphism of groupoids ψ : K → G .
Next we expound the bicategorical constructions beyond the notion of principal groupoids-bisets. A
morphism of left principal (H , G )-bisets F : (X, ς, ϑ) → (X ′ , ς′ , ϑ′ ) is a map F : X → X ′ which is
simultaneously G -equivariant and H -equivariant, that is, the following diagrams
/X
/X
X ς × t G1
H1 s × ϑ X
X ❖❖❖ϑ
ς ♦♦
❖❖❖
♦
♦
♦
'
w♦
F
F
id × F
F × id
F
(11)
G0 g❖❖
♦7 H0
❖❖❖
♦
♦
♦
′
′
′
′
/
❖
♦
/
♦ ϑ′
H1 s × ϑ′ X
X.
X ς′ × t G 1
X
ς′
X′
commute. An isomorphism of left principal bisets is a morphism whose underlying map is bijective. As in
the geometric case we have:
Proposition 2.10. Given two groupoids G and H . Then any morphism between left principal (H , G )bisets is an isomorphism.
ON GEOMETRICALLY TRANSITIVE HOPF ALGEBROIDS
9
Proof. Let F : (X, ς, ϑ) → (X ′ , ς′ , ϑ′ ) be a morphism of left principal (H , G )-bisets. We first show that F
is injective. So take x, y ∈ X such that F(x) = F(y), whence ς(x) = ς(y). By Lemma 2.9, we know that
there exists h ∈ H1 with s(h) = ϑ(x) such that hx = y. Therefore, we have F(hx) = F(y) = hF(x) = F(x)
and so h = ιϑ(x) , since the left action is free. This shows that x = y. The surjectivity of F is derived as
follows. Take an arbitrary element x′ ∈ X ′ and consider its image ς′ (x′ ) ∈ G0 . Since ς is surjective, there
exists x ∈ X such that ς(x) = ς′ F(x) = ς′ (x′ ). This means that F(x) and x′ are in the same orbit, so there
exists h′ ∈ H1 (with s(h′ ) = ϑ(x)) such that h′ F(x) = F(h′ x) = x′ , which shows that F is surjective.
Remark 2.11. By Proposition 2.10, the category of left principal bisets PBl (H , G ) is actually a groupoid
(not necessarly a small category). On the other hand, notice that if (X, ς, ϑ) is a left principal (H , G )-biset,
then its opposite (X o , ϑ, ς) is a right principal (G , H )-biset, where the underlying set still the same set X
while the actions were switched by using the inverse maps of both groupoids. This in fact establishes an
isomorphism of categories between PBl (H , G ) and the category of right principal bisets PBr (G , H ).
Remark 2.12. Given (X, ς, ϑ) an (H , G )-biset and (X ′ , ς′ , ϑ′ ) a (G , K )-biset. One can endow the fibre
product X ς × ϑ′ X ′ within a structure of an (H , K )-biset. Furthermore, G also acts on this set by the
action (x, x′ ).g = (xg, g−1 x′ ), for g ∈ G1 , (x, x′ ) ∈ X ς × ϑ′ X ′ with t(g) = ς(x) = ϑ′ (x′ ). Denote by
X ⊗G X ′ := X ς × ϑ′ X ′ /G its orbit set, then clearly this set inherits a structure of (H , K )-biset. This is the
tensor product of bisets, also known as le produit contracté [15, Définition 1.3.1 page 114], [10, Chap.III,
§4, 3.1]. It turns out that, if (X, ς, ϑ) is a left principal biset and (X ′ , ς′ , ϑ′ ) is a left principal biset, then
X ⊗G X ′ is a left principal (H , K )-biset. Moreover, one can show that the tensor product (over different
groupoids) is associative, up to a natural isomorphism. This defines the bicategory PBl of left principal
bisets. Analougly, we have the bicategories PBr and PBb (of principal bisets).
For a single 0-cell, i.e., a groupoid G , the category PBb (G , G ) turns to be a bigroup (or a categorical
group). Moreover, in analogy with the group case, one can construct with the help of Proposition 2.10
and by using morphisms between left translation groupoids, a presheaf BG : Setsop −→ 2-Grpds to the
category of 2-groupoids known as the classifying 2-stack of the groupoid G (compare with Example 2.1).
2.4. Principal groupoids-biset versus weak equivalences. A morphism of groupoids φ : H → G is
said to be a weak equivalence if it satisfies the following two conditions:
(WE-1) The composition map G1 s × φ0 H0
(WE-2) The following diagram is cartesian
pr1
t
/ G1
/ G0 is surjective.
φ1
H1
/ G1
(s,t)
(s,t)
H0 × H0
φ0 ×φ0
Equivalently there is a bijection Γ : H1 H0
Γ = (s, t).
φ0
/ G0 × G0
× s G1 t × φ0 H0 such that pr2 ◦ Γ = φ0 and (pr1 , pr3 ) ◦
In categorical terms, condition (WE-1) says that φ is an essentially surjective functor: Each object of G
is isomorphic to the image by φ of an object in H . The second condition, means that φ is fully faithful: If
u, v are two objects in H then φ defines a bijection between the sets of arrows H (u, v) and G φ0 (u), φ0(v) .
Both properties classically characterize functors which define equivalences of categories.
Two groupoids G and H are said to be weakly equivalent when there exists a third groupoid K with a
diagram (i.e., a span) of weak equivalences:
G
K PP
PPP
♦♦♦
♦
P(
♦
w♦
H.
For sake of completeness, next we give a result which relate the notion of principal biset with that of
weak equivalence.
10
LAIACHI EL KAOUTIT
Proposition 2.13. Let G and H be two groupoids. Assume that there is (X, ς, ϑ) a principal (H , G )-biset.
Then the canonical morphisms of groupoids
H
H X X Y ❚G❚
Θ ✐✐✐
❚❚❚Σ ❚
✐
❚❚*
t✐✐✐✐
G
are weak equivalences, where Θ, Σ are as in (3) and (4). In particular, G and H are weakly equivalent.
Proof. We only show that if (X, ς, ϑ) is a left principal (H , G )-biset, then the canonical morphism
Σ : H X X Y G −→ G ,
(h, x, g), x 7−→ (g, ς(x)
is a weak equivalence. The proof of the fact that Θ is a weak equivalence follows similarly from the
assumption that (X, ς, ϑ) is right principal (H , G )-biset.
Condition (WE-1) for Σ is clear, since ς is surjective by condition (P-1) of Definition 2.8. Consider the
map
/ X ς × s G1 t × ς X
Γ : H1 s × ϑ X ς × s G 1
✤
/ x, g, hxg−1.
h, x, g
Using the map δ : X ς × ς X → H1 resulting from condition (P-2) on (X, ς, ϑ) and which satisfies equations
(6)-(8), we define the inverse of Γ to be the map:
Γ−1 : X ς × s G1 t × ς X
✤
x, g, y
/ H1 s × ϑ X ς × s G 1
/ δ(y, xg−1 ), x, g,
which gives condition (WE-2) for Σ.
Remark 2.14. As we have seen in Remark 2.11, the opposite of left principal (H , G )-biset is a right
principal (G , H )-biset. Thus the opposite of principal biset is also a principal biset. In this way, Proposition
2.13 says that the “equivalence relation” between groupoids defined by ’being connected by a principal
biset’ is contained in the equivalence relation defined by ’being weakly equivalent’. An interesting question
is then to check if both relations are the same. Precisely, one can ask whether two weakly equivalent
groupoids H and G are connected by a certain principal (H , G )-biset. The complete answer was recently
given in [11, Theorem 2.9] (see Remark 3.10 below for these equivalence relations in Hopf algebroids
context).
2.5. Transitive groupoids are characterized by weak equivalences. This subsection is the main motivation for the forthcoming sections. Here we show perhaps a well known result that characterizes transitive
groupoids by means of weak equivalences and principal groupoids-bisets.
Proposition 2.15. Let G be a groupoid. Then the following are equivalent:
(i) For every map ς : X → G0 , the induced morphism of groupoids φς : G ς → G is a weak equivalence;
(ii) G is a transitive groupoid;
∗
(iii) For every map ς : X → G0 , the pull-back biset φς (U (G )) is a principal (G , G ς )-biset.
Proof. (i) ⇒ (ii). Is immediate.
∗
(ii) ⇒ (iii). By definition φς (U (G )) is a left principal (G , G ς )-biset. We need then to check that, under
∗
condition (ii), it is also right principal (G , G ς )-biset. This biset is given by φς (U (G )) = G1 s × ς X,et, pr2 ,
where et := t ◦ pr1 : G1 s × ς X → G1 → G0 . The left and right actions are given as in equations (9) and (10)
by
g ⇀ ( f, x) = (g f, x) and ( f, x) ↼ (y, h, x) = ( f h, y),
for any ( f, x) ∈ G1 s × ς X, g ∈ G1 with s(g) = t( f ), and h ∈ G1 with s(h) = ς(y), t(h) = ς(x) and s( f ) = t(h).
Both conditions (1)-(2) in Definition 2.7 are then clearly satisfied. The right canonical map is defined by
h
i
∇′ : G1 s × ς X pr2 × t Gς 1 −→ G1 s × ς X et ×et G1 s × ς X ,
f, x , y, h, x 7−→ f, x , f h, y .
ON GEOMETRICALLY TRANSITIVE HOPF ALGEBROIDS
11
The map et is clearly surjective, since G is transitive. This gives condition (P-1) of Definition 2.8 for
∗
φς (U (G )) as a right principal biset. Now, we need to check that ∇′ is bijective, that is, condition (P-2) is
fulfilled. However, the inverse of this map is easily shown to be the following map
h
i
∇′−1 : G1 s × ς X et ×et G1 s × ς X −→ G1 s × ς X pr2 × t Gς 1 ,
f, x , f ′ , x′ 7−→ f, x , x′ , f −1 f ′ , x .
(iii) ⇒ (i). If we assume that φς (U (G )) is a principal (G , G ς )-biset, then the map et above should be
surjective. Therefore, the map
∗
G1 s × ς X
pr1
/ G1
t
/ G0
is also surjective, which is condition (WE-1) for the morphism φς . Condition (WE-2) for this morphism is
trivial, since by definition we know that Gς 1 = X ς × t G1 s × ς X.
2.6. Correspondence between transitive groupoids and principal group-sets. A particular example of
principal groupoids-bisets are, of course, principal group-sets. As we will see below transitive groupoids
are characterized by these group-sets. Precisely, there is a (non canonical) correspondence between transitive groupoids and principal group-sets, as we will show in this subsection.
Let π : P → G0 be a map and G a group which acts on the left side of P. Recall that the triple (P, G, π)
is said to be a left principal G-set, if the following conditions are satisfied:
(P’1) π is surjective;
(P’2) π(gp) = π(p), for every p ∈ P and g ∈ G;
(P’3) The canonical map G × P −→ P π × π P sending (g, p) 7→ (gp, p) is bijective.
Equivalently, the action is free and G0 is the orbit set of P. Comparing with Definition 2.7, this means that
the triple (P, ∗, π) with ∗ : P → {∗}, is a principal left (G, G0 )-biset, where the group G is considered as a
groupoid with one object {∗} and G0 is considered as a groupoid whose underlying category is a discrete
category (i.e., category with only identities arrows) with set of objects G0 , and acts trivially on P along π.
In the previous situation, consider P × P as a left G-set by the diagonal action and denote by G1 :=
(P × P)/G its set of orbits. The pair (G1 , G0 ) admits as follows a structure of transitive groupoid. Indeed,
let (p, p′ ) ∈ P × P and denote by [(p, p′ )] ∈ (P × P)/G its equivalence class. The source and target
are s [(p, p′ )] = π(p′ ) and t [(p, p′ )] = π(p). The identity arrow of an object x ∈ G0 is given, using
conditions (P’1)-(P’2), by the class [(p, p)] where π(p) = x. Let p, q be two equivalence classes in G1 such
that s(p) = t(q). Henceforth, if (p, p′ ) is a representative of p, then q can be represented by (p′ , p′′ ). The
multiplication pq is then represented by (p, p′′ ). This is a well defined multiplication since the action is
free. By conditions (P’1), (P’3), we have that (G1 , G0 ) is a transitive groupoid with a canonical morphism
of groupoids
/
/P
P×P o
G1 := (P × P)/G o
/ P/G := G0 .
/
Conversely, given a transitive groupoid G , and fix an object x ∈ G0 . Set G := G x the isotropy group of
x and let P := t−1 ({x}) be the set of all arrows with target this x, i.e., the left star set of x. Consider the left
G-action G × P → P derived from the multiplication of G . Since G is transitive, the triple (P, G, π) satisfies
then the above conditions (P’1)-(P’3), which means that it is a left principal G-set.
3. Hopf algebroids: comodules algebras, principal bundles, and weak equivalences
This section contains the definitions of commutative Hopf algebroids and theirs bicomodules algebras.
All definitions are given in the algebraic way. Nevertheless, we will use a slightly superficial language of
presheaves, sufficiently enough to make clearer the connection with the contents of Section 2.
Parallel to subsections 2.3 and 2.4, we present a brief contents on principal bibundles between Hopf
algebroids and their connection with weak equivalences. Dualizable objects in the category of (right)
comodules are treated in the last subsection, where we also proof some useful lemmata.
12
LAIACHI EL KAOUTIT
3.1. Preliminaries and basic notations. We work over a commutative base field k. Unadorned tensor
product − ⊗ − stands for the tensor product of k-vector spaces − ⊗k −. By k-algebra, or algebra, we
understand commutative k-algebras, unless otherwise specified. The category of (right) A-modules over an
algebra A, is denoted by ModA . The k-vector space of all A-linear maps between two (right) A-modules M
and N, is denoted by HomA (M, N). When N = A is the regular module, we denote M ∗ := HomA (M, A).
Given two algebras R, S , we denote by S (R) := Algk (S , R) the set of all k-algebra maps from S to R.
In what follows, a presheaf of sets (of groups, or of groupoids) stands for a functor from the category of
algebras Algk to the category of sets (groups, or groupoids). Clearly, to any algebra A, there is an associated
presheaf which sends C → A(C) = Algk (A, C), thus, the presheaf represented by A.
For two algebra maps σ : A → T and γ : B → T we denote by σ T γ (respectively, σ T or T γ , if one of the
algebra maps is the identity) the underlying (A, B)-bimodule of T (respectively, the underlying A-module
of T ) whose left A-action is induced by σ while its right B-action is induced by γ, that is,
a . t = σ(a)t,
t . b = tγ(b),
for every a ∈ A, b ∈ B, t ∈ T.
Assume there is an algebra map x ∈ A(R). The extension functor (−)x : ModA → ModR is the functor
which sends any A-module M to the extended R-module Mx = M ⊗A R. In order to distinguish between two
extension functors, we use the notation Mx := M ⊗x R and My := M ⊗y S , whenever another algebra map
y ∈ A(S ) is given.
In the sequel we will use the terminology coring (or cogébroı̈de as in [9, 4]) for coalgebra with possibly
different left-right structures on its underlying modules over the base ring. We refer to [6] for basic notions
and properties of these objects.
3.2. The 2-category of Hopf algebroids. Recall from, e.g., [26] that a commutative Hopf algebroid, or a
Hopf algebroid over a field k, is a pair (A, H) of two commutative k-algebras, together with algebra maps
η : A ⊗ A → H,
ε : H → A,
∆ : s Ht → s Ht ⊗ A s Ht ,
S : s Ht → t Hs
and a structure (s Ht , ∆, ε) of an A-coring with S an A-coring map to the opposite coring. Here the source
and the target are the algebra maps s : A → H and t : A → H defined by s(a) = η(a⊗ 1) and t(a) = η(1⊗ a),
for every a ∈ A. The map S is called the antipode of H subject to the following equalities:
S 2 = Id,
t(ε(u)) = S (u(1) )u(2),
s(ε(u)) = u(1) S (u(2) ),
for every u ∈ H,
(12)
where we used Sweedler’s notation: ∆(u) = u(1) ⊗A u(2) and summation is understood. The algebras A and
H are called, respectively, the base algebra and the total algebra of the Hopf algebroid (A, H).
As commutative Hopf algebra leads to an affine group scheme, a Hopf algebroid leads to an affine
groupoid scheme (i.e., a presheaf of groupoids). More precisely, given a Hopf algebroid (A, H) and an
algebra C, reversing the structure of (A, H) we have, in a natural way, a groupoid structure
/
s∗
(13)
H (C) : H(C) o εt∗∗
/ A(C).
This structure is explicitly given as follows: the source and the target of a given arrow g ∈ H(C) are,
respectively, s∗ (g) = g ◦ s and t∗ (g) = g ◦ t, the inverse is g−1 = g ◦ S . Given another arrow f ∈ H(C) with
t∗ ( f ) = s∗ (g), then the groupoid multiplication is defined by the following algebra map
g f : H −→ C,
u 7−→ f (u(1) )g(u(2)) ,
summation always understood. The identity arrow of an object x ∈ A(C) is ε∗ (x) = x ◦ ε.
The functor H is referred to as the associated presheaf of groupoids of the Hopf algebroid (A, H), and
the groupoids of equation (13) are called the fibres of H . Depending on the handled situation, we will
employ different notations for the fibres of H at an algebra C:
H (C) := H(C), A(C) := H1 (C), H0 (C) .
The presheaf of groupoids H op is defined to be the presheaf whose fibre at C is the opposite groupoid
H (C)op (i.e., the same groupoid with the source interchanged by the target).
Examples of Hopf algebroids can be then proportioned using well known constructions in groupoids, as
we have seen in subsection 2.1.
ON GEOMETRICALLY TRANSITIVE HOPF ALGEBROIDS
13
Example 3.1. Let A be an algebra and set H := A ⊗ A. Then the pair (A, H) is clearly a Hopf algebroid
with structure s : A → H, a 7→ a ⊗ 1, t : A → H, a 7→ 1 ⊗ a; ∆(a ⊗ a′ ) = (a ⊗ 1) ⊗A (1 ⊗ a′ ), ε(a ⊗ a′ ) = aa′ ,
S (a ⊗ a′ ) = a′ ⊗ a. Clearly, the fibres of the associated presheaf of groupoids H are groupoids of pairs,
as in Example 2.2. Thus, H A × A , A an isomorphism of presheaves of groupoids, where A is the
presheaf of sets attached to the algebra A.
Example 3.2. Let (B, ∆, ε, S) be a commutative Hopf k-algebra and A a commutative right B-comodule
algebra with coaction A → A ⊗ B, a 7→ a(0) ⊗ a(1) . This means that A is right B-comodule and the coaction is
an algebra map, see [23, §, 4]. Let B be the affine k-group attached to B. To any algebra C, one associated
in a natural way, the following action groupoid as in Example 2.3:
/
(A × B)(C) : A(C) × B(C) o
/ A(C),
where the source is given by the action (x, g) 7→ xg sending a 7→ (xg)(a) = x(a(0) )g(a(1)), and the target is the
first projection. Consider, on the other hand, the algebra H = A ⊗ B with algebra extension η : A ⊗ A → H,
a′ ⊗ a 7→ a′ a(0) ⊗ a(1) . Then (A, H) has a structure of Hopf algebroid, known as a split Hopf algebroid:
∆(a ⊗ b) = (a ⊗ b(1) ) ⊗A (1A ⊗ b(2) ), ε(a ⊗ b) = aε(b), S (a ⊗ b) = a(0) ⊗ a(1) S(b).
Obviously, the associated presheaf of groupoids H op (where H is the one associated to (A, H)) is canon
ically isomorphic to the action groupoids A × B. Thus, we have an isomorphism H op A × B, A of
presheaves of groupoids.
Example 3.3. Let A be an algebra and consider the commutative polynomial Laurent ring over A ⊗ A, that
is, H = (A ⊗ A)[X, X −1] with the canonical injection η : A ⊗ A → H. The pair (A, H) is a Hopf algebroid
with structure maps
∆ (a ⊗ a′ )X = (a ⊗A 1)X ⊗A (1 ⊗ a′ )X , ε (a ⊗ a′ )X = aa′ , S (a ⊗ a′ )X = (a′ ⊗ a)X −1.
The fibres of the associated presheaf H are described using the induced groupoid by the multiplicative
affine k-group, in the sense of Example 2.4. Precisely, take an algebra C, then H(C) is canonically bijective
to the set A(C) × Gm (C) × A(C), where Gm is the multiplicative affine k-group. This in fact induces, in a
natural way, an isomorphisms of groupoids (H(C), A(C)) A(C) × Gm (C) × A(C), A(C) , where the later
is the induced groupoid by the group Gm (C), as in Example 2.4. As presheaves of groupoids, we have then
an isomorphism H A × Gm × A , A , where as above A is the presheaf attached to the algebra A.
There is in fact a more general construction: Take any commutative Hopf k-algebra B, then for any
algebra A, the pair (A, A ⊗ B ⊗ A) admits a canonical structure of Hopf algebroid whose associated presheaf
is also of the form A × B × A , A , where A and B are as before.
Example 3.4 (Change of scalars). Let (A, H) be a Hopf algebroid over k and consider L a field extension
of k. Then the pair of algebras (AL , HL ) := (A ⊗k L, H ⊗k L) admits, in a canonical way, a structure of
Hopf algebroid over the field L. The structure maps are denoted using the subscript L, i.e., sL , tL , εL , ... If
we denote by HL : AlgL → Grpds the presheaf of groupoids associated to (AL , HL ), then the usual homtensor adjunction shows that HL factors through the forgetful functor AlgL → Algk . That is we have a
commutative diagram of functors:
AlgL
O
/ Alg
k
H
/ Grpds.
7
HL
The notion of character group in commutative Hopf algebras context is naturally extended to that of
character groupoids in commutative Hopf algebroids:
Definition 3.5. Let (A, H) be a Hopf algebroid over a field k and H its associated presheaf of groupoids.
The character groupoid of (A, H) is the fibre groupoid H (k) = (H(k), A(k)) at the base field k. Notice
that the character groupoid might be empty (i.e., could be a category without objects).
The following definition, which we will frequently used in the sequel, can be found in [9, page 129]. It
is noteworthy to mention that in our case (i.e., the case of affine k-schemes), the base presheaf H0 of the
presheaf H associated to a given Hopf algebroid (A, H) with A , 0, is always non empty. That is, the
condition H0 , ∅ in op.cit., is satisfied since H0 is represented by A and there is always a field extension L
of k such that A(L) , ∅ as A , 0 (i.e., it have maximal ideals).
14
LAIACHI EL KAOUTIT
Definition 3.6. Let (A, H) be a Hopf algebroid and H its associated presheaf of groupoids. Given an
algebra C, consider the fibre groupoid H (C). Two objects x, y ∈ A(C) are said to be locally isomorphic,
in the sense of the fpqc topology (or fpqc locally isomorphic), if there exists a faithfully flat extension
p : C → C ′ and an arrow g ∈ H(C ′ ) such that
p ◦ x = g ◦ s,
and
p ◦ y = g ◦ t.
We say that any two objects of H are fpqc locally isomorphic (without specifying the algebra C), if for
any algebra C and any two objects x, y ∈ A(C), x and y are fpqc locally isomorphic.
Remark 3.7. In case we have H (C) = ∅, for some algebra C, the condition of Definition 3.6 is conventionally assumed to be verified for this C. On the other hand, it is not difficult to check that if there exists
a field extension L of k such that two objects of HL are fpqc locally isomorphic, then any two objects of
H are also fpqc locally isomorphic. The converse is not immediate and follows from Theorem 4.8 below.
More precisely, if H0 (k) , ∅ and any two objects of H are fpqc locally isomorphic, then any two objects
of HL are also fpqc locally isomorphic for any field extension L of k.
If the presheaf H is fibrewise transitive, that is, each of its fibres H (C) is a transitive groupoid, then
obviously any two objects of H are fpqc locally isomorphic. For instance, this is the case for the Hopf
algebroid (A, A ⊗ A), since in this case each of the groupoid’s H (C) is the groupoid of pairs, see Example
3.1. The same holds true for the class of Hopf algebroids described in Example 3.3.
A morphism φ : (A, H) → (B, K) of Hopf algebroids consists of a pair φ = (φ0 , φ1 ) of algebra maps
φ0 : A → B and φ1 : H → K that are compatible, in a canonical way, with the structure maps of both H
and K. That is, the equalities
φ1 ◦ s = s ◦ φ0 ,
∆ ◦ φ1 = χ ◦ (φ1 ⊗A φ1 ) ◦ ∆,
φ1 ◦ t = t ◦ φ0 ,
ε ◦ φ1 = φ0 ◦ ε,
S ◦ φ1 = φ1 ◦ S ,
(14)
(15)
(16)
hold, where χ is the obvious map χ : K ⊗A K → K ⊗B K, and where no distinction between the structure
maps of H, K was made. Clearly, any morphism φ : (A, H) → (B, K) of Hopf algebroids induces (in the
opposite way) a morphism between the associated presheaves of groupoids, which is given over each fibre
by
(φ0 ∗ , φ1 ∗ ) : K (C) = (K(C), B(C)) −→ H (C) = (H(C), A(C)), sending (g, x) 7→ (g ◦ φ1 , x ◦ φ0 ).
In this way, the construction in the following example corresponds to the construction of the induced
groupoid as expounded in Example 2.4.
Example 3.8 (Base change). Given a Hopf algebroid (A, H) and an algebra map φ : A → B, then the pair
of algebras
(B, Hφ) := (B, B ⊗A H ⊗A B)
is a Hopf algebroid known as the base change Hopf algebroid of (A, H), and (φ, φ1 ) : (A, H) → (B, Hφ)
is a morphism of Hopf algebroids, where φ1 : H → Hφ sends u 7→ 1B ⊗A u ⊗A 1B . Moreover, as in the
case of groupoids, see subsection 2.1, any morphism φ : (A, H) → (B, K) factors through the base change
morphism (A, H) → (B, Hφ0 ), by using the map Hφ0 → K sending b ⊗A u ⊗A b′ 7→ bb′ φ1 (u).
The associated presheaf of groupoids Hφ of the Hopf algebroid (B, Hφ) is fiberwise computed as the
induced groupoid (see Example 2.4) of H along the map φ : B → A where B and A are the presheveas
of sets associated to B and A, respectively.
The aforementioned relation with the induced groupoids comes out as follows. Take an algebra C and
consider the associated groupoid H (C). Then the map φ∗ : B(C) → A(C) leads, as in Example 2.4, to the
∗
induced groupoid H (C)φ(C) . This in fact determines a presheaf of groupoids C → H (C)φ which can be
easily shown to be represented by the pair of algebras (B, Hφ).
We finish this subsection by recalling the construction of the 2-category of flat Hopf algebroids.
A Hopf algebroid (A, H) is said to be flat, when s H (or Ht ) is a flat A-module. Notice, that in this case
s as well as t are faithfully flat extensions. As was mentioned before, groupoids, functors, and natural
transformations form a 2-category. Analogously (flat) Hopf algebroids over the ground field k form a
ON GEOMETRICALLY TRANSITIVE HOPF ALGEBROIDS
15
2-category, as was observed in [24, §3.1]. Precisely, 0-cells are Hopf algebroids, or even flat ones, 1cells are morphisms of Hopf algebroids, and for two 1-cells (φ0 , φ1 ), (ψ0 , ψ1 ) : (A, H) → (B, K), a 2-cell
c : (φ0 , φ1 ) → (ψ0 , ψ1 ) is defined to be an algebra map c : H → B that makes the diagrams
HO
s
A
c
/B
④=
④
④
④④
④④ φ0
HO
t
A
c
/B
④=
④
④
④④
④④ ψ0
∆
H
∆
/ H ⊗A H
mK ◦ φ1 ⊗A (t◦c)
H ⊗A H
mK ◦ (s◦c) ⊗A ψ1
/K
(17)
commutative, where mK denotes the multiplication of K. The identity 2-cell for (φ0 , φ1 ) is given by 1φ :=
φ0 ◦ ε. The tensor product (or vertical composition) of 2-cells is given as
c′ ◦ c : (φ0 , φ1 )
c
/ (ψ0 , ψ1 )
c′
/ (ξ0 , ξ1 ),
where
c′ ◦ c : H → B,
u 7→ c(u(1) )c′ (u(2) ).
(18)
3.3. Comodules, bicomodules (algebras), and presheaves of orbit sets. A right H-comodule is a pair
(M, ̺) consisting of right A-module M and right A-linear map (referred to as the coaction) ̺ : M → M ⊗A
s H, m 7→ m(0) ⊗ A m(1) (summation understood) satisfying the usual counitary and coassociativity properties.
Morphisms between right H-comodules (or right H-colinear map) are A-linear maps compatible with both
coactions. The category of all right H-comodules is denoted by ComodH . This is a symmetric monoidal
k-linear category with identity object A endowed with the coaction t : A → Ht A ⊗A s Ht .
The tensor product in ComodH is defined via the so called the diagonal coaction. Precisely, given (M, ̺)
and (N, ̺) two (right) H-comodules. Then the tensor product M ⊗A N is endowed with the following (right)
H-coaction:
(19)
̺M⊗A N : M ⊗A N −→ (M ⊗A N) ⊗A H,
m ⊗A n 7−→ (m(0) ⊗A n(0) ) ⊗A m(1) n(1) .
The vector space of all H-colinear maps between two comodules (M, ̺) and (N, ̺) will be denoted by
HomH (M, N), and the endomorphism ring by EndH (M).
Inductive limit and cokernels do exist in ComodH , and can be computed in A-modules. Furthermore, it
is well known that the underlying module s H is flat if and only if ComodH is a Grothendieck category and
the forgetful functor UH : ComodH → ModA is exact. As it can be easily checked, the forgetful functor
UH has a right adjoint functor − ⊗A s H : ModA → ComodH .
The full subcategory of right H-comodules whose underlying A-modules are finitely generated is denoted by comodH . The category of left H-comodules is analogously defined, and it is isomorphic via the
antipode to the category of right H-comodules.
A (right) H-comodule algebra can be defined as a commutative monoid in the symmetric monoidal
category ComodH . This is a commutative algebra extension σ : A → R where the associated A-module Rσ
is also a (right) H-comodule whose coaction ̺R : Rσ → Rσ ⊗A s H is an algebra map, which means that
̺R (1R ) = 1R ⊗A 1H ,
′
′
̺R (rr′ ) = r(0) r(0)
⊗A r(1) r(1)
, for every r, r′ ∈ R.
This of course induces a (right) H -action on the presheaf of sets R associated to R. Precisely, given an
algebra C, consider the map σ∗ : R(C) → A(C) sending x 7→ x ◦ σ, and set
R(C) σ∗ × s∗ H1 (C) → R(C), (x, g) 7→ xg, where xg : R → C, r 7→ x(r(0) ) g(r(1) ).
(20)
Then this defines, in a natural way, a right action of the groupoid H (C) on the set R(C), in the sense of
Definition 2.6. Equivalently such an action can be expressed as a pair morphism of presheaves R → H0
and R σ∗ × s∗ H1 → R satisfying pertinent compatibilities. In this way, an action of a presheaf of groupoids
on a presheaf of sets, can be seen as a natural generalization to the groupoids framework of the notion of
an action of group scheme on a scheme [10, no 3, page 160], or more formally, as a generalization of the
notion of ”objet à groupe d’opérateurs à droite” [15, Chapitre III, §1.1].
For a right H-comodule algebra (R, σ), we have the subalgebra of coinvariants defined by
o
n
(21)
RcoinvH := r ∈ R| ̺R (r) = r ⊗A 1H .
16
LAIACHI EL KAOUTIT
Denote by R H the presheaf of sets represented by the algebra RcoinvH . On the other hand, we have the
presheaf defining the orbit sets, which is given as follows: Take an algebra C and consider the action (20),
we then obtain the orbits set R(C)/H (C). Clearly this establishes a functor: C −→ R(C)/H (C) yielding
a presheaf OH (R) with a canonical morphism of presheaves OH (R) −→ R H .
Remark 3.9. An important example of the previous construction is the case of the right H-comodule
algebra (A, t). In this case we have a commutative diagram of presheaves:
A ❘❘❘
❘❘❘
)
ζ
τ
OH (A )
/ AH
❦❦5
❦❦❦❦
where as before A is the presheaf represented by the algebra A and A H is represented by AcoinvH .
Notice that the presheaf OH (A ) is not necessarily represented by AcoinvH , thus, the right hand map in the
previous diagram is not in general an isomorphism of presheaves, see [25, page 54].
In this direction, both A τ × τ A and A ζ × ζ A enjoy a structure of presheaves of groupoids with fibres
are groupoids of pairs described in Examples 2.2. Nevertheless, A ζ × ζ A is not necessarily representable.
Furthermore, there is a commutative diagram
/ A τ× τ A
H ❘❘❘
❘❘❘
✐✐4
)
✐✐✐✐
A ζ× ζ A
of presheaves of groupoids.
Given two Hopf algebroids (A, H) and (B, K), the category of (H, K)-bicomodules is defined as follows. An object in this category is a triple (M, λ, ̺) consisting of left H-comodule (M, λ) and right Kcomodule (M, ̺) such that λ is a morphism of right K-comodules, or equivalently ̺ is a morphism of left
H-comodules. Morphisms between bicomodules are simultaneously left and right comodules morphisms.
On the other hand, the pair of tensor product (A⊗ B, H o ⊗ K) admits, in a canonical way, a structure of Hopf
algebroid, where (A, H o ) is the opposite Hopf algebroid (i.e., the source and the target are interchanged, or
equivalently, the fibres of the associated presheaf are the opposite groupoids H (C)op ). This is the tensor
product Hopf algebroid, and its category of right comodules is canonically identified with the category of
(H, K)-bicomodules. Thus bicomodules form also a symmetric monoidal k-linear category.
A bicomodule algebra is a bicomodule which is simultaneously a left comodule algebra and right comodule algebra. As above, by using the actions of equation (20) a bicomodule algebra leads to a presheaf
of groupoid bisets. That is, a presheaf with fibres groupoid-bisets, in the sense of Definition 2.7.
3.4. Weak equivalences and principal bundles between Hopf algebroids. Any morphism φ : (A, H) →
(B, K) of Hopf algebroids induces a symmetric monoidal k-linear functor
φ∗ := UH (−) ⊗A B : ComodH −→ ComodK ,
where, for any H-comodule (M, ̺), the K-comodule structure of M ⊗A B is given by
M ⊗A B −→ (M ⊗A B) ⊗B K,
m ⊗A b 7−→ (m(0) ⊗A 1B ) ⊗B φ1 (m(1) )t(b).
Following [17, Definition 6.1], φ is said to be a weak equivalence whenever φ∗ is an equivalence of categories. In this case, ComodH and ComodK are equivalent as symmetric monoidal k-linear categories.
Notice that if φ is a weak equivalence, then so is the associated morphism between the tensor product
Hopf algebroids φo ⊗ φ : (A ⊗ A, H o ⊗ H) → (B ⊗ B, K o ⊗ K), which induces then a symmetric monoidal
equivalence between the categories of H-bicomodules and K-bicomodules.
Two Hopf algebroids (A, H) and (B, K) are said to be weakly equivalent if there exists a diagram
(A, H)
(C, J) i❙
❙❙❙❙
❦❦5
❦
❦
❙
❦
❦
(B, K),
of weak equivalences.
As was shown in [11] weak equivalences between flat Hopf algebroids are strongly related to principal
bi-bundles. Such a relation is in part a consequence of the analogue one for groupoids as was shown in
Proposition 2.13 (see Remark 2.14).
ON GEOMETRICALLY TRANSITIVE HOPF ALGEBROIDS
17
Recall that, for two flat Hopf algebroids (A, H) and (B, K), a left principal (H, K)-bundle is a threetuple (P, α, β) which consists of diagram of commutative algebras α : A → P ← B : β where the (A, B)bimodule α Pβ enjoys a structure of an (H, K)-bicomodule algebra such that
(PB1) β : B → P is a faithfully flat extension (the local triviality of the bundle in the fpqc topology);
(PB2) the canonical map canP, H : P ⊗B P → H ⊗A P sending p ⊗B p′ 7→ p(0) ⊗A p(1) p′ is bijective.
Observe that these two conditions, in conjunction with the faithfully flat descent theorem [13, Theorem
5.9], show that PcoinvH = B, see (21) for the notation.
The notion of principal bundles is a natural generalization of the notion of Torsor, where the group
object is replaced by groupoid object, see [15, Définition 1.4.1, page 117] and [10, Chapter III, §4]. In case
of Hopf algebras over commutative rings, these objects are termed Hopf Galois extensions, see [23, 28].
Right principal bundles and bi-bundles are clearly understood. For instance to each left principal bundle
(P, α, β), one can define a right principal bundle on the opposite bicomodule Pco . As in the case of groupoids,
see subsection 2.3, a simpler example of left principal bundle is the unit bundle U (H) which is H with
its canonical structure of H-bicomodule algebra. A trivial bundle attached to a given morphism of Hopf
algebroids φ : (A, H) → (B, K), is the one whose underlying bicomodule algebra is of the form P :=
H ⊗A B, that is, the pull-back bundle φ∗ (U (H)) of the unit bundle U (H).
Parallel to subsection 2.3, for any bicomodule algebra, and thus for any left principal bundle, one can
associate the so called the two-sided translation Hopf algebroid, which is denoted by (P, H X P Y K).
The underlying pair of algebras is (P, Hs ⊗ α Pβ ⊗ s K) and its structure of Hopf algebroid is given as follows:
• the source and target are given by
s(p) := 1H ⊗A p ⊗B 1K ,
t(p) := S (p(−1) ) ⊗A p(0) ⊗B p(1) ;
• the comultiplication and counit are given by:
∆(u ⊗A p ⊗B w) := u(1) ⊗A p ⊗B w(1) ⊗P u(2) ⊗A 1P ⊗B w(2) , ε(u ⊗A p ⊗B w) := α ε(u) pβ ε(w) ;
• whereas the antipode is defined as:
S u ⊗A p ⊗B w := S (up(−1) ) ⊗A p(0) ⊗B p(1) S (w).
Furthermore, there is a diagram
(P, H X P Y K)
j❯❯❯β=(β,
❯❯❯❯β1 )
❯
(22)
α=(α, α1✐)✐✐4
(A, H)
✐
✐✐✐✐
(B, K)
of Hopf algebroids, where α1 and β1 are, respectively, the maps u 7→ u ⊗A 1P ⊗B 1K and w 7→ 1H ⊗A 1P ⊗B w.
It is easily checked that (P, H X P Y K) is a flat Hopf algebroid whenever (A, H) and (B, K) they are so.
Remark 3.10. It is noteworthy to mention that the fibres of the presheaf associated to a left principal bundle
are not necessarily principal bisets over the fibres groupoids, in the sense of Definition 2.8 (but possibly the
entry presheaf is locally so in the fpqc topology sense). To be precise, let P denote as before the presheaf
of sets associated to the algebra P. This is a presheaf of (H , K )-bisets, that is, using left and right actions
of equation (20), for any algebra C, we have that the fibre P(C) is actually an (H (C), K (C))-biset as in
Definition 2.7. However, P(C) is not necessarily a left principal biset. Nevertheless, it is easily seen that
the associated presheaf of two-sided translation groupoids is represented by the two-sided translation Hopf
algebroid (P, H X P Y K). Lastly, as in [11], two weakly equivalent flat Hopf algebroids are shown to be
connected by a principal bibundle, for which the diagram (22) becomes a diagram of weak equivalences,
see op. cit. for more characterizations of weak equivalences.
3.5. Dualizable right comodules. Recall that a (right) H-comodule (M, ̺M ) is said to be dualizable, if
there is another (right) H-comodule (N, ̺N ) and two morphisms of comodules
ev : (N ⊗A M, ̺N⊗A M ) → (A, t), and db : (A, t) → (M ⊗A N, ̺M⊗A N )
(23)
satisfying, up to natural isomorphisms, the usual triangle properties. Taking the underlying A-linear maps
(ev, db) and using these triangle properties, one shows that UH (N) M ∗ = HomA (M, A). Thus, the
underlying A-module of any dualizable comodule is finitely generated and projective. Moreover, the dual
18
LAIACHI EL KAOUTIT
of (M, ̺M ) in the category ComodH is, up to an isomorphism, the comodule (M ∗ , ̺M∗ ) with the following
coaction:
ϕ 7−→ e∗i ⊗A t(ϕ(ei, (0)))S (ei, (1) ) ,
(24)
̺M∗ : M ∗ → M ∗ ⊗A H,
P
where {ei , e∗i } is a dual basis for MA , that is, db(1A ) = i ei ⊗A e∗i . The converse holds true as well, that is,
dualizable objects in ComodH are, up to natural isomorphisms, precisely the objects of the subcategory
comodH which are projective as A-modules. This is a well known fact which will be implicitly used below:
Lemma 3.11. Let (M, ̺M ) be a right H-comodule whose underlying A-module M is finitely generated and
projective, and consider (M ∗ , ̺M∗ ) as right H-comodule with the coaction given by (24). Then (M, ̺M )
is a dualizable object in ComodH with dual object (M ∗ , ̺M∗ ). In particular, the full subcategory of dualizable right H-comodules consists of those comodules with finitely generated and projective underlying
A-modules.
The following lemma will be used in the sequel and the k-algebras involved in it are not necessary
commutative.
Lemma 3.12. Let (A, H) and (A′ , H′ ) be two corings. Assume that there are two bimodules B MA and B′ MA′ ′
such that (M, ̺M ) and (M,′ ̺M′ ) are, respectively, (B, H)-bicomodule and (B′, H′ )-bicomodule (here B and
B′ are considered as B-coring and B′ -coring in a trivial way), and that MA , MA′ ′ are finitely generated and
projective modules with dual bases, respectively, {mi , m∗i } and {n j , n∗j }.
(1) If the associated canonical map:
(25)
canM : M ∗ ⊗B M → H,
m∗ ⊗A m 7−→ m∗ (m(0) )m(1)
∗
is injective, then EndM ⊗B M (M) = EndH (M), where M ∗ ⊗B M is the standard A-coring [4], or the comatrix
A-coring [12].
(2) If both canM and canM′ are injective, B M and B′ M ′ are faithfully flat modules, then
′
′
EndH⊗H (M ⊗ M ′ ) B ⊗ B′ EndH (M) ⊗ EndH (M ′ ).
Proof. (1) is a routine computation. Part (2) uses part (1) and the result [12, Theorem 3.10].
3.6. Dualizable comodule whose endomorphism ring is a principal bundle. This subsection is of independent interest. We give conditions under which the endomorphism ring (of linear maps) of a dualizable
right comodule is a left principal bundle.
Let (A, H) be a flat Hopf algebroid and (M, ̺) a dualizable right H-comodule. Denote by B := EndH (M)
its endomorphism ring of H-colinear maps, and consider the endomorphism ring of A-linear maps EndA (M)
as right H-comodule via the isomorphism M ∗ ⊗A M EndA (M) together with the following obvious algebra
maps α : A −→ EndA (M) and β : B ֒→ EndA (M). The proof of the following is left to the reader.
Proposition 3.13. Assume that A M, B M are faithfully flat modules and that the canonical map can M of
equation (25) is bijective (e.g., when (M, ̺) is a small generator in the category of right H-comodules).
Then the triple (EndA (M), α, β) is a right principal (B, H)-bundle (where (B, B) is considered as a trivial
Hopf algebroid).
4. Geometrically transitive Hopf algebroid: Definition, basic properties and the result
In this section we recall the definition of geometrically transitive Hopf algebroids and prove some of
their basic properties. Most of the results presented here are in fact consequences of those stated in [4]. For
sake of completeness, we give below slightly different elementary proofs of some of these results.
4.1. Definition and basic properties. We start by proving the following result which will help us to well
understand the forthcoming definition.
Proposition 4.1 ([4, Proposition 6.2, page 5845]). Let (A, H) be a flat Hopf algebroid. Assume that (A, H)
satisfies the following condition
(GT1) H is projective as an (A ⊗ A)-module.
Then, we have
(GT11) Every H-comodule is projective as an A-module;
ON GEOMETRICALLY TRANSITIVE HOPF ALGEBROIDS
19
(GT12) comodH is an abelian category and the functor UH : comodH → ModA is faithful and exact;
(GT13) Every object in ComodH is a filtrated limit of subobjects in comodH .
Proof. (GT 11). Let M be a right H-comodule. Then, as a right A-module M is a direct summand of
M ⊗A H. Since A HA is a direct summand of a free (A ⊗ A)-module, M is a direct summand of the right
A-module M ⊗A (A ⊗ A) M ⊗ A. Thus MA is projective. The same proof works for left H-comodules.
(GT 12). The category comodH is additive with finite product and cokernels. Let us check that comodH
do have kernels. So, assume a morphism f : N → M in comodH is given. Then the kernel Ker( f ) is a right
H-comodule, since we already know that ComodH is a Grothendieck category. Thus we need to check
that the underlying module of this kernel is a finitely generated A-module. However, this follows from the
fact that f k : Ker( f ) → N splits in A-modules, as we know, by the isomorphism of right H-comodules
N/Ker( f ) Im( f ) and condition (GT 11), that this quotient is projective as an A-modules. The last claim
in (GT 12) is now clear.
(GT 13). Following [6, §20.1, §20.2], since s H is by condition (GT 11) a projective module, we have
that the category of rational left ∗ H-modules is isomorphic to the category of right H-comodules, where
∗
H = HomA (s H, A) is the left convolution A-algebra of H. Since any submodule of a rational module is
also rational, every rational module is then a filtrated limit of finitely generated submodules. Therefore,
any right H-comodule is a filtrated limit of subcomodules in comodH , as any finitely generated rational
module is finitely generated as an A-module.
Recall that a (locally small) k-linear category C is said to be locally of finite type, if any object in C is of
finite length and each of the k-vector spaces of morphisms C(c, c′ ) is finite dimensional.
Definition 4.2. [Bruguières] Let (A, H) be a flat Hopf algebroid. We say that (A, H) is a geometrically
transitive Hopf algebroid (GT for short) if it satisfies the following conditions:
(GT1) H is projective as an (A ⊗ A)-module.
(GT2) The category comodH is locally of finite type.
(GT3) EndH (A) k.
Here EndH (A) denotes the endomorphisms ring of the right H-comodule (A, t) which is identified with the
coinvariant subring AcoinvH = {a ∈ A| t(a) = s(a)}. We are implicitly assuming that A , 0 as a comodule.
The subsequent lemma gives others consequences of the properties stated in Definition 4.2, which will
be used later on.
Lemma 4.3. Let (A, H) be a flat Hopf algebroid.
(a) If (A, H) satisfies (GT 11) and (GT 3), then A is a simple (right) H-comodule.
(b) If (A, H) satisfies (GT 1) and (GT 3), then every (right) H-comodule is faithfully flat as an A-module.
Proof. (a). First let us check that, under the assumption (G11), any subcomodule of (A, t) is a direct
summand in ComodH . So let (I, ̺I ) be an H-subcomodule of (A, t). Then A/I is an H-comodule which is
finitely generated and projective as an A-module, by assumption (GT 11). Therefore, I is a direct summand
of A as an A-submodule. Denotes by π : A → I the canonical projection of A-modules, and let e2 = e be
an idempotent element in A such that I = eA and π(a) = ea, for every a ∈ A.
Next we show that π is a morphism of right H-comodules, which proves that (I, ̺I ) is a direct summand
of (A, t). To this end, it suffices to check that s(e) = t(e), since we know that EndH (A) = AcoinvH = {a ∈
A| t(a) = s(a)}. Clearly the coaction of I is entirely defined by the image of e, and we can write ̺I (e) =
e ⊗A u, for some element u ∈ H, which satisfies the following equalities
s(e) u = t(e)1H ,
e ⊗A u ⊗A u = e ⊗A u(1) ⊗A u(2) ∈ I ⊗A H ⊗A H,
(26)
where the first equation comes from the fact that the inclusion I ֒→ A is a morphism of H-comodules.
On the other hand, we know by Lemma 3.11 that I is a dualizable right H-comodule. Up to canonical
isomorphism, its dual comodule have for the underlying A-module, the module I ∗ = eA with coaction ̺I∗ :
eA → eA ⊗A H, sending ea 7→ e ⊗A t(ea)S (u) given by equation (24). The evaluation map ev : I ∗ ⊗A I → A,
ea ⊗A ea′ 7→ eaa′ of equation (23), is then a morphism of right H-comodules. Therefore, we have the
following equality
1 ⊗A t(e) = e ⊗A t(e)S (u)u ∈ A ⊗A s H s H.
(27)
20
LAIACHI EL KAOUTIT
Combining the first equality of equation (26) and equation (27), we get t(e)S (u) = t(e). Hence s(e)u =
s(e), and so s(e) = t(e), by the first equality in (26).
We have then show that any H-subcomodule of the H-comodule A is a direct summand, since by (GT 3)
the endomorphism ring is a field EndH (A) k, we conclude that A is a simple H-comodule.
(b). By Proposition 4.1, we know that (A, H) satisfies conditions (GT 11)-(GT 13). Let us first show
that any comodule in comodH is faithfully flat as an A-module. By condition (GT 11), we know that any
comodule in this subcategory is finitely generated and projective as A-module, so it is flat as an A-module.
Moreover, we know from Lemma 3.11 that the subcategory comodH consists exactly of dualizable right
H-comodules. Let us then pick a dualizable comodule M ∈ comodH , and assume that M ⊗A X = 0 for
some A-module X. This in particular implies that evM ⊗A X = 0, from which we get that A ⊗A X X = 0,
as evM is surjective, since we already know by item (a) that A is a simple comodule. This shows that every
object in comodH is faithfully flat as an A-module.
For an arbitrary comodule, we know by condition (GT 13) stated in Proposition 4.1, that any right
H-comodule is a filtrated limit of subcomodules in comodH . Therefore, any right H-comodule is a flat
A-module. Given now a right H-comodule M and assume that M ⊗A X = 0, for some A-module X. We have
that M = lim(Mi ) where {τi j : Mi ֒→ M j }i≤ j ∈ Λ is a filtrated system in comodH with structural morphisms
−−→
τi j which are split morphisms of A-modules. This limit is also a filtrated limit of A-modules, and so the
equality lim(Mi ⊗A X) M ⊗A X = 0 implies that there exists some j ∈ Λ, such that M j ⊗A X = 0. Hence
−−→
X = 0, since M j is a faithfully flat A-module by the previous argumentation.
Lemma 4.4. Let (A, H) be a flat Hopf algebroid which satisfies conditions (GT 11) and (GT 3). Then the
k-algebra map η : A ⊗ A → H is injective. In particular, if (A, H) is geometrically transitive, then η is
injective.
Proof. We know from Lemma 4.3(a) that A is a simple H-comodule. Therefore, by [5, Theorem 3.1], the
following map
HomH (A, H) ⊗ A −→ H,
f ⊗k a 7−→ f (a)
is a monomorphism, which is, up to the isomorphism HomH (A, H) A derived from the adjunction
between the forgetful functor UH and the functor − ⊗A s H, is exactly the map η. Hence η is injective. The
particular case is immediately obtained form Definition 4.2 and Proposition 4.1.
Remark 4.5 (Transitive Hopf algebroids). Recall from [4, Définitions pages 5838, 5850] that a Hopf algebroid (A, H) with A , 0, is said to be transitive if it satisfies conditions (GT 12), (GT 13) and (GT 2), (GT 3)
from Proposition 4.1 and Definition 4.2, respectively, and every comodule in comodH is projective. Thus
the geometrically transitive property implies the transitive one. The converse holds true if the the center of
the division ring of any simple comodule (left or right one) is a separable field extension of k, that is, if
(A, H) is a separable Hopf algebroid over k, as introduced in [4, Définition page 5847]. Obviously, over a
perfect field k both notions coincide. It is noteworthy to mention that if we consider the associated presheaf
H of a transitive Hopf algebroid (A, H), it is not clear, at least to us, how to express the transitivity of
(A, H) in terms of certain topology at the level of H . Lastly, let us mention that in general a Hopf algebroid (A, H) is a geometrically transitive if and only if (AL , HL ) is transitive, for any filed extension L of k
(see [4, Proposition 7.3 page 5851]), perhaps this justifies the terminology “geometrically transitive”.
We finish this section by characterizing dualizable objects over GT Hopf algebroids and by making
some useful remarks on these algebroids.
Proposition 4.6. Let (A, H) be a flat Hopf algebroid. Assume that (A, H) satisfies the following condition:
(GT 11)′ Every finitely generated right H-comodule is projective.
Then the full subcategory of ComodH of dualizable objects coincides with comodH . In particular, if (A, H)
is geometrically transitive, then the category comodH consists of all dualizable right H-comodules.
Proof. By Lemma 3.11, every dualizable right H-comodule is finitely generated and projective as an Amodule. This gives the direct inclusion. Conversely, any object in comodH is, by condition (GT 11)′ and
Lemma 3.11, a dualizable right H-comodule, form which we obtain the other inclusion. The particular
case of GT Hopf algebroids follows directly from Proposition 4.1.
ON GEOMETRICALLY TRANSITIVE HOPF ALGEBROIDS
21
Remark 4.7. Let (A, H) be a GT Hopf algebroid. Then, by condition (GT13) of Proposition 4.1 and
Proposition 4.6, the category of comodules ComodH has a set of small projective generators, which we
denote by A. Therefore, by applying [13, Theorem 5.7], we obtain that the canonical map can : L (O) →
H is an isomorphism of Hopf algebroids, where O : A → proj(A) is the forgetful functor to the category
of finitely generated and projective A-modules, and where L (O) is the Hopf algebroid reconstructed from
the pair (A, O), see [4] and also [13] for the explicit description of the underlying A-coring of L (O).
4.2. Characterization by means of weak equivalences. This subsection contains our main result. We
give several new characterizations of geometrically transitive flat Hopf algebroids. The most striking one
is the characterization of these Hopf algebroids by means of weak equivalences, which can be seen as the
geometric counterpart of the characterization of transitive groupoids, as we have shown in subsection 2.5,
precisely in Proposition 2.15.
Theorem 4.8. Let (A, H) be a flat Hopf algebroid over a field k and denote by H its associated presheaf
of groupoids. Assume that H0 (k) , ∅. Then the following are equivalent:
(i) η : A ⊗ A → H is a faithfully flat extension;
(ii) Any two objects of H are fpqc locally isomorphic (see Definition 3.6);
(iii) For any extension φ : A → B, the extension α : A → Ht ⊗A φ B, a 7→ s(a) ⊗A 1B is faithfully flat;
(iv) (A, H) is geometrically transitive (Definition 4.2).
By [11, Proposition 5.1], condition (iii) in Theorem 4.8 is also equivalent to the following ones:
(v) For any extension φ : A → B, the associated canonical morphism of φ : (A, H) → (B, Hφ) is a
weak equivalence;
(vi) The trivial principal left (H, Hφ)-bundle H ⊗A B is a principal bi-bundle.
Given a GT Hopf algebroid (A, H) and an extension φ : A → B. Since Hφ is a flat Hopf algebroid, the
forgetful functor ComodHφ → ModB is exact. Therefore, condition (v) implies that B is Landweber exact
over A, in the sense that the functor UH (−) ⊗A B : ComodH → ModB is exact, see [17, Definition 2.1].
Example 4.9. The following Hopf algebroids (A, A ⊗ A) and (A, (A ⊗ A)[X, X −1]) described, respectively,
in Examples 3.1 and 3.3, are clearly geometrically transitive. This is also the case of (A, A ⊗ B ⊗ A) for any
Hopf algebra B. On the other hand, if A is a right B-comodule algebra whose canonical map A ⊗ A → A ⊗ B
is a faithfully flat extension, then the split Hopf algebroid (A, A ⊗ B) is obviously geometrically transitive.
A more elaborade example of GT Hopf algebroid, by using principal bundles over Hopf algebras (i.e.,
Hopf Galois extensions), is given in Proposition 5.11 below.
Next, we give the proof of Theorem 4.8.
The proof of (i) ⇒ (ii). Let C be an algebra and x, y two objects in A(C). Denote by x ⊗ y : A ⊗ A → C
the associated algebra map and consider the obvious algebra map p : C → C ′ := H ⊗A⊗ A C. By assumption
it is clear that p is a faithfully flat extension. Set the algebra map g : H → C ′ which sends u 7→ u ⊗A⊗ A 1C .
We then have that p ◦ x = g ◦ s and p ◦ y = g ◦ t, which shows that x and y are locally isomorphic.
The proof of (ii) ⇒ (iii). We claim that under hypothesis (ii) the underlying A-module of any (left or
right) H-comodule is faithfully flat. In particular, this implies that the comodule H ⊗A B, with coaction
∆⊗A B, is faithfully flat for every A-algebra B, and this gives us condition (iii). Since there is an isomorphism
of categories between right H-comodules and left H-comodules, which commutes with forgetful functors,
it suffices then to show the above claim for right H-comodules.
So let us fix a right H-comodule M and take two objects in different fibres groupoids x ∈ A(T ) and
y ∈ A(S ), where T, S are algebras. We claim that M ⊗A x T is faithfully flat T -module if and only if M ⊗A y S
is faithfully flat S -module. Clearly our first claim follows from this one since we know that A(k) , ∅ and
over a field any module is faithfully flat.
Let us then check this second claim; we first assume that R = T = S . In this case, we know that any pair
of objects x, y ∈ A(R) are fpqc locally isomorphic, thus there exists a faithfully flat extension p : R → R′
and g ∈ H(R) such that e
x := p ◦ x = g ◦ s and e
y := p ◦ y = g ◦ t. On the other hand, the map
M ⊗A ex R′ −→ M ⊗A ey R′ ,
m ⊗A r′ 7−→ m(0) ⊗A g−1 (m(1) )r′
is clearly an isomorphism of R′ -modules. Therefore, M ⊗A ex R′ is a faithfully flat R′ -module if and only
if M ⊗A ey R′ it is. However, we know that M ⊗A ex R′ (M ⊗A x R) ⊗R p R′ is faithfully flat R′ -module if and
22
LAIACHI EL KAOUTIT
only if M ⊗A x R is faithfully flat R-module, as p is a faithfully flat extension. The same then holds true
interchanging x by y. Therefore, M ⊗A x R is faithfully flat R-module if and only if M ⊗A y R so is.
For the general case, that is, when T , S with x ∈ A(T ) and y ∈ A(S ), we take R := T ⊗ S and
consider the canonical faithfully flat extensions T → R ← S . This leads to the following two objects
x : A → T → R and y : A → S → R. Since M ⊗A x T (resp. M ⊗A y S ) is faithfully flat T -module (resp.
S -module) if and only if M ⊗A x R (resp. M ⊗A y R) is faithfully flat R-module, we have, by the proof of the
previous case, that M ⊗A x T is faithfully flat T -module if and only if M ⊗A y S is faithfully flat S -module,
and this finishes the proof of this implication.
The proof of (iii) ⇒ (iv). Take an object x in A(k) and denote by kx the base field endowed with its
A-algebra structure via the algebra map x : A → k. By assumption A → H ⊗A kx is a faithfully flat
extension. Therefore, by [11, Proposition 5.1], we know that the associated base change morphism x :
(A, H) → (kx , Hx ), where (kx , Hx ) is the Hopf k-algebra Hx = kx ⊗A H ⊗A kx , is actually a weak equivalence.
This means that the induced functor x∗ := UH (−) ⊗A kx : ComodH → ComodHx is a symmetric monoidal
equivalence of categories, and thus transforms, up to natural isomorphisms, dualizable H-comodules into
dualizable Hx -comodules. Similar property hods true for its inverse functor. In particular, taking an object
M ∈ comodH , it is clear that x∗ (M) = M ⊗A kx is finite dimensional k-vector space and so a dualizable
right Hx -comodue, see for instance Lemma 3.11. Therefore, M should be a dualizable right H-comodule.
The converse is obvious and then the full subcategory comodH coincides with the full subcategory of
dualizable right H-comodules, form which we have that comodH and comodHx are equivalent k-linear
categories. Hence comodH is locally of finite type, and the endomorphism ring EndH (A) k. This shows
simultaneously conditions (GT 2) and (GT 3).
To check condition (GT 1) we use the morphism between the tensor product Hopf algebroids, that is,
xo ⊗ x : (A ⊗ A, H o ⊗ H) → (k x ⊗ k x k, Hx ⊗ Hx ). As we have seen in subsection 3.4, this is also a
weak equivalence. Thus the category of right (H o ⊗ H)-comodules is equivalent, as a symmetric monoidal
category, to the category of right comodules over the Hopf k-algebra Hx ⊗ Hx , which as in the case of x
also implies that comodH o ⊗H and comodHx ⊗Hx are equivalent. Therefore, from one hand, we have by the
same reasoning as above that any comodule in comodH o ⊗H is projective as an (A ⊗ A)-module since it
is a dualizable comodule. On the other hand, we have that every right (H o ⊗ H)-comodule is a filtrated
inductive limit of objects in comodH o ⊗H since right (Hx ⊗ Hx )-comodules satisfies the same property
with respect to finite-dimensional right comodules comodHx ⊗Hx . Now, by apply [4, Proposition 5.1(ii)]
to the (A ⊗ A)-coring H o ⊗ H, we then conclude that every right (H o ⊗ H)-comodule is projective as an
(A ⊗ A)-module. Thus, H is projective as an (A ⊗ A)-module, which shows condition (GT 1).
The proof of (iv) ⇒ (i). Set B := A ⊗ A and K := H o ⊗ H. We know that (B, K) is a flat Hopf algebroid.
Since H is projective as (A ⊗ A)-module, we have that K is projective as (B ⊗ B)-module. Now, since the
map η is injective by Lemma 4.4, we can apply Lemma 3.12 by taking M = A as right H-comodule and
M ′ = A as right H o -comodule, to obtain the following chain of isomorphism
EndK (B) = EndH
o ⊗H
o
(A ⊗ A) EndH (A) ⊗ EndH (A) k ⊗ k k.
This means that the Hopf algebroid (B, K) satisfies the conditions of Lemma 4.3(b). Therefore, any right
K-comodule is faithfully flat as a B-module, henceforth, H is a faithfully flat (A⊗ A)-module. This finishes
the proof of Theorem 4.8.
Remark 4.10. Let (A, H) be a flat Hopf algebroid over k with A , 0 and A(k) = ∅. Then the same proof of
the implication (ii) ⇒ (iii) in Theorem 4.8, works for (A, H) by using any field extension L of k such that
A(L) , ∅. Assume now that (A, H) satisfies condition (i) of Theorem 4.8 and take a field extension L such
that A(L) , ∅. Then (AL , HL ) also satisfies this condition and by Theorem 4.8 we have that (AL , HL ) is a
GT Hopf algebroid, as we know that AL (L) = AlgL (AL , L) , ∅. Furthermore, if L is a perfect field, then by
applying [4, Théorème 6.1, page 5845] we can show that (A, H) is a GT Hopf algebroid as well. Summing
up, given a Hopf algebroid (A, H) as above, if its satisfies condition (i) of Theorem 4.8 and k admits a
perfect extension L such that A(L) , ∅, then (A, H) satisfies all the other conditions of this Theorem.
5. More properties of geometrically transitive Hopf algebroids
In this section we give more properties of GT Hopf algebroids. First we set up an analogous property of
transitive groupoids with respect to the conjugacy of their isotropy groups. To this end we introduce here
ON GEOMETRICALLY TRANSITIVE HOPF ALGEBROIDS
23
perhaps a known notion of isotropy Hopf algebra. This is the affine group scheme which represents the
presheaf of groups defined by the isotropy group at each fibre. Next we show that any two isotropy Hopf
algebras are weakly equivalent. The notion of conjugacy between two isotropy Hopf algebras, is not at all
obvious, and the 2-category of flat Hopf algebroids is employed in order to make it clearer. In this direction
we show that two isotropy Hopf algebras are conjugated if and only if the character groupoid is transitive,
and both conditions are fulfilled is the case of GT Hopf algebroids. Lastly, we give an elementary proof
of the fact that any dualizable comodule is locally free of constant rank, which in some sense bear out the
same property enjoyed by finite dimensional k-representations of a given transitive groupoid. The case
when the character groupoid of a GT Hopf algebroid is an empty groupoid, is also analyzed.
5.1. The isotropy Hopf algebras are weakly equivalent. Let (A, H) be a flat Hopf algebroid and H its
associated presheaf of groupoids. Assume as before that the base algebra satisfies A , 0 and A(k) , ∅, and
consider H (k) the character groupoid of (A, H), see Definition 3.5. As before, for each object x ∈ A(k),
we denote by kx the A-algebra k via the extension x, and consider the associated Hopf k-algebra of a base
ring extension (given by the k-algebra map x : A → kx ), that is, Hx := kx ⊗A H ⊗A kx .
Definition 5.1. Given an object x ∈ A(k). The Hopf algebra (kx , Hx ) is called the isotropy Hopf algebra of
(A, H) at the point x.
It noteworthy to mention that the associated affine k-group of (kx , Hx ) coincides with the one called
groupe d’inertie de x relativement à H as referred to in [10, III, §2, no 2; page 303].
The terminology used in Definition 5.1 is, in relation with groupoids, justified by the following lemma.
Fix an object x ∈ A(k), we denoted by 1x the unit element of the A-algebra kx . Take C to be an algebra with
unit map 1C : k → C. Composing with x, we have then an object x∗ (1C ) = 1C ◦ x ∈ A(C). Let us denote by
∗
G x (C) := H (C)x (1C ) the isotropy group of the object x∗ (1C ) in the groupoid H (C), see equation (1). This
construction is clearly funtorial and so leads to a presheaf of groups G x : Algk → Grps, C → G x (C).
Lemma 5.2. For any x ∈ A(k), the presheaf of groups G x is affine, and up to a natural isomorphism, is
represented by the Hopf k-algebra Hx .
Proof. Given an element g in the group G x (C), that is, an algebra map g : H → C such that g ◦ t = g ◦ s =
x∗ (1C), we can define the following algebra map:
κC (g) : Hx −→ C,
k1x ⊗A u ⊗A k′ 1x 7−→ kk′ g(u) ,
which is clearly functorial in C. This leads to a natural transformation κ− : G x (−) −→ Algk (Hx , −).
Conversely, to any algebra map h : Hx → C, one associate the algebra map
νC (h) := h ◦ τx : H −→ Hx −→ C,
where τx : H → Hx sends u 7→ 1 x ⊗A u ⊗A 1x . This construction is also functorial in C, which defines
a natural transformation ν− : Algk (Hx , −) −→ G x (−). It is not difficult now to check that both natural
transformations κ and ν, are mutually inverse.
Recall that for groupoids the transitivity property is interpreted by means of conjugation between theirs
isotropy groups, which means that any two of these groups are isomorphic. Next we show how this last
property is reflected at the level of the isotropy Hopf algebras. The conjugacy of the isotropy Hopf algebras,
in relation with the transitivity of the character groupoid, will be considered in the next subsection.
Proposition 5.3. Let (A, H) be a flat Hopf algebroid with A , 0 and A(k) , ∅. Assume that (A, H) is
geometrically transitive. Then any two isotropy Hopf algebras are weakly equivalent.
Proof. Take two objects x, y ∈ A(k) and consider as before the following diagram
(kx , Hx ) i❙
❙❙❙❙
❙
x
(A, H)
(ky , Hy )
❦❦5
❦
❦
❦
❦ y
(28)
of Hopf algebroids. By Theorem 4.8, both x and y are weak equivalences, in particular, the Hopf algebras
(kx , Hx ) and (ky , Hy ) are Morita equivalent, in the sense that their categories of comodules are equivalent as
symmetric monoidal k-linear categories. Therefore, (kx , Hx ) and (ky , Hy ) are weakly equivalent by applying
[11, Theorem A].
24
LAIACHI EL KAOUTIT
Remark 5.4. In the terminology of [28, Definition 3.2.3], the Hopf algebras (kx , Hx ) and (ky , Hy ) are said
to be monoidally Morita-Takeuchi equivalent. By applying [28, Corollary 3.2.3], there is a Hopf bi-Galois
object, or a principal bi-bundle as in subsection 3.4, connecting Hx and Hy (notice here that the side on
comodules is not relevant since the Hopf algebras are commutative).
Next, we compute explicitly, by using results from [11], the principal bi-bundle connecting (kx , Hx ) and
(ky , Hy ), as was mentioned in the previous Remark. Following [11], any two weakly equivalent flat Hopf
algebroids are connected by a two-stage zig-zag of weak equivalences, and this is the case for the previous
Hopf algebras. That is, in the situation of Proposition 5.3, diagram (28) can be completed to a square by
considering the two-sided translation Hopf algebroid built up by using the principal bibundle connecting
(kx , Hx ) and (ky , Hy ), see subsection 3.4. In more specific way, we have the two trivial principal bibundles
Px := H ⊗A kx and Py := H ⊗A ky which correspond, respectively, to the weak equivalences x and y. Notice
that Px is an (H, Hx )-bicomodule algebra with algebra maps
(29)
αx : A → Px ,
a 7→ s(a) ⊗A 1 and βx : kx → Px ,
k 7→ 1H ⊗A k1x .
Similar notations are applied to the (H, Hy)-bicomodule algebra Py . The cotensor product of these two
bibundles Px co H Py is again a principal (Hx , Hy )-bibundle (recall here that Px co is the opposite bundle of
Px ). The algebra maps defining this structure are βex : kx −→ Px co H Py ←− ky : βey , given by
βex (k) = βx (k) H 1Py ,
βey (k) = 1Px H βy (k),
where the notation is the obvious one. The associated two-sided translation Hopf algebroid is described as
follows. First we observe the following general fact in Hopf algebroids with source equal to the target, i.e.,
Hopf algebras over commutative algebras.
Lemma 5.5. Let (R, L) and (R′ , L′ ) be two commutative Hopf algebras, and assume that there is a diagram
of Hopf algebroids:
′ ′
(R, L) h❘
5 (R , L )
❘❘❘❘
❧
❧
❧
❧❧❧ω′
ω ❘
(A, H)
Then the pair R ⊗A H ⊗A R′ , L ⊗A H ⊗A L′ of algebras, admits a structure of Hopf algebroid with maps:
• the source and target:
s(r ⊗A u ⊗A r′ ) := r1L ⊗A u ⊗A r′ 1L′ ,
t(r ⊗A u ⊗A r′ ) := rω(S (u(1))) ⊗A u(2) ⊗A ω′ (u(3) )r′ ;
• comultiplication and counit:
∆(l ⊗A u ⊗A l′ ) := l(1) ⊗A u ⊗A l′(1) ⊗C l(2) ⊗A 1H ⊗A l′(2) , ε(l ⊗A u ⊗A l′ ) := εL (l) ⊗A u ⊗A εL′ (l′ );
• the antipode:
S (l ⊗A u ⊗A l′ ) := SL l ω(u(1) ) ⊗A u(2) ⊗A ω′ (u(3) )SL′ (l′ ).
Proof. These are routine computations.
Now we come back to the situation of Proposition 5.3. Consider the following algebras:
Px, y := kx ⊗A H ⊗A ky ,
Hx, y := Hx ⊗A H ⊗A Hy ,
with the structure of Hopf algebroid, as in Lemma 5.5. Consider then the following obvious algebra maps
ωx : Hx −→ Hx, y , k1x ⊗A u ⊗A k′ 1x 7−→ k1Hx ⊗A u ⊗A k′ 1Hy ;
and
ωy : Hy −→ Hx,y , k1y ⊗A u ⊗A k′ 1y 7−→ k1Hx ⊗A u ⊗A k′ 1Hy .
Proposition 5.6. Let (A, H) be as in Proposition 5.3, consider x, y ∈ A(k) and their associated isotropy
Hopf algebras (kx , Hx ) and (ky , Hy ). Assume that (A, H) is geometrically transitive. Then there is an
isomorphism
Px co H Py , Hx X (Px co H Py ) Y Hy Px, y , Hx, y
ON GEOMETRICALLY TRANSITIVE HOPF ALGEBROIDS
25
of Hopf algebroids with the following diagram
(Px, y , Hx, y ) ω
i❚❚❚❚y
ω x ❥❥5
❚❚
❥❥❥❥
(kx , Hx ) j❚
4 (ky , Hy )
❚❚❚❚
❥❥❥❥
❚
❥
❥
❚
❥
y
x
(A, H)
of weak equivalences.
Proof. The stated isomorphism follows directly by comparing the structure of the two-sided translation
Hopf algebroid, as given in subsection 3.4, with that of (Px, y , Hx, y ) given in Lemma 5.5. By Proposition 5.3,
we know that x and y are weak equivalences. Therefore, ωx and ωy are weak equivalences by applying [11,
Proposition 6.3] in conjunction with the previous isomorphism of Hopf algebroids.
Remark 5.7. The diagram stated in Proposition 5.6, is not necessarily strictly commutative; however, it is
commutative up to a 2-isomorphism in the 2-category of flat Hopf algebroids described in subsection 3.2.
Precisely, one shows by applying [11, Lemma 6.11] that there is a 2-isomorphism ωx ◦ x ωx ◦ y.
5.2. The transitivity of the character groupoid. Let (A, H) be a flat Hopf algebroid as in the previous
subsection and consider its character groupoid H (k) = (H(k), A(k)). We have seen in Theorem 4.8 that
(A, H) is geometrically transitive if and only if the attached presheaf of groupoids H is locally transitive,
that is, satisfies condition (ii) of that theorem. The aim of this subsection is to characterize the transitivity
of the groupoid H (k), by means of the conjugation between the isotropy Hopf algebras. First we introduce
the notion of conjugacy.
Definition 5.8. Let x, y be two objects in H (k). We say that the isotropy Hopf algebras (k x , Hx ) and
(ky , Hy ) are conjugated, provided there is an isomorphism of Hopf algebras g : (k x , Hx ) → (ky , Hy ) such
that the following diagram
(kx , Hx ) i❘
❘❘❘❘
❘
x
g
(A, H)
/ (ky , Hy )
❧5
❧
❧
❧
❧❧ y
is commutative up to a 2-isomorphism, where Hopf k-algebras are considered as 0-cells in the 2-category
of flat Hopf algebroids described in subsection 3.2.
As in [11, §6.4], this means that there is an algebra map g : H → k such that
g ◦ s = x,
g ◦ t = y, and u(1) − ⊗A u(1) 0 ⊗A u(1) + g(u(2)) = g(u(1)) ⊗A u(2) ⊗A 1y ∈ Hy
(30)
where, by denoting the Hopf algebroids map z := g ◦ x : (A, H) → (kx , Hx ), we have
z0 = x
and
z1 (u) = g(1 x ⊗A u ⊗A 1 x ) := u− ⊗A u0 ⊗A u+ (summation understood).
Proposition 5.9. Let (A, H) be a flat Hopf algebroid with A , 0 and A(k) , ∅. Assume that (A, H) is
geometrically transitive. Then the following are equivalent:
(i) the character groupoid H (k) is transitive;
(ii) for any two objects x, y in H (k), the algebras H ⊗A kx and H ⊗A ky are isomorphic as left H-comodules
algebras;
(iii) any two isotropy Hopf algebras are conjugated.
Furthermore, under the same assumptions, condition (i) is always fulfilled.
Proof. We first check the equivalences between these conditions. So, let x ∈ A(k) = H0 (k) and denote as
before by Px := H ⊗A kx the stated left H-comodule algebra.
(i) ⇒ (ii). Given x, y ∈ A(k), by assumption there is an algebra map h : H → k such that h ◦ s = x and
h ◦ t = y. So we can define the following map
F : Px −→ Py ,
u ⊗A k1x 7−→ u(1) ⊗A h S (u(2) ) k1y .
Clearly F is an A-algebra map, and so it is left A-linear. The fact that F is left H-colinear is also clear, and
this shows condition (ii), since F is obviously bijective.
26
LAIACHI EL KAOUTIT
(ii) ⇒ (iii). Assume for a given x, y ∈ A(k), there is a left H-comodule algebra isomorphism F : Px →
Py . For any u ∈ H, we denote by F(u ⊗A 1x ) = u− ⊗A u+ (summation understood). Consider the k-linear map
g : H → k which sends u 7→ y ε(u− ) u+ . This is a k-algebra map since F it is so. For any a ∈ A, we have
g s(a) = y ε(s(a)− ) s(a)+ = y ε(s(a)) 1 = y(a)
and
g t(a) = y ε(t(a)− ) t(a)+ = y ε(1H ) x(a)1 = x(a),
as F is k-linear. Define the map
g : (k x , Hx ) −→ (ky , Hy ),
(k1x , 1x ⊗A u ⊗A 1x ) 7−→ (k1y , g(S (u(1)))1y ⊗A u(2) ⊗A g(u(3) )1y ) .
By using the characterization given in Lemma 5.2, or a direct computation, one can shows that this map is
an isomorphism of Hopf algebras. Furthermore, it is easily seen that the pair (g, g) satisfies the equalities
of equation (30). Thus, (kx , Hx ) and (ky , Hy ) are conjugated, which means condition (iii).
(iii) ⇒ (i). This implication follows immediately from equation (30).
Let us check that condition (i) is fulfilled under assumption. For a given element u ∈ H there exists, by
the isomorphism of Remark 4.7, a finite family M1 , · · · , Mk of dualizable right H-comodules and finite set
P
of elements {(pl , ϕl )}1≤ l ≤k , pl ∈ Ml and ϕl ∈ Ml∗ , such that u is uniquely written as u = l s ϕl (pl(0) ) pl(1) (see
[13, Section 4] for more details on the
quoted in Remark 4.7). Given now two objects x, y ∈ A(k),
map can
we define g : H → k by g(u) := y s ϕl (pl(0) ) x ε(pl(1) ) . It turns out that g is a well defined algebra map,
which satisfies g ◦ s = y and g ◦ t = x. This shows that H (k) is transitive and finishes the proof.
5.3. GT Hopf algebroids and principal bundles over Hopf algebras. Parallel to subsection 2.6 we
study here the relationship between GT Hopf algebroids and principal bundles over Hopf algebras (i.e.,
commutative Hopf Galois extensions [23, §8], [28], or k-torsor as in [15] and [10]). This is a restricted
notion of principal bundle, as defined in subsection 3.4, to the case of Hopf algebras.
To be precise, let B be a commutative Hopf algebra over k, a pair (P, α) consisting of an algebra extension
α : A → P and a right B-comodule algebra P with left A-linear coaction, is said to be a right principal
B-bundle provided α is faithfully flat and the canonical map canP : P ⊗A P → P ⊗ H, x ⊗A y 7→ xy(0) ⊗ y(1)
is bijective. Notice that if we translate this definition to the associated affine k-schemes, then the outcome
characterizes in fact the notion of torsors as it was shown in [10, Corollaire 1.7, page 362], see also [15,
Définition 1.4.1, page 117].
Let (A, H) be a Hopf algebroid as in subsection 5.1 and H its associated presheaf of groupoids. Take
an object x ∈ A(k) and consider as before Px = H ⊗A kx the right comodule algebra over the isotropy Hopf
algebra (kx , Hx ) with the algebra extension αx : A → Px of equation (29). On the other hand denote by Px
the presheaf of sets which associated to each algebra C the set Px (C) := t−1 {1C ◦ x} where t is the target
of the groupoid H (C).
Lemma 5.10. For any x ∈ A(k), the presheaf of sets Px is affine, and up to a natural isomorphism, is
represented by the algebra Px . Furthermore, if (A, H) is geometrically transitive, then (Px , αx ) is a principal
right Hx -bundle.
Proof. The first claim is an immediate verification. The last one is a consequence of Theorem 4.8.
In contrast with the case of transitive groupoids described in subsection 2.6, the converse in Lemma 5.10
is not obvious. Specifically, it is not automatic to construct a GT Hopf algebroid from a principal bundle
over a Hopf algebra. In more details, let (P, α) be a right principal B-bundle over a Hopf algebra B with
extension α : A → P, and consider P ⊗ P as a right B-comodule algebra via the diagonal coaction and set
o
n
H := (P ⊗ P)coinvB = u ∈ P ⊗ P| ̺P⊗ P (u) = u ⊗ 1B
its coinvariant subalgebra. The map α induces two maps s, t : A → H which going to be the source and
the target. The counity is induced by the multiplication of P. The comultiplication is derived from that of
(P, P ⊗ P), however, not in an immediate way, because slightly technical assumptions are needed for this.
Precisely, consider M := (P ⊗ P) ⊗A (P ⊗ P) as a right B-comodule algebra with the coaction
̺ : M −→ M ⊗ B,
(x ⊗ y) ⊗A (u ⊗ v) 7−→ (x(0) ⊗ y(0) ) ⊗A (u(0) ⊗ v(0) ) ⊗ x(1) y(1) u(1) v(1) .
ON GEOMETRICALLY TRANSITIVE HOPF ALGEBROIDS
27
This is a well defined coaction since we know that PcoinvB A. Clearly we have that H ⊗A H ⊆ McoinvB , and
under the assumption of equality we obtain:
Proposition 5.11. Let (P, α) be a right principal B-bundle over a Hopf algebra B with extension α : A → P.
Denote by υ : H := (P ⊗ P)coinB → P ⊗ P the canonical injection where P ⊗ P is a right B-comodule algebra
via the diagonal coaction. Assume that υ is a faithfully flat extension and that H ⊗A H = McoinB .
Then (A, H) admits a unique structure of Hopf algebroid such that (α, υ) : (A, H) → (P, P ⊗ P) is a
morphism of GT Hopf algebroids.
Proof. First observe that the map s : A → H is a flat extension (and so is t) since α and υ are faithfully flat
extension and we have a commutative diagram:
υ
/H
O O
0
s
t
α
/A
0
/ P⊗ P
O O
/P
of algebra maps. The fact that (A, H) admits a coassociative comultiplication follows essentially form the
second assumption. Indeed, let ∆′ : P ⊗ P → M be the map which sends x ⊗ y 7→ (x ⊗ 1) ⊗A (1 ⊗ y), so
it is easily checked that, under the stated assumption, there is a map H → H ⊗A H which completes the
diagram:
0
/M
O
/ H ⊗A H
O
/H
−⊗1
// M ⊗ B
∆′
∆
0
̺
υ
/ P⊗ P
This gives a coassociative comultiplication on the A-bimodule H using the structure of A-bimodule derived
from the above source and the target s, t. To check that ∆ is counital one uses the following equalities
(p ⊗ 1) ⊗A (1 ⊗ q) = p(0) ⊗ p−(1) q−(1) ⊗A p+(1) q+(1) ⊗ q(0) ∈ (P ⊗ P) ⊗A (P ⊗ P),
together with the properties of the translation map δ : B → P ⊗A P, b 7→ b− ⊗A b+ given by the inverse of
the canonical map canP .
With the previous structure maps, (A, H) is now a Hopf algebroid such that the pair of maps (α, υ) :
(A, H) → (P, P ⊗ P) is a morphism of Hopf algebroids with codomain a GT Hopf algebroid. Lastly, since
α ⊗ α is a faithfully flat extension, s ⊗ t : A ⊗ A → H is also faithfully flat, and hence (A, H) is by Theorem
4.8 a GT Hopf algebroid as well.
5.4. GT Hopf algebroids with empty character groupoid. Let (A, H) be a flat Hopf algebroid over k
with A , 0 and A(k) = ∅. For example, taking any non zero algebra A with A(k) = ∅ and consider the
Hopf algebroids given in Examples 3.1 and 3.3. Now, let L be a field extension of k such that A(L) , ∅
and denote by OL : AlgL → Algk the forgetful functor from the category of commutative L-algebras to
k-algebras. Next, we will use the notations of Example 3.4. So, fix an algebra map q ∈ A(L) and denote
by e
q ∈ AL (L) = AlgL (AL , L) its image, that is, the L-algebra map e
q : AL → L sending a ⊗ l 7→ q(a)l.
Consider the base extension Hopf algebroid (Lq , Lq ⊗A H ⊗A Lq ) := (Lq , Hq ) over k, where Lq is considered
as an algebra extension of A via the map q. The associated presheaf of groupoids is denoted by Hq and its
fq := Hq ◦ OL . In this way, we get a presheaf of groups
composition with OL by H
fq ⋆ : Alg −→ Grps,
fq (R)z
H
R −→ H
(31)
L
fq (R)z is the isotropy
where z : L → R is the k-algebra map defining R as an object in AlgL and where H
fq (R) attached to the object z. Thus, for any pair (R, z) as before, we have by
group of the groupoid H
Example 3.8 that
fq (R)z := (z, g, z)| g ∈ H(R) such that gs = gt = zq
H
where the multiplication is given as in Example 2.4 and the unit is the element (z, zqε, z).
28
LAIACHI EL KAOUTIT
On the other hand, following Lemma 5.2 we can define the presheaf of groups attached to the Hopf
algebroid (AL , HL ) over L, at the point e
q. That is, we can consider the presheaf HL associated to (AL , HL ),
and denote by
GLeq : AlgL −→ Grps,
R −→ GLeq (R)
(32)
where GLeq (R) is the isotropy group of the groupoid HL (R) at the point e
q ∈ AL (L). Thus
GLeq (R) := h ∈ HL (R) = AlgL (HL , R)| hsL = htL = e
q ,
where the group structure comes from the groupoid HL (R).
Proposition 5.12. Given (R, z) as above, then the following morphisms of groups
GLeq (R)
✤
h
φq, R
fq (R)z
/H
/ (z, ĥ, z),
where ĥ ∈ H(R) is the k-algebra map sending u 7→ h(u ⊗ 1), establish a natural isomorphism
fq
φq : GLeq −→ H
⋆
fq ⋆ is represented by the isotropy
of presheaves of groups. In particular, up to a natural isomorphism, H
Hopf L-algebra (L, HL, eq ) of the Hopf algebroid (AL , HL ) at the point e
q.
Proof. Let us first check that φq, R is a well defined map. Take h ∈ GLeq , then, for every a ∈ A, we have
ĥ ◦ s(a) = h(s(a) ⊗ 1) = h ◦ sL (a ⊗ 1) = h ◦ tL (a ⊗ 1) = h ◦ (t(a) ⊗ 1) = ĥ ◦ t(a) = e
q(a) = q(a).1L = zq(a).
fq (R)z . The image by φq, R of the identity element e
Hence (z, ĥ, z) ∈ H
qεL is (z, e
qc
εL , z) = (z, zqε, z) which is
e
q
z
′
f
the identity element of the group Hq (R) . Now, given h, h ∈ GL (R) and u ∈ H, we have that
c′ (u) = (hh′ )(u ⊗ 1) = h′ (u ⊗ 1)(1) h (u ⊗ 1)(2) = h′ u(1) ⊗ 1 h u(2) ⊗ 1 = ĥ′ (u(1) ) ĥ(u(2)) = (ĥ ĥ′ )(u)
hh
c′ = ĥ ĥ′ . Therefore,
which implies that hh
c′ , z),
(z, ĥ, z) (z, ĥ′ , z) = (z, ĥhˆ′ , z) = (z, hh
which shows that φq, R is a morphism of groups. On the other hand, φq, R is clearly injective and if we take
fq (R)z and set h = ĝ : HL → R sending u ⊗ l 7→ g(u) ⊗ l, then we have that
an element (z, g, z) ∈ H
(z, g, z) = φq, R (h). This shows that φq, R is also surjective, and thus an isomorphism of groups. Lastly, it
is immediate to see that φq, − is a natural transformation and so a natural isomorphism as desired. The
particular statement follows directly from Lemma 5.2.
Remark 5.13. Let (Lq , Lq ⊗A H ⊗A Lq ) be as above the base change Hopf algebroid of (A, H) and denote
by (L, Hq ) its quotient Hopf
L-algebra
where Hq := Lq ⊗A H ⊗A Lq /hs − ti is the quotient modulo the Hopf
ideal generated by the set s(l) − t(l)
. Then the following map of L-vector spaces
l∈L
Hq
l ⊗A u ⊗A l′
✤
/ HL, eq
/ 1 ⊗A (u ⊗ ll′ ) ⊗A 1
L
L
is a surjective morphism of Hopf L-algebras. On the other hand, the presheaf of set Peq : AlgL → Sets
defined as in Lemma 5.10 for the Hopf algebroid (AL , HL ) is, up to a natural isomorphisms, represented by
the left H-comodule L-algebra H ⊗A Lq which under condition (i) of Theorem 4.8, becomes a principal
(H, Hq)-bibundle.
Proposition 5.14. Let (A, H) be a flat Hopf algebroid over k with A , 0 and A(k) = ∅, denote by H its
associated presheaf of groupoids. Consider L a field extension of k such that A(L) , ∅. Assume that the
unit map η = s ⊗ t : A ⊗ A → H is a faithfully flat extension. Then
(1) H (L) is a transitive groupoid;
(2) For every p, q ∈ A(L), the base change Hopf algebroids (Lq , Hq ) and (L p , H p ) are weakly equivalent.
ON GEOMETRICALLY TRANSITIVE HOPF ALGEBROIDS
29
Proof. As we have observed in Remark 4.10, if η is a faithfully flat extension, then so is ηL : AL ⊗L AL →
HL . Since AL (L) , ∅, we have by Theorem 4.8, that (AL , HL ) is geometrically transitive Hopf algebroid.
Therefore, by applying Proposition 5.9, we know that HL (L) is a transitive groupoid. Now, given p, q ∈
A(L) we obtain two objects e
p, e
q ∈ AL (L) of this groupoid. Hence, there exists an L-algebra map h ∈ HL (L)
such that h ◦ sL = e
p and h ◦ tL = e
q. Consider the algebra map g = ĥ : H → L sending u 7→ h(u ⊗ 1), so we
have that g ◦ s = p and g ◦ t = q. This proves part (1). As for part (2), we know by Remark 4.10 that (A, H)
satisfies condition (iii) of Theorem 4.8. Henceforth, the canonical base change maps (A, H) → (L p , H p )
and (A, H) → (Lq , Hq ) are weak equivalences by [11, Proposition 5.1]. Therefore, (L p , H p ) and (Lq , Hq )
are weakly equivalent by applying [11, Theorem A] and this finishes the proof.
5.5. Dualizable comodules over GT Hopf algebroids are locally free of constant rank. The aim of this
subsection is to apply Theorem 4.8 in order to give an elementary proof of the well know fact sated in [9,
page 114] which implicitly asserts that over a GT Hopf algebroid with non empty character groupoid, any
comodule which has a locally free fibre with rank n, then so are other fibres. An important consequence of
this fact is that any dualizable comodule over such a Hopf algebroid is locally free with constant rank. This
is an algebraic interpretation of a well known property on representations of transitive groupoid in vector
spaces. Namely, if a given representation over such a groupoid has a finite dimensional fibre, then so are
all other fibres and all the fibres have the same dimension. We start by the following general lemma which
will be needed below.
Lemma 5.15. Let ϕ : R → T be a faithfully flat extension of commutative algebras. Then, for any R-module
P, the following conditions are equivalent.
(i) P is locally free R-module of constant rank n;
(ii) Pϕ := P ⊗R T is locally free T -module of constant rank n.
Proof. By [2, Proposition 12, page 53, and Théorème 1, page 138], we only need to check that P is of a
constant rank n if and only if so is Pϕ . So let us first denote by ϕ∗ : Spec(T ) → Spec(R) the associated continuous map of ϕ. Denote by rRP : Spec(R) → Z and rTPϕ : Spec(T ) → Z, the rank functions corresponding,
respectively, to P and Pϕ .
It suffices to check that rRP is a constant function with value n if and only if rTPϕ is a constant function with
the same value. Given a prime ideal a ∈ Spec(T ), consider the localising algebras T a and Rϕ∗ (a) at the prime
ideals a and ϕ∗ (a). It is clear that we have an isomorphism of T a -modules P ⊗R T a Pϕ∗ (a) ⊗Rϕ∗ (a) T a , where
ϕa : Rϕ∗ (a) → T a is the associated localisation map of the extension ϕ. Therefore, the free modules P ⊗R T a
and Pϕ∗ (a) have the same rank. Hence, we have rRP ϕ∗ (a) = rTPϕ (a), for any a ∈ Spec(T ), and so rRP ◦ ϕ∗ = rTPϕ .
This shows that if rRP is a constant function with value n, then so is rTPϕ . The converse also hods true since
we know that ϕ∗ is surjective, because of the faithfully flatness of ϕ, and this finishes the proof.
Proposition 5.16. Let (A, H) be a flat Hopf algebroid with A , 0 and A(k) , ∅. Assume that (A, H)
is geometrically transitive, and let M be a (right) H-comodule whose underlying A-module is finitely
generated and projective. Given two objects x ∈ A(S ) and y ∈ A(T ), then the following are equivalent
(i) Mx := M ⊗A S is locally free S -module of constant rank n;
(ii) My := M ⊗A T is locally free T -module of constant rank n.
Proof. Let us first show that the stated conditions are equivalent when R = S = T . In this case we know
by Theorem 4.8, that the objects x, y ∈ A(R) are locally isomorphic. Therefore, there exists a faithfully flat
extension p : R → R′ such that Mex = M ⊗ex R′ is isomorphic as R′ -module to Mey = M ⊗ey R′ , where e
x = p◦x
and e
y = p ◦ y. Thus, Mex and Mey they have the same rank function.
On the other hand, by applying Lemma 5.15 to Mx , we get that Mex is locally free R′ -module of constant
rank n if and only if Mx is locally free R-module of constant rank n. The same result hold true using My
and Mey . Therefore, Mx is locally free R-module of constant rank n if and only if so is My .
For the general case S , T , consider R := T ⊗ S and set the algebra maps x := ιS ◦ x, y := ιT ◦ y,
where ιS : S → R ← T : ιT are the obvious maps. By the previous case, we know that Mx is locally free
R-module of constant rank n if and only if so is My . Now by Lemma 5.15, we have, from one hand, that Mx
is locally free R-module of constant rank n if and only if Mx is locally free S -module of constant rank n,
and from the other, we have that My is locally free R-module of constant rank n if and only if My is locally
30
LAIACHI EL KAOUTIT
free T -module of constant rank n. Therefore, Mx is locally free S -module of constant rank n if and only if
My is so as T -module.
As a corollary of Proposition 5.16, we have:
Corollary 5.17. Let (A, H) be a flat Hopf algebroid with A , 0 and A(k) , ∅. Assume that (A, H) is
geometrically transitive. Then every dualizable (right) H-comodule is a locally free A-module of constant
rank. In particular, given a dualizable right H-comodule M and two distinct object x , y ∈ A(k), then Mx
and My have the same dimension as k-vector spaces.
Aknowledgements: The author would like to thank the referee for her/his thorough review and highly
appreciate the comments and suggestions, which significantly contributed to improving the paper.
References
[1] Leovigilio Alonso Tarrı́o, Ana Jeremı́as López, Marta Pérez Rodrı́guez and Marı́a J. Vale Gonsalves, A funtorial formalism for
quasi-coherent sheaves on geometric stacks. Expo. Math. 33 (2015), no. 4, 452–501.
[2] N. Bourbaki, Éléments de mathématique. Algèbre commutative. Chapitre 1-4, Actualités Scientifiques et Industrielles, No. 1290,
Herman, Paris, 1961.
[3] R. Brown, From groups to groupoids: A brief survey. Bull. London Math. Soc. 19 (2), (1987), 113–134.
[4] A. Bruguières, Thèorie Tannakienne non commutative, Commun. in Algebra 22 (1994), 5817–5860.
[5] T. Brzeziński, Galois comodules, J. Algebra 290 (2005), 503–537.
[6] T. Brzeziński and R. Wisbauer, Corings and comodules, London Mathematical Society Lecture Note Series, vol. 309, Cambridge
University Press, Cambridge, 2003.
[7] S. Bouc, Biset Functors for finite Groups, LNM, vol. 1999. Springer-Verlag Berlin Heidelberg 2010.
[8] P. Cartier, Groupoı̈des de Lie et leurs Algèbroı̈des. Séminaire Bourbaki 60e année, 2007-2008, num. 987, 165–196.
[9] P. Deligne, Catégories tannakiennes. In The Grothendieck Festschrift (P. Cartier et al., eds.), Progr. Math., 87, vol. II, Birkhäuser,
Boston, MA. 1990, pp. 111–195.
[10] M. Demazure and P. Gabriel, Groupes algébriques. Tome I: Géométrie algébrique, généralités, groupes commutatifs, Masson
& Cie, Éditeur, Paris; North-Holland Publishing Co., Amsterdam, 1970, Avec un appendice Corps de classes local par Michiel
Hazewinkel.
[11] L. El Kaoutit and N. Kowalzig, Morita theory for Hopf algebroids, principal bibundles, and weak equivalences. Doc. Math. 22
(2017), 551–609.
[12] L. El Kaoutit and J. Gómez-Torrecillas, Comatrix corings: Galois corings, descent theory, and a structure theorem for
cosemisimple corings. Math. Z., 244 (2003), 887–906.
[13] L. El Kaoutit and J. Gómez-Torrecillas, Infinite comatrix corings. Int. Math. Res. Notices, 39 (2004), 2017–2037.
[14] P. Gabriel, Construction de preschemas quotient. Exposé V. Propriétés Générales des Schémas en Groupes LNM Vol.151
Springer Verlag 1970 (SGA 3 -I) pp. 252–286.
[15] J. Giraud, Cohomologie non-abélienne. Die Grundlehren der mathematischen Wissenschaften in Einzeldarstellungen Band 179.
Springer-Verlag Berlin Heidelberg New York 1971.
[16] P. J. Higgins, Notes on categories and groupoids, Van Nostrand Reinhold, Mathematical Studies 32, London 1971.
[17] M. Hovey and N. Strickland, Comodules and Landweber exact homology theories. Adv. Math. 192 (2005), no. 2, 427–456.
[18] B. Jelenc. Serre fibrations in the Morita category of topological groupoids, Topol. Appl. 160 (2013), 9–13.
[19] Peter S. Landweber, Associated prime ideals and Hopf algebras, J. Pure Appl. Algebra 3 (1973), 43–58.
[20] G. Laumon and L. Moret-Baily, Champs algébriques. Vol. 39, Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A
Series of Modern Surveys in Mathematicas. Springer-Verlag, Berlin 2000.
[21] Kirill C. H. Mackenzie, General Theory of Lie Groupoids and Lie Algebroids, London Math. Soc. Lecture Note Ser., vol. 213,
Cambridge Univ. Press, Cambridge, 2005.
[22] I. Moerdijk and J. Mrčun, Lie groupoids, sheaves and cohomology, Poisson geometry, deformation quantisation and group
representations, London Math. Soc. Lecture Note Ser., vol. 323, Cambridge Univ. Press, Cambridge, 2005, pp. 145–272.
[23] S. Montgomery, Hopf algebras and their actions on rings. CBMS Regional Conference Series in Mathematics, 82 American
Mathematical Society, Providence, RI, 1993. xiv+238 pp.
[24] N. Naumann, The stack of formal groups in stable homotopy theory. Adv. Math. 215 (2007), no. 2, 569–600.
[25] Geoffery M. L. Powell, On affine morphisms of Hopf algebroids. Homology, Homotopy Appl. 10 (2008), no. 1, 53–95.
[26] D. C. Ravenel, Complex Cobordism and Stable Homotopy Groups of Spheres. Pure and Applied Mathematics Series, Academic
Press, San Diego, 1986.
[27] J. Renault, A groupoid approach to C ∗ -algebras. Lecture Notes in Mathematics 793, Springer Verlag, 1980.
[28] P. Schauenburg, Hopf-Galois and bi-Galois extensions, Galois theory, Hopf algebras, and semiabelian categories, Fields Inst.
Commun., vol. 43, Amer. Math. Soc., Providence, RI, 2004, pp. 469–515.
Universidad de Granada. Departamento de Álgebra y IEMath-Granada. Facultad de Educación, Econonı́a y Tecnologı́a de
Ceuta. Cortadura del Valle, s/n. E-51001 Ceuta, Spain
E-mail address: [email protected]
URL: http://www.ugr.es/˜kaoutit/
| 4math.GR
|
Scalable Partial Least Squares Regression on
Grammar-Compressed Data Matrices
Yasuo Tabei
Hiroto Saigo
Japan Science and
Technology Agency, Japan
Kyushu University, Japan
[email protected]
Yoshihiro Yamanishi
arXiv:1606.05031v1 [cs.DS] 16 Jun 2016
Kyushu University, Japan
[email protected]
Simon J. Puglisi
University of Helsinki, Finland
[email protected] [email protected]
u.ac.jp
ABSTRACT
With massive high-dimensional data now commonplace in
research and industry, there is a strong and growing demand for more scalable computational techniques for data
analysis and knowledge discovery. Key to turning these data
into knowledge is the ability to learn statistical models with
high interpretability. Current methods for learning statistical models either produce models that are not interpretable
or have prohibitive computational costs when applied to
massive data. In this paper we address this need by presenting a scalable algorithm for partial least squares regression (PLS), which we call compression-based PLS (cPLS),
to learn predictive linear models with a high interpretability
from massive high-dimensional data. We propose a novel
grammar-compressed representation of data matrices that
supports fast row and column access while the data matrix is
in a compressed form. The original data matrix is grammarcompressed and then the linear model in PLS is learned on
the compressed data matrix, which results in a significant
reduction in working space, greatly improving scalability.
We experimentally test cPLS on its ability to learn linear
models for classification, regression and feature extraction
with various massive high-dimensional data, and show that
cPLS performs superiorly in terms of prediction accuracy,
computational efficiency, and interpretability.
1.
INTRODUCTION
Massive data are now abundant throughout research and
industry, in areas such as biology, chemistry, economics, digital libraries and data management systems. In most of these
fields, extracting meaningful knowledge from a vast amount
of data is now the key challenge. For example, to remain
competitive, e-commerce companies need to constantly analyze huge data of user reviews and purchasing histories [24].
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than
ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from [email protected].
KDD ’16, August 13-17, 2016, San Francisco, CA, USA
c 2016 ACM. ISBN 978-1-4503-4232-2/16/08. . . $15.00
DOI: http://dx.doi.org/10.1145/2939672.2939864
In biology, detection of functional interactions of compounds
and proteins is an important part in genomic drug discovery [29, 7] and requires analysis of a huge number of chemical compounds [3] and proteins coded in fully sequenced
genomes [4]. There is thus a strong and growing demand for
developing new, more powerful methods to make better use
of massive data and to discover meaningful knowledge on a
large scale.
Learning statistical models from data is an attractive approach for making use of massive high-dimensional data.
However, due to high runtime and memory costs, learning
of statistical models from massive data — especially models
that have high interpretability — remains a challenge.
Partial least squares regression (PLS) is a linear statistical
model with latent features behind high-dimensional data [26,
35, 36] that greedily finds the latent features by optimizing the objective function under the orthogonal constraint.
PLS is suitable for data mining, because extracted latent
features in PLS provide a low-dimensional feature representation of the original data, making it easier for practitioners
to interpret the results. From a technical viewpoint, the
optimization algorithm in PLS depends only on elementary
matrix calculations of addition and multiplication. Thus,
PLS is more attractive than other machine learning methods that are based on computationally burdensome mathematical programming and complex optimization solvers. In
fact, PLS is the most common chemoinformatics method in
pharmaceutical research.
However, applying PLS to massive high-dimensional data
is problematic. While the memory for the optimization algorithm in PLS depends only on the size of the corresponding
data matrix, storing all high-dimensional feature vectors in
the data matrix consumes a huge amount of memory, which
limits large-scale applications of PLS in practice. One can
use lossy compression (e.g., PCA [14, 9] and b-bit minwise
hashing [12, 20]) to compactly represent data matrices and
then learn linear models on the compact data matrices [21].
However, although these lossy compression-based methods
effectively reduce memory usage [21, 30], their drawback
is that they cannot extract informative features from the
learned models, because the original data matrices cannot
be recovered from the compressed ones.
Grammar compression [2, 27, 16] is a method of lossless
compression (i.e., the original data can be completely re-
Table 1: Summary of scalable learning methods of linear models.
Approach
Compression Type # of parameters Interpretability Optimization
PCA-SL [14, 9]
Orthogonal rotation
Lossy
2
Limited
Stable
bMH-SL [21]
Hashing
Lossy
3
Unable
Stable
SGD [33, 8]
Sampling
1
Limited
Unstable
cPLS (this study) Grammar compression
Lossless
1
High
Stable
covered from grammar-compressed data) that also has a
wide variety of applications in string processing, such as
pattern matching [37], edit-distance computation [13], and
q-gram mining [1]. Grammar compression builds a small
context-free grammar that generates only the input data
and is very effective at compressing sequences that contain
many repeats. In addition, the set of grammar rules has a
convenient representation as a forest of small binary trees,
which enables us to implement various string operations
without decompression. To date, grammar compression has
been applied only to string (or sequence) data; however,
as we will see, there remains high potential for application
to other data representations. A fingerprint (or bit vector)
is a powerful representation of natural language texts [23],
bio-molecules [32], and images [11]. Grammar compression
is expected to be effective for compressing a set of fingerprints as well, because fingerprints belonging to the same
class share many identical features.
Principal component analysis (PCA) [14] is a widely used
machine learning tool, and is a method of lossy compression,
i.e., the original data cannot be recovered from compressed
data. There have been many attempts to extend PCA [31,
28] and present a scalable PCA in distributed settings for analyzing big data [9]. For classification and regression tasks,
a data matrix is compressed by PCA, and linear models
are learned on the compressed data matrix by a supervised
learning method (SL), which is referred to as PCA-SL. Despite these attempts, PCA and its variants do not look at
the correlation structure between data and output variables
(i.e., class labels/response variables), which results in not
only the inability of feature extractions in PCA but also the
inaccurate predictions by PCA-SL.
Li et al. [21] proposed a compact representation of fingerprints for learning linear models by applying b-bit minwise
hashing (bMH). A d-dimensional fingerprint is conceptually
equivalent to the set si ⊂ {1, ..., d} that contains element i
if and only if the i-th bit in the fingerprint is 1. Li et al.’s
Contribution. In this paper, we present a new scalable
method works as follows. We first pick h random permutalearning algorithm for PLS, which we call lossless compressiontions πi , i = 1, .., h, each of which maps [1, d] to [1, d]. We
based PLS (cPLS), to learn highly-interpretable predictive
then
apply a random permutation π on a set si , compute the
linear models from massive high-dimensional data. A key
minimum element as min(π(si )), and take as a hash value
idea is to convert high-dimensional data with fingerprint repits lowest b bits. Repeating this process h times generates h
resentations into a set of sequences and then build grammar
hash values of b bits each. Expanding these h values into a
rules for representing the sequences in order to compactly
(2b × h)-dimensional fingerprint with exactly h 1’s builds a
store data matrices in memory. To achieve this, we propose
compact representation of the original fingerprint.
a novel grammar-compressed representation of a data maLinear models are learned on the compact fingerprints by
trix capable of supporting row and column access while the
SL,
which is referred to as bMH-SL. Although bMH-SL is
data matrix is in a compressed format. The original data
applicable
to large-scale learning of linear models, bMH is
matrix is grammar-compressed, and then a linear model is
a method of lossy compression and cannot extract features
learned on the compressed data matrix, which allows us to
from linear models learned by SL. Other hashing-based apsignificantly reduce working space. cPLS has the following
proaches have been proposed such as Count-Min sketch [5],
desirable properties:
Vowpal Wabbit [34], and Hash-SVM [25]. However, like
1. Scalability: cPLS is applicable to massive high-dimensional bMH-SL, these algorithms cannot extract features, which
data.
is a serious problem in practical applications.
2. Prediction Accuracy: cPLS can achieve high predicStochastic gradient descent (SGD) [8, 33] is a computation accuracies for both classification and regression.
tionally efficient algorithm for learning linear models on a
3. Usability: cPLS has only one hyper parameter, which
large-scale. SGD samples ν feature vectors from an input
enhances the usability of cPLS.
dataset and computes the gradient vector from the sampled
4. Interpretability: Unlike lossy compression-based methfeature vectors. The weight vector in linear models is upods, cPLS can extract features reflecting the correlation
dated using the gradient vector and the learning rate µ, and
structure between data and class labels/response varithis process is repeated until convergence. Unfortunately
ables.
however, learning linear models using SGD is numerically
We experimentally test cPLS on its ability to learn linear
unstable, resulting in low prediction accuracy. This is bemodels for classification, regression and feature extraction
cause SGD has three parameters (ν, µ, and C) that must be
with various massive high-dimensional data, and show that
optimized if high classification accuracy is to be attained.
cPLS performs superiorly in terms of prediction accuracy,
Online learning is a specific version of SGD that loads an
computational efficiency, and interpretability.
input dataset from the beginning and updates the weight
vector in a linear model for each feature vector. AdaGrad [8]
is an efficient online learning that automatically tunes pa2. LITERATURE REVIEW
rameters of ν and µ in SGD. Although online learning is
Several efficient algorithms have been proposed for learnspace-efficient (owing to its online nature), it is also numering linear models on a large scale. We now briefly review
ically unstable. Even worse, AdaGrad is applicable only
the state of the art, which is also summarized in Table 1.
(i) Input sequence
(ii) Compressed sequence and dictionary . The
dictionary corresponds to a forest of grammar trees.
Figure 1: Illustration of grammar compression.
to differentiable loss functions, which limits its applicability
to simple linear models, e.g., SVM and logistic regression,
making the learned model difficult to interpret.
Despite the importance of scalable learning of interpretable
linear models, no previous work has been able to achieve
high prediction accuracy for classification/regression tasks
and high interpretability of the learned models. We present
a scalable learning algorithm that meets both these demands
and is made possible by learning linear models on grammarcompressed data in the framework of PLS. Details of the
proposed method are presented in the next section.
3.
GRAMMAR COMPRESSION
Given a sequence of integers S, a grammar-compressor
generates a context-free grammar (CFG) that generates S
and only S. The grammar consists of a set of rules1 . Each
rule is of the form Zi → ab. Symbols that appear on the lefthand side of any rule are called non-terminals. The remaining symbols are called terminals, all of which are present in
the input sequence. Informally, a rule Zi → ab indicates
that on the way to recovering the original sequence from
its grammar-compressed representation, occurrences of the
symbol Zi should be replaced by the symbol pair ab (the
resulting sequence may then be subject to yet more replacements). A data structure storing a set of grammar rules
is called a dictionary and is denoted by D. Given a nonterminal, the dictionary supports access to the symbol pair
on the right-hand of the corresponding grammar rule, i.e.,
D[Zi ] returns ab for rule Zi → ab. The original sequence can
be recovered from the compressed sequence and D. The set
of grammar rules in D can be represented as a forest of (possibly small) binary trees called grammar trees, where each
node and its left/right children correspond to a grammar
rule. See Figure 1 for an illustration.
The size of a grammar is measured as the number of rules
plus the size of compressed sequence. The problem of finding
the minimal grammar producing a given string is known to
be NP-complete [2], but several approximation algorithms
exist that produce grammars that are small in practice (see,
e.g., [27, 19, 16]). Among these is the simple and elegant
Re-Pair [19] algorithm, which we review next.
3.1
Re-Pair Algorithm
The Re-Pair grammar compression algorithm by Larsson
and Moffat [19] builds a grammar by repeatedly replacing
the most frequent symbol pair in an integer sequence with a
1
In this paper we assume without loss of generality that the
grammar is in Chomsky Normal Form.
new non-terminal. Each iteration of the algorithm consists
of the following two steps: (i) find the most frequent pair of
symbols in the current sequence, and then (ii) replace the
most frequent pair with a new non-terminal symbol, generating a new grammar rule and a new (and possibly much
shorter) sequence. Steps (i) and (ii) are then applied to the
new sequence and iterated until no pair of adjacent symbols
appears twice.
Apart from the dictionary D that stores the rules as they
are generated, Re-Pair maintains a hash table and a priority queue that together allow the most frequent pair to
be found in each iteration. The hash table, denoted by H,
holds the frequency of each pair of adjacent symbols ab in
the current sequence, i.e., H : ab → N. The priority queue
stores the symbol pairs keyed on frequency and allows the
most frequent symbol to be found in step (i). In step (ii),
a new grammar rule Z1 → ab is generated where ab is the
most frequent symbol pair and Z1 is a new non-terminal not
appearing in a sequence. The rule is stored in the dictionary
D. Every occurrence of ab in the sequence is then replaced
by Z1 , generating a new, shorter sequence. This replacement
will cause the frequency of some symbol pairs to change, so
the hash table and priority queue are then suitably updated.
Let sc denote a sequence generated at c-th iteration in
the Re-Pair algorithm. For input sequence s in Figure 1,
the most frequent pair of symbols is 12. Thus, we generate
rule Z1 → 12 to be added to the dictionary D and replace
all the occurrences of 12 by non-terminal Z1 in s. After four
iterations, the current sequence s4 has no repeated pairs, and
thus the algorithm stops. Dictionary D has four grammar
rules that correspond to a forest of two small trees.
As described by Larsson and Moffat [19], Re-Pair can be
implemented to run in linear time in the length of the input
sequence, but it requires the use of several heavyweight data
structures to track and replace symbol pairs. The overhead
of these data structures (at least 128 bits per position) prevents the algorithm from being applied to long sequences,
such as the large data matrices.
Another problem that arises when applying Re-Pair to
long sequences is the memory required for storing the hash
table: a considerable number of symbol pairs appear twice
in a long sequence, and the hash table stores something for
each of them, consuming large amounts of memory.
In the next section, we present scalable Re-Pair algorithms
that achieve both space-efficiency and fast compression time
on large data matrices. Specifically, our algorithms need
only constant working space.
4.
OUR GRAMMAR-COMPRESSED DATA
MATRIX
Our goal is to obtain a compressed representation of a
data matrix X of n rows and d columns. Let xi denote
the ith row of the matrix represented as a fingerprint (i.e.
binary vector). An alternative view of a row that will be
useful to us is as a sequence of integers si = (p1 , p2 , ..., pm ),
p1 < p2 < · · · < pm , where pi ∈ si if and only if xi [pi ] = 1.
In other words the sequence si indicates the positions of the
1 bits in xi .
In what follows we will deal with a differentially encoded
form of si in which the difference for every pair of adjacent
elements in si is stored, i.e., si = (p1 , p2 , ..., pm ) is encoded
as sgi = (p1 , p2 − p1 , p3 − p2 , ..., pm − pm−1 ). This differ-
ential encoding tends to increase the number of repeated
symbol pairs, which allows the sequences sgi to be more effectively compressed by the Re-Pair algorithm. A grammar
compressor captures the underlying correlation structure of
data matrices: by building the same grammar rules for the
same (sequences of) integers, it effectively compresses data
matrices with many repeated integers.
the q-th interval, H(ab) is initialized as qN/` + 1, which ensures that ab is kept until at least the next interval, i.e., the
(qN/` + 1)-th interval. Algorithm 1 shows the pseudo-code
of lossy counting.
The estimated number of symbol pairs in the hash table is
O(`) [22], resulting in O(` log `) bits consumed by the hash
table.
4.1
Algorithm 1 Lossy counting. H: hash table, N : length of
an input string at a time point, `: length of each interval.
Note that lossy counting can be used in step (i) in the RePair algorithm.
1: Initialize N = 0 and ∆ = 0
2: function LossyCounting(a, b)
3:
N =N +1
4:
if H(ab) 6= 0 then
5:
H(ab) = H(ab) + 1
6:
else
7:
H(ab) = ∆ + 1
6 ∆ then
8:
if b N` c =
9:
∆ = b N` c
10:
for each symbol pair ab in H do
11:
if H(ab) < ∆ then
12:
Remove ab from H
Re-Pair Algorithms in Constant Space
We now present two ideas to make Re-Pair scalable without seriously deteriorating its compression performance. Our
first idea is to modify the Re-Pair algorithm to identify topk frequent symbol pairs in all rows scgi in step (i) and replace
all the occurrences of the top-k symbol pairs in all rows scgi
in step (ii), generating new k grammar rules and new rows
sc+1
gi . This new replacement process improves scalability by
reducing the number of iterations required by roughly a factor of k.
Since we cannot replace both frequent symbol pairs ab and
bc in triples abc in step (ii), we replace the first appearing
symbol pair ab, preferentially. However, such preferential
replacement can generate a replacement of a pair only once
and can add redundant rules to a dictionary, adversely affecting compression performance. To overcome this problem, we replace the first and second appearances of each frequent pair at the same time and replace the next successive
appearance of the frequent pair as usual, which guarantees
generating grammar rules that appear at least twice.
Our second idea is to reduce the memory of the hash table
by removing infrequent symbol pairs. Since our modified
Re-Pair algorithm can work storing compressed sequences
scgi at each iteration c in a secondary storage device, the
hash table consumes most of the memory in execution. Our
modified Re-Pair generates grammar rules from only top-k
frequent symbol pairs in the hash table, which means only
frequent symbol pairs are expected to contribute to the compression. Thus, we remove infrequent symbol pairs from the
hash table by leveraging the idea behind stream mining techniques originally proposed in [15, 6, 22] for finding frequent
items in data stream. Our method is a counter-based algorithm that computes the frequency of each symbol pair
and removes infrequent ones from the hash table at each interval in step (i). We present two Re-Pair algorithms using
lossy counting and frequency counting for removing infrequent symbol pairs from the hash table. We shall refer to
the Re-Pair algorithms using lossy counting and frequency
counting as Lossy-Re-Pair and Freq-Re-Pair, respectively.
4.2
Lossy-Re-Pair
The basic idea of lossy counting is to divide a sequence of
symbols into intervals of fixed length and keep symbol pairs
in successive intervals in accordance with their appearance
frequencies in a hash table. Thus, if a symbol pair has appeared h times in the previous intervals, it is going to be
kept in the next h successive intervals.
Let us suppose a sequence of integers made by concatenating all rows sgi of X and let N be the length of the
sequence. We divide the sequence into intervals of fixedlength `. Thus, the number of intervals is N/`. We use
hash table H for counting the appearance frequency of each
symbol pair in the sequence. If symbol pair ab has count
H(ab), it is ensured that ab is kept in hash table H until the
next H(ab)-th interval. If symbol pair ab first appears in
4.3
Freq-Re-Pair
The basic idea of frequency counting is to place a limit,
v, on the maximum number of symbol pairs in hash table H
and then keep only the most frequent v symbol pairs in H.
Such frequently appearing symbol pairs are candidates to
be replaced by new non-terminals, which generates a small
number of rules.
The hash table counts the appearance frequency for each
symbol pair in step (i) of the Re-Pair algorithm. When the
number of symbol pairs in the hash table reaches v, FreqRe-Pair removes the bottom percent of symbol pairs with
respect to frequency. We call the vacancy rate. Algorithm 2 shows the pseudo-code of frequency counting. The
space consumption of the hash table is O(v log v) bits.
Algorithm 2 Frequency counting. H: hash table, |H|:
number of symbol pairs in H, v: the maximum number of
symbol pairs in H, : vacancy rate. Note that frequency
counting can be used in step (i) in the Re-Pair algorithm.
1: function FrequencyCounting(a, b)
2:
if H(ab) 6= 0 then
3:
H(ab) = H(ab) + 1
4:
else
5:
if |H| ≥ v then
6:
while v(1 − /100) < |H| do
7:
for each symbol pair a0 b0 in H do
8:
H(a0 b0 ) = H(a0 b0 ) − 1
9:
if H(a0 b0 ) = 0 then
10:
Remove a0 b0 from H
11:
H(ab) = 1
5.
DIRECT ACCESS TO ROW AND COLUMN
In this section, we present algorithms for directly accessing
rows and columns of a grammar-compressed data matrix,
which is essential for us to be able to apply PLS on the
compressed matrix in order to learn linear regression models.
5.1
Access to Row
Accessing the i-th row corresponds to recovering the original si from grammar-compressed scgi . We compute this operation by traversing the grammar trees. For recovering the
i-th row si , we start traversing the grammar tree having a
node of the q-th symbol scgi [q] as a root for each q from 1
to |scgi |. Leaves encountered in the traversal must have integers in sequence sgi , which allows us to recover sgi via tree
traversals, starting from the nodes with non-terminal scgi [q]
for each q ∈ [1, |scgi |]. We recover the original i-th row si
from sgi by cumulatively adding integers in sgi from 1 to
|sgi |, i.e, si [1] = sgi [1], si [2] = sgi [2] + si [1],...,si [|sgi |] =
sgi [|sgi |] + si [|sgi | − 1].
5.2
Access to Column
Accessing the j-th column of a grammar-compressed data
matrix requires us to obtain a set of row identifiers R such
that xij = 1 for i ∈ [1, n], i.e., R = {i ∈ [1, n]; xij = 1}.
This operation enables us to compute the transpose X| from
X in compressed format, which is used in the optimization
algorithm of PLS.
P [Zi ] stores a summation of terminal symbols as integers at the leaves under the node corresponding to terminal
symbol Zi in a grammar tree. For example, in Figure 1,
P [Z1 ] = 3, P [Z2 ] = 6, P [Z3 ] = 8 and P [Z4 ] = 4. P can
be implemented as an array that is randomly accessed from
a given non-terminal symbol. We shall refer to P as the
weight array. The size of P depends only on the grammar
size.
The j-th column is accessed to check whether or not xij =
1 in compressed sequence scgi , for each i ∈ [1, n]. We efficiently solve this problem on grammar-compressed data matrix by using the weight array P . Let uq store the summation of weights from the first symbol scgi [1] to the q-th symbol
scgi [q], i.e., uq = P [scgi [1]] + P [scgi [2]] + · · · + P [scgi [q]], and let
u0 = 0. If uq is not less than j, the grammar tree with the
node corresponding to a symbol scgi [q] as a root can encode
j at a leaf. Thus, we traverse the tree in depth-first order from the node corresponding to symbol scgi [q] as follows.
Suppose Z = scgi [q] and u = uq−1 . Let Z` (resptively Zr ) be
a (respectively b) of Z → ab in D. (i) if j < u, we go down
to the left child in the tree; (ii) otherwise, i.e., j ≥ u, we
add P [Z` ] to u and go down to the right child. We continue
the traversal until we reach a leaf. If s = j at a leaf, this
should be xij = 1 at row i; thus we add i to solution set R.
Algorithm 3 shows the pseudo-code for column access.
6.
CPLS
In this section we present our cPLS algorithm for learning PLS on grammar-compressed data matrices. We first
review the PLS algorithm on uncompressed data matrices.
NIPALS [35] is the conventional algorithm for learning PLS
and requires the deflation of the data matrix involved. We
thus present a non-deflation PLS algorithm for learning PLS
on compressed data matrices.
6.1
NIPALS
Let us assume a collection of n data samples and their output variables (x1 , y1 ), (x2 , y2 ), ..., (xn , yn ) where yiP
∈ <. The
output variables are assumed to be centralized as n
i=1 yi =
Algorithm 3 Access to the j-th column on grammarcompressed data matrix. R: solution set of row identifiers i
at column j s.t. xij = 1.
1: function AccessColumn(j)
2:
for i in 1..n do
3:
u0 = 0
4:
for q in 1..|scgi | do
c
5:
uq = uq−1 + P [Sgi
[q]]
6:
if j ≤ uq then
7:
Recursion(i, j, scgi [q], uq−1 )
8:
break
1: function Recursion(i,j,Z,u)
2:
if Z is a terminal symbol then
3:
if u + Z = j then
4:
Add i to R
5:
return
6:
Set Zl (resp. Zr ) as a (resp. b) of Z → ab in D
7:
if u + P [Zl ] > j then
8:
Recursion(i,j,Zl ,u)
. Go to left child
9:
else
10:
Recursion(i,j,Zr ,u + P [Zl ]) . Go to right child
0. Denote by y ∈ <n the vector of all the training output
variables, i.e., y = (y1 , y2 , ..., yn )| .
The regression function of PLS is represented by the following special form,
f (x) =
m
X
αi wi| x,
i=1
where the wi are weight vectors reducing the dimensionality
of x; they satisfy the following orthogonality condition:
1 i=j
wi| X| Xwj =
.
(1)
0 i 6= j
We have two kinds of variables wi and αi to be optimized.
Denote by W ∈ <d×m the weight matrix i-th column of
which is weight vector wi , i.e., W = (w1 , w2 , ..., wm ). Let
α ∈ <m be a vector whose i-th element is αi , i.e., α =
(α1 , α2 , ..., αm )| . Typically, W is first optimized and then α
is determined by minimizing the least squares error without
regularization,
min ||y − XWα||22 .
α
(2)
By computing the derivative of equation (2) with respect to
α and setting it to zero, α is obtained as follows:
α = (W| X| XW)−1 W| X| y.
(3)
The weight vectors are determined by the following greedy
algorithm. The first vector w1 is obtained by maximizing the
squared covariance between the mapped feature Xw and the
output variable y as follows: w1 = argmaxw cov 2 (Xw, y),
subject to w| X| Xw = 1, where cov(Xw, y) = y | Xw. The
problem can be analytically solved as w1 = X| y.
For the i-th weight vector, the same optimization problem
is solved with additional constraints to maintain orthogonality,
wi = argmax cov 2 (Xw, y),
(4)
w
subject to w| X| Xw = 1, w| X| X| wj = 0, j = 1, ..., i − 1.
The optimal solution of this problem cannot be obtained analytically, but NIPALS solves it indirectly. Let us define the
i-th latent vector as ti = Xwi . The optimal latent vectors
ti are obtained first and the corresponding wi is obtained
later. NIPALS performs the deflation of design matrix X
to ensure the orthogonality between latent components ti as
follows, X = X − ti t|i X. Then, the optimal solution has the
form, wi = X| y.
Due to the deflation, X = X − ti t|i X, NIPALS completely
destroys the structure of X. Thus, it cannot be used for
learning PLS on grammar-compressed data matrices.
6.2
cPLS Algorithm
We present a non-deflation PLS algorithm for learning
PLS on grammar-compressed data matrices. Our main idea
here is to avoid deflation by leveraging the connection between NIPALS [35] and the Lanczos method [18] which was
originally proposed for recursive fitting of residuals without
changing the structure of a data matrix.
We define residual ri+1 = (ri − (y | ti−1 )ti−1 ) that is initialized as r1 = y. The i-th weight vector is updated as wi =
X| (ri−1 − (y | ti−1 )ti−1 ), which means wi can be computed
without deflating the original data matrix X. The i-th latent vector is computed as ti = Xwi and is orthogonalized by
applying the Gram-Schmidt orthogonalization to the i-th latent vector ti and previous latent vectors t1 ,t2 ,...,ti−1 as follows, ti = (I−Ti−1 T|i−1 )Xwi , where Ti−1 = (t1 , t2 , ..., ti−1 ) ∈
<n×(i−1) . The non-deflation PLS algorithm updates the
residual ri instead of deflating X, thus enabling us to learn
PLS on grammar-compressed data matrices.
cPLS is the non-deflation PLS algorithm that learns PLS
on grammar-compressed data matrices. The input data matrix is grammar-compressed and then the PLS is learned on
the compressed data matrix by the non-deflation PLS algorithm. Our grammar-compressed data matrix supports
row and column accesses directly on the compressed format
for computing matrix calculations of addition and multiplication, which enables us to learn PLS by using the nondeflation PLS algorithm. Let XG be the grammar-compressed
data matrix of X. Algorithm 4 shows the pseudo-code of
cPLS. Since our grammar-compression is lossless, the cPLS
algorithm on grammar-compressed data matrices learns the
same model as the non-deflation PLS algorithm on uncompressed data matrices and so achieves the same prediction
accuracy.
Algorithm 4 The cPLS algorithm. XG : the grammarcompressed data matrix of X.
1: r1 = y
2: for i = 1, ..., m do
3:
wi = X|G ri
. access to column
4:
if i = 1 then
5:
t1 = XG wi
. access to row
6:
else
7:
ti = (I − Ti−1 T|i−1 )XG wi
. access to row
8:
ti = ti /||ti ||2
9:
ri+1 = ri − (y | ti )ti
10: Compute the coefficients α using equation (3).
We perform feature extraction after line 3 at each iteration in Algorithm 4. The features corresponding to the
top-u largest weights wi are extracted. Due to the orthogonality condition (1), the extracted features give users a novel
insight for analyzing data, which is shown in Section 7.
The cPLS algorithm has three kinds of variables to be
optimized: wi , ri , and ti . The memory for wm is O(md)
and the memory for tm and ri is O(mn). Thus, the total
memory for the variables in cPLS is O(m min(n, d)) highly
depending on parameter m. The parameter m controls the
amount of fitting of the model to the training data and is
typically chosen to optimize the cross validation error. Since
the cPLS algorithm learns the model parameters efficiently,
m can be set to a small value, which results in overall spaceefficiency.
7.
EXPERIMENTS
In this section, we demonstrate the effectiveness of cPLS
with massive datasets. We used five datasets, as shown in
Table 2. ”Book-review” consists of 12,886,488 book reviews
in English from Amazon [24]. We eliminated stop-words
from the reviews and then represented them as 9,253,464
dimensional fingerprints, where each dimension of the fingerprint represents the presence or absence of a word. ”Compound” is a dataset of 42,682 chemical compounds that are
represented as labeled graphs. We enumerated all the subgraphs of at most 10 vertices from the chemical graphs by
using gSpan [38] and then converted each chemical graph
into a 52,099 dimensional fingerprint, where each dimension
of the fingerprint represents the presence or absence of a
chemical substructure. ”Webspam” is a dataset of 16,609,143
fingerprints of 350,000 dimensions2 . ”CP-interaction” is a
dataset of 216,121,626 compound-protein pairs, where each
compound-protein pair is represented as a 3,621,623 dimensional fingerprint and 300,202 compound-protein pairs are
interacting pairs according to the STITCH database [17].
We used the above four datasets for testing the binary classification ability. ”CP-intensity” consists of 1,329,100 compoundprotein pairs represented as 682,475 dimensional fingerprints,
where the information about compound-protein interaction
intensity was obtained from several chemical databases (e.g.,
ChEMBL, BindingDB and PDSP Ki). The intensity was
observed by IC50 (half maximal (50%) inhibitory concentration). We used the ”CP-intensity” dataset for testing the
regression ability. The number of all the nonzero dimensions
in each dataset is summarized in the #nonzero column in
Table 2, and the size for storing fingerprints in memory by
using 32bits for each element is written in the memory column in Table 2. We implemented all the methods by C++
and performed all the experiments on one core of a quad-core
Intel Xeon CPU E5-2680 (2.8GHz). We stopped the execution of each method if it had not finished within 24hours
in the experiments. In the experiments, cPLS did not use
a secondary storage device for compression, i.e., cPLS compressed data matrices by loading all data in memory.
7.1
Compression Ability and Scalability
First, we investigated the influence on compression performance of the top-k parameter in our Re-Pair algorithms. For
this setting, we used the Lossy-Re-Pair algorithm, where parameter ` is set to the total length of all rows in an input data
matrix in order to keep all the symbols in the hash table. We
examined k = {1 × 104 , 2.5 × 104 , 5 × 104 , 7.5 × 104 , 10 × 104 }
for the Book-review, Compound and Webspam datasets and
examined k = {1 × 105 , 2.5 × 105 , 5 × 105 , 7.5 × 105 , 10 × 105 }
for the CP-interaction and CP-intensity datasets.
2
The dataset is downloadable from http://www.csie.ntu.
edu.tw/˜cjlin/libsvmtools/datasets/binary.html.
2e+04
4e+04
6e+04
Top−k
8e+04
2e+04
4e+04
6e+04
Top−k
8e+04
1e+05
2e+04
4e+04
6e+04
Top−k
8e+04
1e+05
1000
●
●
●
●
●
6600 6800 7000 7200 7400 7600 7800
Compression Time (sec)
●
●
Compression Size (MB)
400
600
800
●
200
●
Compression Size (MB)
Compression Time (sec)
0
●
60000
70000
80000
Compression Time (sec)
●
50000
●
●
Compression Size (MB)
4800
5000
5200
5400
●
CP-intensity
Compression Size (MB)
Compression Time (sec)
4600
25000
10000
●
Compression Size (MB)
800
1000
1200
1400
●
Memory (mega bytes)
2,665
3,489
4,977
125,243
110,111
CP-interaction
Compression Size (MB)
Compression Time (sec)
600
●
●
15000
20000
Compression Time (sec)
1000
60000
1e+05
Compression Size (MB)
400
600
800
●
●
200
●
Webspam
Compression Size (MB)
Compression Time (sec)
0
●
●
30000
40000
50000
Compression Time (sec)
●
20000
1000
Compound
Compression Size (MB)
Compression Time (sec)
1200
Compression Size (MB)
1400
1600
1800
2000
Book-review
Table 2: Summary of datasets.
Number Dimension
#nonzeros
12,886,488
9,253,464
698,794,696
42,682 52,099,292
914,667,811
350,000 16,609,143
1,304,697,446
216,121,626
3,621,623 32,831,736,508
1,329,100
682,475 28,865,055,991
10000 15000 20000 25000 30000 35000
Compression Time (sec)
Label type
binary
binary
binary
binary
real
Dataset
Book-review
Compound
Webspam
CP-interaction
CP-intensity
2e+05
4e+05
6e+05
Top−k
8e+05
1e+06
2e+05
4e+05
6e+05
Top−k
8e+05
1e+06
Figure 2: Compression size in mega bytes (MB) and compression time in seconds (sec) for various top-k
Figure 2 shows compression size and compression time
for various top-k. We observed a trade-off between compressed size and compression time for all the datasets. The
smaller the compressed size, the larger the compression time
for larger values of k. In particular, significantly faster compression time was possible at the cost of only slightly worse
compression. For example, Lossy-Re-Pair took 57,290 seconds to compress the Book-review dataset and its size was
1,498 mega bytes (MB) for k=10000. When k=100000, compression time dropped to 20,004 seconds (less than half),
while compressed size increased negligibly to 1,502MB.
The same trends for the Book-review dataset were observed in the other datasets, which suggests that in practice a large value of k can be chosen for fast compression,
without adversely affecting compression performance. Notably, we observed our compression method to be particularly effective for the larger datasets: CP-interaction and
CP-intensity. The original sizes of CP-interaction and CPintensity were 125GB and 110GB, respectively, while the
compressed sizes of CP-interaction and CP-intensity were
at most 5GB and at 535MB, respectively. Our compression
method thus achieved compression rates of 4% and less than
1% for CP-interaction and CP-intensity, respectively. Such
significant reductions in data size enable the PLS algorithm
to scale to massive data.
Next, we evaluated the performance of Lossy-Re-Pair and
Freq-Re-Pair, where parameters `={1MB, 10MB, 100MB,
1000MB} were examined for Lossy-Re-Pair, and parameters
v ={1MB, 10MB, 100MB, 1000MB} and = {30} were examined for Freq-Re-Pair. Table 3 shows the compressed size,
compression time and the working space used for the hash
table in Lossy-Re-Pair and Freq-Re-Pair. We observed that
both Lossy-Re-Pair and Freq-Re-Pair achieved high compression rates using small working space. Such efficiency
is crucial when the goal is to compress huge data matrices
that exceed the size of RAM; our Re-Pair algorithm can
compress data matrices stored in external memory (disk).
For compressing the CP-interaction dataset, Lossy-Re-Pair
and Freq-Re-Pair consumed 16GB and 13GB, respectively,
achieving a compressed size of 5GB. We observed the same
tendency for the other datasets (See Table 3).
7.2
Prediction Accuracy
We evaluated the classification and regression capabilities
of cPLS, PCA-SL, bMH-SL and SGD. Following the previous works [39, 21], we randomly selected 20% of samples for
testing and used the remaining 80% of samples for training.
cPLS has one parameter m, so we selected the best parameter value among m = {10, 20, ..., 100} that achieved the
highest accuracy for each dataset. The PCA phase of PCASL has one parameter deciding the number of principal components m, which was chosen from m = {10, 25, 50, 75, 100}
whose maximum value of 100 is the same as that of cPLS’s
parameter m. Linear models were learned with LIBLINEAR [10], one of the most efficient implementations of linear classifiers, on PCA’s compact feature vectors, where the
hinge loss of linear SVM for classification and the squared error loss for regression were used with L2 -regularization. The
learning process of PCA-SL [14, 9] has one parameter C for
L2 -regularization, which was chosen from C = {10−5 , 10−4 ,
..., 105 }. For PCA-SL [14, 9], we examined all possible combinations of two parameters (m and C) and selected the best
combination achieving the highest accuracy for each dataset.
The hashing process of bMH-SL [21] has two parameters
(the number of hashing values h and the length of bits b), so
we examined all possible combinations of h = {10, 30, 100}
and b = {8, 16}. As in PCA-SL, linear models were learned
with LIBLINEAR [10] on bMH’s compact feature vectors,
where the hinge loss of linear SVM for classification and
the squared error loss for regression were used with L2 regularization. The learning process of bMH-SL [21] has
one parameter C for L2 -regularization, which was chosen
from C = {10−5 , 10−4 , ..., 105 }. For bMH-SL, we examined all possible combinations of three parameters (h, b,
and C) and selected the best combination achieving the
highest accuracy for each dataset. We implemented SGD
on the basis of the AdaGrad algorithm [8] using the logistic loss for classification and the squared error loss for
regression with L2 -regularization. SGD [8] has one parameter C for L2 -regularization, which was also chosen from
C = {10−5 , 10−4 , ..., 105 }. We measured the prediction accuracy by the area under the ROC curve (AUC) for classification and Pearson correlation coefficient (PCC) for regression. Note that AUC and PCC return 1 for perfect inference
in classification/regression, while AUC returns 0.5 for random inference and PCC returns 0 for random inference. We
report the best test accuracy under the above experimental
settings for each method below.
Table 3: Compression size in mega bytes (MB), compression time in seconds (sec), and working space for
hash table (MB) for varying parameter ` in Lossy-Re-Pair and v in Freq-Re-Pair for each dataset.
compression size (MB)
compression time (sec)
working space (MB)
compression size (MB)
compression time (sec)
working space (MB)
compression size (MB)
compression time (sec)
working space (MB)
compression size (MB)
compression time (sec)
working space (MB)
compression size (MB)
compression time (sec)
working space (MB)
Book-review
Lossy-RePair `(MB)
Freq-RePair v(MB)
1
10
100
1000
1
10
100
1000
1836
1685
1502
1501
2021
1816
1680
1501
9904
12654 19125 20004 2956
2355
3165 21256
1113
1931
7988
8603
292
616
3856
6724
Compound
Lossy-RePair `(MB)
Freq-RePair v(MB)
1
10
100
1000
1
10
100
1000
1288
859
825
825
1523
1302
825
825
5096
7053
7787
7946
1362
1587
8111
8207
1113
1926
5030
5030
292
616
3535
3535
Webspam
Lossy-RePair `(MB)
Freq-RePair v(MB)
1
10
100
1000
1
10
100
1000
1427
948
940
940
2328
2089
1050
940
6953
10585 10584 10964 2125
2799
7712 11519
1112
1923
7075
7075
292
616
3856
5539
CP-interaction
Lossy-RePair `(MB)
Freq-RePair v(MB)
10
100
1000 10000
10
100
1000 10000
5199
5139
5036 20307 9529
5136
5136
24hours 55919 44853 43756 24565 39647 47230 48653
9914 16650 16635
616
3856 13796 13796
CP-intensity
Lossy-RePair `(MB)
Freq-RePair v(MB)
10
100
1000 10000
10
100
1000 10000
558
543
540
535
588
535
535
535
8103
6479
6494
6657
5423
5848
5859
5923
1936
3552
3722
3738
616
2477
2477
2477
Table 4: Results of cPLS, PCA-SL, bMH-SL and SGD for various datasets. Dspace: the working space for
storing data matrix (MB), Ospace: the working space for optimization algorithm and Ltime: learning time
(sec).
cPLS
Data
m
Dspace(MB) Ospace(MB) Ltime(sec) AUC/PCC
Book-review
100
1288
15082
21628
0.96
Compound
20
786
7955
1089
0.83
Webspam
60
890
7736
4171
0.99
CP-interaction
40
4367
53885
35880
0.77
CP-intensity
60
472
10683
33969
0.67
PCA-SL
Data
m/C
Dspace(MB) Ospace(MB) Ltime(sec) AUC/PCC
Book-review
100/1
14747
110
6820
0.70
Compound
25/0.1
12
1
6
0.65
Webspam
50/1
200
2
129
0.99
CP-interaction
>24hours
CP-intensity
100/0.1
1521
11
42
0.11
bMH-SL
Data
b/h/C
Dspace(MB) Ospace(MB) Ltime(sec) AUC/PCC
Book-review 100/16/0.01
2457
110
1033
0.95
Compound
30/16/10
2
1
1
0.62
Webspam
30/16/10
20
2
2
0.99
CP-interaction
30/16/0.1
12366
1854
10054
0.77
CP-intensity 100/16/0.1
253
11
45
0.54
SGD
Data
C
Dspace(MB) Ospace(MB) Ltime(sec) AUC/PCC
Book-review
10
1694
57
0.96
Compound
10
9539
83
0.82
Webspam
10
3041
85
0.99
CP-interaction
1
663
3163
0.75
CP-intensity
0.1
124
280
0.04
Component1
Component2
Component3
Component10
Component9
Component8
Component7
Component6
Component5
Component4
Figure 3: Extracted features for the top 10 latent components in the application of cPLS to the Compound
dataset. Each column represents the highly weighted features (chemical substructures) of a latent component.
Table 4 shows the prediction accuracy, working space and
training time of cPLS, PCA-SL, bMH-SL, and SGD. The
working space for the storing data matrix and the working
space needed for optimizations were separately evaluated.
While PCA-SL and bMH-SL significantly reduced the working space for storing data matrices, the classification and
regression accuracies were low. Since PCA-SL and bMH-SL
compress data matrices without looking at the correlation
structure between data and output variables for compressing
data matrices, high classification and regression accuracies
were difficult to achieve.
SGD significantly reduced the working space, since it did
not store data matrices in memory. Classification and regression accuracies of SGD were not high, because of the instability of the optimization algorithm. In addition, SGD is
applicable only to simple linear models, making the learned
model difficult to interpret.
Our proposed cPLS outperformed the other methods (PCASL, bMH-SL, and SGD) in terms of AUC and PCC and significantly reduced the working space. The results showed
cPLS’s efficiency for learning PLS on compressed data matrices while looking at the correlation structure between data
and output variables. Such a useful property enables us to
extract informative features from the learned model.
7.3
Interpretability
Figure 3 shows the top-10 highly-weighted features that
were extracted for each component in the application of
cPLS to the Compound dataset, where one feature corresponds to a compound chemical substructure. It was observed that structurally similar chemical substructures were
extracted together as important features in the same component, and the extracted chemical substructures differed
between components. This observation corresponds to a
unique property of cPLS. Analysing large-scale compound
structure data is of importance in pharmaceutical applications, especially for rational drug design. For example, the
extracted chemical substructures are beneficial for users who
want to identify important chemical fragments involved in
therapeutic drug activities or adverse drug reactions.
8.
CONCLUSIONS AND FUTURE WORK
We presented a scalable algorithm for learning interpretable
linear models — called cPLS — which is applicable to largescale regression and classification tasks. Our method has the
following appealing properties:
1. Scalability: cPLS is applicable to large numbers of
high-dimensional fingerprints (see Sections 7.1 and 7.2).
2. Prediction Accuracy: The optimization of cPLS is
numerically stable, which enables us to achieve high
prediction accuracies (see Section 7.2).
3. Usability: cPLS has only one hyperparameter to be
tuned in cross-validation experiments (see Section 6.2).
4. Interpretability: Unlike lossy compression-based methods, cPLS can extract informative features reflecting
the correlation structure between data and class labels (or response variables), which makes the learned
models easily interpretable (see Section 7.3).
In this study, we applied our proposed grammar compression algorithm to scaling up PLS, but in principle it can be
used for scaling up other machine learning methods or data
mining techniques. An important direction for future work
is therefore the development of scalable learning methods
and data mining techniques based on grammar-compression
techniques. Such extensions will open the door for machine
learning and data mining methods to be applied in various
large-scale data problems in research and industry.
9.
ACKNOWLEDGMENTS
This work was supported by MEXT/JSPS Kakenhi (24700140,
25700004 and 25700029), the JST PRESTO program, the
Program to Disseminate Tenure Tracking System, MEXT
and Kyushu University Interdisciplinary Programs in Education and Projects in Research Development, and the Academy
of Finland via grant 294143.
10.
REFERENCES
[1] P. Bille, P. H. Cording, and I. L. Gortz. Compact
q-gram profiling of compressed strings. In CPM, pages
62–73, 2013.
[2] M Charikar, E. Lehman, D. Liu, R. Panigrahy,
M. Prabhakaran, A. Sahai, and A. Shelat. The
smallest grammar problem. IEEE Transactions on
Information Theory, 51:2554–2576, 2005.
[3] B. Chen, D. Wild, and R. Guha. PubChem as a source
of polypharmacology. JCIM, 49:2044–2055, 2009.
[4] The Uniprot Consortium. The universal protein
resource (uniprot) in 2010. NAR, 38:D142–D148, 2010.
[5] G. Cormode and Muthukrishnan S. An improved data
stream summary: the count-min sketch and its
applications. Journal of Algorithms, 55:58–75, 2005.
[6] D. Demaine, A. López-Ortiz, and I. Munro. Frequency
estimation of internet packet streams with limited
space. In ESA, pages 348–360, 2002.
[7] C.M. Dobson. Chemical space and biology. Nature,
432(7019):824–828, 2004.
[8] J. Duchi, E. Hazan, and Y. Singer. Adaptive
subgradient methods for online learning and stochastic
optimization. JMLR, 12:2121–2159, 2011.
[9] T. Elgamal, M. Yabandeh, A. Aboulnaga,
W. Mustafa, and M. Hefeeda. sPCA: Scalable
principal component analysis for big data on
distributed platforms. In SIGMOD, 2015.
[10] R. Fan, K. W. Chang, C. J. Hsieh, X. R. Wang, and
C. J. Lin. LIBLINEAR: A library for large linear
classification. JMLR, pages 1871–1874, 2008.
[11] D. A. Forsyth and J. Ponce. Computer Vision: A
Modern Approach. Prentice Hall Professional
Technical Reference, 2002.
[12] A. Gionis, P. Indyk, and R. Motwani. Similarity search
in high dimensions via hashing. In VLDB, 1999.
[13] D. Hermelin, D. H. Landau, and O. Weimann. A
unified algorithm for accelerating edit-distance
computation via text-compression. In STACS, pages
529–540, 2009.
[14] I. T. Jolliffe. Principal Component Analysis. Springer,
1986.
[15] R. Karp, S. Shenker, and C. Papadimitriou. A simple
algorithm for finding frequent elements in sets and
bags. TODS, 28:51–55, 2003.
[16] J. C. Kieffer, E. Yang, G. J. Nelson, and P. C.
Cosman. Universal lossless compression via multilevel
pattern matching. IEEE Transactions on Information
Theory, 46(4):1227–1245, 2000.
[17] M. Kuhn, D. Szklarczyk, A. Franceschini,
M. Campillos, C. von Mering, L.J. Jensen, A. Beyer,
and P. Bork. STITCH 2: An interaction network
database for small molecules and proteins. NAR,
38(suppl 1):D552–D556, 2010.
[18] C. Lanczos. An iteration method for the solution of
the eigen value problem of linear differential and
integral operators. Journal of Research of the National
Bureau of Standards, 45:255–282, 1950.
[19] J. Larsson and A. Moffat. Offline dictionary-based
compression. In DCC, pages 296–305, 1999.
[20] P. Li and A. C. König. b-bit minwise hashing. In
WWW, pages 671–680, 2010.
[21] P. Li, A. Shrivastava, J. L. Moore, and A. C. König.
Hashing algorithms for large-scale learning. In NIPS,
pages 2672–2680, 2011.
[22] G. Manku and R. Motwani. Approximate frequency
counts over data stream. In VLDB, volume 5, pages
346–357, 2002.
[23] C. D. Manning and H. Schütze. Foundations of
Statistical Natural Language Processing. The MIT
Press, 1999.
[24] J. McAuley and J. Leskovec. Hidden factors and
hidden topics: understanding rating dimensions with
review text. In RecSys, 2013.
[25] Y. Mu, G. Hua, W. Fan, and S. Chang. Hash-SVM:
Scalable kernel machines for large-scale visual
classification. In CVPR, pages 979–986, 2014.
[26] R. Rosipal and N. Krämer. Overview and recent
advances in partial least squares. In Subspace, Latent
Structure and Feature Selection, LNCS, pages 34–51.
Springer, 2006.
[27] W. Rytter. Application of Lempel-Ziv factorization to
the approximation of grammar-based compression.
TCS, 302(1–3):211–222, 2003.
[28] B. Schölkopf, A. Smola, and K. R. Müller. Nonlinear
component analysis as a kernel eigenvalue problem.
Neural Computation, 10:1299–1319, 1998.
[29] B. Stockwell. Chemical genetics: Ligand-based
discovery of gene function. Nature Reviews Genetics,
1:116–125, 2000.
[30] Y. Tabei and Y. Yamanishi. Scalable prediction of
compound-protein interactions using minwise-hashing.
BMC Systems Biology, 7:S3, 2013.
[31] M. E. Tipping and C. M. Bishop. Mixtures of
probabilistic principal component analysers. Neural
Computation, 11:443–482, 1999.
[32] R. Todeschini and V. Consonni. Handbook of
Molecular Descriptors. Wiley-VCH, 2002.
[33] Y. Tsuruoka, J. Tsujii, and S. Ananiadou. Stochastic
gradient descent training for l1-regularized log-linear
models with cumulative penalty. In ACL and AFNLP,
pages 477–485, 2009.
[34] K. Weinberger, A. Dasgupta, J. Langford, A. Smola,
and J. Attenberg. Feature hashing for large scale
multitask learning. In ICML, pages 1113–1120, 2009.
[35] H. Wold. Path models with latent variables: The
NIPALS approach. In Quantitative Sociology:
International Perspectives on Mathematical and
Statistical Model Building, pages 307–357. Academic
Press, 1975.
[36] S. Wold, M. Sjöström, and L. Eriksson.
PLS-regression: a basic tool of chemometrics.
Chemometrics and Intelligent Laboratory Systems,
58:109–130, 2001.
[37] T. Yamamoto, H. Bannai, S. Inenaga, and M. Takeda.
Faster subsequence and don’t-care pattern maching on
compressed texts. In CPM, pages 309–322, 2011.
[38] X. Yan and J. Han. gSpan: graph-based substructure
pattern mining. In ICDM, pages 721–724, 2002.
[39] H.F. Yu, C.J. Hsieh, K.W. Chang, and C.J. Lin. Large
linear classification when data cannot fit in memory.
In KDD, pages 833–842, 2010.
| 8cs.DS
|
Noname manuscript No.
(will be inserted by the editor)
Stochastic Greedy Algorithms For Multiple Measurement Vectors
arXiv:1711.01521v1 [math.OC] 5 Nov 2017
Jing Qin · Shuang Li · Deanna Needell · Anna Ma ·
Rachel Grotheer · Chenxi Huang · Natalie Durgin
Received: date / Accepted: date
Abstract Sparse representation of a single measurement vector (SMV) has been explored in a variety of
compressive sensing applications. Recently, SMV models have been extended to solve multiple measurement vectors (MMV) problems, where the underlying signal is assumed to have joint sparse structures.
To circumvent the NP-hardness of the `0 minimization problem, many deterministic MMV algorithms
solve the convex relaxed models with limited efficiency. In this paper, we develop stochastic greedy algorithms for solving the joint sparse MMV reconstruction problem. In particular, we propose the MMV
Stochastic Iterative Hard Thresholding (MStoIHT) and MMV Stochastic Gradient Matching Pursuit
(MStoGradMP) algorithms, and we also utilize the mini-batching technique to further improve their performance. Convergence analysis indicates that the proposed algorithms are able to converge faster than
Jing Qin was supported by the faculty start-up fund from Montana State University. Shuang Li was supported by NSF
CAREER grant CCF-1149225. Deanna Needell was partially supported by the Alfred P. Sloan Foundation, NSF CAREER
#1348721, and NSF BIGDATA #1740325.
Jing Qin
Department of Mathematical Sciences
Montana State University
Bozeman, MT 59717.
E-mail: [email protected]
Shuang Li
Department of Electrical Engineering
Colorado School of Mines
Golden, CO 80401.
Deanna Needell
Department of Mathematics
University of California, Los Angeles
Los Angeles, CA 90095.
Anna Ma
Department of Mathematics
Claremont Graduate University
Claremont, CA 91711.
Rachel Grotheer
Center for Data, Mathematical, and Computational Sciences
Goucher College
Baltimore, MD 21204.
Chenxi Huang
Center for Outcomes Research and Evaluation
Yale University
New Haven, CT 06511.
Natalie Durgin
Spiceworks, Austin, TX, 78746.
2
Jing Qin et al.
their SMV counterparts, i.e., concatenated StoIHT and StoGradMP, under certain conditions. Numerical experiments have illustrated the superior effectiveness of the proposed algorithms over their SMV
counterparts.
Keywords Multiple measurement vectors (MMV) · stochastic iterative hard thresholding (StoIHT) ·
stochastic gradient matching pursuit (StoGradMP) · joint sparse signal recovery
PACS 02.70.-c · 02.60.-x · 87.19.le
Mathematics Subject Classification (2010) 65K10 · 94A12 · 94A08
1 Introduction
Reconstruction of sparse signals from limited measurements has been studied extensively with a variety
of applications in various imaging sciences, machine learning, computer vision and so on. The major
problem is to reconstruct a signal which is sparse by itself or in some transformed domain from a few
measurements (or observations) acquired by a certain sensing machine. Let x ∈ Rn be the signal to be
reconstructed. Then the sparse signal reconstruction problem can be formulated as an `0 constrained
minimization problem
minn F (x), s.t. kxk0 ≤ k,
(1)
x∈R
where the sparsity kxk0 counts nonzeros in x. Here F (x) is a loss function measuring the discrepancy
between the acquired measurements y ∈ Rm (m n) and the measurements predicted by the estimated
solution. In particular, if the measurements are linearly related to the underlying signal, i.e., there exists
a sensing matrix A ∈ Rm×n such that y = Ax + n where n is the Gaussian noise, then the least squares
loss function is widely used:
1
2
F (x) = ky − Axk2 .
2
In this case, (1) is a single measurement vector (SMV) sparse signal reconstruction problem. The choice
of F depends on the generation mechanism of the measurements. Since the measurements are typically
generated continuously in most imaging techniques, it becomes significantly important in practice to
reconstruct a collection of sparse signals, expressed as a signal matrix, from multiple measurement vectors
(MMV). More precisely, the signal matrix X ∈ Rn×L with k (k ≤ n) nonzero rows can be obtained by
solving the following MMV model
min F (X),
X∈Rn×L
s.t.
kXkr,0 ≤ k,
(2)
where kXkr,0 stands for the row-sparsity of X which counts nonzero rows in X. Note that it is possible that
certain columns of X have more zero components than zero rows of X. The MMV sparse reconstruction
problem was first introduced in magnetoencephalography (MEG) imaging [1, 2], and has been extended
to other applications [3, 4, 5, 6, 7, 8, 9].
Many SMV algorithms can be applied to solve MMV problems. The most straightforward way is to use
SMV algorithms to reconstruct each signal vector sequentially or simultaneously via parallel computing,
and then concatenate all resultant signals to form the estimated signal matrix. We call these types of
SMV algorithms, concatenated SMV algorithms. On the other hand, the MMV problem can be converted
to an SMV one by columnwise stacking the unknown signal matrix X as a vector and introducing a block
diagonal matrix as the new sensing matrix A. However, both approaches do not fully take advantage of
the joint sparse structure of the underlying signal matrix, and lack computational efficiency as well. In
this paper, we develop MMV algorithms without concatenation of the SMV results or vectorization of
the unknown signal matrix.
Since the `0 term in (1) and (2) is non-convex and non-differentiable, many classical convex optimization algorithms fail to produce a satisfactory solution. To handle the NP-hardness of the problem,
many convex relaxation methods and their MMV extensions have been developed, e.g., the `2 -regularized
Stochastic Greedy Algorithms For Multiple Measurement Vectors
3
M-FOCUSS [1], and the `1 -regularized MMV extensions of the alternating direction method of multipliers [10, 11]. By exploiting the relationship between the measurements and the correct atoms, multiple
signal classification (MUSIC) [12] and its improved variants [13, 14] have been developed. However, in
the rank defective cases when the rank of the measurement matrix Y is much smaller than the desired
row-sparsity level, the MUSIC type of methods will mostly fail to identify the correct atoms. The third
category of algorithms for solving the `0 constrained problem is the class of greedy algorithms that seek
the sparsest solution by updating the support iteratively, including Orthogonal Matching Pursuit (OMP)
[15], simultaneous OMP (S-OMP) [16, 17], Compressive Sampling Matching Pursuit (CoSaMP) [18], Regularized OMP (ROMP) [19], Subspace-Pursuit (SP) [20], and Iterative Hard-Thresholding (IHT) [21].
It has been shown that CoSaMP and IHT are more efficient than the convex relaxation methods with
strong recovery guarantees. However, most of these algorithms work for compressive sensing applications
where F is a least squares loss function. Recently, the Gradient Matching Pursuit (GradMP) [22] has been
proposed to extend CoSaMP to handle more general loss functions. To further improve the computational
efficiency and consider the non-convex objective function case, Stochastic IHT (StoIHT) and Stochastic
GradMP (StoGradMP) have been proposed [23]. Nevertheless, the aforementioned greedy algorithms are
designed for solving the SMV problem and the concatenated extension to the MMV versions will result
in limited performance especially for large data sets. In this paper, we propose the MMV Stochastic IHT
(MStoIHT) and the MMV Stochastic GradMP (MStoGradMP) methods for solving the general MMV
joint sparse recovery problem (2). To accelerate convergence, the mini-batching technique is applied to
the proposed algorithms. We also show that the proposed algorithms converge faster than their SMV
concatenated counterparts under certain conditions. A large variety of numerical experiments on joint
sparse matrix recovery and video sequence recovery have demonstrated the superior performance of the
proposed algorithms over their SMV counterparts in terms of running time and accuracy.
Organization. The rest of the paper is organized as follows. Preliminary knowledge and notation
clarifications are provided in Section 2. Section 3 presents the concatenated SMV algorithms, and the
proposed stochastic greedy algorithms, i.e., MStoIHT and MStoGradMP, in detail. Section 4 discusses
how to apply the mini-batching technique to accelerate the proposed algorithms. Convergence analysis is
provided in Section 5. By choosing the widely used least squares loss function as F , joint sparse signal
recovery in distributed compressive sensing is discussed in Section 6. Extensive numerical results are
shown in Section 7. Finally, some concluding remarks are made in Section 8.
2 Preliminaries
To make the paper self-contained, we first introduce some useful notation and definitions, and then briefly
describe the related algorithms, i.e., StoIHT and StoGradMP. Let [m] = {1, 2, . . . , m} and |Ω| be the
number of elements in the set Ω. Consider a finite atom set D = {d1 , . . . , dN } (a.k.a. the dictionary)
with each atom di ∈ Rn .
2.1 Vector Sparsity
Assume that a vector x ∈ Rn can be written as a linear combination of di ’s, i.e., x =
with
D = d1 · · · dN , α = (α1 , . . . , αN )T .
Then the support of x with respect to α and D is defined by
suppα,D (x) = {i ∈ [N ] : αi 6= 0} := supp(α).
The `0 -norm of x with respect to D is defined as the minimal support size
X
kxk0,D = min{|T | : x =
αi di , T ⊆ [N ]} = min | suppα,D (x)|.
α
i∈T
α
PN
i=1
αi di = Dα
4
Jing Qin et al.
Since absolute homogeneity does not hold in general, i.e., kγxk0,D = |γ| kxk0,D holds if and only if |γ| = 1,
this `0 -norm is not a norm. Here the smallest support suppα,D (x) is called the support of x with respect
to D, denoted by suppD (x). Thus
| suppD (x)| = kxk0,D .
Note that the support may not be unique if D is over-complete in that there could be multiple representations of x with respect to the atom set D due to the linear dependence of the atoms in D. The vector
x is called k-sparse with respect to D if
kxk0,D ≤ k.
2.2 Matrix Sparsity
By extending vector sparsity, we define the row sparsity for a matrix X ∈ Rn×L as follows
kXkr,0,D = min{|Ω| : Ω =
Ω
L
[
suppD (X·,i )},
i=1
where X·,i is the i-th column of X. Here the minimal common support Ω is called the (row-wise) joint
support of X with respect to D, denoted by supprD (X), which satisfies
| supprD (X)| = kXkr,0,D .
The matrix X is called k-row sparse with respect to D if all columns of X share a joint support of size
at most k with respect to D, i.e.,
kXkr,0,D ≤ k.
2.3 Functions Defined on A Matrix Space
Given a function f : Rn×L → R, the matrix derivative is defined by concatenating gradients [24]
∂f
∂f
=
= ∇X·,1 f · · · ∇X·,L f ,
∂X
∂Xi,j n×L
(3)
where Xi,j is the (i, j)-th entry of X. Notice that
2
kXkF =
n
X
2
kXi,· k2 =
L
X
2
kX·,j k2 = Tr(X T X),
j=1
i=1
where Xi,· is the i-th row vector of X, and Tr(·) is the trace operator to add up all the diagonal entries
of a matrix. The inner product for any two matrices U, V ∈ Rn×L is defined as
hU, V i = Tr(U T V ).
Note that the equality
2
2
2
kU + V kF = kU kF + kV kF + 2hU, V i
(4)
and the Cauchy-Schwartz inequality
hU, V i ≤ kU kF kV kF
(5)
still hold. By generalizing the concepts in [23], we define the D-restricted strong convexity property and
the strong smoothness property (a.k.a. the Lipschitz condition on the gradient) for the functions defined
on a matrix space.
Stochastic Greedy Algorithms For Multiple Measurement Vectors
5
Definition 1 The function f : Rn×L → R satisfies the D-restricted strong convexity (D-RSC) if there
exists ρ−
k > 0 such that
f (X 0 ) − f (X) −
E ρ−
2
(X), X 0 − X ≥ k kX 0 − XkF
∂X
2
D ∂f
(6)
for all matrices X 0 , X ∈ Rn×L with | supprD (X) ∪ supprD (X 0 )| ≤ k.
Definition 2 The function f : Rn×L → R satisfies the D-restricted strong smoothness (D-RSS) if there
exists ρ+
k > 0 such that
∂f
∂f
0
(X) −
(X 0 ) ≤ ρ+
(7)
k kX − X kF
∂X
∂X
F
for all matrices X 0 , X ∈ Rn×L with | supprD (X) ∪ supprD (X 0 )| ≤ k.
2.4 Related Work
StoIHT (see Algorithm 1) and StoGradMP (see Algorithm 2) have been proposed to solve the `0 constrained SMV problem [23]
minn
x∈R
M
1 X˜
fi (x),
M i=1
subject to
kxk0,D ≤ k.
(8)
At each iteration of StoIHT, one component function f˜i is first randomly selected with probability p(i).
Here the input discrete probability distribution p(i)’s satisfy
M
X
and p(i) ≥ 0, i = 1, . . . , M.
p(i) = 1,
i=1
Next in the “Proxy” step, gradient descent along the selected component is performed. Then the last two
steps, i.e., “Identify” and “Estimate”, essentially project the gradient descent result to its best k-sparse
approximation. Given w = (w1 , . . . , wn )T and η ≥ 1, the best k-sparse approximation operator acted on
w and η, denoted by approxk (w, η), constructs an index set Γ with |Γ | = k such that
kPΓ w − wk2 ≤ η w − w(k)
2
where
w(k) = argmin kw − yk2 .
y∈R(DΓ )
|Γ |≤k
Here PΓ w is the orthogonal projection of w onto the subspace R(DΓ ) in Rn spanned by the atoms with
indices in Γ , and w(k) is the best k-sparse approximation of w in the subspace R(DΓ ). In particular, if
η ≥ 1 and D = {ei : i = 1, 2, . . . , n} with ei = [0, . . . , 1 , . . . , 0]T , then approxk (w, η) returns the index
(i)
set of the first k largest entries of w in absolute value, i.e.,
approxk (w, 1) = {i1 , i2 , . . . , ik : |wi1 | ≥ . . . ≥ |wik | ≥ . . . ≥ |win |} := Γb.
Then the projection PΓ w reads as in componentwise form
(
wj if j ∈ Γb,
PΓb (w) j =
0
if j ∈
/ Γb.
There are two widely used stopping criteria:
xt+1 − xt
kxt k2
2
< ε,
and
M
1 X˜ t
fi (x ) < ε,
M i=1
6
Jing Qin et al.
Algorithm 1 Stochastic Iterative Hard Thresholding (StoIHT)
Input: k, γ, η, p(i), ε.
b = xt .
Output: x
Initialize: x = 0.
for t = 1, 2, . . . , T do
Randomly select an index it ∈ {1, 2, . . . , M } with probability p(it )
γ
∇f˜it (xt )
Proxy: bt = xt − M p(i
)
t
Identify: Γ t = approxk (bt , η).
Estimate: xt+1 = PΓ t (bt ).
If the stopping criteria are met, exit.
end for
where ε > 0 is a small tolerance. It is well known that the first stopping criteria is more robust in practice.
Different from StoIHT, StoGradMP involves the gradient matching process, i.e., to find the best k-sparse
approximation of the gradient rather than the estimated solution. At the solution estimation step, the
original problem is restricted to the components from the estimated support. It has been empirically shown
that StoGradMP converges faster than StoIHT due to the more accurate estimation of the support. But
StoGradMP requires that the sparsity level k is no more than n/2.
Algorithm 2 Stochastic Gradient Matching Pursuit (StoGradMP)
Input: k, η1 , η2 , p(i), ε.
b = xt .
Output: x
Initialize: x0 = 0, Λ = 0.
for t = 1, 2, . . . , T do
Randomly select an index it ∈ {1, 2, . . . , M } with probability p(it )
Calculate the gradient rt = ∇f˜it (xt )
Γ = approx2k (rt , η1 )
Γb = Γ ∪ Λ
1 PM ˜
bt = argminx M
x ∈ R(DΓ̂ )
i=1 fi (x),
Λ = approxk (bt , η2 )
xt+1 = PΛ (bt )
If the stopping criteria are met, exit.
end for
3 Proposed Stochastic Greedy Algorithms
In this section, we present concatenated SMV algorithms, and develop stochastic greedy algorithms for
MMV problems based on StoIHT and StoGradMP. Suppose that there are M differentiable and convex
functions fi : Rn×L → R that satisfy the D-restricted strong smoothness property (see Definition 2), and
their mean
M
1 X
fi (X)
(9)
F (X) =
M i=1
satisfies the D-restricted strong convexity property (see Definition 1). These assumptions will be used
extensively throughout the entire paper. Consider the following row-sparsity constrained MMV problem
min
X∈Rn×L
M
1 X
fi (X),
M i=1
subject to
kXkr,0,D ≤ k.
(10)
By vectorizing X, i.e., rewriting X as a vector x ∈ RnL by columnwise stacking, we can relax (10) to
a sparsity constrained SMV problem of the form (8) where the sparsity level k is replaced by kL. Since
kxk0,D ≤ kL does not necessarily guarantee kXkr,0,D ≤ k, the solution to the relaxed problem may
not be the same as the vectorization of the solution to (10). On the other hand, the iterative stochastic
Stochastic Greedy Algorithms For Multiple Measurement Vectors
7
algorithms such as StoIHT and StoGradMP, can be developed to the concatenated versions for solving
(10) under the following assumption on the objective function fi ’s.
Assumption 1. The objective function in (10) is separable, in the sense that a collection of functions
gi,j : Rn → R exist with
fi (X) =
L
X
gi,j (X·,j ),
i = 1, . . . , M.
(11)
j=1
Under this assumption, the concatenated algorithms, i.e., CStoIHT in Algorithm 3 and CStoGradMP
in Algorithm 4, can be applied to solve (10), which essentially reconstruct each column of X by solving
the SMV problem (8). Notice that the outer loops of CStoIHT and CStoGradMP can be executed in
a parallel manner on a multi-core computer, when the order of the inner loop and the outer loop in
Algorithm 3 can be swapped. However, if the sparsity level k is very large, then the support sets of
X·,j ’s are prone to overlap less initially which results in the less accurate estimation of the joint support
and larger errors in the initial iterates. In addition, for some nonlinear function fi (X) which can not be
separated as a sum of functions for columns of X, it will be challenging to find an appropriate objective
function for the corresponding SMV problem.
Algorithm 3 Concatenated Stochastic Iterative Hard Thresholding (CStoIHT)
Input: k, γ, η, p(i), ε.
b = Xt.
Output: X
Initialize: X = 0 ∈ Rn×L .
for j = 1, . . . , L do
for t = 1, 2, . . . , T do
Randomly select an index it ∈ {1, 2, . . . , M } with probability p(it )
γ
t −
t )
Proxy: bt = X·,j
∇git ,j (X·,j
M p(i )
t
Identify: Γ t = approxk (bt , η).
t+1
Estimate: X·,j
= PΓ t (bt ).
If the stopping criteria are met, exit.
end for
end for
Algorithm 4 Concatenated Stochastic Gradient Matching Pursuit (CStoGradMP)
Input: k, η1 , η2 , p(i), ε.
b = Xt.
Output: X
Initialize: X 0 = 0 ∈ Rn×L , Λ = 0.
for j = 1, . . . , L do
for t = 1, 2, . . . , T do
Randomly select an index it ∈ {1, 2, . . . , M } with probability p(it )
t )
Calculate the gradient rt = ∇git ,j (X·,j
t
Γ = approx2k (r , η1 )
Γb = Γ ∪ Λ
1 PM
bt = argminx M
x ∈ R(DΓ̂ )
i=1 gi,j (x),
Λ = approxk (bt , η2 )
t+1
X·,j
= PΛ (bt )
If the stopping criteria are met, exit.
end for
end for
To circumvent the aforementioned issues, we first propose the MMV Stochastic Iterative Hard Thresholding algorithm (MStoIHT) detailed in Algorithm 5. Compared to StoIHT, MStoIHT replaces the gradient by the matrix derivative (3). The second significant difference lies in the “Identify” and “Estimate”
steps, especially the operator approxk (·, ·). Now we extend the operator approxk (·, ·) from sparse vectors
8
Jing Qin et al.
to row-sparse matrices. Given X ∈ Rn×L and η ≥ 1, the best k-row sparse approximation operator acted
on X and η, denoted by approxrk (X, η), constructs a row index set Γ such that
kPΓ X·,j − X·,j k2 ≤ η kX·,j − (X·,j )k k2 ,
j = 1, . . . , L,
where (X·,j )k is the best k-sparse approximation of the column vector X·,j with respect to D. In particular,
if D = {ei : i = 1, . . . , n} and η = 1, approxrk (X, 1) returns the row index set of the first k largest `2 row
norms in X, i.e.,
approxrk (X, 1) = {i1 , i2 , . . . , ik : kXi1 ,· k2 ≥ kXi2 ,· k2 ≥ . . . ≥ kXin ,· k2 } := Γe.
By abusing the notation, we define PΓe (X) to be the projection of X onto the subspace of all row-sparse
matrices with row indices restricted to Γe. Therefore, we have
PΓe X = PΓe X·,1 PΓe X·,2 . . . PΓe X·,L .
Due to the common support Γe, the projection PΓe (X) can be written as in row-wise form
(
Xj,· if j ∈ Γe,
PΓe (X) j,· =
0
if j ∈
/ Γe.
Here PΓe (X) returns a k-row sparse matrix, whose nonzero rows correspond to the k rows of X with
largest `2 row norms.
Algorithm 5 MMV Stochastic Iterative Hard Thresholding (MStoIHT)
Input: k, γ, η, p(i), ε.
b = Xt.
Output: X
Initialize: X = 0.
for t = 1, 2, . . . , T do
Randomly select an index it ∈ {1, 2, . . . , M } with probability p(it )
∂fit (X t )
γ
M p(it )
∂X
Identify: Γ t = approxrk (B t , η).
Estimate: X t+1 = PΓ t (B t ).
Proxy: B t = X t −
If the stopping criteria are met, exit.
end for
Next, we propose the MMV Stochastic Gradient Matching Pursuit (MStoGradMP) detailed in Algorithm 6, where the two gradient matching steps involve the operator approxrk (·, ·). The stopping criteria
in all proposed algorithms can be set as the same as those in Algorithm 1 and Algorithm 2.
Algorithm 6 MMV Stochastic Gradient Matching Pursuit (MStoGradMP)
Input: k, η1 , η2 , p(i), ε.
b = Xt.
Output: X
Initialize: X 0 = 0, Λ = 0.
for t = 1, 2, . . . , T do
Randomly select an index it ∈ {1, 2, . . . , M } with probability p(it )
Calculate the generalized gradient Rt =
Γ = approxr2k (Rt , η1 )
Γb = Γ ∪ Λ
B t = argminX F (X), X ∈ R(DΓ̂ )
Λ = approxrk (B t , η2 )
X t+1 = PΛ (B t )
If the stopping criteria are met, exit.
end for
∂fit (X t )
∂X
Stochastic Greedy Algorithms For Multiple Measurement Vectors
9
4 Batched Acceleration
To accelerate computations and improve performance, we apply the mini-batching technique to obtain
batched variants of Algorithms 5 and 6. We first partition the index set {1, 2, . . . , M } into a collection of
equal-sized batches τ1 , . . . , τd where the batch size |τi | = b for all i = 1, 2, . . . , dM/be := d. For simplicity,
we assume that d is an integer. Similar to the approach in [25], we reformulate (9) as
d
1 X 1 X
fj (X) .
F (X) =
d i=1 b j∈τ
(12)
i
Based on this new formulation, we get the batched version of Algorithm 5, which is termed as BMStoIHT,
described in Algorithm 7. Here the input probability p(i) satisfies
d
1X
p(i) = 1
d i=1
and p(i) ≥ 0, i = 1, . . . , d.
Likewise, we get a batched version of MStoGradMP, termed as BMStoGradMP, which is detailed in
Algorithm 8. It is empirically shown in Section 7 that the increase of the batch size greatly speeds up the
convergence of both algorithms, which is more obvious in BMStoIHT. As a by-product, mini-batching
can also improve the recovery accuracy. However, there is a trade-off between the batch size and the
performance improvement [25].
Algorithm 7 Batched MMV Stochastic Iterative Hard Thresholding (BMStoIHT)
Input: k, γ, η, b and p(i).
b = Xt.
Output: X
Initialize: X = 0.
for t = 1, 2, . . . , T do
Randomly select a batch index τt ⊆ {1, 2, . . . , d} of size b with probability p(τt )
∂fτt (X t )
γ
dp(τt )
∂X
Identify Γ t = approxrk (B t , η).
Estimate X t+1 = PΓ t (B t ).
Proxy: B t = X t −
If the stopping criteria are met, exit.
end for
Algorithm 8 Batched MMV Stochastic Gradient Matching Pursuit (BMStoGradMP)
Input: k, η1 , η2 , b and p(i).
b = Xt.
Output: X
Initialize: X 0 = 0, Λ = 0.
for t = 1, 2, . . . , T do
Randomly select a batch index τt ⊆ {1, 2, . . . , d} of size b with probability p(τt )
Calculate the generalized gradient Rt =
Γ = approxr2k (Rt , η1 )
Γb = Γ ∪ Λ
B t = argminX F (X), X ∈ span(DΓ̂ )
Λ = approxrk (B t , η2 )
X t+1 = PΛ (B t )
If the stopping criteria are met, exit.
end for
∂fτt (X t )
∂X
10
Jing Qin et al.
5 Convergence Analysis
In this section, we provide the convergence guarantees for the proposed MStoIHT and MStoGradMP,
together with their SMV counterparts, i.e., CStoIHT and CStoGradMP. To simplify the discussion, the
result at the t-th iteration of CStoIHT/CStoGradMP refers to the result obtained after t inner iterations and L outer iterations of Algorithm 3/Algorithm 4, or equivalently the maximum number of inner
iterations is set as t. Furthermore, all convergence results can be extended to their batched versions,
i.e., BStoIHT and BMStoGradMP. Comparison of contraction coefficients shows that the proposed algorithms have faster convergence under the D-RSC, D-RSS and separability of the objective function (see
Assumption 1).
By replacing the `2 -norm and vector inner product with the Frobenius norm and the matrix inner product respectively and using (4) and (5), we get similar convergence results for MStoIHT and
MStoGradMP as in [23]:
Theorem 1 Let X ∗ be a feasible solution of (10) and X 0 be the initial solution. At the (t+1)-th iteration
of Algorithm 5, the expectation of the recovery error is bounded by
E X t+1 − X ∗
where
F
≤ κt+1 X 0 − X ∗
F
σX ∗
,
1−κ
+
(13)
q
q
−
−
κ = 2 1 − γ(2 − γα3k )ρ3k + (η 2 − 1)(1 + γ 2 α3k ρ̄+
3k − 2γρ3k ),
αk = max
1≤i≤M
ρ+
k (i)
,
M p(i)
+
ρ+
k = max ρk (i),
ρ̄+
k =
1≤i≤M
(14)
M
1 X +
ρ (i).
M i=1 k
Thus Algorithm 7 converges linearly. In particular, if η = γ = 1, then
q
−
κ = 2 1 − 2ρ−
3k + α3k ρ3k .
(15)
To solve the problem (10), CStoIHT uses StoIHT to restore each column of X separately and then
concatenate all column vectors to form a matrix. To analyze the convergence of CStoIHT, we first derive
∗ 2
an upper bound for E X·,j − X·,j
following the proof in [23, Section 8.2]. Note that we look for the
2
∗
contraction coefficient of E X·,j − X·,j
2
2
∗
rather than E X·,j − X·,j
2
.
Lemma 1 Let X ∗ be a feasible solution of (10) and X 0 be the initial solution. Under Assumption 1, there
exist κj , σj > 0 such that the expectation of the recovery error squared at the t-th iteration of Algorithm 3
for estimating the j-th column of X ∗ is bounded by
t+1
∗
EIt X·,j
− X·,j
2
2
0
∗
≤ κt+1
X·,j
− X·,j
j
2
2
+
σj
,
1 − κj
(16)
t
∗
0
where X·,j
is the approximation of X·,j
at the t-th iteration of StoIHT with the initial guess X·,j
, i.e., the
result at the t-th inner iteration and j-th outer iteration of Algorithm 3 with the initial guess X 0 . Here
It is the set of all indices i1 , . . . , it randomly selected at or before the t-th step of the algorithm.
Proof Due to the separable form of fi in Assumption 1, we consider L problems of the form
min
w
M
1 X
gi,j (w),
M i=1
kwk0,D ≤ k,
j = 1, . . . , L,
(17)
∗
where gi,j are given in (11). This relaxation is also valid since X·,j
is also a feasible solution of (17) if X ∗
t
∗
is a feasible solution of (10). Let wt = X·,j
, w∗ = X·,j
, x = wt+1 − w∗ 2 ,
u = wt − w∗ −
γ
PΩ (∇git ,j (wt ) − ∇git ,j (w∗ ))
M p(it )
+
2
γ
PΩ ∇git ,j (w∗ )
M p(it )
:= u1 + u2 ,
2
Stochastic Greedy Algorithms For Multiple Measurement Vectors
11
and
v = (η 2 − 1) wt − w∗ −
γ
∇git ,j (wt )
M p(it )
2
.
2
Let
v1 = (η 2 − 1) wt − w∗ −
v2 = (η 2 − 1)
γ
(∇git ,j (wt ) − ∇git ,j (w∗ ))
M p(it )
γ
∇git ,j (w∗ )
M p(it )
2
,
2
2
.
2
The inequality (a + b)2 ≤ 2a2 + 2b2 yields that
v ≤ 2v1 + 2v2 .
By the direct computation, we can show that
x2 − 2ux − v ≤ 0,
which implies that
Thus the inequality a + b ≤
√
x≤u+
p
u2 + v.
2a2 + 2b2 yields
x2 ≤ 2u2 + 2(u2 + v) = 4u2 + 2v ≤ 8(u21 + u22 ) + 4v1 + 4v2 .
By taking the conditional expectation Eit |It−1 on both sides, we obtain
Eit |It−1 x2 ≤ 8Eit |It−1 u21 + 8Eit |It−1 u22 + 4Eit |It−1 v1 + 4Eit |It−1 v2
8γ 2
2
Ei |I kPΩ ∇git ,j (w∗ )k2
minit M 2 (p(it ))2 t t−1
4γ 2 (η 2 − 1)
2
−
t
∗ 2
+ 4(η 2 − 1)(1 + γ 2 α3k,j ρ̄+
Ei k∇git ,j (w∗ )k2
3k,j − 2γρ3k,j ) w − w 2 +
minit M 2 (p(it ))2 t
2
+
−
2
2
wt − w∗ 2
= 8(1 − (2γ − γ 2 α3k,j )ρ−
3k,j ) + 4(η − 1)(1 + γ α3k,j ρ̄3k,j − 2γρ3k,j )
4γ 2
∗ 2
2
∗ 2
+
2E
kP
∇g
(w
)k
+
(η
−
1)E
k∇g
(w
)k
i
Ω
i
,j
i
i
,j
t
t
t
t
2
2
minit M 2 (p(it ))2
t
∗
≤ 8(1 − (2γ − γ 2 α3k,j )ρ−
3k,j ) w − w
:= κj wt − w∗
2
2
2
2
+
+ σj .
+
Note that all coefficients α3k,j , ρ−
3k,j and ρ̄3k,j depend on the function gij in (11). Therefore
EIt wt+1 − w∗
2
2
≤ κj EIt−1 wt − w∗
2
2
+ σj .
which implies that
t+1
∗
E X·,j
− X·,j
2
2
0
∗
≤ κt+1
X·,j
− X·,j
j
2
2
+
σj
,
1 − κj
j = 1, . . . , L.
Here the contraction coefficient is
+
−
2
2
κj = 8(1 − (2γ − γ 2 α3k,j )ρ−
3k,j ) + 4(η − 1)(1 + γ α3k,j ρ̄3k,j − 2γρ3k,j ),
(18)
and the tolerance parameter is
σj =
4γ 2
2
2
2Eit kPΩ ∇git ,j (w∗ )k2 + (η 2 − 1)Eit k∇git ,j (w∗ )k2 .
2
2
min M (p(it ))
1≤i≤M
In particular, if γ = η = 1 and p(i) = 1/M , then
−
κj = 8(1 − 2ρ−
3k,j + α3k,j ρ3k,j ),
σj =
M
8 X
∗
PΩ ∇gi,j (X·,j
)
M i=1
2
2
.
(19)
12
Jing Qin et al.
Theorem 2 Let X ∗ be a feasible solution of (10) and X 0 be the initial solution. Under Assumption 1,
at the (t + 1)-th iteration of Algorithm 3, the expectation of the recovery error is bounded by
E X t+1 − X ∗
F
≤κ
bt+1 X 0 − X ∗
F
+ δ,
(20)
0
where X t is the approximation of X ∗ at the t-th iteration of Algorithm 3 with the initial guess X·,j
. Here
the contraction coefficient κ
b and the tolerance parameter δ are defined as
v P
u
L
u
q
j=1 σj
t
κ
b=
max κj , δ =
,
1≤j≤L
1 − max κj
1≤j≤L
where κj is the contraction coefficient for each StoIHT defined in (18).
t+1
0
Proof For each j = 1, 2, . . . , L, StoIHT with the initial guess X·,j
generates X·,j
after t iterations, i.e.,
the result at the (t + 1)-th inner iteration and j-th outer iteration of Algorithm 3. Then the expectation
of the recovery error squared is bounded by
2
t+1
∗
E X·,j
− X·,j
2
0
∗
≤ κt+1
X·,j
− X·,j
j
2
2
+
σj
,
1 − κj
where κj and σj are defined in Lemma 1. Note that κj depends only on the constants in the D-RSC
and D-RSS properties of the objective function or its component function gi,j while σj depends on the
∗
feasible solution X·,j
. By combining all L components of X t+1 , we get
q
2
E X t+1 − X ∗ F ≤ E kX t+1 − X ∗ kF
v
u L
uX
t+1
∗ 2
=t
E X·,j
− X·,j
2
j=1
v
u L
uX t+1
0 − X∗
κj
X·,j
≤t
·,j
j=1
σj
+
2
1 − κj
2
s
( max κj )t+1 kX 0 −
≤
1≤j≤L
t+1
≤
q
max κj
1≤j≤L
2
X ∗ kF
X0 − X∗
PL
+
F
j=1
σj
1 − max1≤j≤L κj
v P
u
L
u
j=1 σj
+t
1 − max κj
1≤j≤L
t+1
:= κ
b
0
X −X
∗
F
+ δ.
In particular, if γ = η = 1, then
r
√ r
q
−
−
max κj =
max 8(1 − 2ρ−
+
α
ρ
)
=
2
2 max (1 − 2ρ−
κ
b=
3k,j 3k,j
3k,j
3k,j + α3k,j ρ3k,j ).
1≤j≤L
1≤j≤L
1≤j≤L
+
In the case that ρ−
3k,j = ρ3k , α3k,j = α3k and ρ3k,j = ρ3k , we have
√
κ
b = 2κ,
where κ is defined Theorem 1. This shows that the CStoIHT converges slower than the proposed MStoIHT
provided the same coefficients in the strong smoothness and convexity of the objective function, which
we show holds for the distributed compressive sensing setting in Section 6.
By using the same proof techniques as in Theorem 1, we can get the following convergence result for
MStoGradMP.
Stochastic Greedy Algorithms For Multiple Measurement Vectors
13
Theorem 3 Let X ∗ be a feasible solution of (10) and X 0 be the initial solution. At the (t+1)-th iteration
of Algorithm 6, the expectation of the recovery error is bounded by
σX ∗
,
(21)
E X t+1 − X ∗ F ≤ κt+1 X 0 − X ∗ F +
1−κ
where
s
!
p
2
ρ+
η12 − 1
4k (2η1 − 1)
M p(i)
−1+
,
max
2
1≤i≤M
η1
ρ−
4k η2
r
α4k
∂fi
(1 + η2 )
(X ∗ )
2 max M p(i)
max PΩ
− +3
|Ω|≤4k
1≤i≤M
∂X
ρ4k
min M p(i)
1≤i≤M
r
α4k
κ = (1 + η2 )
ρ−
4k
σX ∗ =
ρ−
4k
p
1≤i≤M
.
F
Thus Algorithm 6 converges linearly. In particular, if η1 = η2 = 1 and p(i) = 1/M , then
q
−
2 α4k (ρ+
4k − ρ4k )
κ=
.
ρ−
4k
(22)
Similar to CStoIHT, we start the convergence analysis for CStoGradMP by finding the contraction
coefficient for the expectation of recovery error squared at each iteration of CStoGradMP.
Lemma 2 Let X ∗ be a feasible solution of (12) and X 0 be the initial solution. Under Assumption 1,
the expectation of the recovery error squared at the t-th iteration of Algorithm 4 for estimating the j-th
column of X ∗ is bounded by
∗
EIt bt − X·,j
where
β1 =
α4k
,
− α4k
2ρ−
4k
δ1 =
2
2
∗
≤ β1 EIt PΓ̂ (bt − X·,j
)
2
α4k (2ρ−
4k
− α4k ) min M 2 (p(i))2
2
2
+ δ1 ,
(23)
EIt Ei PΓb ∇gi,j (X·,j )
2
2
.
(24)
1≤i≤M
Here It is the set of all indices i1 , . . . , it randomly selected at or before the t-th step of the algorithm.
Proof Consider the problem (17). By the proof in [23, Lemma 2], we get
∗
PΓb (bt − X·,j
)
2
2
t
∗
≤ 2(1 − (2γ − γ 2 α4k )ρ−
4k b − X·,j
2
2
+
2γ 2
Ei PΓb ∇gi,j (X·,j )
min M 2 (p(i))2
2
2
,
1≤i≤M
where we use the inequality (a + b)2 ≤ 2a2 + 2b2 for a, b ∈ R and the expectation inequality (EX)2 ≤
E(X 2 ). Then we have
∗
bt − X·,j
2
2
∗
= PΓb (bt − X·,j
)
2
2
∗
)
+ PΓbc (bt − X·,j
t
∗
≤ 2(1 − (2γ − γ 2 α4k )ρ−
4k ) b − X·,j
2
2
+
2
2
2γ 2
∗
Ei PΓb ∇gi,j (X·,j
)
min M 2 (p(i))2
2
2
∗
+ PΓbc (bt − X·,j
)
2
2
.
1≤i≤M
Moving the first term on the right hand side to the left hand side leads to
∗
bt − X·,j
2
2
≤
2γ 2
Ei PΓb ∇gi,j (X·,j )
φ min M 2 (p(i))2
1≤i≤M
2
2
+
1
∗
PΓbc (bt − X·,j
)
φ
2
2
,
−
2
where φ = 2ρ−
4k (2γ − γ α4k ) − 1. Maximizing φ with respect to γ yields γ = 1/α4k and φmax = (2ρ4k −
α4k )/α4k . By choosing the optimal value of γ and taking the expectation with respect to It on the both
sides of the above inequality, we get (23).
p
Similarly, using the inequality EX ≤ E(X 2 ) and the fact that a ≤ b + c yields a2 ≤ 2b2 + 2c2 ,
we are able to get the following result, which is different from Lemma 3 in [23] in that we consider the
expectation for the `2 -norm squared here rather than that for the `2 -norm.
14
Jing Qin et al.
Lemma 3 Let X ∗ be a feasible solution of (12) and X 0 be the initial solution. Under Assumption 1,
the expectation of the recovery error squared at the t-th iteration of Algorithm 4 for estimating the j-th
column of X ∗ is bounded by
∗
Eit PΓbc (bt − X·,j
)
2
2
2
2
t
∗
≤ β2 X·,j
− X·,j
+ δ2 ,
(25)
where it is the index randomly selected at the t-th iteration of the CStoGradMP and
2 −
(2η12 − 1)ρ+
2(η12 − 1)
4k − η1 ρ4k
+
,
−
2
1≤i≤M
η12
η1 ρ4k
2
max p(i)
1≤i≤M
2
∗
max PΩ ∇gi,j (X·,j
δ2 = 8 −
) 2.
|Ω|≤4k
ρ4k min p(i)
1≤i≤M
β2 = 4 max M p(i)
(26)
1≤i≤M
Theorem 4 Let X ∗ be a feasible solution of (12) and X 0 be the initial solution. Under Assumption 1,
at the (t + 1)-th iteration of Algorithm 4, there exist κ
e, η > 0 such that the expectation of the recovery
error is bounded by
et+1 X 0 − X ∗ F + η,
(27)
E X t+1 − X ∗ F ≤ κ
t
∗
0
where X·,j
is the approximation of X·,j
at the t-th iteration of CStoGradMP with the initial guess X·,j
.
Proof At the t-th iteration of CStoGradMP, i.e., Algorithm 4, we have
t+1
X·,j
− bt
2
2
≤ η22 bt(k) − bt
2
2
∗
≤ η22 X·,j
− bt
2
2
,
where bt(k) is the best k-sparse approximation of bt with respect to the atom set D. Therefore, we get
t+1
∗
X·,j
− X·,j
2
2
t+1
∗
− bt + bt − X·,j
≤ X·,j
≤2
t+1
X·,j
−
2
bt 2
t
2
2
∗
+ 2 b − X·,j
t
∗
≤ (2 + 2η22 ) b − X·,j
2
2
2
2
.
Next we establish the relationships among various expectations
t+1
EIt X·,j
− X·,j
2
2
∗
≤ (2 + 2η22 )EIt bt − X·,j
2
2
2
+ δ1 )
2
∗ 2
X·,j 2 + δ2 ) +
∗
≤ (2 + 2η22 )(β1 EIt PΓ̂ (bt − X·,j
)
t
≤ (2 + 2η22 )β1 (β2 EIt X·,j
−
t
∗
:= κj X·,j
− X·,j
2
2
(2 + 2η22 )δ1
+ σj ,
where the first inequality is guaranteed by Lemma 2 and the second inequality is guaranteed by Lemma 3.
Here the contraction coefficient κj and the tolerance parameter σj are defined by
κj = (2 + 2η22 )β1 β2 ,
σj = (2 + 2η22 )β1 δ2 + (2 + 2η22 )δ1 ,
where β1 , δ1 are defined in (24) and β2 , β2 are defined in (26). Then similar to the proof of Theorem 2,
we can derive that
E X t+1 − X ∗ F ≤ κ
et+1 X 0 − X ∗ F + η
where
κ
e=
q
max κj ,
1≤j≤L
v
u
u
and η = t
PL
j=1
σj
1 − max κj
1≤j≤L
.
Stochastic Greedy Algorithms For Multiple Measurement Vectors
15
In particular, if η1 = η2 = 1 and p(i) = 1/M , then
s
κ
e=4
−
α4k (ρ+
4k − ρ4k )
.
−
ρ−
4k (2ρ4k − α4k )
If, in addition, α4k = ρ−
4k , then we have
κ
e = 2κ,
where the contraction coefficient κ for MStoGradMP is given in (22), which implies that MStoGradMP
converges faster than CStoGradMP in this case due to the smaller contraction coefficient. Compared with
MStoIHT, MStoGradMP has even larger convergence improvement in terms of recovery accuracy and
running time.
6 Distributed Compressive Sensing Application
In this section, we show that the objective function commonly used in the distributed compressive sensing
problem satisfies the D-RSC and D-RSS properties, which paves the theoretical foundation for using
the proposed algorithms in this application. Suppose that there are L underlying signals xj ∈ Rn for
j = 1, 2, . . . , L, and their measurements are generated by
yj = A(j) xj ,
j = 1, 2, . . . , L
where A(j) ∈ Rm×n (m n) is the measurement matrix (a.k.a. the sensing matrix). For discussion
simplicity, we assume all measurement matrices are the same, i.e., A(j) = A = [A·,1 , . . . , A·,n ]. By concatenating all vectors as a matrix, we rewrite the above equation as follows
Y = AX,
Y = [y1 , . . . , yL ] ∈ Rm×L ,
X = [x1 , . . . , xL ] ∈ Rn×L .
Now assume that the atom set is finite and denote D = {d1 , . . . , dN } with the corresponding dictionary
D = [d1 , . . . , dN ]. Consider the following distributed compressive sensing model with common sparse
supports [26]
L
min
X
1 X
2
kyj − Axj k2
2m j=1
s.t. xj = Dθj
supp(θj ) = Ω ⊆ {1, 2, . . . , N }.
(28)
Here the objective function has the form
F (X) =
1
2
kY − AXkF .
2m
(29)
Then F (X) can be written as
F (X) =
M
1 X
fi (X),
M i=1
where M = m/b and
L
fi (X) =
n
X
1 X
1
2
(Yi,j −
Ai,k Xk,j )2 =
kYi,· − Ai,· Xk2 .
2b j=1
2b
k=1
(30)
16
Jing Qin et al.
The above expression shows that fi ’s satisfy the Assumption 1 and thereby the concatenated algorithms
in Section 3 can be applied. We first compute the partial derivative. For s = 1, 2, . . . , n and t = 1, 2, . . . , L,
we have
! n
L
n
X
X
1 X
∂fi (X)
∂Xk,j
=
2
Ai,k Xk,j − Yi,j
Ai,k
∂Xs,t
2b j=1
∂Xs,t
k=1
k=1
!
L
n
n
X
1X X
=
Ai,k Xk,j − Yi,j
Ai,k δk,s δj,t
b j=1
k=1
k=1
!
n
1 X
Ai,k Xk,t − Yi,t Ai,s .
=
b
k=1
Here δi,j = 1 if i = j and zero otherwise. Thus the generalized gradient of fi (X) with respect to X has
the form
1
∂fi (X)
= ATi,· (Ai,· X − Yi,· ).
∂X
b
Lemma 4 If the sensing matrix A ∈ Rm×n satisfies the Restricted Isometry Property (RIP), i.e., there
exists δk > 0 such that
2
2
2
(1 − δk ) kxk2 ≤ kAxk2 ≤ (1 + δk ) kxk2
for any k-sparse vector x ∈ Rn , then the function F (X) defined in (29) satisfies the D-restricted strong
convexity property.
Proof Let X ∈ Rn×L with k nonzero rows, which implies that each column of X has at most k nonzero
components. By the RIP of A, we have
2
2
2
(1 − δk ) kX·,j k2 ≤ kAX·,j k2 ≤ (1 + δk ) kX·,j k2 ,
2
Note that kXkF =
PL
j=1
j = 1, . . . , L.
2
kX·,j k2 . Thus we get
2
2
2
(1 − δk ) kXkF ≤ kAXkF ≤ (1 + δk ) kXkF .
For any two X, X 0 ∈ Rn×L with | supprD (X) ∪ supprD (X 0 )| ≤ k, we have
E
D ∂F (X)
, X0 − X
F (X 0 ) − F (X) −
∂X
D1
E
1
2
2
=
kY − AX 0 kF − kY − AXkF −
AT (AX − Y ), X 0 − X
2m
m
1
2
=
kA(X 0 − X)kF
2m
1 − δk
2
≥
kX 0 − XkF .
2m
Thus F (X) satisfies the D-restricted strong smoothness property with ρ−
k =
1−δk
2m .
Lemma 5 If the sensing matrix A ∈ Rm×n satisfies the following property: for any k-sparse vector
x ∈ Rn , there exists δk > 0 such that
1
ATτi ,· Aτi ,· x
b
2
≤ (1 + δk ) kxk2
where Aτi ,· is formed by extracting rows of A with row indices in τi . Then the function fi (X) defined in
(30) satisfies the D-restricted strong convexity property.
Stochastic Greedy Algorithms For Multiple Measurement Vectors
17
Proof Let X ∈ Rn×L have k nonzero rows. Then
1
ATτi ,· Aτi ,· X·,j
b
2
≤ (1 + δk ) kX·,j k2 ,
j = 1, . . . , L,
which implies that
1
ATτi ,· Aτi ,· X
b
F
≤ (1 + δk ) kXkF .
For any two X, X 0 ∈ Rn×L with | supprD (X) ∪ supprD (X 0 )| ≤ k, we have
∂fi (X) ∂fi (X 0 )
−
∂X
∂X
1
ATi,· (Ai,· X − Yi,· ) − ATi,· (Ai,· X 0 − Yi,· )
b
1
ATi,· Ai,· (X − X 0 ) F
=
b
≤ (1 + δk ) kX − X 0 kF .
=
F
F
Therefore fi (X) satisfies the D-restricted strong convexity with ρ+
k (i) = 1 + δk .
By Lemma 4, Lemma 5 and the convergence analysis in Section 5, the contraction coefficient in the
proposed algorithms depends on the coefficient in the RIP condition, whose infimum for some special
type of matrices are available [27].
7 Numerical Experiments
In this section, we conduct a variety of numerical experiments to validate the effectiveness of the proposed
algorithms. More specifically, our tests include reconstruction of row sparse signals from a linear system
and joint sparse video sequence recovery. To compare different results quantitatively, we use the relative
error defined as follows
kX t − X ∗ kF
,
ReErr =
kX ∗ kF
where X ∗ is the ground truth and X t is the estimation of X ∗ at the t-th iteration. Regarding the
computational efficiency, we also record the running time which counts all the computation time over a
specified number of iterations excluding data loading or generation. Here we use the commands tic and
toc in Matlab. To assess the concatenated SMV algorithms, we apply the SMV algorithm sequentially
to the same sensing matrix and all columns of the measurement matrix Y , and save all intermediate
approximations of each column of X for further computation of the relative error. In all tests, we use
the discrete uniform distribution, i.e., p(i) = 1/M for i = 1, 2, . . . , M in the non-batched version and
p(i) = 1/d for i = 1, 2, . . . , d in the batched version. The parameter η is fixed as 1. By default, each
algorithm is stopped when either the relative error between two subsequent approximations of X ∗ reaches
the tolerance or the maximum number of iterations is achieved.
All our experiments are performed in a desktop with an Intel R Xeon R CPU E5-2650 v4 @ 2.2GHz and
64GB RAM in double precision. The algorithms are implemented in Matlab 2016a running on Windows
10.
7.1 Joint Sparse Matrix Recovery
In the first set of experiments, we compare the proposed algorithms and their concatenated SMV counterparts in terms of reconstruction error and running time. In particular, we investigate the impact of
the sparsity level k, and the number of underlying signals L to be reconstructed on the performance
of the BStoIHT and BStoGradMP in both concatenated SMV and MMV versions, in terms of relative
error and the running time. To reduce randomness in the results, we run 50 trials for each test with fixed
parameters and then take the average over the number of trials.
18
Jing Qin et al.
1.2
1.05
StoIHT, k=5
StoIHT, k=10
StoIHT, k=15
StoIHT, k=20
MStoIHT, k=5
MStoIHT, k=10
MStoIHT, k=15
MStoIHT, k=20
0.95
0.8
relative error
relative error
1
BStoIHT, k=5
BStoIHT, k=10
BStoIHT, k=15
BStoIHT, k=20
BMStoIHT, k=5
BMStoIHT, k=10
BMStoIHT, k=15
BMStoIHT, k=20
1
0.9
0.6
0.4
0.85
0.2
0
0.8
0
0.05
0.1
0.15
0.2
running time
0.25
0.3
0.35
0.4
0
0.05
0.1
0.15
0.2
running time
Fig. 1 Comparison of StoIHT and MStoIHT in both non-batched and batched versions for various sparsity levels k of the
signal matrix. From left to right: batch sizes are 1 and 10.
First, we compare StoIHT and MStoIHT in both non-batched and batched versions, and fix the
maximum number of iterations as 1000 and γ = 1 in both algorithms. To start with, we create a sensing
matrix A ∈ R100×200 where each entry follows the normal distribution with zero mean and variance of
1/100 and each column of A is normalized by dividing its `2 -norm. In this way, it can be shown that the
spark of A, i.e., the smallest number of linearly dependent columns of A, is 100 with probability one.
To create a signal matrix X ∗ ∈ R200×40 , we first generate a Gaussian distributed random matrix of size
200 × 40, and then randomly zero out (200 − k) rows where k is the row sparsity of X ∗ . The measurement
matrix Y is created by AX for the noise-free cases. By choosing the sparsity level k ∈ {5, 10, 15, 20}
and the batch size b ∈ {1, 10}, we obtain the results shown in Figure 1. Since the initial guess for the
signal matrix is set a zero matrix, all the error curves start with the point (0, 1). Notice that to show
the computational efficiency, we use the running time in seconds as the horizonal axis rather than the
number of iterations. It can be seen that as the sparsity level grows, i.e., the signal matrix is less joint
sparse, more running time (or iterations) is required to achieve the provided tolerance in terms of the
relative reconstruction error. Meanwhile, as the batch size increases, BMStoIHT performs better than the
sequential BStoIHT. With large sparsity levels, the inaccurate joint support obtained in the concatenated
SMV algorithms cause large relative errors in the first few iterations (see Figure 1).
Next, we fix the sparsity level k as 5 and choose the number of signals as L ∈ {20, 40, 60, 80}.
Figure 2 compares the results obtained by BStoIHT and BMStoIHT when the batch size is 1 and 10.
In general, BMStoIHT takes less running time than its sequential concatenated SMV counterpart. We
can see that mini-batching significantly improves the reconstruction accuracy and reduces the running
time of BMStoIHT. After a large number of tests, we also find that the computational speedup of
BMStoIHT is almost linear with respect to the number of signals to be reconstructed. Lastly, to test
the robustness to noise, we add the Gaussian noise with zero mean and standard deviation (a.k.a. noise
level) σ ∈ {0.02, 0.04, 0.06, 0.08} to the measurement matrix Y . The relative errors for all BStoIHT and
BMStoIHT results versus running time are shown in Figure 3. It is worth noting that the change of
sparsity and noise levels have insignificant impact on the running time, which explains that the curve
corresponding to the same algorithm stops almost at the same horizontal coordinate in Figure 1 and
Figure 3. By contrast, the running time grows as the number of signals to be recovered increases which
suggests that the endpoint of each curve has different horizontal coordinates in Figure 2.
In the second set of tests, we compare StoGradMP and MStoGradMP in non-batched and batched
versions. It is known that StoGradMP usually converges much faster than StoIHT. We fix the maximum
number of iterations as 30, γ = 1, and the batch size as 1 (non-batched version). Similar to the previous
tests, we create a 100 × 200 random matrix whose columns are normalized, and fix the number of signals
L = 40. Figure 4 shows the results obtained by StoGradMP and MStoGradMP with sparsity level
Stochastic Greedy Algorithms For Multiple Measurement Vectors
1
1.05
StoIHT, L=20
StoIHT, L=40
StoIHT, L=60
StoIHT, L=80
MStoIHT, L=20
MStoIHT, L=40
MStoIHT, L=60
MStoIHT, L=80
0.95
BStoIHT, L=20
BStoIHT, L=40
BStoIHT, L=60
BStoIHT, L=80
BMStoIHT, L=20
BMStoIHT, L=40
BMStoIHT, L=60
BMStoIHT, L=80
0.9
0.8
0.7
relative error
1
relative error
19
0.9
0.85
0.6
0.5
0.4
0.3
0.2
0.8
0.1
0
0.75
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0
0.4
0.05
0.1
0.15
0.2
0.25
0.3
0.35
running time
running time
Fig. 2 Comparison of StoIHT and MStoIHT in both non-batched and batched versions for various numbers of signals L
to be reconstructed. From left to right: batch sizes are 1 and 10.
1
1.05
StoIHT, <=0.02
StoIHT, <=0.04
StoIHT, <=0.06
StoIHT, <=0.08
MStoIHT, <=0.02
MStoIHT, <=0.04
MStoIHT, <=0.06
MStoIHT, <=0.08
relative error
0.95
BStoIHT, <=0.02
BStoIHT, <=0.04
BStoIHT, <=0.06
BStoIHT, <=0.08
BMStoIHT, <=0.02
BMStoIHT, <=0.04
BMStoIHT, <=0.06
BMStoIHT, <=0.08
0.9
0.8
0.7
relative error
1
0.9
0.85
0.6
0.5
0.4
0.3
0.8
0.2
0.75
0.1
0
0.7
0
0.1
0.2
0.3
0.4
running time
0.5
0.6
0.7
0.8
0
0.05
0.1
0.15
0.2
0.25
0.3
running time
Fig. 3 Comparison of StoIHT and MStoIHT in both non-batched and batched versions for various noise levels σ to the
measurement matrix. From left to right: batch sizes are 1 and 10.
k ∈ {60, 70, 80, 90}. Note that for better visualization, we skip the starting point (0,1) for all relative error
plots, and use the base 10 logarithmic scale for the horizontal axis of running time since the MStoGradMP
takes much less running time than StoGradMP after the same number of iterations. Unlike StoIHT and
MStoIHT, both StoGradMP and MStoGradMP require that the sparsity level is no more than n/2,
i.e., 100 in our case. As the sparsity k increases, the operator pinv for computing the pseudo-inverse
matrix becomes more computationally expensive for matrices with more columns than their rank, which
results in the significant growth of running time. For sparse signal matrices, StoGradMP performs better
than StoIHT in terms of convergence. Next, we set the number of signals as 20, 40, 60, 80, and get the
results shown in Figure 5. It can be seen that MStoGradMP always takes less running time with even
higher accuracy than the sequential StoGradMP. We also discovered that the computation speedup of
MStoIHT is almost constant with respect to the number of signals to be reconstructed. The robustness
comparison is shown in Figure 6, where the noise level ranges in {0.02, 0.04, 0.06, 0.08}. Furthermore, it
is empirically shown that the BMStoGradMP performs much better than BMStoIHT considering their
respective convergence behavior and robustness.
20
Jing Qin et al.
1.6
1.6
StoGradMP, k=60
StoGradMP, k=70
StoGradMP, k=80
StoGradMP, k=90
MStoGradMP, k=60
MStoGradMP, k=70
MStoGradMP, k=80
MStoGradMP, k=90
relative error
1.2
1.2
1
1
0.8
0.8
0.6
0.6
0.4
10 -3
10 -2
10 -1
10 0
BStoGradMP, k=60
BStoGradMP, k=70
BStoGradMP, k=80
BStoGradMP, k=90
BMStoGradMP, k=60
BMStoGradMP, k=70
BMStoGradMP, k=80
BMStoGradMP, k=90
1.4
relative error
1.4
0.4
10 -3
10 1
10 -2
10 -1
running time
10 0
10 1
running time
1.2
1
1.1
0.9
1
0.8
relative error
relative error
Fig. 4 Comparison of BStoGradMP and BMStoGradMP with various sparsity levels k. From left to right: batch sizes are
1 and 10.
0.9
0.8
StoGradMP, L=20
StoGradMP, L=40
StoGradMP, L=60
StoGradMP, L=80
MStoGradMP, L=20
MStoGradMP, L=40
MStoGradMP, L=60
MStoGradMP, L=80
0.7
0.6
0.5
10 -3
10 -2
10 -1
running time
0.7
0.6
BStoGradMP, L=20
BStoGradMP, L=40
BStoGradMP, L=60
BStoGradMP, L=80
BMStoGradMP, L=20
BMStoGradMP, L=40
BMStoGradMP, L=60
BMStoGradMP, L=80
0.5
0.4
10 0
0.3
10 -3
10 -2
10 -1
10 0
running time
Fig. 5 Comparison of BStoGradMP and BMStoGradMP with various numbers of signals L. From left to right: batch sizes
are 1 and 10.
7.2 Joint Sparse Video Sequence Recovery
In this set of experiments, we compare the proposed Algorithm 8 and the split Bregman algorithm
for constrained MMV problem (SBC) [11, Algorithm 2], on joint sparse video sequence reconstruction.
We first download a candle video consisting of 75 frames from the Dynamic Texture Toolbox in http:
//www.vision.jhu.edu/code/. In order to make the test video sequence possess a joint sparse structure,
we extract 11 frames of the original data, i.e., frames 1 to 7, 29, 37, 69 and 70, each of which is of size
80 × 30. Then we create a data matrix X ∈ R2400×11 , whose columns are a vectorization of all video
frames. To further obtain a sparse representation of X, the K-SVD algorithm [28] is applied to obtain a
dictionary Ψ ∈ R2400×50 for X. The K-SVD dictionary Ψ and the support of the corresponding coefficient
matrix Θ for the extracted 11 frames are shown in Figure 7 (a) and (b). Some selected columns of
the dictionary Ψ , namely atoms, are reshaped as an image of size 80 × 30 illustrated in Figure 7 (c)
and (d). It can be seen that these 11 frames are nearly joint sparse under the learned dictionary. The
ΘkF
= 0.0870.
relative error of using this K-SVD dictionary Ψ to represent the data matrix X is kX−Ψ
kXkF
60×2400
A Gaussian random matrix Φ ∈ R
with zero mean and unit variance is set as a sensing matrix,
Stochastic Greedy Algorithms For Multiple Measurement Vectors
21
1.2
1
StoGradMP, <=0.02
StoGradMP, <=0.04
StoGradMP, <=0.06
StoGradMP, <=0.08
MStoGradMP, <=0.02
MStoGradMP, <=0.04
MStoGradMP, <=0.06
MStoGradMP, <=0.08
relative error
0.8
0.8
0.7
relative error
1
BStoGradMP, <=0.02
BStoGradMP, <=0.04
BStoGradMP, <=0.06
BStoGradMP, <=0.08
BMStoGradMP, <=0.02
BMStoGradMP, <=0.04
BMStoGradMP, <=0.06
BMStoGradMP, <=0.08
0.9
0.6
0.4
0.6
0.5
0.4
0.3
0.2
0.2
0.1
0
10 -3
10 -2
10 -1
running time
10 0
10 1
0
10 -3
10 -2
10 -1
10 0
10 1
running time
Fig. 6 Comparison of BStoGradMP and BMStoGradMP with various noise levels σ. From left to right: batch sizes are 1
and 10.
which is used to measure this data matrix. In other words, the measurements Y ∈ R60×11 are generated
via Y = ΦX. Given the measurements Y and the new sparse representation dictionary A = ΦΨ , we then
apply Algorithm 8 and SBC to recover the joint sparse coefficient matrix Θ̂. In Algorithm 8, the sparsity
level k is set as 10, and the block size b is set as 3, which implies that there are d = 20 blocks. In addition,
we set η1 = η2 = 1. Both algorithms stop when the residual error reaches a tolerance threshold τ = 10−6 ,
i.e., kY − AΘ̂kF ≤ τ . The first four recovered frames are shown in Figure 8. Regarding the computation
time, the proposed Algorithm 8 is about ten times faster than SBC according to our experiments. The
proposed algorithm takes less running time than SBC since it stops at the 10-th iteration while SBC stops
at the 1691-th iteration. Even surprisingly, although the same error tolerance is set for both algorithms,
kX −Ψ Θ̂ k
the proposed algorithm yields a smaller relative error ( :,ikX:,i k2:,i 2 ) than SBC. Figure 9 compares the
relative discrepancy kY − AΘ̂kF / kY kF versus running time for both algorithms, where we can see the
zigzag pattern in the SBC result.
8 Conclusions
In this paper, we study the multiple measurement vector sparse reconstruction problem, which is of great
importance in a large amount of signal processing applications. We propose two stochastic greedy algorithms, MStoIHT and MStoGradMP, together with their respective accelerated versions by applying
the mini-batching technique. Our convergence analysis has shown theoretically that the proposed algorithms converge faster than their concatenated SMV counterparts. A variety of numerical experiments
on linear systems and video frame processing have shown that the proposed algorithms outperform the
concatenated SMV algorithms in terms of efficiency, accuracy and robustness to the noise.
Acknowledgements The initial research for this effort was conducted at the Research Collaboration Workshop for Women
in Data Science and Mathematics, July 17-21 held at ICERM. Funding for the workshop was provided by ICERM, AWM
and DIMACS (NSF grant CCF-1144502).
References
1. S. F. Cotter, B. D. Rao, K. Engan, and K. Kreutz-Delgado, “Sparse solutions to linear inverse problems with multiple
measurement vectors,” IEEE Transactions on Signal Processing, vol. 53, no. 7, pp. 2477–2488, 2005.
22
Jing Qin et al.
Support of sparse coefficient
K-SVD Dictionary
1
0.1
5
0.9
10
0.8
15
0.7
20
0.6
25
0.5
30
0.4
35
0.3
40
0.2
45
0.1
0.08
500
0.06
0.04
1000
0.02
0
1500
-0.02
-0.04
2000
-0.06
-0.08
50
10
20
30
40
50
0
2
(a)
20-th column
24-th column
(c)
4
6
8
10
(b)
28-th column
29-th column
33-th column
47-th column
(d)
Fig. 7 (a) K-SVD dictionary learned from the total 75 frames from the candle video. (b) Support of sparse coefficient
matrix for extracted 11 frames. The sparse coefficient matrix has non-zero entries on white area. (c-d) Some columns of the
learned K-SVD dictionary.
2. B. D. Rao, K. Engan, and S. Cotter, “Diversity measure minimization based method for computing sparse solutions
to linear inverse problems with multiple measurement vectors,” in Acoustics, Speech, and Signal Processing, 2004.
Proceedings.(ICASSP’04). IEEE International Conference on, vol. 2, pp. ii–369, IEEE, 2004.
3. Z. He, A. Cichocki, R. Zdunek, and J. Cao, “CG-M-FOCUSS and its application to distributed compressed sensing,”
in International Symposium on Neural Networks, pp. 237–245, Springer, 2008.
4. J. A. Bazerque and G. B. Giannakis, “Distributed spectrum sensing for cognitive radio networks by exploiting sparsity,”
IEEE Transactions on Signal Processing, vol. 58, no. 3, pp. 1847–1862, 2010.
5. A. Majumdar and R. K. Ward, “Joint reconstruction of multiecho mr images using correlated sparsity,” Magnetic
resonance imaging, vol. 29, no. 7, pp. 899–906, 2011.
6. A. Majumdar and R. K. Ward, “Face recognition from video: An MMV recovery approach,” in Acoustics, Speech and
Signal Processing (ICASSP), 2012 IEEE International Conference on, pp. 2221–2224, IEEE, 2012.
7. A. Majumdar and R. Ward, “Rank awareness in group-sparse recovery of multi-echo mr images,” Sensors, vol. 13,
no. 3, pp. 3902–3921, 2013.
8. M. E. Davies and Y. C. Eldar, “Rank awareness in joint sparse recovery,” IEEE Transactions on Information Theory,
vol. 58, no. 2, pp. 1135–1146, 2012.
9. S. Li, D. Yang, G. Tang, and M. B. Wakin, “Atomic norm minimization for modal analysis from random and compressed
samples,” arXiv preprint arXiv:1703.00938, 2017.
10. H. Lu, X. Long, and J. Lv, “A fast algorithm for recovery of jointly sparse vectors based on the alternating direction
methods,” in Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pp. 461–
469, 2011.
11. Z. Jian, F. Yuli, Z. Qiheng, and L. Haifeng, “Split bregman algorithms for multiple measurement vector problem,”
Multidim Syst Sign Process, 2015.
Stochastic Greedy Algorithms For Multiple Measurement Vectors
Ground truth
BMStoGradMP
SBC
23
Ground truth
BMStoGradMP
SBC
10
10
10
10
10
10
20
20
20
20
20
20
30
30
30
30
30
30
40
40
40
40
40
40
50
50
50
50
50
50
60
60
60
60
60
60
70
70
70
70
70
70
80
80
10
20
30
80
10
20
30
80
10
20
30
80
10
20
30
(a) Frame 1
Ground truth
80
10
20
30
10
BMStoGradMP
SBC
Ground truth
BMStoGradMP
10
10
10
10
10
20
20
20
20
20
20
30
30
30
30
30
30
40
40
40
40
40
40
50
50
50
50
50
50
60
60
60
60
60
60
70
70
70
70
70
70
80
10
20
30
80
10
20
(c) Frame 3
30
SBC
10
80
20
(b) Frame 2
30
80
10
20
30
80
10
20
30
80
10
20
30
10
20
30
(d) Frame 4
Fig. 8 The first four frames reconstructed by BMStoGradMP and SBC. The relative reconstruction errors of all BMStoGradMP results, i.e., kX:,i − Ψ Θ̂:,i k2 /kX:,i k2 , are: (a) 1.3819 × 10−15 , (b) 1.3819 × 10−15 , (c) 1.1547 × 10−15 , (d)
4.8385 × 10−15 . By contrast, the relative reconstruction errors of all SBC results are: (a) 5.0906 × 10−15 , (b) 5.0906 × 10−15 ,
(c) 6.8482 × 10−10 , and (d) 3.5012 × 10−8 .
12. P. Feng and Y. Bresler, “Spectrum-blind minimum-rate sampling and reconstruction of multiband signals,” in Acoustics,
Speech and Signal Processing (ICASSP), 1996 IEEE International Conference on, vol. 3, pp. 1688–1691, IEEE, 1996.
13. J. M. Kim, O. K. Lee, and J. C. Ye, “Compressive MUSIC: Revisiting the link between compressive sensing and array
signal processing,” IEEE Transactions on Information Theory, vol. 58, no. 1, pp. 278–301, 2012.
14. K. Lee, Y. Bresler, and M. Junge, “Subspace methods for joint sparse recovery,” IEEE Transactions on Information
Theory, vol. 58, no. 6, pp. 3613–3641, 2012.
15. J. A. Tropp, “Greed is good: Algorithmic results for sparse approximation,” IEEE Transactions on Information theory,
vol. 50, no. 10, pp. 2231–2242, 2004.
16. J. Chen and X. Huo, “Theoretical results on sparse representations of multiple-measurement vectors,” IEEE Transactions on Signal Processing, vol. 54, no. 12, pp. 4634–4643, 2006.
17. J. A. Tropp, A. C. Gilbert, and M. J. Strauss, “Algorithms for simultaneous sparse approximation. part i: Greedy
pursuit,” Signal Processing, vol. 86, no. 3, pp. 572–588, 2006.
18. D. Needell and J. A. Tropp, “CoSaMP: Iterative signal recovery from incomplete and inaccurate samples,” Applied and
Computational Harmonic Analysis, vol. 26, no. 3, pp. 301–321, 2009.
19. D. Needell and R. Vershynin, “Signal recovery from incomplete and inaccurate measurements via regularized orthogonal
matching pursuit,” IEEE Journal of selected topics in signal processing, vol. 4, no. 2, pp. 310–316, 2010.
20. W. Dai and O. Milenkovic, “Subspace pursuit for compressive sensing signal reconstruction,” IEEE Transactions on
Information Theory, vol. 55, no. 5, pp. 2230–2249, 2009.
21. T. Blumensath and M. E. Davies, “Iterative hard thresholding for compressed sensing,” Applied and computational
harmonic analysis, vol. 27, no. 3, pp. 265–274, 2009.
24
Jing Qin et al.
10 0
10
10 -2
-5
^ F =kY kF
kY ! A#k
^ F =kY kF
kY ! A#k
10
10 0
10 -4
-10
10 -6
10
-15
10
-20
10 -8
0
0.2
0.4
0.6
0.8
running time
(a) BMStoGradMP
1
1.2
1.4
#10
-3
10
-10
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
running time
(b) SBC
Fig. 9 Plot of the relative discrepancy kY − AΘ̂kF / kY kF versus the running time for BMStoGradMP and SBC.
22. N. Nguyen, S. Chin, and T. Tran, “A unified iterative greedy algorithm for sparsityconstrainted optimization. 2013.”
23. N. Nguyen, D. Needell, and T. Woolf, “Linear convergence of stochastic iterative greedy algorithms with sparse constraints,” IEEE Transactions on Information Theory, 2017.
24. J. R. Magnus, “On the concept of matrix derivative,” Journal of Multivariate Analysis, vol. 101, no. 9, pp. 2200–2206,
2010.
25. D. Needell and R. Ward, “Batched stochastic gradient descent with weighted sampling,” in Approximation Theory XV:
San Antonio 2016, pp. 279–306, Springer, 2016.
26. D. Baron, M. B. Wakin, M. F. Duarte, S. Sarvotham, and R. G. Baraniuk, “Distributed compressed sensing,” IEEE
Transactions on Information Theory, vol. 52, no. 12, pp. 5406–5425, 2006.
27. S. Foucart and H. Rauhut, A mathematical introduction to compressive sensing, vol. 1. Birkhäuser Basel, 2013.
28. M. Aharon, M. Elad, and A. Bruckstein, “K-svd: An algorithm for designing overcomplete dictionaries for sparse
representation,” IEEE Transactions on signal processing, vol. 54, no. 11, pp. 4311–4322, 2006.
| 8cs.DS
|
arXiv:1609.08474v1 [math.GT] 27 Sep 2016
Cubulable Kähler groups
Thomas Delzant and Pierre Py
September 2016
Abstract
We prove that a Kähler group which is cubulable, i.e. which acts properly discontinuously
and cocompactly on a CAT(0) cubical complex, has a finite index subgroup isomorphic to a
direct product of surface groups, possibly with a free Abelian factor. Similarly, we prove that a
closed aspherical Kähler manifold with a cubulable fundamental group has a finite cover which
is biholomorphic to a topologically trivial principal torus bundle over a product of Riemann
surfaces. Along the way, we prove a factorization result for essential actions of Kähler groups
on irreducible, locally finite CAT(0) cubical complexes, under the assumption that there is no
fixed point in the visual boundary.
Contents
1 Introduction
2
2 Groups acting on CAT(0) cubical complexes
4
2.1
Cubical complexes and hyperplanes . . . . . . . . . . . . . . . . . . . . . . . . . . .
5
2.2
Existence of strongly separated hyperplanes . . . . . . . . . . . . . . . . . . . . . .
6
2.3
Non-amenability of certain Schreier graphs
8
. . . . . . . . . . . . . . . . . . . . . .
3 Fibering Kähler groups acting on CAT(0) cubical complexes
12
3.1
Filtered ends . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
3.2
The first pluriharmonic map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
3.3
A second plurisubharmonic map . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
3.4
Producing fibrations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
18
4 Cubulable Kähler manifolds and groups
19
4.1
Cubulable Kähler groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
4.2
Cubulable Kähler manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
22
5 Comments
23
1
1
Introduction
The purpose of this text is to characterize fundamental groups of compact Kähler manifolds which
are cubulable, i.e. which act properly discontinuously and cocompactly on a CAT(0) cubical complex. For a survey concerning fundamental groups of compact Kähler manifolds, referred to as
Kähler groups below, we refer the reader to [2, 13]. For basic definitions about cubical complexes,
we refer the reader to [8, I.7] or [50]. We only mention here that cubical complexes form a particular type of polyhedral complexes, and provide an important source of examples in geometric
group theory.
From now on we will only deal with finite dimensional cubical complexes. Such complexes have
a natural complete and geodesic metric. In the 80’s, Gromov gave a combinatorial criterion for
this metric to be CAT(0) [28, §4.2]. His work on hyperbolization procedures [28], as well as Davis’
work on reflection groups [18] drew attention to these complexes; see also [3, 19, 20]. Later, Sageev
established a link between actions of a group G on a CAT(0) cubical complex and the existence of
a multi-ended subgroup of G [49]. More recently, under the influence of Agol, Haglund and Wise,
CAT(0) cubical complexes received a lot of attention as the list of examples of cubulated groups
increased dramatically and thanks to their applications to 3-dimensional topology. We refer the
reader to [1, 5, 6, 32, 33, 54, 55] for some of these developments.
On the other hand, as we will see, actions of Kähler groups on CAT(0) cubical complexes are
very constrained. Even more, Kähler groups are in a sense orthogonal to groups acting on CAT(0)
cubical complexes. The first results in this direction go back to the work of the first author, together
with Gromov [23]. Note that most of the results of [23] are not formulated in terms of actions on
CAT(0) cubical complex but can be interpreted in terms of such actions thanks to the work of
Sageev [49]. In [23], the authors studied codimension one subgroups of hyperbolic Kähler groups.
More generally they studied filtered ends, or cuts of hyperbolic Kähler groups, see section 3.1 for the
definitions. Some of their results were later generalized by Napier and Ramachandran [46]. Stated
informally, the results of [23] show that under suitable hyperbolicity assumptions, the presence of
sufficiently many subgroups of a Kähler group Γ whose numbers of filtered ends are greater than
2 implies that a finite index subgroup Γ0 of Γ is a subdirect product of a certain number of surface
groups with a free Abelian group. All along this text, the expression surface group stands for
the fundamental group of a closed oriented hyperbolic surface. Recall that a subdirect product
of groups G1 , . . . , Gm is a subgroup of G1 × · · · × Gm which surjects onto each factor. So, in the
previous statement, we mean that Γ0 is a subdirect product of G1 , . . . , Gm where all the Gi ’s are
surface groups, except possibly for one of them which could be free Abelian.
There is also a particular class of actions of Kähler groups on CAT(0) cubical complexes which
are easy to describe. These are the ones given by homomorphisms into right-angled Artin groups.
For homomorphisms from Kähler groups to right-angled Artin groups, it is easy to obtain a factorization result through a subdirect product of surface groups, possibly with a free Abelian factor.
This relies on the facts that right-angled Artin groups embed into Coxeter groups and that Coxeter
groups act properly on the product of finitely many trees. This was observed by the second author
in [48].
These results left open the question of describing actions of Kähler groups on general CAT(0)
cubical complexes. We answer this question here assuming that the cubical complexes are locally
finite and that the actions have no fixed point in the visual boundary. We will briefly discuss
in section 5 the need for these two hypothesis. From this, one deduces easily a description of
cubulable Kähler groups.
The following statements involve various more or less standard notions about cubical complexes:
essential actions, irreducible complexes, visual boundaries... We recall all these definitions in
section 2.1.
2
Theorem A Let M be a closed Kähler manifold whose fundamental group π1 (M ) acts on a finite
dimensional locally finite irreducible CAT(0) cubical complex X. Assume that:
• the action is essential,
• π1 (M ) has no fixed point in the visual boundary of X,
• π1 (M ) does not preserve a Euclidean flat in X.
Then there exists a
nected fibers F : M1
π1 (Σ) is surjective.
fixed point set in X
finite cover M1 of M which fibers: there exists a holomorphic map with con→ Σ to a hyperbolic Riemann surface such that the induced map F∗ : π1 (M ) →
Moreover the kernel of the homomorphism F∗ acts elliptically on X, i.e. its
is nonempty.
Hence, up to restricting the action to a convex subset of X, the action factors through the
fundamental group of a hyperbolic Riemann surface. Indeed, as a convex subspace one can take
the fixed point set of the subgroup Ker(F∗ ); this is a subcomplex of the first cubic subdivision of
X.
From Theorem A, and using results due to Caprace and Sageev [15] and to Bridson, Howie,
Miller and Short [10], we deduce a characterization of cubulable Kähler groups. In the next three
statements, we implicitly assume that the cubical complexes are locally finite.
Theorem B Suppose that a Kähler group Γ acts properly discontinuously and cocompactly on a
CAT(0) cubical complex X. Let X = X1 × · · · × Xr be the decomposition of X into irreducible
factors. Assume moreover that the action of Γ on X is essential and that each Xi is unbounded.
Then Γ has a finite index subgroup Γ∗ which is isomorphic to a direct product
Γ∗ ' H1 × · · · × Hk × G1 × · · · × Gm ,
(1)
where each Hj is isomorphic to Z, each Gj is a surface group, and k + m = r is the number
of irreducible factors of X. Moreover the isomorphism (1) can be chosen in such a way that for
each i, Xi contains a convex Γ∗ -invariant subset Yi on which the action factors through one of the
projections Γ∗ → Gj or Γ∗ → Hj .
If a group G acts properly discontinuously and cocompactly on a CAT(0) cubical complex X,
one can always find a G-invariant subcomplex Y inside X, on which the action is essential and all
of whose irreducible factors are unbounded. This follows from Corollary 6.4 in [15]. This implies
the following corollary:
Corollary C Suppose that a Kähler group Γ acts properly discontinuously and cocompactly on a
CAT(0) cubical complex X. Then Γ has a finite index subgroup which is isomorphic to a direct
product of surface groups with a free Abelian group. The conclusion of Theorem B holds after
replacing X by a suitable invariant subcomplex.
In what follows, we will say that a closed manifold M is cubulable if it has the same homotopy
type as the quotient of a finite dimensional CAT(0) cubical complex by a properly discontinuous,
free, cocompact group action. Using an argument going back to Siu [53], and the fact that a
cubulable manifold is aspherical, we will finally prove the following theorem.
Theorem D If a Kähler manifold M is cubulable, it admits a finite cover which is biholomorphic
to a topologically trivial principal torus bundle over a product of Riemann surfaces. If M is
algebraic, it admits a finite cover which is biholomorphic to a direct product of an Abelian variety
with a product of Riemann surfaces.
3
We mention that the factorization result in Theorem A does not rely on the use of harmonic
maps unlike most of the factorization results for actions of Kähler groups on symmetric spaces,
trees or more general buildings. Indeed, although there is a general theory of harmonic maps with
values into CAT(0) spaces [31, 39], the crucial step
harmonic ⇒ pluriharmonic
namely the use of a Bochner formula to prove that a harmonic map from a Kähler manifold to a
certain nonpositively curved space is automatically pluriharmonic is not available when the target
space is a CAT(0) cubical complex. Thus, our proof follows a different scheme. The idea is to
produce fibrations of our Kähler manifold M over a Riemann surface, in such a way that the kernel
N π1 (M ) of the homomorphism induced by the fibration preserves a certain hyperplane ĥ of the
cubical complex. Since N is normal it will have to preserve each hyperplane of the π1 (M )-orbit of
ĥ. From this we will deduce that N actually acts elliptically on the CAT(0) cubical complex, i.e.
has a fixed point.
From the Kähler point of view, the new idea of this paper appears in Section 3.4. It can be
summarized as follows. Consider an infinite covering Y of a closed Kähler manifold. Given a proper
pluriharmonic function u on Y , one can look for conditions under which the foliation induced by
the holomorphic form du1,0 is actually a fibration. Our observation is that it is enough to find on Y
a second plurisubharmonic function which is not a function of u. This differs from other fibration
criterions in the study of Kähler groups as the second function we need is only required to be
plurisubharmonic. In the various Castelnuovo-de Franchis type criterions known so far, one needs
two pluriharmonic functions, or two holomorphic 1-forms, see [47] and the references mentioned in
the introduction there.
The text is organized as follows. In section 2, we recall basic facts on CAT(0) cubical complexes
as well as more advanced results due to Caprace and Sageev. Given a group G acting on a CAT(0)
cubical complex and satisfying suitable hypothesis, wee also explain how to construct a hyperplane
ĥ whose stabilizer H in G has the property that the Schreier graph of H\G is non-amenable.
In section 3, we explain how to construct some nontrivial plurisubharmonic functions on certain
covering spaces of a Kähler manifold whose fundamental group acts on a CAT(0) cubical complex.
The proof of Theorem A is concluded in Section 3.4. In section 4, we prove Theorem B and D.
Finally, section 5 contains a few comments about possible improvements of our results.
Acknowledgements. We would like to thank Pierre-Emmanuel Caprace, Yves de Cornulier
and Misha Sageev for many explanations about CAT(0) cubical complexes as well as Martin
Bridson for discussions motivating this work.
2
Groups acting on CAT(0) cubical complexes
In this section we recall some basic properties of CAT(0) cubical complexes and some advanced
results due to Caprace and Sageev [15]. For a more detailed introduction to these spaces, we refer
the reader to [8, I.7] and [50] for instance. From now on and until the end of the text, all CAT(0)
cubical complexes will be finite dimensional; we will not mention this hypothesis anymore. By
convention, we assume that all cubes of our cubical complexes are isometric to [−1, 1]n for some
n.
4
2.1
Cubical complexes and hyperplanes
A cubical complex X can be naturally endowed with a distance as follows. If x and y are in X,
one considers chains
x0 = x, x1 , . . . , xn = y
where each pair (xi , xi+1 ) is contained in a cube Ci of X and one defines
d(x, y) = inf
n−1
X
dCi (xi , xi+1 )
i=0
where the inf is taken over all possible chains. Here dCi is the intrinsic distance of the cube;
the number dCi (xi , xi+1 ) does not depend on the choice of the cube containing xi and xi+1 .
The function d is a distance which is complete and geodesic thanks to the finite dimensionality
hypothesis [8, I.7]. From now on, we will always assume that the cube complexes we consider
are CAT(0) spaces, when endowed with the previous metric. According to a classical theorem by
Gromov [28, §4.2], this is equivalent to the fact that the complex is simply connected and that the
link of each vertex is a flag complex. The visual boundary [8, II.8] of a CAT(0) cubical complex
X is denoted by ∂∞ X.
We now recall the definition of hyperplanes, first introduced by Sageev [49]. We fix a CAT(0)
cubical complex X. Let 2 be the equivalence relation on edges of our complex, defined as follows.
One says that two edges e and f are equivalent, denoted e2f , if there exists a chain
e1 = e, . . . , en = f
such that for each i, ei and ei+1 are opposite edges of some 2-dimensional cube. If e is an edge we
will denote by [e] its equivalence class.
A midcube of a cube C, identified with [−1, 1]n , is the subset of C defined by setting one of the
coordinates equal to 0. One says that a midcube and an edge e of a cube C ' [−1, 1]n are transverse
if the midcube is defined by ti = 0 and the coordinate ti is the only nonconstant coordinate of
the egde e. Now the hyperplane associated to an equivalence class of edges [e] is the union of all
midcubes which are transverse to an edge of the class [e]. It is denoted by ĥ([e]). If we want to
denote a hyperplane without referring to the edge used to define it, we will denote it by ĥ. Finally
we will denote by N (ĥ) the union of the interiors of all cubes which intersect a hyperplane ĥ.
One can prove that hyperplanes enjoy the following properties [49]:
1. Each hyperplane intersects a cube in at most one midcube.
2. Each hyperplane ĥ separates X into two connected components, the closures of the two
connected components of X − ĥ are called the two half-spaces associated to ĥ.
3. Every hyperplane as well as every halfspace is a convex subset of X for the distance d.
4. For every hyperplane ĥ, the set N (ĥ) is convex and naturally isometric to ĥ × (−1, 1). If
t ∈ (−1, 1), the subset of N (ĥ) corresponding to ĥ × {t} under this isometry is called a
translated hyperplane of ĥ.
5. If a geodesic segment intersects a hyperplane in two points, it is entirely contained in it.
The group Aut(X) of automorphisms of a CAT(0) cubical complex X is the group of all permutations X → X which send isometrically a cube of X to another cube of X. An automorphism
of X is automatically an isometry of the distance d introduced earlier. In what follows, every
5
time we will say that a group G acts on a CAT(0) cubical complex, we will mean that we have a
homomorphism from the group G to the group Aut(X) of automorphisms of X. In this case, if ĥ
is a hyperplane, we will denote by
StabG (ĥ)
the subgroup of G of elements which globally preserve ĥ and which also preserve each component
of the complement of ĥ.
Following [15], we say that an action of a group G on a CAT(0) cubical complex X is essential
if no G-orbit is contained in a bounded neighborhood of a halfspace. This implies in particular
that G has no fixed point in X. We will also use the following convention. If h is a halfspace
of X, we will denote by ĥ the associated hyperplane (the boundary of h) and by h∗ the opposite
halfspace, i.e. the closure of X \ h.
Finally, we mention that there is a natural notion of irreducibility for a CAT(0) cubical complex,
see [15, §2.5]. Any finite dimensional CAT(0) cube complex X has a canonical decomposition as a
product of finitely many irreducible CAT(0) cubical complexes. Moreover every automorphism of
X preserves this decomposition, up to a permutation of possibly isomorphic factors. We refer the
reader to [15, §2.5] for a proof of these facts. We will use this decomposition to deduce Theorem B
from Theorem A.
2.2
Existence of strongly separated hyperplanes
The following definition is due to Behrstock and Charney [4] and is a key notion to study rank 1
phenomena for group actions on CAT(0) cubical complexes.
Definition 1 Two hyperplanes ĥ1 and ĥ2 in a CAT(0) cubical complex are strongly separated if
they are disjoint and if there is no hyperplane meeting both ĥ1 and ĥ2 .
If X is a CAT(0) cubical complex and Y is a closed convex subset of X, we will denote by
πY : X → Y the projection onto Y [8, II.2]. The following proposition is also taken from [4].
Proposition 1 If the hyperplanes ĥ1 and ĥ2 are strongly separated, there exists a unique pair of
point (p1 , p2 ) ∈ ĥ1 × ĥ2 such that d(p1 , p2 ) = d(ĥ1 , ĥ2 ). The projection of any point of ĥ2 (resp. ĥ1 )
onto ĥ1 (resp. ĥ2 ) is p1 (resp. p2 ).
Proof. The first claim is Lemma 2.2 in [4]. The proof given there also shows that if ĥ is a hyperplane
distinct from ĥ1 and ĥ2 , no translated hyperplane of ĥ can intersect both ĥ1 and ĥ2 . We now prove
the last assertion of the proposition. It is enough to prove that the projection on ĥ1 of each point
of ĥ2 is the middle of an edge of the cubical complex. Since ĥ1 is connected, this implies that all
points of ĥ1 have the same projection, which must necessarily be the point p1 . Let x ∈ ĥ2 . If the
projection q of x onto ĥ1 is not the middle of an edge, than there exists a cube C of dimension at
least 3 which contains q as well as the germ at q of the geodesic from q to x. One can identify C
with [−1, 1]n in such a way that
q = (0, s, s3 , . . . , sn )
with |s| < 1. Since the germ of geodesic going from q to x is orthogonal to ĥ1 , it must be contained
in [−1, 1] × {s} × [−1, 1]n−2 . We call m̂ the hyperplane associated to any edge of C parallel to
{0} × [−1, 1] × {0}n−2 . Hence the germ of [q, x] at q is contained in a translated hyperplane of
m̂. This implies that [q, x] is entirely contained in this translated hyperplane. This contradicts the
fact that no translated hyperplane can intersect both hˆ1 and hˆ2 .
2
6
Note that the second point of the proposition can be stated in the following slightly stronger way.
If h1 is the half-space defined by ĥ1 and which does not contain ĥ2 , then the projection of any point
of h1 onto ĥ2 is equal to p2 . Indeed if q ∈ h1 and if γ is the geodesic from q to πĥ2 (q), there must
exist a point q 0 on γ which belongs to ĥ1 . Hence πĥ2 (q) = πĥ2 (q 0 ) = p2 . This proves the claim.
The following theorem is due to Caprace and Sageev [15, §1.2]. It gives sufficient conditions for
the existence of strongly separated hyperplanes in a CAT(0) cubical complex X.
Theorem 2 Assume that the CAT(0) cubical complex is irreducible and that the group Aut(X)
acts essentially on X and without fixed point in the visual boundary. Then X contains two strongly
separated hyperplanes.
In the end of this section, we consider a CAT(0) cubical complex X, two strongly separated
hyperplanes ĥ and k̂ in X, and a group G acting on X. We prove a few lemmas which will be used
in the next sections. We denote by k the halfspace delimited by k̂ which does not contain ĥ.
Lemma 3 Let p be the projection of k̂ (or k) on ĥ. Let h be an element of StabG (ĥ). If h(k) ∩ k is
nonempty, then h fixes p.
Proof. This is an easy consequence of Proposition 1. Let πĥ : X → ĥ be the projection. We have
seen that πĥ (k) = p. Since h ∈ StabG (ĥ), the map πĥ is h-equivariant. Let x ∈ k be such that
h(x) ∈ k. We have:
πĥ (h(x)) = p
since h(x) ∈ k. But we also have πĥ (h(x)) = h(πĥ (x)) = h(p) since x ∈ h. Hence h(p) = p.
We now define
2
Σ = {h ∈ StabG (ĥ), h(k) ∩ k 6= ∅}
and let A be the subgroup of StabG (ĥ) generated by Σ. According to the previous lemma, every
element of A fixes p. Let
[
U=
a(k)
a∈A
be the union of all the translates by A of the halfpsace k.
Lemma 4 Let h be an element of StabG (ĥ). If h(U ) ∩ U is nonempty, then h ∈ A and h(U ) = U .
Proof. Let h be such that h(U ) ∩ U is nonempty. Then there exist a1 and a2 in A such that
ha1 (k) ∩ a2 (k) 6= ∅. This implies that a−1
2
2 ha1 is in Σ, hence in A. In particular h is in A.
If the space X is proper, there are only finitely many hyperplanes passing through a given ball
of X. Since the group A fixes the point p, in this case we get that the family of hyperplanes
a(k̂)
a∈A
is actually a finite family. This implies that the family of halfspaces (a(k))a∈A is also finite. Note
that the same conclusion holds if we assume that the action of G has finite stabilizers instead of
assuming the properness of X. Indeed, the whole group A is finite in that case. We record this
observation in the following:
Proposition 5 If X is locally finite or if the G-action on X has finite stabilizers, then there exists
a finite set A1 ⊂ A such that for all a ∈ A, there exists a1 ∈ A1 such that a1 (k) = a(k).
We will make an important use of this Proposition in section 3.3. Although we will only use it
under the hypothesis that X is locally finite, we decided to remember the fact that the conclusion
still holds for actions with finite stabilizers as this might be useful to study proper actions of Kähler
groups on non-locally compact CAT(0) cubical complexes.
7
2.3
Non-amenability of certain Schreier graphs
In this section we consider an irreducible CAT(0) cubical complex and a finitely generated group G
which acts essentially on X, does not preserve any Euclidean subspace of X, and has no fixed point
in ∂∞ X. A consequence of the work of Caprace and Sageev [15] is that under these hypothesis,
the group G contains a nonabelian free group. We will need the following slight modification of
this important fact; see also [38] for a similar statement.
To state the next theorem, we need the following definition. Following Caprace and Sageev [15],
we say that three halfspaces a, b, c form a facing triple of halfspaces if they are pariwise disjoint.
Theorem 6 Let X be an irreducible CAT(0) cubical complex. Assume that G is a finitely generated
group of automorphisms of X which satisfies the following three conditions: G does not fix any point
in the visual boundary of X, does not preserve any Euclidean subspace of X and acts essentially.
Then X contains a facing triple of halfspaces k, h, l such that the three hyperplanes k̂, ĥ, l̂ are
strongly separated and such that there exists a non-Abelian free group F < G with the property that
F ∩ gStabG (ĥ)g −1 = {1} for all g ∈ G.
Besides the facts about CAT(0) cubical complexes already recalled in the previous sections,
we will further use the following three results from [15], which apply under the hypothesis of the
previous theorem.
1. The space X contains a facing triple of halfspaces, see Theorem E in [15].
2. We will use the flipping lemma from [15, §1.2]: for any halfspace h, there exists g ∈ G such
that h∗ ( g(h).
3. Finally we will also use the double skewer lemma from [15, §1.2]: for any two halfspaces k ⊂ h,
there exists g ∈ G such that g(h) ( k ⊂ h.
We now turn to the proof of the theorem.
Proof of Theorem 6. By 1 above, one can choose a facing triple of halfspaces
h, h1 , h2
in X. By the flipping lemma, there exists an element k ∈ G such that k(h∗ ) ( h. We now define
h3 = k(h1 ) and h4 = k(h2 ). By construction, h1 , h2 , h3 , h4 is a facing quadruple of halfspaces. We
will need to assume moreover that these four halfspaces are strongly separated. This can be done
thanks to the following lemma.
Lemma 7 There exists half-spaces h0j ⊂ hj (1 ≤ j ≤ 4) such that the h0j s are strongly separated.
Proof of Lemma 7. According to Theorem 2, we can find two halfspaces a1 ⊂ a2 such that the
corresponding hyperplanes âi are strongly separated. We claim the following:
Up to replacing the pair (a1 , a2 ) by the pair (a∗2 , a∗1 ), there exists i ∈ {1, 2, 3, 4} and x ∈ G such
that x(a1 ) ⊂ x(a2 ) ⊂ h∗i .
Let us prove this claim. First we prove that one of the four halfspaces a1 , a2 , a∗1 , a∗2 is contained
in h∗j for some j. If this is false, each of these four halfspaces intersects the interior of each of the
hj . In particular ak and a∗k intersect the interior of hj . Since hj is convex, âk intersects it also.
Considering two indices j 6= j 0 and a geodesic from a point in âk ∩ hj to a point in âk ∩ hj 0 , one
8
sees that âk intersects each of the hyperplanes ĥj . Since this is true for k = 1, 2, this contradicts
the fact that the hyperplanes aˆ1 and aˆ2 are strongly separated. This concludes the proof that one
of the four halfspaces a1 , a2 , a∗1 , a∗2 is contained in h∗j for some j. If a2 or a∗1 is contained in one
of the h∗j , this proves the claim (with x = 1). Otherwise we assume that a1 ⊂ h∗j (the last case
a∗2 ⊂ h∗j being similar). In this case the double skewer lemma applied to a1 ⊂ a2 implies that there
exists x ∈ G such that:
x(a1 ) ⊂ x(a2 ) ⊂ a1 ⊂ h∗j .
This proves our claim. We now write b1 := x(a1 ), b2 := x(a2 ).
Since h1 ⊂ h∗2 , the double skewer lemma implies that there exists g ∈ G such that g(h∗2 ) ( h1 .
Similarly there exists h ∈ G such that h3 ) h(h∗4 ). Applying one of the four elements g, g −1 , h, h−1
to b1 and b2 , we obtain two halfspaces which are contained in one of the hj . For instance if i = 2
in the claim above, one has g(b1 ) ⊂ g(b2 ) ⊂ h1 . In what follows we continue to assume that we
are in this case, the other ones being similar. Since h1 ⊂ h∗4 and since h(h∗4 ) ⊂ h3 , one has
hg(b1 ) ⊂ hg(b2 ) ⊂ h3 .
Finally by a similar argument we have:
g −1 hg(b1 ) ⊂ g −1 hg(b2 ) ⊂ h2
and
h−1 g(b1 ) ⊂ h−1 g(b2 ) ⊂ h4 .
We now define h01 = g(b1 ), h02 = g −1 hg(b1 ), h03 = hg(b1 ) and h04 = h−1 g(b1 ). We check that the
h0j s are strongly separated halfspaces. It is clear that they are pairwise disjoint. We check that
the corresponding hyperplanes are strongly separated. We do this for the pair {hˆ01 , hˆ02 }, the other
cases being similar. If a hyperplane û intersects both hˆ01 and hˆ02 , it will have to intersect also g(b̂2 ).
This contradicts the fact that the pair
g(b̂1 ), g(b̂2 )
is strongly separated. Hence hˆ0 and hˆ0 are strongly separated.
2
1
2
We now continue our proof using the facing quadruple of strongly separated hyperplanes constructed in the previous lemma. So, up to replacing hj by h0j , we assume that
(hj )1≤j≤4
is a facing quadruple of strongly separated halfspaces which moreover do not intersect ĥ.
Exactly as in the proof of the previous lemma, the double skewer lemma implies that there
exists g ∈ G such that g(h∗2 ) ( h1 and h ∈ G such that h3 ) h(h∗4 ). Define U = h1 ∪ h2 and
V = h3 ∪ h4 . We now see that we have a Schottky pair: since g(V ) ⊂ h1 ⊂ h∗2 , for any positive
integer n, we have g n (V ) ⊂ h1 ⊂ U . Also, since g −1 (V ) ⊂ h2 ⊂ h∗1 we have g −n (V ) ⊂ h2 ⊂ U for
any positive integer n. Similarly hn (U ) ⊂ V for any non zero integer n. The ping-pong lemma
implies that g and h generate a free subgroup of G. Note that this argument is borrowed from the
proof of Theorem F in [15].
Now we observe that ĥ ⊂ U c ∩ V c . If we apply one of the four elements {g, g −1 , h, h−1 } to ĥ we
obtain a subset of U or a subset of V . This implies that for any nontrivial element γ of hg, hi we
have γ(ĥ) ⊂ U or γ(ĥ) ⊂ V . In particular γ(ĥ) ∩ ĥ = ∅, and the intersection of the groups hg, hi
and StabG (ĥ) is trivial.
We need to prove something slightly stronger. Namely we are looking for a free subgroup of G
which intersects trivially every conjugate of StabG (ĥ). So we first make the following observation.
Let x be an element of hg, hi which is not conjugate to a power of g or of h. We will prove that
x does not belong to any conjugate of StabG (ĥ). Up to changing x into x−1 and up to conjugacy,
we can assume that this element has the form:
x = g a1 hb1 · · · g ar hbr
9
with ai , bi nonzero integers and r ≥ 1. This implies that any positive power xm of x satisfies
xm (U ) ⊂ U . Better, since x(U ) ⊂ g a1 (V ) it follows from the properties of g and of the hi ’s that
the distance of x(U ) to the boundary of U is bounded below by some positive number δ. This
implies that
d(xm (U ), ∂U ) ≥ mδ.
(2)
Similarly one proves that
d(x−m (V ), ∂V ) ≥ mδ 0
(3)
0
for any m ≥ 1 and for some positive number δ . Suppose now by contradiction that x stabilizes a
hyperplane û. Assume that û does not intersect U . For p ∈ û we have
d(p, U ) = d(xm (p), xm (U ))
for any integer m ≥ 1. But since xm (p) is never in U the quantity d(xm (p), xm (U )) is greater or
equal to the distance from xm (U ) to ∂U . Hence we obtain, using Equation (2):
d(p, U ) ≥ mδ
which is a contradiction since the left hand side does not depend on m. This proves that û must
intersect U . In a similar way, using Equation (3), one proves that û must intersect V . But this
contradicts the fact that the halfspaces (hj )1≤j≤4 are pairwise strongly separated. Hence x does
not preserve any hyperplane. In particular x is not contained in any G-conjugate of StabG (ĥ).
Now one can consider a normal subgroup N of the free group hg, hi which does not contain any
nontrivial power of g or h (for instance its derived subgroup). Any finitely generated subgroup
F < N has the property that it intersects trivially every G-conjugate of StabG (ĥ). This proves our
claim.
We finally construct two halfspaces k and l as in the statement of the Theorem. We simply have
to repeat the arguments used in the proof of Lemma 7. Exactly as in the proof of that lemma,
one can find two halfspaces b1 ⊂ b2 which are contained in h∗j for some j and such that b̂1 and
b̂2 are strongly separated. We assume that j = 2 to simplify. We continue as in the proof of the
lemma. Applying the element g to b1 and b2 , we obtain two halfspaces which are contained in h1 :
g(b1 ) ⊂ g(b2 ) ⊂ h1 . Since h1 ⊂ h∗4 and since h(h∗4 ) ⊂ h3 , one has
hg(a1 ) ⊂ hg(a2 ) ⊂ h3 .
Finally by a similar argument we have:
g −1 hg(b1 ) ⊂ g −1 hg(b2 ) ⊂ h2 .
We now define k = g(b1 ) and l = g −1 hg(b1 ). One now checks that ĥ, l̂, k̂ are strongly separated
exactly as in the end of the proof of Lemma 7.
2
In what follows, we will use the following definition.
Definition 2 Let G be a finitely generated group acting on a CAT(0) cubical complex X. We
say that a hyperplane ĥ of X is stable for G if the Schreier graph StabG (ĥ)\G is nonamenable,
i.e. satisfies a linear isoperimetric inequality. We say that a halfspace h is stable for G if the
corresponding hyperplane ĥ is stable for G.
Note that to define the Schreier graph of StabG (ĥ)\G, one needs to pick a finite generating set
for G, but the (non)amenability of this graph is independent of this choice. We refer the reader
to [36] for the discussion of various equivalent notions of amenability for Schreier graphs. Here we
will only need the following:
10
Lemma 8 Let G be a group, H < G a subgroup, and F < G a finitely generated free group
such that F ∩ gHg −1 = {1} for all g ∈ G. Then any Schreier graph of H\G satisfies a linear
isoperimetric inequality i.e. is nonamenable.
Proof. It is well-known that the Schreier graph of H\G satisfies a linear isoperimetric inequality
if and only if the G-action on `2 (G/H) does not have almost invariant vectors. So we will prove
this last fact. For this it is enough to prove that the F -action on `2 (G/H) does not have almost
invariant vectors. But the hypothesis on F implies that F acts freely on G/H. Hence, as an
F -representation, `2 (G/H) is isomorphic to a direct sum of copies of the regular representation of
F on `2 (F ). This proves the claim.
2
We will need to record the following corollary of the previous theorem.
Corollary 1 Under the hypothesis of Theorem 6, we have:
• Any halfspace which is part of a facing triple of halfspaces is stable.
• For any halfspace h which is part of a facing triple of halfspaces, and any finite index subgroup
G2 of G, there exists γ ∈ G2 such that h and γ(h) are strongly separated.
Proof. This is contained in the proof of the previous theorem. Indeed for the first point of the
corollary, we observe that we started the proof of the previous theorem with any facing triple
of halfspaces and proved that a given halfspace among the three is stable, as a consequence of
Lemma 8.
For the second point, we consider the triple h, l, k constructed in the previous theorem. We have
l ⊂ h∗ . Applying the double skewer lemma to this last pair, we find x ∈ G such that x(h∗ ) ( l.
This implies xn (h∗ ) ⊂ l for all n ≥ 1. We pick n0 ≥ 1 such that xn0 ∈ G2 . In particular the
hyperplane xn0 (ĥ) is contained in l. Any hyperplane meeting both xn0 (ĥ) and ĥ would have to
meet l̂, which is impossible since ĥ and l̂ are strongly separated. Hence xn0 (ĥ) and ĥ are strongly
separated.
2
We will also use the following classical fact.
Lemma 9 Let M be a closed Riemannian manifold with fundamental group G. Let H < G be a
subgroup and let M1 → M be the covering space associated to H. Then the Schreier graph H\G
satisfies a linear isoperimetric inequality if and only if M1 satisfies a linear isoperimetric inequality.
A proof can be found in [30, Ch. 6]. Note that the proof in [30] is given only is the case when H
is trivial, but the arguments apply in general. Combining Theorem 6, Corollary 1, and Lemma 9,
we obtain:
Proposition 10 Let M be a closed Riemannian manifold with fundamental group G. Suppose that
G acts on a CAT(0) cubical complex X, satisfying the hypothesis of Theorem 6. Let h be a halfspace
of X which is part of a facing triple of halfspaces and let ĥ be the corresponding hyperplane. Let
Mĥ be the covering space of M corresponding to the subgroup
StabG (ĥ) < G.
Then, Mĥ satisfies a linear isoperimetric inequality.
11
3
Fibering Kähler groups acting on CAT(0) cubical complexes
In this section we first give a criterion to produce fibrations of certain open Kähler manifolds over
Riemann surfaces (Proposition 11). Although this criterion is well-known to experts, we explain
how to deduce its proof from known results about filtered ends of Kähler manifolds. This serves
as a pretext to survey this notion and its applications in the study of Kähler groups. Later, in
sections 3.2 and 3.3, we explain how to construct pluriharmonic or plurisubharmonic functions on
certain covering spaces of a compact Kähler manifold, starting from an action of its fundamental
group on a CAT(0) cubical complex. We finally prove Theorem A in section 3.4.
3.1
Filtered ends
The aim of this subsection is to recall the proof of the following classical proposition.
Proposition 11 let M be a closed Kähler manifold. Let M1 → M be a covering space of M and
f → M1 be the universal cover. We assume that there exists a proper, pluriharmonic map
let π : M
f.
e : M1 → I where I is an open interval of R. Let ee := e ◦ π be the lift of e to the universal cover M
If some level set of ee is not connected, then there exists a proper holomorphic map with connected
fibers
M1 → Σ
where Σ is a Riemann surface. This applies in particular if some levet set of e is not connected.
Before turning to the proof of the proposition, we recall briefly various notions of ends in
topology and group theory.
Let Y be a noncompact manifold. Recall that the number of ends of Y , denoted by e(Y ), is the
supremum of the number of unbounded connected components of Y − K where K runs over the
relatively compact open sets of Y with smooth boundary. Now if M is a closed manifold, M1 → M
f → M1 the universal covering space, one can look at the number of
a covering space, and π : M
ends of various spaces, each of which also admits a purely group theoretical description.
f; this is also the number of ends of the fundamental group
• There is the number of ends of M
of M .
• There is the number of ends of the space M1 , which is an invariant of the pair π1 (M1 ) <
π1 (M ) [34, 51]. When this number is greater than 1, one says that π1 (M1 ) has codimension
1 in π1 (M ) or that it is multi-ended.
• There is also a third, less known notion, of filtered ends, which can be associated to a
continuous map between two manifolds. Here we will only consider the case where this map
is the universal covering map. This leads to the following definition.
f → M1 be its universal covering space. A
Definition 3 Let M1 be an open manifold and let π : M
f → M1 ) is an element of the set
filtered end of M1 (or of π : M
f − π −1 (K))
lim π0 (M
←
where K runs over the relatively compact open sets of M1 with smooth boundary. The number of
filtered ends of M1 is denoted by ee(M1 ).
12
As for the previous notions, one can show that in the case where M1 covers a compact manifold
M , this number only depends on the pair π1 (M1 ) < π1 (M ). Also, one always has ee(M1 ) ≥ e(M1 ).
f → M1 can be greater than
The interest in this notion is that the number of filtered ends of π : M
1 even if M1 is 1-ended. A simple example of this situation is obtained as follows. Take M to be
a genus 2 closed surface and Σ to be a subsurface of genus 1 with one boundary component. The
covering space M1 of M defined by the subgroup π1 (Σ) < π1 (M ) has this property.
This notion was first introduced by Kropholler and Roller [40] in a purely group theoretical
context. A topological approach to it was later given in the book by Geoghegan [27]. Filtered ends
were studied in the context of Kähler groups by Gromov and the first author [23]. There, the name
cut was used instead of filtered ends, or rather the term cut was used to indicate the presence of at
least two filtered ends for a certain map or covering space, whereas the term Schreier cut referred
to the classical notion of relative ends of a pair of groups in [23].
With the notion of filtered ends at our disposal, Proposition 11 will be a simple application of
the following theorem.
Theorem 12 Let M be a closed Kähler manifold and let M1 → M be an infinite covering space
of M . If the number of filtered ends of M1 is greater or equal to 3, then there exists a proper
holomorphic mapping with connected fibers M1 → Σ, where Σ is a Riemann surface.
This result was proved in [23] under certain additional hypothesis of “stability”. A more general
version was later proved by Napier and Ramachandran [46]. Their version includes the theorem
stated above but also more general ones which apply to Kähler manifolds which are not necessarily
covering spaces of a compact manifold.
Before turning to the proof of Proposition 11, we make the following easy observation.
Lemma 13 Let V be a complex manifold and f : V → R be a nonconstant smooth pluriharmonic
function. Denote by Crit(f ) the set of critical points of f . Then for each t ∈ R, the set Crit(f ) ∩
f −1 (t) is nowhere dense in f −1 (t).
The proof is straightforward, once one remembers that the function f is locally the real part of
a holomorphic function F and that the critical set of f locally coincides with that of F .
Proof of Proposition 11. Note that if M1 has at least three ends, the result follows from much older
results, see [44]. Thus, we could assume that M1 has only two ends, although this is not necessary.
Let t1 be a real number such that ee−1 (t1 ) is not connected. Let x and y be two points in
distinct connected components of ee−1 (t1 ). By Lemma 13, we can assume that x and y are not
critical point of f . We claim that at least one of the two sets
{e
e > t1 },
{e
e < t1 }
is not connected. Let us assume that this is false. Then one can find a path α from x to y contained
in the set {e
e ≥ t1 } and a path β from y to x contained in the set {e
e ≤ t1 }. We can assume that
α and β intersect ee−1 (t1 ) only at their endpoints. Let γ be the path obtained by concatenating
α and β. Let γ1 : S 1 → M1 be a smooth map freely homotopic to γ, intersecting ee−1 (t1 ) only at
f is simply connected, we can find a
the points x and y and transverse to ee−1 (t1 ) there. Since M
smooth map
f
u : D2 → M
coinciding with γ1 on the boundary. Pick ε > 0 small enough, such that t1 + ε is a regular value
of ee ◦ u. The intersection of the level ee ◦ u = t1 + ε with the boundary of the disc is made of two
points aε and bε such that u(aε ) → x and u(bε ) → y when ε goes to 0. Let Iε be the connected
13
component of the set {e
e ◦ u = t1 + ε} whose boundary is {aε , bε }. Up to taking a subsequence, we
can assume that Iε converges as ε goes to 0 to a compact connected subset I0 ⊂ {e
e ◦ u = t1 }. Now
the set u(I0 ) is connected, contained in ee−1 (t1 ) and contains x and y. This contradicts the fact
that x and y are in distinct components of ee−1 (t1 ). Hence at least one of the two sets {e
e > t1 } and
{e
e < t1 } is not connected. Note that the maximum principle implies that no connected component
of these open sets can be at bounded distance from the set ee−1 (t1 ). This is easily seen to imply
that M1 has at least three filtered ends. The conclusion now follows from Theorem 12.
2
Remark 1 Note that our Proposition 11 can be applied to the case where e is the primitive of
the lift of a harmonic 1-form α on M with integral periods and where M1 is the covering space
associated to the kernel of the homomorphism π1 (M ) → Z induced by α. In this case one recovers
a particular case of a result due to Simpson [52]. Hence Theorem 12 about filtered ends implies this
particular case of the result of Simpson. Proposition 11 can also be thought of as a nonequivariant
version of Simpson’s result.
3.2
The first pluriharmonic map
We now consider a CAT(0) cubical complex X, assumed to be irreducible and locally finite. Let M
be a compact Kähler manifold and let Γ be its fundamental group. We consider a homomorphism
% : Γ → Aut(X).
We suppose that the Γ-action on X satisfies the three hypothesis of Theorem A, although this will
f the universal cover of M . If ĥ is a hyperplane of
not be used untill section 3.3. We denote by M
f
X we will write Mĥ for the quotient of M by the group StabΓ (ĥ):
f/StabΓ (ĥ).
Mĥ = M
Since X is contractible, one can choose a Lipschitz
hyperplane ĥ of X and a halfspace h associated to ĥ.
d(x, ĥ)
dh (x) =
−d(x, ĥ)
(4)
f → X. We fix a
%-equivariant map ψ : M
Define a function dh on X by
if x ∈ h,
if x ∈ h∗ .
f → R be the composition f = dh ◦ ψ. Finally let f be the function induced by f on M .
Let f : M
ĥ
Proposition 14 The map f is proper. In particular the manifold Mĥ has at least two ends.
f. We consider the
Proof. In the following proof, we let H = StabG (ĥ). We pick a point x0 in M
map
F : H\G → R
defined by F (Hg) = f (g · x0 ). It is enough to prove that F is proper, this is easily seen to imply
that f is proper. So let (gn ) be a sequence of elements of G such that
|F (Hgn )| ≤ C,
(5)
for some constant C. We need to prove that the set {Hgn } is a finite subset of H\G. But
equation (5) implies that the distance of %(gn )(ψ(x0 )) to ĥ is bounded by C. Put differently this
says that each of the hyperplanes (%(gn )−1 (ĥ)) intersects the ball of radius C centered at ψ(x0 ).
By our hypothesis of local finiteness of X this implies that the family of hyperplanes (%(gn )−1 (ĥ))
14
is finite. Hence there exists element x1 , . . . , xr in G such that for all n there exists i such that
%(gn )−1 (ĥ) = %(xj )(ĥ). This implies Hgn = Hx−1
2
i .
In what follows we will denote by Ends(V ) the space of ends of an open manifold V . We recall
that V ∪ Ends(V ) carries a natural topology, see for instance [26, Ch.6]. We now make use of the
following classical theorem.
Theorem 15 Let V be a complete Riemannian manifold of bounded geometry and satisfying a
linear isoperimetric inequality. Let χ : Ends(V ) → {−1, 1} be a continuous map. Then there
exists a unique continuous map hχ : V ∪ Ends(V ) → [−1, 1] which extends χ, is harmonic on V ,
and has finite energy:
Z
|∇hχ |2 < ∞.
V
Note that we will not need the precise definition of a Riemannian manifold of bounded geometry;
the only important point we need is that a covering space of a closed Riemannian manifold is of
bounded geometry. This theorem was proved by Kaimanovich and Woess [35] and by Li and
Tam [42] independently. A simple beautiful proof due to M. Ramachandran can be found in the
paper by Kapovich [37, §9]. In the case when V is Kähler, the map hχ is pluriharmonic. This
follows from a standard integration by part argument, which is valid here because hχ has finite
energy, see Lemma 3.1 in [41] or [29, §1.1.B].
We now assume that ĥ is stable for Γ and apply the above to the manifold V = Mĥ . The
proper function f defines a partition of the space of ends of this manifold into two open sets: the
set of ends of {f ≥ 0} and the set of ends of {f ≤ 0}. Let χ : Ends(V ) → {−1, 1} be the function
taking the value 1 on the first open set and −1 on the second open set. Since Mĥ satisfies a linear
isoperimetric inequality, we obtain:
Proposition 16 There exists a unique function
uh : Mĥ → (−1, 1)
which is pluriharmonic, of finite energy and satisfies uh (x) → 1 when x → ∞ and f (x) > 0 and
uh (x) → −1 when x → ∞ and f (x) < 0.
f to M . For t0 ∈ (0, 1) we define:
We denote by p the projection from M
ĥ
f, uh (p(x)) > t0 }.
Uh,t0 = {x ∈ M
We have:
Lemma 17 If t0 is close enough to 1 the set Uh,t0 satisfies ψ(Uh,t0 ) ⊂ h.
Proof. By the continuity of uh on Mĥ ∪ Ends(Mĥ ), we have that if t0 is close enough to 1, one has
the inclusion {uh > t0 } ⊂ {f ≥ 0}. This implies the conclusion of the lemma.
2
We will ultimately produce a fibration of the manifold Mĥ onto a Riemann surface. To achieve
this, we will construct a second plurisubharmonic function defined on Mĥ , which is not a function
of uh . We build this second map in the next section.
3.3
A second plurisubharmonic map
We consider a facing triple of strongly separated halfspaces h, k, l as given by Theorem 6. According
to Corollary 1, h, k and l are stable for Γ. Thus, we can apply the results of section 3.2 to anyone
15
of the three covering spaces Mĥ , Mk̂ , Ml̂ . We recall that these covering spaces were defined by
Equation (4). We also make use of the various lemmas proved in section 2.2. In particular we
consider the subgroup
A < StabΓ (ĥ)
generated by Σ = {h ∈ StabΓ (ĥ), h(k) ∩ k 6= ∅}. If a ∈ A, the hyperplane a(k̂) is stable for Γ, as k̂
is. Thus, we can also apply the results of section 3.2 to the covering space Ma(k̂) . So we denote by
uh : Mĥ → (−1, 1)
and
ua(k) : Ma(k̂) → (−1, 1)
the proper pluriharmonic maps of finite energy provided by Proposition 16.
f
f
For each a ∈ A, let u
]
a(k) be the lift of ua(k) to M . Now define a function βk : M → R by:
βk (x) = sup u
]
a(k) (x).
a∈A
Proposition 5 implies that the A-orbit of k is finite, hence the supremum above is actually the
supremum of a finite number of smooth pluriharmonic functions. Hence βk is a continuous plurisubharmonic function. We now fix some constant C in (0, 1). Let Vk,C = {βk > C}.
f → R is A-invariant. In particular the open set Vk,C ⊂ M
f
Lemma 18
1. The function βk : M
is A-invariant.
2. For each a ∈ A, there exists Ca ∈ (0, 1) such that ψ({ug
a(k) > Ca }) ⊂ a(k).
3. If C is close enough to 1, one has ψ(Vk,C ) ⊂ ∪a∈A a(k).
Proof. Note that by the uniqueness of the map ua(k) , we have:
u
]
ek ◦ a−1 .
a(k) = u
So the function βk is also equal to
sup uek ◦ a−1 ,
a∈A
which is A-invariant. This proves the first point. The second point was already proved in Lemma 17
for the harmonic function uh : Mĥ → (−1, 1) and the proof is similar here. As for the third point,
if A1 ⊂ A is a finite set such that {a(k)}a∈A = {a(k)}a∈A1 , the constant C = max Ca has the
a∈A1
2
desired property.
We now assume that the constant C satisfies the conclusion of the previous lemma. Note that
∗
f/A. This open set actually
the open set Vk,C naturally defines an open set Vk,C
= Vk,C /A inside M
f by the bigger subgroup StabΓ (ĥ). Namely:
embeds into the quotient of M
f/A → M is injective on the closure of V ∗ .
Proposition 19 The natural map π : M
k,C
ĥ
Proof. We will check that if h ∈ StabΓ (ĥ) − A then h(Vk,C ) ∩ Vk,C is empty. Let us suppose
by contradiction that this is not empty for some h ∈ StabΓ (ĥ) − A. Let x ∈ Vk,C be such that
h(x) ∈ Vk,C . By Lemma 18 (3), there exist a1 , a2 ∈ A such that ψ(x) ∈ a1 (k) and ψ(h(x)) ∈ a2 (k).
This implies that h(a1 (k)) ∩ a2 (k) 6= ∅. By Lemma 4, h must be in A, a contradiction.
2
∗
We now define a function γ : Mĥ → R+ as follows. If x ∈ π(Vk,C
), define γ(x) = βk (y) − C,
∗
where y is a lift of x inside Vk,C . If x ∈
/ π(Vk,C ), define γ(x) = 0. We have:
16
Lemma 20 If C is close enough to 1, the function γ is continuous and plurisubharmonic. Hence,
∗
Vk,C
is a plurimassive set in the sense of [46, Def. 1.1].
f/A → M . Up to taking
Proof. As in the previous Proposition, we denote by π the projection M
ĥ
C closer to 1, we can always assume that there exists C0 ∈ (0, C) such that the conclusion of the
third point of Lemma 18 holds for C0 . We claim that F := π(Vk,C /A) is closed. We first conclude
the proof of the lemma using this claim. We consider the two open sets
π(Vk,C0 /A)
and
Mĥ − F.
They cover Mĥ , hence it is enough to prove that γ is continuous and plurisubharmonic on each of
them. On Mĥ − F , γ is 0 so there is nothing to prove. On π(Vk,C0 /A), the function γ is constructed
f. It descends to a
as follows: one considers the plurisubharmonic function max(βk − C, 0) on M
f/A → R+ . The restriction of γ to π(Vk,C /A) is obtained by composing the inverse
function q : M
0
of the map π : Vk,C0 /A → π(Vk,C0 /A) with q. Hence it is continuous and plurisubharmonic.
We finally prove that F := π(Vk,C /A) is closed. It is enough to see that
[
h(Vk,C )
h∈StabΓ (ĥ)
is closed. For this it is enough to check that there exists ε > 0 such that for any h ∈ StabΓ (ĥ) − A,
the distance from h(Vk,C ) to Vk,C is greater or equal to ε. Applying the map ψ, we see that it is
enough to find a positive lower bound, independent of h ∈ StabΓ (ĥ) − A, for the distance from
h(U ) to U in X, where U = ∪a∈A a(k), as in section 2.2. But this follows from the fact that there
is a uniform positive lower bound for the distance between two disjoint halfspaces in a CAT(0)
cubical complex.
2
We started this section considering a facing triple of strongly separated hyperplanes h, k and l.
So far, we only used h and k. In the next proposition, we make use of the third hyperplane l.
f) are connected. Then, there
Proposition 21 Assume that the level sets of u
fh (the lift of uh to M
exists a finite cover M2 → Mĥ (possibly equal to Mĥ ) and a continuous plurisubharmonic function
δ : M2 → R+ such that there exists a level set of the lift of uh to M2 on which δ is not constant.
Proof. Note that the function uh is surjective. Assume that the conclusion of the proposition is
false when M2 = Mĥ and δ = γ. Then there exists a function
ϕ : (−1, 1) → R+
such that γ = ϕ ◦ uh . We claim that the function ϕ is continuous, convex, and vanishes on [t0 , 1)
for t0 close enough to 1. Let us prove these claims.
First we note that ϕ is continuous in a neighborhood of every real number t such that u−1
h (t)
is not contained in the critical set of uh . But every real number t has this property according
to Lemma 13. Hence ϕ is continuous on (−1, 1). Note that this argument proves that for each
t ∈ (−1, 1), one can find q ∈ Mĥ and local coordinates (z1 , . . . , zn ) centered at q such that γ(z) =
ϕ(uh (q) + Re(z1 )). The convexity of ϕ then follows from [24, 5.13]. For the last claim, we pick a
f such that uh (p(x)) > t0 . Here and as before p denotes the projection M
f → M . If t0
point x ∈ M
ĥ
is close enough to 1, Lemma 17 implies that ψ(x) ∈ h. Since h and
[
a(k)
a∈A
17
are disjoint, Lemma 18 implies that h(x) ∈
/ Vk,C for any h in StabΓ (ĥ). Hence γ(p(x)) = 0. This
proves that ϕ vanishes on [t0 , 1).
These three properties of ϕ imply that the level sets of ϕ are connected. In fact the level ϕ = c
is a point for c > 0 and is an interval of the form [a0 , 1) for c = 0. This implies, together with the
f → R+ are connected. But this
hypothesis on the level sets of u
fh , that the level sets of γ ◦ p : M
implies StabΓ (ĥ) = A. Now since the A-orbit of k is finite, the group
H2 = StabΓ (ĥ) ∩ StabΓ (k̂)
is of finite index in StabΓ (ĥ). Let M2 → M1 be the corresponding cover and u2 be the lift of uh
to M2 . Let δ be the lift of the function uk to M2 . We claim that δ satisfies the conclusion of the
proposition.
If this is not the case, then δ is a function of u2 . As before one sees that δ = ϕ0 ◦ u2 where ϕ0
is continuous and convex. Actually, since δ is pluriharmonic and not only plurisubharmonic, ϕ0
must be affine. So there exists real numbers a and b such that:
δ = au2 + b.
Since u2 and δ both take values into (−1, 1) and are onto, one sees that b must be 0 and that
a = ±1. We now obtain a contradiction from this fact, making use of the hyperplane l. Let sn be
f such that d(ψ(sn ), l̂) goes to infinity and such that ψ(sn ) ∈ l. Such a
a sequence of points of M
sequence exists because the action is essential. Since d(ψ(sn ), h) ≥ d(ψ(sn ), l̂) we must have that
f (p(sn )) → −∞, hence also uh (p(sn )) → −1.
f → M2 .
In the next paragraph, we denote by x 7→ [x] the covering map M
If a = −1, one sees that δ([sn ]) → 1 which implies that ψ(sn ) ∈ k for n large enough. Hence
Ψ(sn ) ∈ k ∩ l. This is a contradiction since k and l are disjoint. If a = 1 we argue in a similar way
f such that ψ(qn ) ∈ h and d(ψ(qn ), ĥ) → ∞.
with the pair (h, k). We take a sequence qn of points of M
This implies u2 ([qn ]) → 1. Since δ = u2 this implies that ψ(qn ) ∈ k for n large enough. Since k ∩ h
is empty we get a contradiction again. This proves the proposition.
2
3.4
Producing fibrations
We continue with the notations and hypothesis from section 3.3. Our aim is now to prove the
following:
Proposition 22 The manifold Mĥ fibers: there exists a proper holomorphic map Mĥ → Σ with
connected fibers onto a Riemann surface Σ.
Proof. By Proposition 11, we can assume that the level sets of the lift of the map
uh : Mĥ → (−1, 1)
to the universal cover are connected, otherwise the conclusion is already known. By Proposition 21,
we can first replace Mĥ by a finite cover p : M2 → Mĥ such that there exists a function δ : M2 → R+
which is continuous, plurisubharmonic and not constant on the set {uh ◦p = t} for some real number
t. Note that if this is true for some number t, one can always find a number t0 , close to t, which
has the same property and moreover satisfies that t0 is a regular value of uh ◦ p. Let F be the
foliation of M2 defined by the (1, 0) part of the differential of uh ◦ p. Note that d(uh ◦ p) might
have zeros; we refer to [2, p.55] for the precise definition of F . We now consider the manifold
Xt0 = (uh ◦ p)−1 (t0 ).
18
It is invariant by the foliation F . Hence on Xt0 we have a real codimension 1 foliation defined by
the nonsingular closed 1-form which is the restriction of dC (uh ◦ p) to Xt0 . Such a foliation has
all its leaves closed or all its leaves dense; this is an elementary particular case of the theory of
Riemannian foliations. We must thus show that the restriction of F to Xt0 cannot have all its
leaves dense. Let q be a point where the restriction of δ to Xt0 reaches its maximum m. Let L(q)
be the leaf of F through q. The maximum principle implies that δ is constant on L(q). Hence
the closure of L(q) is contained in the set δ = m. Since δ is not constant on Xt0 , L(q) cannot be
dense in Xt0 , hence is closed. We have found a compact leaf of the foliation F . This compact leaf
projects to a compact leaf of the foliation Fĥ defined by du1,0
h on Mĥ . But it is now classical that
this implies that the foliation Fĥ is actually a fibration. See for instance [14, (7.4)] or [23, §4.1].2
Now we will apply the following result:
Let V be a closed Kähler manifold. Assume that V has a covering space V1 → V which admits
a proper holomorphic mapping to a Riemann surface, with connected fibers: π1 : V1 → Σ1 . Then
π1 descends to a finite cover of V : there exists a finite cover V2 of V such that V1 → V decomposes
as V1 → V2 → V , V2 admits a holomorphic mapping π2 : V2 → Σ2 with connected fibers and there
exists a holomorphic map Σ1 → Σ2 which makes the following diagram commutative:
V1
/ Σ1
V2
/ Σ2 .
This fact is now well-known, we refer the reader to [23, §5.6] or [45, Prop. 4.1] for a proof.
Applying this result to V = M and to the cover V1 = Mĥ we obtain a finite cover M2 → M
and a fibration π2 : M2 → Σ onto a Riemann surface. By replacing M2 by another finite cover,
we can assume that the fundamental group of a smooth fiber of π2 surjects onto the kernel of
the map (π2 )∗ : π1 (M2 ) → π1 (Σ). Note that a smooth fiber of the fibration of Mĥ obtained in
Proposition 22 projects onto a smooth fiber of π2 . This implies that the normal subgroup
N := Ker((π2 )∗ : π1 (M2 ) → π1 (Σ))
is contained in the stabilizer of ĥ inside π1 (M2 ). In what follows, we write Γ2 = π1 (M2 ). To
conclude the proof of Theorem A, we only have to establish the next proposition.
Proposition 23 The normal subgroup N acts as an elliptic subgroup of Aut(X) i.e. the fixed
point set of N in X is nonempty.
Proof. We know that N is contained in the group StabΓ2 (ĥ). But since N is normal in Γ2 , N is
contained in
StabΓ2 (γ(ĥ))
for all γ in Γ2 . By Corollary 1, we can pick γ ∈ Γ2 such that ĥ and γ(ĥ) are strongly separated.
Since N preserves ĥ and γ(ĥ) it must preserve the projection of ĥ onto γ(ĥ), which is a point
according to Proposition 1. Hence N is elliptic.
2
4
4.1
Cubulable Kähler manifolds and groups
Cubulable Kähler groups
We first recall a few definitions concerning finiteness properties of groups as well as a result by
Bridson, Howie, Miller and Short [10]. These will be used in the proof of Theorem B. A group G
19
is of type FPn if there is an exact sequence
Pn → Pn−1 → · · · → P0 → Z → 0
of ZG-modules, where the Pi are finitely generated and projective and where Z is considered as a
trivial ZG-module. It is of type FP∞ if it is of type FPn for all n. See [12, §VIII.5] for more details
on these notions. We simply mention that the fundamental group of a closed aspherical manifold
is of type FP∞ .
It is proved in [10] that if H1 , . . . , Hn are either finitely generated free groups or surface groups
and if G is a subgroup of the direct product
H1 × · · · × Hn
which is of type FPn , then G is virtually isomorphic to a direct product of at most n groups,
each of which is either a surface group of a finitely generated free group. We also refer the reader
to [9, 11] for more general results. Note that the idea of applying the results from [10] to Kähler
groups is not new. This possibility was already discussed in [9, 11], and put into use in [48]. We
also mention here that there exist Kähler groups which are subgroups of direct products of surface
groups but which are not of type FP∞ , see [25, 43].
We now prove Theorem B. So let Γ and X be as in the statement of the theorem. Let
X = X1 × · · · × Xr
be the decomposition of X into a product of irreducible CAT(0) cubical complexes. There is a
finite index subgroup Γ1 of Γ which preserves this decomposition. Note that the action of Γ1 on
each of the Xi is essential since the original action is essential on X. We will make use of the
following two results from [15]:
• The group Γ1 contains an element γ0 which acts as a rank 1 isometry on each irreducible
factor.
• The group Γ1 contains a copy of Zr .
See [15], Theorem C and Corollary D for these statements. We recall here that a rank 1 isometry
of a CAT(0) space is a hyperbolic isometry none of whose axis bounds a flat half-plane.
Proposition 24 Let i ∈ {1, . . . , r}. Exactly one of the following two cases occurs:
1. The action of Γ1 preserves a geodesic line in Xi .
2. The action of Γ1 has no invariant Euclidean flat in Xi and fixes no point in the viusal
boundary of Xi .
Proof. Since Γ1 contains a rank 1 isometry, if Γ1 preserves a Euclidean flat in Xi , this flat must be
a geodesic line. According to Proposition 7.3 from [15] (see also the proof of that proposition), if
Γ1 does not preserve a geodesic line in Xi , it does not have any fixed point in the visual boundary
of Xi .
2
Up to replacing Γ1 by a finite index subgroup, we assume that whenever Γ1 preserves a geodesic
line Li in some Xi , it acts by translations on it. In this case, the translation group is discrete as
follows from [7] for instance. Hence the action of Γ1 on Li factors through a homomorphism to Z.
We now continue the proof of Theorem B. We change the numbering of the factors Xi ’s so that
for 1 ≤ i ≤ k, Γ1 preserves a geodesic line in Xi and acts by translations on it, whereas for i > k,
20
it fixes no point in the visual boundary of Xi and preserves no flat. Hence for j > k, the Γ1 -action
on Xj satisfies all the hypothesis of Theorem A. Using Theorem A we get a finite index subgroup
Γ2 < Γ1 and a homomorphism
ψ : Γ2 → Zk × π1 (Σ1 ) × · · · × π1 (Σr−k )
(6)
with the following properties:
1. The homomorphism Γ2 → π1 (Σj ) induced by ψ is surjective for each k + 1 ≤ j ≤ r.
2. For each i, there exists a convex cobounded subset Yi ⊂ Xi on which the Γ2 -action factors
through the coordinate number i of the homomorphism Γ2 → Zk if i ≤ k or through the
homomorphism Γ2 → π1 (Σi−k ) if i > k.
Since the action of Γ2 is proper, the kernel N of ψ is finite. By replacing Γ2 by a finite index
subgroup Γ3 , we can assume that N ∩ Γ3 is central in Γ3 . We thus have a central extension:
{0}
/ N ∩ Γ3
/ Γ3
/ ψ(Γ3 )
/ {0}.
Lemma 25 The group ψ(Γ3 ) is of type FP∞ .
Proof. Let Y ⊂ X be the fixed point set of N . This is a subcomplex of the first cubical subdivision
of X. Since ψ(Γ3 ) is torsion-free, it acts freely on Y . The quotient Y /ψ(Γ3 ) is a finite complex,
hence ψ(Γ3 ) is of type FL, in particular FP∞ . See [12, VIII.6] for the definition of the FL condition
and its relation to the FPn and FP∞ conditions.
2
Now the result of [10] implies that ψ(Γ3 ) itself is isomorphic to a direct product of surface
groups and finitely generated free groups. No non-Abelian free factor can appear, but since this is
not needed for the next lemma, we will postpone a little bit the explanation of this fact.
Lemma 26 The group Γ3 has a finite index subgroup Γ4 which does not intersect N . In particular
Γ4 ' ψ(Γ4 ).
Proof. It is enough to prove that the central extension of ψ(Γ3 ) by N ∩Γ3 appearing above becomes
trivial on a finite index subgroup of ψ(Γ3 ). Being isomorphic to a direct product of surface groups
and free groups, ψ(Γ3 ) has torsion-free H1 . Hence the universal coefficient theorem implies that
H 2 (ψ(Γ3 ), N ∩ Γ3 )
is isomorphic to Hom(H2 (ψ(Γ3 ), Z), N ∩ Γ3 ). If H < ψ(Γ3 ) is a subgroup of finite index such that
every element in the image of the map H2 (H, Z) → H2 (ψ(Γ3 ), Z) is divisible by a large enough
integer p, the pull-back map Hom(H2 (ψ(Γ3 ), Z), N ∩ Γ3 ) → Hom(H2 (H, Z), N ∩ Γ3 ) is trivial.
Hence the desired central extension is trivial on H.
2
As in Lemma 25, one proves that Γ4 is of type FP∞ . Applying again the result of [10], we get
that Γ4 is isomorphic to a direct product of surface groups, possibly with a free Abelian factor. To
obtain the more precise statement of Theorem B, and to justify the fact Γ4 does not contain any
free non-Abelian factor, we argue as follows. We will call direct factor of the product
Zk × π1 (Σ1 ) × · · · × π1 (Σr−k ).
(7)
either one of the groups π1 (Σs ) or one of the Z copy of Zk . The intersection of Γ4 with each of the
r direct factors in (7) must be nontrivial because Γ4 contains a copy of Zr . Indeed if one of these
intersections was trivial, Γ4 would embed into a direct product which does not contain Zr . We call
L1 , . . . , Lr these intersections, where Li < Z for 1 ≤ i ≤ k and where Li < π1 (Σi−k ) for i ≥ k + 1.
The proof of [10, p. 101] shows that for i ≥ k + 1, Li is finitely generated and of finite index inside
π1 (Σi−k ). This implies that Γ4 contains a finite index subgroup Γ∗ which is the direct product
of each of its intersections with the factors in (7). The group Γ∗ now satisfies the conclusion of
Theorem B.
21
4.2
Cubulable Kähler manifolds
We now turn to the proof of Theorem D. Let M be as in the statement of the theorem. Note
that in particular, M is aspherical. Applying Corollary C, we see that a finite cover M1 of M has
fundamental group isomorphic to a product of the form
Z2l × π1 (S1 ) × · · · × π1 (Sm )
(8)
where the Si ’s are closed surfaces of genus greater than 1. We fix such an isomorphism and we
denote by πi (1 ≤ i ≤ m) the projection from π1 (M1 ) onto π1 (Si ). From now on the proof will not
make any further use of CAT(0) cubical complexes. We only use arguments from Kähler geometry.
First, we will need the following classical result, see Theorem 5.14 in [16]:
Theorem 27 Let X be a Kähler manifold, S a topological surface of genus > 2, and π : π1 (X) →
π1 (S) a surjective homomorphism with finitely generated kernel. Then, there exists a complex
structure on S and a holomorphic map with connected fibers p : X → S such that the map
p∗ : π1 (X) → π1 (S)
induced by p is equal to π.
Applying this theorem to the various πi ’s we obtain that the surfaces Si ’s can be endowed with
complex structures such that one has holomorphic maps pi : M1 → Si inducing the homomorphisms
πi at the level of fundamental group.
Let A be the Albanese torus of M1 and let α : M1 → A be the Albanese map, which is welldefined up to translation. We also denote by Ai the Albanese torus (or Jacobian) of the Riemann
surface Si and by αi : Si → Ai the corresponding map. By definition of the Albanese maps, for each
i there exists a holomorphic map ϕi : A → Ai which makes the following diagram commutative:
M1
α
/A
αi
/ Ai .
pi
Si
ϕi
We denote by ϕ : A → A1 × · · · × Am the product of the maps ϕi and by β : S1 × · · · × Sm →
A1 × · · · × Am the product of the maps αi . Up to composing the maps αi ’s with some translations,
we can and do assume that ϕ maps the origin of A to the origin of A1 × · · · × Am , hence is a group
homomorphism.
Let Y be the following submanifold of A:
Y = {y ∈ A, ϕ(y) ∈ Im(β)}
This is indeed a submanifold since β is an embedding and ϕ is a submersion. Now by construction
the Albanese map α of M1 can be written as:
α=i◦Φ
where i is the inclusion of Y in A and Φ : M1 → Y is holomorphic. We now have:
Lemma 28 The map Φ is a homotopy equivalence between M1 and Y .
Proof. Let B be the kernel of the map ϕ. The complex dimension of B is equal to
m
X
1
b1 (M1 ) −
b1 (Sj )
2
j=1
22
which equals the number l appearing in Equation (8). One can find a C ∞ diffeomorphism θ : A →
B × A1 × · · · × Am such that ϕ ◦ θ−1 is equal to the natural projection
B × A1 × · · · × Am → A1 × · · · × Am .
This implies that the complex manifold Y is C ∞ diffeomorphic to B × S1 × · · · × Sm . In particular,
Y and M1 have isomorphic fundamental groups. We finally prove that Φ induces an isomorphism
on fundamental groups. It follows from our description of Y that one can choose identifications of
π1 (M1 ) and π1 (Y ) with
Z2l × π1 (S1 ) × · · · × π1 (Sm )
in such a way that Φ∗ preserves each of the projections onto the groups π1 (Sj ). This implies that
Φ∗ induces an isomorphism between the quotients of π1 (M1 ) and π1 (Y ) by their respective centers.
To conclude, it is enough to check that Φ∗ induces an isomorphism between the centers of π1 (M1 )
and π1 (Y ). But the composition of Φ∗ with the projection from π1 (Y ) onto its abelianization
coincides with the map α∗ : π1 (M1 ) → π1 (A) ' H1 (M1 , Z). Since the center of π1 (M1 ) injects in
H1 (M1 , Z), the restriction of Φ∗ to the center of π1 (M1 ) must be injective. Now using the fact
that the quotient of H1 (M1 , Z) by the image of the center of π1 (M1 ) in H1 (M1 , Z) is torsionfree,
one sees easily that Φ∗ (Z(π1 (M1 ))) must be equal to the center of π1 (Y ). This concludes the proof
that Φ∗ is an isomorphism.
The manifold M1 is aspherical by hypothesis. The manifold Y is also aspherical since it is
diffeomorphic to B × S1 × · · · × Sm . Since Φ induces an isomorphism on fundamental group, it is
a homotopy equivalence. This concludes the proof of the lemma.
2
We now conclude the proof using the following fact due to Siu and contained in the proof of
Theorem 8 from [53]:
Let f : Z1 → Z2 be a holomorphic map between two compact Kähler manifolds of dimension n.
Assume that f is of degree 1 and that the induced map H2n−2 (Z1 , R) → H2n−2 (Z2 , R) is injective.
Then f is a holomorphic diffeomorphism. The conclusion holds in particular if f is a homotopy
equivalence.
Applying Siu’s result to the map Φ : M1 → Y we obtain that M1 and Y are biholomorphic. This
proves the first statement in Theorem D. When the original manifold M is algebraic, an easy
application of Poincaré’s reducibility theorem [21, VI.8] shows that a finite cover of M1 is actually
biholomorphic to a product of a torus with finitely many Riemann surfaces. This concludes the
proof of Theorem D.
5
Comments
We discuss here some possible improvements to our results.
First, one would like to remove the hypothesis of local finiteness in Theorem A. We summarize
at which points this hypothesis was used:
1. It was used for the first time in Proposition 5. However in this place, we have seen that it can
be replaced by the hypothesis that the group action under consideration has finite stabilizers.
Note that Proposition 5 is used later in section 3.3 to prove that a certain function defined as
a supremum of continuous plurisubharmonic functions is actually the supremum of a finite
number of continuous functions, hence is continuous.
23
2. It is also used in the proof of Proposition 14 to show that the signed distance function to
a hyperplane induces a proper function on a suitable covering space of the manifold under
consideration.
Second, one could try to remove the hypothesis that there is no fixed point in the visual
boundary in Theorem A. There is a well-known strategy to achieve this, see [17, §2.H] as well
as appendix B by Caprace in [17]. If a group G acts without fixed point on a CAT(0) cubical
complex X but with a fixed point in the visual boundary ∂∞ X, one can construct another CAT(0)
cubical complex X1 , of smaller dimension, on which G acts and such that X1 embeds in the Roller
boundary of X. By applying this construction repeatedly, one can obtain a description of actions
having a fixed point in the visual boundary. The reason why we cannot use this method here is that
the passage from X to X1 does not preserve the local finiteness of the complex. We thank PierreEmmanuel Caprace for useful discussions concerning this point. Thus, one sees that removing the
hypothesis of local finiteness from Theorem A should also allow to describe parabolic actions of
Kähler groups on CAT(0) cubical complexes. Note that parabolic actions of Kähler groups on
trees are already understood [22].
References
[1] I. Agol, The virtual Haken conjecture, with an appendix by I. Agol, D. Groves and J. Manning,
Doc. Math. 18 (2013), 1045–1087. – cited on p. 2
[2] J. Amoros, M. Burger, K. Corlette, D. Kotschick and D. Toledo, Fundamental groups of compact Kähler manifolds, Mathematical Surveys and Monographs 44, American Mathematical
Scoiety, Providence, RI, 1996. – cited on p. 2, 18
[3] W. Ballmann and J. Swiatkowski, On groups acting on nonpositively curved cubical complexes,
Enseign. Math. (2) 45, No. 1-2 (1999), 51–81. – cited on p. 2
[4] J. Behrstock and R. Charney, Divergence and quasimorphisms of right-angled Artin groups,
Math. Ann. 352, No. 2 (2012), 339–356. – cited on p. 6
[5] N. Bergeron, F. Haglund and D. T. Wise, Hyperplane sections in arithmetic hyperbolic manifolds, J. Lond. Math. Soc. (2) 83, No. 2 (2011), 431–448. – cited on p. 2
[6] N. Bergeron and D. T. Wise, A boundary criterion for cubulation, Amer. J. Math. 134, No. 3
(2012), 843–859. – cited on p. 2
[7] M. Bridson,On the semisimplicity of polyhedral isometries, Proc. Amer. Math. Soc. 127, No. 7
(1999), 2143–2146. – cited on p. 20
[8] M. Bridson and A. Haefliger, metric spaces of nonpositive curvature, Grundlehren der Mathematischen Wissenschaften 319, Springer-Verlag, Berlin (1999). – cited on p. 2, 4, 5, 6
[9] M. Bridson and J. Howie, Subgroups of direct products of elementarily free groups, Geom.
Funct. Anal. 17, No. 2 (2007), 385–403. – cited on p. 20
[10] M. Bridson, J. Howie, C. F. Miller III and H. Short, The subgroups of direct products of surface
groups, Dedicated to John Stallings on the occasion of his 65th birthday, Geom. Dedicata 92
(2002), 95–103. – cited on p. 3, 19, 20, 21
[11] M. Bridson, J. Howie, C. F. Miller III and H. Short, Subgroups of direct products of limit
groups, Ann. of Math. (2) 170, No. 3 (2009), 1447–1467. – cited on p. 20
24
[12] K. S. Brown, Cohomology of groups, Corrected reprint of the 1982 original, Graduate Texts
in Mathematics 87, Springer-Verlag, New York, (1994). – cited on p. 20, 21
[13] M. Burger, Fundamental groups of Kähler manifolds and geometric group theory, Bourbaki
Seminar No. 1022, Astérisque No. 339 (2011). – cited on p. 2
[14] J. Carlson and D. Toledo, Harmonic mappings of Kähler manifolds to locally symmetric spaces,
Inst. Hautes Études Sci. Publ. Math. No. 69 (1989), 173–201. – cited on p. 19
[15] P.-E. Caprace and M. Sageev, Rank rigidity for CAT(0) cube complexes, Geom. Funct.
Anal. 21, No. 4 (2011), 851–891. – cited on p. 3, 4, 6, 7, 8, 9, 20
[16] F. Catanese, Differentiable and deformation type of algebraic surfaces, real and symplectic
structures, Symplectic 4-manifolds and algebraic surfaces, Lecture Notes in Math. 1938, Srpinger, Berlin (2008), 55–167. – cited on p. 22
[17] I. Chatterji, T. Fernos and A. Iozzi, The median class and superrigidity of actions on CAT(0)
cubes complexes, with an appendix by P.-E. Caprace, J. Topol. 9, No. 2 (2016), 349–400. –
cited on p. 24
[18] M. Davis, Groups generated by reflections and aspherical manifolds not covered by Euclidean
space, Ann. of Math. (2) 117, No. 2 (1983), 293–324. – cited on p. 2
[19] M. Davis and T. Januszkiewicz, Hyperbolization of polyhedra, J. Differential Geom. 34, No. 2
(1991), 347–388. – cited on p. 2
[20] M. Davis, T. Januszkiewicz and R. Scott, Nonpositive curvature of blow-ups, Selecta Math.
N.S. 4, No. 4 (1998), 491–547. – cited on p. 2
[21] O. Debarre, Tores et variétés abéliennes complexes, Cours Spécialisés 6, Société Mathématique
de France, Paris, EDP Sciences, Les Ulis (1999). – cited on p. 23
[22] T. Delzant, Trees, valuations and the Green-Lazarsfeld set, Geom. Funct. Anal. 18, No. 4
(2008), 1236–1250. – cited on p. 24
[23] T. Delzant and M. Gromov, Cuts in Kähler groups, Infinite groups: geometric, combinatorial
and dynamical aspects, Progr. Math. 248, Birkhäuser, Basel (2005), 31–55. – cited on p. 2,
13, 19
[24] J.-P. Demailly, Complex analytic and differential geometry, book available on
https://www-fourier.ujf-grenoble.fr/∼demailly/documents.html. – cited on p. 17
[25] A. Dimca, S. Papadima and A. Suciu, Non-finiteness properties of fundamental groups of
smooth projective varieties, J. Reine Angew. Math. 629 (2009), 89–105. – cited on p. 20
[26] C. Drutu and M. Kapovich, Lectures on Geometric Group Theory, book in preparation, available at https://www.math.ucdavis.edu/∼kapovich/papers.html. – cited on p. 15
[27] R. Geoghegan, Topological methods in group theory, Graduate Texts in Mathematics 243,
Springer, New York (2008). – cited on p. 13
[28] M. Gromov, Hyperbolic groups, Essays in group theory, Math. Sci. Res. Inst. Publ. 8, Springer,
New York (1987), 75–263. – cited on p. 2, 5
[29] M. Gromov, Kähler hyperbolicity and L2 -Hodge theory, J. Differential Geom. 33, No. 1 (1991),
263–292. – cited on p. 15
25
[30] M. Gromov, Metric structures for Riemannian and non-Riemannian spaces, based on the
1981 french original, with appendices by M. Katz, P. Pansu and S. Semmes, Progress in
Mathematics 52, Birkhäuser Boston (1999). – cited on p. 11
[31] M. Gromov and R. Schoen, Harmonic maps into singular spaces and p-adic superrigidity for
lattices in groups of rank one, Inst. Hautes Études Sci. Publ. Math. 76 (1992), 165–246. –
cited on p. 4
[32] M. Hagen and D. Wise, Cubulating hyperbolic free-by-cyclic groups: the general case, Geom.
Funct. Anal. 25, No. 1 (2015), 134–179. – cited on p. 2
[33] F. Haglund and D. Wise, Special cube complexes, Geom. Funct. Anal. 17, No. 5 (2008), 1551–
1620. – cited on p. 2
[34] C. H. Houghton, Ends of locally compact groups and their coset spaces, J. Austral. Math.
Soc. 17 (1974), 274–284. – cited on p. 12
[35] V. Kaimanovich and W. Woess, The Dirichlet problem at infinity for random walks on graphs
with a strong isoperimetric inequality, Probab. Theory Related Fields 91, No. 3-4 (1992),
445–466. – cited on p. 15
[36] I. Kapovich, The nonamenability of Schreier graphs for infinite index quasiconvex subgroups
of hyperbolic groups, Enseign. Math. (2) 48, No. 3-4 (2002), 359–375. – cited on p. 10
[37] M. Kapovich, Energy of harmonic functions and Gromov’s proof of Stallings’ theorem, Georgian Math. J. 21, No. 3 (2014), 281–296. – cited on p. 15
[38] A. Kar and M. Sageev, Ping-pong on CAT(0) cube complexes, preprint 2016, arXiv:1507.05744.
– cited on p. 8
[39] N. J. Korevaar and R. M. Schoen, Sobolev spaces and harmonic maps for metric spaces targets,
Comm. Anal. Geom. 1, No. 3-4 (1993), 561–659. – cited on p. 4
[40] P. H. Kropholler and M. A. Roller, Relative ends and duality groups, J. Pure Appl. Algebra 61,
No. 2 (1989), 197–210. – cited on p. 13
[41] P. Li, On the structure of complete Kähler manifolds with nonnegative curvature near infinity,
Invent. Math. 99, No. 3 (1990), 579–600. – cited on p. 15
[42] P. Li and L.-F. Tam Harmonic functions and the structure of complete manifolds, J. Differential Geom. 35, No. 2 (1992), 359–383. – cited on p. 15
[43] C. Llosa Isenrich, Branched covers of elliptic curves and Kähler groups with exotic finiteness
properties, preprint 2016, arXiv:1606:03142. – cited on p. 20
[44] T. Napier and M. Ramachandran, Structure theorems for complete Kähler manifolds and
applications to Lefschetz type theorems, Geom. Funct. Anal. 5, No. 5 (1995), 809–851. – cited
on p. 13
[45] T. Napier and M. Ramachandran, Hyperbolic Kähler manifolds and proper holomorphic mappings to Riemann surfaces, Geom. Funct. Anal. 11, No. 2 (2001), 382–406. – cited on p. 19
[46] T. Napier and M. Ramachandran, Filtered ends, proper holomorphic mappings of Kähler
manifolds to Riemann surfaces, and Kähler groups, Geom. Funct. Anal. 17, No. 5 (2008),
1621–1654. – cited on p. 2, 13, 17
[47] T. Napier and M. Ramachandran, L2 Castelnuovo-de Franchis, the cup product lemma, and
filtered ends of Kähler manifolds, J. Topol. Anal. 1, No. 1 (2009), 29–64. – cited on p. 4
26
[48] P. Py, Coxeter groups and Kähler groups, Math. Proc. Cambridge Philos. Soc. 155, No. 3
(2013), 557–566. – cited on p. 2, 20
[49] M. Sageev, Ends of group pairs and non-positively curved complexes, Proc. London Math. Soc.
(3) 71, No. 3 (1995), 585–617. – cited on p. 2, 5
[50] M. Sageev, CAT(0) cube complexes and groups, Geometric group theory, IAS/Park City Math.
Ser. 21, Amer. Math. Soc. Providence, RI (2014), 7–54. – cited on p. 2, 4
[51] P. Scott, Ends of pairs of groups, J. Pure Appl. Algebra 11, No. 1–3 (1977-1978), 179–198. –
cited on p. 12
[52] C. Simpson, Lefschetz theorems for the integral leaves of a holomorphic one-form, Compositio
Math. 87, No. 1 (1993), 99–113. – cited on p. 14
[53] Y.-T. Siu, The complex-analyticity of harmonic maps and the strong rigidity of compact Kähler
manifolds, Ann. of Math. (2) 112, No. 1 (1980), 73–111. – cited on p. 3, 23
[54] D. Wise, Cubulating small cancellation groups, Geom. Funct. Anal. 14, No. 1 (2004), 150–214.
– cited on p. 2
[55] D. Wise, From riches to raags: 3-manifolds, right-angled Arting groups, and cubical geometry, CBMS Regional Conference Series in Mathematics 117, Published for the Conference
Board of the Mathematical Sciences, Washington, DC, by the American Mathematical Society,
Providence, RI (2012). – cited on p. 2
Thomas Delzant
Pierre Py
IRMA
Instituto de Matemáticas
Université de Strasbourg & CNRS
Universidad Nacional Autónoma de México
67084 Strasbourg, France
Ciudad Universitaria, 04510 México DF, México
[email protected]
[email protected]
27
| 4math.GR
|
arXiv:1703.07809v1 [math.NA] 22 Mar 2017
Empirical Risk Minimization as Parameter
Choice Rule for General Linear Regularization
Methods
Housen Li
[email protected]
National University of Defense Technology, Changsha, China
Frank Werner1
[email protected]
Max Planck Institute for Biophysical Chemistry, Göttingen, Germany
and
Felix Bernstein Institute for Mathematical Statistics in the Bioscience,
University of Goettingen, Germany
We consider the statistical inverse problem to recover f from noisy measurements
Y = T f +σξ where ξ is Gaussian white noise and T a compact operator between Hilbert
spaces. Considering general reconstruction methods of the form fˆα = qα (T ∗ T ) T ∗ Y
with an ordered filter qα , we investigate the choice of the regularization
parameter
h
i
α by minimizing an unbiased estimate of the predictive risk E kT f − T fˆα k2 . The
corresponding parameter αpred and its usage are well-known in the literature, but oracle
inequalities and optimality results in this general setting areh unknown. We
i prove a
(generalized) oracle inequality, which relates the direct risk E kf − fˆαpred k2 with the
i
h
oracle prediction risk inf α>0 E kT f − T fˆα k2 . From this oracle inequality we are then
able to conclude that the investigated parameter choice rule is of optimal order.
Finally we also present numerical simulations, which support the order optimality
of the method and the quality of the parameter choice in finite sample situations.
Keywords: statistical inverse problem, regularization method, filter-based inversion, a-posteriori
parameter choice rule, order optimality, exponential bounds, oracle inequality.
AMS classification numbers: Primary 62G05; Secondary 62G20, 65J22, 65J20.
1 Introduction
Suppose we want to recover an unknown function f ∈ X from noisy measurements
Y = T f + σξ
(1)
where T : X → Y is an operator between Hilbert spaces X and Y, ξ is a standard Gaussian white
noise process and σ > 0 denotes the noise level. In fact, (1) has to be understood in a weak sense
as ξ ∈
/ Y, i.e. for each y ∈ Y we have access to observations of the form
Yy := hT f, yiY + σ hξ, yiY ∗ ×Y
2
where hξ, yi ∼ N 0, kykY and E [hξ, y1 i hξ, y2 i] = hy1 , y2 i for all y1 , y2 ∈ Y. Models of the form
(1) underly a plenitude of applications, see e.g. (O’Sullivan, 1986), and have been considered by
Mathé and Pereverzev (2001); Bissantz et al. (2007); Ingster et al. (2012, 2014); Werner (2015).
1 Corresponding
author
1
Throughout the paper we will assume that the operator T is injective, compact and HilbertSchmidt, i.e. the squares of its singular values are summable. Especially, this implies that its singular values tend to 0 and hence the inversion of T is unstable, i.e. the problem to recover f from (1)
is ill-posed and regularization is needed, see Cavalier (2011); Engl et al. (1996) and the references
therein. In the literature many different approaches for estimation of f can be found, including
methods based on a singular value decomposition (SVD) of T , see e.g. (Johnstone and Silverman,
1991; Mair and Ruymgaart, 1996; Johnstone et al., 2004; Cavalier and Golubev, 2006), waveletvaguelette (Donoho, 1995) and vaguelette-wavelet methods (Abramovich and Silverman, 1998),
and Galerkin-type methods (Cohen et al., 2004).
In this paper we follow a common paradigm and consider regularization methods written in
terms of an ordered filter qα : [0, kT ∗ T k] → R parametrized by α > 0 (see e.g. Definition 1 for the
precise requirements on qα ), meaning that the regularized solution is given by
fˆα = qα (T ∗ T ) T ∗ Y.
(2)
Regularization methods of the form (2) include famous examples like spectral cut-off and
Tikhonov regularization, and have been studied extensively in the literature, see (Engl et al.,
1996) and the references therein, and (Bissantz et al., 2007).
Choosing an appropriate parameter α in (2) is an important problem in regularization theory
as it dramatically influences the performance of the estimator fˆα . Bissantz et al. (2007) show
that estimators of the form (2) are order-optimal over certain smoothness classes W ⊂ X , if
the parameter α = α∗ is chosen in a reasonable a-priori way (depending on properties of W).
ˆ
This
means fα∗ achieves the best possible rate of convergence w.r.t. the direct risk R (α, f ) :=
E
fˆα − f
2
X
in these classes. In practice, the parameter α has to be chosen without any
knowledge of f (and hence of W), which makes a-priori parameter choice rules useless. Therefore,
a-posteriori parameter choice rules are of interest, as they make only use of the data Y and the
noise level σ > 0. As a-posteriori parameter choice rules have to adapt to the unknown properties
of W (and hence of f ), this issue is also known as adaptivity. For simplicity, we will assume here
that σ is known. In practice, the parameter σ can typically be estimated sufficiently fast from the
data or the measurement setting, see e.g. (Rice, 1984; Hall et al., 1990; Dette et al., 1998). We
will discuss this situation in more detail in Section 6.
A variety of a-posteriori parameter choice rules have been proposed in the literature, including
the discrepancy principle (Morozov, 1966; Davies and Anderssen, 1986; Lukas, 1995; Blanchard et al.,
2016), generalized cross-validation (Wahba, 1977; Golub et al., 1979; Lukas, 1993), the Lepskiı̆type balancing principle (Lepskiı̆, 1991; Mathé, 2006; Mathé and Pereverzev, 2006; Werner and Hohage,
2012) and many more. We refer to (Bauer and Lukas, 2011) for a recent overview and numerical comparison. General adaptivity in statistical inverse problems has also been treated in
(Goldenshluger, 1999; Tsybakov, 2000; Cavalier et al., 2003).
In this paper we deal with a specific method based on empirical risk minimization originally
introduced by Mallows (1973) for model selection in linear regression and therefore known as
2
Mallow’s CL . Consider the prediction risk r (α, f ) := E T fˆα − f
. Following Stein (1981),
Y
we find that an (up to a constant independent of α) unbiased estimator for this quantity is given
by
2
r̂(α, Y ) := T fˆα − 2hT fˆα , Y i + 2σ 2 Trace (T ∗ T qα (T ∗ T )) .
(3)
Y
Now the idea is to choose α as a minimizer of r̂ (α, Y ), i.e.
αpred ∈ argmin r̂ (α, Y ) .
α>0
Note that the functional (3) penalizes the the misfit between the model fˆα and the data Y , and
furthermore 2σ 2 Trace (T ∗ T qα (T ∗ T )) penalizes the number of degrees of freedom of the model.
2
We refer to Birgé and Massart (2001, 2007) for details and a discussion of other possible penalty
terms in (3).
It is known that choosing α = αpred in combination with certain regularization schemes leads to
an order optimal method w.r.t. the prediction risk r (α, f ), see e.g. (Li, 1987; Vogel, 1986; Lukas,
1993). A very precise result about its performance which is also a central ingredient of this paper
can be found in the seminal paper by Kneip (1994), who proves exponential deviation bounds for
r (αpred , f ).
Due to ill-posedness, optimality w.r.t. the prediction risk is however a very weak statement,
and consequently we are interested in order optimality w.r.t. the direct risk. This question
has hardly been touched in literature, and to the authors best knowledge the only result is due
to Chernousova and Golubev (2014) who restrict to finite dimensional spaces and spectral cutoff regularization. Nevertheless, the choice α = αpred has successfully been applied in image
denoising applications, see e.g. (Luisier et al., 2007; Candès et al., 2013; Wang and Morel, 2013;
Deledalle et al., 2014). Furthermore, the distributional behavior of αpred has recently been studied
by Lucka et al. (2017). There it has been argued that the choice αpred and choices based on
unbiased risk minimization in general do not seem suitable for inverse problems. Besides, it remains
one of the most popular parameter selection rules, due to its favorable practical performance (see
e.g. Bauer and Lukas (2011); Chernousova and Golubev (2014)). Supporting this fact, we will
prove an oracle inequality of the form
R (αpred , f ) ≤ Θ min r (α, f )
(4)
α>0
for all f ∈ W as σ ց 0 where Θ is some (explicit) functional and W ⊂ X some smoothness
class. More specifically, the functional Θ will be such that we can derive order optimality of fˆαpred
under weak assumptions over many classes W ⊂ X . Moreover, we provide a general strategy to
prove estimates of the form (4) which is of independent interest and might be used for the analysis
of other a-posteriori parameter choice rules as well. This makes our analytical methodology
substantially different from that in (Chernousova and Golubev, 2014), since the crucial techniques
(such as optional stopping of martingales) there do not apply to general regularization schemes.
For more details on oracle inequalities in statistical inverse problems we refer to (Cavalier et al.,
2002; Blanchard et al., 2016).
The rest of this paper is organized as follows. In the next section we introduce ordered filter
based regularization methods and the empirical parameter choice rule via unbiased prediction
risk minimization. The convergence analysis of such a rule is established by means of an oracle
inequality in Section 3 with corresponding conclusions on order optimality. In Section 4 we present
a general methodology for proving oracle inequalities of the form (4) and apply this methodology
for proving the results from Section 3. The performance, as well as the convergence behavior, of
the analyzed parameter choice rule is examined by comparison with other empirical parameter
choice rules in a simulation study in Section 5. We end this paper with some conclusions in
Section 6 .
2 Filter based regularization and empirical risk minimization
2.1 Ordered filter based regularization methods
Suppose that X and Y are Hilbert spaces, T : X → Y is an injective and compact Hilbert-Schmidt
operator, and ξ in (1) is a standard Gaussian white noise process as described in the introduction.
To simplify the notation we will always assume that dim (X ) = ∞, but stress that the case of
finite dimensional X (and Y)
can be treated
similarly. By assumption, there exists a singular
√
λk , ek , gk k∈N of T where λ1 ≥ λ2 ≥ · · · > 0 are the eigenvalues
value decomposition (SVD)
of T ∗ T and e1 , e2 , . . . are the corresponding normalized eigenvectors. By introducing the notation
Yk := hgk , Y i, ξk := hgk , ξi and fk = hf, ek i, we equivalently transform the model (1) to the
3
Gaussian sequence model
Yk =
iid
p
λk fk + σξk ,
k = 1, 2, . . . ,
(5)
with ξk ∼ N (0, 1).
As mentioned in the introduction, we focus on regularization methods of the form (2), which
by meanings of (5) can be equivalently formulated as
p
fˆα = λk qα (λk )Yk ,
k ∈ N.
k
P∞
Remark 1. Note that T being Hilbert-Schmidt implies that k=1 λk < ∞, and hence T ∗ Y can
be interpreted as a random variable with values in X as
"∞
#
∞
h
i
X
X
2
∗ 2
E kT ξkX = E
hξ, T ek iY =
λk E hξ, gk i2Y < ∞.
k=1
k=1
Consequently, fˆα as in (2) is well-defined.
Estimators of this form can be understood as stable
√ approximations of the well-known least
squares estimate fˆ := (T ∗ T )−1 T ∗ Y (or fˆk = Yk / λk ) in the sense of replacing (·)−1 with a
function qα (·). To obtain a well-defined and reasonable regularization method, the functions qα (·)
should satisfy proper conditions. We are particularly interested in case that qα (·) is an ordered
filter.
Definition 1. Let qα : [0, λ1 ] → R, indexed by α ∈ A ⊂ R+ , be a sequence of functions. We
always assume that 0 is an accumulation point of A and that A is bounded.
(i) The family qα , α ∈ A, is called a filter, if there exist constants Cq′ , Cq′′ > 0 such that for every
α ∈ A and every λ ∈ [0, λ1 ]
α |qα (λ)| ≤ Cq′
and
λ |qα (λ)| ≤ Cq′′ .
(ii) The filter qα , α ∈ A, is called ordered, if the sequence {qα (λk )}∞
k=1 is strictly monotone, i.e.
α<β
⇒
∀ k ∈ N : qα (λk ) ≤ qβ (λk ) and ∃ k0 ∈ N such that qα (λk0 ) < qβ (λk0 ),
2
and continuous as A ∋ α 7→ {qα (λk )}∞
k=1 ∈ ℓ .
The requirement of an ordered filter is rather weak, as it is satisfied by various regularization
methods. In Table 1 we give several examples of such. Note that for spectral cut-off regularization,
the set A has to be chosen as {λk : k = 1, 2, . . .} in order to guarantee the strict monotonicity
and the continuity required by condition (ii) in Definition 1. In Table 1 we also indicate whether
the method can be implemented without SVD. This property is extremely crucial in practice,
especially for large-scale applications, where the computation of an SVD is often impossible given
limited time and resources. The implementation of Showalter’s method, for instance, can avoid
SVD by employing Runge-Kutta schemes, see e.g. (Rieder, 2005). For a further discussion of
these and other methods we refer to the monograph (Engl et al., 1996).
In this paper, we focus on the asymptotic properties of ordered filter based regularization methods as the noise level σ tends to zero. As noticed by Bakushinskiı̆ (1984), the convergence rates of
regularization methods can be arbitrarily slow if the underlying problem ill-posed. In order to derive rates of convergence we need assume some smoothness about the unknown truth f . Typically,
the smoothness of f is measured relative to the smoothing properties of the forward operator T
in terms of a source condition, i.e. we assume that
f ∈ Wφ := {f ∈ X : f = φ(T ∗ T )w, kwkX ≤ C}
for some constant C,
(6)
where φ is a so-called index function, i.e. φ is continuous, strictly increasing, and φ(0) = 0. Note
that for any f ∈ X there exists a function φe such that f ∈ Wφe, cf. (Mathé and Hofmann, 2008).
4
To take advantage of (6) we furthermore assume that φ is a qualification of the filter qα , this is
sup φ(λ) |1 − λqα (λ)| ≤ Cφ φ(α)
for all α ∈ A,
λ∈[0,λ1 ]
(7)
with Cφ being a constant depending only on φ. For further details on general source conditions
and corresponding qualifications we refer to Mathé and Pereverzev (2003).
As an example consider φ(t) = tv , which is known as Hölder type source condition of order
v > 0:
Wv := {f ∈ X : f = (T ∗ T )v w, kwkX ≤ C} .
(8)
The function φ(t) = tv is a qualification of the filter qα if
sup λv |1 − λqα (λ)| ≤ Cv αv
for all α ∈ A.
λ∈[0,λ1 ]
(9)
In this case, the largest possible v such that (9) is satisfied, is called the classical or polynomial
qualification index v0 of the ordered filter qα . For the methods discussed in Table 1, v0 as well as
Cv is also depicted.
Table 1: Examples of ordered filters.
Method
qα (λ)
Cq′
Cq′′
v0
Cv
SVD required
Spectral cut-off
1
λ 1[α,∞) (λ)
1
1
∞
1
Yes
Tikhonov
1
λ+α
1
1
1
No
m-iterated Tikhonov
(λ+α)m −αm
λ(λ+α)m
v v (1 − v)1−v
m
1
m
(v/m)v (1 − v/m)m−v
No
1
1
∞
(v/e)v
No
1
1
∞
(v/e)v
No
Landweber (kT k ≤ 1)
Showalter
P⌊1/α⌋−1
(1 − λ)j
λ
1−exp(− α
)
j=0
λ
For further reference, we collect the assumed properties of f and qα as follows:
Assumption 1.
(i) The true solution f satisfies f ∈ Wφ as in (6).
(ii) The function φ is a qualification of the filter qα as in (7).
√
(iii) The function ψ(x) := xφ−1 ( x), with φ−1 being the inverse function of φ, is convex.
Remark 2. We stress that, being a standard assumption for convergence analysis, Assumption 1
(iii) actually imposes no restriction, since one can always work on a slightly larger source set Wφe
with another index function φe for which Assumption 1 (iii) is satisfied.
2.2 Empirical prediction risk minimization
As discussed in the Introduction, the optimal regularization parameter α will in general depend
on Y , σ and f , but the latter is unknown and hence this α cannot be realized in practice. Recall
that we always assume σ > 0 to be known. By meanings of the prediction risk r (α, f ), the optimal
α ∈ A is given by
αo = argmin r(α, f ).
(10)
α∈A
5
As a common remedy, we will try to estimate αo from the observations Y in (1) by minimizing
an unbiased estimator of r (α, f ), which can be derived as follows. Let us introduce the shorthand
notation
sα (λ) := λqα (λ).
Then we have
h
i
2
r(α, f ) = E kT (qα (T ∗ T )T ∗ (T f + σξ) − f )kY
h
i
2
2
= k(I − T qα (T ∗ T )T ∗ )T f kY + σ 2 E kT qα (T ∗ T )T ∗ξkY
=
∞
X
k=1
λk (1 − sα (λk ))2 fk2 + σ 2
∞
X
sα (λk )2 .
(11)
k=1
and furthermore
2
E T fˆα − 2hT fˆα, Y iY
Y
i
h
= E kT qα (T ∗ T )T ∗ (T f + σξ)k2Y − 2hT qα (T ∗ T )T ∗(T f + σξ), T f + σξiY
=
∞
X
k=1
λk (1 − sα (λk ))2 fk2 −
∞
X
λk fk2 + σ 2
k=1
∞
X
k=1
sα (λk )2 − 2σ 2
∞
X
sα (λk ).
k=1
Consequently for r̂ (α, Y ) as in (3) we have
E [r̂ (α, Y )] = r(α, f ) −
∞
X
λk fk2 ,
k=1
i.e. up to a constant independent of α, r̂ (α, Y ) is an unbiased estimator of r (α, f ). Hence we
define
αpred = argmin r̂(α, Y ).
(12)
α∈A
Remark 3. For the clarity of our notation, we stress that α 7→ r(α, f ) is a deterministic
function, whereas α 7→ r̂ (α, Y ) is a random function and consequently, αpred as in (11) is
a random variable. To obtain bounds for the estimator fˆαpred we will also need bounds for
r(αpred , f ), i.e. a deterministic
function
evaluated at a random variable. In particular we stress
that E [r(αpred , f )] 6= E
T fˆαpred − T f
2
Y
.
Note that αpred is computable in practice as it only relies on the data and the forward operator.
As discussed in the Introduction, there are many results available for the performance of fˆαpred
measured by the prediction risk, among which the most precise one is due to Kneip (1994) given
below.
Theorem 2.1 (Deviation bound of prediction risk (Kneip, 1994)). Assume the model (1). Let
fˆα = qα (T ∗ T )T ∗Y with an ordered filter qα , αo as in (10), and αpred as in (12). Then there exist
positive constants C1 , C2 such that for all x ≥ 0 and for all f ∈ X ,
(
)!
2
2
√
x
C
2
x, p
min
,
P T fˆαpred − T f
− T fˆαo − T f
≥ x ≤ C1 exp −
σ
Y
Y
r(αo , f )
which remains true when replacing T fˆαpred − T f
6
2
Y
− T fˆαo − T f
2
Y
by r(αpred , f ) − r(αo , f ).
Proof. Recall that dim X = ∞, which, together with the injectivity of T , implies dim Y = ∞. In
fact, the assertion for finite dimensional Y follows directly from (Kneip, 1994, Proposition 1(i))
and (Kneip, 1994, Theorem 1) by chasing the dependency of the constants on the noise level σ.
Concerning the key technical tools in Kneip’s proof, we note, that the Lemma 2 there actually
holds for infinite sequences a ∈ ℓ2 and bounded linear trace operators A : ℓ2 → ℓ2 , and that the
Lemma 3 there can be extended to infinite dimensional ordered linear smoothers, as long as they
are Hilbert-Schmidt. Thus, the proof by Kneip (1994) carries over to the case that dim Y = ∞.
The above theorem, in particular, implies (cf. (Chernousova and Golubev, 2014, Theorem 1))
2
p
E T fˆαpred − T f
≤ r(αo , f ) + Cσ r(αo , f )
for every f ∈ X ,
Y
which guarantees the order optimality of fˆαpred in terms of the prediction risk.
3 MISE estimates
The section is devoted to the convergence analysis of fˆα = qα (T ∗ T )T ∗Y with α = αpred as in
(12). In what follows, we will prove that fˆαpred also possesses an optimality property in terms of
the direct risk, i.e. in terms of the mean integrated square error (MISE).
3.1 Assumptions and merit discussions
We start with some technical assumptions. As we have already seen in the previous section, many
calculations involve summations over {λk }∞
k=1 , which can be formulated as Lebesgue-Stieltjes
integrals with respect to
Σ(x) := #{k : λk ≥ x}.
Following Bissantz et al. (2007), we assume that Σ can be approximated by a smooth function S
to avoid the difficulty caused by the non-smoothness of Σ:
Assumption 2.
(i) There exists a surrogate function S ∈ C 2 ((0, ∞)) of Σ satisfying
lim S(α)/Σ(α) = 1,
αց0
S ′ (α) < 0
for all α > 0,
lim S(α) = lim S ′ (α) = 0,
αր∞
αր∞
lim αS(α) = 0.
αց0
(ii) There exist constants α1 ∈ (0, λ1 ] and CS ∈ (0, 2) such that
− αS ′ (α) is integrable on (0, α1 ],
S ′′ (α)
CS
≤
−S ′ (α)
α
for all α ∈ (0, α1 ].
(iii) There exists a constant Cq > 0 such that
r
Z ∞
x
dx < ∞
Ψ′ (Cq x) exp −C2
2
1
Here C2 is as in Theorem 2.1.
7
with Ψ(x) :=
x
(S −1
2.
(x))
Remark 4. Note that Assumption 2 (i) and (ii) are essentially (Bissantz et al., 2007, Assumption 2) with a proper extension of S to a larger domain (0, ∞). They are rather weak requirements,
and cover a wide range of situations, see e.g. (Bissantz et al., 2007, Section 5). The additional
assumption (iii) is needed to control certain general moments of r(αpred , f ) with the help of Theorem 2.1.
The second assumption concerns the choice of regularization parameter α for (ordered) filters
qα .
Assumption 3.
(i) For α ∈ A, the function λ 7→ sα (λ) is non-decreasing.
−1/2
(ii) There exists cq > Cq
with Cq as in Assumption 2 (iii) such that
sα (α) = αqα (α) ≥ cq
as α ց 0.
(13)
Remark 5. It is easy to see that the above condition with some cq > 0 is satisfied by all the
regularization methods in Table 1, which indicates that Assumption 3 is fairly general. More
precisely, it can be understood√as a parametrization√condition. For instance, the Tikhonov method
with re-parametrization α 7→ α, i.e. qα (λ) = 1/( α + λ), still defines an ordered filter, but does
−1/2
not satisfy (13) anymore. The condition cq > Cq
requires some compatibility of the ill-posedness
with the parametrization of the filter.
Under these assumptions, upper bounds for fˆα with an a-priori choice of α have been proven
by Bissantz et al. (2007):
Theorem 3.1 (A-priori parameter choice (Bissantz et al., 2007)). Consider the model (1), and
let fˆα := qα (T ∗ T )T ∗ Y with a filter qα , and Assumptions 1 (i) and 2 (i)-(ii) hold. Let also α∗
satisfy
α∗ φ(α∗ )2 = σ 2 S(α∗ ).
(14)
(i) If Assumption 1 (ii) holds, then for some constant C1 we have
2
S(α∗ )
sup E fˆα∗ − f
≤ C1 φ(α∗ )2 = C1 σ 2
α∗
X
f ∈Wφ
(ii) If φ̃(t) :=
as σ ց 0.
√
tφ(t) is a qualification of the filter qα , then for some constant C2 we have
2
sup E T fˆα∗ − T f
≤ C2 α∗ φ(α∗ )2 = C2 σ 2 S(α∗ )
as σ ց 0.
f ∈Wφ
Y
3.2 Oracle inequality
We are now in position to derive a crucial oracle inequality in the general form of (4) for the
empirical parameter choice α = αpred .
Theorem 3.2 (Oracle inequality). Assume the model (1). Let fˆα := qα (T ∗ T )T ∗Y with an ordered
filter qα , and Assumptions 1, 2 and 3 hold. Let also αo be given by (10), and αpred by (12). Then
there are positive constants Ci , i = 1, . . . , 6, such that for all f ∈ Wφ it holds
p
2
r(αo , f ) + σ r(αo , f )
2
−1
ˆ
+ C6 σ 2 .
(15)
E fαpred − f
C2 r(αo , f ) + C3 σ + C4
≤ C1 ψ
r(αo ,f )
X
−1
C5 σ2
S
The proof is based on a general strategy based on technical lemmas, and is given in Section 4.
To ease the understanding of this paper, we will now start with conclusions from Theorem 3.2.
8
3.3 Convergence rates and examples
The derived oracle inequality in Theorem 3.2 readily provides error estimates for the estimator
fˆαpred given proper upper bounds of the oracle prediction risk r(αo , f ) as in Theorem 3.1.
Theorem 3.3
√ (Convergence rates). Assume the same setting as in Theorem 3.2, and additionally
that φ̃(t) = tφ(t) is a qualification of the filter qα . Let α∗ be given by (14). Then as σ ց 0 we
have
2
α φ(α∗ )2
σ 2 S(α∗ )
2 S(α∗ )
ˆ
∗
=
C
σ
sup E fαpred − f
≤ C1 φ(α∗ )2 + C2
+
C
1
2
2
∗)
α∗
S −1 (C3 S(α∗ ))
X
f ∈Wφ
S −1 C3 α∗ φ(α
σ2
for some positive constants Ci , i = 1, 2, 3.
Remark 6. If, in addition, there is a constant C4 > 0 such that
S(C4 x) ≥ C3 S(x)
then it follows from Theorem 3.3 that
sup E fˆαpred − f
f ∈Wφ
2
X
for all x > 0,
≤ Cφ(α∗ )2 = Cσ 2
S(α∗ )
.
α∗
This coincides with the convergence rate we obtain under an a-priori parameter choice in Theorem 3.1, which turns out to be order optimal in most cases, see e.g. Section 3.3.1 and Bissantz et al.
(2007).
We note that there are two additional important assumptions for a-posteriori parameter choice
αpred compared to the a-priori choice (cf. Theorem 3.1 (i) and Theorem 3.3). The one is Assumption 2 (iii), which concerns the control of general moments of r(αpred , f ), due to the randomness
√
of αpred . The other is that not only φ but also φ̃(t) = tφ(t) is a qualification of the filter qα .
The latter seems to be typical for parameter choice rule relying on residuals in image space, as e.g.
for the discrepancy principle in case of deterministic inverse problems, see (Engl et al., 1996,
Section 4.3) and (Mathé and Pereverzev, 2006), or for generalized cross-validation (GCV), see
(Lukas, 1993; Vogel, 2002). Still we stress that the qualification assumption plays no role in the
proof of the oracle inequality (Theorem 3.2), and it only kicks in for the derivation of convergence
rates for fˆαpred through r(αo , f ) in the convergence analysis under the a-priori parameter choice.
3.3.1 Mildly ill-posed problems
We now consider a particular mildly ill-posed problem. More precisely, we assume
Assumption 4.
(i) polynomial decay of eigenvalues of T ∗ T
λk = Ca k −a
(ii) smoothness of the truth
(
f ∈ Wb/(2a) :=
k = 1, 2, . . .
f ∈X :
∞
X
wk fk2
k=1
with some a > 1 and Ca > 0;
≤1
)
with wk = Cb k b
k = 1, 2, . . . ,
for some positive constants b and Cb .
In the above assumption, the requirement of a > 1 is to ensure that the forward operator T is
Hilbert-Schmidt, and the notation Wb/(2a) is consistent with (8) as the above smoothness condition
is equivalent to the Hölder type source condition of order b/(2a).
In this simple setting the convergence rates of ordered filter based methods with empirical
parameter choice αpred can be explicitly computed.
9
Corollary 3.4 (Order optimality). Assume the model (1). Let fˆα := qα (T ∗ T )T ∗ Y
filter qα , and Assumptions 3 and 4 with any cq > 0 hold. If the qualification index
ordered filter satisfies v0 ≥ b/(2a) + 1/2, then there are positive constants C1 and
2
2
2b
ˆ
ˆ
sup E fαpred − f
≤ C1 inf sup E f − f
≤ C2 σ a+b+1
fˆ f ∈Wb/(2a)
X
f ∈Wb/(2a)
X
with an ordered
v0 in (9) of the
C2 such that
as σ ց 0.
The infimum above is taken over all possible estimators fˆ (including both linear and nonlinear
ones).
−1/a
Proof. Define S(α) := α/Ca
, which satisfies Assumption 2 (i) and (ii). Note that φ(x) =
b/(2a)
−1
x
and then ψ(x) := xφ (x) = x(a+b)/b , so Assumption 1 is satisfied. Since Ψ(x) =
2
−1
x/ S (x) = Ca−2 x1+2a , Assumption 2 (iii) clearly holds. By definition (14), it follows that
1/(1+a+b) 2a/(1+a+b)
α∗ = Ca
σ
. Thus, by Theorem 3.3 and Remark 6, we have
b
2
b
ˆ
as σ ց 0,
sup E fαpred − f
≤ C1 φ(α∗ )2 = C1 Caa(1+a+b) (σ 2 ) 1+a+b
X
f ∈Wb/(2a)
for some positive constant C1 .
Moreover, it is well-known that (see e.g. Pinsker, 1980; Golubev and Khas’minskiı̆, 1999)
2
b
ˆ
inf sup E f − f
≥ C2 (σ 2 ) a+b+1
for some C2 > 0.
fˆ f ∈Wb/(2a)
X
Remark 7. The above proposition in particular implies that all the methods in Table 1 are order
optimal with α = αpred when b/(2a) + 1/2 ≤ v0 in the minimax sense. This reproduces the result
in Chernousova and Golubev (2014) for the spectral cut-off method as a special case.
Note that the parameter choice αpred depends only on the data, and is completely independent of
the unknown truth. Thus, the ordered filter based regularization methods with αpred automatically
adapt to the unknown smoothness of the truth, and achieve the best possible rates up to a constant. In other words, the ordered filter based regularization methods with parameter choice αpred
are adaptively minimax optimal (cf. Pinsker, 1980) over a range of smoothness classes, and the
adaptation range is determined by the qualification index of the filter and the smoothing property of
the forward operator. Importantly, we point out that the order optimality here is in sharp contrast
to the Lepskiı̆-type balancing principle (Lepskiı̆, 1991), where one typically losses a log-factor in
the asymptotic convergence rates.
In addition, we stress again the price we pay for a-posteriori parameter choice αpred is a stronger
qualification assumption b/(2a) + 1/2 ≤ v0 , as the convergence rates of direct risk for a-priori
parameter choice only asks for b/(2a) ≤ v0 by Theorem 3.1 (i), see also Remark 6.
3.3.2 Exponentially ill-posed problems
Next we consider an exponentially ill-posed setting: The eigenvalues of T ∗ T satisfy
λk ≍ exp (−µk a )
k = 1, 2, . . .
for some a, µ > 0,
and the smoothness of the truth is characterized by (6) with
φ(x) := xb (− log x)c
for some b > 0, c ∈ R, or b = 0, c > 0.
P
2
−2ac
exp(2µbk a ).
That is, we assume that the truth lies in f ∈ X : ∞
k=1 wk fk ≤ 1 with wk ≍ k
It is easy to see that the assumptions of Theorem 3.2 are satisfied for all the regularization
methods listed in Table 1 provided that 0 < a < 1/2, or a = 1/2 and µ is sufficiently small.
Hence, if in addition b + 1/2 < v0 , or b = v0 , c ≤ 0 (recall that v0 is the qualification index),
10
we can obtain certain error bounds by means of Theorem 3.3. However, it turns out that our
bounds are too rough to guarantee order optimality. In fact, in case of b = 0, the error bound on
the right-hand side of (15) even diverges as the noise level σ tends to 0. In summary, our oracle
inequality is applicable for exponentially ill-posed problems as discussed here, but is not strong
enough to derive rates of convergence or even show optimality of the investigated parameter choice
αpred .
4 A general methodology for proving oracle inequalities of the
form (15)
To prove the oracle inequality (15), we will proceed as follows. First, we obtain bounds for general
moments of the prediction risk (cf. Corollary 4.1). Secondly, we proceed with a standard estimate
for the (deterministic) bias of the estimator fˆα under the smoothness Assumption 1 (cf. Lemma
4.2). Finally, we prove a comparison Lemma on the variance terms (cf. Lemma 4.3). Putting these
three ingredients together, the proof of Theorem 3.2 is then straightforward. Rephrased differently,
if these three ingredients are given for some parameter choice rule ᾱ, an oracle inequality similar
to (15) can be derived.
4.1 Main ingredients
We start with bounds for general moments of the prediction risk.
Corollary 4.1 (General moments of prediction risk). Assume the same setting as Theorem 2.1,
and let Ψ : R+ → R+ be an increasing and continuously differentiable function with Ψ(0) = 0 such
that
r
Z ∞
x
dx < ∞
(16)
Ψ′ (x) exp −C2
2
1
with C2 being the same as in Theorem 2.1. Then
2r(αo , f )
r(αpred , f )
≤Ψ
+ CΨ
E Ψ
σ2
σ2
for all f ∈ X ,
where CΨ is a constant depending only on Ψ.
Proof. Let ro := r(αo , f ). Then
Z ∞
r(αpred , f )
r(αpred , f )
′
E Ψ
=
Ψ (x)P
≥ x dx
σ2
σ2
0
r
Z 2ro /σ2
Z ∞
ro
≤
Ψ′ (x) dx + C1
Ψ′ (x) exp −C2 x − 2 dx
σ
0
2ro /σ2
r
Z ∞
x
2ro
dx
+ C1
Ψ′ (x) exp −C2
≤Ψ
2
σ
2
2ro /σ2
2r(αo , f )
≤Ψ
+ CΨ .
σ2
[by Fubini’s theorem]
[by Theorem 2.1]
This concludes the proof.
Remark 8. A simple example is Ψ(x) := xθ for some θ > 0. By Corollary 4.1, it leads to
E r(αpred , f )θ ≤ 2θ r(αo , f )θ + Cθ σ 2θ ,
where the second term on the right hand side is typically negligible compared to the first as σ tends
to 0.
11
Based on the smoothness Assumption 1, we can prove the following estimate for the bias of fˆα :
Lemma 4.2 (Source condition). Under Assumption 1 (i) and (iii), it holds that
!
∞
∞
X
1 X
2 2
2 −1
2 2
(1 − sα (λk )) fk ≤ C ψ
λk (1 − sα (λk )) fk ,
C2
k=1
k=1
with C being the same constant in (6).
Proof. Based on spectral analysis, condition (6) can be equivalently written as
fk = φ(λk )wk ,
k = 1, 2, . . . ,
with
∞
X
k=1
wk2 ≤ C 2 .
p
Thus, fk2 ≤ C 2 φ(λk )2 and then φ−1 ( C −2 fk2 ) ≤ λk . By Jensen’s inequality, we have
ψ
∞
∞
X
X
1
1 2
2
2
2
(1 − sα (λk ))
≤
ψ
f
(1
−
s
(λ
))
f
α k
C2 k
C2 k
k=1
k=1
∞
1 X
2
≤ 2
λk fk2 (1 − sα (λk )) .
C
k=1
The assertion follows by applying ψ −1 to the above inequality.
Finally, we need certain comparison relations, which are required for bounding the variance part
of the risk:
Lemma 4.3 (Comparison). Let qα be a filter, and Assumptions 2 and 3 hold. Then there are
positive constants C1 and C2 such that for every α ∈ A
!
!
∞
∞
∞
∞
X
X
X
X
2
2
4
2
2
sα (λk ) ,
λk qα (λk ) ≤ C2 Ψ2 Cq
and
sα (λk )
λk qα (λk ) ≤ C1 Ψ1 Cq
k=1
k=1
k=1
k=1
with Cq being the same as in Assumption 2 (iii), and the functions Ψ1 and Ψ2 are defined as
Ψ1 (x) =
x
,
(x)
S −1
Ψ2 (x) =
x
2,
(S −1 (x))
x > 0.
(17)
Proof. To ease notation, let C be a generic positive constant, whose value may vary from place to
place. We consider two separate cases: α is sufficiently small, and α 6ց 0.
Case I: α is sufficiently small. Let cq be the same as in Assumption 3. Then as α ց 0
∞
X
k=1
2
sα (λk ) = −
≥−
Z
∞
2
t qα (t) dΣ(t) ≥ −
0
Z
2
∞
Z
∞
t2 qα (t)2 dΣ(t)
α
α2 qα (α)2 dΣ(t)
[by Assumption 3 (i)]
α
≥ −c2q
Z
∞
α
dΣ(t) = c2q Σ(α)
[by Assumption 3 (ii)]
= c2q S(α).
[by Assumption 2 (i)]
It follows from Assumption 2 (ii) and (Bissantz et al., 2007, Lemma 12) that as α ց 0
Z α
Z α
Z α1
1
dS(t) ≤ CαS(α).
t dS(t) ≤ CαS(α) − Cα2
S(t) dt = αS(α) −
t
α
0
0
12
(18)
(19)
Then as α ց 0 we have
∞
X
k=1
2
λk qα (λk ) = −
Z
∞
tqα (t)2 dΣ(t)
0
Z α
Z ∞
1
t
dΣ(t)
−
dΣ(t)
≤C −
2
t
α
0 α
Z α
Z α
α
1
1
1
t
Σ(t) dt + Σ(α) = C 2
Σ(t) dt
≤ C − 2 Σ(t) + 2
α
α 0
α
α 0
0
Z α
1
=C 2
S(t) dt
α 0
S(α)
,
≤C
α
[by Definition 1 (i)]
[by Assumption 2 (i)]
[by (19)]
and similarly
∞
X
λ2k qα (λk )4
k=1
=−
Z
∞
t2 qα (t)4 dΣ(t)
0
Z α 2
Z ∞
1
t
dΣ(t) −
dΣ(t)
≤C −
4
t2
α
0 α
2
Z
Z α
α
α
t
1
1
1
≤ C − 4 Σ(t) + 4
2tΣ(t) dt + 2 Σ(α) ≤ C 3
Σ(t) dt
α
α 0
α
α 0
0
Z α
1
=C 3
S(t) dt
α 0
S(α)
≤C 2 .
α
[by Definition 1 (i)]
[by Assumption 2 (i)]
[by (19)]
These together with (18) prove the assertion if α is sufficiently small, by noticing that Cq > 1/c2q .
Case II: α 6ց 0. Note that α 6ց 0 and the boundedness of A imply that 1/α is bounded below
from 0, and above from ∞. Then by Assumption 3 it holds that
∞
X
k=1
∞
sα (λk )2 ≥ C
1X
λk ≥ C.
α
k=1
By Definition 1 (i) we have
∞
X
k=1
λk qα (λk )2 ≤ C
∞
1 X
λk ≤ C,
α2
and
k=1
∞
X
k=1
λ2k qα (λk )4 ≤ C
∞
1 X
λk ≤ C.
α3
k=1
Thus the assertion clearly holds for α 6ց 0.
Combining the above two cases concludes the proof.
4.2 Proof of Theorem 3.2
Now we are in position to prove Theorem 3.2. To obtain bounds, we will split the variance part
by exploiting a technique from Golubev (2010) dealing with ordered processes:
Lemma 4.4 (Ordered processes (Golubev, 2010)). Assume that a sequence of functions ck : A →
R, k = 1, 2, . . ., with A ⊆ R+ , satisfy that
κ(α) :=
∞
X
k=1
2
ck (α)
!1/2
13
<∞
for α ∈ A,
and that κ is continuous and strictly monotone on A. Define (x)+ := max{x, 0}, and
ζ(α) :=
∞
X
k=1
ck (α)(ξk2 − 1)
iid
with ξk ∼ N (0, 1).
Then there exists a constant C such that
C
2
E sup ζ(α) − xκ(α) + ≤
x
α∈A
for all x > 0.
Proof. Note that ζ(α) almost surely finite since κ(α) < ∞. In the terminology of Golubev (2010),
ζ(α) is an ordered process, but with a possibly nonzero mean. The assertion can be proven in
exactly the same way as (Golubev, 2010, Lemmas 1 and 2).
Proof of Theorem 3.2. Consider the bias-variance decomposition
h
i
h
i
2
2
2
ˆ
E fαpred − f
≤ 2E (qαpred (T ∗ T )T ∗ T − I)f X + 2σ 2 E qαpred (T ∗ T )T ∗ ξ X
X
"∞
#
"∞
#
X
X
2 2
2
2 2
= 2E
1 − sαpred (λk ) fk + 2σ E
λk qαpred (λk ) ξk .
k=1
(20)
k=1
For the first term (i.e. bias part) in (20), we have for all f ∈ W
#
"∞
#
"
∞
1 X
X
2
2 2
−1
2
λk (1 − sαpred (λk )) fk
E
≤ C1 E ψ
fk 1 − sαpred (λk )
C1
k=1
k=1
1
−1
r(αpred , f )
≤ C1 E ψ
C1
1
≤ C1 ψ −1
E [r(αpred , f )]
C1
2
≤ C1 ψ −1
r(αo , f ) + C2 σ 2 .
C1
[by Lemma 4.2]
[by (11)]
[by Jensen’s inequality]
[by Remark 8]
For the second term (i.e. variance part) in (20), we further split it into two terms
"∞
#
"∞
#
"∞
#
X
X
X
2
2
2 2
2
2 2
2
σ E
λk qαpred (λk ) ξk = σ E
λk qαpred (λk ) (ξk − 1) . (21)
λk qαpred (λk ) + σ E
k=1
k=1
k=1
Note that Ψ′1 (x) ≤ S −1 (x)Ψ′2 (x) for the functions Ψ1 and Ψ2 in (17), so by monotonicity of S and
Assumption 2 (iii), condition (16) is satisfied with Ψ(x) = Ψ1 (Cq x) as well. Then, for the first
term in (21), it holds that
!#
"∞
#
"
∞
X
X
2
2
2
2
[by Lemma 4.3]
sαpred (λk )
σ E
λk qαpred (λk ) ≤ C3 σ E Ψ1 Cq
k=1
k=1
r(αpred , f )
2
≤ C3 σ E Ψ1 Cq
σ2
r(αo , f )
2
≤ C3 σ Ψ1 2Cq
+ C4
σ2
r(α , f )
+ C3 C4 σ 2 .
o
= 2C3 Cq
S −1 2Cq r(ασo2,f )
14
[by (11)]
[by Corollary 4.1]
For the second term in (21), we apply Lemma 4.4 with ck (α) = λk qα (λk )2 , which leads to (with
ζ and κ defined in Lemma 4.4)
"∞
#
h
X
i
2
2 2
σ E
λk qαpred (λk ) (ξk − 1) ≤ σ 2 E ζ(αpred ) − xκ(αpred )2 + + σ 2 E xκ(αpred )2
k=1
≤ σ E sup ζ(α) − xκ(α) + + σ 2 E xκ(αpred )2
α∈A
"∞
#
X
C
5
+ xσ 2 E
λ2k qαpred (λk )4
for all x > 0.
≤ σ2
x
2
2
k=1
We minimize the right hand side of the above equation over x > 0, and then obtain
σ2 E
"
∞
X
k=1
#
λk qαpred (λk )2 (ξk2 − 1) ≤ 2σ 2
≤ C6 σ
C5 E
2
"
"
∞
X
λ2k qαpred (λk )4
k=1
E Ψ2
Cq
∞
X
k=1
#!1/2
2
sαpred (λk )
!#!1/2
[by Lemma 4.3]
1/2
r(αpred , f )
≤ C6 σ 2 E Ψ2 Cq
[by (11)]
σ2
1/2
r(αo , f )
≤ C6 σ 2 Ψ2 2Cq
[by Corollary 4.1]
+
C
7
σ2
p
σ r(αo , f )
1/2
+ C6 (C7 )1/2 σ 2 .
≤ C6 (2Cq )
S −1 2Cq r(ασo2,f )
Combining all these estimates concludes the proof.
5 Numerical simulations
In this section, we will investigate the behavior of the following parameter choice methods by
means of a simulation study:
h
i
(i) The oracle parameter choice αor = argmin E fˆα − f † , which is not available in practice
α∈A
and evaluated here for comparison only,
(ii) the a-posteriori parameter choice rule which is investigated in this paper, given by αpred =
argmin r̂ (α, Y ) with r̂ (α, Y ) as in (3),
α∈A
(iii) and the Lepskiı̆-type balancing principle originally introduced by Lepskiı̆ (1991), and was
further developed for usage in statistical inverse problems by Bauer and Hohage (2005);
Mathé (2006); Mathé and Pereverzev (2006); Werner and Hohage (2012). Given estimators
fˆα1 , ..., fˆαn for a suitably fine grid α1 < ... < αn , it consists in choosing
n
o
p
fˆαj − fˆαk
j ∗ = max 1 ≤ j ≤ n
≤ 4σ Trace (qαk (T ∗ T ) T ∗ ) for all 1 ≤ k ≤ j .
X
(22)
p
and αLEP := αj ∗ . Note that the term σ Trace (qα (T ∗ T ) T ∗ ) is in fact an estimator for the
standard deviation of fˆα . For an explanatory derivation of this choice we refer to Mathé
(2006). Unfortunately, the computation of αLEP is computationally expensive (see e.g. our
simulations below).
15
In all above methods, the set of considered parameters is
n
j
ko
−1
A = σ 2 · rk k = 0, 1, ..., (log (r)) log σ −2 kT ∗ T k
for some r > 1, i.e. it discretizes the range of possible α’s σ 2 , kT ∗ T k in a logarithmically
equispaced way. In our simulations we use r = 1.2. We also tried different values of r which did
not influence the results significantly. Note that it can readily be seen from the error decomposition
(11) that the discrete parameter set A is – under appropriate conditions on the filter which are
satisfied by all filters in Table 1 – able to resemble the optimal behavior of a continuous parameter
set σ 2 ≤ α < ∞ up to a constant depending on r.
Let us briefly comment on the implementation of the parameter choice rule αpred . Even though
the minimization is not performed over the continuum α ∈ A but over a discrete set here, the
computation of αpred can be numerically challenging. In Figure 1 we depict the function α 7→
r̂ (α, Y ) in an example using Tikhonov regularization, which shows that the function is quite flat
around its absolute minimum. We observed this behavior in many situations, especially if σ is
not too small. Finally we mention that the evaluation of the trace operator in r̂ (α, Y ) can be
expensive, but this can be overcome by different techniques (cf. (Engl et al., 1996, Sec. 9.4) or
(Vogel, 2002, Sec. 7.1)).
10−1
r̂ (α, Y )
10−3
10−5
10−7
σ = 10−2
σ = 10−4
σ = 10−6
10−9
10−12 10−11 10−10
10−9
10−8
10−7
α
10−6
10−5
10−4
10−3
10−2
Figure 1: The function α 7→ r̂ (α, Y ) for different noise levels. The red crosses mark the actual
minima. The operator T is as in Section 5.1, and f is as in Example 1. The chosen
regularization method is Tikhonov regularization.
5.1 Convergence rates
At first we investigate the empirical rate of convergence in a mildly ill-posed situation. Therefore
we consider the linear integral operator T : L2 ([0, 1]) → L2 ([0, 1]) defined by
(T f ) (x) =
Z1
x ∈ [0, 1]
k (x, y) f (y) dy,
0
with kernel k (x, y) = min {x · (1 − y) , y · (1 − x)} , x, y ∈ [0, 1], i.e. (T f )′′ = −f for all f ∈
L2 ([0, 1]). Obviously, the eigenvalues λk of T ∗ T satisfy λk ∼ k −4 .
1
3
We discretize T by choosing equidistant points x1 = 2n
, x2 = 2n
, . . . , xn = 2n−1
2n and using the
composite midpoint rule
(T f ) (x) =
Z1
0
n
k (x, y) f (y) dy ≈
16
1X
k (x, xi ) f (xi )
n i=1
on the grid points x = xj , 1 ≤ j ≤ n. To avoid an inverse crime, the exact data T f is always
calculated analytically.
We consider two different scenarios varying in the smoothness of the unknown solution f :
Example 1. As the first example, we consider the continuous function
(
x
if 0 ≤ x ≤ 12 ,
f (x) =
1 − x if 12 ≤ x ≤ 1.
It can readily be seen by straightforward computations that the Fourier coefficients fk of f are
given by
k
(−1) − 1
.
fk =
4π 3 k 2
3
Consequently f ∈ W3/8−ε and the optimal rate of convergence is O σ 4 −ε for any ε > 0, see
Proposition 3.4.
Example 2. In the second example we choose
(
1 if 14 ≤ x ≤ 34 ,
f (x) =
0 else.
As this function can be written as the derivative of functions as in the first example, it is clear
1
that f ∈ H 2 −ε ([0, 1]) for any ε > 0. This is also evident by the fact that the Fourier coefficients
fk of f are given by
k
(−1) sin πk
2
.
fk =
2π 2 k
1
Consequently, f ∈ W1/8−ε and the optimal rate of convergence is O σ 3 −ε for any ε > 0, see
Proposition 3.4.
In Figures 2-4 we plot several empirical risks against the noise level σ ∈ 2−15 , ..., 2−25 . The
optimal rate of convergence is also indicated. We consider spectral cut-off (cf. Figure 2), Tikhonov
regularization (cf. Figure 3), and Showalter regularization (cf. Figure 4). Using Monte Carlo
4
simulations with
per noise level we compute empirical
versions of the oracle risk
10 experiments
2
2
Ror (σ) := E fˆαor − f †
, the prediction risk Rpred (σ) := E fˆαpred − f †
, and the Lepskiı̆
2
2
2
risk RLEP (σ) := E fˆαLEP − f †
.
2
In all plots we find a good agreement of our theoretical predictions and the empirical results.
Compared with the Lepskiı̆-type balancing principle, it seems that αpred performs order-optimal
with a slightly smaller constant. The loss of a log-factor by using αLEP cannot be visible in such
a small simulation study.
5.2 Efficiency simulations
Besides the convergence rates simulations above we also want to numerically infer on the constant
in the oracle inequality which will be done by efficiency simulations. Therefore (inspired by
Bauer and Lukas (2011); Chernousova and Golubev (2014)) we consider the following setup. The
forward operator is a 300 × 300 diagonal matrix with singular values λ (k) = k −a with a fixed
parameter a > 0. Then we repeat the following experiment 104 times : Given
a parameter ν
we generate a random ground truth f ∈ R300 by f (k) = ±k −ν · 1 + N 0, 0.12 where the sign
is independent and uniformly distributed for each component.
From this ground truth, data is
generated according to Y (k) = λ (k) · f (k) + N 0, σ 2 where the noise is again independent in
each component. Based on the data compute empirical versions of the oracle risk Ror (σ), the
17
Example 2
Example 1
101
10−1
10−2
10−3
100
Ror (σ)
Rpred (σ)
RLEP (σ)
10−4
Ror (σ)
Rpred (σ)
RLEP (σ)
1
3
σ3
σ4
10−5 −8
10
10−7
10−6
σ
10−5
10−1 −8
10
10−4
10−7
10−6
σ
10−5
10−4
Figure 2: Simulation results for spectral cut-off.
Example 1
Example 2
101
10−1
10−2
10−3
100
Ror (σ)
Rpred (σ)
RLEP (σ)
10−4
Ror (σ)
Rpred (σ)
RLEP (σ)
3
1
σ4
10−5 −8
10
10−7
10−6
σ
10−5
σ3
10−1 −8
10
10−4
10−7
10−6
σ
10−5
10−4
Figure 3: Simulation results for Tikhonov regularization.
Example 1
Example 2
101
10−1
10−2
10−3
100
Ror (σ)
Rpred (σ)
RLEP (σ)
10−4
Ror (σ)
Rpred (σ)
RLEP (σ)
1
3
σ4
10−5
10−8
10−7
10−6
σ
10−5
σ3
10−1
10−4
10−8
10−7
10−6
σ
10−5
10−4
Figure 4: Simulation results for Showalter regularization.
prediction risk Rpred (σ), and the Lepskiı̆ risk RLEP (σ) for Tikhonov regularization. In Figure 5
we depict the fractions of the oracle risk with the different a-posteriori risks for various parameters
ν and a to compare the average behavior of these parameter choice methods.
In conclusion we empirically find that both choices αpred and αLEP perform nearly as good as
the oracle choice if σ is not too small. Comparing the performance of αpred and αLEP it seems
18
1
1
0.9
0.9
a = 3, ν = 0.3
0.8
0.7
a = 4, ν = 0.4
0.8
ror (σ)
rpred (σ)
ror (σ)
rLEP (σ)
0
1
2
3
4
σ
0.7
5
0
1
2
3
5
·10−2
1
0.9
0.9
a = 5, ν = 0.5
0.8
0.7
4
σ
·10−2
1
ror (σ)
rpred (σ)
ror (σ)
rLEP (σ)
a = 6, ν = 0.6
0.8
ror (σ)
rpred (σ)
ror (σ)
rLEP (σ)
0
1
2
3
4
σ
0.7
5
ror (σ)
rpred (σ)
ror (σ)
rLEP (σ)
0
1
2
3
σ
·10−2
4
5
·10−2
Figure 5: Efficiency simulations for Tikhonov regularization with different smoothness parameters
a and ν.
that αLEP behaves better for small values of σ, even though this has not been observed in Figures
2–4. Nevertheless, the computational effort for αpred is significantly smaller: in our efficiency
simulations around 90% of the computation time were spent for computing αLEP in (22).
6 Conclusion and outlook
In this study we have analyzed a parameter choice method for filter-based regularization methods
applied to linear statistical inverse problems. Therefore we have proven an oracle inequality, which
generalizes the one from Chernousova and Golubev (2014) to general ordered filters satisfying
weak assumptions (cf. Definition 1 and Assumption 3). From this oracle inequality we derived
convergence rates of the investigated parameter choice, which are actually order optimal in a wide
range of cases.
We point out that our techniques do not seem to be limited to the specific choice of α investigated
here. Quite the contrary, we provide a general strategy to prove an oracle inequality, which might
be used for other choices as well. If one would consider a different parameter choice rule, then
an oracle inequality can be carried out the same way as in this study provided the following
ingredients are available:
• general moment bounds for the prediction risk as in Corollary 4.1 (in our case based on the
exponential bounds proven by Kneip (1994)),
• a deterministic optimality result as in Lemma 4.2,
• estimates for the behavior of the regularization algorithm as in Lemma 4.3.
Note that the latter two assumptions do not rely on the parameter choice rule at all but only
on the interplay of the operator, the regularization method, and the source condition. Consequently, whenever the general moments can be bounded, the analysis can basically be carried over
from standard results, see e.g. Bellec and Tsybakov (2016) for a useful deviation inequality for
regularization methods with convex penalties.
19
Even though we have always assume that the noise level σ is known, the generalization to
unknown σ is straightforward. For the choice of parameter αpred , we simply replace σ by a proper
estimator σ̂ in (12). The only affected part in our argumentation is the Kneip’s deviation bound
in Theorem 2.1, which still holds if we further assume the smoothness of T f (this is usually the
case due to the blessing of ill-posedness), see (Kneip, 1994, Section 6) for further details.
Another possible generalization concerns the errors in (1). If ξ is such that the ξk ’s in (5) are
independent sub-Gaussian errors, then there are two crucial parts of the proofs which have to be
generalized: Theorem 2.1 and Lemma 4.4. In fact it turns out that both also hold for independent sub-Gaussian errors (see Kneip, 1994; Golubev, 2010), so the whole analytical methodology
remains valid in such a case as well.
The general analytical strategy advocated in this paper, of course, has its own limitations, as
the resulting oracle inequality might turn out to be inadequate or even trivial in certain cases
(see Section 3.3.2 for instance). Future questions include generalizations to exponentially ill-posed
problems, nonlinear problems and noise models with heavier tails.
Acknowledgements
FW wants to thank Yu. Golubev for pointing his attention to the inspiring paper (Chernousova and Golubev,
2014) during a visit in Göttingen. FW gratefully acknowledges financial support by the German Research Foundation DFG through subproject A07 of CRC 755, and HL acknowledges support through RTG 2088, subproject B2, and the National Nature Science Foundation of China
(61571008, 61402495).
References
Abramovich, F. and Silverman, B. W. (1998). Wavelet decomposition approaches to statistical
inverse problems. Biometrika, 85:115–129.
Bakushinskiı̆, A. B. (1984). Remarks on the choice of regularization parameter from quasioptimality and relation tests. Zh. Vychisl. Mat. i Mat. Fiz., 24(8):1258–1259.
Bauer, F. and Hohage, T. (2005). A Lepskij-type stopping rule for regularized Newton methods.
Inverse Probl., 21(6):1975.
Bauer, F. and Lukas, M. A. (2011). Comparing parameter choice methods for regularization of
ill-posed problems. Math. Comput. Simulation, 81(9):1795–1841.
Bellec, P. C. and Tsybakov, A. B. (2016). Bounds on the prediction error of penalized least squares
estimators with convex penalty. arXiv:1609.06675.
Birgé, L. and Massart, P. (2001). Gaussian model selection. J. Eur. Math. Soc. (JEMS), 3(3):203–
268.
Birgé, L. and Massart, P. (2007). Minimal penalties for Gaussian model selection. Probab. Theory
Related Fields, 138(1-2):33–73.
Bissantz, N., Hohage, T., Munk, A., and Ruymgaart, F. (2007). Convergence rates of general
regularization methods for statistical inverse problems and applications. SIAM J. Numer. Anal.,
45(6):2610–2636.
Blanchard, G., Hoffmann, M., and Reiß, M. (2016). Optimal adaptation for early stopping in
statistical inverse problems. arXiv:1606.07702.
Candès, E. J., Sing-Long, C. A., and Trzasko, J. D. (2013). Unbiased risk estimates for singular
value thresholding and spectral estimators. IEEE Trans. Signal Process., 61(19):4643–4657.
20
Cavalier, L. (2011). Inverse problems in statistics. In Inverse problems and high-dimensional
estimation, volume 203 of Lect. Notes Stat. Proc., pages 3–96. Springer, Heidelberg.
Cavalier, L., Golubev, G. K., Picard, D., and Tsybakov, A. B. (2002). Oracle inequalities for
inverse problems. Ann. Statist., 30(3):843–874. Dedicated to the memory of Lucien Le Cam.
Cavalier, L. and Golubev, Y. (2006). Risk hull method and regularization by projections of illposed inverse problems. Ann. Statist., 34(4):1653–1677.
Cavalier, L., Golubev, Y., Lepski, O., and Tsybakov, A. (2003). Block thresholding and sharp
adaptive estimation in severely ill-posed inverse problems. Teor. Veroyatnost. i Primenen.,
48(3):534–556.
Chernousova, E. and Golubev, Y. (2014). Spectral cut-off regularizations for ill-posed linear
models. Math. Methods Statist., 23(2):116–131.
Cohen, A., Hoffmann, M., and Reiß, M. (2004). Adaptive wavelet Galerkin methods for linear
inverse problems. SIAM J. Numer. Anal., 42(4):1479–1501.
Davies, A. R. and Anderssen, R. S. (1986). Improved estimates of statistical regularization parameters in Fourier differentiation and smoothing. Numer. Math., 48(6):671–697.
Deledalle, C.-A., Vaiter, S., Fadili, J., and Peyré, G. (2014). Stein Unbiased GrAdient estimator
of the Risk (SUGAR) for multiple parameter selection. SIAM J. Imaging Sci., 7(4):2448–2487.
Dette, H., Munk, A., and Wagner, T. (1998). Estimating the variance in nonparametric regressionwhat is a reasonable choice? J. R. Stat. Soc. Ser. B. Stat. Methodol., 60(4):751–764.
Donoho, D. L. (1995). Nonlinear solution of linear inverse problems by wavelet-vaguelette decomposition. Appl. Comput. Harmon. Anal., 2(2):101–126.
Engl, H., Hanke, M., and Neubauer, A. (1996). Regularization of Inverse Problems. Springer.
Goldenshluger, A. (1999).
5(5):907–925.
On pointwise adaptive nonparametric deconvolution.
Bernoulli,
Golub, G. H., Heath, M., and Wahba, G. (1979). Generalized cross-validation as a method for
choosing a good ridge parameter. Technometrics, 21(2):215–223.
Golubev, G. K. and Khas’minskiı̆, R. Z. (1999). A statistical approach to some inverse problems
for partial differential equations. Problemy Peredachi Informatsii, 35(2):51–66.
Golubev, Y. (2010). On universal oracle inequalities related to high-dimensional linear models.
Ann. Statist., 38(5):2751–2780.
Hall, P., Kay, J. W., and Titterinton, D. M. (1990). Asymptotically optimal difference-based
estimation of variance in nonparametric regression. Biometrika, 77:521–528.
Ingster, Y., Laurent, B., and Marteau, C. (2014). Signal detection for inverse problems in a
multidimensional framework. Math. Methods Statist., 23(4):279–305.
Ingster, Y. I., Sapatinas, T., and Suslina, I. A. (2012). Minimax signal detection in ill-posed
inverse problems. Ann. Statist., 40(3):1524–1549.
Johnstone, I. M., Kerkyacharian, G., Picard, D., and Raimondo, M. (2004). Wavelet deconvolution
in a periodic setting. J. R. Stat. Soc. Ser. B Stat. Methodol., 66(3):547–573.
Johnstone, I. M. and Silverman, B. W. (1991). Discretization effects in statistical inverse problems.
J. Complexity, 7:1–34.
Kneip, A. (1994). Ordered linear smoothers. Ann. Statist., 22(2):835–866.
21
Lepskiı̆, O. V. (1991). On a problem of adaptive estimation in Gaussian white noise. Theory
Probab. Appl., 35(3):454–466.
Li, K.-C. (1987). Asymptotic optimality for Cp , CL , cross-validation and generalized crossvalidation: discrete index set. Ann. Statist., 15(3):958–975.
Lucka, F., Proksch, K., Brune, C., Bissantz, N., Burger, M., Dette, H., and Wübbeling, F. (2017).
Risk estimators for choosing regularization parameters in ill-posed problems – properties and
limitations. arXiv:1701.04970.
Luisier, F., Blu, T., and Unser, M. (2007). A new sure approach to image denoising: Interscale
orthonormal wavelet thresholding. IEEE Trans. Image Process., 16(3):593–606.
Lukas, M. A. (1993). Asymptotic optimality of generalized cross-validation for choosing the regularization parameter. Numer. Math., 66(1):41–66.
Lukas, M. A. (1995). On the discrepancy principle and generalised maximum likelihood for regularisation. Bull. Austral. Math. Soc., 52(3):399–424.
Mair, B. A. and Ruymgaart, F. H. (1996). Statistical inverse estimation in Hilbert scales. SIAM
J. Appl. Math., 56(5):1424–1444.
Mallows, C. L. (1973). Some comments on Cp . Technometrics, 15(4):661–675.
Mathé, P. (2006). The Lepskiı̆ principle revisited. Inverse Probl., 22(3):L11–L15.
Mathé, P. and Hofmann, B. (2008). How general are general source conditions? Inverse Problems,
24(1):015009.
Mathé, P. and Pereverzev, S. V. (2001). Optimal discretization of inverse problems in Hilbert
scales. Regularization and self-regularization of projection methods. SIAM J. Numer. Anal.,
38(6):1999–2021.
Mathé, P. and Pereverzev, S. V. (2003). Geometry of linear ill-posed problems in variable Hilbert
scales. Inverse Probl., 19(3):789–803.
Mathé, P. and Pereverzev, S. V. (2006). Regularization of some linear ill-posed problems with
discretized random noisy data. Math. Comp., 75(256):1913–1929 (electronic).
Morozov, V. A. (1966). On the solution of functional equations by the method of regularization.
Soviet Math. Dokl., 7:414–417.
O’Sullivan, F. (1986). A statistical perspective on ill-posed inverse problems.
1(4):502–527. With comments and a rejoinder by the author.
Statist. Sci.,
Pinsker, M. S. (1980). Optimal filtration of square-integrable signals in Gaussian noise. Probl.
Inf. Transm., 16(2):52–68. (Russian).
Rice, J. (1984). Bandwidth choice for nonparametric regression. Ann. Statist., 12(4):1215–1230.
Rieder, A. (2005). Runge-Kutta integrators yield optimal regularization schemes. Inverse Probl.,
21(2):453–471.
Stein, C. M. (1981). Estimation of the mean of a multivariate normal distribution. Ann. Statist.,
9(6):1135–1151.
Tsybakov, A. (2000). On the best rate of adaptive estimation in some inverse problems. C. R.
Acad. Sci. Paris Sér. I Math., 330(9):835–840.
Vogel, C. R. (1986). Optimal choice of a truncation level for the truncated SVD solution of linear
first kind integral equations when data are noisy. SIAM J. Numer. Anal., 23(1):109–117.
22
Vogel, C. R. (2002). Computational Methods for Inverse Problems, volume 23 of Frontiers in
Applied Mathematics. Society for Industrial and Applied Mathematics (SIAM), Philadelphia,
PA.
Wahba, G. (1977). Practical approximate solutions to linear operator equations when the data
are noisy. SIAM J. Numer. Anal., 14(4):651–667.
Wang, Y.-Q. and Morel, J.-M. (2013). SURE guided Gaussian mixture image denoising. SIAM J.
Imaging Sci., 6(2):999–1034.
Werner, F. (2015). On convergence rates for iteratively regularized Newton-type methods under
a Lipschitz-type nonlinearity condition. J. Inverse Ill-Posed P., 23(1):75–84.
Werner, F. and Hohage, T. (2012). Convergence rates in expectation for Tikhonov-type regularization of Inverse Problems with Poisson data. Inverse Probl., 28(10):104004.
23
| 10math.ST
|
REDUCING THE RANK OF A MATROID
arXiv:1211.4853v2 [cs.DS] 5 Sep 2015
GWENAËL JORET AND ADRIAN VETTA
Abstract. We consider the rank reduction problem for matroids: Given a matroid M and an
integer k, find a minimum size subset of elements of M whose removal reduces the rank of M
by at least k. When M is a graphical matroid this problem is the minimum k-cut problem,
which admits a 2-approximation algorithm. In this paper we show that the rank reduction
problem for transversal matroids is essentially at least as hard to approximate as the densest
k-subgraph problem. We also prove that, while the problem is easily solvable in polynomial
time for partition matroids, it is NP-hard when considering the intersection of two partition
matroids. Our proof shows, in particular, that the maximum vertex cover problem is NP-hard
on bipartite graphs, which answers an open problem of B. Simeone.
1. Introduction
Consider the well-known minimum k-cut problem: Given a graph G and an integer k, find a
minimum size subset of edges whose removal increases the number of connected components by
at least k. This problem is NP-hard, assuming k is part of the input, and several 2-approximation
algorithms have been developed for it over the years [14, 17, 18, 20]. Notice that the minimum
k-cut problem has a simple formulation in terms of matroids: Given a graph G and an integer
k, find a minimum size subset of elements of the graphical matroid of G whose removal reduces
its rank by at least k.
This observation motivates the study of the rank reduction problem in other classes of matroids. For example, is the rank reduction problem computationally hard and, if so, does it
admit approximation algorithms with good approximation guarantees (as is the case for graphical matroids)? Moreover, as we will see, many fundamental problems can be formulated in this
rank reduction framework.
In this paper, our focus is on the case of transversal matroids. First, we show that the rank
reduction problem in transversal matroids is roughly at least as hard to approximate as the
densest k-subgraph problem: Given a graph G and an integer k, find a subset of k vertices inducing a subgraph with a maximum number of edges. (Note that in our reduction the parameter
k is not necessarily the same one as in the rank reduction problem.) The densest k-subgraph
1
problem can be approximated to within a factor of O(n 4 + ) due to a recent breakthrough result
of [4]. Moreover, it is widely believed [1, 5, 10, 11] that its hardness is also close to this upper
bound – indeed, Bhaskara et al. [5] present nΩ(1) lower bounds for lift and project methods
based upon the Sherali-Adams and the Lassere hierarchies. We will show in particular that an
O(mε )-approximation algorithm for the rank reduction problem on transversal matroids (where
m denotes the number of elements) implies an O(n4ε )-approximation algorithm for the densest
k-subgraph (where n is the number of vertices).
Secondly, we prove that while the rank reduction problem is easily solvable in polynomial time
for partition matroids (a special class of transversal matroids), it is NP-hard when considering
the intersection of two partition matroids. Our proof shows in particular that the maximum
vertex cover problem—also known as the partial vertex cover problem—is NP-hard on bipartite
graphs. Here, one is given a graph and a positive integer k, and the goal is to find a set of
k vertices hitting as many edges as possible. The problem is obviously NP-hard on arbitrary
graphs since it contains the vertex cover problem as a special case. Whether it remained NPhard on bipartite graphs was an open problem of B. Simeone (see [15]). We note that we learned
1
after finishing this paper that Apollonio and Simeone [2] independently obtained a proof of this
result.
2. Preliminaries
In this section we give the necessary definitions and notations. All graphs and matroids in
this paper are finite, and “graph” will always mean an undirected simple graph. We use the
shorthands |G| and ||G|| for the number of vertices and edges of a graph G, respectively. We
denote by µ(G) the maximum size of a matching in G, which we call the matching number of
G.
A matroid M is a pair (E, I) where I is a family of subsets, called the independent sets, of
the ground set E satisfying the following three axioms:
• the empty set ∅ is independent;
• every subset of an independent is again independent, and
• if X and Y are two independent sets with |X| > |Y | then there exists x ∈ X \ Y such
that Y ∪ {x} is independent.
The inclusion-wise maximal independent sets are the bases of the matroid M ; as follows from
the third axiom the bases all have the same cardinality. The rank function of M is the function
r : 2E → N that assigns to each subset X of elements of E the maximum size r(X) of an
independent set contained in X, called the rank of X. In particular, r(E) is the cardinality of
a basis of M , which is called the rank of M .
The rank reduction problem for matroids is defined as follows: Given a positive integer k
and a matroid on a set A of elements with rank function r, the goal is to find a minimum size
subset X ⊆ A such that r(A \ X) 6 r(A) − k.
For example consider the case of graphical matroids: Given a graph G = (V, E), the graphical
matroid of G is obtained by taking E as ground set, and letting a subset F of edges be independent if and only if the corresponding subgraph is acyclic. Here the rank reduction problem
is the minimum k-cut problem.
As stated, here we study transversal matroids. A bipartite graph G with bipartition (A, B)
induces a matroid M as follows: The matroid M has A as ground set, and X ⊆ A is independent
in M if and only if there exists a matching of G covering X. The fact that this is indeed a
matroid is well-known; see for instance [19]. Any matroid M that can be obtained this way is
called a transversal matroid, and the bipartite graph G is said to be a model for M . Observe
that, letting r denote the rank function of M , the rank r(X) of X ⊆ A is equal to µ(G[X ∪ B]).
Also note that being a transversal matroid is a hereditary property, in the sense that for each
set X ⊆ A, taking the restriction I ∩ X of all independent sets I yields a transversal matroid
on ground set X.
A special case of transversal matroids are partition matroids. Here we are given a collection
E1 , . . . , Ep of disjoint sets and integers d1 , . . . , dp such that 0 6 di 6 |Ei | for each i ∈ {1, . . . , k}.
One can define a corresponding matroid with ground set E := E1 ∪ · · · ∪ Ek by letting X ⊆ E
be independent if and only if |Ei ∩ X| 6 di for each i ∈ {1, . . . , k}. Such a matroid is called a
partition matroid, with model ((E1 , d1 ), . . . , (Ep , dp )). This corresponds to a transversal matroid
on a bipartite graph G with bipartition (E, B), where B has di vertices that are adjacent to
all vertices in Ei , and none other, for each set Ei . Notice that partition matroids are also
hereditary.
Throughout, since we restrict ourselves to specific families of matroids, we assume that the
matroid is given concisely and not given explicitly as a set system in input. Specifically, a
corresponding model of the matroid is provided: a bipartite graph for a transversal matroid, a
graph for a graphical matroid, etc.
More generally, the rank reduction problem can be considered on the intersection of matroids.
Given two matroids M1 = (E, I1 ) and M2 = (E, I2 ) with common ground set E, the intersection
M1 ∩ M2 of M1 and M2 is the pair (E, I) where I is the family of sets X ⊆ E that are
independent in both M1 and M2 , which are said to be the independent sets of M1 ∩ M2 . While
2
the independence system (E, I) is not necessarily a matroid anymore, it enjoys several of the
nice properties of matroids (see [19]). In particular, letting as before the rank r(X) of X ⊆ E be
the maximum size of an independent set of M1 ∩ M2 contained in X, the rank r(E) of M1 ∩ M2
can be computed in polynomial time given access to the rank functions r1 and r2 of M1 and
M2 , respectively, by a classical result of Edmonds (see [19]). We examine the rank reduction
problem for the intersection of two partition matroids in Section 4.
3. Transversal Matroids
We start our investigation of the rank reduction problem with an easy observation, namely
that the problem can be solved in polynomial time if the input matroid is a partition matroid.
Theorem 3.1. The rank reduction problem can be solved in polynomial time on partition matroids.
Proof. Let M be a given partition matroid with model ((E1 , d1 ), . . . , (Ep , dp )) P
and rank function
r. Let E := E1 ∪ · · · ∪ Ep denote the ground set of M . Observe that r(E) = pi=1 di . Let k be
the given parameter for the rank reduction problem on M . We may assume that 1 6 k 6 r(E).
Let ci :=P
|Ei | − di for each i ∈ {1, . . . , p}. Given X ⊆ E, the rank r(E \ X) of the set E \ X is
equal to pi=1 min(|Ei \ X|, di ).
Let X ⊆ E be such that r(E \ X) 6 r(E) − k. Moreover, assume X is inclusion-wise minimal
with this property. Then, for each i ∈ {1, . . . , p}, either |Ei ∩ X| > ci + 1 or Ei ∩ X = ∅.
Moreover,
letting J be the subset
of indices i ∈ {1, . . . , p} such that Ei ∩ X 6= ∅, we have that
P
P
i∈J ci .
i∈J di > k and |X| = k +
P
Conversely, suppose J 0 ⊆ {1, . . . , p} is such that i∈J 0 di > k. S
Then choosing arbitrarily ci
elements ofP
Ei , for each i ∈ J 0 , plus k additional elements from i∈J 0 Ei gives a set X 0 with
|X 0 | = k + i∈J 0 ci such that r(E \ X 0 ) 6 r(E) − k.
Therefore, computing an optimal solution to the rank
reduces to the
P reduction problem
P
problem of finding a subset J ⊆ {1, . . . , p} such that i∈J di > k and i∈J ci is minimum.
Thus we obtain a knapsack problem. Moreover, as ci and di are at most |Ei |, they are of
polynomial size when encoded in unary. Thus the knapsack problem can be solved easily in
polynomial time using dynamic programming.
While the rank reduction problem admits a simple polynomial-time algorithm on partition
matroids, the problem turns out to be more difficult on the broader class of transversal matroids.
In fact, up to some degree, the problem can be viewed as a generalization of the densest ksubgraph problem. In the latter problem, one is given a graph G and a positive integer k, and
the aim is to find a subgraph H of G with |H| = k and ||H|| maximum. Towards this goal, we
consider a closely related problem, the minimum t-edge subgraph problem: Given a graph G
and a positive integer t, the goal is to find a subgraph H of G with ||H|| = t and |H| minimum.
We start by drawing a connection between the rank reduction problem on transversal matroids
and the minimum t-edge subgraph problem. Then we will extend the connection to the densest
k-subgraph problem.
Lemma 3.2. For each constant ε with 0 < ε < 1/2, every O(mε )-approximation algorithm
for the rank reduction problem on transversal matroids with m elements can be turned into an
O(n2ε )-approximation algorithm for the minimum t-edge subgraph problem on graphs with n
vertices.
Proof. Let G = (V, E) be an instance of the minimum t-edge subgraph problem. Let n := |V |.
Let V1 , V2 , . . . , Vn be n disjoint copies of V . Let E 0 be a disjoint copy of E. Let H be the
bipartite graph with bipartition (A, B) where A = V1 ∪ V2 ∪ · · · ∪ Vn ∪ E 0 and B = E, and where
u ∈ A is adjacent to v ∈ B if either u corresponds to a vertex of G that is incident to the edge
corresponding to v in G, or if u and v correspond to the same edge of G.
Let r denote the rank function of the transversal matroid induced by H on A; thus for X ⊆ A,
r(X) is the maximum size of a matching in H[X ∪ B]. Obviously, r(A) = |E|, since every v ∈ B
3
can be matched to its copy in A. Let m := |A| denote the number of elements of the transversal
matroid. Now consider the rank reduction problem on this matroid with k = t. Recall that a
feasible solution is a subset X ⊆ A such that r(A \ X) 6 r(A) − t = |E| − t.
As is well known, we have that r(A \ X) 6 r(A) − t for X ⊆ A if and only if there exists
Y ⊆ B such that |NH (Y ) \ X| 6 |Y | − t, where NH (Y ) denotes the set of vertices of H that
have a neighbor in Y .1 Such a set Y is said to be a witness for X. The set Y defines in turn a
corresponding subgraph GY of G consisting of all the edges of G included in Y , and the vertices
of G incident to those edges. By definition of H, the set NH (Y ) consists of the n copies of each
vertex of GY , along with the copies in E 0 of each edge of GY . Observe that any set X 0 obtained
by taking the n copies in A of each vertex of GY and t arbitrarily chosen edges of GY in A is
such that |NH (Y ) \ X 0 | = |Y | − t. Moreover, since |NH (Y ) \ X| 6 |Y | − t and X 0 ⊆ NH (Y ), it
follows that |X 0 | 6 |X|, that is, X 0 is a solution of size no greater than X and having the same
witness Y . Such a pair (X 0 , Y ) is called a canonical pair.
Now, if a canonical pair (X, Y ) is such that |Y | > t, then NH (Y ) \ X consists of exactly
|Y | − t > 0 edges of GY (or more precisely, their copies in E 0 ). For each u ∈ NH (Y ) \ X with
corresponding copy v in B, we have that X \ {u} is again a solution to the rank reduction
problem, with witness Y \ {v}, and of size smaller than X.
To summarize the above discussion, given an arbitrary set X 0 ⊆ A such that r(A \ X 0 ) 6
r(A) − t, one can in polynomial time compute a canonical pair (X, Y ) with |Y | = ||GY || = t
and |X| = n|GY | + t 6 |X 0 |.
Conversely, for every subgraph G0 ⊆ G with ||G0 || = t, there is a natural corresponding
canonical pair (X, Y ), where Y contains the copies in B of the t edges of G0 , and where X =
NH (Y ). Letting x∗ and j ∗ denote the size of an optimal solution for the rank reduction and
minimum t-edge subgraph problems, respectively, it follows that x∗ = nj ∗ + t.
Now suppose that the rank reduction problem admits a cmε -approximation algorithm, where
0 < ε < 1/2 and c > 0 are absolute constants. Letting (X, Y ) be a canonical pair with |Y | = t
obtained using this algorithm, we have
n|GY | + t = |X| 6 cmε x∗ = c(n2 + |E|)ε (nj ∗ + t) 6 c(2n2 )ε (nj ∗ + t)
and hence
t
c(2n2 )ε (nj ∗ + t) − t
2ε
∗
6 2cn
j +
6 2cn2ε (j ∗ + j ∗ ) = 4cn2ε j ∗ .
|GY | 6
n
n
∗ 2
∗
(In the last inequality we used the fact that (j ∗ )2 > t, and thus nt 6 (jn) 6 j nn = j ∗ .)
Therefore, GY is a t-edge subgraph whose order is within a 4cn2ε -factor of optimal.
As pointed out to us by an anonymous referee, the following lemma is implicit in the recent work of Chlamtac, Dinitz, and Krauthgamer [8] (in [8], the minimum t-edge subgraph
problem is called the smallest m-edge subgraph problem). We include a proof nevertheless, for
completeness.
Lemma 3.3. For each constant ε with 0 < ε < 1/2, every O(nε )-approximation algorithm for
the minimum t-edge subgraph problem can be turned into an O(n2ε )-approximation algorithm
for the densest k-subgraph problem.
1This is a consequence of Hall’s Marriage Theorem, as we now explain for completeness. Add t−1 new vertices
to the set A, yielding a set A0 , and make each of them adjacent to every vertex in B. Let H 0 be the bipartite
graph obtained from H in this manner. Let X ⊆ A. Then every matching M in H \ X with |M | 6 |B| − (t − 1)
can be extended to a matching M 0 of H 0 \ X with |M 0 | = |M | + t − 1. Conversely, every matching M 0 in H 0 \ X
with |M 0 | > t − 1 yields a matching M of H \ X with |M | > |M 0 | − (t − 1) by discarding the at most t − 1 edges
of M 0 incident to the vertices in A0 \ A. Hence, H \ X has a matching of size |B| − (t − 1)—or equivalently,
r(A \ X) > r(A) − (t − 1)—if and only if B can be completely matched in H 0 \ X. By Hall’s theorem, the latter
happens if and only if |NH 0 −X (Y )| > |Y | for every Y ⊆ B, which is equivalent to |NH (Y ) \ X| > |Y | − (t − 1) for
every Y ⊆ B. Therefore, r(A \ X) 6 r(A) − t if and only if there exists Y ⊆ B such that |NH (Y ) \ X| 6 |Y | − t.
4
Proof. Suppose that the minimum t-edge subgraph problem admits a c0 nε -approximation algorithm, which we denote A, where 0 < ε < 1 and c0 > 1 are absolute constants. Let G be an
instance of the densest k-subgraph problem. As before, we let n and m denote the number of
vertices and edges of G, respectively. We may assume n > k > 2. Since c0 > 1 and nε > 1, there
exists c with c0 6 c 6 2c0 such that cnε is an integer. We will consider the approximation factor
of A to be cnε in what follows, to avoid cumbersome floors and ceilings in the calculations.
Run algorithm A on G with t = 1, 2, . . . , m. Let Hi be the ith subgraph returned by the
algorithm. Clearly, we may suppose that |Hi | 6 |Hi+1 | for each i ∈ {1, . . . , m − 1}.
Let z ∗ denote the number of edges in an optimal solution to the densest k-subgraph problem
on G. If |Hm | 6 kcnε then let t0 := m, otherwise let t0 be the index in {1, . . . , m − 1} such
that |Ht0 | 6 kcnε and |Ht0 +1 | > kcnε . Since algorithm A is a cnε -approximation algorithm, and
since either t0 = m or |Ht0 +1 | > kcnε , it follows that every subgraph of G with exactly k vertices
has at most t0 edges, that is, z ∗ 6 t0 .
Let q := d|Ht0 |/bk/2ce. Observe that 2 6 q 6 3cnε . Let V1 , V2 , . . . , Vq be a partition
of the vertex set of Ht into q subsets with |V1 | = |V2 | = · · · = |Vq−1 | = bk/2c and |Vq | =
|Ht0 | − (q − 1)bk/2c 6 bk/2c. Let (Vi , Vj ) be a pair with i 6= j such that ||Ht0 [Vi ∪ Vj ]|| is
maximized. By the pigeonhole principle,
||Ht0 [Vi ∪ Vj ]|| >
t0
||Ht0 ||
=
q
q >
2
2
t0
z∗
>
.
(3cnε )2
9c2 n2ε
If |Vi ∪ Vj | = k, then let H := Ht0 [Vi ∪ Vj ]. If, on the other hand, |Vi ∪ Vj | < k, then let
H := Ht0 [Vi ∪ Vj ∪ X] where X is an arbitrary subset of V (Ht0 ) \ (Vi ∪ Vj ) of size k − |Vi ∪ Vj |.
Thus in both cases |H| = k and ||H|| > z ∗ /9c2 n2ε . Hence, H is a solution to the densest
k-subgraph problem on G whose number of edges is within a 9c2 n2ε -factor of the optimum.
Combining Lemma 3.2 and 3.3 gives:
Theorem 3.4. For each constant ε with 0 < ε < 1/4, every O(mε )-approximation algorithm
for the rank reduction problem on transversal matroids with m elements can be turned into an
O(n4ε )-approximation algorithm for the densest k-subgraph problem on graphs with n vertices.
As discussed in the introduction, the best approximation algorithm for the densest k-subgraph
problem currently known has an approximation ratio of O(n1/4+δ ) for any fixed δ > 0 [4] and
it is conjectured that the inapproximability of the problem is of a similar magnitude. It would
be nice to obtain strong inapproximability bounds for the rank reduction problem that do not
rely on this conjecture. One approach may be to analyze hypergraphs as the rank reduction
problem in transversal matroids incorporates the hypergraph version of the minimum t-edge
subgraph problem. That is, we wish to select as few vertices as possible that induce at least t
hyperedges. Perhaps surprisingly, little is known about this problem. As far as we are aware,
the only specific hardness result is NP-hardess due Vinterbo [21] who studied the problem in
the context of making medical databases anonymous.
We conclude this section with a remark about the approximability of the minimum t-edge
subgraph problem itself. Given the existence of a O(n1/4+δ )-approximation algorithm for the
densest k-subgraph problem, in view of Lemma 3.3 it is perhaps natural to wonder whether one
could achieve a O(n1/8+δ )-approximation for the former problem. While this is still open as far as
we know, Chlamtac et al. [8] recently made progress in that direction by describing an algorithm
√
for the minimum t-edge subgraph problem with an approximation ratio of O(n3−2 2+δ ) =
O(n0.1716+δ ) for fixed δ > 0.
4. The Maximum Vertex Cover Problem in Bipartite Graphs
As we have seen, the rank reduction problem admits a fairly simple polynomial-time algorithm
on partition matroids but becomes much harder on transversal matroids, in the sense that
approximation algorithms offering good guarantees seem unlikely to exist. Another interesting
5
generalization of the case of partition matroids is to consider the intersection of two partition
matroids.
As is well-known, the set of matchings of a bipartite graph G = (V, E) with bipartition (A, B)
can be modeled as the family of common independent sets of two partition matroids M1 and
M2 defined on E: Take M1 to be the partition matroid with model ((E(u1 ), 1), . . . , (E(ua ), 1))
and M2 the partition matroid with model ((E(v1 ), 1), . . . , (E(vb ), 1)), where A = {u1 , . . . , ua },
B = {v1 , . . . , vb }, and for w ∈ V the set E(w) denotes the set of edges incident to w. Hence,
in this specific case the rank reduction problem on M1 ∩ M2 amounts to finding a subset F of
edges of G of minimum size such that µ(G − F ) 6 µ(G) − k. In this section we show that this
problem is NP-hard. More accurately, we show that a problem polynomially equivalent to it,
the maximum vertex cover problem on bipartite graphs, is NP-hard; see Theorem 4.1. This
solves an open problem of B. Simeone (see [15]).
The maximum vertex cover problem (also known as the partial vertex cover problem) is
defined as follows: Given a graph G = (V, E) and a positive integer k 6 |G|, find a subset X
of vertices of G with |X| = k such that the number of edges covered by X is maximized. (An
edge e of G is covered by X if e has at least one endpoint in X.)
Now, if G is bipartite, t is a positive integer with t 6 µ(G), and F is a subset of edges of G
such that µ(G − F ) 6 µ(G) − t =: k, then by Kőnig’s theorem G − F has a vertex cover X of size
k, and hence X covers at least |E| − |F | edges of G. (We remark that X could cover some edges
of F too, and that X can be computed in polynomial time given F .). Conversely, for every set
X ⊆ V with |X| = k, the set F of edges of G not covered by X is such that µ(G − F ) 6 k.
Therefore, for bipartite graphs, the maximum vertex cover problem is polynomially equivalent
to that of finding a minimum-size set of edges decreasing the matching number by a prescribed
amount.
It should be noted that two recent works [6, 9] with an overlapping set of authors claim
that the NP-hardness of the maximum vertex cover problem on bipartite graphs can be derived
directly from the reduction of Corneil and Perl [7] showing that the densest k-subgraph problem
is NP-hard on bipartite graphs. However, the argument relating the latter reduction to the
maximum vertex cover problem, described explicitly in [9, Lemma 4], is flawed.2 We also
mention that the proof of Theorem 1 in [6], showing that a related problem called the maximum
quasi-independent set problem is NP-hard on bipartite graphs, relies on the assumption that
the maximum vertex cover problem is NP-hard on bipartite graphs. Thus our result also fills a
gap in that proof.
Theorem 4.1. The maximum vertex cover problem is NP-hard on bipartite graphs.
Before proving Theorem 4.1, we need to introduce a technical lemma.
Lemma 4.2. Let ` be an integer with ` > 6. Then the integer program
minimize
s. t.
x + 2y + 3z
`
x+y+z−s=
−`
2
s
x6
2
x, y, z, s ∈ N
2As mentioned in [9, Lemma 3], the maximum vertex cover problem in bipartite graphs is polynomially
equivalent to the densest k-subgraph problem in complements of bipartite graphs. Thus one may equivalently
consider the complexity of the latter problem. In the proof of Lemma 4 in [9], the authors point out that the
reduction of [7] implies that the problem
of finding a densest k-subgraph in the complement of a bipartite graph
with bipartition (A, B) with k = q + 2q and with the extra requirement that it contains exactly q vertices from
A and 2q vertices from B is NP-hard. From this they wrongly conclude that the densest k-subgraph problem,
without this extra constraint, is also NP-hard on complements of a bipartite graphs. (In fact, the instances
obtained via the reduction in [7] satisfy |B| > k, and thus a densest k-subgraph is trivially obtained by taking k
vertices in the clique B.)
6
has a unique optimal solution given by x =
`
2
, y = 0, z = 0, s = `.
Proof. The proof is a straightforward case analysis. Consider an optimal solution x, y, z, s to
the integer program and, arguing by contradiction, assume it differs from the solution described
above. Let f (x, y, z) := x + 2y + 3z.
Case 1: s 6 2. We have x 6 s and thus
`
`
y+z =
−`+s−x>
− `.
2
2
It follows that f (x, y, z) > 2(y + z) > 2 2` − 2`. But 2 2` − 2` > 2` since ` > 6, contradicting
the optimality of the solution.
Case 2: 3 6 s < `.
`
`
s
`
s
y+z =
−`+s−x>
−`+s−
=
−
− (` − s) > 2.
2
2
2
2
2
s
Here, the last inequality follows from the fact that s+1
−
2
2 > 3.
Now, increment s by 1, x by 3, and decrement y and z in such a way that they remain
non-negative integers and that the sum y + z decreases by exactly 2. The modified solution is
still feasible and f (x, y, z) decreases by at least 1, a contradiction.
Case 3: s = `. Then y + z > 1, since otherwise
we would have the solution described in the
lemma statement. It follows that x 6 2` − 1. Thus we can increment x by 1 and decrement
by 1 a positive variable among y, z. This
f (x, y,
z), sa contradiction.
strictly decreases
`
`
`
Case 4: s > `. Then x + y + z = 2 + s − l > 2 . But x 6 2 < 2 , otherwise the solution
cannot be minimum. Therefore y + z > 1. Thus we improve the solution by incrementing x by
1 and decrementing by 1 a positive variable among y, z.
Now we may turn to the proof of Theorem 4.1.
Proof of Theorem 4.1. The reduction is from the NP-complete problem Clique: Given a graph
H and an integer `, decide whether H contains a clique on ` vertices or not. We may assume
` > 6 (otherwise, we simply check the existence of an `-clique by brute force). We may also
suppose that H has minimum degree at least 2. Indeed, a vertex with degree at most 1 cannot
be part of an `-clique, and thus those vertices can iteratively be removed from the graph.
Finally, we assume that ||H|| > |H| + 2` . This last assumption can also be made without loss
of generality. Indeed, if ||H|| is too small then one can simply consider the disjoint union of H
with a large enough 3-regular graph; since ` > 6 no vertex from this new 3-regular component
can be part of an `-clique.
We build an instance (G, k) of the maximum vertex cover problem as follows. First, create
two adjacent vertices av and bv for every vertex v ∈ V (H), and similarly two adjacent vertices
ae and be for every edge e ∈ E(H). Next, for every edge uv ∈ E(H), add the edges auv bu , buv au
and auv bv , buv av . Finally, let
`
k := |H| + ||H|| −
+ `.
2
Observe that G is bipartite with bipartition
{ax : x ∈ V (H) ∪ E(H)}, {bx : x ∈ V (H) ∪ E(H)} .
See Figure 1 for an illustration of the construction.
A feasible solution for this instance of the maximum vertex cover problem is a subset X of
vertices of G with |X| = k, which we call a partial vertex cover for short. We let c(X) denote
the number of edges covered by such a set X. Let OP T denote the maximum of c(X) over
every partial vertex cover X of G.
A partial vertex cover X of G is nice if
X ∩ {au , bu } ∈ {au }, {au , bu }
7
ay
by
ax
bx
aw
bw
av
bv
au
bu
auv
buv
auy
buy
avy
bvy
avw
bvw
awx
bwx
axy
bxy
u
y
v
x
w
H
G
Figure 1. Illustration of the construction of the bipartite graph G. (We note
that this graph H cannot be a valid instance
of the problem because ` > 6 and
thus H cannot satisfy ||H|| > |H| + 2` ; our aim here is only to illustrate how G
is obtained from H on a small example.)
for every u ∈ V (H) and
X ∩ {ae , be } ∈ ∅, {ae }
for every e ∈ E(H).
Claim 4.3. Given a partial vertex cover X of G, one can find a nice partial vertex cover X 0 of
G with c(X 0 ) > c(X).
Proof. First we define a partial vertex cover X̃ based on X which is close to being nice: Let
X̃ := au , bu : u ∈ V (H), au , bu ∈ X ∪ au : u ∈ V (H), |{au , bu } ∩ X| = 1
∪ ae , be : e ∈ E(H), ae , be ∈ X ∪ ae : e ∈ E(H), |{ae , be } ∩ X| = 1 .
By construction |X̃| = |X|. Clearly, an edge ax bx with x ∈ V (H) ∪ E(H) is covered by X̃ if
and only if it is covered by X. Also, given a pair (u, e) of vertex u ∈ V (H) and edge e ∈ E(H)
such that u is incident to e in H, the set X̃ covers at least as many edges in {ae bu , be au } as X
(though not necessarily the same ones). It follows that c(X̃) > c(X).
A useful property of the set X̃ is that if bx ∈ X̃ for some x ∈ V (H) ∪ E(H) then necessarily
ax ∈ X̃. For simplicity we call this property the a-property of X̃.
We need to introduce an additional definition. An element x ∈ V (H) ∪ E(H) is said to be
bad in a partial vertex cover Y of G if either x ∈ V (H) and ax , bx ∈
/ Y (x is a bad vertex), or
x ∈ E(H) and ax , bx ∈ Y (x is a bad edge). Observe that Y is nice if and only if Y has the
a-property and there is no bad element.
Suppose e = uv is an edge of H which is bad in X̃. If u or v is also bad in X̃, say u, then let
X̃ 0 := X̃ − {be } ∪ {au }.
We have ae ∈ X̃, thus the edge ae be is still covered by X̃ 0 . Since X̃ 0 covers also be au , there is
at most one edge incident to be in G (namely, be av ) which is not covered by X̃ 0 . On the other
8
hand, X̃ 0 covers the previously uncovered edge au bu . Hence, X̃ 0 is a partial vertex cover with
c(X̃ 0 ) > c(X̃). Observe that X̃ 0 still has the a-property, and the edge e is no longer bad in X̃ 0 .
If, on the other hand, none of u, v is bad in X̃, then au , av ∈ X̃ by the a-property. Since
ae ∈ X̃, it follows that c(X̃ − {be }) = c(X̃). There
exists an element x ∈ V (H) ∪ E(H) such
`
that ax ∈
/ X̃, because |X̃| = k = |H| + ||H|| − 2 + ` < |H| + ||H|| (since ` > 6). Let then
X̃ 0 := X̃ − {be } ∪ {ax }.
The set X̃ 0 is a partial vertex cover with the a-property and with c(X̃ 0 ) > c(X̃). Moreover, the
edge e is no longer bad in X̃ 0 .
Now apply iteratively the above modifications on X̃ as long as there exists a bad edge.
b with the a-property, without bad edges, and with
This results in a partial vertex cover X
b
c(X) > c(X̃).
b Suppose u ∈ V (H) is such a vertex, that is, au , bu ∈
b
Next we deal with bad vertices in X.
/ X.
Consider two edges e, f incident to u in H. (Recall that H has minimum degree at least 2.)
b = k = |H| + ||H|| − ` + ` and ||H|| > |H| + ` by our assumption on H, we have
Since |X|
2
2
b > 2|H|. Together with the a-property of X,
b it follows that ae0 ∈ X
b for some edge e0 ∈ E(H)
|X|
b because otherwise e0 would be a bad edge for X.
b
(possibly e0 = e or e0 = f ). Note that be0 ∈
/ X,
Let
0
b
b
0
X := X − {ae } ∪ {au }.
b − {ae0 }) > c(X)
b − 3. Furthermore, be , bf ∈
b as there
Since ae0 has degree 3 in G we have c(X
/X
b Thus, the three edges au bu , be au , bf au of G were not covered by X
b but
are no bad edges in X.
0
0
b , so we have c(X
b ) > c(X).
b Similarly as before, the partial vertex cover X
b 0 has
are covered by X
b Therefore, iterating this procedure as long as
the a-property and one less bad vertex than X.
b > c(X)
there is a bad vertex, we eventually obtain a partial vertex cover X 0 with c(X 0 ) > c(X)
having the a-property and no bad element, as desired.
Consider a nice partial vertex cover X of G. Let S(X) be the set of vertices u ∈ V (H) such
that au , bu ∈ X, and let s(X) := |S(X)|. An edge e = uv of H satisfies exactly one of the
following three conditions:
(1) u, v ∈ S(X);
(2) exactly one of u, v belongs to S(X),
(3) u, v ∈
/ S(X).
We say that edge e is of type i (i ∈ {1, 2, 3}) if e satisfies the ith condition above and moreover
ae , be ∈
/ X. (We will focus on edges e of H such that ae , be ∈
/ X in what follows, which is why
the other ones do not get assigned a type.) Let Ei (X) be the set of edges of H with type i, and
let ei (X) := |Ei (X)|.
Claim 4.4. Let X be a nice partial vertex cover. Then
c(X) = ||G|| − e1 (X) − 2e2 (X) − 3e3 (X).
Proof. As X is nice, av ∈ X for all v ∈ V . Therefore every edge of the form av bv or av be is
covered. Also, for each edge e = uv ∈ E(H) \ ∪16i63 Ei (X) we have that ae be , ae bu , ae bv are
all covered. Thus the only uncovered edges are incident to vertices ae where e ∈ ∪16i63 Ei (X).
Suppose e ∈ ∪16i63 Ei (X) and let us consider which edges among the three edges ae be , ae bu , ae bv
are covered by X. If e ∈ E1 (X), then X covers ae bu and ae bv but not ae be . If e ∈ E2 (X), then X
covers exactly one of ae bu , ae bv , and avoids ae be . If e ∈ E3 (X), then X covers none of the three
edges. Hence, the total number of edges not covered by X is exactly e1 (X)+2e2 (X)+3e3 (X).
Claim 4.5. Let X be a nice partial vertex cover. Then
`
c(X) 6 ||G|| −
,
2
9
with equality if and only if s(X) = `, e1 (X) =
`
2
, e2 (X) = e3 (X) = 0.
Proof. Let x := e1 (X), y := e2 (X), z := e3 (X) and s := s(X). Then c(X) = ||G|| − f (x, y, z),
where f (x, y, z) = x + 2y + 3z by the previous claim.
Every edge in E1 (X) has its two endpoints in S(X); hence,
s
(1)
> x.
2
Also,
`
|H| + ||H|| − (x + y + z − s) = |X| = k = |H| + ||H|| −
−` ,
2
and thus
(2)
`
x+y+z−s=
− `.
2
Since ` > 6 and x,
y, z, s, ` are non-negative integers satisfying (1) and (2), by Lemma 4.2 we
`
have f (x, y, z) > 2 , or equivalently,
`
c(X) 6 ||G|| −
.
2
Moreover, equality holds if and only if x = 2` , y = 0, z = 0, s = ` by the same lemma.
It follows from Claims 4.3 and 4.5 that
`
OP T 6 ||G|| −
.
2
If H has an `-clique K ⊆ H, then the subset X ⊆ V (G) defined by
X := {au , bu : u ∈ V (K)} ∪ {au : u ∈ V (H) − V (K)}
∪ {ae : e ∈ E(H) − E(K)}
is a partial vertex cover of G with c(X) = ||G|| − 2` , implying OP T = ||G|| − 2` .
Conversely, if OP T = ||G|| − 2` , then there exists a partial vertex cover X of G with
c(X) = ||G|| − 2` , and by Claim 4.3 we may assume that X is nice. From Claim 4.5 we then
have s(X) = `, e1 (X) = 2` , e2 (X) = e3 (X) = 0, implying that S(X) induces an `-clique in H.
Therefore,
we can decide in polynomial time if H has an `-clique by checking if OP T =
`
||G|| − 2 . This concludes the proof.
The NP-hardness of the maximum vertex cover problem on bipartite graphs motivates the
search for non-trivial approximation algorithms for this class of graphs. A recent result in
this direction is due to Apollonio and Simeone [3], who gave an LP-based (4/5)-approximation
algorithm for bipartite graphs.
Acknowledgments
We are grateful to Attila Bernath and Tamás Kiraly for their comments on the proof of
Theorem 4.1, and to the two anonymous referees for their helpful remarks and suggestions. We
also thank Mohit Singh, Bill Cunningham and Anupam Gupta for interesting discussions, and
Nicola Apollonio for providing us with a preliminary version of [2].
10
References
[1] R. Andersen and K. Chellapilla, “Finding dense subgraphs with size bounds”, Proceedings
of the 6th International Workshop on Algorithms and Models for the Web-Graph, pp. 25–37,
2009.
[2] N. Apollonio and B. Simeone, “The maximum vertex coverage problem on bipartite
graphs”. Discrete Applied Mathematics, 165, pp. 37–48, 2014.
[3] N. Apollonio and B. Simeone, “Improved Approximation of Maximum Vertex Coverage
Problem on Bipartite Graphs”. SIAM J. Discrete Math., 28(3), pp. 1137–1151, 2014.
[4] A. Bhaskara, M. Charikar, E. Chlamtac, U. Feige, and A. Vijayaraghavan, “Detecting
1
high log-densities: an O(n 4 ) approximation for densest k-subgraph”, Proceedings of 42nd
Symposium on Theory of Computing (STOC), pp. 201–210, 2010.
[5] A. Bhaskara, M. Charikar, V. Guruswami, A. Vijayaraghavan, and Y. Zhou “Polynomial
Integrality gaps for Strong relaxations of Densest k-subgraph”, Proceedings of 23rd Symposium on Discrete Algorithms (SODA), pp. 388–405, 2012.
[6] N. Bourgeois, A. Giannakos, G. Lucarelli, I. Milis, V. Paschos, and O. Pottie, “The max
quasi-independent set problem”, Journal of Combinatorial Optimization, 23(1), pp. 94–
117, Jan. 2012.
[7] D. Corneil and Y. Perl. “Clustering and domination in perfect graphs”. Discrete Applied
Mathematics, 9(1), pp. 27–39, 1984.
[8] E. Chlamtac, M. Dinitz, and R. Krauthgamer. “Everywhere-Sparse Spanners via Dense
Subgraphsh”, Proceedings of the 2012 IEEE 53rd Annual Symposium on Foundations of
Computer Science, pp. 758–767, 2012.
[9] F. Della Croce and V. Paschos, “On the max k-vertex cover problem”, Cahiers du LAMSADE, 307, Université Paris-Dauphine, 2011.
[10] U. Feige, D. Peleg, and G. Kortsarz, “The dense k-subgraph problem”, Algorithmica, 29(3),
pp. 410–421, 2001.
[11] U. Feige and M. Seltser, “On the densest k-subgraph problem”, Technical Report CS97-16,
Department of Mathematics and Computer Science, Weizmann Institute Of Science, 1997.
[12] O. Goldschmidt and D. Hochbaum, “A polynomial time algorithms for the k-cut problem
for fixed k”, Mathematics of Operations Research, 19(1), pp. 24–37, 1994.
[13] S. Khot, “Ruling out PTAS for graph min-bisection, densest subgraph and bipartite clique”,
Proceedings of the 44th Symposium on the Foundations of Computer Science (FOCS), pp.
136–145, 2004.
[14] J. Naor and Y. Rabani, “Tree packing and approximating k-cuts”, Proceedings of the 12th
Symposium on Discrete Algorithms (SODA), pp. 26–27, 2001.
[15] Open Problem Forum of the Egerváry Research Group, http://lemon.cs.elte.hu/
egres/open/Partial_vertex_cover_in_bipartite_graphs.
[16] J. Oxley, “What is a matroid?”, Cubo, 5, pp. 179–218, 2003.
[17] R. Ravi and A. Sinha, “Approximating k-cuts via network strength”, Proceedings of the
13th Symposium on Discrete Algorithms (SODA), pp. 621–622, 2002.
[18] H. Saran and V. Vazirani, “Finding k-cuts within twice the optimal”, SIAM Journal on
Computing, 24, pp. 101–108, 1995.
[19] A. Schrijver, Combinatorial Optimization: Polyhedra and Efficiency, Springer, 2003.
[20] L. Zhao, H. Nagamochi and T. Ibaraki, “Approximating the minimum k-way cut in a graph
via minimum 3-way cuts”, Journal of Combinatorial Optimization, 5, pp. 397–410, 2001.
[21] S. Vinterbo, “A note on the hardness of the k-ambiguity problem”, Technical Report, 2002.
11
Département d’Informatique
Université Libre de Bruxelles
Brussels, Belgium
E-mail address: [email protected]
Department of Mathematics and Statistics, and School of Computer Science
McGill University
Montreal, Canada
E-mail address: [email protected]
12
| 8cs.DS
|
Secrecy Capacity of Colored Gaussian Noise
Channels with Feedback
arXiv:1801.06702v1 [cs.IT] 20 Jan 2018
Chong Li, Yingbin Liang, H. Vincent Poor and Shlomo Shamai (Shitz)
Abstract
In this paper, the k-th order autoregressive moving average (ARMA(k)) Gaussian wiretap channel with noiseless
causal feedback is considered, in which an eavesdropper receives noisy observations of the signals in both forward
and feedback channels. It is shown that a variant of the generalized Schalkwijk-Kailath scheme, a capacity-achieving
coding scheme for the feedback Gaussian channel, achieves the same maximum rate for the same channel with the
presence of an eavesdropper. Therefore, the secrecy capacity is equal to the feedback capacity without the presence
of an eavesdropper for the feedback channel. Furthermore, the results are extended to the additive white Gaussian
noise (AWGN) channel with quantized feedback. It is shown that the proposed coding scheme achieves a positive
secrecy rate. As the amplitude of the quantization noise decreases to zero, the secrecy rate converges to the capacity
of the AWGN channel.
Index Terms
Secrecy Capacity, Feedback, Colored Gaussian, Schalkwijk-Kailath Scheme
I. I NTRODUCTION
It has been more than a half century since the information theorists started to investigate the capacity of feedback
Gaussian channels. As the pioneering studies on this topic, Shannon’s 1956 paper [2] showed that feedback does
not increase the capacity of the memoryless AWGN channel, and Elias [3] [4] proposed some simple corresponding
feedback coding schemes. Schalkwijk and Kailath [5] [6] then developed a notable linear feedback coding scheme to
achieve the capacity of the feedback AWGN channel. Thereafter, the problem of finding the feedback capacity and the
capacity-achieving codes for the memory Gaussian channels (e.g. ARMA(k)) has been extensively studied. Butman
[7] [8], Wolfowitz [9] and Ozarow [10] [11] extended Schalkwijk’s scheme to ARMA(k) Gaussian channels, leading
Chong Li is with Qualcomm Research, Bridgewater, NJ, 08807, USA ([email protected]). Yingbin Liang is with the Department of
Electrical and Computer Engineering, The Ohio State University, Columbus, OH, 43220, USA (email: [email protected]). H. Vincent Poor is
with the Department of Electrical Engineering, Princeton University, Princeton, NJ, 08544, USA (email: [email protected]). Shlomo Shamai
(Shitz) is with the Department of Electrical Engineering, Technion-Israel Institute of Technology, Technion City, Haifa 32000, Israel (email:
[email protected])
The work of Y. Liang has been supported by the U.S. National Science Foundation under the grant CCF-1618127. The work of H. V. Poor
has been supported by the U.S. National Science Foundation under the grants CNS-1702808 and ECCS-1647198. The work of S. Shamai has
been supported by the European Union’s Horizon 2020 Research And Innovation Programme, grant agreement no. 694630.
This paper was presented (in part) in [1].
to several valuable upper and lower bounds on the capacity. Motivated by these elegant results/insights, in 1989 Cover
and Pombra [12] made a breakthrough on characterizing the n-block capacity of the feedback Gaussian channel.
In 2010, Kim [13] provided a characterization (in the form of an infinite dimensional optimization problem) of the
capacity of the ARMA(k) feedback Gaussian channel based on Cover-Pombra’s n-block capacity characterization.
Unfortunately, except for the first-order ARMA noise, it is non-trivial to compute the capacity by solving this
infinite dimensional optimization. Recently, Gattami [14] showed that the capacity of the stationary Gaussian noise
channel with finite memory can be found by solving a semi-definite programming problem. In addition, Li and Elia
[15] [16] utilized the control-theoretic tools to compute the capacity of the ARMA(k) feedback Gaussian channel
and construct the capacity-achieving feedback codes.
As a natural extension of the above results, the ARMA(k) feedback Gaussian channel with an eavesdropper
that has noisy access of channel transmissions is of much interest. Concretely, two fundamental questions can be
asked: 1). would the feedback capacity of such a channel decrease subject to the secrecy constraint? 2). what would
be the secrecy capacity-achieving codes? In fact, secure communications over feedback channels have attracted a
lot of attention over the last decade. Substantial progresses have been made towards understanding this type of
channels. In particular, although the feedback may not increase the capacity of open-loop AWGN channels, [17]–
[26] showed that feedback can increase the secrecy capacity by sharing a secret key between legitimate users or
roiling the codewords from eavesdroppers’ observations. For instance, [17] and [18] showed the achievement of a
positive secrecy rate by using noiseless feedback even when the secrecy capacity of the feed-forward channel is
zero. Furthermore, [27] presented an achievable scheme for the wiretap channel with generalized feedback, which is
a generalization and unification of several relevant previous results in the literature. Very recently, [28] proposed an
improved feedback coding scheme for the wiretap channel with noiseless feedback, which was shown to outperform
the existing ones in the literature.
On the other hand, the multiple-access (MA) wiretap channel with feedback has also been studied in recent a few
years. [29] derived achievable secrecy rate regions for discrete memoryless Gaussian channels where two trusted
users send independent confidential messages to an intended receiver in the presence of a passive eavesdropper. [30]
developed inner and outer bounds on the secrecy rate region for the MA wiretap channel with noiseless feedback.
In a more general setting where users have multiple-input multiple-output (MIMO) equipments, [31] characterized
the pre-log factor of the secrecy rate in the case with the number of antennas at the source being no larger than
that at the eavesdropper. [32] and [33] investigated the benefits of state-feedback to increase the secrecy degrees of
freedom for the two-user Gaussian MIMO wiretap channel.
However, it is noteworthy that most of the aforementioned results on both the point-to-point and the MA
communications only considered memoryless wiretap channels. Thus motivated, in this paper we consider a memory
wiretap channel (i.e. ARMA(k) Gaussian noise channel) with feedback and make two major contributions.
1) We propose a variant of the generalized S-K scheme, which achieves the feedback capacity of the ARMA(k)
Gaussian noise channel without an eavesdropper and show that the proposed coding scheme also achieves
the same maximum rate of such a channel with an eavesdropper.
2) We further study the AWGN channel with quantized feedback, which is a more realistic channel model for
Fig. 1. ARMA(k) Gaussian wiretap channel with feedback.
the feedback link. We show that the proposed coding scheme provides non-trivial positive secrecy rates and
achieves the feedback capacity of the AWGN channel as the amplitude of the quantization noise vanishes to
zero.
The rest of the paper is organized as follows. In Section II and III, we introduce the system model and the
preliminary results, respectively. Section IV presents the main results of our paper. Section V provides the technical
proofs. Finally, in Section VI, we conclude the paper and outline possible research venues down the road.
Notation: Uppercase and the corresponding lowercase letters (e.g.,Y, Z, y, z) denote random variables and their
realizations, respectively. We use log to denote the logarithm with base 2, and 0 log 0 = 0. We use x0 to denote
∞
the transpose of a vector or matrix x. The symbol C[a,b]
denotes the set of bounded continuous functions on [a, b].
RH 2 denotes the set of stable and proper rational filters in Hardy space H2 , and L2 denotes the 2-norm Lebesgue
space.
II. S YSTEM M ODEL
In this section, we present the mathematical system model. First of all, we consider a discrete-time Gaussian
channel with noiseless feedback (Fig. 1). The additive Gaussian channel is modeled as
y(k) = u(k) + w(k),
k = 1, 2, · · · ,
(1)
jθ
where the Gaussian noise {w(k)}∞
k=1 is assumed to be stationary with power spectrum density Sw (e ) for ∀θ ∈
[−π, π). Unless the contrary is explicitly stated, “stationary” without specification refers to stationary in wide sense.
Moreover, we assume the power spectral density satisfies the Paley-Wiener condition
1
2π
Z π
−π
| log Sw (e jθ )|dθ < ∞.
Assumption 1. (ARMA(k) Gaussian Channel) In this paper, noise w is assumed as the output of a finite-dimensional
linear time invariance (LTI) minimum-phase stable system H ∈ RH 2 , driven by the white Gaussian noise with
zero mean and unit variance. The power spectral density (PSD) of w is colored (nonwhite)1 , bounded away from
zero and has a canonical spectral factorization given by Sw (e jθ ) = |H(e jθ )|2 .
1 Both
the (Shannon) capacity [2] and the secrecy capacity [23] for the AWGN channel with noiseless causal feedback have been known.
As shown in Fig.1, the feedback wiretap channel of interest includes a forward channel from Alice to Bob as
described by (1), a causal noiseless feedback ŷ from Bob to Alice, and three noisy observation channels to the
eavesdropper Eve. Note that a classical wiretap channel model can be recovered if the eavesdropper’s channel input
u and ŷ are removed. In this paper, we assume that the eavesdropper is powerful and can access three inputs 2 . The
noisy wiretap channels are modeled as
z(k) =u(k) + v(k),
z̃(k) =y(k) + ṽ(k),
k = 1, 2, · · · .
ẑ(k) =ŷ(k) + v̂(k),
The additive noises v, ṽ and v̂ are assumed to be arbitrarily finite-memory channels, i.e.,
k−1
p(v(k)|vk−1
1 ) = p(v(k)|vk−d ),
k ≥ d,
k−1
p(ṽ(k)|ṽk−1
1 ) = p(ṽ(k)|ṽk−d˜),
˜
k ≥ d,
k−1
p(v̂(k)|v̂k−1
1 ) = p(v̂(k)|v̂k−dˆ),
ˆ
k ≥ d,
(2)
where d, d˜ and dˆ respectively represent the size of finite memories and the notation vba represents a sequence
{va , va+1 , · · · , vb } in a compact form. In this paper, we assume these noises have strictly positive and bounded
variance for all k. But they are not necessarily uncorrelated.
We specify a sequence of (n, 2nRs ) channel codes with an achievable secrecy rate Rs as follows. We denote
the message index by xu 0 , which is uniformly distributed over the set {1, 2, 3, · · · , 2nRs }. The encoding process
ui (xu 0 , ŷi−1 ) at Alice satisfies the average transmit power constraint P, where ŷi−1 = {ŷ0 , ŷ1 , · · · , ŷi−1 } (ŷ0 = 0)
/ for i =
1, 2, · · · , n, and u1 (xu 0 , ŷ0 ) = u1 (xu 0 ). Bob decodes the message to be x̂u 0 by following a decoding function g : yn →
(n)
{1, 2, · · · , 2nRs } with an error probability satisfying Pe =
1
2nRs
nRs
∑2xu 0 =1 p(xu 0 6= g(yn )|xu 0 ) ≤ εn , where limn→∞ εn = 0.
Meanwhile, the information received by Eve should asymptotically vanish, i.e., limn→∞ n1 I(xu 0 ; zn1 , z̃n1 , ẑn1 ) = 0. The
objective of secure communications is to send xu 0 to Bob at an as high rate Rs as possible. The secrecy capacity
Csc is defined as the supremium of all achievable rates Rs . Mathematically,
Csc =
sup
Rs
feasible coding schemes
(3)
1
n n n
s.t. lim I(xu 0 ; z1 , z̃1 , ẑ1 ) = 0,
n→∞ n
where the “feasible coding schemes” refer to all feedback codes satisfying the secrecy requirements and the power
constraint. Note that the feedback capacity (without the secrecy constraint) from Alice to Bob, denoted as C f b , can
be recovered by removing the secrecy constraint. This implies Csc ≤ C f b .
III. P RELIMINARIES OF FEEDBACK CAPACITY AND CAPACITY- ACHIEVING CODING SCHEME
In this section, we review the characterization of the feedback capacity C f b and then propose a variant of the
generalized S-K scheme, which is a C f b -achieving feedback codes without the presence of an eavesdropper. The
materials here are useful for us to further investigate the channel model with an eavesdropper.
2 Note
that in [1] only the access to channel input u and channel output y is considered. Based on the generalized results in this paper, however,
the results in [1] with three noisy observation channels to the eavesdropper as assumed in this paper still hold.
A. Feedback Capacity C f b Revisited
Firstly, we present the feedback capacity characterization for the Gaussian channel under Assumption 1. As
proved in [13], the feedback capacity from Alice to Bob for such a channel with the average power budget P can
be characterized by
1 π
C f b = max
log |1 + Q(eiθ )|dθ ,
Q 2π −π
Z
1 π
|Q(e jθ )|2 Sw (e jθ )dθ ≤ P,
s.t.
2π −π
Z
(4)
Q ∈ RH 2 is strictly causal.
Remark 1. Under Assumption 1, the optimal Q has no zeros on the unit circle (Proposition 5.1 (ii) in [13]).
When specified to ARMA(1) channel, the above characterization can be simplified and has a closed-form solution.
Lemma 1. (Theorem 5.3 in [13]) For the ARMA(1) feedback Gaussian channel with the average channel input
power budget P > 0, the feedback capacity C f b is given by
C f b = − log x0 ,
where x0 is the unique positive root of the fourth-order polynomial in x
Px2 =
(1 − x2 )(1 + σ αx)2
,
(1 + σ β x)
and
σ = sign(β − α).
Furthermore, an optimal Q to achieve C f b is given by
Q(eiθ ) = −Pσ x0
1 + β αx0
eiθ (1 + β eiθ )
.
1 + ασ x0 (1 + αeiθ )(1 − σ x0 eiθ )
(5)
Since the optimization in (4) has infinite dimensional search space, except for the ARMA(1) Gaussian channels,
neither the analytical nor the numerical solutions to Q(e jθ ) was known in the literature. One recent result in [14]
casted the above optimization into a semi-definite programming and then used the convex tools to compute C f b .
In addition, [15] [16] provided a numerical approach to compute C f b and explicitly construct the optimal Q(e jθ ),
which can be efficiently solved by the standard convex optimization tools. We refer the interested reader to [15]
[16] for details. To make our paper self-contained, we re-state the main results in [15] [16] as follows.
Proposition 1. (Lemma 4, [16]) Consider a non-white additive noise w under Assumption 1 in the forward channel
and let Q(e jθ ) = a(θ ) + jb(θ ). The h-upper-bound on the feedback capacity C f b , denoted by C f b (h), can be
characterized by
1
4π
C f b (h) = max
Γ
s.t.
1
2π
Z π
log((1 + a(θ ))2 + b(θ )2 )dθ
−π
Z π
a2 (θ ) + b2 (θ ) Sw (θ )dθ ≤ P,
−π
Z π
(6)
Z π
a(θ ) cos(nθ )dθ +
−π
b(θ ) sin(nθ )dθ = 0
(strict causality constraint)
−π
n = 0, 1, 2, · · · , h,
where the maximum is taken over a functional set Γ defined as
Γ ={a(θ ), b(θ ) : [−π, π] → R
| a(θ ), b(θ ) ∈ L2 }.
(7)
Furthermore, we have C f b (h) ≥ C f b (h + 1),C f b (h) ≥ C f b for any h ≥ 0, and
C f b = lim C f b (h).
h→∞
Remark 2. With a bit abuse of notation, we use Sw (θ ) instead of Sw (eiθ ) for simplicity. The basic idea to obtain the
characterization (6) is that the strict causality constraint (4) can be equivalently imposed on (the infinite number
of) the non-positive index coefficients of the inverse Fourier transform of Q(eiθ ) by setting them to zeros. The
upper bound follows as we herein only impose a finite number of the Fourier coefficients with non-positive index
−h, −h + 1, · · · , −1, 0 to be zero.
Notice that C f b (h) still turns out to be a semi-infinite dimensional problem, in which there are finite number
of constraints but infinite number of feasible solutions in L2 . However, one important result, as shown below,
establishes the strong Lagrangian dual of C f b (h). In other words, there is no duality gap between the semi-infinite
dimensional primal problem and its finite dimensional dual problem, a fact that provides a convex optimization
approach to compute C f b (h).
Theorem 1. (Theorem 2 in [15]) Under Assumption 1, let
A(θ ) = [cos(θ ), cos(2θ ), · · · , cos(hθ )]0 ,
B(θ ) = [sin(θ ), sin(2θ ), · · · , sin(hθ )]0 .
For λ > 0, η ∈ Rh , and η0 ∈ R, define
r2 (θ ) = (2λ Sw (θ ) + η 0 A(θ ) + η0 )2 + (η 0 B(θ ))2 .
Then, the following statements hold.
a) The Lagrangian dual of (6) is given by
(D) : µh = −
max
λ >0,η∈Rh ,η0 ∈R
g(λ , η, η0 ),
(8)
where
g(λ , η, η0 ) =
1
2π
Z π
1
−π
2
log(2λ Sw (θ ) − ν(θ )) −
r2 (θ )
1
+ λ Sw (θ ) dθ − λ P + η0 + ,
2ν(θ )
2
(9)
with
ν(θ ) =
−r2 (θ ) +
p
r4 (θ ) + 8λ Sw (θ )r2 (θ )
.
2
(10)
b) The dual problem (D) in (8) is equivalent to the following convex optimization problem
µh = −
max
g̃(λ , η, η0 , ν(θ )),
h
λ > 0, η ∈ R , η0 ∈ R
(11)
∞
ν(θ ) ≥ 0 ∈ C[−π,π]
where
1
g̃(λ , η, η0 , ν(θ )) =
2π
r2 (θ )
1
log(2λ Sw (θ ) − ν(θ )) −
+ λ Sw (θ ) dθ − λ P + η0 + ,
2
2ν(θ )
2
Z π
1
−π
(12)
and the optimal ν(θ ) is characterized by (10).
c) Furthermore, C f b (h) = µh , and an optimal strictly casual filter Qh (e jθ ) = a(θ ) + jb(θ ) for C f b (h) exists and is
characterized by
2λ Sw (θ ) + η 0 A(θ ) + η0
−1
ν(θ )
η 0 B(θ )
b(θ ) =
a.e.,
ν(θ )
a(θ ) =
a.e.,
(13)
where (λ , η 0 , η0 , ν(θ )) are obtained by solving (8) (or (11)).
Note that the above optimization is not easily computable due to the integral in the objective function. A natural
approach is to approximate the integral with a finite sum by discretizing θ . We apply such discretization to (11)
(with spacing
π
m)
and introduce the following finite dimensional convex problem. Given m, consider
µh (m) = −
max
λ >0,η,η0 ,νi ≥0
g̃m (λ , η, η0 , νi )
(14)
where
g̃m (λ , η, η0 , νi ) =
1 2m 1
r2 (θi )
1
log(2λ
S
(θ
)
−
ν
)
+
λ
S
(θ
)
−
−
λ
P
+
η
+
,
w
i
i
w
i
0
∑ 2
2m i=1
2νi
2
π
and θi = −π + m
(i − 1).
It is argued in [16] that µh (m) can be an arbitrarily well approximation of µh , i.e., limm→∞ µh (m) = µh . Notice that
the optimization (14) is in a simple convex form. In particular, the log of an affine function is concave. The term
r2 (θi )
νi
is quadratic (composed with an affine function of the variables) over linear function, therefore convex. Thus,
(14) can be efficiently solved with the standard convex optimization tools, e.g. CVX, a package for specifying and
solving convex programs [34], [35].
In what follows, we describe, given an optimal Q in (4), how to construct an implementable coding scheme that
achieves the feedback capacity from Alice to Bob.
Fig. 2. Representation I: Decomposition of filter K into the feedback encoder (Alice) and decoder (Bob). The eavesdropper channels are not
included. z-transform is used to represent the dynamics of LTI systems.
B. C f b -achieving Feedback Coding Scheme
First of all, once an optimal Q is found for the above optimization, we construct a feedback filter K = −Q(1 +
Q)−1 stabilizing the channel within the prescribed input average power budget (see [15] for proofs). Next, based
on the transfer function K, we construct an explicit feedback coding scheme as follows, which is deterministic
(time-invariant) and has doubly exponential decaying decoding error probability.
We first present controller K as a LTI single-input-single-output (SISO) finite-dimensional discrete-time unstable
system with the following state-space model:
xs (k + 1)
A
= s
K:
xu (k + 1)
0
h
u(k) = Cs
0
xs (k)
Bs
+ y(k)
xu (k)
Bu
i xs (k)
.
Cu
xu (k)
Au
(15)
Based on Remark 1, we assume that the eigenvalues of Au are strictly outside the unit disc while the eigenvalues
of As are strictly inside the unit disc. Without loss of generality, we assume that As and Au are in Jordan form.
Assume Au has m eigenvalues, denoted by λi (Au ), i = 1, 2, · · · , m. Next, we review the coding scheme in [16] [36]
(Fig. 2), which decompose K into an encoder (Alice) and a decoder (Bob) with an estimated signal fed back to the
encoder via the noiseless feedback channel.
Representation I: Decoder-estimation-based Feedback Coding Scheme [16] [36] (Fig. 2).
Decoder: The decoder runs K driven by the channel output y.
xs (k + 1)
= As xs (k) + Bs y(k), xs (0) = 0,
x̂u (k + 1) = Au x̂u (k) + Bu y(k), x̂u (0) = 0.
It produces two signals: an estimate of the initial condition of the encoder
x̂u 0 (k) = A−k−1
x̂u (k + 1),
u
and a feedback signal
h
i xs (k)
.
û(k) = Cs Cu
x̂u (k)
Encoder: The encoder runs the following dynamics
x̃u (k + 1) = Au x̃u (k), x̃u (0) = xu 0 ,
ũu (k) = Cu x̃u (k).
It receives û and produces the channel input
u(k) = ũu (k) − û(k).
As will be seen later, the above coding scheme cannot be directly applied to the feedback ARMA(k) Gaussian
channel with an eavesdropper. Therefore, we next propose an equivalent representation of the above decoderestimation-based feedback coding scheme. It will be proved that a variant of this new representation achieves the
same maximum rate of the same channel without the eavesdropper.
Representation II: Channel-output-based Feedback Coding Scheme (Fig. 3).
Decoder: The decoder runs K driven by the channel output y.
x̂u (k + 1) = Au x̂u (k) + Bu y(k), x̂u (0) = 0.
It only produces an estimate of the initial condition of the encoder
x̂u 0 (k) = A−k−1
x̂u (k + 1).
u
Encoder: The encoder runs the following dynamics driven by the initial state, i.e., the message,
x̃u (k + 1) = Au x̃u (k), x̃u (0) = xu 0 ,
ũu (k) = Cu x̃u (k).
It receives y and runs dynamics driven by the received feedback y,
xs (k + 1)
= As xs (k) + Bs y(k), xs (0) = 0
x̂u (k + 1) = Au x̂u (k) + Bu y(k), x̂u (0) = 0,
and produces a signal
h
i xs (k)
.
û(k) = Cs Cu
x̂u (k)
Fig. 3. Representation II: Decomposition of filter K into the feedback encoder (Alice) and decoder (Bob). The eavesdropper channels are not
included.
Then, the encoder produces the channel input
u(k) = ũu (k) − û(k).
By comparing the two representations, we see that the only difference comes from the feedback signal. In
Representation I, the feedback signal û is generated by the decoder, while in Representation II, the feedback signal
is simply the raw channel output. The equivalence can be directly verified by comparing the channel inputs u
(encoder) and the estimate of the message x̂u 0 (decoder) of the two representations.
Proposition 2. For a given message xu 0 and a sequence of additive noise wk1 (k ≥ 1), the Representation I and
Representation II of the proposed coding scheme produce identical channel input u(k) and message estimate x̂u 0 (k)
for ∀k.
Remark 3. It is important to notice that the “equivalence” only holds for such a channel without an eavesdropper.
This is because in our model the eavesdropper can access the feedback link. In Representation II, since the channel
output is directly fed back to Alice, the eavesdropper’s access to both the channel output and the feedback link
has no difference from the access to the channel output only. However, this is clearly not true for Representation I,
in which the eavesdropper can extract more useful information from the decoding process (Bob) by accessing the
feedback link.
We next provide some insight on the proposed coding scheme. All the discussions in the rest of this section hold
for both representations.
First of all, in the above proposed coding schemes, since the closed loop is stable, u(k) goes to zero with time
if the noise is not present. Given that
u(k) =ũu (k) − û(k)
(16)
=Cu (x̃u (k) − x̂u (k)) −Cs xs (k),
and the system is observable, we have x̂u (k) → −x̃u (k). Thus, −x̂u 0 (k) is an estimate at time k of x̃u (0) =
xu,0 . In the presence of noise, Theorem 4.3 in [36] shows that the above coding scheme leads to x̂u 0 (k − 1) v
0
−k 0
0
N (−xu,0 , A−k
u E[x̂u (k)x̂u (k) ](Au ) ) for large k, where E[x̂u (k)x̂u (k) ] represents the state covariance matrix. Note
that, since the stable closed-loop system is observable and controllable, the matrix E[x̂u (k)x̂u (k)0 ] is positive definite
(which converges to a steady state value and is independent from the initial state of the system) and so is
0
−k 0
A−k
u E[x̂u (k)x̂u (k) ](Au ) . Since the eigenvalues of Au are strictly outside the unit disc, it is straightforward to
see that the state covariance matrix converges to zero (e.g. element-wise) as k goes to infinity.
In addition, in the above coding scheme the message index xu0 ∈ Rm is allocated at the centroid of an unit
hypercube in the coordinate system depending on Au ∈ Rm×m . We refer interested readers to Theorem 4.3 in [36]
for details. For scalar Au ∈ C, the unit hypercube becomes an interval, e.g. [−0.5, 0.5]. That is, 2nRs messages are
represented by the middle points of equally divided 2nRs subintervals within [−0.5, 0.5]. In the end, we note that
the causality of the feedback channel is captured by the one-step delay of the state transition on xs and x̂u .
The next lemma indicates that the proposed coding scheme K is C f b -achieving. The proof is omitted as it follows
directly from Theorem 4.3 and Theorem 4.6 in [36].
Proposition 3. Consider stationary Gaussian channels in (4). Given a filter Q(eiθ ) ∈ RH 2 , the proposed coding
scheme based on the decomposition of K = −Q(1 + Q)−1 achieves a reliable transmission rate (in the sense of
Shannon) at
1
2π
Z π
−π
m
log |1 + Q(eiθ )|dθ = ∑ log |λi (Au )| bits/channel use,
i=1
and has double exponential decaying error probability.
This proposition indicates that the achievable rate of the proposed coding schemes (Representation I and Representation II) is explicitly characterized by the objective function in (4), leading to the fact that the proposed
coding schemes are capacity-achieving if Q(eiθ ) is an optimal solution to (4). Moreover, this achievable rate is
only determined by the unstable eigenvalues of the system, implying that the capacity-achieving coding scheme K
must be unstable although the closed-loop system is stable.
Thus far, we have presented explicit feedback codes (in the state-space representation) that can achieve the
feedback capacity. In the next section, we show that a variant of this coding scheme K can lead to the asymptotic
zero leakage of the message to Eve. This implies Csc = C f b .
IV. M AIN R ESULTS
In this section, we first present our results for the ARMA(k) feedback channel with an eavesdropper, and then
extend our results to the case with quantization noise in feedback.
A. ARMA(k) Feedback Channel with an Eavesdropper
We first present our new development on the properties of the feedback coding scheme for ARMA(k) Gaussian
channels without the presence of an eavesdropper. We then use these properties to establish our main theorem that
characterizes the feedback secrecy capacity and its achieving coding scheme.
The following result shows that, by choosing the particular m-step initializations (in the state-space representation)
for the proposed coding scheme, the channel inputs (k ≥ m + 1) are only determined by the past additive Gaussian
noise w, a fact that is vital to guarantee the asymptotic secrecy from Eve.
Proposition 4. For the proposed coding scheme in Fig. 2 or Fig. 3, assume the first m-step channel inputs
m+1 x
m+1 refers to matrix A with power m + 1), the estimate message x̂ (m) = A−m−1 ym
um
u
u 0 (where Au
u0
u
1 = Au
1
(or equivalently, x̂u (m + 1) = ym
1 ) and xs (m + 1) = 0, where m is the number of the eigenvalues of matrix Au . Then
the induced channel inputs u(k) for k ≥ m + 1 are only determined by the past Gaussian noise wk−1
1 .
Proof. See Section V-A.
Proposition 5. With the initializations defined in Proposition 4, the coding scheme K in Fig. 2 or Fig. 3 remains
to be C f b -achieving.
Proof. The proof follows directly from two facts. On the one hand, all these initials have no effect on the average
transmission power on channel inputs, which only depend on the steady state of the underlying LTI systems. On
the other hand, these initials do not change the reliable transmission rate of the coding scheme K, which is defined
in an asymptotic manner and only determined by the unstable eigenvalues in Au (Proposition 3).
This result provides a variant of the coding scheme K, in which the first m steps (k = 1, 2, · · · , m) of the channel
input u are different from others (k ≥ m + 1) and the states x̂u and xs (m + 1) of the coding dynamics are particularly
initialized at m + 1. By doing so, after the initialization phase (k ≤ m), Alice only transmits the statistical learning
information of “noise” rather than the message for k ≥ m + 1 in the forward channel, such that if only having access
to the forward channel (which is essentially the case for Representation II) Eve cannot learn innovative information
of the message through its noisy wiretap channels except in the first m-steps.
Theorem 2. Consider the ARMA(k) Gaussian wiretap channel with feedback (Fig. 1) under the average channel
input power constraint P > 0. Then,
1) the feedback secrecy capacity equals the feedback (Shannon) capacity, i.e., Csc = C f b , where C f b is obtained
from Section III-A;
2) the feedback secrecy capacity is achieved by the C f b -achieving feedback coding scheme K ( Representation
m+1 x , x̂ (m) = A−m−1 ym (or equivalently, x̂ (m + 1) = ym ), and x (m + 1) = 0.
II) with um
u
s
u0 u0
u
1 = Au
1
1
Proof. See Section V-B.
This theorem shows that there exists a feedback coding scheme such that the secrecy requirement can be achieved
without loss of the communication rate of the legitimate users. In addition, Section III-B provides such a feedback
coding scheme that achieves the feedback secrecy capacity. In particular, a Csc -achieving feedback code can be
constructed from the optimal Q in (4) by following the procedures in Section III-B (Representation II) with the
initializations defined in Proposition 4. The next corollary shows that the well-known S-K scheme [6] is a special
case of our proposed coding scheme.
Corollary 1. Consider the AWGN wiretap channel with feedback (Fig. 1) under the average channel input power
constraint P > 0. Assume that the additive
mean and variance σw2 > 0. Then the proposed coding
r noise w has zero
√ 2
p
A −1
P+σw2
scheme K (Representation II) with Au =
, Bu = − Auu , Cu = − A2u − 1, and As = Bs = Cs = 0 becomes
σ2
w
the original S-K scheme, and achieves the secrecy capacity Cs f = C f b = 12 log(1 + σP2 ).
w
Proof. See Section V-C.
This corollary recovers Theorem 5.1 in [23], showing that the well-known S-K scheme not only achieves the
feedback capacity but also automatically provides the secrecy from the eavesdropper.
Remark 4. As discussed in Remark 3, if the eavesdropper has no access to the feedback link, then the above results
hold for the coding scheme in Representation I as well. That is, the feedback secrecy capacity Csc is achieved by
the C f b -achieving feedback coding scheme K (Representation I) with the selected initializations. The proof of this
argument follows directly from the proof of Theorem 2 by deleting the eavesdropper’s access to the feedback link.
Therefore, we conclude that the channel-output-based feedback coding scheme Representation II is highly necessary
to remove the advantage of the eavesdropper’s access to the feedback link. If no such access, the “equivalence”
between Representation I and Representation II holds for channels with the eavesdropper.
B. Feedback with Quantization Noise
In this section, we extend our result to Gaussian channels with quantized feedback. It is noteworthy that the
capacity of colored Gaussian channels with noisy feedback remains an open problem [37] [38] [39], even when
simplified to the quantization feedback. Therefore, in this paper, as an initial step towards the secrecy capacity
of noisy feedback Gaussian channels, we focus on AWGN channels with quantized feedback. In [40], the authors
presented a linear coding scheme featuring a positive information rate and a positive error exponent for AWGN
channels with feedback corrupted by quantization or bounded noise. In what follows, we show that our proposed
linear coding scheme Representation II, when specified to the AWGN channel with quantized feedback, converges
to the scheme in [40] and, more importantly, leads to a positive secrecy rate. Furthermore, this achievable secrecy
rate converges to the capacity of the AWGN channel as the amplitude of the quantization noise decreases to zero.
Firstly, we define a memoryless uniform quantizer with sensitivity σq as follows [40].
Definition 1. Given a real parameter σq > 0, a uniform quantizer with sensitivity σq is a function Φσq : R → R
defined as
Φσq (y) = 2σq b
y + σq
c,
2σq
Fig. 4. Coding Structure for the AWGN channel with quantized feedback. The quantization noise q can be recovered by the decoder to help
decoding.
where b·c represents the floor function. Then, the quantization error at instant k, i.e., the feedback noise, is given
by
q(k) = Φσq (y(k)) − y(k).
Notice that, for a given channel output y(k), the quantization noise q(k) can be recovered by the decoder as we
assume the decoder knows the quantization rule. In other words, the decoder can get access to both the channel
outputs and the feedback noise while the encoder can only get access to the corrupted channel output. On the
other hand, note that with the quantized feedback the coding schemes Representation I and Representation II are
not equivalent any more due to the different feedback signals. The Representation I may not be applicable here.
Therefore, we tailor the coding scheme Representation II to the AWGN channel with quantized feedback as follows
(Fig. 4). We first let As = Bs = Cs = 0.
Decoder: The decoder runs K driven by the sum of the channel output y and the quantization noise q.
x̂u (k + 1) = Au x̂u (k) + Bu (y(k) + q(k)), x̂u (0) = 0.
It produces an estimate of the initial condition of the encoder
x̂u (k + 1).
x̂u 0 (k) = A−k−1
u
Encoder: The encoder runs the following dynamics
x̃u (k + 1) = Au x̃u (k), x̃u (0) = xu 0 ,
ũu (k) = Cu x̃u (k).
It receives y + v and duplicates the decoding dynamics,
x̂u (k + 1) = Au x̂u (k) + Bu (y(k) + q(k)), x̂u (0) = 0,
and produces a signal,
û(k) = Cu x̂u (k).
Then, it produces the channel input
u(k) = ũu (k) − û(k).
We next show that the above coding scheme can achieve a positive secrecy rate, which converges to the AWGN
capacity as the feedback noise σq decreases. The following definition will be used to characterize this secrecy rate.
Definition 2. [40] For the given positive real parameters σw2 , σq and the power constraint P, define a parameter
rq as follows.
1) If 4σq ≤ P, rq is the nonnegative real solution of the following equation,
p
√
σw 22rq − 1 = P − σq (1 + 2rq ).
2) If 4σq > P, then rq = 0.
It is easy to check that rq satisfies the following three properties [40]
1) rq converges to the AWGN capacity as σq decreases, i.e.,
lim rq =
σq →0+
2) If σq =
√
P
2 ,
we have rq = 0.
P
1
log(1 + 2 ).
2
σw
√
√
P
P
3) If P max{σw2 , σq2 }, rq ' log( σw +σ
). In other words, the ratio of rq and log( σw +σ
) converges to 1 as
q
q
P → ∞.
Theorem 3. Consider an AWGN channel with uniformly memoryless quantized feedback defined in Definition
1, where the channel input power constraint is P > 0 and the noise variance of the AWGN channel and the
quantization sensitivity in the feedback link are assumed to be σw2 and σq , respectively. Assume u(1) = A2u xu 0 , and
x̂u 0 (1) = A−2
u (y(1) + q(1)) (or equivalently, x̂u (2) = y(1) + q(1)). Then, the above proposed coding scheme with
1
Au = 2r , Bu = −1,Cu = Au − Au
and As = Bs = Cs = 0 achieves a secrecy rate r for all r < rq (rq is defined in
Definition 2).
Proof. See Section V-D.
Combined with the aforementioned property (1) on rq , this theorem implies that the achievable feedback secrecy
rate of the proposed coding scheme converges to the AWGN capacity as σq decreases to zero.
V. T ECHNICAL P ROOFS
In this section, we present the omitted proofs in Sections IV.
A. Proof of Proposition 4
Based on the Representation I of the proposed feedback coding scheme, we have
u(k) =ũu (k) − û(k)
=Cu (x̃u (k) − x̂u (k)) −Cs xs (k)
=Cu (Au x̃u (k − 1) − Aku x̂u 0 (k − 1)) −Cs xs (k)
(17)
=···
=Cu (Aku xu 0 − Aku x̂u 0 (k − 1)) −Cs xs (k)
=Cu Aku (xu 0 − x̂u 0 (k − 1)) −Cs xs (k).
Next, for k ≥ m + 1 where the initial steps have passed and the signals evolve as described in the coding scheme,
it yields
x̂u 0 (k) =A−k−1
x̂u (k + 1)
u
=A−k−1
(Au x̂u (k) + Bu y(k))
u
−k−1
=A−k
Bu y(k)
u x̂u (k) + Au
=x̂u 0 (k − 1) + A−k−1
Bu y(k)
u
=x̂u 0 (k − 1) + A−k−1
Bu (u(k) + w(k))
u
(18)
(a)
= x̂u 0 (k − 1) + A−k−1
Bu (Cu Aku xu 0 −Cu Aku x̂u 0 (k − 1) −Cs xs (k) + w(k))
u
=x̂u 0 (k − 1) − A−k−1
BuCu Aku x̂u 0 (k − 1) + A−k−1
BuCu Aku xu 0 + A−k−1
Bu (w(k) −Cs xs (k)) + xu 0 − xu 0
u
u
u
=(I − A−k−1
BuCu Aku )x̂u 0 (k − 1) − (I − A−k−1
BuCu Aku )xu 0 + A−k−1
Bu (w(k) −Cs xs (k)) + xu 0
u
u
u
=(I − A−k−1
BuCu Aku )(x̂u 0 (k − 1) − xu 0 ) + A−k−1
Bu (w(k) −Cs xs (k)) + xu 0 ,
u
u
where step (a) follows from (17). Let αk = I − A−k−1
BuCu Aku and βk = A−k−1
Bu . Moving xu 0 to the left side, we
u
u
have
x̂u 0 (k) − xu 0 = αk (x̂u 0 (k − 1) − xu 0 ) + βk (w(k) −Cs xs (k)).
By iterating the above equation, for k ≥ m + 1, we obtain
k
k
x̂u 0 (k) − xu 0 =
αi (x̂u 0 (m) − xu 0 ) +
∏
k
∑ ∏
α j βi (w(i) −Cs xs (i)).
i=m+1 j=i+1
i=m+1
m+1 x
−m−1 ym , we have
where we assume αk+1 = 1. Given um
u 0 and x̂u 0 (m) = Au
1 = Au
1
m
−m−1 m
x̂u 0 (m) = A−m−1
(um
w1 .
u
1 + w1 ) = xu 0 + Au
Then, it yields
k
x̂u 0 (k) − xu 0 =
∏
i=m+1
αi A−m−1
wm
u
1 +
k
k
∑ ∏
i=m+1 j=i+1
α j βi (w(i) −Cs xs (i)).
Furthermore, given xs (m + 1) = 0, from (17) we have
u(m + 1) = −Cu Am+1
(x̂u 0 (m) − xu 0 ) −Cs xs (m + 1)
u
= −Cu Am+1
(x̂u 0 (m) − xu 0 )
u
(19)
= −Cu wm
1.
Now, for k ≥ m + 2, we present the channel inputs u(k) and recall the evolution of xs (k) as follows,
u(k) = −Cu Aku (x̂u 0 (k − 1) − xu 0 ) −Cs xs (k)
k−1
k−1
= −Cu Aku ∏ αi A−m−1
wm
u
1 + ∑
i=m+1
k−1
∏
α j βi (w(i) −Cs xs (i)) −Cs xs (k),
i=m+1 j=i+1
(20)
and
xs (k) =As xs (k − 1) + Bs (u(k − 1) + w(k − 1)).
Starting with u(m + 1) = −Cu wm
1 and xs (m + 1) = 0, the above coupled iterations induce values of u(k) and xs (k)
that depend only on noise wk−1
1 . Therefore, for k ≥ m + 2
u(k) , φk (wk−1
1 ),
(21)
where the mapping φk : Rk−1 → R is defined by the iterations in (20). Combining with (19), we conclude that, for
k ≥ m + 1, u(k) only depends on wk−1
1 . Due to the equivalence between Representation I and Representation II
of the coding scheme, this result directly holds for Representation II.
B. Proof of Theorem 2
Following from Proposition 3, the proposed coding scheme derived from the optimal filter Q can achieve the
feedback capacity C f b . Then, in what follows, we only need to show that under the selected initializations of the
proposed coding scheme, the following secrecy requirement is satisfied,
1
I(xu 0 ; zn1 , z̃n1 , ẑn1 ) = 0.
n→∞ n
lim
In this proof, we use the Representation II of the proposed coding scheme. Following from the model in Fig. 1
and (19), (21) in the proof of Proposition 4, the three inputs z(k), z̃(k) and ẑ(k) to Eve for k ≥ 1 are given by
m
m
m+1
zm
xu 0 + vm
1 = u1 + v1 = Au
1,
z(m + 1) = u(m + 1) + v(m + 1) = −Cu wm
1 + v(m + 1),
z(k) = φk (wk−1
1 ) + v(k),
(22)
k ≥ m + 2,
and
m
m
n
m+1
n
z̃m
xu 0 + ṽm
1 = u1 + ṽ1 + w1 = Au
1 + w1 ,
z̃(m + 1) = u(m + 1) + ṽ(m + 1) + w(m + 1) = −Cu wm + ṽ(m + 1) + w(m + 1),
z̃(k) = φk (wk−1
1 ) + ṽ(k) + w(k),
k ≥ m + 2,
(23)
and
m
m
n
m+1
n
ẑm
xu 0 + v̂m
1 = u1 + v̂1 + w1 = Au
1 + w1 ,
ẑ(m + 1) = u(m + 1) + v̂(m + 1) + w(m + 1) = −Cu wm + v̂(m + 1) + w(m + 1),
ẑ(k) = φk (wk−1
1 ) + v̂(k) + w(k),
(24)
k ≥ m + 2,
m+1 x
where um
u 0 is the selected initializations. Recall that noises v(k), ṽ(k) and v̂(k) are additive noises defined
1 = Au
in (2).
˜ d}
ˆ + 1 and k ≥ m + 1 we have
Then, for n ≥ k + max{d, d,
(a)
h(xu 0 |zn1 , z̃n1 , ẑn1 ) ≥h(xu 0 |zn1 , z̃n1 , ẑn1 , wn1 , vnm+1 , ṽnm+1 , v̂nm+1 )
(b)
m m n n
n
n
=h(xu 0 |zm
1 , z̃1 , ẑ1 , w1 , vm+1 , ṽm+1 , v̂m+1 )
(c)
m+1
m m+1
m n n
n
n
xu 0 + ṽm
xu 0 + v̂m
=h(xu 0 |Am+1
xu 0 + vm
u
1 , Au
1 + w1 , Au
1 + w1 , w1 , vm+1 , ṽm+1 , v̂m+1 )
(25)
m+1
m+1
n n
n
n
xu 0 + ṽm
xu 0 + v̂m
=h(xu 0 |Am+1
xu 0 + vm
u
1 , Au
1 , Au
1 , w1 , vm+1 , ṽm+1 , v̂m+1 )
(d)
˜
˜
m+1
m+1
n m+d m+d m+d
= h(xu 0 |Am+1
xu 0 + vm
xu 0 + ṽm
xu 0 + v̂m
u
1 , Au
1 , Au
1 , w1 , vm+1 , ṽm+1 , v̂m+1 )
˜
˜
m+1
m+1
m+d m+d m+d
=h(xu 0 |Am+1
xu 0 + vm
xu 0 + ṽm
xu 0 + v̂m
u
1 , vm+1 , ṽm+1 , v̂m+1 ),
1 , Au
1 , Au
where (a) follows from the fact that conditioning does not increase entropy, step (b) follows from Proposition 4,
step (c) follows from (22), (23) and (24), step (d) follows from the finite memory of wiretap channel noise (v, ṽ
,v̂) and the last step follows from the fact that the noise w is assumed to be independent from others.
Then, we obtain
I(xu 0 ; zn1 , z̃n1 , ẑn1 )
=h(xu 0 ) − h(xu 0 |zn1 , z̃n1 , ẑn1 )
˜
˜
m+1
m+1
m+d m+d m+d
xu 0 + ṽm
xu 0 + v̂m
≤h(xu 0 ) − h(xu 0 |Am+1
xu 0 + vm
u
1 , Au
1 , Au
1 , vm+1 , ṽm+1 , v̂m+1 )
˜
(26)
˜
m+1
m+1
m+d m+d m+d
xu 0 + ṽm
xu 0 + v̂m
=I(xu 0 ; Am+1
xu 0 + vm
u
1 , Au
1 , Au
1 , vm+1 , ṽm+1 , v̂m+1 )
=I(xu 0 ; Axu 0 + B),
˜
ˆ
ˆ × m zero matrix) and B = [vm , ṽm , v̂m , vm+d , ṽm+d , v̂m+d ].
where A = [Am+1
, Am+1
, Am+1
,00]T (00 is an (d + d˜ + d)
u
u
u
1 1 1 m+1 m+1 m+1
Recall that message xu 0 is uniformly selected from {1, 2, · · · , 2nRs } that are equally spaced in an m-dimensional
unit hypercube. The covariance matrix of xu 0 is
1
12 Im
as n → ∞. Following from the fact that for a fixed covariance
a vector Gaussian input distribution maximizes the mutual information, we obtain the following upper bound,
1
1
lim I(xu 0 ; zn1 , z̃n1 , ẑn1 )) ≤ lim I(xu 0 ; Axu 0 + B)
n→∞ n
n→∞ n
1
= lim
h(Axu 0 + B) − h(Axu 0 + B|xu 0 )
n→∞ n
1
(27)
= lim
h(Axu 0 + B) − h(B)
n→∞ n
1
1
T
T
≤ lim
log det E[BB ] + AA − h(B)
n→∞ 2n
12
= 0.
The last step follows from the fact that the right-hand side upper bound is independent from index n. The proof is
complete.
C. Proof of Corollary 1
Based on the Schalkwijk’s scheme in [6], the channel input (encoder) and the message estimate (decoder) for
k = 1 are given below by using the notations in this paper.
u(1) =Au xu 0 ,
(28)
−1
x̂u 0 (1) =A−1
u y(k) = xu 0 ) + Au w(1).
The dynamics of the Schalkwijk’s coding scheme for k ≥ 2 can be summarized as follows,
q
u(k) = A2u − 1Ak−1
u (x̂u 0 (k − 1) − xu 0 ),
q
x̂u 0 (k) =x̂u 0 (k − 1) − A−k−1
A2u − 1y(k),
u
r
P+σw2
where Au =
and σw is the variance of the additive white Gaussian noise in the forward channel.
σ2
(29)
w
In our coding scheme, for k = 1,
u(1) =A2u xu 0 ,
(30)
−2
x̂u 0 (1) =A−2
u y(1) = xu 0 + Au w(1),
Plugging these selected parameters into (17) and (18), we have channel inputs and the message estimate of our
proposed coding scheme as follows,
q
u(k) = A2u − 1Aku (x̂u 0 (k − 1) − xu 0 ),
q
x̂u 0 (k) =x̂u 0 (k − 1) − A−k−2
A2u − 1y(k).
u
(31)
By scaling the message xu 0 and the corresponding estimate x̂u 0 by factor Au , we recover the dynamics of the
Schalkwijk’s scheme. Note that this constant scaling on the message index xu 0 have no effect on the reliable
transmission rate and the power cost at channel input. The proof is complete.
D. Proof of Theorem 3
Starting from the decoder with Au = 2r and Bu = −1, we have the decoding dynamics (k ≥ 2)
x̂u (k) = Au x̂u (k − 1) + Bu (y(k − 1) + q(k − 1))
= Au (Au x̂u (k − 2) + Bu (y(k − 2) + q(k − 2))) + Bu (y(k − 1) + q(k − 1))
= A2u x̂u (k − 2) + Au Bu (y(k − 2) + q(k − 2)) + Bu (y(k − 1) + q(k − 1))
= ···
k−1
= Aku x̂u (0) + Bu ∑ Ak−1−i
(y(i) + q(i))
u
i=0
k−1
= Bu ∑ Ak−1−i
(y(i) + q(i))
u
i=0
k−1
= − ∑ 2r(k−1−i) (y(i) + q(i)).
i=0
(32)
Then, the estimate of the initial state of the encoder (i.e., the message index) is
k−1
−r(i+1)
x̂u 0 (k − 1) = A−k
(y(i) + q(i)).
u x̂u (k) = − ∑ 2
(33)
i=0
Next, based on (17) with Cs = 0 and Cu = Au − A1u , we have the dynamics of channel inputs as
u(k) =Cu Aku (xu 0 − x̂u 0 (k − 1))
=Cu Aku (xu 0 − A−k
u x̂u (k))
=Cu Aku (xu 0 − A−k
u (Au x̂u (k − 1) + Bu (y(k − 1) + q(k − 1))))
(34)
=Cu Aku (xu 0 − A−k+1
x̂u (k − 1)) +Cu Bu (y(k − 1) + q(k − 1))
u
(a)
=Au u(k − 1) +Cu Bu (y(k − 1) + q(k − 1))
=2r u(k − 1) + (2−r − 2r )(y(k − 1) + q(k − 1)),
where step (a) follows from the above first two steps. Notice that our coding schemes of the decoder (33) and the
encoder (34) are identical to the coding schemes (2) and (4) in [40]. In addition, Theorem 3.2 in [40] shows that,
for a given rv (Definition 2), the proposed scheme can achieve any transmission rate r with r < rv .
Now, we need to prove that the proposed coding scheme achieves secrecy with regard to the eavesdropper. In
fact, we can directly follow the proof of Proposition 4 and characterize the channel inputs as
u(1) =A2u xu 0 ,
u(2) = −Cu (w(1) + q(1)),
k−1
−i−1
−1
k−1−i
k−2 w(1)
u(k) = −Cu Aku (1 − A−1
B
C
)
+
A
(1
−
A
B
C
)
B
(w(i)
+
q(i))
,
u
u
u
u
u
∑ u
u
u
A2u
i=2
(35)
k ≥ 3.
As a consequence, we note that the channel inputs of the proposed coding scheme only depend on the past forward
channel noise w and feedback quantization noise q. This fact enables us to show that, by following the proof of
Theorem 2, this coding scheme satisfies the secrecy requirement
lim
n→∞
1
I(xu 0 ; zn1 , z̃n1 , ẑn1 ) = 0.
n
To avoid redundancy, we herein only provide sketchy arguments. Details can be directly obtained by following
(25) to (27).
h(xu 0 |zn1 , z̃n1 , ẑn1 ) ≥h(xu 0 |zn1 , z̃n1 , ẑn1 , wn1 , qn1 , vn2 , ṽn2 , v̂n2 )
=h(xu 0 |z1 , z̃1 , ẑ1 , wn1 , qn1 , vn2 , ṽn2 , v̂n2 )
=h(xu 0 |A2u xu 0 + v1 , A2u xu 0 + ṽ1 + w1 , A2u xu 0 + v̂1 + w1 , wn1 , qn1 , vn2 , ṽn2 , v̂n2 )
=h(xu 0 |A2u xu 0 + v1 , A2u xu 0 + ṽ1 , A2u xu 0 + v̂1 , wn1 , qn1 , vn2 , ṽn2 , v̂n2 )
˜
ˆ
1+d 1+d
=h(xu 0 |A2u xu 0 + v1 , A2u xu 0 + ṽ1 , A2u xu 0 + v̂1 , wn1 , qn1 , v1+d
2 , ṽ2 , v̂2 )
˜
ˆ
d
=h(xu 0 |A2u xu 0 + v1 , A2u xu 0 + ṽ1 , A2u xu 0 + v̂1 , v21+d , ṽ21+d , v̂1+
2 ).
(36)
Then, we obtain
I(xu 0 ; zn1 , z̃n1 , ẑn1 )
˜
ˆ
1+d 1+d
≤I(xu 0 ; A2u xu 0 + v1 , A2u xu 0 + ṽ1 , A2u xu 0 + v̂1 , v1+d
2 , ṽ2 , v̂2 )
(37)
=I(xu 0 ; Axu 0 + B),
ˆ × 1 zero matrix) and B = [v1 , ṽ1 , v̂1 , v1+d , ṽ1+d˜, v̂1+dˆ]. The rest of the
where A = [A2u , A2u , A2u ,00]T (00 is an (d + d˜ + d)
2
2
2
proof is omitted as it directly follows from (27).
VI. C ONCLUSION
In this paper, we considered the ARMA(k) Gaussian wiretap channel with feedback and showed that a variant
of the generalized S-K scheme, which is a feedback capacity-achieving code, secures transmissions by itself from
the eavesdropper. Namely, the feedback secrecy capacity equals the feedback capacity without the presence of an
eavesdropper. We further extended our scheme to the AWGN channel with quantized feedback and proved that our
scheme can achieve a positive secrecy rate, which converges to the AWGN channel capacity as the quantization
noise decreases to zero.
We conclude this paper by listing a few related research topics, which can facilitate to illustrate the complete
picture of the secrecy communications with feedback. First of all, it is known that the S-K coding scheme nicely
unifies communications, control and estimation for feedback systems. In this paper, by leveraging the tools from both
control and communications, we showd that (a variant of) the S-K scheme automatically provides secrecy for the
legitimate users. Therefore, understanding the secrecy nature of the S-K scheme from an estimation perspective could
be a missing piece of the work. One possible investigation along this line is to extend the fundamental relation
between the derivative of the mutual information and the MMSE [41], known as I-MMSE, from the open-loop
channels to the feedback channels by invoking the direct information [42], [43] rather than the mutual information
[44]. Furthermore, extend the current results to channels with noisy feedback can be very valuable. Toward this end,
it is necessary to construct a feedback coding scheme with a good achievable rate for noisy feedback channels with
no eavesdropper, which itself is a quite nontrivial problem in general. Finally, extensions to the multi-access colored
Gaussian channels with feedback can be of much interest to the community in the field of secrecy communications.
R EFERENCES
[1] C. Li and Y. Liang, “Secrecy capacity of the first-order autoregressive moving average gaussian channel with feedback,” in Proc. IEEE
International Symposium on Information Theory, pp. 1963–1967, 2017.
[2] C. Shannon, “The zero error capacity of a noisy channel,” Inst.Radio Eng. Trans.Information Theory, vol. IT-2, pp. S8–S19, 1956.
[3] P. Elias, “Channel capacity without coding,” MIT Research Lab. of Electronics, Cambridge, MA, Quarterly Progr. Rep., 1956.
[4] ——, “Networks of Gaussian channels with applications to feedback systems,” IEEE Trans. Inf. Theory, vol. 13, pp. 493 – 501, 1967.
[5] J. P. M. Schalkwijk and T. Kailath, “A coding scheme for additive noise channels with feedback i: No bandwidth constraint,” IEEE Trans.
Inf. Theory, vol. IT-12, no. 2, pp. 172–182, 1966.
[6] J. P. M. Schalkwijk, “A coding scheme for additive noise channels with feedback ii: Band-limited signals,” IEEE Trans. Inf. Theory, vol.
IT-12, no. 2, pp. 183–189, 1966.
[7] S. Butman, “A general formulation of linear feedback communication systems with solutions,” IEEE Trans. Inf. Theory, vol. 15, no. 3,
pp. 392 – 400, 1969.
[8] ——, “Linear feedback rate bounds for regressive channels,” IEEE Trans. Inf. Theory, vol. 22, no. 3, pp. 363 – 366, 1976.
[9] J. Wolfowitz, “Signalling over a Gaussian channel with feedback and autoregressive noise,” J. Appl. Probability, vol. 12, no. 4, pp. 713 –
723, 1975.
[10] L. H. Ozarow, “Random coding for additive gaussian channels with feedback,” IEEE Trans. Inf. Theory, vol. 36, no. 1, pp. 17 – 22, 1990.
[11] ——, “Upper bounds on the capacity of gaussian channels with feedback,” IEEE Trans. Inf. Theory, vol. 36, no. 1, pp. 151 – 161, 1990.
[12] T. M. Cover and S. Pombra, “Gaussian feedback capacity,” IEEE Trans. Inf. Theory, vol. 35, no. 1, pp. 37–43, 1989.
[13] Y. H. Kim, “Feedback capacity of stationary gaussian channels,” IEEE Trans. Inf. Theory, vol. 56, no. 1, pp. 57–85, 2010.
[14] A. Gattami, “Feedback capacity of gaussian channels revisited,” https://arxiv.org/abs/1511.06866, 2015.
[15] C. Li and N. Elia, “Control approach to computing the feedback capacity for stationary finite dimensional gaussian channels,” in Proc.
53th Annual Allerton Conference on Communication, Control, and Computing, pp. 1038–1045, 2015.
[16] ——, “Youla coding and computation of gaussian feedback capacity,” IEEE Trans. Inf. Theory, to appear, 2018.
[17] U. M. Maurer, “Secret key agreement by public discussion from common information,” IEEE Trans. Inf. Theory,, vol. 39, no. 3, pp.
733–742, 1993.
[18] R. Ahlswede and I. Csiszár, “Common randomness in information theory and cryptography, part i: Secret sharing,” IEEE Trans. Inf.
Theory,, vol. 39, no. 4, pp. 1121–1132, 1993.
[19] R. Ahlswede and N. Cai, “Transmission, identification and common randomness capacities for wire-tape channels with secure feedback
from the decoder,” in General Theory of Information Transfer and Combinatorics. Springer Berlin, vol. 4123, pp. 258 – 275, 2006.
[20] L. Lai, H. E. Gamal, and H. V. Poor, “Secrecy capacity of the wiretap channel with noisy feedback,” Proceedings of the 45th Annual
Allerton Conference on Communication, Control and Computing, Monticello, IL,, 2007.
[21] ——, “The wiretap channel with feedback: encryption over the channel,” IEEE Trans. Inf. Theory, vol. 54, no. 11, pp. 5059–5067, 2008.
[22] E. Ekrem and S. Ulukus, “Wiretap channel with secure rate-limited feedback,” in Proc. 42nd Annual Conference on Information Sciences
and Systems, pp. 791 – 796, 2008.
[23] D. Gündüz, D. R. Brown, and H. V. Poor, “Secret communication with feedback,” in Proc. International Symposium on Information Theory
and Its Applications, pp. 1 – 6, 2008.
[24] E. Ardestanizadeh, M. Franceschetti, T. Javidi, and Y.-H. Kim, “Wiretap channel with secure rate-limited feedback,” IEEE Trans. Inf.
Theory, vol. 55, no. 12, pp. 5353 – 5361, 2009.
[25] B. Dai, Z. Ma, and L. Yu, “Feeding back the output or sharing state, which is better for the state-dependent degraded wiretap channel
with noncausal csi at the transmitter?” Entropy, vol. 17, no. 12, pp. 7900–7925, 2015.
[26] B. Dai, A. J. H. Vinck, Y. Luo, and Z. Ma, “Capacity-equivocation region of a special case of wiretap channel with noiseless feedback,”
Journal of Communications and Networks, vol. 17, no. 1, 2015.
[27] G. Bassi, P. Piantanida, and S. Shamai, “On the capacity of the wiretap channel with generalized feedback,” in Proc. IEEE International
Symposium on Information Theory, pp. 1154 – 1158, 2015.
[28] B. Dai, “An improved feedback coding scheme for the wire-tap channel,” ArXiv, https://arxiv.org/pdf/1712.04981.pdf, 2017.
[29] X. Tang, R. Liu, P. Spasojevic, and H. V. Poor, “Multiple access channels with generalized feedback and confidential messages,” Proceedings
of the 2007 IEEE Information Theory Workshop on Frontiers in Coding Theory, Lake Tahoe, CA,, 2007.
[30] B. Dai and Z. Ma, “Multiple access wiretap channel with noiseless feedback,” ArXiv, https://arxiv.org/pdf/1701.04052.pdf, 2017.
[31] T. Kim and H. V. Poor, “The gaussian wiretap channel with noisy public feedback: Breaking the high-snr ceiling,” Proceedings of the
43rd Asilomar Conference on Signals, Systems and Computers, Pacific Grove, CA, 2009.
[32] S. Yang, M. Kobayashi, P. Piantanida, and S. Shamai, “Secrecy degrees of freedom of mimo broadcast channels with delayed csit,” IEEE
Trans. Inf. Theory, vol. 59, no. 9, pp. 5244 – 5256, 2013.
[33] R. Tandon, P. Piantanida, and S. Shamai, “On multi-user miso wiretap channels with delayed csit,” in Proc. IEEE International Symposium
on Information Theory, pp. 211 – 215, 2014.
[34] M. Grant and S. Boyd, “Cvx: Matlab software for disciplined convex programming, version 2.0 beta,” 2013.
[35] ——, “Graph implementations for nonsmooth convex programs, recent advances in learning and control,” Lecture Notes in Control and
Information Sciences, Springer, pp. 95–110, 2008.
[36] N. Elia, “When bode meets shannon: control-oriented feedback communication schemes,” IEEE Transactions on Automatic Control, vol. 49,
no. 9, pp. 1477–1488, 2004.
[37] C. Li and N. Elia, “The information theoretic characterization of the capacity of channels with noisy feedback,” IEEE International
Symposium on Information Theory, pp. 174–178, 2011.
[38] ——, “Upper bound on the capacity of gaussian channels with noisy feedback,” 49th Annual Allerton Conference, pp. 84–89, 2011.
[39] C. Li, “Fundamental limitations on communication channels with noisy feedback: information flow, capacity and bounds,” Iowa State
University Ph.D. dissertation, 2013.
[40] N. C. Martins and T. Weissman, “Coding for additive white noise channels with feedback corrupted by quantization or bounded noise,”
IEEE Trans. Inf. Theory, vol. 54, no. 9, pp. 4274–4282, 2008.
[41] D. Guo, S. Shamai, and S. Verdu, “The interplay between information and estimation measures,” Foundations and Trends in Signal
Processing, vol. 6, no. 4, pp. 243–429, 2012.
[42] J. L. Massey, “Causality, feedback and directed information,” In Proc. Int. Symp. Inf. Theory Applic., pp. 303–305, 1990.
[43] G. Kramer, “Directed information for channels with feedback,” Ph.D. dissertation, Swiss Federal Institute of Technology (ETH), Zurich,
Switzeland, 1998.
[44] H. Asnani, K. Venkat, and T. Weissman, “Relations between information and estimation in the presence of feedback,” In: Como G.,
Bernhardsson B., Rantzer A. (eds) Information and Control in Networks. Lecture Notes in Control and Information Sciences, vol 450.
Springer, 2014.
| 7cs.IT
|
F -JUMPING AND F -JACOBIAN IDEALS FOR HYPERSURFACES
arXiv:1302.3327v2 [math.AC] 30 Apr 2013
LUIS NÚÑEZ-BETANCOURT AND FELIPE PÉREZ
Abstract. We introduce two families of ideals, F -jumping ideals and F -Jacobian ideals,
in order to study the singularities of hypersurfaces in positive characteristic. Both families
are defined using the D-modules Mα that were introduced by Blickle, Mustaţă and Smith.
Using strong connections between F -jumping ideals and generalized test ideals, we give a
characterization of F -jumping numbers for hypersurfaces. Furthermore, we give an algorithm that determines whether certain numbers are F -jumping numbers. In addition, we
use F -Jacobian ideals to study intrinsic properties of the singularities of hypersurfaces. In
particular, we give conditions for F -regularity. Moreover, F -Jacobian ideals behave similarly to Jacobian ideals of polynomials. Using techniques developed to study these two
new families of ideals, we provide relations among test ideals, generalized test ideals, and
generalized Lyubeznik numbers for hypersurfaces.
Contents
1. Introduction
1.1. F-jumping ideals.
1.2. F-Jacobian ideals.
1.3. Further consequences and relations
2. Preliminaries
2.1. Tight closure
2.2. Generalized test ideals
2.3. D-modules
2.4. F e -modules
3. F -jumping ideals and F -jumping numbers
4. F -Jacobian ideals
4.1. Definition for unique factorization domains
4.2. Definition for rings essentially of finite type over an F –finite local ring.
4.3. Examples
5. Further consequences and relations
Acknowledgments
References
1
2
2
3
4
4
5
6
6
7
15
16
25
27
29
31
31
1. Introduction
The aim of this note is to introduce two families of ideals, F -jumping ideals and F Jacobian ideals. Both families of ideals measure singularities in positive characteristic. The
F -jumping ideal is closely related to the test ideal introduced by Hara and Yoshida [HY03],
while the F -Jacobian ideal is strongly connected with the original test ideal defined by
1
Hochster and Huneke [HH90]. These two families are related through Mα , the modules
introduced by Blickle, Mustaţă and Smith [BMS09], and the theory of D-modules and F modules [Lyu97, Yek92]. Roughly speaking, the F -jumping ideal associated to an element
f in a regular F -finite ring, R, and to a rational number α “ per´1 , gives the unique simple
D-submodule of Mα . If α “ 1, then Mα – Rf with the usual D-module structure. Moreover,
the unique simple D-submodule of Mα is R. Under additional hypotheses, the F -Jacobian
ideal determines the sum of the simple F -submodules of Rf {R.
Several of our results are obtained by investigating properties of F e -submodules of Mα
e
through ideals I Ă R such that f r I Ă I rp s . We pay particular attention to ideals such that
e
I “ pI rp s : f r q. This is given by Lyubeznik’s study of F -modules via root maps [Lyu97].
1.1. F-jumping ideals. Let R be an F -finite regular ring, f an element of R and α a
positive real number. The test ideal, τ pf α q, was introduced by Hara and Yoshida as an
analogue of the multiplier ideal in positive characteristic [HY03]. The F -jumping numbers
for f are defined as the positive real numbers c, such that τ pf c´ǫ q “ τ pf c q for every ǫ ą 0.
Blickle, Mustaţă and Smith showed that the F -jumping numbers are rational and form a
discrete set [BMS09]. These numbers encode important information about the singularity of
f (cf. [BFS12]).
Let α be the rational number per´1 , with r a positive integer. Mα is Rf ¨eα as an R-module,
where eα is a formal symbol. The D-module structure depends on the number α [BMS09,
Remark 2.4]. However, it does not depend on the on the representation of α “ per´1 .
One of the main aims of this work is to study the simplicity of Mα as a D-module and
as a F e -module. In Proposition 3.15 and Remark 3.16 we show that Nα “ D ¨ f rαs eα is the
unique simple D-submodule of Mα . This raises the definition of the F -jumping ideal Jpf α q
as the ideal such that
Jpf α qeα “ R ¨ eα X Nα .
We characterize the simplicity of Mα in the following theorem:
Theorem 1.1. Let R be an F –finite regular domain and f P R be a nonzero element. The
following are equivalent:
(i) α is not an F -jumping number;
(ii) Nα “ Mα ;
(iii) Mα is a simple D-module;
(iv) Mα is a simple F e -module;
(v) JF pf α q “ R.
Moreover, the test ideal τ pf α q is the minimal root for Nα and τ pf α´ǫ q is the minimal
root for Mα (see Propositions 3.17 and 3.18). Additionally, in Algorithm 3.19 we present a
process to decide whether α is an F -jumping number.
1.2. F-Jacobian ideals. Suppose that S “ Krx1 , . . . , xn s is a polynomial ring over a perfect
, . . . , BBf
field, K, and f P S. The Jacobian ideal is defined by Jacpf q “ pf, BBf
q. This ideal
x1
xn
plays a fundamental role in the study of singularity in zero and positive characteristics. In
this case, Jacpf q “ R if and only if R{f R is a regular ring. Another important property
is that Jacpf gq Ă f Jacpgq ` g Jacpf q for f, g P S, which is given by the Leibniz rule. The
equality in the previous containment holds only in specific cases [Fab13, Proposition 8] and
it is used to study transversality of singular varieties [Fab13, FA12].
2
Let R be an F -finite regular local ring. We define the F -Jacobian ideal, JF pf q, to be
the intersection of R ¨ e1 Ď M1 with the sum of the minimal F -submodules in M1 properly
containing the unique simple F -module. The F -Jacobian ideal behaves similarly to the
Jacobian ideal of a polynomial. As the Jacobian ideal, they determine singularity:
‚ if R{f R is F -regular, then JF pf q “ R (Corollary 4.40);
‚ if R{f R is F -pure, then R{f R is F -regular if and only if JF pf q “ R (Corollary 4.42).
‚ If f has an isolated singularity and R{f R is F -pure, then JF pf q “ R if R{f R is
F -regular, and JF pf q “ m otherwise (Proposition 4.43).
In addition, the F -Jacobian ideal also satisfies a Leibniz rule: JF pf gq “ f JF pgq ` gJF pf q
for relatively primes elements f, g P R (Proposition 4.14). The Leibniz rule in characteristic
zero is important in the study of transversality of singular varieties and free divisors over
the complex numbers [Fab13, FA12].
e
e
The F -Jacobian ideals behave well with pe -th powers; JF pf p q “ JF pf qrp s (Proposition
4.19). This is a technical property that was essential in several proofs. This contrasts with
how the Jacobian ideal changes with pe -th powers:
pe ´1
e
pe ´1
e
Bf p
Bf p
e
,...,
q “ Jacpf qrp s
,...,p
q “ f R ‰ pf ,
Jacpf q “ pf , p
Bx1
Bxn
Bx1
Bxn
The F -Jacobian ideal can be computed from the test ideal in certan cases and they are
strongly related (see Proposition 4.33). However, they are not the same (see Example 4.45).
Moreover, the F -Jacobian can be defined for elements such that R{f R is not reduced and
satifies properties that the test ideal does not (eg. Propositions 4.14 and 4.19).
Furthermore, we define the F -Jacobian ideal for a regular F -finite UFD, R, such that
Rf {R has finite length as a D-module (Section 4.1). We also define the F -Jacobian ideal
for a ring which is essentially of finite type over an F -finite local ring (Section 4.2). Both
definitions agree for rings that belog to the previous families (Corollary 4.36).
pe
pe
e Bf
e Bf
pe
pe
1.3. Further consequences and relations. Using ideas and techniques developed to define and study F -jumping ideals and F -Jacobian ideals, we give bounds for the length of
Rf {R as an F -module. One of these bounds is in terms of a flag of test ideals previously
defined by Vassilev [Vas98]. If R is an F -finite regular local ring and R{f R is F -pure, there
exists a strictly ascending chain of ideals
rps
rps
f R “ τ0 Ă τ1 Ă . . . Ă τℓ “ R
such that pτi : τi q Ă pτi`1 : τi`1 q and τi`1 is the pullback of the test ideal of R{τi .
Another bound is given by the generalized Lyubeznik numbers. These are invariants associated to a local ring of equal characteristic defined by the first author and Witt [NBW12].
These numbers are defined using the D-module structure of local cohomology modules. For
a hypersurface of particular interest is the Lyubeznik number
dimpR{f Rq
λ0
p
pR{f R; Lq :“ lengthDpR,Lq´module
Hf1Rp pRq,
p
p is the completion of R with respect to its maximal ideal and L is a coefficient field.
where R
Namely, the relations are:
Theorem 1.2. Let pR, m, Kq be an F -finite regular local ring and f P R be an element,
such that R{f R is reduced. Let f “ f1 . . . fs be a factorization of f into irreducible elements.
3
Then,
lengthF ´mod Rf {R ď lengthR pτ pf 1´ǫ q{τf q ` s,
where τf is the pullback of the test ideal, τ pR{f Rq. Moreover, if R{f R is F -pure, and
0 Ă f R “ τ0 Ă τ1 Ă . . . Ă τℓ “ R
is the flag of ideals previously defined, then
dimpR{f Rq
ℓ ď lengthF ´mod Rf {R ď λ0
pR{f R; Lq
for every coefficient field, L.
Suppose that R is local, R{f R F -pure and K perfect. In this case, λdimpR{f Rq pR{f R; Lq “ 1
if and only if R{f R is F -regular [NBW12, Bli04]. This fact and Theorem 1.2 say that
λdimpR{f Rq pR{f R; Lq is measuring “how far” an F -pure hypersurface is from being F -regular.
2. Preliminaries
In this section, we review the basic facts that we will need about tight closure, generalized
test ideals, D-modules, and F -modules.
2.1. Tight closure. We recall some definitions in tight closure introduced by Hochster
and Huneke [HH90, HH94a], and mention some properties of this theory [Fed87, LS01].
Throughout this section, we assume that R is an F -finite ring.
We say that R is F -pure if for every R-module, M, the morphism induced by the inclusion
of R in R1{p , M bR R Ñ M bR R1{p , is injective. If R Ñ R1{p splits, we say that R is F split. These two properties, F -purity and F -splitting, are equivalent when R is F -finite. In
addition, if pR, m, Kq is a regular local ring and I Ă R is an ideal, then R{I is F -pure if and
only if pI rps : Iq Ć mrps (Fedder’s Criterion, [Fed87, Theorem 1.12]).
If I Ă R is an ideal, the tight closure I ˚ of I is the ideal of R consisting of all elements
z P R for which there exists some c P R that is not in any minimal prime of R, such that
cz q P I rqs for all q “ pe " 0.
We say that R is weakly F -regular if I “ I ˚ for every ideal. If every localization of R
is weakly F -regular, we say the R is F -regular. In general, tight closure does not commute
with localization [BM10], and it is unknown if the localization of a weakly F -regular ring is
again a weakly F -regular ring. That is why the adjective weakly is used.
R is strongly F -regular if for all c P R not in any minimal prime, there exists some q “ pe
such that the morphism of R-modules, R Ñ R1{q , that sends 1 to c1{q splits. Strong F regularity is preserved after localization. In a Gorenstein ring, strong and weak F-regularity
are equivalent.
We define the F -singular locus of R by
SingF pRq “ tP P SpecpRq | RP is not F -regularu.
We define the test ideal of R by
τ pRq “
č
pI : I ˚ q.
IĂR
4
If R is a Gorenstein ring, we have that
τ pRq “
č
I parameter ideal
pI : I ˚ q.
[HH90, Theorem 8.23] [Mat89, Theorem 18.1].
Remark 2.1. Let R be a reduced ring essentially of finite type over an excellent local ring
of prime characteristic. Let τ pRq denote the test ideal of R. We know that for every multiplicative system W Ă R, W ´1 τ pRq “ τ pW ´1 Rq [Smi94, Proposition 3.3] [LS01, Theorem
2.3]. It is worth pointing out that, in this case, τ pRq contains a nonzero-divisor [HH94a,
Theorem 6.1].
If P P SingF pRq, then RP is not F -regular, and so τ pRP q “ τ pRqRP ‰ RP . Therefore,
τ pRq Ă P and P P Vpτ pRqq. On the other hand, if P P Vpτ pRqq, then τ pRP q “ τ pRqRP ‰
RP , and then R is not F -regular. Hence, P P SingF pRq. Therefore, SingF pRq “ Vpτ pRqq.
2.2. Generalized test ideals. Test ideals were generalized by Hara and Yoshida [HY03]
in the context of pairs pR, I c q, where I is an ideal in R and c is a real parameter. Blickle,
Mustaţă, and Smith [BMS08] gave an elementary description of these ideals in the case of a
regular F -finite ring, R. We give the definition introduced by them:
e
e
Given an ideal I in R we denote by I r1{p s the smallest ideal J such that I Ď J rp s [BMS08,
Definition 2.2]. The existence of the smallest such ideal is a consequence of the flatness of
the Frobenius map in the regular case.
We recall some properties that we will use often
e
and
e
pIJq1{p Ă I r1{p s ¨ J r1{p
es
` rpe s ˘1{ps
` s ˘1{pe
e s
I
“ I rp {p s Ă I rp s
¯rpe s
´
r1{pe s
“ D peq f [ÀMBL05, Proposition 3.1],
[BMS08, Proposition 2.4]. In addition, pf q
where D peq “ HomRpe pR, Rq.
Given a non-negative number, c, and a nonzero ideal, I, we define the generalized test
ideal with exponent c by
ď
e
e
τ pI c q “
pI rcp s qr1{p s ,
eą0
where rcs stands for the smallest integer greater than or equal to c.
The ideals in the union above form an increasing chain of ideals; therefore, they stabilize
e
e
because R is Noetherian. Hence for large enough e, τ pI c q “ pI rcp s qr1{p s . In particular,
s
e
τ pf pe q “ pf s qr1{p s [BMS09, Lemma 2.1].
An important property of test ideals is given by Skoda’s Theorem [BMS08, Theorem 2.25]:
if I is generated by s elements and c ď s, then τ pI c q “ I ¨ τ pI c´1 q.
For every nonzero ideal I and every non-negative number c, there exists ǫ ą 0 such that
1
τ pI c q “ τ pI c q for every c ă c1 ă c ` ǫ [BMS08, Corollary 2.16].
A positive real number c is an F-jumping number for I, if τ pI c q ‰ τ pI c´ǫ q for all ǫ ą 0.
All F -jumping numbers of an ideal I are rational, and they form a discrete set; that
is, there are no accumulation points of this set. In fact, they form a sequence with limit
infinity [BMS08, Theorem 3.1]. Then for every positive number α, there is a positive rational
number, β ă α, such that τ pf β q “ τ pf γ q for every γ P pβ, αq. We denote τ pf β q by τ pf α´ǫ q.
5
2.3. D-modules. Given two commutative rings, A and R, such that A Ă R, we define the
ring of A-linear differential operators on R, DpR, Aq, as the subring of HomA pR, Rq, defined
inductively as follows. The differential operators of order zero are induced by multiplication
by elements in R (HomR pR, Rq “ R). An element θ P HomA pR, Rq is a differential operator
of order less than or equal to k ` 1 if θ ¨ r ´ r ¨ θ is a differential operator of order less than
or equal to k for every r P R.
If A, B, and R are commutative rings rings such that A Ă B Ă R, then DpR, Bq Ă
DpR, Aq.
Notation 2.2. If A “ Z, we write DR for DpR, Zq, and we write only D if it is clear which
ring we are working on.
If M is a DpR, Aq-module, then Mf has a structure of a DpR, Aq-module, such that the
natural morphism M Ñ Mf is a morphism of DpR, Aq-modules. Thus, Rf {R is an DpR, Aqmodule.
Ť
If R is a reduced F -finite ring, we have that DR “ ePN HomRpe pR, Rq [Yek92]. We denote
peq
HomRpe pR, Rq by DR . Moreover, if R is an F -finite domain, then R is a strongly F -regular
ring if and only if R is F -split and a simple DR -module [Smi95a, Theorem 2.2].
If R is an F -finite reduced ring, W Ă R a multiplicative system and M a simple DR module, then W ´1 M is either zero or a simple DW ´1 R -module. As a consequence, for every
DR -module of finite length, N,
lengthDW ´1 R W ´1 N ď lengthDR N.
2.4. F e -modules. In this section, we recall some definitions and properties of the Frobenius
functor introduced by Peskine and Szpiro [PS73]. We assume that R is regular. This allows
us to use the theory of F -modules introduced by Lyubeznik [Lyu97].
Every morphism of rings ϕ : R Ñ S defines a functor from R-modules to S-modules,
where ϕ˚ M “ S bR M. If S “ R and ϕ is the Frobenius morphism, F M denotes ϕ˚ M. If
R is a regular ring, F is an exact functor. We denote the e-th iterated Frobenius functor by
F e.
e
p
q. In
Example 2.3. If M is the cokernel of a matrix pri,j q, then F e M is the cokernel of pri,j
e
rpe s
particular, if I Ă R is an ideal, then F R{I “ R{I .
We say that an R-module, M, is an F e -module if there exists an isomorphism of R-modules
ν : M Ñ F e M.
If M is an R-module and β : M Ñ F e M is a morphism of R-modules, we consider
β
Fβ
F 2β
M “ limpM Ñ F e M Ñ F 2e M Ñ . . .q.
Ñ
β
Then, M is an F e -module and M Ñ M is the structure isomorphism. In this case, we say
that M is generated by β : M Ñ FRe M. If M is a finitely generated R-module, we say that
M is an F e -finite F e -module. If β is an injective map, then M injects into M. In this case,
we say that β is a root morphism and that M is a root for M.
Example 2.4.
(i) Since F e R “ R, we have that R is an F e -module, where the structure
morphism ν : R Ñ R is the identity.
6
(ii) For every element f P R and r, e P N, we take α “
F e -module that is generated by
fp
fr
er
fp
r
pe ´1
and define Mα as the F e -finite
2e r
R Ñ R Ñ R Ñ ....
Remark 2.5. This structure of Mα as F e -module depends of the representation per´1 . How1
ever, if we take another representation per1 ´1 , both structures will induce the same structure
1
as F e¨e -module.
We say that φ : M Ñ N is a morphism of F -modules if the following diagram commutes:
φ
M
//
N
νN
νM
FR φ
// F N
FM
The F e -modules form an Abelian category, and the F e -finite F e -modules form a full
Abelian subcategory. Moreover, if M is F e -finite then Mf is also an F e -finite F e -module
for every f P R. In addition, if R is a local ring, then every F e finite F e -module has finite
length as F e -module and has a minimal root [Bli04, Lyu97].
Example 2.6. The localization map R Ñ Rf is a morphism of F -modules for every f P R.
Example 2.7. The quotient of localization map R Ñ Rf is an FR -finite FR -module for
f p´1
every f P R. Rf {R is generated by R{f R Ñ FR pR{f Rq “ R{f p R.
We recall that every F e -submodule M Ă Rf {R is a D-module [Lyu97, Examples 5.2].
We have that Rf {R has finite length as an F -module, because Rf {R has finite length as a
D-module. Let R be an F -finite regular ring. If Rf {R has finite length as D-module, then
Rf {R has finite length as an F -module for every f P R. Therefore, if Rf {R has finite length
as a D-module, then Rf {R has finitely many F -submodules [Hoc07].
3. F -jumping ideals and F -jumping numbers
In this section we define F -jumpings ideals and give some basic properties. In particular,
we relate them with the generalized test ideals and F -jumping numbers.
Notation 3.1. Throughout this section R denotes an F -finite regular domain of characteristic p ą 0 and α denotes a rational number whose denominator is not divisible by p. This
means that α has the form per´1 for some e.
We note that ppaℓ´1 ` . . . ` pe ` 1qr ` α “ peℓ α for every ℓ P N. Let R be a regular local
e
e
e
F -finite ring. Then, φpx1{p q P I for all φ P HompR1{p , R) if and only if x P I rp s .
Lemma 3.2. τ pf pλ q Ă τ pf λ qrps
Proof. We have that
rppj λs
pj
ď
prpj λs
pj
prpj λs
j
jÑ8 p
and lim
“ pλ. Then,
¯r1{pj s ď ´ j ¯r1{pj´1 s ď ˆ´ j ¯r1{pj s ˙rps
ď´
prpj λs
f rp λs
“ τ pf λ qrps
f rp λs
Ă
f
“
τ pf q “
pλ
jPN
by the properties of the ideals I
jPN
jPN
r1{pj s
[BMS08, Lemma 2.4].
7
We will study the F e -module Mα “ Rf eα introduced in by Blickle, Mustaţă and Smith
[BMS09]. As an Rf -modules, Mα is free of rank one generated by eα . Here eα is thought
formally as 1{f α . As an F e -module Mα is generated by the morphism
fr
R ÝÑ F e R “ R,
as in Example 2.4. Moreover, the Frobenius morphism acts on Mα by
e
bp
b
F p m ¨ eα q “ mpe `r ¨ eα .
f
f
e
The structural isomorphism v e : Re b Mα Ñ Mα is given by
e
v e pa b
b
abp
q
“
¨ eα
fm
f mpe `r
whose inverse morphism is
pv e q´1 p
1
c
sppe ´1q`r
¨
e
q
“
cf
b
¨ eα .
α
fs
fs
Mα carries a natural structure as a DR -module, which does not depend of the presentation
se
of α [BMS09, Remark 2.4]. Given P P DR
, we take
c
c
P ¨ m “ pv se q´1 pP ¨ v se p m qq,
f
f
where P acts in Rse b Mα by P pa b uq “ P paq b u. We note that
es
abp
b
v pa b m q “ mpes `rp1`pe `...`peps´1q q ¨ eα
f
f
se
and
c
1
pmppe ´1q`rqp1`pe `...`peps´1q q
q
“
cf
b
¨ eα
fm
fm
1
es
e
eps´1q q
“ cf mpp ´1q`rp1`p `...`p
b m ¨ eα
f
pv se q´1 p
thus
es
e
eps´1q q
P pcf mpp ´1q`rp1`p `...`p
c
P¨ m “
f
f mpes `rp1`pe `...peps´1q q
When m “ 0, the previous expression is equal to
e
q
¨ eα .
eps´1q
q
P pcf rp1`p `...p
q
P ¨c“
¨ eα .
e `...`peps´1q q
rp1`p
f
Definition 3.3. We denote DR f rαs ¨ eα by Nα . We define the F -jumping ideal associated to
f and α as the ideal JFR pf α q of R such that
JFR pf α qeα “ Nα X Reα .
If it clear in which ring we are working, we only write JF pf α q.
Lemma 3.4. The morphism Mα Ñ Mα`1 defined by sending eα ÞÑ f eα`1 is an isomorphism
of F e -modules (as well as D-modules). In particular, Nα – Nα`1 “ DR f rα`1s ¨ eα`1 .
8
Proof. We have the following commutative diagram of R-modules,
0
0
fr
R
f
//
f
R
R{f
f r`p
e ´1
f r`p
e ´1
f rp
R
e
R
R{f
pe
f pr`p
e ´1qpe
f pr`p
e ´1qpe
f
f rp
R
R
//
2e
...
//
p2e
//
0
//
pe
//
//
0
R{f p
2e
f pr`p
e ´1qp2e
f pr`p
e ´1qp2e
//
//
...
...
0
0
We note that
R{f
f r`p
e ´1
Ñ
R{f p
e
f
is zero map because r ě 1 as α ą 0. By taking the limits, we obtain that Mα Ñ Mα`1 is an
isomorphism of F -modules, and hence of DR -modules.
Proposition 3.5. Let ℓ P N. Then, f ℓ JF pf α q Ă JF pf α`ℓ q.
Proof. Let φ : Mα Ñ Mα`ℓ be the morphism of DR -modules given by eα ÞÑ f ℓ eα`ℓ . We have
that φ is an isomorphism by Lemma 3.4. Then,
f ℓ JF pf α qeα`ℓ “ JF pf α qf ℓ eα`ℓ “ φpJF pf α qeα q
“ φpNα X Reα q “ φpNα q X φpReα q “ Nα`ℓ X Rf ℓ eα`ℓ
“ Nα`ℓ X Rf ℓ eα`ℓ Ă Nα`ℓ X Reα`ℓ “ JF pf α`ℓ qeα`ℓ .
Therefore, we obtain that f ℓ JF pf α q Ă JF pf α`ℓ q.
Remark
3.6. Given any ideal I Ď R, we consider the D-module generated by I, DR ¨ Ieα “
Ť
es
sě1 D ¨ Ieα . This is equal to
¯
ď´
es
rp1`...`peps´1q q
rp1`...`peps´1q q
¨ eα .
D pf
Iq{f
sě0
This DR -module intersects with Reα at the ideal
¯
ď´
es
rp1`...`peps´1q q
rp1`...`peps´1q q
D pf
Iq : f
sě0
and contains I.
We define inductively a sequence of ideals I j associated to f α . We take I 1 “ τ pf α q. Given
e
e
I j , we take I j`1 “ ppI j qrp s : f r q. We note that f r I 1 “ f r τ pf α q “ τ pf r`α q “ τ pf p α q Ă
e
e
τ pf α qrp s “ I 1 pf α qrp s , where the second equality follows from Skoda’s Theorem and Lemma
e
3.2. We have that I 1 Ă I 2 and f r I 2 Ă pI 1 qrp s . Inductively, we obtain that I j Ă I j`1 and
e
f r I j`1 Ă pI j qrp s for every j P N.
Since R is Noetherian, there is an N P N such that I N “ I N `1 . By our definition of I j ,
N
I “ I N `j for every j ě N.
9
Definition 3.7. We call the previous sequence of ideals the F -flag of ideals associated to f
and α, and we denote the ideals by IRj pf α q. If it is clear in which ring we are taking this flag,
we write only I j pf α q.
Remark 3.8. By the properties of colon ideals and the flatness of Frobenius, we have that
I j pf α q “ pτ pf α qrp
je s
: f rp1`p
e `...`pepj´1q q
q.
As a consequence, the F -flag of ideals associated to f and α depends on the presentation of
α “ per´1 . However, the union of this flag does not depend on this presentation as the next
proposition shows.
Ť
Proposition 3.9. JF pf α q “ j I j pf α q “ I N pf α q for N " 0.
Proof. By the definition of τ pf α q and properties of test ideals in Section 2.2, τ pf α qrp
´`
˘r1{pse s ¯rpse s
e
se
f rp αs
“ D se pf rp αs q for s ě N for a positive integer N P N.
Then, by Remark 3.6,
č
JF pf α qeα “ Nα Reα
č
“ DR f rαs ¨ eα Reα
¯
ď´
eps´1q q rαs
eps´1q q
eα
“
D es pf rp1`...`p
f q : f rp1`...`p
se s
“
sě0
¯
ď´
es
rp1`...`peps´1q q rαs
rp1`...`peps´1q q
eα
D pf
f q:f
“
sěN
“
¯
ď´
eps´1q q`αs
eps´1q q
eα
D es pf rp1`...`p
q : f rp1`...`p
sěN
¯
ď´
es
pes rαs
rp1`...`peps´1q q
eα
D pf
q:f
“
sěN
“
“
¯
ď´
se
eps´1q q
eα
τ pf α qrp s : f rp1`...`p
sěN
ď
sěN
I s pf α qeα .
Proposition 3.10. Let W Ă R be a multiplicative system. Then,
JFW ´1 R pf α q “ W ´1 JF pf α q
Proof. There exists N such that
e
e
JFR pf α q “ IRN pf p α q “ IRN `1 pf p α q.
We have that IR1 pf p α q “ τ pf α q and that IRj`1 pf p α q “ pIRj pf p α qrp s : f r q. We have that
e
1
pe α
W ´1 IR1 pf p α q “ W ´1 τ pf α q “ τ pf α W ´1 Rq “ IW
q [BMS08, Proposition 2.13], Thus,
´1 R pf
j
´1 j
pe α
pe α
´1
N
pe α
W IR pf q “ IW ´1 R pf q because ´ bR W R is a flat functor. Since IW
q“
´1 R pf
N `1
pe α
IW
pf
q,
we
have
that
´1 R
e
e
e
e
N
p α
JFW ´1 R pf α q “ IW
q “ W ´1 JFR pf α q.
´1 R pf
10
e
Proposition 3.11. Suppose that R is a local ring. Then
p
JFRp pf α q “ JFR pf α qR,
p denotes the completion of R with respect to the maximal ideal.
where R
e
e
e
Proof. We have that IR1 pf p α q “ τ pf α q and that IRj`1 pf p α q “ pIRj pf p α qrp
Proposition 2.13]. In addition,
es
: f r q [BMS08,
e
p “ τ pf α qR
p “ τ pf α Rq
pR
p “ I 1p pf pe α q
IR1 pf p α qR
R
e
p is a flat functor.
p “ I j pf pe α q because ´ bR R
[BMS08, Proposition 2.13]. Thus, IRj pf p α qR
p
R
The rest is analogous to the proof of Proposition 3.10.
fr
e
Every F e -submodule of Mα is given by an ideal I Ă R such that I Ñ I rp s makes sense
e
[Lyu97, Corollary 2.6]. This is f r I Ă I rp s . We set
˙
ˆ
e
2e
f r rpe s f p r rp2e s f p r
Ñ I
Ñ ... .
NI “ lim I Ñ I
Ñ
e
e
Lemma 3.12. Let I, J Ă R be ideals such that f r I Ă I rp s , f r J Ă J rp s and I Ă J. Then the
F e -submodule Mα generated by I is equal to the one generated by J if and only if there exits
epℓ´1q q
eℓ
ℓ P N such that f rp1`...`p
J Ă I rp s .
Proof. Let NI and NJ be the F e -submodules of Mα generated by I and J respectively. We
e
e
fr
have that J{I Ñ J rp s {I rp s generates the F e -module NJ {NI . Since J{I is a finitely generated
R-module, NJ “ NI if and only if there exist there exists ℓ such that
J{I
f rp1`...`p
epℓ´1q q
Ñ
eℓ
J rp s {I rp
eℓ s
is the zero morphism. Therefore, NI “ NJ if and only if there exits ℓ P N such that
epℓ´1q q
e
f rp1`...`p
J Ă I rp s .
fr
Proposition 3.13. DR f rαs ¨ eα is an F e -submodule of Mα . Moreover, JF pf α q Ñ JF pf α qrp
e
fr
and τ pf α q Ñ τ pf α qrp s generate DR f rαs ¨ eα .
Proof. By the definition of τ pf α q and properties of test ideals in Section 2.2, τ pf α qrp
´`
˘r1{pse s ¯rpse s
e
se
f rp αs
“ D se pf rp αs q for s ě N for an integer N P N. Then,
ď
DR ¨ f rαs eα “
D es ¨ f rαs eα
sě0
ď
“
sěN
“
sěN
D es ¨ f rαs eα
¯
ď´
eps´1q q rαs
eps´1q q
¨ eα
D es pf rrp1`...`p
f q{f rp1`...`p
¯
ď´
es
rpes αs
rp1`...`peps´1q q
¨ eα
D pf
q{f
“
sěN
11
se s
es
“
“
¯
ď´
se
eps´1q q
¨ eα .
τ pf α qrp s {f rp1`...`p
sěN
e
As f r τ pf α q “ τ pf r`α q “ τ pf p α q Ă τ pf α qrp
submodule of Mα generated by the morphism
es
by Lemma 3.2, it is a root for the F e -
fr
e
τ pf α q Ñ τ pf α qrp s ,
which is
¯
ď´
se
eps´1q q
¨ eα .
τ pf α qrp s {f rp1`...`p
s
fr
e
Hence, DR ¨f rαs eα is the F e -submodule of Mα generated by the morphism τ pf α q Ñ τ pf rα qrp s .
e
As f r I s pf α q Ă I s pf α qrp s for every s, we have that I s pf α q also generates an F e -submodule
e
fr
of Mα . Since I s`1 pf α q{I s pf α q is the kernel of the morphism R{I s pf α q Ñ R{I s pf α qrp s ,
fr
fr
e
e
I s pf α q Ñ I s pf α qrp s and I s`1 pf α q Ñ I s`1 pf α qrp s
generate the same F e -submodule by Lemma 3.12. Therefore, DR ¨f rαs eα is the F e -submodule
e
fr
generated by JF pf α q Ñ JF pf α qrp s because JF pf α q “ I s pf α q for s " 0.
?
e
Lemma 3.14. Let I Ă R be a nonzero ideal such that I Ď pI rp s : f r q. Then f P I.
se
e
eps´1q
q
Proof. We have that for every s P N, I Ă pI rp s : f rp1`p `...`p
q by flatness of Frobenius
se
rpse s
rp1`pe `...`peps´1q q
and properties of colon ideals. Thus, IRf Ă ppIRf q
:f
q “ pIRf qrp s . Let
es
P Ă R be a prime ideal of R that does not contain f. We have that IRP “ I rp s RP for every
s ą 0. We claim that IRP “ RP for every P ; otherwise,
č
č
se
se
IRP Ă pIRP qrp s Ď pP RP qp “ 0
s
and we know that IRP ‰ 0. Therefore, IRf “ Rf and f P
?
I.
Proposition 3.15. For any nonzero F e -submodule N of Mα , Nα Ă N. In particular, Nα is
a simple F -module.
Proof. As any two R-modules intersect
in Rf , we have that there is a minimal simple F e Ş
e
submodule N. We know that Reα N “ Ieα , where I “ pI rp s : f r q [Lyu97, Corollary 2.6].
Then f P I by Lemma 3.14.
´`
˘r1{pse s ¯rpse s
se
e
se
There exists an N such that τ pf α qrp s “
f rp αs
“ D se pf rp αs q and that
n
τ pf pse `α q “ τ pf α q for every s ě N (cf. Section 2.2).
DR ¨ f n eα “
“
“
ď
sě0
ď
D es ¨ f n eα
sěN
D es ¨ f n eα
ď´
sěN
es
D pf
rp1`...`peps´1q q n
f q{f
12
rp1`...`peps´1q q
¯
¨ eα
“
“
ď´
sěN
ď
sěN
D es pf n`rp1`...`p
˜ˆ
´
f
eps´1q q
n`rp1`...`peps´1q q
q{f rp1`...`p
eps´1q q
¯r1{pse s ˙rpes s
{f
¯
¨ eα
rp1`...`peps´1q q
˙
ďˆ
rp1`...`peps´1q q
n
es s
eps´1q q
`
rp
rp1`...`p
es
es
p
τ pf p
“
¨ eα
q
{f
sěN
Ą
Ą
ď´
sě0
n
es
τ pf pes `α qrp s {f rp1`...`p
ď´
sěN
es
τ pf α qrp s {f rp1`...`p
“ DR ¨ f rαs eα
“ Nα
eps´1q q
eps´1q q
¯
¯
¸
¨ eα
¨ eα
¨ eα
Therefore, Nα Ă N.
Corollary 3.16. Nα is a simple D-module.
Proof. Since JF pf α q commutes with localization, we have that the construction of Nα commutes with localization. Suppose that N Ă Nα is a nonzero D-submodule. After localizing
at a maximal ideal, m, we have that pRzmq´1 N Ă pRzmq´1 Nα . It suffices to prove the
statement in the local case.
If R is a local ring, Nα is a simple F e -module and it is a direct sum of simple DR -modules
[Lyu97, Theorem 3.2 and 5.6]. Since every two R-modules in Rf intersect, Nα must be a
simple DR -module.
Proposition 3.17. Let I Ă τ pf α q be a nonzero ideal such that I Ă pI rp
I “ τ pf α q. In particular, τ pf α qeα is a minimal root for Nα .
es
pes ´1
es
: f r q. Then,
Proof. We have that I Ă pI rp s : f pe ´1 rq for every s P N by flatness of Frobenius and
properties of colon ideals. There exists n P N such that f n P I by Lemma 3.14. We have
that
´
es ´1
es ´1 ¯r1{pes s
n` ppe ´1
r
n` ppe ´1
r
rpes s
f
PI
, and so f
Ă I.
In addition,
n`
and
ppe ´ 1qn ` ppes ´ 1qr
r
pes ´ 1
r
“
ą e
,
e
e
p ´1
p ´1
p ´1
r
ppe ´ 1qn ` ppes ´ 1qr
“ e
“ α.
e
sÑ8
p ´1
p ´1
There exists an N such that
˙r1{peN s
ˆ
eN ´1
r
n` ppe ´1
α
R Ă I Ă τ pf α q.
τ pf q “ f
lim
13
fr
e
Therefore, I “ τ pf α q. The claim about being a minimal root follows because I Ñ I rp s
generates an nonzero F e -submodule of Nα , which is simple by Proposition 3.15.
Using ideas analogous to the previous proof, we recover a result previously obtained by
Blickle [Bli08, Proposition 3.5]
e
Proposition 3.18. Let β “ ab P Qą0 . If α ą β, then f r τ pf β q Ă τ pf β qrp s and it generates
Mα as F -module. Moreover, τ pf α´ǫ qeα is the minimal root of Mα as a F e -module.
Proof. Since
r
pe ´1
ą ab , we have that br ` a ą pe a. Then
pe a
a`br
e
e
f r τ pf β q “ τ pf β`r q “ τ pf pe ´1 q Ă τ pf pe ´1 q “ τ pf p β q Ă τ pf β qrp s .
Thus τ pf β q generates an F -submodule of Mα .
epℓ´1q q
“ α, we can pick an ℓ P N such that rp1 ` . . . pepℓ´1q q ą peℓ β.
Since lim rp1`...p
peℓ
ℓÑ8
We have that
f
rp1`...pepℓ´1q q
Then τ pf
ˆ´
¯r1{pe s ˙rpe s
¯rpe s ˙r1{pe s ˆ´
rp1`...pepℓ´1q q
e
rp1`...pepℓ´1q q
rp1`...pepℓ´1q q
peℓ
qrp s .
“ τ pf
f
f
R
Ă
R“
rp1`pepℓ´1q q
peℓ
pepℓ´1q `...1qr
peℓ
q generates Mα as a F e -module by Lemma 3.12 because
ppepℓ´1q `...1qr
peℓ
ă α.
q Ă τ pf β q, τ pf β q also generates Mα as a F e -module.
Since τ pf
ℓ´1
ℓ
Let I Ă R be an ideal such that f r I Ă I, and that f rpq `...1q R Ă I rq s for some ℓ P N.
Then,
¯r1{peℓ s
´
rp1`...pepℓ´1q q
f
Ă I,
and so,
τ pf α´ǫ q Ă τ pf
rp1`...pepµ´1q q
peµ
¯ eµ
´
epµ´1q q r1{p s
q “ f rp1`...p
Ă I.
Thus, τ pf α´ǫ q is contained in every ideal that generates Mα . Since τ pf α´ǫ q generates Mα ,
τ pf α´ǫ qeα is the minimal root of Mα as an F e -module.
Proof of Theorem 1.1. We have that α “ per´1 is not an F -jumping number if and only if
τ pf α q ‰ τ pf α´ǫ q. This happens if and only if Nα “ Mα by Propositions 3.17 and 3.18, which
is equivalent to JF pαf q “ R. The rest follows from Proposition 3.15 and Corollary 3.16.
Theorem 1.1 gives the following algorithm to decide whether a number of the form
is an F -jumping number for f P R :
Algorithm 3.19.
Input: α “ per´1 and f P R.
Output: R if α is not F -jumping number for f and a proper ideal otherwise.
Process: Compute τ pf α q;
Take J1 “ τ pf α q;
rpe s
Compute Jn`1 “ pJn : f r q until there is an N such that JN “ JN `1 ;
Return: JN .
Example 3.20. Let R “ F13 rx, ys and f “ x2 ` y 3. Therefore:
14
r
pe ´1
11
(i) If α “ 12
, τ pf α q “ px, yqR, J1 “ px, yqR, and J2 “ ppx13 , y 13 q : f 11 q “ R because f 11
is equal to x22 ´ 2x20 y 3 ` 3x18 y 6 ´ 4x16 y 9 ` 5x14 y 12 ´ 6x12 y 15 ´ 6x10 y 18 ` 5x8 y 21 ´
is not an F -jumping number.
4x6 y 24 ` 3x4 y 27 ´ 2x2 y 30 ` y 33 . Then, 11
12
10
α
(ii) If α “ 12 , τ pf q “ px, yqR, J1 “ px, yqR, and J2 “ ppx13 , y 13q : f 10 q “ px, yq because
f 10 is equal to x20 ´ 3x18 y 3 ` 6x16 y 6 ` 3x14 y 9 ` 2x12 y 12 ` 5x10 y 15 ` 2x8 y 18 ` 3x6 y 21 `
is an F -jumping number.
6x4 y 24 ´ 3x2 y 27 ` y 30 . Then, 10
12
Remark
a 3.21. Since
a F -jumping ideals and test ideals commute with localization, we have
α
that JF pf q “ pτ pf α q : τ pf α´ǫ qq. In general, JF pf α q is not equal to pτ pf α q : τ pf β qq. Let
R “ F7 rx, ys, f “ x3 y 2 and α “ 64 . Then, β “ 36 is the biggest F -jumping number smaller
than α. We have that τ pf β q “ xyR and τ pf α q “ x2 yR, and so px2 yR : xyRq “ xR. However,
JF pf α q “ x2 .
e
e
Proposition 3.22. I j pf p α q “ I j´1 pf α qrp s for j ě 2.
e
e
Proof. We will prove that I j pf p α q “ I j´1 pf α qrp s for j ě 2 by induction on j.
If j “ 2,
e
e
e
I 2 pf p α q “ I 2 pf α`r q “ pτ pf α`r qrp s : f p r q
e
e
e
“ pτ pf α`r q : f r qrp s “ pf r τ pf α q : f r qrp s “ τ pf α qrp s “ I 1 pf α q.
Suppose that the claim is true for j. Then,
e
e
e
e
e
I j`1 pf p α q “ pI j pf p α qrp s : f p r q “ pI j pf p α q : f r qrp
e
e
es
e
“ pI j´1 pf α qrp s : f r qrp s “ I j pf α qrp s .
Proposition 3.23. JF pf pα q “ JF pf α qrps .
e
e
Proof. We first note that JF pf p α q “ JF pf α qrp s because
ď
ď
e
e
e
e
JF pf p α q “
I j pf p α q “
I j pf α qrp s “ JF pf α qrp s
j
j
by Lemma 3.22.
e`1
In addition, we have that JF pf α qrps “ pJF pf α qrp s : f pr q. Then, JF pf α qrps defines the
F e -submodule of Mpα generated by the morphism
f pr
JF pf α qrps Ñ JF pf α qrp
e`1 s
.
Since JF pf pα q defines the unique simple F -submodule of Mpα by Propositions 3.13 and
3.15, JF pf pαq Ă JF pf α qrps .
Combining these two observations we have that
e
JF pf p α q Ă JF pf p
e´1 α
qrps Ă . . . Ă JF pf pα qrp
e´1 s
e
e
Ă JF pf α qrp s “ JF pf p α q.
Hence, JF pf pαq “ JF pf α qrps because the Frobenius map is faithfully flat.
4. F -Jacobian ideals
The F -Jacobian ideal of an element f is connected with the minimal F -module of the first
local cohomology of R supported at f . In this section we define the F -Jacobian ideal and
deduce some of its properties.
15
4.1. Definition for unique factorization domains.
Notation 4.1. Throughout this section R denotes an F -finite regular UFD of characteristic
p ą 0 such that Rf {R has finite length as a DpR, Zq-module for every f P R.
This hypothesis is satisfied for every F -finite regular local ring and for every F -finite
polynomial ring [Lyu97, Theorem 5.6].
Lemma 4.2. Let S be a UFD and f P S be an irreducible element. Then, N X M ‰ 0 for
any nonzero S-submodules M, N Ă Sf {S.
Proof. Let a{f β P Mzt0u and b{f γ P Nzt0u, where β, γ ě 1. Since S is a UFD and
f is irreducible, we may assume that gcdpa, f q “ gcdpb, f q “ 1. Then, gcdpab, f q “ 1,
and so ab{f ‰ 0 in Sf {S. We have that ab{f “ bf β´1 pa{f β q “ af γ´1 pb{f γ q ‰ 0. Then,
ab{f P N X M..
Lemma 4.3. Let S be a regular ring of characteristic p ą 0, f P S an element and π : S Ñ
S{f S be the quotient morphism. Let
and
I : tI Ă S | I is an ideal, f P I, pI rps : f p´1q “ Iu
N “ tN Ă Sf {S|N is an F -submoduleu.
Then, the correspondence given by sending N to IN “ π ´1 pN X R{f Rq is bijective, with
f p´1
inverse defined by sending the ideal I P I to the F -module NI generated by I{f S Ñ
F pI{f Sq “ I rps {f p S.
f p´1
Proof. Since φ : R{f R Ñ R{f p R is a root for Rf {R, its F -submodules are in correspondence
with ideals J Ă R{f R such that φ´1 pF pJqq “ J [Lyu97, Corollary 2.6]. We have the following
generating morphisms:
0
Ó
f p´1
0
Ó
fp
2 ´p
fp
2 ´p
fp
2 ´p
0
Ó
fp
3 ´p2
fp
3 ´p2
fp
3 ´p2
I{f R Ñ F pIR{f Rq Ñ F 2 pIR{f Rq Ñ . . .
Ó
Ó
Ó
f p´1
R{f R Ñ
Ó
f p´1
R{I Ñ
Ó
0
R{f p R
Ó
R{I rps
Ó
0
Ñ
Ñ
2
R{f p R
Ó
R{I rp
Ó
0
2s
Ñ ...
Ñ ...
Since J is a quotient I{f R of an ideal, F pJq “ I rps {f p R. Then,
I{f R “ φ´1 pI rps {f p q
“ th P R{f R | f p´1h P I rps {f p u
“ th P R | f p´1h P I rps u{f R
16
“ pI rps : f p´1 q{f
and the result follows.
Lemma 4.4. Let f P R be an irreducible element. Then, there is a unique simple F -module
in Rf {R.
Proof. Since Rf {R is an F -module of finite length, there exists a simple F -submodule M
Rf {R. Let N be an F -submodule of Rf {R. Since M X N ‰ 0 by Lemma 4.2 and M
a simple F -module, M “ M X N. Hence, M is the only nonzero simple F -submodule
Rf {R.
Ă
is
of
Proposition 4.5. Let g P R be an irreducible element and f “ g n for some integer n ě 1.
Then there exists a unique ideal I Ă R such that:
(i) f P I,
(ii) I ‰ f R,
(iii) pI rps : f p´1 q “ I, and
(iv) I is contained in any other ideal satisfying (i),(ii) and (iii).
Proof. We note that Rf {R “ Rg {R. Let I be the ideal corresponding to the minimal simple
F -submodule in given in Lemma 4.4 under the bijection in Lemma 4.3. Then, it is clear
from Lemma 4.3 that I satisfies (i)-(iv).
Definition 4.6. Let g P R be an irreducible element and f “ g n for some integer n ě 1. We
denote the minimal simple submodule of Rf {R by minFR pf q, and we called it the minimal
F -module of f . Let σ : R{f R Ñ Rf {R be the morphism defined by σprasq “ a{f which is
well defined because R is a domain. Since image of σ is R f1 , we abuse notation and consider
R{f R Ă Rf {R. We denote pφσq´1 pminF pf q X R f1 q by JF pf q, and we call it the F -Jacobian
ideal of f . If f is a unit, we take minF pf q “ 0 and JF pf q “ R.
Notation 4.7. If it is clear in which ring we are working, we write JF pf q instead of JFR pf q
and minF pf q instead of minFR pf q.
Proposition 4.8. Let f P R be an irreducible element. Then minF pf q is the only simple
D-submodule of Rf {R.
Proof. We claim that Rf {R has only one simple DR -module. Since Rf {R has finite length
as a D-module, there is a simple D-submodule, M. It suffices to show that for any other
DR -submodule, N Ă M. We have that M X N ‰ 0 by Lemma 4.2, and so M “ M X N Ă N
because M is a simple DR -module. Since minFR pf q is an DR -module [Lyu97, Examples 5.1
and 5.2], we have that M Ă minF pf q.
It suffices to prove that M is an F -submodule of Rf {R. Since R{f R is a domain, we have
that the localization morphism, R{f R Ñ Rm {f Rm , is injective. Then, SuppR pR{f Rq “
SuppR pJq for every nonzero ideal J Ă R{f R. Then,
SuppR pR{f Rq “ SuppR pR{f R X Nq Ă SuppR pNq Ă SuppR pRf {Rq “ SuppR pR{f q
for every R-submodule of Rf {R by Lemma 4.2. Let m denote a maximal ideal such that
f P m. Thus, Mm ‰ 0, and then, Mm is the only simple DRm -module of pRf qm {Rm . Since
17
Rm is a regular local F -finite ring, we have that minFRm pf q is a finite direct sum of simple
DRm -modules [Lyu97, Theorem 5.6]. Therefore, Mm “ minFRm pf q by Lemma 4.2.
Let π : R Ñ R{f R denote the quotient morphism, and I “ π ´1 pR{f R X Mq. We note
rps
that I ‰ f R because R{f X M ‰ 0 by Lemma 4.2. We claim that pIm : f q “ Im for every
maximal ideal. If f P m,
Im {f “ pRm {f Rm q X Mm “ pRm {f Rm q X minFRm pf q “ JFRm pf q{f ;
otherwise, Im “ Rm “ JFRm pf q because f is a unit in Rm Then, pI rps : f p´1q “ I and so I
corresponds to an FR -submodule of Rf {R, NI by Lemma 4.3. Moreover,
fp
f p´1
2 ´p
NI “ limpI{f R Ñ I rps {f p R Ñ . . .q.
Ñ
Since localization commutes with direct limit, we have that for every maximal ideal such
that f P m,
fp
f p´1
2 ´p
rps
{f p Rm Ñ . . .q “ NIm “ NI bR Rm .
Mm “ minFRm pf q “ limpIm {f Rm Ñ Im
Ñ
Therefore, M “ NI because SuppR pMq “ SuppR pR{f q, and it is an F -submodule of Rf {R.
Hence, M “ minFR pf q.
Remark 4.9. If f P R is an irreducible element, then:
(i) minpf q “ minpf n q for every n P N because Rf n {R “ Rf {R,
(ii) JF pf q is the minimal of the family of ideals I containing properly f R such that
pI : f p´1q “ I by Proposition 4.5.
(iii) JF pf q is not the usual Jacobian ideal of f . If S “ F3 rx, y, z, ws and f “ xy ` zw, we
have that the Jacobian of f is m “ px, y, z, wqS. However, m ‰ pmrps : f 2 q.
(iv) JF pf q “ R if and only if Rf {R is a simple F -module by the proof of Proposition 4.5
and Lemma 4.3.
(v) JF pf q “ R if and only if Rf {R is a simple DR -module by Proposition 4.8.
Proposition 4.10. Let fi , . . . fℓ P R be irreducible relatively prime elements and f “
f1 ¨ ¨ ¨ fℓ . Then minF pfi q is an F -submodule of Rf {R. Moreover, all the simple F -submodules
of Rf {R are minF pf1 q, . . . , minF pfℓ q.
Proof. The morphism Rfi {R Ñ Rf {R, induced by the localization map Rf1 Ñ Rf , is a
morphism of F -finite F -modules given by the diagram:
0
0
R{fi R
fip´1
//
fip
R{fip R
2
f1p ...fˆip ¨¨¨fℓp
R{f R
f p´1
0
//
2
fip
2
R{fip R
f1p ...fˆip ¨¨¨fℓp
//
2 ´p
R{f p R
fp
2 ´p
//
3
2
3 ´p2
//
3
...
f1p ...fˆip ¨¨¨fℓp
2
R{f p R
fp
3
3 ´p2
//
...
Then minF pfi q is a simple F -submodule of Rf {R. Let N be an F -submodule of Rf {R, and
a{f1β1 ¨ ¨ ¨ fℓβℓ P Nzt0u. Since fi is irreducible, we may assume that gcdpa, fi q “ 1 and βi ‰ 0
for some i “ 1 . . . , ℓ. Thus, a{fi P N XRfi {R and a{fi ‰ 0. Then, minF pfi q Ă N XRfi {R Ă N.
In particular, if N is a simple F -submodule, then N “ minF pfi q.
18
Remark 4.11. As a consequence of Lemma 4.10, we have that
minF pf1 q ‘ . . . ‘ minF pfℓ q P Rf {R
because Rg X Rh “ R for all elements g, h P R such that gcdpg, hq “ 1.
Definition 4.12. Let fi , . . . fℓ P R be irreducible relatively prime elements, f “ f1β1 ¨ ¨ ¨ fℓβℓ ,
and π : R Ñ R{f R be the quotient morphism. We define minF pf q by
minF pf1 q ‘ . . . ‘ minF pfℓ q,
and we called it the minimal F -module of f . Let σ : R{f R Ñ Rf {R be the morphism
defined by σprasq “ a{f which is well defined because R is a domain. Since image of σ is
R f1 , we will abuse notation and consider R{f R Ă Rf {R. We denote pφσq´1 pminF pf q X R f1 q
by JF pf q, and we call it the F -Jacobian ideal of f .
Remark 4.13. In the local case, minF pf q is the intersection homology D-modules LpR{f, Rq
previously defined by Blickle [Bli04, Theorem 4.5].
Proposition 4.14. Let f, g P R be relatively prime elements. Then,
JF pf gq “ f JF pgq ` gJF pf q.
Moreover, f JF pgq X gJF pf q “ f gR.
Proof. We consider Rf {R and Rg {R as F -submodules of Rf g {R, where the inclusion is given
by the localization maps, ιf : Rf Ñ Rf g and ιg : Rg Ñ Rf g . Let π : R Ñ R{f gR and
ρ : R Ñ R{f R be the quotient morphisms. The limit of the morphism induced by the
diagram
0
0
0
f p´1
R{f R
//
R{f p R
g
gp
R{f gR
//
2 ´p
2
R{f p R
//
2
pf gqp´1
fp
R{f p g p R
gp
p2 ´p
pf gq
//
fp
3 ´p2
...
//
3
2
R{pf gqp R
pf gqp
3 ´p2
//
...
is ιf . Moreover, under this correspondence
0
0
JF pf q{f R
f p´1
//
JF pf q
rp2 s
g
gJF pf q{f gR
0
{f R
gp
pf gqp´1
//
fp
p
2 ´p
//
JF pf q
rp2 s
2
gp
g p JF pf qrps {f p g p R
pf gqp
2 ´p
//
2
2
2
{f p R
...
//
3
2
g p JF pf qrp s {pf gqp R
//
...
induces the isomorphism of F -modules, ιf : minF pf q Ñ ιf pminF pf qq. We have that
gpJF pf qq “ π ´1 pminF pf q X R{f gRq Ă π ´1 pminF pf q X R{f gRq “ JF pf gq.
In addition,
pg p JF pf qrps : pf gqp´1q “ gJF pf q,
19
and it defines minF pf q as a F -submodule of Rf g {R. Likewise,
f JF pgq Ă JF pf gq, pf p JF pgqrps : pf gqp´1q “ f JF pgq,
and it defines minF pgq as a F -submodule of Rf g {R. Then,
f JF pgq ` gJF pf q Ă JF pf gq.
Since minF pf q X minF pgq “ 0, we have that f JF pgq X gJF pgq “ f gR.
We claim that
pf p JF pgqrps ` g p JF pf qrps : f p´1g p´1q “ f JF pgq ` gJF pf q.
To prove the first containment, take
h P pf p JF pgqrps ` g p JF pf qrps : f p´1 g p´1q.
Then f p´1g p´1h “ f p v ` g p w for some v P pJF pgqqrps and w P JF pgqrps . Since f and g are
relatively prime, f p´1 divides w and g p´1 divides v. Thus, there exist a, b P R such that v “
g p´1a and w “ g p´1b. Then, a P pJF pgqrps : g p´1q “ JF pgq and b P pJF pf qrps : f p´1 q “ JF pf q.
Since,
f p´1 g p´1h “ f p v ` g p w “ f p g p´1a ` g p g p´1b,
h “ f a ` gb P f JF pgq ` gJF pf q.
For the other containment, it is straightforward to check that
f JF pgq ` gJF pf q Ă pf p JF pgqrps ` g p JF pf qrps : f p´1g p´1q.
Since Nf JF pgq`gJF pf q , the F -module generated by f JF pgq ` gJF pf q, contains minF pf q and
minF pgq,
minF pf q ‘ minF pgq Ă Nf JF pgq`gJF pf q .
Therefore, JF phq Ă f JF pgq ` gJF pf q and the result follows.
Proposition 4.15. Let β, γ P N be such that β ă γ. Then,
f γ´β JF pf β q Ă JF pf γ q Ă JF pf β q.
Proof. Let σℓ : R{f ℓ Ñ Rf {R be the injection defined by sending ras Ñ a{f ℓ . We note that
the image of σℓ is R f1ℓ We have that the following commutative diagram,
R{f β R
■■
■■ σβ
■■
■■
■$$
Rf {R
f γ´β
::
✉✉
✉✉
✉
✉✉ σ
✉✉ γ
R{f γ R
Then, R f1β X minF pf q Ă R f1γ X minF pf q, and this corresponds to
f γ´β JF pf β q{f γ R Ă JF pf γ q{f γ R.
Hence, f γ´β JF pf β q Ă JF pf γ q because f γ belongs to both ideals.
20
f γ´β
The morphism R f1γ X minF pf β q Ñ R f1β X minF pf q is well defined and it is equivalent to
the morphism JF pf γ q{f γ R Ñ JF pf β q{f β given by the quotient morphism R{f γ R Ñ R{f β R.
Then, JF pf γ q ` f β R Ă JF pf β q and the result follows.
Remark 4.16. There are examples in which the containment in Proposition 4.15 is strict.
Let R “ Fp rxs and f “ x. In this case, Rf {R is a simple F -module. Then, JF pxβ q “ R for
every β P N and f γ´β JF pf β q Ă JF pf γ q for every γ ą β.
Corollary 4.17. Let f, g P R be such that f divides g. Then, JF pgq Ă JF pf q.
Proof. This follows from Propositions 4.15 and 4.14.
Proposition 4.18. Let f P R and W Ă R be a multiplicative system. Then, JFW ´1 R pf q “
W ´1 JFR pf q.
Proof. By Proposition 4.14, it suffices to prove the claim for f “ g n , where g is an irreducible
element. We note that g is either a unit or a irreducible element in W ´1 R. We have that
minFW ´1 R pf q “ minFW ´1 R pgq is either zero or a simple F -module by Lemma 4.8. Then,
minFW ´1 R pf q “ W ´1 minFR pf q, and so
JFW ´1 R pf q{f W ´1R “ W ´1 R{f W ´1 R X minFW ´1 R pf q
“ W ´1 pR{f R X minFR pf qq
“ W ´1 JFR pf q{f W ´1R,
and the result follows because f belongs to both ideals.
Proposition 4.19. Let f P R. Then, JFR1{pe pf q “ JFR pf qR
e
JFR pf qrp s .
1{pe
pe
. Moreover, JFR pf q “
Proof. By Proposition 4.14, we may assume that f “ g n where g is irreducible. Let q denote
pe and h denote the length of Rf {R in the category of F -modules. Let G : R1{q Ñ R be the
1{q
isomorphism defined by r Ñ r q . Under the isomorphism G, Rf {R1{q corresponds to Rf q {R.
1{q
Then, the length of Rf {R1{q in the category of FR1{q -modules is h. Let 0 “ M0 Ă . . . Ă
Mh “ Rf {R be a chain of FR -submodules of Rf {R such that Mi`1 {Mi is a simple FR -module.
Let f R “ J0 Ă . . . Ă Jh “ R be the corresponding chain of ideals under the bijection given
in Lemma 4.3. Since f “ g n and g is irreducible, M1 “ minFR pf q and J1 “ JFR pf q. We
note that pJip R1{q : f p´1 q “ Ji R1{q and Ji R1{q ‰ Ji`1 R1{q because R1{q is a faithfully flat
R-algebra.
Then, we have a strictly ascending chain of ideals
f R1{q “ J0 R1{q Ă . . . Ă Jh R1{p “ R1{q
1{q
that corresponds to a strictly ascending chain of FR1{q -submodules of Rf {R1{q .
1{q
Since f “ pg 1{q qqn , g 1{q is irreducible and the length of Rf {R1{q is h, we have that
JFR pf qR1{q “ J1 R1{p “ JFR1{q pf q.
After applying the isomorphism G to the previous equality, we have that
JFR pf qrqs “ GpJFR pf qR1{q q “ GpJFR1{q pf qq “ JFR pf q q.
21
Proposition 4.20. Let R Ñ S be a flat morphism of UFDs and let f P R. If S is as in
Notation 4.1, then JFS pf q Ă JFR pf qS.
Proof. We may assume that f “ g β where g is an irreducible element in R by Proposition
4.14. Since S is flat, pJFR pf qrps S : f p´1 q “ JFR pf qS. Let M denote the FS -submodule of
Sf {S given by JFR pf qS under the correspondence in Lemma 4.3. If f is a unit in S, then
JF pf qS “ S and the result is immediate. We may assume that f is not a unit in S. Since
JF pf q ‰ f R, we can pick a P JF pf qzf R. Then, a “ bg γ for some 0 ď γ ă β and b P R such
b
b
that gcdpb, gq “ 1. Then, R{g Ñ R{g is injective, and so S{gS Ñ S{gS is also injective.
Thus, gcdpb, gq “ 1 in S. Hence, b{g is not zero in Sg {S. Moreover, b{g “ g β´γ´1a{f P M
and it is not zero. Let g1 , . . . , gℓ P S irreducible relatively prime elements such that g “
g1β1 ¨ ¨ ¨ g1βℓ . We have that b{gi “ hb{g P Sgi {S X Mzt0u, where h “ g1β1 ¨ ¨ ¨ giβi´1 ¨ ¨ ¨ g1βℓ . Then,
minFS pgi q Ă M and so minFS pf q Ă M. Therefore, JFS pf q Ă JFR pf qS.
Proposition 4.21. Suppose that R is a local ring. Let f P R. Then
p
JFRp pf q “ JFR pf qR,
p denotes the completion of R with respect to the maximal ideal.
where R
p [Bli04, Theorem 4.6]. Then,
Proof. We have that minFRp pf q “ minFR pf q bR R
´
¯
p R
p X minF pf q “ ppR{f Rq X minF pf qq bR R
p “ JF pf qR
p
JF “ R{f
p
R
p
R
R
R
Lemma 4.22. Let R “ Krx1 , . . . , xn s, where K is a perfect field. Let K Ñ L be an algebraic
field extension of K, S “ Lrx1 . . . , xn s, and R Ñ S the map induced by the extension. Then,
JFS pf q “ JFR pf qS.
Proof. We can assume that f “ g β where g is an irreducible element in R by 4.14. It suffices
to show that JFR pRqS Ă JFS pSq by Proposition 4.20.
There is an inclusion φ : Rf {R Ñ Sf {S, which is induced by R Ñ S. We take M “
pminFS pf qq X Rf {R. We claim that M is a DR -module of Rf {R. Since K is perfect,
ď
1 Bt
HomS pe pS, Sq “ DpR, Kq “ Rr
DR “
s.
t
t!
Bx
i
ePN
t
B
We note that DR Ă DS , and that for every m P Rf {R, φp Bx
t mq “
i
Bt
m
Bxti
Bt
φpmq.
Bxti
As a consequence,
P M for every m P M. Therefore, M is a DR -module.
Let I “ M X R{f R. We note that
I “ minFS pf q X R{f R “ pJFS pf q{f Sq X R{f R
and that S{f S is an integral extension of R{f R because L is an algebraic extension of K.
Let r P JFS pf q{f S not zero, and aj P R{f R such that a0 ‰ 0
in S{f S. Then,
r n ` an´1 r n´1 ` . . . ` a1 r ` a0 “ 0
rpan´1 r n´1 ` . . . ` a1 q “ ´a0 ,
22
and so a0 P I “ pJFS pf q{f Sq X R{f R, and then M ‰ 0. Therefore, minFR pf q Ă M and so
JF pf q{f Ă I. Let π : R Ñ R{f R be the quotient morphism. Then,
JFR pf q Ă π ´1 pIq “ JFS pf q X R,
and
JFR pf qS Ă pJFS pf q X RqS Ă JFS pf q.
Lemma 4.23. Let R “ Krx1 , . . . , xn s, where K is an F -finite field. Let L “ K 1{p , S “
Lrx1 . . . , xn s, and R Ñ S the map induced by the extension K Ñ L. Then JFS pf q “ JFR pf qS.
Proof. We have that R Ă S Ă R1{p . Then, by Proposition 4.20,
JFR1{p pf q Ă JFS pf qR1{p Ă pJFR pf qSqR1{p “ JFR pf qR1{p .
Since JFR1{p pf q “ JFR pf qR1{p by Proposition 4.19,
0 “ JFS pf qR1{p {pJFR pf qSqR1{p “ pJFS pf q{JFR pf qSq bS R1{p .
Therefore, JFS pf q “ JFR pf qS because R1{p is a faithfully flat S-algebra.
Lemma 4.24. Let R “ Krx1 , . . . , xn s, where K is an F -finite field. Let L be the perfect
closure of K, S “ Lrx1 . . . , xn s, and R Ñ S the map induced by the extension K Ñ L. Then
JFS pf q “ JFR pf qS.
Proof. We may assume that f “ g n for an irreducible g P R by Proposition 4.14. Let
e
S e “ K 1{p rx1 , . . . , xn s. Let h1 , . . . , hℓ denote a set of generators for JFS pf q. In this case,
pJFS pf qrps : f p´1 q “ JFS pf q. Then there exist ci,j P S such that
ÿ
f p´1hj “
ci,j hpj .
Ť
Since S “ e S e , there exists N such that ci,j , hj P S N . Let I Ă RN be the ideal generated
by h1 , . . . , hℓ . We note that IS “ JFS pf q; moreover, JFS pf q X S N “ I because S e Ñ S splits
for every e P N.
We claim that pI rps : f p´1 q “ I. We have that f p´1hℓ P I rps by our choice of N and so
I Ă pI rps : f p´1q. If g P pI rps : f p´1q, then f p´1 g P I rps Ă JFS pf qrps and g P JFS pf q X S N “ I.
As in the proof of Lemma 4.22, pJFS pf q{f Sq X pS N {f S N q ‰ 0 and then JFS pf q X S N “
I ‰ f S. Therefore, JFS N pf q Ă I by Proposition 4.5. Hence,
JFS N pf qS Ă IS “ JFS pf q Ă JFS N pf qS,
and the result follows because
JFR pf qS “ pJFR pf qS N qS “ JFS N pf qS.
Theorem 4.25. Let R “ Krx1 , . . . , xn s, where K is an F -finite field. Let L be an algebraic
extension of K, S “ Lrx1 . . . , xn s, and R Ñ S the map induced by the extension K Ñ L.
Then JFS pf q “ JFR pf qS.
23
Proof. It suffices to show JFR pf qS Ă JFS pf q by Proposition 4.20. Let K ˚ and L˚ denote the
perfect closure of K and L respectively. Let R˚ “ K ˚ rx1 , . . . , xn s and S ˚ “ L˚ rx1 , . . . , xn s.
Then,
pJFR pf qSqS ˚JFR pf qS ˚ “ pJFR pf qR˚ qS ˚ “ JFR˚ pf qS ˚ “ JFS ˚ pf q “ JFS pf qS ˚
by Lemma 4.22 and 4.24. Therefore,
pJFR pf qS{JFS pf qq bS S ˚ “ pJFR pf qSqS ˚ {pJFS pf qqS ˚ “ 0.
Hence JFR pf qS{JFS pf q “ 0 because S ˚ is a faithfully flat S-algebra.
Example 4.26. Let R “ F3 rx, ys, and f “ x2 ` y 2 and m “ px,?yq. We have that pmrps :
f p´1 q “ m. Then, JFR pf q Ă m. Let F3 ris the extension of F3 by ´1, S “ Lrx, ys and φ :
R Ñ S be the inclusion given by the field extension. Then, JFS pf q “ px, yqS by Proposition
4.14 because x2 ` y 2 “ px ` iyqpx ´ iyq. Since φ is a flat extension, JFS pf q Ă JFR pf qS. Then,
m “ R X JFS pf q Ă R X JFR pf qS. Hence, JF pf q “ m.
Proposition 4.27. Let f P R be an irreducible element. Then,
˜ ˆ
˙¸
č
ď ´` e ˘r1{pe s ¯rpe s
e
f p ´1 a
,f
JF pf q “
: f p ´1
gcdpa,f q“1
ePN
Proof. We have that minF pf q is the intersection of all nonzero D-submodules of Rf {R by
Proposition 4.8. In particular, minF pf q is the intersection of all nonzero cyclic D-modules
generated by elements a{f P Rf {R. Hence,
č
JF pf q{f “
ppD ¨ a{f q X R{f q
gcdpa,f q“1
“
č
gcdpa,f q“1
˜
ď`
˘
D peq ¨ a{f X R{f
ePN
¸
˘
Ş
Ť `
We have that b P JF pf q if b{f P gcdpa,f q“1 ePN D peq ¨ a{f , so, for every a P R such
that gcdpa, f q “ 1, there exists an e P N such that b{f P D peq ¨ a{f. Thus, there exists
φ P HomRpe pR, Rq such that
e
φpa{f q “ 1{f p φpf p
pe
e ´1
aq “ b{f ` r
Therefore, after multiplying by f , we have that
˜ ˆ
˙¸
č
ď ´` e ˘r1{pe s ¯rpe s
e
f p ´1 a
,f
bP
: f p ´1
gcdpa,f q“1
ePN
` e ˘
` e
e ˘rpe s
because pf p ´1 aqr1{p s
“ D peq f p ´1 a .
On the other hand, if
˙
č ď ˆ´` e ˘r1{pe s ¯rpe s
p ´1
pe ´1
f
a
,f
,
bP
:f
gcdpa,f q“1 ePN
24
then for every a P R such that gcdpa, f q “ 1, there exists an e P N and φ P HomRpe pR, Rq
such that
e
e
e
f p ´1 b “ φpf p ´1 aq ` f p r
´`
˘r1{pe s ¯rpe s
e
e
e
because f p ´1 aq, f
“ D peq pf p ´1 aq. Therefore, after dividing by f p , we have that
˘
Ş
Ť `
b{f P gcdpa,f q“1 ePN D peq ¨ a{f , and then b P JF pf q.
Theorem 4.28. Let f P R be such that R{f R is a F -pure ring. If JF pf q “ R, then R{f R
is strongly F -regular.
Proof. We may assume that pR, m, Kq is local because being reduced, F -pure, and F regular are local properties for R{f R. Then, f is irreducible by Proposition 4.14. Since
J
a such˙that gcdpa, f q “ 1 there exits an e P N such that R “
ˆF´pf q “ R, for every
´`
˘r1{pe s ¯rpe s
` pe ´1 ˘r1{pe s ¯rpe s pe ´1
e
e
f p ´1 a
,f
f
a
,f
.
by Lemma 4.27. Then, f p ´1 P
:f
` e ˘r1{pe s
e
e
Since f p ´1 R mrp s for every e P N by Fedder’s Criterion , R “ f p ´1 a
; otherwise,
` pe´1 ˘r1{pe s
pe
pe
p f
a
, f q Ă m. Then, there exist a morphism φ : R Ñ R of R -modules such that
e
pe ´1
aq “ 1. Let ϕ : R{f R Ñ R{f R be the morphism defined by ϕprxsq “ rφpf p ´1 xqs.
φpf
We note that ϕ is a well defined morphism of pR{f Rqp -modules such that ϕprasq “ 1. Then,
R{f R is a simple DpR{f Rq-module. Hence, R{f R is strongly F -regular [Smi95a, Theorem
2.2].
Remark 4.29.
‚ The result of the previous theorem is a consequence of a result of
Blickle [Bli04, Corollary 4.10], as R{f R is a Gorenstein ring. However, our proof is
different from the one given there.
‚ JFR pf q “ R does not imply that R{f R is F -pure. Let K “ FracpF2 rusq be the fraction
field of the polynomial ring F2 rus, R “ Krrx, yss, and f “ x2 ` uy 2. Then, f is an
irreducible element such that R{f R is not pure because f P px, yqr2s R. Let L “ K 1{2 ,
S “ Lrrx, yss and R Ñ S be the inclusion given by the extension K Ă L. Thus,
f “ px ´ u1{2 yq2 in S, and then JFS pSq “ S Ă JFR pf qS. Then, R “ JFS pSq X R “
JFR pf qS X R “ JFR pf q because R Ñ S splits. Hence, JFR pf q “ R and R{f R is not
F -pure.
4.2. Definition for rings essentially of finite type over an F –finite local ring.
Notation 4.30. Throughout this section R denotes a ring essentially of finite type over
an F –finite local ring. Let f P R, π : R Ñ R{f R be the quotient morphism. If R{f R is
reduced, τf denotes the pullback of the test ideal of R{f R, π ´1 pτ pR{f Rqq.
Under the hypotheses on R in Notation 4.30, there is an F -module and a D-module of
Rf {R called the intersection homology LpR, R{f Rq [Bli01, Bli04]. We have that for every
maximal ideal m Ă R, pRzmq´1 LpR, R{f Rq “ minFRm pf q.
Definition 4.31. Let R{f R Ă Rf {R be the inclusion morphism 1 ÞÑ
F -Jacobian, Jf pf q as the pullback to R of pR{f Rq X LpR, R{f Rq.
rpj´1 s
Lemma 4.32. Suppose that R{f R is reduced. Let I j pf q “ pτf
I j`1 pf q and I j`1 pf q “ pI j pf qrps : f p´1 q.
25
: fp
1
.
f
j´1 ´1
We define the
q. Then I j pf q Ă
Proof. Since, in this case, the test ideal of R{f R commutes with localization, we may assume
that R is a local ring. We have that τf {f R is the minimal root for minF pf q [Bli04, Theorem
e
rpe s
rps
4.6]. Then, f p´1I 1 pf q “ f p´1 τf Ă τf “ I 1 pf qrp s . Thus, I 1 Ă I 2 and f p´1 I 2 Ă I1 .
Moreover, I 2 {f R is also a root for minF pf q because I 2 pf q{I 1 pf q is the kernel of the map
f p´1
R{I 1 pf q Ñ R{I 1 pf qrps .
Inductively, we obtain that I j Ă I j`1, f r I j`1 Ă I j and that I j {f R is a root for minF pf q for
every j P N and the result follows.
Ť j
Proposition 4.33. Suppose that R{f R is reduced. Then, JF pf q “ j IR pf q.
f p´1
Proof. We have that τf f1 is the minimal root for LpR, R{f Rq. Moreover, any ideal I j pf q Ñ
I j pf qrps also generates LpR, R{f Rq as an F -module. Moreover,
ď j
č
IR pf q “ LpR, R{f Rq R{f R “ JF pf q.
j
Remark 4.34. In general, we do not have τf “ JF pf q. Let R “ Krxs, where K is any
perfect field of characteristic p ą 0. Let f “ x2 . Then, τf “ xR ‰ R “ JF pf q. In addition,
Example 4.45, shows another situation where τf ‰ JF pf q.
Remark 4.35. If R is an F -finite local ring, then
f p´1
JF pf q Ñ JF pf qrps
is a generating morphism for minF pf q because in this case minF pf q “ LpR, R{f Rq.
Corollary 4.36. Let S be a ring that is as in Notation 4.1 and as in Notation 4.30. Let
f P S. Let J denote the F -Jacobian ideal of f as in Definition 4.12 and let J 1 the F -Jacobian
ideal of f as in Definition 4.31. Then, J “ J 1 .
Proof. We have that in both contexts the F -Jacobian ideal commutes with localization. We
rps
f p´1
may assume that R is a regular local F -finite ring. As J2 “ pJ2 : f p´1 q and J2 {f R Ñ
rps
J2 {f p R is a root for minF pf q by Lemma 4.35, we have that J1 “ J2 .
Remark 4.37. As for every maximal ideal m Ă R, Rm is as in Notation 4.1, we have that
‚ f n JF pf q Ă JF pf n q,
e
e
‚ JF pf p q “ JF pf qrp s , and
‚ if gcdpf, gq “ 1, JF pf q “ f JF pgq ` gJF pf q.
because those properties localize.
Proposition 4.38. Suppose that pR, m, Kq is local. Let pS, η, Lq denote a regular F -finite
ring. Let R Ñ S be a flat local morphism such that the closed fiber S{mS is regular and
L{K is separable. Then, JFS pf q “ JFR pf qS.
Proof. It suffices to proof that minFR pf qS “ minFS pf q. We can assume without loss of
p and JF pf q “ JF pf qSp
generalization tat R{f R is reduced. We have that JFRp pf q “ JFR pf qR
S
p
S
p Ñ Sp is still a
by Proposition 4.21. In addition, the induced morphism in the completion R
26
flat local morphism. Since JFS pf q Ă JFR pf qS and JFSp pf q Ă JFRp pf qSp by Proposition 4.20,
p F pf q “ pJF pf qS{JF pf qq bS S.
p Therefore, we can assume that R and S are
JFRp pf qS{J
R
S
p
S
complete.
We note that R{f R Ñ S{f S is again a flat local morphism such that the closed fiber S{mS
is regular L{K is separable by flat base change. Then S{f S is reduced and τ pR{f RqS “
τ pS{f Sq [HH94a, Theorem 7.2], and so IFj S pf q “ IFj R pf qS. Hence, JFS pf q “ JFR pf qS by
Proposition 4.33.
Corollary 4.39. Suppose that R is a Zh -graded ring. Let f P R be a homogeneous element.
Then, JF pf q is a homogeneous ideal.
Proof. It suffices to proof that minpf q is a Zh -graded submodule of Rf {R. We can assume
that R{f R is reduced. We have that τ pR{f Rq is a homogeneous ideal [HH94b, Theorem 4.2].
This means that IRj pf q is a homogeneous ideal for every j. Therefore, JF pf q is homogeneous
and that minF pf q Zn -graded submodule of Rf {R.
Corollary 4.40. Let S be a ring that is as in Notation 4.1 or as in Notation 4.30. Let
f P S be such that R{f R is reduced. Then, VpJF pf qq Ă SingF pS{f Sq. Moreover, if S{f S is
an F -pure ring, then VpJF pf qq “ SingF pS{f Sq.
Proof. For every prime ideal P P VpJF pf qq, JFSP pf q ‰ SP . Since SP is as in Notation
4.30, we have that τ pSP {f SP q Ă JFSP pf q Ă P SP . Then, SP is not F -regular and then
P P SingF pS{f Sq.
Now, we suppose that S{f S is F -pure. For every prime ideal P P SingF pS{f Sq, SP {f SP
is not F -regular. Then, JFRP pf q ‰ RP by Theorem 4.28. Then, P P VpJF pf qq.
Lemma 4.41. Let S be a ring that is as in Notation 4.1 and as in Notation 4.30. Let f P S
be an element and Q Ă S be a prime ideal. If SQ {f SQ is F -pure, then SQ {JFSQ pf q is F -pure.
Proof. We may replace S by SQ . Since S{f S is F -pure, we have that f p´1 R Qrps by Fedder’s
Criterion. We have that f p´1 P pJF pf qrps : JF pf qq, and so pJF pf qrps : JF pf qq Ć Qrps .
Therefore, S{JF pf q is F -pure.
Corollary 4.42. Let f P R. If R{f R is an F -pure ring, then JF pf q “ τf .
a
a
Proof. We have that JF pf q “ π ´1 pR{f Rq by Corollary 4.40 because SingF pR{f Rq “
Vpτ pR{f Rqq in this case. Since R{JF pf q is F -pure by Lemma 4.41, JF pf q is a radical ideal.
In addition, τ pR{f Rq is a radical ideal [FW89, Proposition 2.5]. Hence, JF pf q “ τf .
4.3. Examples.
Proposition 4.43. Let f P R be an element with an isolated singularity at the maximal
ideal m. If Rm {f Rm is F -pure, then
#
R if R{f R is F ´ regular
JF pf q “
m otherwise
Proof. Since R{f R has an isolated singularity
a at m, we have that JF pf qRP “ RP for every
prime ideal different from m. Then, m Ă JF pf q.
If Rm {f Rm is F -regular, then R{f R is F -regular, and so JF pRq “ R by Theorem 4.28.
27
a
If Rm {f Rm is not F -regular, then JF pRq ‰ R by Theorem 4.28. Then, m “ JF pf q.
Since Rm {f Rm is F -pure, we have that Rm {JF pf qRm is F -pure by Lemma 4.41. Then,
Rm {JF pf qRm is a reduced ring. Hence, JF pf q “ m.
Example 4.44. Let K is an F -finite field. Let E be an elliptic curve over K. We choose a
closed immersion of E in P2K and set R “ Krx, y, zs, the completed homogeneous co-ordinate
ring of P2K . We take f P R as the cubic form defining E. We know that f has an isolated
singularity at m “ px, y, zqR. If the elliptic curve is ordinary, then R{f R is F -pure [Har77,
Proposition 4.21] [Bha12, Theorem 2.1] [BS]. We know that R{f R is never an F -regular
ring [HH94b, Discussion 7.3bpbq, Theorem 7.12]. Then, JF pf q “ m by Proposition 4.43.
Example 4.45. Let R “ Krx, y, zs, where is an F -finite field of characteristic p ą 3. Let
f “ x3 ` y 3 ` z 3 P R, and π : R Ñ R{f R be the quotient morphism and m “ px, y, zqR. We
have that τf “ m [Smi95b, Example 6.3]. Then, m Ă JF pf q by Proposition 4.33.
We have that R{f R is F -pure if and only if p ” 1 mod 3. We have that pmrps : f p´1 q “ m
if p ” 1 mod 3, and pmrps : f p´1 q “ R if p ” 1 mod 2. Hence,
#
R p ” 2 mod 3
JF pf q “
m p ” 1 mod 3.
Example 4.46. Let R “ Krx1 , . . . , xn s, where K is an F –finite field of characteristic p ą 0.
Let f “ a1 xd11 ` . . .` an xdnn , be such that a‰ 0. We have that R{f R has an isolated singularity
at the maximal ideal m “ px1 , . . . , xn q.
If d11 ` . . . ` d1n “ 1 and pp ´ 1q{d1 is an integer for every i, then R{f R is F -pure for p " 0
[Her12, Theorem 3.1] and not F -regular [Gla96, Theorem 3.1] because f p´1 is congruent to
e
e
cpx1 ¨ ¨ ¨ xn qp ´1 module mrp s for a nonzero element c P K. Hence, JF pf q “ R for p " 0 by
Proposition 4.43.
Remark 4.47. Let R “ Krx1 , . . . , xn s be a polynomial ring and f P R be such that R{f R
is reduced. We can obtain JF pf q from τ pR{f Rq by Proposition 4.33. In the case where
n ą 3, f “ xd1 ` . . . xdn and d is not divided by the characteristic of K, there is an algorithm
to compute the test ideal of R{f R [McD03]. Therefore, there is an algorithm to compute
JF pf q.
Example 4.48. Let R “ Krx1 , . . . , xn s, where K is a field of characteristic p ą 0. Let
f “ xd1 ` . . . ` xdn . This examples are based in computations done by McDermott [McD03,
Example 11, 12 and 13].
If p “ 2, n “ 5 and d “ 5,
τf “ px2i xj q1ďi,jď5.
r2s
r4s
Then, px21 , x22 , x23 , x24 , x25 , x1 x2 x3 x4 x5 qR “ pτf : f q and R “ pτf : f 3 q. Hence, JF pf q “ R.
If p “ 3, n “ 4 and d “ 7,
τf “ px2i x2j q1ďi,jď4.
r3s
Then R “ pτf : f 2 q and JF pf q “ R.
If p “ 7, n “ 5 and d “ 4,
τf “ px1 , . . . , x5 qR.
r7s
Then R “ pτf : f 6 q and JF pf q “ R.
28
5. Further consequences and relations
Using ideas and techniques developed in the previous sections, we obtain relations among
test ideals, generalized test ideals and generalized Lyubeznik numbers for hypersurfaces.
Notation 5.1. Throughout this section R denotes a ring essentially of finite type over an F –
finite local ring. Let f P R denote an element such that R{f R is reduced. Let π : R Ñ R{f R
be the quotient morphism. τf denotes π ´1 pτ pR{f Rqq, the pullback of the test ideal of R{f R.
Remark 5.2. We have that τ pf 1´ǫ q f1 is the minimal root for Rf as an F -module. Then
f q´1
τ pf 1´ǫ q{f R Ñ τ pf 1´ǫ qrqs {f q R
generates Rf {R (see Proposition 3.17 or [Bli08, Proposition 3.5]. We have that τ pR{f Rq “
τf {f is the minimal root of minF pf q [Bli04, Proposition 4.5 and Theorem 4.6]. Therefore,
τf f1 Ă τ pf 1´ǫ q f1 Ă R1 {R. Then,
τf Ă τ pf 1´ǫ q.
This containment can also be obtained using F -adjunction [Tak08, Sch09].
Suppose that R is a finitely generated polynomial ring over an F -finite field and f is a
polynomial. Since the Jacobian ideal of f is contained in τf when R{f R is reduced [HH99],
the previous containment of test ideals gives implies Jpf q Ă τ pf 1´ǫ q. The containment of
the Jacobian ideal in the test ideal τ pf 1´ǫ q was proved using different techniques [KLZ11]
and still holds when R{f R is not reduced.
We assume that pR, m, Kq is a local ring and F -finite. Vassilev proved that if I Ă R is an
rps
ideal and τI is the pullback of the test ideal of R{I, then pI rps : Iq Ă pτI : τI q [Vas98]. In
particular, if R{I is F -pure, then R{τI is also F -pure. Using this criteria, she showed that
if R{I is F -pure, there is a sequence of ideals
(5.2.1)
I “ τ0 Ă τ1 Ă . . .
such that τi`1 is the pull back of the test ideal of R{τi . The sequence in 5.2.1 stabilizes when
τi “ R, because if τi ‰ 0, R{τi , then is an F -finite reduced local ring and its test ideal is not
zero [HH94a, Theorem 6.1]. We apply this ideas to bound the length of M1 {N1 “ Rf {R as
an F -module.
Lemma 5.3. Suppose that pR, m, Kq is a local ring and F -finite and R{f R is F -pure. Let
0 Ă f “ τ0 Ă τ1 Ă . . . Ă τℓ “ R
be the sequence of pullback of test ideals for f R defined in 5.2.1. If τi`1 ‰ τi , then Nτi`1 ‰ Nτi .
rps
Proof. We have that f p´1 P pf p : f q Ă pτj : τj q for every j. We will proceed by contradiction.
e
rpe s
Suppose that τi`1 ‰ τi , and Nτi`1 “ Nτi . Then there exists e such that f p ´1 τi`1 Ă τi by
Lemma 3.12. Since R{τj is F -pure for every j, we have that both τi and τi`1 are radical
ideals. Therefore, we can choose a minimal prime Q of τi such that pτi`1 qQ “ RQ . Hence
e
e
e
rpe s
f p ´1 pτi`1 qQ “ f p ´1 RQ Ă pτi qQ Ă pQRQ qrp s , and we obtain a contradiction because
RQ {f RQ is F -pure
29
Proposition 5.4. Suppose that pR, m, Kq is local and R{f R is F -pure. If
0 Ă f R “ τ0 Ă τ1 Ă . . . Ă τℓ “ R
is the sequence of ideals for f R defined in 5.2.1, then
ℓ ď lengthF ´mod Rf {R
Proof. We have that every pullback of test ideal τi defines a different F -module by Lemma
5.3, which proves the proposition.
The Lyubeznik numbers are invariants of a local ring of equal characteristic [Lyu93, Theorem/Definition 4.1]. The first author and Witt defined a generalization these invariants using
the theory of D-modules and local cohomology [NBW12, Definition 4.3]. We only recall the
definition of hypersurfaces; however, the generalized Lyubeznik numbers are defined for any
local ring of equal characteristic and a sequence of integers and ideals.
Suppose that R “ Krrx1 , . . . , xn ss Let d “ dimpR{f Rq “ n ´ 1 and L be a coefficient
field of R{f R. We define the generalized Lyubeznik λd0 pR{f R; Lq by lengthDpR,Lq Rf {R. This
number is well defined and depends only on R{f R and L.
Corollary 5.5. Suppose that R “ Krrx1 , . . . , xn ss, where K is an F -finite field, and that
R{f R is F -pure. Let L be a coefficient field of R{f R. If
0 Ă f R “ τ0 Ă τ1 Ă . . . Ă τℓ “ R
is the sequence of ideals in 5.4, then
dimpR{f Rq
Proof. We have that
ℓ ď λ0
pR{f R; Lq.
dimpR{f Rq
ℓ ď lengthF ´mod Rf {R ď lengthDpR,Zq Rf {R ď lengthDpR,Lq Rf {R “ λ0
pR{f R; Lq.
dimpR{f Rq
λ0
pR{f R; Lq
Remark 5.6. If R{f R is F -pure and
“ 1, then R{f R is F -regular
dimpR{f Rq
[Bli04, NBW12]. In addition, if R{f R is F -pure and K perfect, then λ0
pR{f R; Lq “ 1
if and only if R{f R is F -regular [Bli04, NBW12]. Then, Corollary 5.5 is telling us that for
dimpR{f Rq
F -pure hypersurfaces λ0
pR{f R; Lq measures how far is R{f R from being F -regular.
Proposition 5.7. Suppose that pR, m, Kq is local and F -finite, and that R{f R is reduced
and lengthR τ pf 1´ǫ q{τf is finite. Let f “ f1 . . . fℓ be a factorization of f into irreducible
elements. Then,
lengthF ´mod Rf {R ď lengthR pτ pf 1´ǫ q{τf q ` ℓ
Proof. We have that minF pf q “ minF pf q ‘ . . .‘ minF pfℓ q. In addition, the length of minF pf q
as a F -module is ℓ. We recall that τ pf 1´ǫ q f1 is a root for Rf {R by Proposition 3.18 and τf f1
is the minimal root for minF pf q. Therefore, for every F -submodule N Ă Rf {R that contains
minF pf q, NXτ pf 1´ǫ q f1 is a root for N and it must contain τf f1 . Then for every strict ascending
chain of F -modules minF pf q “ N0 Ă N1 Ă . . . Ă Nt “ Rf {R, t ď lengthR τ pf 1´ǫ q{τf . Hence
lengthF ´mod Rf {R ď lengthR pτ pf 1´ǫ q{τf q ` ℓ.
Proof of Theorem 1.2. This is consequence of Propositions 5.4 and 5.7 and Corollary 5.5.
30
Acknowledgments
We thank Josep Àlvarez Montaner, Angélica Benito, Daniel J. Hernández, Mel Hochster,
Mircea Mustaţă, Emily E. Witt and Wenliang Zhang for insightful mathematical conversations related to this work. We thank Manuel Blickle, Karl schwede and Kevin Tucker for
poitnting out missed references. We also thank Alexis Cook for her careful reading of this
manuscript. The first author thanks the National Council of Science and Technology of
Mexico (CONACyT) for its support through grant 210916.
References
[ÀMBL05] Josep Àlvarez Montaner, Manuel Blickle, and Gennady Lyubeznik. Generators of D-modules in
positive characteristic. Math. Res. Lett., 12(4):459–473, 2005.
[BFS12]
Angélica Benito, Eleonore Faber, and Karen Smith. Measuring singularities with Frobenius: the
basics. Springer, 2012.
[Bha12]
Bargav Bhatt. F –pure threshold of an elliptic curve. Preprint, 2012.
[Bli01]
Manuel Blicle. The intersection of homology D-module in finite characteristic. PhD thesis, University of Michigan, 2001.
[Bli04]
Manuel Blickle. The intersection homology D-module in finite characteristic. Math. Ann.,
328(3):425–450, 2004.
[Bli08]
Manuel Blickle. Minimal γ-sheaves. Algebra Number Theory, 2(3):347–368, 2008.
[BM10]
Holger Brenner and Paul Monsky. Tight closure does not commute with localization. Ann. of
Math. (2), 171(1):571–588, 2010.
[BMS08] Manuel Blickle, Mircea Mustaţǎ, and Karen E. Smith. Discreteness and rationality of F thresholds. Michigan Math. J., 57:43–61, 2008. Special volume in honor of Melvin Hochster.
[BMS09] Manuel Blickle, Mircea Mustaţă, and Karen E. Smith. F -thresholds of hypersurfaces. Trans.
Amer. Math. Soc., 361(12):6549–6565, 2009.
[BS]
Bhargav Bhatt and Anurag K. Singh. F -thresholds of Calabi-Yau hypersurfaces. preprint.
[FA12]
Eleonore Faber and Paolo Aluffi. Splayed divisors and their chern classes. Preprint, 2012.
[Fab13]
Eleonore Faber. Towards transversality of singular varieties: splayed divisors. Publ. RIMS, To
appear, 2013.
[Fed87]
Richard Fedder. F -purity and rational singularity in graded complete intersection rings. Trans.
Amer. Math. Soc., 301(1):47–62, 1987.
[FW89]
Richard Fedder and Kei-ichi Watanabe. A characterization of F -regularity in terms of F -purity.
In Commutative algebra (Berkeley, CA, 1987), volume 15 of Math. Sci. Res. Inst. Publ., pages
227–245. Springer, New York, 1989.
[Gla96]
Donna Glassbrenner. Strong F -regularity in images of regular rings. Proc. Amer. Math. Soc.,
124(2):345–353, 1996.
[Har77]
Robin Hartshorne. Algebraic geometry. Springer-Verlag, New York, 1977. Graduate Texts in
Mathematics, No. 52.
[Her12]
Daniel J. Hernández. F -invariants of diagonal hypersurfaces. Preprint, 2012.
[HH90]
Melvin Hochster and Craig Huneke. Tight closure, invariant theory, and the Briançon-Skoda
theorem. J. Amer. Math. Soc., 3(1):31–116, 1990.
[HH94a] Melvin Hochster and Craig Huneke. F -regularity, test elements, and smooth base change. Trans.
Amer. Math. Soc., 346(1):1–62, 1994.
[HH94b] Melvin Hochster and Craig Huneke. Tight closure of parameter ideals and splitting in modulefinite extensions. J. Algebraic Geom., 3(4):599–670, 1994.
[HH99]
Melvin Hochster and Craig Huneke. Tight closure in equal characteristic zero. 1999.
[Hoc07]
Melvin Hochster. Some finiteness properties of Lyubeznik’s F -modules. In Algebra, geometry and
their interactions, volume 448 of Contemp. Math., pages 119–127. Amer. Math. Soc., Providence,
RI, 2007.
[HY03]
Nobuo Hara and Ken-Ichi Yoshida. A generalization of tight closure and multiplier ideals. Trans.
Amer. Math. Soc., 355(8):3143–3174 (electronic), 2003.
31
[KLZ11]
[LS01]
[Lyu93]
[Lyu97]
[Mat89]
[McD03]
[NBW12]
[PS73]
[Sch09]
[Smi94]
[Smi95a]
[Smi95b]
[Tak08]
[Vas98]
[Yek92]
Mordechai Katzman, Gennady Lyubeznik, and Wenliang Zhang. An upper bound on the number
of F -jumping coefficients of a principal ideal. Proc. Amer. Math. Soc., 139(12):4193–4197, 2011.
Gennady Lyubeznik and Karen E. Smith. On the commutation of the test ideal with localization
and completion. Trans. Amer. Math. Soc., 353(8):3149–3180 (electronic), 2001.
Gennady Lyubeznik. Finiteness properties of local cohomology modules (an application of Dmodules to commutative algebra). Invent. Math., 113(1):41–55, 1993.
Gennady Lyubeznik. F -modules: applications to local cohomology and D-modules in characteristic p ą 0. J. Reine Angew. Math., 491:65–130, 1997.
Hideyuki Matsumura. Commutative ring theory, volume 8 of Cambridge Studies in Advanced
Mathematics. Cambridge University Press, Cambridge, second edition, 1989. Translated from
the Japanese by M. Reid.
Moira A. McDermott. Test ideals in diagonal hypersurface rings. II. J. Algebra, 264(1):296–304,
2003.
Luis Núñez-Betancourt and Emily E. Witt. Generalized Lyubeznik numbers. Preprint, 2012.
C. Peskine and L. Szpiro. Dimension projective finie et cohomologie locale. Applications à la
démonstration de conjectures de M. Auslander, H. Bass et A. Grothendieck. Inst. Hautes Études
Sci. Publ. Math., (42):47–119, 1973.
Karl Schwede. F -adjunction. Algebra Number Theory, 3(8):907–950, 2009.
Karen E. Smith. Tight closure of parameter ideals. Invent. Math., 115(1):41–60, 1994.
Karen E. Smith. The D-module structure of F -split rings. Math. Res. Lett., 2(4):377–386, 1995.
Karen E. Smith. Test ideals in local rings. Trans. Amer. Math. Soc., 347(9):3453–3472, 1995.
Shunsuke Takagi. A characteristic p analogue of plt singularities and adjoint ideals. Math. Z.,
259(2):321–341, 2008.
Janet Cowden Vassilev. Test ideals in quotients of F -finite regular local rings. Trans. Amer.
Math. Soc., 350(10):4041–4051, 1998.
Amnon Yekutieli. An explicit construction of the Grothendieck residue complex. Astérisque,
(208):127, 1992. With an appendix by Pramathanath Sastry.
Department of Mathematics, University of Michigan, Ann Arbor, MI 48109-1043,
USA.
Email address: [email protected]
Department of Mathematics, University of Michigan, Ann Arbor, MI 48109-1043,
USA.
Email address: [email protected]
32
| 0math.AC
|
1
Distributed Information-Theoretic Clustering
Georg Pichler, Student Member, IEEE, Pablo Piantanida,
Senior Member, IEEE, and Gerald Matz, Senior Member, IEEE
arXiv:1602.04605v4 [cs.IT] 17 Nov 2017
Abstract
We study a novel multi-terminal source coding setup motivated by the biclustering problem. Two separate
n
encoders observe two i.i.d. sources Xn and Zn , respectively. The
goal is to find rate-limited encodings f (x )
n
n
n
and g(z ) that maximize the mutual information I f (X ); g(Z ) /n. We provide inner and outer bounds on the
achievable region and discuss connections of this problem with hypothesis testing against independence, pattern
recognition, the information bottleneck method, and lossy source coding with logarithmic-loss distortion. Improving
previous cardinality bounds allows us to thoroughly study the special case of a binary symmetric source and to
quantify the gap between the inner and the outer bound in this special case. Furthermore, we generalize our results to
the case of more than two i.i.d. sources. As a special case of this generalization we investigate a multiple description
(MD) extension of the CEO problem with log-loss distortion. Surprisingly, this MD-CEO problem permits a tight
single-letter characterization of the achievable region.
Index Terms
Source coding; Mutual information; Information bottleneck; CEO problem; Logarithmic loss.
I. I NTRODUCTION
The recent decades witnessed a rapid proliferation of digital data in a myriad of repositories such as internet fora,
blogs, web applications, news, emails and the social media bandwagon. A significant part of this data is unstructured
and it is thus hard to extract relevant information. This results in a growing need for a fundamental understanding
and efficient methods for analyzing data and discovering valuable and relevant knowledge from it in the form of
structured information.
When specifying certain hidden (unobserved) features of interest, the problem then consists of extracting those
relevant features from a measurement, while neglecting other, irrelevant features. Formulating this idea in terms of
lossy source compression [4], we can quantify the complexity of the encoded data via its rate and the quality via
the information provided about specific (unobserved) features.
In this paper, we introduce and study the distributed clustering problem from a formal information-theoretic
perspective. Given correlated samples Xn1 , Xn2 , . . . , XnK observed at different encoders, the aim is to extract a
description from each sample such that the descriptions are maximally informative about each other. In other words,
the k th encoder tries to find a (lossy) description Wk = fk (Xnk ) of its observation Xnk subject to a complexity
requirement (coding rate) such that the mutual information between two disjoint subsets of descriptions (Wk )k∈A
and (Wk )k∈B is maximized. Our goal is to characterize the optimal tradeoff between relevance (mutual information
between the descriptions) and complexity (encoding rate).
A. Distributed Clustering
Biclustering (or co-clustering) was first explicitly considered by Hartigan [5] in 1972. A historical overview of
biclustering including additional background can be found in [6, Section 3.2.4]. In general, given an N × M data
matrix (anm ), the goal of a biclustering algorithm [7] is to find partitions Bk ⊆ {1, . . . , N } and Cl ⊆ {1, . . . , M },
k = 1 . . . K , l = 1 . . . L such that all the elements of the “biclusters” (anm )n∈Bk ,m∈Cl are homogeneous in a certain
sense. The measure of homogeneity of the biclusters depends on the specific application. The method received
Parts of the material in this paper were previously presented in [1]–[3].
G. Pichler and G. Matz are with the Institute of Telecommunications, TU Wien, Vienna, Austria. P. Piantanida is with Laboratoire des
Signaux et Systèmes (L2S, UMR8506), CentraleSupélec-CNRS-Université Paris-Sud, Gif-sur-Yvette, France.
Part of this work was supported by the FP7 Network of Excellence in Wireless COMmunications NEWCOM# and by the WWTF under
grants ICT12-054 (TINCOIN) and ICT15-119 (Co3-iGraB).
2
renewed attention when Cheng and Church [8] applied it to gene expression data. Many biclustering algorithms
have been developed since (e. g., see [9] and references therein). An introductory overview of clustering algorithms
for gene expression data can be found in the lecture notes [10]. The information bottleneck (IB) method, which
can be viewed as a uni-directional information-theoretic variant of biclustering, was successfully applied to gene
expression data as well [11].
In 2003, Dhillon et. al. [12] adopted an information-theoretic approach to biclustering. They used mutual information
to characterize the quality of a biclustering. Specifically, for the special case when the underlying matrix represents
the joint probability distribution of two discrete random variables X and Y, i. e., anm = P{X = n, Y = m}, their
goal was to find
clustering functions f : {1, . . . , N } → {1, . . . , K} and g : {1, . . . , M } → {1, . . . , L} that maximize
I f (X); g(Y) for specific K and L. In the present work, we investigate an extension of the approach of Dhillon
et. al. [12], where we consider stationary, memoryless processes and N , K , M , and L scale exponentially in the
blocklength. The resulting problem is addressed using information-theoretic tools.
The aim of the present paper is to characterize the achievable region of this information-theoretic biclustering
problem and connect it to known problems in network information theory. Furthermore, we provide an extension
to more than two stationary, memoryless sources. This distributed information-theoretic clustering problem offers
a formidable mathematical complexity. It is fundamentally different from “classical” distributed source coding
problems like distributed lossy compression [13, Chapter 12]. Usually, one aims at reducing redundant information,
i. e., information that is transmitted by multiple encoders, as much as possible, while still guaranteeing correct
decoding. By contrast, in the clustering problem we are interested in maximizing this very redundancy. In this
sense, the clustering problem is complementary to conventional distributed source coding. We also point out that
the extension of biclustering to multiple variables contains the Körner-Marton problem [14] as a special case, which
implies that in general Berger-Tung coding is suboptimal.
B. Contributions
We first study the case of two sources in Section II. This problem turns out to be closely related to a number of
information-theoretic problems, including hypothesis testing against independence [15], pattern recognition [16],
the information bottleneck method [17] and lossy source coding with logarithmic-loss distortion [18]. We study
these connections in detail in Section II-B (Theorem 5, Corollary 6, and Proposition 8) and exploit them to provide
an inner bound on the achievable region (Theorem 11). The outer bound (Theorem 9) follows from standard
information-theoretic manipulations. In Section II-E we extensively study the doubly symmetric binary source as
a special case. In order to perform this analysis, we require stronger cardinality bounds (Propositions 10 and 13)
than the ones usually obtained using the convex cover method [13, Appendix C]. We are able to improve upon
the state-of-the-art cardinality bounding techniques by combining the convex cover method with the perturbation
method [13, Appendix C] and leveraging ideas similar to [19], which allow us to restrict our attention to the
extreme points of the achievable region. The resulting bounds, proved in Appendix A2, allow for the us binary
auxiliary random variables in the case of binary sources. We subsequently show that [16, Conjecture 1] does not
hold (Propositions 20 and 21). Based on a weaker conjecture (Conjecture 19), we then argue that there is indeed a
gap between the outer and the inner bound for a doubly symmetric binary source (Conjecture 18). In Section III
we extend the inner and the outer bound to the case of multiple sources (Theorems 24 and 25) by incorporating a
binning strategy into the achievability part of the proof. In Section III-B we investigate the chief executive officer
(CEO) problem under a mutual information constraint, a special case of the information-theoretic clustering problem
with multiple sources. We show that it is equivalent to classical multiterminal lossy source coding under logarithmic
loss distortion (Lemma 32). By leveraging this equivalence, we obtain tight bounds for a special case (Lemma 33)
using results from [18]. In this problem, the CEO relies on the full set of descriptions obtained from all agents. If
an agent fails to deliver a description, the CEO’s estimate may be severely degraded. We thus propose an extension
of the CEO problem that takes into account the possibility that descriptions are not delivered by using multiple
description (MD) coding [20] (see [21] for applications). In Section III-D we study this MD-CEO problem. Using
tools from submodularity theory and convex analysis, we are able to provide a complete single-letter characterization
of the resulting achievable region (Theorem 37), which has the remarkable feature that it allows to exploit rate
which is in general insufficient for successful typicality decoding.
3
C. Notation and Conventions
1) Non-Standard Conventions: We will use the shorthand [l : k] := {l, l + 1, . . . , k − 1, k}. For convenience, we
define the sets K := [1 : K], J := [1 : J], and L := [1 : L] for K, J, L ∈ N. Furthermore, let Ω denote the set of
all pairs (A, B), where A, B ⊂ K are nonempty and disjoint. Note that Ω contains 3K − 2K+1 + 1 different pairs
(A, B). We also define Π as the set of all pairs (A, B), where A ⊆ J and B ⊆ L are nonempty. Hence, the set Π
contains 2J+L − 2J − 2L + 1 elements. For a total order < on a set E (cf. [22, Definition 1.5]) and e ∈ E we will
use the notation =e := {e0 ∈ E : e0 = e} and accordingly for w, < and v. e. g., given the total order < on {1, 2, 3}
with 3 < 1 < 2, we have =3 = {1, 2}, =1 = {2} and =2 = ∅.
2) Standard Conventions: We use the notation A, conv(A), and |A| for the topological closure, the convex hull,
and the cardinality of a set A, respectively. The superscript Ac is used to denote the complement of a set (or
event) A. Let 1A denote the indicator function of a set (or event) A. When there is no possibility of confusion
we identify singleton set with its element, e. g., we write {1, 2, 3} \ 1 = {2, 3}. Let R+ be the set of non-negative
reals and R− the set of non-positive reals. We denote random quantities and their realizations by capital, sans-serif
and lowercase letters, respectively. Furthermore, vectors are indicated by bold-face type and have length n, if not
otherwise specified. Random variables are assumed to be supported on finite sets. We use the same letter for the
random variable and for its support set, e. g., Y takes values in Y and X3 takes values in X3 . Given a random variable
X, we write pX ∈ P(X ) for its probability mass function (pmf), where P(X ) is the set of all pmfs on X . We use
the notation X ∼ p to indicate that X is distributed according to p . The notation X ∼ B(p) and X ∼ U(X ) is used
to denote a Bernoulli distribution with parameter p and a uniform distribution, respectively. We use E [X] and P{A}
for the expectation of the random variable X and the probability of an event A, respectively. Subscripts indicate
parts of vectors, e. g., xA := (xi )i∈A for a vector x = (x1 , x2 , . . . , xn ) and A ⊆ [1 : n]. We further use the common
notation xji := x{i,...,j} , xj := xj1 . If a vector is already carrying a subscript, it will be separated by a comma, e. g.,
x53,1 = (x3 )51 = (x3 )5 . Let 0 denote the all-zeros vector and ei = (ei,1 , ei,2 , . . . , ei,n ) ∈ Rn the ith canonical base
vector, i. e., ei,j = 1i (j). Additionally we use subscript sets to denote tuples of real values and their components,
K
K+1
e. g., xK ∈ RK or xΩ ∈ R3 −2 +1 . Naturally, slices of tuples are indexed by subsets, e. g., xA for a set A ⊆ K
is a slice of xK . This notation extends naturally to tuples of vectors, where the subscript indices are separated
by a comma, e. g., for xK ∈ RnK , we have xkA,l = (xA )kl ∈ R(k−l+1)|A| if A ⊆ K. We use the notation of [23,
Chapter 2] for information-theoretic quantities. In particular, given random variables (X, Y, Z) and pmfs p and q ,
H(X), H(X|Y), I(X; Y), I(X; Y|Z), and D(pkq) denote entropy, conditional entropy, mutual information, conditional
mutual information, and Kullback-Leibler divergence, respectively. All logarithms in this paper are to base e and
therefore all information theoretic quantities are measured in nats. The notation h2 (p) := −p log p − (1 − p) log(1 − p)
is used for the binary entropy function, a ∗ b := a(1 − b) + (1 − a)b is the binary convolution operation and the
symbol ⊕ denotes binary addition. The notation X −◦− Y −◦− Z indicates that X, Y, and Z form a Markov chain in
this order and X ⊥ Y denotes that X and Y are independent random variables. When generating codebooks we will
assume that the codebook size is an integer to keep the notation simple. We will use superscript to indicate that a
(42)
relation follows from a specific equation, e. g., the inequality a ≤ b follows from equation (42). Our achievability
proofs are based on the notion of robust typicality [24], also used in [15]. We will adopt the usual notation for types
and typical sequences and make use of the δ -convention [25, Convention 2.11]. For convenience, the necessary
notation and relevant results are summarized in Appendix C.
II. I NFORMATION -T HEORETIC B ICLUSTERING
A. Problem Statement
In this section we will introduce the information-theoretic biclustering problem (or biclustering problem for
short) with two sources and provide bounds on its achievable region. A schematic overview of the problem is
presented in Figure 1. Let (X, Z) be two random variables. The random vectors (X, Z) consist of n i.i.d. copies of
(X, Z). Given a block length n ∈ N and coding rates R1 , R2 ∈ R+ , an (n, R1 , R2 )-code for (X, Z) consists of two
functions f : X n → M1 and g : Z n → M2 such that the finite sets Mk satisfy log |Mk | ≤ nRk , k ∈ {1, 2}. Thus,
the coding rates R1 and R2 limit the complexity of the encoders.
Definition 1. For an (n, R1 , R2 )-code (f, g), we define the co-information of f and g as
1
Θ(f ; g) := I f (X); g(Z) .
n
(1)
4
X
Enc. 1 (rate R1 )
f (X)
nµ ≤ I f (X); g(Z)
Z
Enc. 2 (rate R2 )
g(Z)
Figure 1: Biclustering of two memoryless sources.
This co-information serves as a measure of the mutual relevance of the two encodings f (X) and g(Z). In contrast
to rate-distortion theory, we do not require a specific distortion measure; rather, we quantify the quality of a code in
pure information-theoretic terms, namely via mutual information. The idea is to find functions f and g that extract
a compressed version of the common randomness in the observed data X and Z.
Definition 2. A triple (µ, R1 , R2 ) is achievable if for some n ∈ N there exists an (n, R1 , R2 )-code (f, g) such that
Θ(f ; g) ≥ µ.
(2)
The achievable region R is defined as the closure of the set R of achievable triples.
Remark 1. Note that a standard time-sharing argument shows that R is a convex set (cf. [13, Section 4.4]).
We also point out that stochastic encodings cannot enlarge the achievable region as any stochastic encoding can be
represented as the convex combination of deterministic encodings and R is convex.
B. Related Works
The biclustering problem with two sources turns out to be connected to a hypothesis testing problem and a pattern
recognition problem. In this section we will clarify these connections explicitly, using the “multi-letter” region R∗ .
Definition 3. Let R∗ be the set of triples (µ, R1 , R2 ) such that there exist n ∈ N and random variables U, V
satisfying U −◦− X −◦− Z −◦− V and
nR1 ≥ I(U; X),
(3)
nµ ≤ I(U; V).
(5)
nR2 ≥ I(V; Z),
(4)
We will now showcase the connection between the biclustering problem and several other information-theoretic
problems. These equivalences will provide us with the achievability of R∗ .
1) Hypothesis Testing (HT): Consider the hypothesis testing problem with data compression when testing against
independence, which was studied in [15, Section 6]. For completeness sake we briefly describe the problem setup.
Given the potentially dependent sources (X, Z), define the independent random variables X ∼ pX and Z ∼ pZ .
Definition 4. An (n, R1 , R2 ) hypothesis test consists of an (n, R1 , R2 )-code (fn , gn ) for (X, Z) and a set An ⊆
M1 × M2 , where M1 and M2 are the ranges of fn and gn , respectively. The type I and type II error probabilities
of (fn , gn , An ) are defined as
αn := P fn (X), gn (Z) ∈ An ,
(6)
βn := P fn (X), gn (Z) ∈
/ An ,
(7)
respectively. A triple (µ, R1 , R2 ) is HT-achievable if, for every ε > 0, there is a sequence of (n, R1 , R2 ) hypothesis
tests (fn , gn , An ), n ∈ N such that
lim αn ≤ ε,
n→∞
(8)
5
Let RHT
1
lim − log βn ≥ µ.
n→∞ n
denote the set of all HT-achievable triples.
(9)
Theorem 5. RHT = R∗ .
Proof. Assume (µ, R1 , R2 ) ∈ RHT . For ε > 0, pick an (n, R1 , R2 ) hypothesis test (fn , gn , An ) such that αn ≤ ε
and log βn ≤ −n(µ − ε). The random variables U := fn (X) and V := gn (Z) satisfy the required Markov chain as
well as (3) and (4). We apply the log-sum inequality [23, Theorem 2.7.1] and obtain for any ε0 > 0, provided that ε
is small enough and n is large enough,
1 − αn
αn
+ αn log
βn
1 − βn
1
1
+ αn log
= −h2 (αn ) + (1 − αn ) log
βn
1 − βn
≥ −(1 − ε) log βn − ε0
I(U; V) ≥ (1 − αn ) log
≥ (1 − ε)n(µ − ε) − ε
0
0
≥ n(µ − ε ).
(10)
(11)
(12)
(13)
(14)
This shows that (µ − ε0 , R1 , R2 ) ∈ R∗ and consequently (µ, R1 , R2 ) ∈ R∗ .
Consider the bound [15, Corollary 6]. It shows that (nµ, nR1 , nR2 ) is asymptotically achievable in the hypothesis
testing problem for the vector source (X, Z) if (µ, R1 , R2 ) ∈ R∗ . I. e., for any ε, ε0 > 0, there is a sequence of
(k, nR1 + ε0 , nR2 + ε0 ) hypothesis tests (fk , gk , Ak ) for (X, Z), k ∈ N such that
lim αk ≤ ε,
(15)
1
lim − log βk ≥ nµ − ε0 .
k
(16)
k→∞
This shows that µ −
proof.
ε0
n , R1
+
ε0
n , R2
+
ε0
n
k→∞
∈ RHT for the source (X, Z) and as ε0 was arbitrary, this completes the
We can leverage the equivalence with hypothesis testing to show that R∗ is indeed a multi-letter characterization
of R.
Corollary 6. R = R∗ .
Proof. To prove R ⊆ R∗ , assume (µ, R1 , R2 ) ∈ R and choose n, f , and g according to Definition 2. Defining
U := f (X) and V := g(Z) yields inequalities (3)–(5) and satisfies the required Markov chain.
We will show RHT ⊆ R, which is equivalent to R∗ ⊆ R by Theorem 5. Assuming (µ, R1 , R2 ) ∈ RHT , choose an
arbitrary ε > 0 and pick an (n, R1 , R2 ) hypothesis test (fn , gn , An ) such that αn ≤ ε and − log βn ≥ n(µ − ε). By
the same reasoning as in (14), the (n, R1 , R2 )-code (fn , gn ) achieves Θ(fn ; gn ) ≥ µ − ε0 for any ε0 > 0, implying
(µ, R1 , R2 ) ∈ R.
2) Pattern Recognition (PR): Consider the pattern recognition problem introduced in [16]. For completeness sake
we restate the problem
here.
Let X(i), Z(i) be i.i.d. copies of (X, Z), independently generated for each i ∈ N.
Definition 7. A triple (µ, R1 , R2 ) is said to be PR-achievable if, for any ε > 0, there exists an (n, R1 , R2 )-code
nµ
(f, g) for (X, Z) and a function φ : (M1 )e × M2 → [1 : enµ ] such that,
P W = φ C, g(Z(W)) ≥ 1 − ε,
(17)
where C := f (X(i))i∈[1 : enµ ] is the compressed codebook and X(i), Z(i) i∈N ⊥ W ∼ U([1 : enµ ]). Let RPR denote
the set of all PR-achievable triples.
Remark 2. The variant of the inner bound for the pattern recognition problem stated in [16, Theorem 1] is flawed.
To see this, note that (using the notation of [16]) the point (Rx = 0, Ry = b, Rc = b) is contained in Rin (choose
U = V = ∅) for any b > 0 even if the random variables X and Y are independent. But this point is clearly not
6
achievable in general. However, the region Rin defined in the right column of [16, p. 303] coincides with our
findings and the proof given in [16, Appendix A] holds for this region.
Proposition 8. RPR = R∗ .
Proof. Assume (µ, R1 , R2 ) ∈ RPR and for an arbitrary ε > 0 and sufficiently large n ∈ N choose appropriate
functions f , g , φ satisfying (17). The random variables U := f (X) and V := g(Z) satisfy the required Markov
chain as well as (3) and (4). Furthermore,
I(U; V) = I f (X); g(Z)
(18)
= I C; g(Z(W)) W
(19)
= I C; g(Z(W)), W
(20)
≥ I C; W g(Z(W))
(21)
= H W g(Z(W)) − H W C, g(Z(W))
(22)
≥ nµ − H W φ C, g(Z(W))
(23)
(17)
≥ nµ − h2 (ε) − εnµ.
(24)
The equality in (19) holds as X(i) ⊥ Z(j) for i 6= j , (20) follows from W ⊥ C, (23) follows from W ⊥ Z(W), the
fact that H(W) = nµ and the data processing inequality [23, Theorem 2.8.1]. Fano’s inequality [23, Theorem 2.10.1]
was used in (24). This shows (µ, R1 , R2 ) ∈ R∗ as ε was arbitrary.
To show the other direction, we apply the achievability result [16, Theorem 1] to the multi-letter source (X, Z).
Assuming (µ, R1 , R2 ) ∈ R∗ , we know that for some n ∈ N there are random variables (U, V) satisfying the Markov
chain U −◦− X −◦− Z −◦− V and (3)–(5) hold. By [16, Theorem 1], the triple (nµ, nR1 , nR2 ) is asymptotically
achievable for the pattern recognition problem for the source (X, Z) with an arbitrary error probability ε > 0. For
any ε0 > 0 we can find k ∈ N, a (k,
nR1 + ε0 , nR2 + ε0 )-code (f, g) for
is
(X, Z), and a function φ, such that (17)
0
ε0
ε0
ε0
k(nµ−ε
)
0
satisfied with W ∼ U 1 : e
. Thus, µ − n , R1 + n , R2 + n ∈ RPR for the source (X, Z) and as ε was
arbitrary, this completes the proof.
C. Bounds on the Achievable Region
We first provide outer bounds on the set of achievable triples for biclustering with two sources.
Theorem 9. We have R ⊆ Ro ⊆ R0o , where the two regions Ro and R0o are given by
[
(µ, R1 , R2 ) : R1 ≥ I(U; X), R2 ≥ I(V; Z),
Ro :=
U,V
R0o
:=
[
U,V
µ ≤ I(U; X) + I(V; Z) − I(UV; XZ) ,
(µ, R1 , R2 ) : R1 ≥ I(U; X), R2 ≥ I(V; Z), µ ≤ min{I(U; Z), I(V; X)} ,
(25)
(26)
with U and V any pair of random variables satisfying U −◦− X −◦− Z and X −◦− Z −◦− V.
Theorem 9 follows from the outer bound for the pattern recognition problem [16, Appendix B] via the equivalence
shown in Section II-B2. Nonetheless, we provide a short, self-contained proof in Appendix A1 for the sake of
completeness.
The regions Ro and R0o are both convex since a time-sharing variable can be incorporated into U and V. Furthermore,
0
Ro remains unchanged when U and V are required to satisfy the complete Markov chain U −◦− X −◦− Z −◦− V.
The numerical computation of the outer bounds requires the cardinalities of the auxiliary random variables to be
bounded. We therefore complement Theorem 9 with the following result, whose proof is provided in Appendix A2.
Proposition 10. We have Ro = conv(So ) and R0o = conv(So0 ), where the regions So and So0 are defined as Ro
and R0o , respectively, but with the additional cardinality bounds |U| ≤ |X | and |V| ≤ |Z|.
The cardinality bounds in this result are tighter than the usual bounds obtained with the convex cover method [13,
Appendix C], where the cardinality has to be increased by one. Thus, when dealing with binary sources in
Section II-E, binary auxiliary random variables suffice. The smaller cardinalities come at the cost of convexification
in Proposition 10 since the regions So and So0 themselves are not necessarily convex.
7
We next state an inner bound for the biclustering problem. A more general inner bound will be proved in Theorem 25
in Section III-A for the extension of the biclustering problem to more than two sources.
Theorem 11. We have Ri ⊆ R where
[
Ri :=
(µ, R1 , R2 ) : R1 ≥ I(U; X), R2 ≥ I(V; Z), µ ≤ I(U; V) ,
(27)
U,V
with auxiliary random variables U, V satisfying U −◦− X −◦− Z −◦− V.
Theorem 11 follows from the achievability result of Han on the hypothesis testing problem [15, Corollary 6],
leveraging the equivalence detailed in Theorem 5 and Corollary 6. Alternatively, it also follows from the inner
bound for the pattern recognition problem [16, Appendix A] using Corollary 6 and Proposition 8. The main
differences between the outer and the inner bound lies in the Markov conditions, a phenomenon that also occurs
with Berger-Tung type bounds (cf. [26, Chapter 7] or [13, Section 12.2]). Note that Ro and Ri would coincide if
the Markov condition U −◦− X −◦− Z −◦− V were imposed in the definition of Ro .
Employing a binning scheme would not enlarge the inner bound. The intuition is that binning reduces redundant
information transmitted by both encoders, whereas in information-theoretic biclustering this quantity should actually
be maximized.
A tight bound on the achievable region can be obtained if µ is not greater than the common information (cf.
[27]–[29]) of X and Z, as stated in the following corollary.
Corollary 12. If Y = ζ1 (X) = ζ2 (Z) is common to X and Z in the sense of [28] and 0 ≤ µ ≤ H(Y) then
(µ, R1 , R2 ) ∈ R if and only if µ ≤ min{R1 , R2 }.
Proof. Theorem 9 entails µ ≤ min{R1 , R2 } for any (µ, R1 , R2 ) ∈ R. With U = V = Y, Theorem 11 implies
(H(Y), H(Y), H(Y)) ∈ R. Using time-sharing with 0 ∈ R we obtain (µ, µ, µ) ∈ R for 0 ≤ µ ≤ H(Y) and hence
(µ, R1 , R2 ) ∈ R if µ ≤ min{R1 , R2 }.
The inner bound Ri can be improved by convexification. Furthermore, we incorporate the same strong cardinality
bounds as for the outer bound (cf. Proposition 10), thereby enabling us to use binary auxiliary random variables
when dealing with binary sources in Section II-E.
Proposition 13. We have Si0 := conv(Si ) = conv(Ri ) ⊆ R where Si is defined as Ri , but with the additional
cardinality bounds |U| ≤ |X |, and |V| ≤ |Z|. Furthermore, Si0 can be explicitly expressed as
[
Si0 =
(µ, R1 , R2 ) : R1 ≥ I(U; X|Q), R2 ≥ I(V; Z|Q), µ ≤ I(U; V|Q) ,
(28)
U,V,Q
where U, V, and Q are random variables such that pX,Z,U,V,Q = pQ pX,Z pU|X,Q pV|Z,Q , |U| ≤ |X |, |V| ≤ |Z|, and
|Q| ≤ 3.
The proof of this result is given in Appendix A3.
D. The Information Bottleneck Method
The information-theoretic problem posed by the information bottleneck (IB) method [17] can be obtained as a
special case from the biclustering problem. We will introduce the problem setup and subsequently show how it
can be derived as a special case of Definition 2. Note that the definition slightly differs from [12, Definition 1].
However, the achievable region is identical.
Definition 14. A pair (µ, R1 ) is IB-achievable if, for some n ∈ N, there exists f : X n → M1 with log|M1 | ≤ nR1
and
1
µ ≤ I(f (X); Z).
(29)
n
Let RIB be the set of all IB-achievable pairs.
Furthermore, Definition 14 is equivalent to noisy lossy source coding [30] with logarithmic loss (LL) distortion [18].
8
Definition 15. A pair (µ, R1 ) is LL-achievable if, for some n ∈ N, there exist f : X n → M1 and φ : M1 → P(Z n )
with log|M1 | ≤ nR1 such that
1
E [dLL (φ ◦ f (X), Z)],
n
where dLL (p, z) := − log p(z) is LL distortion. Let RLL be the set of all LL-achievable pairs.
µ≥
(30)
Using results from [18] we can show the following equivalences.
Proposition 16. For a pair (µ, R1 ), the following are equivalent:
1) (µ, R1 ) ∈ RIB .
2) (H(Z) − µ, R1 ) ∈ RLL .
3) (µ, R1 , log|Z|) ∈ R .
4) There exists a random variable U such that U −◦− X −◦− Z, I(X; U) ≤ R and I(Z; U) ≥ µ.
Proof. The equivalence “1 ⇔ 3” holds as Definition 2 collapses to Definition 14 for R2 = log|Z|. “2 ⇔ 4” is
shown in [18, Section III.F]. To show “3 ⇔ 4” apply the outer bound R0o of Theorem 9 and Theorem 11 (identifying
V = Z).
The tradeoff between “relevance” and “complexity” can equivalently be characterized by the IB function (cf. [18],
[31]) µIB (R) := sup{µ : (µ, R) ∈ RIB }. Proposition 16 provides
µIB (R) =
max
U : I(U;X)≤R
U−
◦−X−◦−Z
I(U; Z).
(31)
Interestingly, the function (31) is the solution to a variety of different problems in information theory. As mentioned
in [31], (31) is the solution to the problem of loss-less source coding with one helper [32], [33]. Witsenhausen and
Wyner [34] investigated a lower bound for a conditional entropy when simultaneously requiring another conditional
entropy to fall below a threshold. Their work was a generalization of [35] and furthermore related to [32], [36]–[38].
The conditional entropy bound in [34] turns out to be an equivalent characterization of (31). Furthermore, µIB
characterizes the optimal error exponent, when testing against independence with one-sided data compression [39,
Theorem 2]. Also in the context of gambling in the horse race market, (31) occurs as the maximum incremental
growth in wealth when rate-limited side-information is available to the gambler [40, Theorem 3].
E. Doubly Symmetric Binary Source
We will consider the special case where (X, Z) ∼ DSBS(p) is a doubly symmetric binary source [13, Example 10.1]
with parameter p, i. e., X ∼ B( 21 ), X ⊥ N ∼ B(p), and Z := X ⊕ N. The cardinality bounds in Propositions 10
and 13 will enable us to use binary auxiliary random variables. We first show that the inner bound Si0 and the outer
bound R0o do not coincide.
Proposition 17. For the source (X, Z) ∼ DSBS(p) with p ∈ (0, 1) and p 6= 21 , we have Si0 6= R0o .
The proof of this proposition is given in Appendix A4. We conjecture that there is also a gap between Si0 = conv(Si )
and the outer bound Ro .
Conjecture 18. There exists p ∈ [0, 1], such that Si0 6= Ro for the source (X, Z) ∼ DSBS(p).
To support Conjecture 18, we will introduce a region Sb ⊆ Si and show that conv(Sb ) 6= Ro . Let Sb be defined as
[
Sb :=
(µ, R1 , R2 ) : R1 ≥ log 2 − h2 (α),
0≤α,β≤ 12
R2 ≥ log 2 − h2 (β),
µ ≤ log 2 − h2 (α ∗ p ∗ β) .
(32)
By choosing U = X ⊕ N1 and V = Z ⊕ N2 , where N1 ∼ B(α) and N2 ∼ B(β) are independent of (X, Z) and of
each other, it follows that Sb ⊆ Si . To illustrate the tradeoff between complexity (R1 , R2 ) and relevance (µ), the
boundary of Sb is depicted in Figure 2 for p = 0.1.
Based on numerical experiments, we conjecture the following.
9
I(X; Z)
0.2
0.1
µ
0
log 2
0.4
0.2
0.4
0.2
log 2
R2
R1
Figure 2: Boundary of Sb for p = 0.1.
Conjecture 19. For the source (X, Z) ∼ DSBS(p) with p ∈ [0, 1] we have Si0 = conv(Sb ), or equivalently
Si ⊆ conv(Sb ).
The natural, stronger conjecture that Sb = Si already appeared in [16, Conjecture 1, Eq. (14)]. However, there is a
counterexample [41]1 .
Proposition 20. For the source (X, Z) ∼ DSBS(0) we have Sb 6= Si .
U
a
0
X=Z
1
0
a
0
1−a
V
0
1
1
1−a
1
Figure 3: Random variables for the proof of Proposition 20.
Proof. For a ∈ [0, 1] we define (U, V) by the binary channels
depicted in Figure 3, satisfying U −◦− X −◦− Z −◦− V. We
obtain (µ, R, R) ∈ Si with R = I(U; X) = I(V; Z) = h2 a2 − 21 h2 (a) and µ = I(U; V) = 2R−a log(2). For a = 0.8
we have µ ≈ 0.291103 and R ≈ 0.42281. On the other hand, we obtain µb := max{µ̂ : (µ̂, R, R) ∈ Sb } < 0.285594
using (32) with α = β ≈ 0.07658. As µb < µ we have (µ, R, R) ∈
/ Sb .
This argument can be verified numerically using interval arithmetic [42]. Code written in the Octave Programming
Language [43] using its interval package [44] can be found at [45].
Note that Proposition 20 does not
impact Conjecture 19 as it concerns the case p = 0. For p = 0 we have X = Z
and Corollary 12 implies R = (µ, R1 , R2 ) : R1 , R2 ≥ 0 and µ ≤ min{R1 , R2 , log 2} . It is easily verified that
R = conv(Sb ) and thus Conjecture 19 holds for p = 0 by Proposition 13.
In fact, the entire statement [16, Conjecture 1] does not hold. The second part [16, Conjecture 1, Eq. (15)] claims
that conv(Sb ) = Ro . In what follows, we show how to construct a counterexample.
Proposition 21. For the source (X, Z) ∼ DSBS(0.1), we have conv(Sb ) 6= Ro .
Proposition 21 shows that Conjecture 18 follows directly from Conjecture 19.
1
https://mathoverflow.net/questions/213084/do-binary-symmetric-channels-maximize-mutual-information
10
I(X; Z)
µ
µ̂o
µ̂b
0.365
0.36
0.675
0.68
0.685
R
0.69 log 2
Figure 4: Numeric evaluation of µ̂o and µ̂b for p = 0.1.
To prove Proposition 21 we construct a point (µ, R, R) ∈ Ro that satisfies (µ, R, R) ∈
/ conv(Sb ). To this end,
define the concave functions µ̂b (R) := max{µ : (µ, R, R) ∈ conv(Sb )} and µ̂o (R) := max{µ : (µ, R, R) ∈ Ro }
for R ∈ [0, log 2]. In order to show conv(Sb ) 6= Ro , it suffices to find R̂ ∈ [0, log 2] with µ̂b (R̂) < µ̂o (R̂).
We can numerically compute an upper bound for the function µ̂b . For α, β ∈ [0, 21 ], we calculate
e1 := log 2 − h2 (α),
R
e2 := log 2 − h2 (β), and
R
µ
e := log 2 − h2 (α ∗ p ∗ β)
(33)
(34)
(35)
e1 , R
e 2 ).
on a suitably fine grid and upper bound the upper concave envelope of the implicitly defined function µ
e(R
e
e
Evaluating it at R = R1 = R2 yields an upper bound for µ̂b (R).
On the other hand, we can obtain a lower bound for µ̂o by computing (25) for specific probability mass functions
that satisfy the Markov constraints in Theorem 9. Note that based on the cardinality bound in Proposition 10,
we can restrict the auxiliary random variables U and V to be binary. We randomly sample the binary probability
mass functions that satisfy the Markov constraints in Theorem 9 (but not necessarily the long Markov chain
U −◦− X −◦− Z −◦− V) and in doing so encounter points strictly above the graph of µ̂b . Figure 4 shows the resulting
bounds for p = 0.1 in the vicinity of R = log 2. Albeit small, there is clearly a gap between µ̂b and µ̂o outside the
margin of numerical error.
Proof of Proposition 21. We observed the largest gap between the two bounds at a rate of R̂ ≈ 0.675676. The
particular distribution of (U, V) at this rate, resulting from optimizing over the distributions that satisfy the Markov
constraints in Theorem 9 is given in Table I for reference. Note that this is an exact conditional pmf (i. e., not
a numerical approximation) that satisfies the Markov chains U −◦− X −◦− Z and X −◦− Z −◦− V. It achieves
I(V; Z) + I(U; X) − I(UV; XZ) ≈ 0.35972 which is ∆ ≈ 1.985673 · 10−4 above the inner bound, thus proving
Proposition 21. Using interval arithmetic [42] this claim can be verified numerically. Code written in the Octave
Programming Language [43] using its interval package [44] can be found at [45]. It uses the distribution given in
Table I.
We firmly believe that a tight characterization of the achievable region requires an improved outer bound. However,
using current information theoretic tools, it appears very challenging to find a manageable outer bound based on the
full Markov chain U −◦− X −◦− Z −◦− V.
Remark 3. Recently, Kumar and Courtade introduced a conjecture [46], [47] concerning Boolean functions that
maximize mutual information. Their work was inspired by a similar problem in computational biology [48]. A
weaker form of their conjecture [47, Section IV, 2)], which was solved in [49], corresponds to a zero-rate/one-bit
variant of the binary example studied here.
11
u
v
x
z
P{U = u, V = v|X = x, Z = z}
0
0
0
0
0
0
0
0
1
1
1
1
1
1
1
1
0
0
0
0
1
1
1
1
0
0
0
0
1
1
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0.995358146217353406525
0.00249767559844423319075
0.002498344003957310643325
0.00034313919194834475
0.002142603857654094275
0.99500307447656326760925
0.000000905921035188556675
0.00215611073304415445
0.002142603857654094275
0.00000157432654826600925
0.995002406071050190156675
0.00215611073304415445
0.000356646067338404925
0.00249767559844423319075
0.002498344003957310643325
0.99534463934196334635
Table I: Distribution resulting from random search.
III. C LUSTERING WITH M ULTIPLE S OURCES
A. Problem Statement and Results
In this section we extend the information-theoretic biclustering problem introduced in Section II-A to the case
of multiple sources and we provide bounds on the associated achievable region. A schematic illustration of the
problem is shown in Figure 5.
X1
W1 = f1 (X1 ) (rate R1 )
X2
W2 = f2 (X2 ) (rate R2 )
X3
W3 = f3 (X3 ) (rate R3 )
X4
W4 = f4 (X4 ) (rate R4 )
X5
W5 = f5 (X5 ) (rate R5 )
..
.
..
.
XK
WK = fK (XK ) (rate RK )
nµA,B ≤ I WA ; WB
Figure 5: Clustering of multiple sources.
Let XK (K = [1 : K]) be K random variables, taking values in the finite sets XK . The random vectors XK
consist of n i.i.d. copies of XK . For n ∈ N and RK ∈ RK
+ , an (n, RK )-code fK for XK consists of K functions
n
fk : Xk → Mk , where Mk is an arbitrary finite set with log |Mk | ≤ nRk for each k ∈ K. Recall that the symbol
12
K
K+1
µΩ refers to a tuple µΩ ∈ R3 −2
and hence, |Ω| = 3K − 2K+1 + 1.
+1 ,
where Ω is the set of all pairs (A, B) with A, B ⊂ K nonempty and disjoint
Definition 22. Consider an (n, RK )-code fK with Wk := fk (Xk ); for any (A, B) ∈ Ω we define the co-information
of fA and fB as
1
I(WA ; WB ).
(36)
n
Definition 23. A point (µΩ , RK ) is achievable if, for some n ∈ N, there exists an (n, RK )-code fK such that for
all (A, B) ∈ Ω,
Θ(fA ; fB ) :=
Θ(fA ; fB ) ≥ µA,B .
(37)
The set of all achievable points is denoted R and we refer to its closure R as achievable region.
Remark 4. A standard time sharing argument can be used to show that R is a convex set (cf. [13, Section 4.4]).
We first state an outer bound for the achievable region, whose proof is provided in Appendix B1.
Theorem 24. We have the outer bounds R ⊆ Ro ⊆ R0o . Here, the region R0o is defined as
X
[
Rk ≥ I(UA ; XK |UC ) for A, C ⊆ K,
(µΩ , RK ) :
R0o :=
k∈A
UK
and µA,B ≤ I(UA ; XB ) for (A, B) ∈ Ω ,
(38)
where the auxiliary random variables UK satisfy UA −◦− XA −◦− XK\A for every A ⊆ K. The region Ro is defined
just as R0o only that the inequality for the relevance µA,B is replaced with
µA,B ≤ I(UA ; XA ) + I(UB ; XB ) − I(UA UB ; XA XB ).
(39)
µA,B ≤ I(UA ; UB ) − I(UA ; UB |XA XB ).
(40)
Remark 5. In particular we have the Markov chains UA −◦− XA −◦− XB and UB −◦− XB −◦− XA . Using Lemma 43
we can write (39) equivalently as
The next result, whose proof is detailed in Section III-C, provides an inner bound.
Theorem 25. We have Ri ⊆ R where the region Ri consists of all points (µΩ , RK ) for which there exist random
variables UK satisfying Uk −◦− Xk −◦− (XK\k , UK\k ) for all k ∈ K and for all (A, B) ∈ Ω there exist subsets Ae ⊆ A
and Be ⊆ B such that
X
Rk ≥ I XA0 ; UA0 UA\A0 for all A0 ⊆ A with A0 ∩ Ae =
6 ∅,
(41)
k∈A0
X
k∈B
0
Rk ≥ I XB0 ; UB0 UB\B0
µA,B ≤ I UAe; UBe .
for all B 0 ⊆ B with B 0 ∩ Be 6= ∅,
(42)
(43)
Remark 6. In contrast to the case of two sources, binning does help for K > 2 sources. For illustration, consider
the case K = 3 and assume we are only interested in maximizing Θ(f1 , f2 ; f3 ). Then any information encoded by
both f1 and f2 is redundant as it does not increase Θ(f1 , f2 ; f3 ). The corresponding rate loss can be reduced by a
quantize-and-bin scheme (cf. [26], [32], [50]).
The proof that Ri is indeed achievable uses typicality coding and binning. The conditions (41) and (42) ensure that
the messages of encoders Ae and Be can be correctly decoded from the output of the encoders A and B , respectively.
By (43), these suffice to ensure that µA,B is achievable. Intuitively, the encoders A \ Ae and B \ Be act as helpers.
The special case Ae = A, Be = B for every (A, B) ∈ Ω corresponds to no binning at all, as (41) and (42) then imply
Rk ≥ I(Xk ; Uk ) for all k ∈ K.
The inner bound in Theorem 25 cannot be tight, as evidenced by the Körner-Marton problem [14]. For K = 3
consider independent random variables X1 ∼ B( 12 ) and X3 ∼ B(p) with p ∈ (0, 1) and p 6= 12 and define
X2 := X1 ⊕ X3 . The point (µΩ , RK ) where R3 = log 2, R1 = R2 = H(X3 ) = h2 (p), and µA,B = 0 except for
13
µ{1,2},{3} = H(X3 ) = h2 (p) is achievable [14, Theorem 1]. However, the quantize-and-bin scheme used in the proof
of Theorem 25 cannot achieve this point [14, Proposition 1].
Finally, note that in general Ri is not convex and thus Theorem 25 can be strengthened to conv(Ri ) ⊆ R.
However, characterizing conv(Ri ) using a time-sharing random variable is tedious due to the freedom of choosing
the index sets Ae, and Be for each (A, B) ∈ Ω in Theorem 25.
The following cardinality bounds show that Ri is computable (see Appendix B2 for the proof).
Proposition 26. The region Ri remains unchanged if the cardinality bound |Uk | ≤ |Xk | + 4K is imposed for every
k ∈ K.
B. A Special Case: The CEO Problem
In this section we study a special case of the clustering problem that corresponds to a variant of the CEO problem
[51] in which the usual distortion criterion is replaced with mutual information (MI). This problem turns out to
be equivalent to the classical CEO problem with logarithmic loss distortion as analyzed in [18]. The equivalence
follows in the same fashion as the equivalence between the IB problem and lossy source coding with log-loss
distortion, shown in Section II-D. Using results from [18], we will show that our inner bound becomes tight in this
special case.
We consider a CEO problem under a mutual information constraint where random variables XJ are encoded
to be maximally informative about another set of random variables YL (remember, J = [1 : J] and L = [1 : L]).
Recall that we use the symbol νΠ to denote a tuple νΠ ∈ RΠ , where Π consists of all pairs (A, B) of nonempty
sets A ⊆ J and B ⊆ L.
Definition 27. A point (νΠ , RJ ) is MI-achievable if for some n ∈ N there exists an (n, RJ )-code fJ for XJ such
that for all (A, B) ∈ Π,
1
I(WA ; YB ) ≥ νA,B ,
n
where Wj := fj (Xj ) for j ∈ J . Denote the set of all MI-achievable points by RMI .
(44)
The region RMI is a special case of R.
Proposition 28. Letting K := J + L and XK := (XJ , YL ), we have (νΠ , RJ ) ∈ RMI if and only if (µΩ , RK ) ∈ R,
where
(45)
µA,J+B = 1Π (A, B)νA,B ,
(46)
Rk = log|Xk | for k > J.
Proof. Let (νΠ , RJ ) ∈ RMI be achieved by the (n, RJ )-code fJ in the sense of Definition 27. We extend fJ
to an (n, RK )-code fK by setting fk (xk ) = xk for k > J . This code achieves (µΩ , RK ) ∈ RMI in the sense of
Definition 23.
Let (µΩ , RK ) ∈ R be achieved by the (n, RK )-code fK in the sense of Definition 23. The restriction fJ then
achieves (νΠ , RJ ) ∈ RMI in the sense of Definition 27.
Remark 7. The convexity of R (Remark 4) and Proposition 28 imply that RMI is a convex set.
To shorten notation we will introduce the set of random variables
P∗ := {UJ , Q : Q ⊥ (XJ , YL ), Uj −◦− (Xj , Q) −◦− (XJ \j , YL , UJ \j ) for all j ∈ J }.
(47)
We can obtain an inner bound on RMI directly from Theorem 25 as stated in the following corollary.
(MI)
(MI)
Corollary 29. An inner bound for the achievable region is given by Ri
⊆ RMI where the region Ri
consists
of all points (νΠ , RJ ) such that there exist random variables (UJ , ∅) ∈ P∗ and for all (A, B) ∈ Π there exists a
subset Ae ⊆ A such that
X
Rk ≥ I XA0 ; UA0 UA\A0 for all A0 ⊆ A with A0 ∩ Ae =
6 ∅,
(48)
k∈A0
14
νA,B ≤ I UAe; YB .
(49)
Proof. The result follows from Theorem 25 and Proposition 28 with auxiliary random variables Uk = Xk for
k > J.
We next argue that the CEO problem introduced in Definition 27 is equivalent to the log-loss distortion approach
in [18]. For A ⊆ J and B ⊆ L we consider a decoding function gA,B : MA → P(YBn ) that produces a probabilistic
estimate of YB given the output of the encoders A. The quality of this probabilistic estimate is measured by log-loss
distortion.
Definition 30. A point (νΠ , RJ ) is LL-achievable if, for some n ∈ N, there exists an (n, RJ )-code fJ such that
for all (A, B) ∈ Π there is a decoding function gA,B : MA → P(YBn ) with
1
E dLL (gA,B (WA ), YB ) ≤ νA,B ,
(50)
n
where Wj := fj (Xj ), j ∈ J . Let RLL be the set of all LL-achievable points.
We note that [18] considers the case where νA,B 6= ∞ only if A = J and B = L = {1}. To show that RMI and
RLL are equivalent, we first state an auxiliary lemma which is essentially [18, Lemma 1] and provided just for the
sake of completeness (see Appendix B3 for the proof).
Lemma 31. For two random variables (X, Y) and two functions f : X → M and g : M → P(Y), where M is an
arbitrary set, we have
E dLL g(f (X)), Y ≥ H(Y|f (X))
(51)
with equality if and only if g(m) = P{Y = · |f (X) = m} for all m ∈ M with P{f (X) = m} =
6 0.
Loosely speaking, Lemma 31 implies that the “MI performance” of an encoder is at least as good as its “LL
performance” – independent of the decoder. On the other hand, the optimum can always be achieved when equality
holds in (51). This is the basis for the next result.
0 , R ) ∈ R , where ν 0
Lemma 32. (νΠ , RJ ) ∈ RLL if and only if (νΠ
J
MI
A,B := H(YB ) − νA,B .
0 ,R ) ∈ R
Proof. Assume that (νΠ
J
MI is achieved by the (n, RJ )-code fJ , i. e., (44) holds for all (A, B) ∈
Π with Wj := fj (Xj ). By Lemma 31, choosing the decoding functions gA,B (mA ) = pYB |WA ( · |mA ) implies
(44)
1
1
n E [dLL (gA,B (WA ), Y B )] = n H(Y B |WA ) ≤ νA,B
To show RLL ⊆ RMI , assume that (νΠ , RJ )
and thus (νΠ , RJ ) ∈ RLL .
∈ RLL is achieved by the (n, RJ )-code fJ and the decoding
0
function gA,B , i. e., (50) holds for all (A, B) ∈ Π with Wj := fj (Xj ). Lemma 31 then implies I(WA ; YB ) ≥ nνA,B
0
and hence (νΠ , RJ ) ∈ R.
For the rest of this section, assume L = 1 and Xj −◦− Y1 −◦− XJ \j for all j ∈ J . For brevity we will write
Y := Y1 and νA := νA,1 in the following. As a consequence of Lemma 32, the results in [18] directly apply to the
CEO problem with a mutual information constraint.
Lemma 33. Assume νA = 0 whenever A 6= J . Then (νΠ , RJ ) ∈ RMI if and only if there exist random variables
(UJ , Q) ∈ P∗ and the following inequalities hold:
X
Rk ≥ I XA0 ; UA0 UJ \A0 , Q for all A0 ⊆ J ,
(52)
k∈A0
νJ ≤ I(UJ ; Y|Q).
(53)
Proof. (νΠ , RJ ) ∈ RMI follows by applying Corollary 29 with Ae = ∅ for A =
6 J and Ae = J for A = J , taking
into account the convexity of RMI (Remark 7). The converse follows from [18, Lemma 5] and Lemma 32.
Remark 8. The achievable region of the multiterminal source coding problem with logarithmic-loss distortion,
introduced in [18, Section II], can be obtained as a special case of RMI as well. Choose J = L = 2 and set Yj = Xj ,
(MI)
j ∈ {1, 2}. The inner bound Ri
is also tight (up to convexification) due to the results in [18].
15
C. Proof of Theorem 25
The proof of Theorem 25 extends the methods developed in [15] for the hypothesis testing problem (cf.
Section II-B1) to a setup with multiple sources. We begin by extending [15, Lemma 8] and incorporating a binning
strategy.
Lemma 34 (Existence of a code). Let ε > 0, Uk −◦− Xk −◦− (XK\k , UK\k ) for all k ∈ K, and RK ∈ RK
+ . Then,
for sufficiently large n ∈ N we can obtain an (n, RK + ε)-code fK with Wk := fk (Xk ) and decoding functions
gA,Ae : MA → U ne for each A, Ae ⊆ K with ∅ 6= Ae ⊆ A such that the following two properties hold.
A
For every (A, B) ∈ Ω and ∅ 6= Ae ⊆ A as well as ∅ 6= Be ⊆ B we have
gA,Ae(MA ) × gB,Be(MB ) ∩ T[Un eU e] ≤ exp n(I UAeUBe; XAeXBe + ε) .
(54)
A
B
Furthermore, if (41) and (42) hold, then
n
o
P gA,Ae(WA ), XA , XB , gB,Be(WB ) ∈
/ T[Un eXA XB U e] ≤ ε.
A
B
(55)
The proof of Lemma 34 is provided in Appendix B4.
Furthermore we will need the following set of random variables.
Definition 35. For random variables (A, B, C, D) and δ ≥ 0, define the set of random variables
e B,
e C,
e D
e : (A,
e B)
e ∈T
e e
e e
Sδ (A, B, C, D) := {A,
[AB]δ , (C, D) ∈ T[CD]δ , (A, D) ∈ T[AD]δ }.
(56)
Consider (µΩ , RK ) ∈ Ri and choose UK as in Theorem 25. Fix ε > 0 and apply Lemma 34 to obtain encoding
functions fK and decoding functions gA,Ae. For any pair (A, B) ∈ Ω, find the nonempty subsets Ae ⊆ A ⊆ A and
Be ⊆ B ⊆ B such that (41)–(43) hold. (The case Ae = ∅ or Be = ∅ can be ignored since due to (43) it leads to
µA,B ≤ 0, which is achieved by any code.) Define the functions h1 := gA,Ae ◦ fA and h2 := gB,Be ◦ fB . To analyze
S
Θ(fA ; fB ), we define D1 := h1 (XAn ) and partition XAn as XAn = u e∈D1 h−1
e). We may assume without loss
1 (uA
A
−1
n
n
of generality that h1 (uAe) ⊆ T[XA |U e] (uAe) whenever uAe ∈ T[U e] as this does not interfere with the properties
A
A
(54) and (55) of the code. Defining D2 accordingly, we set F := (D1 × D2 ) ∩ T[Un eU e] . Let us use the shorthand
A B
notation Û1 := h1 (XA ) and Û2 := h2 (XB ), and define pF := P (Û1 , Û2 ) ∈ F and p̄F := P (U1 , U2 ) ∈ F
with U1 := h1 (XA ), U2 := h2 (XB ), where (XA , XB ) are n i.i.d. copies of (XA , XB ) ∼ pXA pXB . We then have
nΘ(fA ; fB ) ≥ nΘ(h1 ; h2 ) = I(h1 (XA ); h2 (XB ))
n
o
n
o
P Û1 = uAe, Û2 = uBe
X
o n
o
P Û1 = uAe, Û2 = uBe log n
=
P
Û
=
u
P
Û
=
u
1
2
e
uAe ∈D1 ,uBe∈D2
A
Be
n
o
n
o
P Û1 = uAe, Û2 = uBe
X
o n
o
=
P Û1 = uAe, Û2 = uBe log n
P
Û
=
u
P
Û
=
u
1
2
e
e
(uAe ,uBe)∈F
A
B
n
o
n
o
P Û1 = uAe, Û2 = uBe
X
o n
o
+
P Û1 = uAe, Û2 = uBe log n
Û
=
u
P
Û
=
u
P
c
1
2
e
(uAe ,uBe)∈F
A
Be
≥ pF log
pF
1−pF
+ (1−pF ) log
,
p̄F
1− p̄F
(57)
(58)
(59)
(60)
where (57) follows from the data processing inequality [23, Theorem 2.8.1] and (60) is a consequence of the log-sum
inequality [23, Theorem 2.7.1]. Expression (60) can be further bounded by
pF log
pF
1−pF
+ (1−pF ) log
= −h2 (pF ) − pF log p̄F − (1−pF ) log(1− p̄F )
p̄F
1− p̄F
≥ −h2 (pF ) − pF log p̄F
(55)
≥ − log 2 − (1−ε) log p̄F .
(61)
(62)
(63)
16
For each uAe ∈ D1 and uBe ∈ D2 , let us define
and
−1
S(uAe, uBe) := {uAe} × h−1
e) × h2 (uBe) × {uBe}
1 (uA
S :=
[
(uAe ,uBe)∈F
(64)
(65)
S(uAe, uBe).
Now, pick any (ûAe, x̂A , x̂B , ûBe) ∈ S . Let ÛAe, X̂A , X̂B , and ÛBe be the type variables corresponding to ûAe, x̂A ,
x̂B , and ûBe, respectively. From part 1 of Lemma 54 we know
P XA = x̂A , XB = x̂B = exp − n H(X̂A X̂B ) + D(X̂A X̂B kXA XB ) .
(66)
Let κ(uAe, uBe; ÛAe, X̂A , X̂B , ÛBe) denote the number of elements in S(uAe, uBe) that have type (ÛAe, X̂A , X̂B , ÛBe).
Then, by part 2 of Lemma 54
κ(uAe, uBe; ÛAe, X̂A , X̂B , ÛBe) ≤ exp nH(X̂A X̂B |ÛAeÛBe) .
(67)
Letting κ(ÛAe, X̂A , X̂B , ÛBe) be the number of elements of S with type (ÛAe, X̂A , X̂B , ÛBe), we have
X
κ(ÛAe, X̂A , X̂B , ÛBe) =
κ(uAe, uBe; ÛAe, X̂A , X̂B , ÛBe)
(68)
(uAe ,uBe)∈F
X
(67)
≤
(uAe ,uBe)∈F
exp nH(X̂A X̂B |ÛAeÛBe)
(69)
≤ exp n I UAeUBe; XA XB + H(X̂A X̂B |ÛAeÛBe) + ε .
(54)
Thus,
P (UAe, UBe) ∈ F
(66)
=
X
ÛAe ,X̂A ,X̂B ,ÛBe
(70)
≤
(70)
κ(ÛAe, X̂A , X̂B , ÛBe)
· exp −n H(X̂A X̂B ) + D(X̂A X̂B kXA XB )
X
exp −n k(ÛAe, X̂A , X̂B , ÛBe) − ε ,
(71)
(72)
ÛAe ,X̂A ,X̂B ,ÛBe
where the sum is over all types that occur in S and
k(ÛAe, X̂A , X̂B , ÛBe) := I(ÛAeÛBe; X̂A X̂B ) − I UAeUBe; XA XB
+ D(X̂A X̂B kXA XB ).
Using a type counting argument (Lemma 53), we can further bound
P (UAe, UBe) ∈ F ≤ (n + 1)|UAe||XA ||XB ||UBe|
·
max
exp −n k(ÛAe, X̂A , X̂B , ÛBe) − ε ,
(73)
(74)
ÛAe ,X̂A ,X̂B ,ÛBe
where the maximum is over all types occurring in S . For any type (ÛAe, X̂A , X̂B , ÛBe) in S , we have by construction
(ÛAe, X̂A , X̂B , ÛBe) ∈ Sδ (UAe, XA , XB , UBe) (following the δ -convention, Remark 12) and we thus conclude
P (UAe, UBe) ∈ F ≤ (n + 1)|UK ||XK |
e e, X
eA, X
eB, U
e e) − ε .
·
max
exp −n k(U
(75)
A
B
e e ,X
e A ,X
e B ,U
e e)∈Sδ (U e ,XA ,XB ,U e)
(U
A
B
A
B
17
Combining (63) and (75) we have shown that for n large enough
log 2 1 − ε
Θ(fA ; fB ) ≥ −
−
log P (UAe, UBe) ∈ F
n
n
e e, X
eA, X
eB, U
e e)
≥ −2ε + (1 − ε) · min k(U
A
B
e
e
e
e
≥ min k(U e, XA , XB , U e) − 2 + I(XA ; XB ) ε,
A
B
(76)
(77)
(78)
e e, X
eA, X
eB, U
e e) ∈ Sδ (U e, XA , XB , U e). As k(U
e e, X
eA, X
eB, U
e e) is continuous as a
where the minimum is over all (U
A
B
A
B
A
B
function of pUe e,Xe A ,Xe B ,Ue e , by letting n → ∞,
A
B
Θ(fA ; fB ) ≥
min
e e ,X
e A ,X
e B ,U
e e)∈S0 (U e ,XA ,XB ,U e)
(U
A
B
A
B
e e, X
eA, X
eB, U
e e) − Cε
k(U
A
B
e e, X
eA, X
eB, U
e e) ∈ S0 (U e, XA , XB , U e) we have
for some fixed constant C . Observe that for (U
A
B
A
B
e e, X
eA, X
eB, U
e e) = I(U
e eX
e e e
k(U
B
A A ; XB UBe)
A
e e; U
e e) = I(U e; U e).
≥ I(U
Combining (79) and (81), we have
A
B
A
(80)
(81)
B
(43)
Θ(fA ; fB ) ≥ I UAe; UBe − Cε ≥ µA,B − Cε.
(79)
(82)
We hence obtain (µΩ − Cε, RK + ε) ∈ R; since ε was arbitrary, this completes the proof.
D. Multiple Description CEO Problem
We continue the discussion of the CEO problem of Section III-B and assume L = 1 as well as Xj −◦− Y :=
Y1 −◦− XJ \j for all j ∈ J . To simplify notation we will again use νA := νA,1 . Extending the setup discussed in
Section III-B, we will allow νj > 0 for any j ∈ J . Loosely speaking, this requires a multiple description (MD)
code for the CEO problem, enabling the CEO to obtain valuable information from the message of the j th agent
alone. Surprisingly, this extension also permits a single-letter characterization. In particular for the case J = 2 this
allows us to give a full single-letter characterization of the achievable region, which will be explicitly stated in
Corollary 38.
(<,I)
Definition 36. For a total order2 < on J and a set I ⊆ J , let the region RMI
such that there exist random variables (UJ , ∅) ∈ P∗ with
be the set of tuples (νΠ , RJ )
Rj ≥ I(Uj ; Xj |U=j ),
j ∈ J,
(83)
νj ≤ I(Uj ; Y|U=j ),
j∈
/ I,
(85)
j ∈ I,
(84)
νj ≤ I(Uj ; Y),
j ∈ I,
(86)
νA ≤ 0,
1 < |A| < |J |.
(88)
Rj ≥ I(Uj ; Xj ),
νK ≤ I(UJ ; Y),
(87)
Remark 9. The purpose of the order < is to determine the order of the messages for successive decoding. Equivalently,
Definition 36 could be rephrased in terms of a permutation of J in place of a total order.
We are now able to state the single-letter characterization of RMI with the additional condition that (88) holds.
S
(<,I)
Theorem 37. We have {(νΠ , RJ ) ∈ RMI : (88) holds} = conv <,I RMI
, where the union is over all total
orders < on J and all sets I ⊆ J .
The proof of Theorem 37 is provided at the end of this section. In particular, Theorem 37 provides a single-letter
characterization of RMI for the case of J = 2 agents. We state this special case separately in the following corollary
to showcase some interesting features of this single-letter region.
2
For the notation regarding total orders refer to Section I-C1.
18
(1)
(2)
(3)
(i)
Corollary 38. For J = 2, we have RMI = conv RMI ∪ RMI ∪ RMI , where (νΠ , RJ ) ∈ RMI if, for some
(UJ , ∅) ∈ P∗ , the following inequalities are satisfied
(1)
(2)
RMI :
(3)
RMI :
ν1 ≤ I(Y; U1 )
RMI :
ν1 ≤ I(Y; U1 |U2 )
ν2 ≤ I(Y; U2 |U1 )
ν2 ≤ I(Y; U2 )
ν{1,2} ≤ I(Y; U1 U2 )
ν{1,2} ≤ I(Y; U1 U2 )
R1 ≥ I(U1 ; X1 )
R1 ≥ I(U1 ; X1 |U2 )
R2 ≥ I(U2 ; X2 |U1 )
ν1 ≤ I(Y; U1 )
(89)
ν{1,2} ≤ I(Y; U1 U2 )
(91)
R2 ≥ I(U2 ; X2 )
(<,I)
(2)
ν2 ≤ I(Y; U2 )
(90)
R1 ≥ I(U1 ; X1 )
(92)
R2 ≥ I(U2 ; X2 ).
(<,I)
Proof. Assuming 1 < 2, we obtain RMI = RMI if 1 ∈
/ I and otherwise RMI
(<,I)
(1)
(<,I)
(3)
2 < 1, we obtain RMI = RMI if 2 ∈
/ I and otherwise also RMI = RMI .
(93)
(3)
= RMI . On the other hand, if
(1)
Remark 10. Note that the total available rate of encoder 2 is R2 = I(X2 ; U2 |U1 ) to achieve a point in RMI .
Interestingly, this rate is in general less than the rate required to ensure successful typicality decoding of U2 .
However, ν2 = I(Y; U2 |U1 ) can still be achieved.
(i)
Remark 11. On the other hand, fixing the random variables U1 , U2 in the definition of RMI shows another interesting
feature of this region. The achievable values for ν1 and ν2 vary across i ∈ {1, 2, 3} and hence do not only depend
on the chosen random variables U1 and U2 , but also on the specific rates R1 and R2 .
It is worth mentioning that by setting ν1 = ν2 = 0 the region RMI reduces to the rate region in Lemma 33.
(<,I)
The following proposition shows that RMI is computable, at least in principle. The given cardinality bound is
(<,I)
not optimal, but it implies RMI
(<,I)
= RMI . The proof of Proposition 39 is provided in Appendix B5.
(<,I)
Proposition 39. The region RMI
every j ∈ J .
remains unchanged if the cardinality bound |Uj | ≤ |Xj | + 4J is imposed for
The following two theorems provide an inner and an outer bound for RMI . In order to show that Theorem 37
holds, we subsequently prove that these bounds are indeed tight, assuming that (88) holds.
(<,I)
Theorem 40. We have RMI
⊆ RMI for any I ⊆ J and any total order < on J .
Theorem 41. If (νΠ , RJ ) ∈ RMI then
X
Rj − νA ≥ I(XB ; UB |YQ) − I Y; UA\B Q
(94)
j∈B
for all A, B ⊆ J and some random variables (UJ , Q) ∈ P∗ .
The proof of Theorems 40 and 41 are given in Appendices B6 and B7, respectively. We will, however, only require
the following simple corollary of Theorem 41.
Corollary 42. For any (νΠ , RJ ) ∈ RMI there are random variables (UJ , Q) ∈ P∗ with
X
j∈A
Rj ≥ 0,
Rj − νJ ≥ I(XA ; UA |YQ) − I Y; UJ \A
Rj − νj ≥ I(Xj ; Uj |YQ),
νj ≤ I(Y; Uj |Q),
Q ,
for all j ∈ J ,
(95)
for all A ⊆ J ,
(96)
for all j ∈ J ,
(97)
for all j ∈ J .
(98)
Proof of Theorem 37. We will make use of some rather technical results on convex polyhedra, derived in
Appendix B9.
Assume (νΠ , RJ ) ∈ RMI . We can then find (UJ , Q) ∈ P∗ such that (95)–(98) hold. We define νeΠ := −νΠ to
simplify notation. It is straightforward to check that the inequalities (95)–(98) define a sequence of closed convex
polyhedra H(j) in the variables (RJ , νeΠ ) that satisfy Assumptions 1 and 2 of Lemma 51. H(0) is defined by (95)
19
and (96) alone, and for k ∈ [0 : J] the polyhedron H(k) is given in the K + k variables (RJ , νeJ , νe1 , νe2 , . . . , νek ) by
adding constraints (97) and (98) for each j ∈ [1 : k]. The set H(0) is a supermodular polyhedron [52, Section 2.3] in
the K variables (RJ , νeJ ) on (K, 2K ) with rank function
(
0,
K∈
/ A,
ϑ(A) =
(99)
I XA\K ; UA\K YQ − I Y; UJ \A Q , K ∈ A,
where supermodularity follows via standard information-theoretic arguments. By the extreme point theorem [52,
Theorem 3.22], every extreme point of H(0) is associated with a total order < on K. Such an extreme point is given
by
(<)
= 0 for j < K,
(100)
= I(Y; U=K |Q) − I(Y; U<K |YQ).
(102)
≤ I(Xj ; Uj |YQ) + I(Y; Uj |Q) = I(Xj ; Uj |Q).
(103)
Rj
(<)
Rj
(<)
νJ
= I(Uj ; Xj |U=j Q) for j = K,
(101)
Assumption 3 of Lemma 51 is now verified by
(<)
Rj
By applying Lemma 51 we find that every extreme point of H(J) is given by a subset I ⊆ J and an order < of K
as
(<,I)
Rj
(<,I)
Rj
(<,I)
Rj
(<,I)
νj
(<,I)
νj
(<,I)
νj
(<,I)
νJ
= I(Xj ; Uj |Q),
(104)
j ∈ I,
= 0,
= I(Uj ; Xj |U=j Q),
= I(Uj ; Y|Q),
= −I(Uj ; Xj |YQ),
= I(Uj ; Y|U=j Q),
j∈
/ I and j < K,
(105)
j ∈ I,
(107)
j∈
/ I and j = K,
(109)
j∈
/ I and j = K,
(106)
j∈
/ I and j < K,
(108)
(110)
= I(Y; U=K |Q) − I(Y; U<K |YQ).
(<,I,q)
(<,I,q)
For each q ∈ Q with P{Q = q} > 0 let the point (νΠ
, RJ
) be defined by (104)–(110), but given
(<,I,q)
(<,I,q)
(<,I)
{Q = q}. By substituting
/ I and j < K , we see that (νΩ
, RJ
) ∈ RMI and consequently
Uj →∅ if j ∈
(<,I)
(<,I)
(<,I)
(νΠ , RJ ) ∈ conv RMI
. Defining the orthant O := {(νΠ , RJ ) : νΠ ≤ 0, RJ ≥ 0}, this implies
[
[
(<,I)
(<,I)
(νK , RJ ) ∈ conv conv RMI
+ O = conv RMI + conv(O)
(111)
<,I
<,I
= conv
[
<,I
= conv
[
<,I
(<,I)
RMI
+ O
(112)
(<,I)
RMI ,
(<,I)
where (112) follows from [53, Theorem 1.1.2] and in (113) we used that RMI
Together with Theorem 40 and the convexity of RMI (Remark 7) we obtain
[ (<,I)
RMI ⊆ conv RMI ⊆ RMI .
<,I
(113)
(<,I)
+ O = RMI
by definition.
(114)
20
S
(<,I)
(<,I)
It remains to show that conv <,I RMI
is closed. Using Proposition 39, we can write RMI = F (<,I) (P∗0 ) +
O, where P∗0 := {pY,XJ ,UJ : (UJ , ∅) ∈ P∗ , |Uj | = |Xj | + 4J , j ∈ J } is a compact subset of the probability simplex
(<,I)
and F (<,I) is a continuous function, given by the definition of RMI , (83)–(88). We can thus write
[ (<,I)
[
conv RMI = conv F (<,I) (P∗0 ) + O,
(115)
which is closed by Lemma 44.
<,I
<,I
IV. S UMMARY AND D ISCUSSION
We introduced a novel multi-terminal source coding problem termed information-theoretic biclustering. Interestingly,
this problem is related to several other problems at the frontier of statistics and information theory and offers
a formidable mathematical complexity. Indeed, it is fundamentally different from “classical” distributed source
coding problems where the encoders usually aim at reducing, as much as possible, redundant information among
the sources while still satisfying a fidelity criterion. In the considered problem, however, the encoders are interested
in maximizing precisely such redundant information.
While an exact characterization of the achievable region is mathematically very challenging and still remains
elusive, we provided outer and inner bounds to the set of achievable rates. We thoroughly studied the special case of
two symmetric binary sources for which novel cardinality bounding techniques were developed. Based on numerical
evidence we formulated a conjecture that entails an explicit expression for the inner bound. This conjecture provides
strong evidence that our inner and outer bound do not meet in general. We firmly believe that an improved outer
bound, satisfying the adequate Markov chains, is required for a tight characterization of the achievable region.
We further established analogous bounds to the achievable rate region of information-theoretic biclustering with
more than two sources. However, these bounds cannot be tight since the famous Körner-Marton problem constitutes
a counterexample. For an analogue of the well-known CEO problem we showed that our bounds are tight in a
special case, leveraging existing results from multiterminal lossy source coding. Furthermore we considered a
multiple description CEO problem which surprisingly also permits a single-letter characterization of the achievable
region. The resulting region has the remarkable feature that it allows to exploit rate that is in general insufficient to
guarantee successful typicality encoding.
The interesting challenge of the biclustering problem lies in the fact that one needs to bound the mutual information
between two arbitrary encodings solely based on their rates. Standard information-theoretic manipulations seem
incapable of handling this requirement well.
ACKNOWLEDGMENT
The authors would like to thank Shlomo Shamai (Shitz) and Emre Telatar for insightful discussions regarding
the binary example. We would also like to thank Christian Chapman from the School of Electrical, Computer and
Energy Engineering at Arizona State University for providing the counterexample to [16, Conjecture 1, Eq. (14)].
A PPENDIX
A. Proofs from Section II
We start this section with two simple lemmas, which will be used in several proofs.
Lemma 43. If U −◦− X −◦− Z and X −◦− Z −◦− V, then
I(U; X) + I(V; Z) − I(UV; XZ) = I(U; V) − I(U; V|XZ),
(116)
I(U; X) + I(V; Z) − I(UV; XZ) ≤ I(V; X).
(118)
I(U; X) + I(V; Z) − I(UV; XZ) ≤ I(U; Z),
(117)
Provided that U −◦− X −◦− Z −◦− V then I(U; X) + I(V; Z) − I(UV; XZ) = I(U; V).
Proof. By standard information-theoretic manipulations we obtain (116) as
I(U; X) + I(V; Z) − I(UV; XZ) = I(U; XZ) + I(V; XZ) − I(UV; XZ)
(119)
21
= I(U; V) − I(U; V|XZ).
(120)
We obtain (117), and by a symmetric argument also (118), as
I(U; X) + I(V; Z) − I(UV; XZ) = I(V; Z) − I(V; XZ|U)
= I(U; Z) + I(V; Z) − I(U; Z) − I(V; XZ|U)
= I(U; Z) + I(V; Z) − I(U; Z) − I(V; Z|U) − I(V; X|ZU)
= I(U; Z) + I(V; Z) − I(UV; Z) − I(V; X|ZU)
= I(U; Z) − I(U; Z|V) − I(V; X|ZU)
≤ I(U; Z).
(121)
(122)
(123)
(124)
(125)
(126)
Lemma 44. For a compact set C ⊆ Rn and a closed, convex set B ⊆ Rn ,
A := conv(C + B) = conv(C) + B = A.
(127)
Proof. We have A = conv(C) + conv(B) = conv(C) + B by [53, Theorem 1.1.2] and the convexity of B . Note that
conv(C) is compact by [54, Theorem 2.3.4]. A is the sum of a compact set and a closed set and, hence, closed [55,
Exercise 1.3(e)].
1) Proof of Theorem 9: For (µ, R1 , R2 ) ∈ R, let (f, g) be an (n, R1 , R2 )-code such that Θ(f ; g) ≥ µ. Defining
the random variables Ui := Xi−1 , f (X) and Vi := Zi−1 , g(Z) for i ∈ [1 : n] we obtain
nR1 ≥ H f (X) = I f (X); X
(128)
n
X
=
I f (X); Xi Xi−1
(129)
i=1
=
n
X
I(Ui ; Xi ),
(130)
i=1
where (130) holds because X are i.i.d. Analogously, we have
nR2 ≥
n
X
I(Vi ; Zi ).
(131)
i=1
From Lemma 43 and the Markov chain f (X) −◦− X −◦− Z −◦− g(Z), we obtain
nµ ≤ I f (X); g(Z)
= I f (X); X + I g(Z); Z − I f (X)g(Z); XZ
n h
i
X
=
I(Ui ; Xi ) + I(Vi ; Zi ) − I(Ui Vi ; Xi Zi ) .
(132)
(133)
(134)
i=1
Now a standard time-sharing argument shows R ⊆ Ro . Lemma 43 provides Ro ⊆ R0o .
2) Proof of Proposition 10: Most steps in the proof apply to both Ro and R0o . We thus state the proof for Ro
and point out the required modifications where appropriate.
For fixed random variables (X, Z) define the set of pmfs (with finite, but arbitrarily large support)
Q := {pU,X,Z,V : U −◦− X −◦− Z, X −◦− Z −◦− V},
(135)
and the compact set of pmfs with fixed alphabet size
Q(a, b) := {pU,X,Z,V ∈ Q : |U| = a, |V| = b}.
(136)
Define the continuous vector valued function F := (F1 , F2 , F3 ) as
F1 (pU,X,Z,V ) := I(X; U) + I(Z; V) − I(UV; XZ),
(137)
22
(138)
F2 (pU,X,Z,V ) := I(U; X),
F3 (pU,X,Z,V ) := I(V; Z).
(139)
In the proof of R0o =conv(So0 ) let F1 (pU,X,Z,V ) := min{I(U; Z), I(V; X)}. We can now write Ro = F (Q) + O and
So = F Q(|X |, |Z|) + O where O := (R− × R+ × R+ ). Since Ro is convex, we may define the extended real
function ψ(λ) := inf x∈Ro λ · x and obtain [54, Theorem 2.2, 3.]
o
\ n
conv(Ro ) = Ro =
x ∈ R3 : x · λ ≥ ψ(λ) .
(140)
λ∈R3
From the definition of Ro , we have ψ(λ) = −∞ if λ ∈
/ O, and ψ(λ) = inf p∈Q λ · F (p) otherwise. This shows,
that
o
\ n
Ro =
x ∈ R3 : x · λ ≥ ψ(λ)
(141)
λ∈O
and using the same argument, one can also show that
o
\ n
e
conv(So ) =
x ∈ R3 : x · λ ≥ ψ(λ)
,
e
ψ(λ)
=
λ∈O
min
p∈Q(|X |,|Z|)
λ · F (p).
(142)
e
We next show that ψ(λ) = ψ(λ)
for λ ∈ O. For arbitrary λ ∈ O and δ > 0, we can find random variables
e X, Z, V)
e ∼ e
(U,
p ∈ Q with λ · F (e
p ) ≤ ψ(λ) + δ . By compactness of Q(a, b) and continuity of F , there is
e
e
p ∈ Q(|U|, |V|) with
λ · F (p) =
min
e
p̂∈Q(|Ue|,|V|)
λ · F (p̂) ≤ λ · F (e
p ) ≤ ψ(λ) + δ.
(143)
We now show that there exists p̂ ∈ Q(|X |, |Z|) with
λ · F (p̂) = λ · F (p).
(144)
As a consequence of the inequalities F1 ≤ F2 and F1 ≤ F3 we have λ · F (p) = 0 if λ1 + max{λ2 , λ3 } ≥ 0. Thus,
we only need to show (144) for λ ∈ O with λ1 + λ2 < 0 and λ1 + λ3 < 0. To this end we use the perturbation
method [56], [57] and perturb p , obtaining the candidate
(Û, X, Z, V̂) ∼ p̂(u, x, z, v) = p(u, x, z, v) 1 + εφ(u) .
(145)
We require
1 + εφ(u) ≥ 0,
E [φ(U)] = 0,
E[φ(U)|X = x, Z = z] = 0,
for every u ∈ U,
(146)
if p(x, z) > 0.
(148)
(147)
The conditions (146) and (147) ensure that p̂ is a valid pmf and (148) implies p̂ ∈ Q. Observe that there is an
ε0 > 0 for any φ, such that (146) is satisfied for ε ∈ [−ε0 , ε0 ]. Furthermore, (148) is equivalent to
E[φ(U)|X = x] = 0,
for every x ∈ X
(149)
due to the Markov chain U −◦− X −◦− Z. Note also that (149) already implies (147). If |U| ≥ |X | + 1 there is a
non-trivial solution to (149), which means there exists φ 6≡ 0 such that (146)–(148) are satisfied. We have
h
λ · F (p̂) = λ1 I(X; U) − I(UV; XZ) + H(Z) + εHφ (U) − εHφ (UX)
i
− εHφ (UV) + εHφ (UXZV) + H(V̂) − H(ZV̂)
+ λ2 [I(X; U) + εHφ (U) − εHφ (UX)]
+ λ3 [H(Z) + H(V̂) − H(ZV̂)].
(150)
P
Here, we used the shorthand Hφ (UX) := − u,x p(u, x)φ(u) log p(u, x) and analogous for other combinations of
23
random variables. By (143), we have
∂2
∂ε2 λ
· F (p̂)
ε=0
≥ 0. Observe that
∂
∂ X
p̂(z, v)
H(V̂) − H(ZV̂) =
p̂(z, v) log
∂ε
∂ε z,v
p̂(v)
=
X ∂p̂(z, v)
∂ε
z,v
and consequently,
log
∂p̂(v)
p̂(z, v) ∂p̂(z, v) p̂(z, v) ∂ε
+
−
p̂(v)
∂ε
p̂(v)
∂2
∂2
λ
·
F
(p̂)
=
(λ
+
λ
)
(H(V̂) − H(ZV̂))
1
3
∂ε2
∂ε2
X ∂p̂(z, v) 2 1
= (λ1 + λ3 )
∂ε
p̂(z, v)
z,v
∂p̂(v) 2 p̂(z, v)
∂p̂(z, v) ∂p̂(v) 1
.
−2
+
∂ε
∂ε p̂(v)
∂ε
p̂(v)2
Here we already used that
∂ 2 p̂(v)
∂ε2
≡
∂ 2 p̂(z,v)
∂ε2
(151)
(152)
(153)
(154)
≡ 0. It is straightforward to calculate
∂p̂(v)
= p(v)E[φ(U)|V = v],
∂ε
∂p̂(z, v)
= p(z, v)E[φ(U)|V = v, Z = z],
∂ε
p̂(z, v)|ε=0 = p(z, v),
(155)
(156)
(157)
(158)
p̂(v)|ε=0 = p(v),
and thus, taking into account that λ1 + λ3 < 0,
X
2
0≥
p(z, v) E[φ(U)|V = v, Z = z] − E[φ(U)|V = v] ,
(159)
z,v
which implies for any (z, v) ∈ Z × V with p(z, v) > 0,
X
X
p(u|z, v)φ(u) =
p(u|v)φ(u).
u
(160)
u
From (160) we can conclude
X
p̂(z, v)
p̂(v)
z,v
P
X
û p(û, z, v) 1 + εφ(û)
=
p(u, z, v)(1 + εφ(u)) log P
p(û,
v)
1
+
εφ(û)
û
z,v,u
P
X
p(z, v) 1 + ε û p(û|z, v)φ(û)
P
=
p(u, z, v)(1 + εφ(u)) log
p(v)
1
+
ε
p(û|v)φ(û)
û
z,v,u
p(z, v)
(160) X
=
p(u, z, v)(1 + εφ(u)) log
p(v)
z,v,u
X
X
p(z, v)
p(z, v)
=
p(z, v) log
+ε
φ(u)p(u, z, v) log
p(v)
p(v)
z,v
z,v,u
H(V̂) − H(ZV̂) =
p̂(z, v) log
= H(V) − H(ZV) + εHφ (V) − εHφ (ZV),
where we used
Hφ (V) := −
X
u,v
p(u, v)φ(u) log p(v),
(161)
(162)
(163)
(164)
(165)
(166)
(167)
24
Hφ (ZV) := −
X
p(u, z, v)φ(u) log p(z, v).
(168)
u,z,v
Substituting in (150) shows that λ · F (p̂) is linear in ε. And by the optimality of p it must be constant. We may now
choose ε maximal, i. e., such that there is at least one u ∈ U with p(u)(1 + εφ(u)) = 0. This effectively reduces
the cardinality of Û by at least one and may be repeated until |Û| = |X | (as then φ ≡ 0). The same process can be
carried out for V and yields p̂ ∈ Q(|X |, |Z|) such that (144) holds.
In the proof of R0o = conv(So0 ), we show (144) by applying the support lemma [13, Appendix C] with |X | − 1
test functions tx (pX̂ ) := pX̂ (x) (x ∈ X \ x0 ) and with the function
f (pX̂ ) := λ1 min{I(V; X), H(Z) − H(Ẑ)} + λ2 H(X) − H(X̂) + λ3 I(Z; V),
(169)
e and
where (Ẑ, X̂) ∼ pX̂ pZ|X . Consequently there exists a random variable Û with (Û, X, Z, V) ∼ p̂ ∈ Q(|X |, |V|)
λ · F (p̂) = λ · F (p). By applying the same argument to V, we obtain p̂ ∈ Q(|X |, |Z|) such that (144) holds.
By combining (143) and (144) we obtain
λ · F (p̂) = λ · F (p) ≤ ψ(λ) + δ.
(170)
e , which implies Ro = conv(So ) using
As δ > 0 was arbitrary and Q(|X |, |Z|) is compact, we proved ψ(λ) = ψ(λ)
(141) and (142). We find Ro = conv(So ) by writing
Ro = conv(F (Q(|X |, |Z|)) + O)
(171)
= conv(F (Q(|X |, |Z|))) + O
(172)
= Ro ,
(174)
⊆ F (Q) + O
(173)
where (172) follows from Lemma 44. The relation (173) is a consequence of Q(|X |, |Z|) ⊆ Q and the convexity of
F (Q).
3) Proof of Proposition 13: We only need to show conv(Si ) = conv(Ri ) as the cardinality bound |Q| ≤ 3
follows directly from the strengthened Carathéodory theorem [58, Theorem 18(ii)] because conv(Ri ) is the convex
hull of a connected set in R3 . We will only show the cardinality bound |U| ≤ |X |. The corresponding bound for
|V| follows analogously. We note that the weaker bounds |U| ≤ |X | + 1 and |V| ≤ |Z| + 1 can be shown directly
using the convex cover method [13, Appendix C], [32], [59]. Define the continuous vector-valued function
e e
e e
e e
F (pU,
(175)
e X,
e Z,
eV
e ) := I(U; V), I(X; U), I(Z; V) ,
and the compact, connected sets of pmfs
n
o
e
e
Q := pU,
,
e X,
e Z,
eV
e : pU,
e X,
e Z,
eV
e = pU|X
e pX,Z pV|Z
e , U = 0 : |X | , V = 0 : |Z|
n
o
e := p e e e e ∈ Q : Ue = 1 : |X | .
Q
U,X,Z,V
To complete the proof of the proposition, it suffices to show
e ,
conv F (Q) ⊆ conv F (Q)
(176)
(177)
(178)
since we then have with O := (R− × R+ × R+ ),
conv(Ri ) = conv F (Q) + O
= conv F (Q) + O
(178)
e +O
⊆ conv F (Q)
e +O
= conv F (Q)
= conv(Si ),
(179)
(180)
(181)
(182)
(183)
3
where (180) and (182) follow
from Lemma 44. The region F (Q) ⊆ R is compact [22, Theorem 4.22]. Therefore,
its convex hull conv F (Q) is compact [60, Corollary 5.33] and can be represented as an intersection of halfspaces
25
in the following manner [54, Proposition 2.2, 3.]: Defining V (λ) := maxx∈F (Q) λ · x for λ = (λ1 , λ2 , λ3 ) ∈ R3 ,
we have
o
\ n
conv F (Q) =
x ∈ R3 : λ · x ≤ V (λ) .
(184)
λ∈R3
With the same reasoning we obtain
o
\ n
e =
conv F (Q)
x ∈ R3 : λ · x ≤ Ve (λ) ,
λ∈
(185)
R3
e
where Ve (λ) := maxx∈F (Q)
e λ · x. We next show V (λ) ≥ V (λ) which already implies (178) due to (184) and (185).
Let t = (tx )x∈X \x0 be |X | − 1 test functions tx (pXe ) := pXe (x) for all but one x ∈ X . Choose any λ ∈ R3 and
fix (U, X, Z, V) ∼ p ∈ Q that achieve λ · F (p) = V (λ). Define the continuous function
e + λ2 (H(X) − H(X))
e + λ3 I(Z; V)
f (pXe ) := λ1 (H(V) − H(V))
(186)
e Z,
e X)
e ∼ pV|Z pZ|X p e . The point (pX (x))x∈X \x , V (λ) lies in the convex hull of the compact [61,
where (V,
0
X
Theorem 26.5] and connected [22, Theorem 4.22] set (t, f ) Q . Therefore, by the strengthened Carathéodory
theorem [58, Theorem 18(ii)],
i. e., there exists a random variable U0 with |U 0 | = |X | and thus
|X | points 0suffice,
e, such that EU0 f (pX|U0 ( · |U )) = λ · F (pU0 ,X,Z,V ) = V (λ). This shows Ve (λ) ≥ V (λ).
pU0 ,X,Z,V ∈ Q
By applying the same reasoning to V, one can show that |V| = |Z| also suffices.
e = X ⊕ N1 and V
e = Z ⊕ N2 , where N1 , N2 ∼ B(α) are independent of (X, Z)
4) Proof of Proposition 17: With U
and of each other, it follows that (µ, R, R) := log 2 − h2 (α ∗ p), log 2 − h2 (α), log 2 − h2 (α) ∈ R0o for α ∈ (0, 12 ).
Assume (µ, R, R) ∈ Si0 and choose U, V, and Q according to Proposition 13. We then have
H(X|UQ) ≥ h2 (α),
(187)
I(U; V|Q) ≥ log 2 − h2 (α ∗ p).
(189)
H(Z|VQ) ≥ h2 (α),
(188)
Using Mrs. Gerber’s Lemma (MGL) [35, Theorem 1], we obtain
(188)
H(X|VQ) ≥ h2 h−1
H(Z|VQ)
∗ p ≥ h2 (α ∗ p).
2
(190)
Thus, I(X; V|Q) ≤ log 2 − h2 (α ∗ p) and furthermore I(X; V|Q) ≥ I(U; V|Q) due to U −◦− (X, Q) −◦− V. These two
inequalities in combination with (189) imply I(X; V|Q) = I(U; V|Q), which amounts to X −◦− (U, Q) −◦− V. We can
therefore write the joint pmf of (U, X, V, Q) in two ways, as
pU,X,V,Q (u, x, v, q) = pX (x)pQ (q)pU|X,Q (u|x, q)pV|X,Q (v|x, q)
(191)
= pX (x)pQ (q)pU|X,Q (u|x, q)pV|U,Q (v|u, q).
(192)
Assume without loss of generality that pQ (q) > 0 for all q ∈ Q. If pU|X,Q (u|x, q) > 0 then (192) necessitates
pV|U,Q (v|u, q) = pV|X,Q (v|x, q)
for v ∈ {0, 1}. Next, we partition Q into three disjoint subsets
Q1 := q ∈ Q : P{U = X|Q = q} = 1 or P{U = X ⊕ 1|Q = q} = 1 ,
Q2 := q ∈ Q : P{U = 0|Q = q} = 1 or P{U = 1|Q = q} = 1 ,
Q3 := q ∈ Q : pU|X,Q (0|x, q) > 0 and pU|X,Q (1|x, q) > 0 for some x ∈ {0, 1} .
(193)
(194)
(195)
(196)
Given q ∈ Q3 , we apply (193) twice and obtain
pV|U,Q (v|0, q) = pV|X,Q (v|x, q) = pV|U,Q (v|1, q),
(197)
i. e., I(U; V|Q = q) = 0, which is also true for q ∈ Q2 . We can thus write
(189)
log 2 − h2 (α ∗ p) ≤ I(U; V|Q) ≤ P{Q ∈ Q1 } I(X; Z) = P{Q ∈ Q1 } log 2 − h2 (p) .
(198)
26
On the other hand we have
(187)
(194)
h2 (α) ≤ H(X|UQ) ≤ P Q ∈ (Q2 ∪ Q3 ) log 2 = 1 − P{Q ∈ Q1 } log 2.
(199)
Combination of the previous two inequalities leads to
(199)
log 2 − h2 (α ∗ p) (198)
h2 (α)
≤ P{Q ∈ Q1 } ≤ 1 −
,
log 2 − h2 (p)
log 2
which is a contradiction since
log 2−h2 (α∗p)
log 2−h2 (p)
>1−
(200)
h2 (α)
log 2 .
B. Proofs from Section III
1) Proof of Theorem 24: If (µΩ , RK ) ∈ R we obtain an (n, RK )-code fK for
some n ∈ N such that (37) holds.
Define Wk := fk (Xk ) and the auxiliary random variables Uk,i := Wk , Xi−1
for k ∈ K and i ∈ [1 : n]. For any
K
two sets A, C ⊆ K we have
X
n
Rk ≥ H(WA )
(201)
k∈A
(202)
= I(WA ; XK )
≥ I(WA ; XK |WC )
n
X
=
I WA ; XK,i WC Xi−1
K,1
(203)
(204)
i=1
=
n
X
i=1
I(UA,i ; XK,i |UC,i ),
(205)
where (203) follows from WA −◦− XK −◦− WC . Furthermore, for any pair (A, B) ∈ Ω we have by Lemma 43 and
WA −◦− XA −◦− XB −◦− WB ,
nµA,B ≤ I WA ; WB
(206)
= I WA ; XA + I WB ; XB − I WA WB ; XA XB
(207)
n
i
Xh
=
I(UA,i ; XA,i ) + I(UB,i ; XB,i ) − I(UA,i UB,i ; XA,i XB,i ) .
(208)
i=1
Now a standard time-sharing argument shows R ⊆ Ro . Lemma 43 implies Ro ⊆ R0o .
2) Proof of Proposition
26: Pick an
arbitrary k ∈ K. For nonempty A, B ⊆ K with k ∈ B we can write
H(XA |UB ) = EUk fA,B (pXk |Uk ( · |Uk )) where
fA,B (pXk |Uk ( · |uk )) := H XA UB\k , Uk = uk .
(209)
Furthermore, H(UA |UB ) = EUk gA,B (pXk |Uk ( · |Uk )) where
gA,B (pXk |Uk ( · |uk )) := H UA UB\k , Uk = uk .
(210)
Observe that both fA,B and gA,B are continuous functions of pXk |Uk ( · |uk ). Apply the support lemma [13, Appendix C]
with the functions fA,B and gA,B for all nonempty A, B ⊆ K such that k ∈ B , and |Xk | − 1 test functions,
0
which guarantee
that the marginal distribution p0Xk does not change. We obtain a new random variable Uk with
0
H XA UB\k Uk = H(XA |UB ) and H UA UB\k Uk = H(UA |UB ). By rewriting (41)–(43) in terms of conditional
entropies, it is evident that the defining inequalities for Ri remain the same when replacing Uk by U0k . The support
of U0k satisfies the required cardinality bound:3
Uk0 ≤ |Xk | − 1 + 2(2K − 1)2K−1
= |Xk | − 1 + 2
3
2K
There are (2K − 1) ways to choose A and 2K−1 ways to choose B.
−2
K
(211)
(212)
27
≤ |Xk | + 4K .
(213)
The same process is repeated for every k ∈ K.
3) Proof of Lemma 31: Defining U = f (X), we have for every m ∈ M with P{U = m} =
6 0
E dLL g(f (X)), Y U = m = −E[log g(m)(Y)|U = m]
X
=−
pY|U (y|m) log g(m)(y)
(214)
(215)
y∈Y
= H(Y|U = m) + D pY|U ( · |m) g(m)( · )
≥ H(Y|U = m),
(216)
(217)
where (217) follows as KL-divergence is non-negative [23, Theorem 2.6.3]. The final result follows by calculating
the expectation over U. By [23, Theorem 2.6.3], g(m) = pY|U ( · |m) is a necessary and sufficient condition for
equality.
ek = I(Xk ; Uk ) + ε00 /2 for each k ∈ K.
4) Proof of Lemma 34: Fix 0 < ε0 , ε00 < ε and set R
e
fk := enRk and M
fk := [1 : M
fk ]. We apply the generalized Markov lemma
• Encoding: For n ∈ N define M
(k)
(Lemma 59) and consider the random codebooks Ck := (Vm )m∈M
fk , which are drawn independently uniform
n
e k = fek (Xk , Ck ) and the
from T[Uk ] for each k ∈ K. Denote the resulting randomized coding functions as W
(k)
e k := V . If n is chosen large enough we have therefore
corresponding decoded value as U
ek
W
n
e K , XK ) ∈
P (U
/ Tn
[UK XK ]
o
≤ ε0 .
(218)
fk into Mk := en(Rk +ε00 ) equally
Next, we introduce (deterministic) binning. If Rk < I(Xk ; Uk ), partition M
sized, consecutive bins, each of size en∆k with
ek − Rk − ε00 = I(Xk ; Uk ) − Rk −
∆k := R
•
ε00
.
2
(219)
fk → Mk := [1 : Mk ] maps a codeword index to the index of the bin, it
The deterministic function βk : M
belongs to. In total the encoding function becomes fk := βk ◦ fek . If Rk ≥ I(Xk ; Uk ), we do not require binning
fk and hence fk := fek .
and let βk be the identity on M
n for each ∅ 6= A
e ⊆ A ⊆ K is
Decoding: Given the codebooks, the decoding procedure gA,Ae : MA → UA
e
−1
fA be all indices that belong to the bin indices
e := βA
carried out as follows: Given w ∈ MA , let m
(w) ⊆ M
(A)
n
n
w. Consider only the typical sequences Vm
e ∩ T[UA ] =: Φ ⊆ UA . If Φ 6= ∅, choose the lexicographically
e
(A)
smallest element of4 ΦAe, otherwise choose the lexicographically smallest element of Vm
e .
e
e
Let A, A, B, B ⊆ K be sets of indices such that the conditions (41) and (42) are satisfied. Using Wk := fk (Xk , Ck )
and the randomized decodings Û1 := gA,Ae(WA , CA ), Û2 := gB,Be(WB , CB ), consider the error event E0 :=
{(Û1 , XA , XB , Û2 ) ∈
/ T[Un eXA XB U e] }. Define the other events
A
B
e A , X A , XB , U
e B) ∈
E1 := {(U
/ T[Un eXA XB U e] },
B
n
oA
(A)
n
:=
E2
VW ∩ T[UA ]
>1 ,
e
n
A
o
(B)
E3 :=
VM ∩ T[UnB ]
>1 ,
Be
(220)
(221)
(222)
−1
where we used the random sets of indices W := βA
(WA ) and M := βB−1 (WB ). We clearly have E0 ⊆ E1 ∪ E2 ∪ E3
and thus
P{E0 } ≤ P{E1 } + P{E2 |E1c } + P{E3 |E1c }
(218)
≤ P{E2 |E1c } + P{E3 |E1c } + ε0 .
4
e
Note that ΦAe is the restriction of Φ to the indices in A.
(223)
(224)
28
S
−1
We can partition the random set of indices W = βA
(WA ) = A0 ⊆A DA0 into (random) subsets
o
n
e A00 and m
e k , ∀k ∈ A0 ,
e ∈W:m
e A00 = W
e k 6= W
DA0 := m
(225)
e A }. For each set ∅ 6= A0 ⊆ A we define the error event
where we used A00 := A \ A0 . Observe that D∅ = {W
n
o
(A)
EA0 := VDA0 ∩ T[UnA ] 6= ∅
(226)
and obtain
E2 ⊆
which implies
P{E2 |E1c } ≤
[
EA0 ,
(227)
X
P{EA0 |E1c }.
(228)
0
A ⊆A:
e=∅
A0 ∩A6
0
A ⊆A:
e=∅
A0 ∩A6
0
Q
)
e ∈ DA0 , we have that V(A
By construction, DA0 has k∈A0 (en∆k − 1) elements. For m
are uniformly distributed
e
m
Q
n
c
n
e
e
e A00 = WA00 . Given E1 we have in particular UA ∈ T[UA ] . Thus, for any uA00 ∈ T[Un 00 ] , we
on k∈A0 T[Uk ] and m
A
can conclude,
[
n
o
(A)
n
c e 00
e A00 = uA00 = P
00
P EA0 E1c , U
{Vm
∈
T
}
E
,
U
=
u
(229)
A
A
1
[UA ]
e
e ∈DA0
m
o
X n (A)
n
c e 00
00
≤
P Vm
(230)
e ∈ T[UA ] E1 , UA = uA
e ∈DA0
m
≤ exp n
X
0
k∈A
∆k
!!
!!
T[UnA0 |UA00 ] (uA00 )
Q
n
k∈A0 T[Uk ]
exp n(H(UA0 |UA00 ) + ε0 (n))
P
≤ exp n
∆k
exp n( k∈A0 H(Uk ) − εk (n))
k∈A0
!!
X
≤ exp n ε(n) + H(UA0 |UA00 ) +
∆k − H(Uk )
,
X
(231)
(232)
(233)
k∈A0
P
where ε(n) = k∈A0 ∪0 εk (n) goes to zero as n → ∞. Here, (232) follows from parts 2 and 3 of Lemma 58. We
ek and (41) imply for any ∅ 6= A0 ⊆ A with A0 ∩ Ae =
observe that the definition of R
6 ∅ that
X
X
ε00
∆k ≤ − − H(UA0 |UA00 ) +
H(Uk ).
(234)
2
0
0
k∈A
k∈A
e A00 in (233) and use (234) to obtain
Marginalize over U
ε00
P{EA0 |E1c } ≤ exp n ε(n) −
≤ ε0
2
(235)
for n large enough. Applying the same arguments to P{E3 |E1c } and combining (224), (228) and (235), we have
P{E0 } ≤ ε0 + 2|A| ε0 + 2|B| ε0 ≤ 2K ε0 .
(236)
For a set ∅ 6= A ⊆ K, we next analyze the random quantity LA := CA ∩ T[UnA ] . For n large enough, we have
29
e A ∈ CA
for any V
E [LA ] ≤ E
=
"
X
VA ∈CA
Y
k∈A
=
Y
k∈A
≤
≤
Y
k∈A
Y
k∈A
h
E 1T[Un ] (VA ) CA
!
A
h
i
e
f
n
Mk E 1T[U ] (VA )
fk
M
Q
fk
M
!
#
(237)
(238)
A
!
!
i
T[UnA ]
k∈A
en
fk en
M
en
(239)
T[Unk ]
H(UA )+ε0 (n)
P
k∈A
(240)
H(Uk )−εk (n)
H(UA )−
P
k∈A
H(Uk )+ε(n)
!!
ε00
= exp n H(UA ) + ε(n) +
I(Uk ; Xk ) +
− H(Uk )
2
k∈A
!!
ε00 X
= exp n H(UA ) + ε(n) + |A| −
H(Uk |Xk )
2
k∈A
ε00
= exp n I(UA ; XA ) + ε(n) + |A|
,
2
X
(241)
(242)
(243)
(244)
P
where ε(n) = k∈A∪0 εk (n) goes to zero as n → ∞. Here, (240)follows from parts 1 and 2 of Lemma
58. Assume
that ε00 is such that Kε00 /2 < ε. Defining the error event E4 = LA ≥ exp n(I(UA ; XA ) + ε) , we know from
Markov’s inequality that for n large enough
ε00
P{E4 } ≤ exp n ε(n) − ε + |A|
≤ ε0 .
(245)
2
Using (236) and (245) we can apply Lemma 45 and obtain deterministic encoding functions fk : Xkn → Mk , and
n such that (55) holds whenever the conditions (41) and (42)
deterministic decoding functions gA,Ae : MA → UA
e
are satisfied. Taking into account that gA,Ae(MA ) × gB,Be(MB ) ⊆ CA∪
e Be, we also have (54). (Note that, given a
specific code, P{E4 |CK = cK } < 1 already implies P{E4 |CK = cK } = 0 as the event E4 is determined by the code
CK alone.)
5) Proof of Proposition
39: Pick arbitrary j, k ∈ J .For nonempty B ⊆
J with j ∈ B we can write H(Xk |UB ) =
EUj fk,B pXj |Uj ( · |Uj ) as well as H(Y|UB ) = EUj gB pXj |Uj ( · |Uj ) , where
fk,B pXj |Uj ( · |uj ) := H Xk UB\j , Uj = uj ,
(246)
gB pXj |Uj ( · |uj ) := H Y UB\j , Uj = uj .
(247)
Observe that fk,B and gB are continuous functions of pXj |Uj ( · |uj ). Apply the support lemma [13, Appendix C]
with the functions fk,B and gB for all k ∈ J , j ∈ B ⊆ J , and |Xj | − 1 test functions, which guarantee that the
marginal distribution pXj does not change. We obtain a new random variable Ûj with H(Xk |UB\j Ûj ) = H(Xk |UB )
and H(Y|UB\j Ûj ) = H(Y|UB ). By rewriting (83)–(88) in terms of conditional entropies, it is evident that the
(<,I)
defining inequalities for RMI remain the same when replacing Uj by Ûj . The support of Ûj satisfies the required
cardinality bound5
|Ûj | ≤ |Xj | − 1 + J2J−1 + 2J−1
≤ |Xj | + 4J .
5
There are J ways to choose k and 2J−1 ways to choose B.
(248)
(249)
30
The same process is repeated for every j ∈ J .
6) Proof of Theorem 40: Pick a total order < on J , a set I ⊆ J and (UJ , ∅) ∈ P∗ . To obtain a code we
apply Lemma 34 with K = J + 1, XK = UK = Y, B = Be = {K}, Ae = A for all ∅ 6= A ⊆ J , and rates
Rj = I(Uj ; Xj |U=j ), RK = log|Y|, as suggested by Proposition 28. As in the proof of Lemma 34 let fej denote
the encoding function without binning and with rate n−1 log fej ≤ I(Uj ; Xj ) + 2ε . Furthermore, let fj0 be the
encoding function including binning, obtaining a rate of n−1 log |fj0 | ≤ I(Uj ; Xj |U=j ) + ε. Finally we obtain the
(n, RJ + ε)-code fJ by setting fj := fej for j ∈ I and fj := fj0 for j ∈
/ I . Let the decoding functions be gA := gA,A
e j := fej (Xj )
for all ∅ 6= A ⊆ J . Furthermore, for each j ∈ J , we define the decoding function gej , which maps W
(j)
onto its codebook entry, i. e., gej (w) = Vw (using the notation from Appendix B4). Also let Wj := fj (Xj ) and
e j.
Wj0 := fj0 (Xj ). For later use, we note that Wj0 is a function of Wj , which is in turn a function of W
0
n
0
Let the event SA be the success event that joint typicality Y, XA , gA (WA ) ∈ T[YXA UA ] holds. Also let Sej be
e j) ∈ T n
the event that Y, Xj , gej (W
. For any A = wk , k ∈ J , and A0 ⊆ A, we have
[YXj Uj ]
X
X
Rj =
j∈A0
j∈A0
≥
X
0
j∈A
I(Uj ; Xj |U=j )
I Uj ; XA0 U=j , UA\A0
= I UA0 ; XA0 UA\A0 .
(250)
(251)
(252)
0 } ≥ 1 − ε by Lemma 34.
Thus, (252) shows that condition (41) is satisfied and for n large enough we have P{SA
Clearly also P{Sej } ≥ 1 − ε for each j ∈ J and n large enough, using Lemmas 56 and 57.
Pick an arbitrary6 ε0 > 0. Provided that n is large enough and ε small enough, we have for any A = wk
1
1
0
I(Y; WA ) ≥ I Y; WA
(253)
n
n
1
0
)
(254)
≥ I Y; gA (WA
n
1
0
= H(Y) − H Y gA (WA
)
(255)
n
1
0
≥ H(Y) − H Y, 1SA0 gA (WA
)
(256)
n
1
1
0
0
= H(Y) − H 1SA0 gA (WA
) − H Y gA (WA
), 1SA0
(257)
n
n
1
0
0
≥ H(Y) − ε0 − (1 − ε)H Y gA (WA
), SA
− εH(Y)
(258)
n
1
0
0
≥ H(Y) − ε0 − H Y gA (WA
), SA
(259)
n
1X
0
0
n
≥ H(Y) − ε0 −
(uA )
(260)
P gA (WA
) = uA SA
log T[Y|U
A]
n u
A
≥ H(Y) − H(Y|UA ) − ε0
0
= I(UA ; Y) − ε .
(261)
(262)
Here, (253) and (254) follow from the data processing inequality [23, Theorem 2.8.1], we applied the entropy bound
[23, Theorem 2.6.4] in (260), and part 3 of Lemma 58 in (261). For A = J we specifically obtain
(87)
1
I(Y; WJ ) ≥ I(UJ ; Y) − ε0 ≥ νJ − ε0 .
(263)
n
For k ∈ J and A = =k we obtain the following chain of inequalities, where (268) and (269) will be justified
6
In what follows, we will routinely merge expressions that can be made arbitrarily small (for n large and ε sufficiently small) and bound
them by ε0 .
31
subsequently.
1
1
1
0
I(Y; Wk ) ≥ I Y; Wk0 ≥ I Y; Wk0 WA
n
n
n
1
1
0
0
0
= I Y; Wk WA − I Y; WA
n
n
(262)
1
0
≥ I(UA Uk ; Y) − ε0 − I Y; WA
n
1
1
0
0
0
+ I XA ; WA
Y
= I(UA Uk ; Y) − ε − I XA ; WA
n
n
1
0
≥ I(UA Uk ; Y) − ε0 − I(XA ; UA ) + H(XA |Y) − H XA WA
,Y
n
≥ I(UA Uk ; Y) − ε0 − I(XA ; UA ) + H(XA |Y) − H(XA |UA , Y)
= I(Uk ; Y|UA ) − ε
0
(85)
≥ νk − ε0 .
(264)
(265)
(266)
(267)
(268)
(269)
(270)
(271)
0 −
Equality in (267) follows from the Markov chain WA
◦− XA −◦− Y. In (268), we used that for ε small enough and
n large enough, we have
1
1
0
0
I XA ; WA
= H WA
(272)
n
n
X
1
≤
H Wj0
(273)
n
j∈A
X
≤
I(Uj ; Xj |U=j ) + ε
(274)
j∈A
≤ I(UA ; XA ) + ε0 ,
(275)
where (273) follows from the chain rule for entropy [23, Theorem 2.2.1] and the data processing inequality [23,
Theorem 2.8.1] and (274) follows from the entropy bound [23, Theorem 2.6.4] and the fact that n−1 log fj0 ≤
I(Uj ; Xj |U=j ) + ε. The inequality (269) follows similar to (262) as for n large enough and ε small enough,
1
1
0
0
H XA WA
, Y ≤ H XA gA (WA
), Y
(276)
n
n
1
0
≤ H XA , 1SA0 gA (WA
), Y
(277)
n
1
0
0
≤ ε0 + H XA gA (WA
), Y, SA
(278)
n
X
1
0
0
≤ ε0 +
P gA (WA
) = uA , Y = y SA
log T[XnA |UA ,Y] (uA , y)
(279)
n u ,y
A
0
≤ ε + H(XA |UA , Y).
(280)
For k ∈ I , we have similarly to (262) that
1
1
ek
I(Y; Wk ) = I Y; W
n
n
1
e k)
≥ I Y; gek (W
n
1
e k)
= H(Y) − H Y gek (W
n
1
e k)
≥ H(Y) − H Y, 1Sek gek (W
n
1
1
e k ), 1 e
≥ H(Y) − H 1Sek − H Y gek (W
S
k
n
n
(281)
(282)
(283)
(284)
(285)
32
1
e k ), Sek
H Y gek (W
n
o
1X n
0
e k ) = uk Sek log T n (uk )
P geA (W
≥ H(Y) − ε −
[Y|Uk ]
n u
≥ H(Y) − ε0 −
(286)
(287)
k
≥ H(Y) − ε0 − H(Y|Uk )
(288)
= I(Uk ; Y) − ε0 ≥ νk − ε0 .
(289)
(86)
7) Proof of Theorem 41: For (νΠ , RJ ) ∈ RMI we apply Definition 27, choosing an (n, RJ )-code fJ for XJ
and define Uj := fj (Xj ) for j ∈ J . For any A ⊆ J we thus have
1
I(UA ; Y) ≥ νA .
n
i−1
With Uj,i := (Uj , Xi−1
, Yni+1 ) we have
j,1 ) and Qi := (Y
X
n
Rj ≥ H(UB )
(290)
(291)
j∈B
(292)
= I(UB ; XB )
(293)
= I(UB ; XB , Y)
= I(UB ; Y) + I(UB ; XB |Y)
(294)
= I(UA UB ; Y) − I UA\B ; Y UB + I(UB ; XB |Y)
= I(UA ; Y) + I UB\A ; Y UA − I UA\B ; Y UB + I(UB ; XB |Y)
≥ nνA + I UB\A ; Y UA − I UA\B ; Y UB + I(UB ; XB |Y)
≥ nνA − I UA\B ; Y + I(UB ; XB |Y)
n
X
νA − I UA\B ; Yi Yi−1 + I UB ; XB,i YXi−1
=
B
(295)
≥
(300)
=
i=1
n
X
i=1
n
X
i=1
i−1
νA − I UA\B,i ; Yi Qi + I UB ; XB,i YXB
νA − I UA\B,i ; Yi Qi + I(UB,i ; XB,i |Yi Qi ) .
(296)
(297)
(298)
(299)
(301)
The result follows by a standard time-sharing argument. Note that the required Markov chain and the independence
are satisfied.
8) A Random Coding Lemma:
(δ)
Lemma 45. For any δ > 0 let Cδ be a random code and (Ei )i∈I finitely many error events associated with the
(δ)
code Cδ . If we have P Ei
≤ δ for every i ∈ I , then, for any ε > 0 we can find δ > 0 such that there is a code
(δ)
c with P Ei Cδ = c ≤ ε for every i ∈ I .
(δ)
Proof. We apply Markov’s inequality to the random variable P Ei Cδ and obtain
n
√ o
√
δ
(δ)
P P Ei Cδ ≥ δ ≤ √ = δ.
(302)
δ
Applying the union bound yields
(
)
[ n (δ)
X n (δ)
√ o
√ o
P
P Ei Cδ ≥ δ
≤
P P Ei Cδ ≥ δ
(303)
i∈I
i∈I
√
≤ |I| δ.
(304)
√
√
(δ)
In particular,
exists at least one code c such that P Ei Cδ = c < δ for all i ∈ I if |I| δ < 1. Choosing
2 there
1
δ = min ε , 2|I|2 yields the desired result.
33
9) Results on Convex Polyhedra: Let H be the convex polyhedron H := {x ∈ Rn : Ax ≥ b} for an m × n
matrix A = (a(1) , a(2) , . . . , a(m) )T and b ∈ Rm , where aT
(j) is the j th row of A. In this section we will use the
notation of [54]. In particular, we shall call a closed convex set line-free if it does not contain a (straight) line. The
characteristic cone of a closed convex set C is defined as cc(C) := {y : x + λy ∈ C for all λ ≥ 0} (x ∈ C arbitrary)
and ext(C) is the set of all extreme points of C , i. e., points x ∈ C that cannot be written as x = λy + (1 − λ)z
with y, z ∈ C , y 6= z and λ ∈ (0, 1).
Lemma 46. A point y is in cc(H) if and only if Ay ≥ 0.
Proof. If Ay ≥ 0, x ∈ H and λ ≥ 0, A(x + λy) ≥ Ax ≥ b. On the other hand, for aT
(j) y < 0, we have
aT
(j) (x + λy) < bj for λ >
bj −aT
(j) x
aT
(j) y
> 0.
Lemma 47. If, for every i ∈ [1 : n], there exists j ∈ [1 : m] such that ei = a(j) and for every j ∈ [1 : m], a(j) ≥ 0,
then H is line-free and cc(H) = Rn+ .
Proof. For any y ∈ Rn+ , clearly Ay ≥ 0 and hence y ∈ cc(H) by Lemma 46. If y ∈
/ Rn+ we have yi < 0 for some
T
i ∈ [1 : n] and choose j ∈ [1 : m] such that a(j) = ei , resulting in a(j) y = yi < 0. To show that H is line-free
assume that x + λy ∈ H for all λ ∈ R. This implies ±y ∈ cc(H), i. e., y = 0.
Definition 48. A point x is on an extreme ray of the cone cc(H) if the decomposition x = y + z with y, z ∈ cc(H)
implies that y = λz for some λ ∈ R.
It is easy to see that the points on extreme rays of O are given by x = λei for λ ≥ 0 and i ∈ [1 : n].
Define A(x) := {j ∈ [1 : m] : aT
(j) x = bj }. We say that exactly n0 linearly independent inequalities are satisfied
with equality at x, if Ax ≥ b and (a(j) )j∈A(x) has rank n0 .
Lemma 49. x ∈ ext(H) if and only if exactly n linearly independent inequalities are satisfied with equality at x.
Proof. Assuming that less than n linearly independent inequalities are satisfied with equality at x, we find
0 6= c ⊥ (a(j) )j∈A(x) and thus x ± εc ∈ H for a small ε > 0, showing that x ∈
/ ext(H).
0
00
Conversely assume x ∈
/ ext(H), i. e., x = λx + (1 − λ)x for λ ∈ (0, 1) and x0 , x00 ∈ H, x0 6= x00 . For
0
T
00
T
0
T
00
any j ∈ A(x), we then have λaT
(j) x + (1 − λ)a(j) x = bj , which implies a(j) x = a(j) x = bj and therefore
0 6= x0 − x00 ⊥ (a(j) )j∈A(x) .
Lemma 50. Assuming that H is line-free, and that exactly n − 1 linearly independent inequalities are satisfied with
equality at x. Then either x = λc + (1 − λ)d where λ ∈ (0, 1) and c, d ∈ ext(H) or x = c + d where c ∈ ext(H)
and d 6= 0 lies on an extreme ray of cc(H).
Proof. We obtain 0 6= r ⊥ (a(j) )j∈A(x) . Define λ1 := inf{λ : x + λr ∈ H} and λ2 := sup{λ : x + λr ∈ H}.
Clearly λ1 ≤ 0 ≤ λ2 . As H is line-free, we may assume without loss of generality λ1 = −1 (note that x ∈
/ ext(H))
and set c = x − r . We now have c ∈ ext(H) as otherwise c − εr ∈ H for some small ε > 0.
2
If λ2 < ∞, define d = x + λ2 r which yields d ∈ ext(H) and x = λc + (1 − λ)d with λ = λ2λ+1
. Note that
λ2 6= 0 as x ∈
/ ext(H).
If λ2 = ∞ we have x − c = r ∈ cc(H). We need to show that r is also on an extreme ray of cc(H). Assuming
0
00
T 0
T 00
r = r 0 + r 00 with r 0 , r 00 ∈ cc(H) yields aT
(j) (r + r ) = 0, which implies a(j) r = a(j) r = 0 for every j ∈ A(x)
by Lemma 46.
For each j ∈ [0 : J], define the closed convex polyhedron H(j) := {x ∈ RK+j : A(j) x ≥ b(j) }, where A(j) is a
matrix and b(j) a vector of appropriate dimension. We make the following three assumptions:
1) A(j) and b(j) are defined recursively as
(j−1)
b(j−1)
A
0
A(j) := 0T
b(j) = c(j)
(305)
1 ,
,
1
(j)
eT
1
c
j
2
(j)
(j)
where c1 and c2 are arbitrary reals.
34
2) Each entry of A(0) equals 0 or 1 and for all k ∈ K at least one row of A(0) is equal to eT
k . Due to Assumption 1,
this also implies that each entry of A(j) is in {0, 1} and for all k ∈ [1 : K + j] at least one row of A(j) is
equal to eT
k.
(j)
(j)
3) For any extreme point x ∈ ext(H(0) ) and any j ∈ J , assume xj ≤ c2 − c1 .
Lemma 51. Under Assumptions 1 to 3, for every k ∈ [0 : J] and every extreme point y ∈ ext(H(k) ) there is an
extreme point x ∈ ext(H(0) ) and a subset Ik ⊆ [1 : k] such that yK = xK and for every j ∈ J ,
(
xj ,
j∈
/ Ik ,
yj =
(306)
(j)
(j)
c2 − c1 , j ∈ Ik ,
and for every j ∈ [1 : k],
yK+j =
( (j)
c2 − xj ,
(j)
c1 ,
j∈
/ Ik ,
j ∈ Ik .
(307)
Proof. For every j ∈ J , H(j) is line-free
and can be written [54, Lemma 6, p. 25]
by Assumption 2 and Lemma 47,
(j)
(j)
(j)
(j)
as H = cc(H ) + conv ext(H ) . Lemma 47 also implies cc(H) = O.
Let us proceed inductively over k ∈ [0 : J]. For k = 0 the statement is trivial. Given any y ∈ ext(H(k) ), we
need to obtain x ∈ ext(H(0) ) and Ik such that y is given according to (306) and (307). Let z = y K+k−1
be
1
(k)
(k)
the truncation of y . Exactly K + k linear independent inequalities of A y ≥ b are satisfied with equality by
Lemma 49, which is possible in only two different ways:
(k−1)
z ≥ b(k−1) are satisfied with
• Construction I: Exactly K + k − 1 linear independent inequalities of A
(k−1)
equality, i. e., z ∈ ext(H
) by Lemma 49, and at least one of
(k)
(308)
(k)
c2 ,
(309)
yK+k ≥ c1 ,
yk + yK+k ≥
•
is satisfied with equality.
As z ∈ ext(H(k−1) ), there exists x ∈ ext(H(0) ) and Ik−1 such that (306) holds for j ∈ J and (307) holds for
j ∈ [1 : k − 1] by the induction hypothesis. In particular yk = xk . Assuming that (309) holds with equality, we
(k)
have yK+k = c2 − xk . Thus, the point x together with Ik = Ik−1 yields y from (306) and (307). Equality in
(308) implies equality in (309) by Assumption 3.
Construction II: Exactly K + k − 2 linear independent inequalities of A(k−1) z ≥ b(k−1) are satisfied with
equality and (308) and (309) are both satisfied with equality as well. Additionally, these K + k inequalities
together need to be linearly independent. This can occur in two different ways by Lemma 50.
(k)
Assume z = λx + (1 − λ)x0 for x, x0 ∈ ext(H(k−1) ), x 6= x0 and λ ∈ (0, 1). This implies yK+k = c1 and
(k)
(k)
(k)
(k)
yk = λxk + (1 − λ)x0k = c2 − c1 , which by Assumption 3 already implies xk = x0k = c2 − c1 . Thus,
(308) and (309) are satisfied (with equality) for every λ ∈ [0, 1] and y cannot be an extreme point as it can be
written as a non-trivial convex combination.
We can thus focus on the second option which is that z is on an extreme ray of H(k−1) , i. e., z = x + λek0
for some x ∈ ext(H(k−1) ), λ > 0 and k 0 ∈ [1 : K + k − 1]. If k 0 6= k , (308) and (309) are satisfied for all
λ > 0 and thus y cannot be an extreme point because it can be written as a non-trivial convex combination.
For k 0 = k the point x with Ik = Ik−1 ∪ k yields the desired extreme point.
C. Types, Typical Sequences and Related Results
In this section we introduce notions and results needed for the mathematical developments and proofs in this
work. The results can be easily derived from the standard formulations provided in [13] and [25].
Definition 52 (Type; [25, Definition 2.1]). The type of a vector x ∈ X n is the random variable X̂ ∼ pX̂ ∈ P(X )
defined by
pX̂ (x) =
1
N (x|x), for every x ∈ X ,
n
(310)
35
where N (x|x) denotes the number of occurrences of x in x. For a random variable X̂, the set of n-sequences with
type X̂ is denoted TX̂n .
n .
For a pair of random variables (X, Y), we say that y ∈ Y n has conditional type Y given x ∈ X n if (x, y) ∈ TXY
n
n
The set of all n-sequences y ∈ Y with conditional type Y given x will be denoted TY|X (x).
A key property of types is the following result, known as type counting.
Lemma 53 (Type counting; [25, Lemma 2.2]). The number of different types of sequences in X n is less than
(n + 1)|X | .
Some important properties of types are listed in the following lemma.
Lemma 54 ([25, Lemmas 2.5 and 2.6]).
e on X , and x ∈ T n
1) For any two random variables X, X
X
h
i
e = x} = exp − n H(X) + D(XkX)
e ,
P{X
e is a sequence of n i.i.d. copies of X
e.
where X
n (x) 6= ∅
2) For a pair of random variables (X, Y) on X × Y and x ∈ X n such that TY|X
n
(n + 1)−|X ||Y| exp nH(Y|X) ≤ TY|X
(x) ≤ exp nH(Y|X) .
(311)
(312)
Definition 55 (Typicality; [13, Section 2.4]). Consider X ∼ pX ∈ P(X ) and δ ≥ 0. We call the random variable
Y ∼ pY ∈ P(X ) δ -typical if Y ∈ T[X]δ with
e ∼ p e ∈ P(X ) : |p e (x) − pX (x)| ≤ δpX (x), ∀x ∈ X }.
T[X]δ := {X
X
X
(313)
n .
A sequence x ∈ X n is δ -typical if its type X̂ is δ -typical. The set of all δ -typical n-sequences is denoted T[X]δ
n
Given pX,Y ∈ P(X × Y) we call the elements of T[XY]δ
the jointly δ -typical n-sequences. We also define the
n
n
conditionally typical n-sequences T[Y|X]δ (x) := {y ∈ Y n : (x, y) ∈ T[XY]δ
}. Typical sequences have several useful
properties, which are presented in the following.
Lemma 56 (Asymptotic equipartition property; [13, Sections 2.4]). Let X be n i.i.d. copies of X. For any δ > 0 we
have
n
o
n
lim P X ∈ T[X]δ
= 1.
(314)
n→∞
n , and let the random
Lemma 57 (Conditional typicality lemma; [13, Section 2.5]). Choose δ 0 > 0, assume x ∈ T[X]δ
0
Qn
0
vector Y be distributed according to pY (y) = i=1 pY|X (yi |xi ). For any δ > δ we have
n
o
n
lim P Y ∈ T[Y|X]δ
(x) = 1.
(315)
n→∞
Lemma 58 (Size of typical sets; [13, Sections 2.4 and 2.5]). The following properties hold for (X, Y) ∼ pX,Y :
1) Using ε(δ) = δH(X),
n
T[X]δ
≤ en(H(X)+ε(δ)) .
(316)
2) For δ > 0, ε0 > 0, n sufficiently large (as a function of ε0 and pX ), and ε(δ) = δH(X),
n
T[X]δ
≥ (1 − ε0 )en(H(X)−ε(δ)) .
(317)
n
(x) ≤ en(H(Y|X)+ε(δ)) .
T[Y|X]δ
(318)
3) For x ∈ X n and ε(δ) = δH(Y|X),
36
n . For δ > δ 0 , ε0 > 0, n sufficiently large (as a function of ε0 and p
4) Let δ 0 > 0 and x ∈ T[X]δ
0
X,Y ), and
ε(δ) = δH(Y|X),
n
T[Y|X]δ
(x) ≥ (1 − ε0 )en(H(Y|X)−ε(δ)) .
(319)
Remark 12. We shall adopt the δ -convention [25, Convention 2.11] and assume the existence of an adequate sequence
(δn )n∈N approaching 0 for every set of random variables. We will omit δ in the notation, e. g., we will write T[X] ,
n , and T
T[X]
[X|Y] (y).
Lemma 59 (Generalized Markov lemma; [62, Lemma 3.4]). Let XK and UK be such that Uk −◦− Xk −◦− (XK\k , UK\k )
for every k ∈ K and fix ε > 0. For n ∈ N and for each k ∈ K let Mk ∈ N with Mk > enI(Xk ;Uk ) . Furthermore,
e k,M (Mk := [1 : Mk ]) bemutually independent random vectors, also independent of X[1 : K] , drawn uniformly
let U
k
from T[Unk ] . Then, for sufficiently large n there exist K functions fk : Xkn × (Ukn )Mk → Mk such that, using
e k,M ) and U∗ = U
e k,W , we have
Wk = fk (Xk , U
k
k
k
n
o
P (XK , U∗K ) ∈ T[XnK UK ] ≥ 1 − ε.
(320)
R EFERENCES
[1] G. Pichler, P. Piantanida, and G. Matz, “Distributed information-theoretic biclustering of two memoryless sources,” in Proc. 53rd Annual
Allerton Conference on Communication, Control, and Computing, Monticello, IL, Sep. 2015, pp. 426–433.
[2] ——, “Distributed information-theoretic biclustering,” in Proc. IEEE Int. Symp. on Inform. Theory, Barcelona, Spain, Jul. 2016, pp.
1083–1087.
[3] ——, “A multiple description CEO problem with log-loss distortion,” in Proc. IEEE Int. Symp. on Inform. Theory, Aachen, Germany,
Jun. 2017, pp. 111–115.
[4] C. E. Shannon, “Coding theorems for a discrete source with a fidelity criterion,” in Claude Elwood Shannon: collected papers, N. J. A.
Sloane and A. D. Wyner, Eds. IEEE Press, 1993, pp. 325–350.
[5] J. A. Hartigan, “Direct clustering of a data matrix,” Journal of the American Statistical Association, vol. 67, no. 337, pp. 123–129, Mar.
1972.
[6] B. Mirkin, Mathematical Classification and Clustering. Kluwer Academic Publisher, 1996.
[7] S. C. Madeira and A. L. Oliveira, “Biclustering algorithms for biological data analysis: A survey,” IEEE/ACM Trans. Comput. Biol.
Bioinformatics, vol. 1, no. 1, pp. 24–45, Aug. 2004.
[8] Y. Cheng and G. M. Church, “Biclustering of expression data,” in Proc. 8th Int. Conf. Intelligent Syst. for Molecular Biology, vol. 8,
San Diego, CA, Aug. 2000, pp. 93–103.
[9] A. Tanay, R. Sharan, and R. Shamir, “Biclustering algorithms: A survey,” Handbook of Computational Molecular Biology, vol. 9, no.
1-20, pp. 122–124, 2005.
[10] R. Sharan, “Analysis of biological networks: Network modules – clustering and biclustering,” lecture notes, 2006. [Online]. Available:
http://www.cs.tau.ac.il/~roded/courses/bnet07.html
[11] N. Slonim, G. S. Atwal, G. Tkačik, and W. Bialek, “Information-based clustering,” Proc. of the Nat. Academy of Sciences of the United
States of America, vol. 102, no. 51, pp. 18 297–18 302, Dec. 2005.
[12] I. S. Dhillon, S. Mallela, and D. S. Modha, “Information-theoretic co-clustering,” in Proc. 9th ACM SIGKDD Int. Conf. on Knowledge
Discovery and Data Mining, Washington, DC, Aug. 2003, pp. 89–98.
[13] A. El Gamal and Y.-H. Kim, Network Information Theory. Cambridge University Press, 2011.
[14] J. Körner and K. Marton, “How to encode the modulo-two sum of binary sources,” IEEE Trans. Inf. Theory, vol. 25, no. 2, pp. 219–221,
Mar. 1979.
[15] T. S. Han, “Hypothesis testing with multiterminal data compression,” IEEE Trans. Inf. Theory, vol. 33, no. 6, pp. 759–772, Nov. 1987.
[16] M. B. Westover and J. A. O’Sullivan, “Achievable rates for pattern recognition,” IEEE Trans. Inf. Theory, vol. 54, no. 1, pp. 299–320,
Jan. 2008.
[17] N. Tishby, F. C. Pereira, and W. Bialek, “The information bottleneck method,” in Proc. 37th Annual Allerton Conference on Communication,
Control, and Computing, Monticello, IL, Sep. 1999, pp. 368–377. [Online]. Available: https://arxiv.org/abs/physics/0004057
[18] T. A. Courtade and T. Weissman, “Multiterminal source coding under logarithmic loss,” IEEE Trans. Inf. Theory, vol. 60, no. 1, pp.
740–761, Jan. 2014.
[19] C. Nair, “Upper concave envelopes and auxiliary random variables,” Int. J. of Advances in Eng. Sciences and Appl. Math., vol. 5, no. 1,
pp. 12–20, Mar. 2013.
[20] A. A. El Gamal and T. M. Cover, “Achievable rates for multiple descriptions,” IEEE Trans. Inf. Theory, vol. 28, no. 6, pp. 851–857,
Nov. 1982.
[21] V. K. Goyal, “Multiple description coding: Compression meets the network,” IEEE Signal Process. Mag., vol. 18, no. 5, pp. 74–93, Sep.
2001.
[22] W. Rudin, Principles of Mathematical Analysis, 3rd ed. McGraw-Hill, 1976.
[23] T. M. Cover and J. A. Thomas, Elements of Information Theory. John Wiley & Sons, 2006.
[24] A. Orlitsky and J. R. Roche, “Coding for computing,” IEEE Trans. Inf. Theory, vol. 47, no. 3, pp. 903–917, Mar. 2001.
37
[25] I. Csiszár and J. Körner, Information Theory: Coding Theorems for Discrete Memoryless Systems. Cambridge University Press, Aug.
2011.
[26] S.-Y. Tung, “Multiterminal source coding,” Ph.D. dissertation, Cornell University, May 1978.
[27] P. Gács and J. Körner, “Common information is far less than mutual information,” Problems of Control and Inform. Theory, vol. 2, pp.
149–162, 1973.
[28] A. B. Wagner, B. G. Kelly, and Y. Altug, “Distributed rate-distortion with common components,” IEEE Trans. Inf. Theory, vol. 57,
no. 7, pp. 4035–4057, 2011.
[29] H. S. Witsenhausen, “On sequences of pairs of dependent random variables,” SIAM Journal on Applied Mathematics, vol. 28, no. 1, pp.
100–113, Jan. 1975.
[30] R. Dobrushin and B. Tsybakov, “Information transmission with additional noise,” IRE Trans. on Inform. Theory, vol. 8, no. 5, pp.
293–304, Sep. 1962.
[31] R. Gilad-Bachrach, A. Navot, and N. Tishby, “An information theoretic tradeoff between complexity and accuracy,” in Learning Theory
and Kernel Machines. Springer, 2003, pp. 595–609.
[32] R. Ahlswede and J. Körner, “Source coding with side information and a converse for degraded broadcast channels,” IEEE Trans. Inf.
Theory, vol. 21, no. 6, pp. 629–637, Nov. 1975.
[33] A. D. Wyner, “On source coding with side information at the decoder,” IEEE Trans. Inf. Theory, vol. 21, no. 3, pp. 294–300, May 1975.
[34] H. S. Witsenhausen and A. D. Wyner, “A conditional entropy bound for a pair of discrete random variables,” IEEE Trans. Inf. Theory,
vol. 21, no. 5, pp. 493–501, Sep. 1975.
[35] A. Wyner and J. Ziv, “A theorem on the entropy of certain binary sequences and applications: Part I,” IEEE Trans. Inf. Theory, vol. 19,
no. 6, pp. 769–772, Nov. 1973.
[36] H. Witsenhausen, “Entropy inequalities for discrete channels,” IEEE Trans. Inf. Theory, vol. 20, no. 5, pp. 610–616, Sep. 1974.
[37] A. Wyner, “A theorem on the entropy of certain binary sequences and applications: Part II,” IEEE Trans. Inf. Theory, vol. 19, no. 6, pp.
772–777, Nov. 1973.
[38] R. Ahlswede and J. Körner, “On the connection between the entropies of input and output distributions of discrete memoryless channels,”
in Proc. 5th Conf. Probability Theory, Sep. 1974, Brasov, Romania, 1977, pp. 13–23.
[39] R. Ahlswede and I. Csiszár, “Hypothesis testing with communication constraints,” IEEE Trans. Inf. Theory, vol. 32, no. 4, pp. 533–542,
Jul. 1986.
[40] E. Erkip and T. M. Cover, “The efficiency of investment information,” IEEE Trans. Inf. Theory, vol. 44, no. 3, pp. 1026–1040, May
1998.
[41] C. Chapman, personal communication, Aug. 2017. [Online]. Available: https://mathoverflow.net/questions/213084/
do-binary-symmetric-channels-maximize-mutual-information
[42] R. E. Moore, R. B. Kearfott, and M. J. Cloud, Introduction to Interval Analysis. SIAM, 2009.
[43] “GNU Octave,” Free Software Foundation. [Online]. Available: https://www.gnu.org/software/octave/
[44] O. Heimlich, “GNU Octave Interval Package.” [Online]. Available: https://octave.sourceforge.io/interval/
[45] G. Pichler, “DSBS-MutInf-counterexample,” 2017, Program code. [Online]. Available: https://doi.org/10.5281/zenodo.1042588
[46] G. R. Kumar and T. A. Courtade, “Which Boolean functions are most informative?” in Proc. IEEE Int. Symp. on Inform. Theory,
Istanbul, Turkey, Jul. 2013, pp. 226–230.
[47] T. A. Courtade and G. R. Kumar, “Which Boolean functions maximize mutual information on noisy inputs?” IEEE Trans. Inf. Theory,
vol. 60, no. 8, pp. 4515–4525, Aug. 2014.
[48] J. G. Klotz, D. Kracht, M. Bossert, and S. Schober, “Canalizing Boolean functions maximize mutual information,” IEEE Trans. Inf.
Theory, vol. 60, no. 4, pp. 2139–2147, Apr. 2014.
[49] G. Pichler, G. Matz, and P. Piantanida, “A tight upper bound on the mutual information of two Boolean functions,” in Proc. Inform.
Theory Workshop, Cambridge, UK, Sep. 2016, pp. 16–20.
[50] T. Berger, “Multiterminal source coding,” in The Information Theory Approach to Communications, G. Longo, Ed. Springer, 1977, pp.
171–231.
[51] T. Berger, Z. Zhang, and H. Viswanathan, “The CEO problem,” IEEE Trans. Inf. Theory, vol. 42, no. 3, pp. 887–902, May 1996.
[52] S. Fujishige, Submodular Functions and Optimization, 2nd ed., ser. Annals of Discrete Mathematics. Elsevier Science, 2005.
[53] R. Schneider, Convex Bodies: The Brunn-Minkowski Theory, 2nd ed. Cambridge University Press, 2014.
[54] B. Grünbaum, Convex Polytopes. Springer, New York, 2003.
[55] W. Rudin, Functional Analysis, 2nd ed. McGraw-Hill, 1991.
[56] A. A. Gohari and V. Anantharam, “Evaluation of Marton’s inner bound for the general broadcast channel,” IEEE Trans. Inf. Theory,
vol. 58, no. 2, pp. 608–619, Feb. 2012.
[57] V. Jog and C. Nair, “An information inequality for the BSSC broadcast channel,” in Inform. Theory and Applicat. Workshop (ITA), San
Diego, CA, Feb. 2010, pp. 1–8.
[58] H. G. Eggleston, Convexity, P. Hall and F. Smithies, Eds. Cambridge University Press, 1958.
[59] A. D. Wyner and J. Ziv, “The rate-distortion function for source coding with side information at the decoder,” IEEE Trans. Inf. Theory,
vol. 22, no. 1, pp. 1–10, Jan. 1976.
[60] C. D. Aliprantis and K. C. Border, Infinite Dimensional Analysis: A Hitchhiker’s Guide, 3rd ed. Springer, 2006.
[61] J. R. Munkres, Topology. Prentice Hall, 2000.
[62] T. S. Han and K. Kobayashi, “A unified achievable rate region for a general class of multiterminal source coding systems,” IEEE Trans.
Inf. Theory, vol. 26, no. 3, pp. 277–288, May 1980.
| 7cs.IT
|
1
Application of Microgrids in Supporting
Distribution Grid Flexibility
Alireza Majzoobi, Student Member, IEEE, and Amin Khodaei, Senior Member, IEEE
S
Abstract—Distributed renewable energy resources have
attracted significant attention in recent years due to the falling cost
of the renewable energy technology, extensive federal and state
incentives, and the application in improving load-point reliability.
This growing proliferation, however, is changing the traditional
consumption load curves by adding considerable levels of
variability and further challenging the electricity supply-demand
balance. In this paper, the application of microgrids in effectively
capturing the distribution network net load variability, caused
primarily by the prosumers, is investigated. Microgrids provide a
viable and localized solution to this challenge while removing the
need for costly investments by the electric utility on reinforcing the
existing electricity infrastructure. A flexibility-oriented microgrid
optimal scheduling model is proposed and developed to coordinate
the microgrid net load with the aggregated consumers/prosumers
net load in the distribution network with a focus on ramping
issues. The proposed coordination is performed to capture both
inter-hour and intra-hour net load variabilities. Numerical
simulations on a test distribution feeder with one microgrid and
several consumers and prosumers exhibit the effectiveness of the
proposed model.
Index Terms--Distributed generation, duck curve, flexibility,
microgrid optimal scheduling, renewable energy resource.
NOMENCLATURE
Indices:
c
Superscript for distribution network consumers and
prosumers.
ch
Superscript for energy storage charging.
dch
Superscript for energy storage discharging.
d
Index for loads.
i
Index for DERs.
j
Index for consumers/prosumers at the distribution
network.
k
Index for intra-hour time periods.
s
Index for scenarios.
t
Index for inter-hour time periods.
u
Superscript for the utility grid.
Sets:
D
F
G
N
O
Set of adjustable loads.
Set of flexibility constraints.
Set of dispatchable units.
Set of consumers/prosumers.
Set of operation constraints.
The authors are with the Department of Electrical and Computer
Engineering, University of Denver, Denver, CO 80210 USA (email:
[email protected]; [email protected]).
Set of energy storage systems.
Parameters:
DR
Ramp down rate.
DT
Minimum down time.
E
Load total required energy.
F(.)
Generation cost.
MC
Minimum charging time.
MD
Minimum discharging time.
MU
Minimum operating time.
UR
Ramp up rate.
UT
Minimum up time.
w
Binary islanding indicator (1 if grid-connected, 0
if islanded).
α, β
Specified start and end times of adjustable loads.
ρ
Market price.
η
Energy storage efficiency.
λ
Value of lost load (VOLL).
ψ
Probability of islanding scenarios.
τ
Time period.
Δ1
Intra-hour flexibility limit.
Δ2
Inter-hour flexibility limit.
low
up
Δ1 /Δ1
Microgrid time-dependent intra-hour
lower/upper flexibility limit.
Δ2low/Δ2up
Microgrid time-dependent inter-hour
lower/upper flexibility limit.
Variables:
C
Energy storage available (stored) energy.
D
Load demand.
I
Commitment state of dispatchable units.
LS
Load curtailment.
P
DER output power.
PM
Utility grid power exchange with the microgrid.
SD
Shut down cost.
SU
Startup cost.
Tch
Number of successive charging hours.
Tdch
Number of successive discharging hours.
Ton
Number of successive ON hours.
off
T
Number of successive OFF hours.
u
Energy storage discharging state (1 when
discharging, 0 otherwise).
2
v
z
Energy storage charging state (1 when charging, 0
otherwise).
Adjustable load state (1 when operating, 0
otherwise).
I. INTRODUCTION
T
HE GROWING trend of renewable generation installations
in the United States, driven primarily by current renewable
portfolio standards in 27 states, efficiency incentives and net
metering in 43 states, and the falling cost of renewable
generation technologies [1], [2], challenges traditional practices
in balancing electricity supply and demand and calls for
innovative methods to reduce impacts on grid stability and
reliability. Fig. 1 shows daily net load (i.e., the consumer load
minus local generation) variations in California ISO, the socalled duck curve, as an example of this challenge [3]. As
renewable generation increases, to reach the 33% renewable
target by 2020, the power grid would require increased levels
of fast ramping units to address abrupt changes (as much as 13
GW in three hours) in the net load, caused by concurrent fall in
renewable generation and increase in demand.
Fig. 1. The current and future estimates of over-generation and ramping effect
in California [3].
To maintain system supply-demand balance, grid operators
traditionally rely on bulk power generation resources, such as
fast ramping hydro and thermal units, that can be quickly
dispatched and ramped up. These units, however, are limited in
number and capacity, capital-intensive, time-consuming to be
constructed, and subject to probable transmission network
congestions. Addressing the variability of renewable
generation, on the other hand, has long been an attractive area
of research to complement renewable generation forecasting
efforts [4]. Uncertainty considerations in power system
operation and planning have significantly increased in the past
few years as a large amount of uncertainty sources are
integrated in power systems as a result of renewable generation
proliferation. The renewable generation integration problem
can be investigated under two contexts of large-scale (which
attempts to manage the generation of wind and solar farms) [5][8], and small-scale (which deals with renewable generation in
the distribution level). Small-scale coordination approaches
mainly focus on various methods of demand side management,
such as demand response [9]-[12], energy storage [13]-[17],
and aggregated electric vehicles [18], [19]. However, these
methods each encounter obstacles that may prevent a viable
application, such as the need for advanced metering
infrastructure in deploying demand response, privacy and
customer willingness issues in electric vehicle applications, and
financial aspects in energy storage deployment. Leveraging
available flexibility in existing microgrids for addressing
renewable generation integration, as proposed in [20] and
extended in this paper, will offer a potentially more viable
solution to be used in distribution networks, and thus calls for
additional studies. The microgrid, as defined by the U.S.
Department of Energy, is “a group of interconnected loads and
distributed energy resources (DER) within clearly defined
electrical boundaries that acts as a single controllable entity
with respect to the grid and can connect and disconnect from
the grid to enable it to operate in both grid-connected or islandmode” [21]. The microgrid, as a novel distribution network
architecture with local generation, control, and consumption,
offers noticeable benefits to both consumers and utility
companies such as enhanced reliability and resilience, reduced
environmental impact, power quality improvement, improved
energy efficiency by loss reduction, and network congestion
relief. Microgrids can be operated in grid-connected and
islanded modes. In the grid-connected mode, which is the
default operation mode, the microgrid can import, export, or
have zero power exchange with the utility grid to achieve the
least-cost supply schedule (i.e., an economic operation).
Capability to switch to the islanded mode is the salient feature
of the microgrids which isolates the microgrid from faults
and/or disturbances in the upstream network to achieve the least
load curtailment (i.e., a reliable operation) [22]-[28].
Microgrids have been significantly deployed over the past few
years and are anticipated to grow even more in the near future
[29], [30], in both national and international levels [31], where
future power grids can be pictured as systems of interconnected
microgrids [32].
This paper focuses on the flexibility advantages of
microgrids as a complementary value proposition in grid
support. The microgrid capability in managing its power
exchange with the utility grid in the grid-connected mode is
specifically considered in this paper for mitigating the net load
ramping in the distribution network and to further ensure that
the power seen by the utility has manageable ramps. There have
been several studies that investigate how a microgrid can
participate in the upstream network market and offer services
to the grid. In [33], an optimal bidding strategy via a microgrid
aggregator is proposed to involve all small-scale microgrids
into an electricity market via real-time balancing market
bidding. In [34], an optimal bidding strategy based on two-stage
stochastic linear programming for an electric vehicle aggregator
who participates in the day-ahead energy and regulation
markets is proposed. Furthermore, it goes on to consider market
conditions and the associated uncertainty of the electric vehicle
fleet. A two-stage market model for microgrid power exchange
with the utility grid, via an aggregator, is proposed in [35] to
achieve an efficient market equilibrium. A risk-constrained
optimal hourly bidding model for microgrid aggregator is
proposed in [36] to consider various uncertainties and
maximize the microgrid benefit. The study in [37] proposes an
3
optimal dispatch strategy for the residential loads via artificial
neural network for calculating the demand forecast error when
the demand changes are known one hour ahead with respect to
the day-ahead forecasted values. The study in [38] presents a
stochastic bidding strategy for microgrids participating in
energy and spinning reserve markets, considering the load and
renewable generation uncertainty. In [39], a stochastic lookahead economic dispatch model for near-real-time power
system operation is proposed and its benefits and
implementability for assessing the power system economic risk
are further explored. These works primarily rely on a market
mechanism to procure microgrids’ flexibility and accordingly
capture the unbalanced power in the day-ahead market as well
as the ramping and variabilities caused by forecast errors or
unforeseen real-time events. In this paper, however, this
problem is studied from a microgrid perspective, i.e., how a
microgrid controller can manage local resources to offer
required/desired services to the utility grid. This work is
particularly important in networks that a market mechanism
cannot be established but grid operators are interested in lowcost and distributed solutions in managing grid flexibility. The
main contributions of this paper are listed as follows:
A flexibility-oriented microgrid optimal scheduling model
is developed to optimally manage local microgrid
resources while providing flexibility services to the utility
grid. This model is achieved by transforming the
distribution net load variability limits into constraints on
the microgrid net load.
A coordinated grid-connected and islanded operation is
considered in the model development to take into account
microgrid’s potential islanding while supporting the utility
grid in the grid-connected mode.
A high resolution operation is modeled via consideration
of both intra-hour and inter-hour time periods, which is
capable of integrating quick variations in renewable
generation.
Unlike existing studies on distribution network flexibility
procurement, which focus on microgrid participation in grid
support via a market mechanism, this paper investigates the
problem from a microgrid’s perspective, i.e., how a microgrid
controller can manage local resources to offer required/desired
services to the utility grid.
The rest of the paper is organized as follows. Section II
describes the outline of proposed flexibility-oriented microgrid
optimal scheduling model. Section III develops the model
formulation, including operation and flexibility constraints.
Section IV presents numerical simulations to show the merits
and the effectiveness of the proposed model applied to a test
distribution network. Section V discusses the specific features
of the proposed model, and finally, Section V concludes the
paper.
II. MODEL OUTLINE
Consider a distribution feeder consisting of a set N = {1, 2,
…, N} customers (both consumers and prosumers) and one
microgrid. The net load of each customer j ϵ N and the
𝑐
𝑀
microgrid are respectively denoted by 𝑃𝑗𝑡𝑘
and 𝑃𝑡𝑘
, where t is
the inter-hour time index and k is the intra-hour time index as
demonstrated in Fig. 2.
Fig. 2. The schematic diagram of inter-hour and intra-hour time intervals.
To fully supply the total net load in this feeder, a power of
𝑢
𝑃𝑡𝑘
needs to be provided by the utility grid where:
u
M
Ptks
Ptks
c
Pjtks
∀ t ,∀ k ,∀ s.
(1)
jN
To address the net load variability seen by the grid operator,
the intra-hour variability (2) and inter-hour variability (3) in the
utility grid power will need to be constrained:
u
Ptks
Ptu( k 1) s 1
∀ t , ∀ s, k 1,
(2)
Pt1us P(ut 1) Ks 2
∀ t ,∀ s.
(3)
These limits are selected by the grid operator based on the dayahead net load forecasts and desired grid flexibility during each
time interval. There are various methods to determine the grid
flexibility [6], [40]-[42]. If this calculated flexibility is less than
the required grid flexibility, which is obtained based on net load
forecasts, the grid operator can utilize distributed resources,
such as microgrids, to compensate the shortage in grid
flexibility. Therefore, intra- and inter-hour limits will be
obtained by comparing the available and required grid
flexibility. Considering the importance of grid flexibility limits
on the microgrid operation, a system-level study needs to be
performed by the utility company. This topic will be
investigated in a follow up research. The grid operator
furthermore can calculate these limits using a cost-benefit
analysis, i.e., to upgrade the current infrastructure to address
increasing flexibility requirements or to procure the flexibility
of existing microgrids and in turn pay for their service. This
topic, however, requires further analysis and modeling which
will be carried out in follow up research.
Fig. 3 shows the schematic diagram of a feeder consisting of
a microgrid along with other connected loads. The microgrid
can be scheduled based on price considerations, i.e., local
resources are scheduled in a way that the microgrid operation
cost is minimized during the grid-connected mode (Fig. 3-top).
The only factor impacting the microgrid scheduling results
from the utility grid side is the real-time electricity price (hence
the term price-based scheduling). The price-based scheduling
can potentially exacerbate the consumption variability. On the
other hand, microgrid resources can be scheduled in
coordination with other loads in the same distribution feeder,
and thus support the utility grid in mitigating potential
variabilities and ensuring supply-load balance (Fig. 3-bottom).
Although the objective is still to minimize the operation cost
during the grid-connected mode, this scheduling is primarily
based on the grid flexibility requirements (hence the term
flexibility-oriented scheduling).
4
operation. Microgrid DERs, loads, and the main grid power
transfer are further subject to operation and flexibility
constraints, respectively represented by sets Os and Fs in (7)(8).
Fig. 3. Impact of the microgrid in increasing the distribution net load
variabilities (top) or capturing the variabilities (bottom).
III. FLEXIBILITY-ORIENTED MICROGRID SCHEDULING –
PROBLEM FORMULATION
The microgrid optimal scheduling problem aims at
determining the least-cost schedule of available resources
(DERs and loads) while considering prevailing operational
constraints, i.e.,
min [ Fi ( Pitk 0 ) tM PtkM0 ] s LStks
t
k
iG
t
k
(4)
s
Subject to
Pitks PtksM LStks Ddtks
A. Operation Constraints (Os)
The microgrid components to be modeled in the optimal
scheduling problem include DERs (i.e., generation units and
energy storage) and loads. Microgrid loads are categorized into
two types of fixed (which cannot be altered and must be
satisfied under normal operation conditions) and adjustable
(which are responsive to price variations and/or controlling
signals). Generation units in a microgrid are either dispatchable
(i.e., units which can be controlled by the microgrid controller)
or nondispatchable (i.e., wind and solar units which cannot be
controlled by the microgrid controller since the input source is
uncontrollable). The primary applications of the energy storage
are to coordinate with generation units for guaranteeing the
microgrid generation adequacy, energy shifting, and islanding
support. From these microgrid components, only dispatchable
DGs, energy storage, and adjustable loads can provide
flexibility benefits for the microgrid due to their controllability.
Microgrid component constraints are formulated as follows:
(9)
Pi min I it Pitks Pi max I it
i G, t , k , s,
(10)
Pitks Pit ( k 1) s URi
i G, t , s, k 1,
t , k , s,
(5)
M
P M , max wtks Ptks
P M , max wtks
t , k , s,
(6)
Pit1s Pi (t 1) Ks URi
i G, t , k , s,
(11)
{Pitks , Ddtks} O s
i, t , k , s,
(7)
Pit ( k 1) s Pitks DRi
i G, t , s, k 1,
(12)
M
Ptks
Fs
t , k , s.
Pi (t 1) Ks Pit1s DRi
i G, t , k , s,
(13)
i G, t ,
(14)
Tioff DTi ( I i (t 1) I it )
i G, t ,
(15)
dch,max
ch,min
Pitks Pitk
uit Pitk
vit
i S, t , k , s,
(16)
ch,max
Pitk
vit
i S, t , k , s,
(17)
i S, t ,
(18)
i
d
(8)
The objective (4) minimizes the microgrid daily operation
cost, which includes the local generation cost, cost of energy
transfer with the utility grid, and the outage cost. The outage
cost (also known as the cost of unserved energy) is defined as
the load curtailment times the value of lost load (VOLL). The
VOLL represents the customers’ willingness to pay for reliable
electricity service and to avoid power outages, which can also
be perceived as the energy price for compensating curtailed
loads. The VOLL depends on the type of customers, time and
duration of outage, time of advanced notification of outage, and
other specific traits of an outage. The VOLL is generally
considered between $0/MWh and $17,976/MWh for residential
customers, while for commercial and industrial customers is
estimated between $3,000/MWh and $53,907/MWh [43, page
7]. The load balance equation (5) ensures that the sum of the
injected/withdrawn power from the utility grid and local DERs
(i.e., dispatchable units, nondispatchable units, and the
distributed energy storage) would match the microgrid load.
The load curtailment variable is used to ensure a feasible
solution in the islanded operation if adequate generation is not
available. The power of energy storage can be negative
(charging), positive (discharging) or zero (idle). Since the
power can be exchanged between the utility grid and the
𝑀
microgrid, 𝑃𝑡𝑘𝑠
can be positive (power import), negative (power
export) or zero. The power transfer with the utility grid is
limited by (6). The binary islanding parameter (which is 1 when
grid-connected and 0 when islanded) ensures that the microgrid
interacts with the utility grid only during the grid-connected
Ti
on
UTi ( I it I i (t 1) )
Pitks
dch,min
Pitk
uit
uit vit 1
Citks Cit ( k 1) s ( Pitksuit / i ) Pitksvit
i S, t , s, k 1,
(19)
Cit1s Ci (t 1) Ks ( Pit1s uit / i ) Pit1s vit
i S, t , s,
(20)
i S, t , k , s,
(21)
Titch MCi (uit ui (t 1) )
i S, t ,
(22)
Titdch MDi (vit vi (t 1) )
i S, t ,
(23)
Ddmin zdtk Ddtks Ddmax zdtk
Tdon MU d ( z dt z d (t 1) )
d D, t , k , s,
(24)
d D, t ,
(25)
d D, s.
(26)
Cimin
[ , ]
Citks
Cimax
Ddtks Ed
Constraint (9) represents the maximum and minimum
generation capacity of dispatchable units. The binary variable I
represents the unit commitment state which would be one when
the unit is committed and zero otherwise. Dispatchable
generation units are also subject to ramp up and ramp down
5
constraints which are defined by (10)-(13). Equations (10) and
(12) represent the ramping constraints for intra-hour intervals,
while (11) and (13) represent the ramping constraint for interhour intervals. The minimum up and down time limits are
imposed by (14) and (15) respectively. The minimum and
maximum limits of the energy storage charging and
discharging, based on the operation mode, are defined by (16)
and (17), respectively. While charging, the binary charging
state v is one and the binary discharging state u is zero; while in
the discharging mode, the binary charging state v is zero and the
binary discharging state u is one. The energy storage charging
power is a negative value which is compatible with the negative
amount for limitations of constraints (16) and (17) for the
charging mode. Only one of the charging or discharging modes
at every time period is possible, which is ensured by (18). The
energy storage stored energy is calculated based on the
available stored energy and the amount of charged/discharged
power, which is represented in (19) and (20) for intra-hour and
inter-hour intervals, respectively. The time period of charging
and discharging is considered to be τ=(1/K)h, where K is the
number of intra-hour periods and h represents a time period of
one hour. The amount of stored energy in energy storage is
restricted with its capacity (21). The minimum charging and
discharging times are represented in (22) and (23), respectively.
Adjustable loads are subject to minimum and maximum rated
powers (24), where binary operating state z is 1 when load is
consuming power and 0 otherwise. The minimum operating
time (25), and the required energy to complete an operating
cycle (26) are further considered for adjustable loads. It is worth
mentioning that t=0, which would appear in (3), (14), (15), (22),
and (23), represents the last hour of the previous scheduling
horizon, here t=24.
B. Flexibility Constraints (Fs)
Flexibility constraints represent additional limits on the
microgrid power exchange with the utility grid. These
constraints are defined in a way that the microgrid net load is
matched with the aggregated net load of connected
prosumers/consumers, so as to capture likely variations. To
u
obtain the flexibility constraints, the value of Ptks
, i.e.,
u
c , is substituted in (2) and (3). By proper
Ptks
PtksM Pjtks
jN
rearrangements, the inter-hour and intra-hour flexibility
constraints will be accordingly obtained as in (27) and (28):
c
M
1 ( Pjtk
Pjtc ( k 1) ) ≤ Ptks
Pt M
( k 1) s
j
(27)
j
c
≤ 1 ( Pjtk
Pjtc ( k 1) )
j
t , s, k 1,
j
2 ( Pjtc1 Pjc(t 1) K ) ≤ Pt1Ms P(M
t 1) Ks
j
(28)
j
≤ 2 ( Pjtc1 Pjc(t 1) K )
j
t , s.
j
Accordingly, new time-dependent flexibility limits can be
defined as follows
c
c
(29)
low
t , k 1,
1,tk 1 ( Pjtk Pjt ( k 1) )
j
j
c
c
up
1,tk 1 ( Pjtk Pjt ( k 1) )
j
c
c
low
2,t 2 ( Pjt1 Pj (t 1) K )
j
(30)
t ,
(31)
t.
(32)
j
c
c
up
2, t 2 ( Pjt1 Pj (t 1) K )
j
t , k 1,
j
j
These new constraints convert the required flexibility by the
grid operator to a limit on the microgrid net load. Although
utility grid flexibility limits, i.e., Δ1 and Δ2, are constant and
determined by the grid operator, the limits on the microgrid net
load are highly variable as they comprise the aggregated net
load of all N customers in the distribution feeder. Depending on
the considered time resolution for forecasts, these limits can
change from every 1 minute to every 1 hour in the scheduling
horizon. The flexibility limits can be adjusted by the grid
operator to achieve the desired net load in the distribution
network. For example, a value of zero for Δ1 would eliminate
intra-hour variations.
It is worth mentioning that connected prosumers/consumers
are considered as given parameters (forecasted) in the
optimization problem. There will be no direct communications
between
the
microgrid
and
the
connected
prosumers/consumers, where all communications will be
through the grid operator. Therefore, the microgrid only
communicates with the grid operator and sends/receives the
required data for capturing and mitigating the distribution
network net load variabilities.
C. Islanding Considerations
The islanding is performed to rapidly disconnect the
microgrid from a faulty distribution network, safeguard the
microgrid components from upstream disturbances, and protect
voltage sensitive loads when a quick solution to utility grid
voltage problems is not imminent. The time and the duration of
such disturbances, however, are not known to microgrids in
advance. Islanding is considered in this paper via a Θ-k
islanding criterion, where Θ(=T×K) represents the total number
of intra-hour time periods in the scheduling horizon and k
represents the number of consecutive intra-hour periods that the
microgrid should operate in the islanded mode. To apply this
criterion to the proposed model, the binary islanding indicator
w is defined and added to the microgrid power exchange
constraint (6). Several scenarios are defined based on the
number of intra-hour time periods (for instance 144 scenarios
for 10-minute intra-hour periods), and the value of w in each
scenario is obtained based on the Θ-k islanding criterion, i.e., in
each scenario w will be 0 for k consecutive intra-hour time
periods (imposing an islanded operation) and 1 in other periods
(representing the grid-connected operation). Fig. 4 shows the
first five islanding scenarios, from a total of 144 scenarios,
associated with a Θ-4 islanding criterion, which requires that
the microgrid be able to operate in the islanded mode for any 4
consecutive intra-hour periods once it is switched to the
islanded mode. Further discussions on the Θ-k islanding
criterion can be found in [23]. It should be noted that the
proposed model is generic and can be applied to any microgrid
size without loss of generality.
6
t
k
Scenario 1
Scenario 2
Scenario 3
Scenario 4
Scenario 5
1
1
0
1
1
1
1
2
0
0
1
1
1
3
0
0
0
1
1
2
4
0
0
0
0
1
5
1
0
0
0
0
6
1
1
0
0
0
1
1
1
1
0
0
2
1
1
1
1
0
3
1
1
1
1
1
3
4
1
1
1
1
1
5
1
1
1
1
1
6
1
1
1
1
1
1
1
1
1
1
1
2
1
1
1
1
1
3
1
1
1
1
1
4
1
1
1
1
1
5
1
1
1
1
1
6
1
1
1
1
1
...
...
...
...
...
...
Fig. 4. First five islanding scenarios associated with a Θ-4 islanding criterion.
IV. NUMERICAL SIMULATIONS
A microgrid with four dispatchable units, two
nondispatchable units including wind and solar, one energy
storage, and five adjustable loads is used to study the
performance of the proposed model. The characteristics of the
microgrid DERs and loads, and the hourly market price are
borrowed from [23]. The maximum ramping capability of the
microgrid, based on the maximum ramping capacity of DERs,
is 18 MW/h and the capacity of the line connecting the
microgrid to the distribution feeder is assumed to be 10 MW. A
VOLL of $10,000/MWh is considered for the microgrid.
The aggregated consumption profile of consumers/prosumers
connected to the system in the same feeder as the microgrid is
shown in Fig. 5. This figure consists of aggregated values for
the distributed solar generation, consumption, and the net load
(i.e., difference between the local consumption and generation).
The net load should be supplied by the utility grid, and as the
figure demonstrates, it includes considerable variabilities due to
the local solar generation. The maximum ramping of this net
load is 3.3 MW/10-min and the peak net load is 12.9 MW. This
net load variability should be satisfied by either fast response
units deployed by the utility or locally by the microgrid, where
the latter is discussed here. The proposed flexibility-oriented
microgrid optimal scheduling model is developed using mixedinteger programming and solved using CPLEX 12.6. It should
be noted that the computation time for the studied cases was
between 3 and 4 minutes, with an average of 3 min and 22 s.
that the unit is not committed. The energy storage charging,
discharging, and idle states are represented by -1, 1, and 0,
respectively. The bold values represent changes in the schedule
due to the islanding requirements. Dispatchable unit 1 has the
lowest operation cost, so it is committed in all scheduling hours,
while other units are committed and dispatched when required
based on economic and reliability considerations. It should be
noted that the amount of load curtailment during the islanded
operation is considered as a measure of microgrid reliability.
The energy storage is charged in low price hours and discharged
in high price hours, i.e., an energy arbitrage, to maximize the
benefits and minimize the operation cost. As the table shows,
the islanding criterion leads to the commitment of more units in
the grid-connected mode to guarantee a seamless islanding.
TABLE I
DER SCHEDULE IN CASE 1
Hours (1-24)
G1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
G2
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
G3
1 1 1 1 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
G4
0 0 0 0 0 0 0 0 0 0 0 1 1 1 0 1 1 1 1 1 1 1 0 0
DES
-1 -1 -1 -1 -1 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 0 0 0 0
Fig. 6 depicts the microgrid net load and the distribution
feeder net load (i.e., the microgrid net load plus the aggregated
consumer/prosumer net load in Fig. 5). As this figure shows,
the microgrid imports the power from the utility grid in low
price hours and switches over to local generation when the
utility grid price is high. This scheduling causes a 21.58 MW
peak load for the utility grid between hours 9 and 10 (that is a
new morning peak), and also exacerbates the distribution feeder
ramping requirement (which is increased to 8.9 MW/10-min
between hours 11 and 12 in this case). In addition, the net load
variability is significantly increased in this case.
Fig. 5. Aggregated prosumers solar generation, consumption, and the net load
in the distribution feeder.
Fig. 6. Distribution feeder net load, and microgrid net load for the 24-hour
horizon in Case 1.
Case 1: The grid-connected, price-based optimal scheduling is
analyzed for a 24-hour horizon. The price-based scheduling
denotes that the microgrid seeks to minimize its operation cost
and does not have any commitment to support the utility grid in
capturing distribution network net load variabilities. Table I
shows the schedule of dispatchable units and the energy storage
for 24 hours of operation in this case. A commitment state of 1
represents that the dispatchable unit is ON while 0 represents
Therefore, the utility grid encounters severe net load
ramping and variations, caused by the microgrid to a great
extent. This result advocates that the microgrid can potentially
have a negative impact on the distribution network net load
when scheduled only based on the price data and economic
considerations. The microgrid operation cost in this case is
$11748.3.
7
Case 2: In this case, the flexibility-oriented microgrid optimal
scheduling is carried out, rather than the price-based
scheduling, to support the utility grid in addressing net load
variations. A Θ-1 islanding criterion with 10-min intra-hour
periods is considered. This islanding criterion ensures that the
microgrid is capable of switching to the islanded mode to
reliably supply local loads (for any 10-min islanding during the
scheduling horizon), while supporting the utility grid by
providing required flexibility during the grid-connected
operation. The flexibility limits of 0.5 MW/10-min are
considered for inter-hour and intra-hour ramping. The intrahour and inter-hour ramping constraints are accordingly
developed, as proposed in (27)-(32) and added to the developed
model. Table II shows the schedule of dispatchable units and
the energy storage for the scheduling horizon. The bold values
represent changes in the schedule, while the highlighted cells
represent changes in the dispatched power compared to Case 1.
This table shows that the commitment of unit 4 and the energy
storage, as well as the dispatched power of all DERs, are
changed compared to Case 1 to satisfy the flexibility
constraints. These changes in the schedules increase the
microgrid operation cost to $12077. The difference between
this cost and the microgrid operation cost in Case 1 should be
paid to the microgrid, as a minimum, to incentivize the
microgrid for providing flexibility and supporting the utility
grid. Fig. 7 shows the distribution feeder net load and the
microgrid net load in this case.
TABLE II
DER SCHEDULE IN CASE 2
Hours (1-24)
G1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
G2
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
G3
1 1 1 1 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
G4
0 0 0 0 0 0 0 0 0 1 1 0 1 1 1 1 1 1 1 1 1 1 0 0
DES
0 -1 -1 -1 -1 -1 -1 0 0 0 0 0 0 1 1 1 1 1 1 1 1 0 0 0
Fig. 7. Distribution feeder net load, and microgrid net load for 0.5 MW/10-min
inter-hour and intra-hour utility ramping in Case 2.
Comparison of Figs. 6 and 7 shows the positive impact of
the microgrid in changing the distribution network net load in a
way that is desirable for the utility grid. As Fig. 6 illustrates, the
distribution feeder net load, which should be supplied by the
utility grid, consists of several rampings in the order of a few
MW/10-min as well as a severe ramping of 8.9 MW/10-min
between hours 11 and 12. In Fig. 7, however, all these
variabilities are reduced to 0.5 MW/10-min as targeted by the
grid operator. Moreover, Fig. 8 depicts the ramping of the utility
grid in both studied cases. This figure clearly demonstrates the
effectiveness of the proposed model in reducing the distribution
network net load ramping, as the obtained data from Case 2 is
efficiently confined between the desired ramping values.
Fig. 8. Utility grid net load ramping in the two studied cases.
The results in Case 2 advocate that to obtain the desired
ramping the microgrid needs to deviate from its price-based
schedule. This deviation results in a $328.7 increase in the
microgrid operation cost (i.e., $12077–$11748.3). This increase
represents the microgrid lost revenue. To incentivize the
microgrid to opt in for offering flexibility services to the utility
grid, the amount of incentive that should be paid to the
microgrid must be equal to or greater than this amount. If less,
the microgrid would prefer to find its price-based schedule
while disregarding the grid requirements. However, it would be
extremely beneficial for the utility grid to incentivize the
microgrid, otherwise the microgrid may exacerbate the
distribution network net load variability as discussed in Case 1.
It is worth mentioning that the microgrid lost revenue is a
function of the consumers/prosumers net load variations as well
as values of Δ1 and Δ2 which are further investigated in the
following.
Case 3: After proving the effectiveness of the proposed model
by comparing Cases 1 and 2, the impact of ramping limits is
studied in this case. To show that the microgrid is also capable
of meeting tight ramping limits, a value of zero is considered
for the intra-hour ramping and 2 MW/10-min for the inter-hour
ramping. Fig. 9 depicts the solution of this case. Considering a
value of zero for intra-hour ramping completely eliminates the
intra-hour variabilities in the distribution network net load,
hence the obtained consumption is constant within each
operation hour while it can change by up to 2 MW between any
two consecutive operation hours.
To closely follow the limits, the microgrid imported power
from the utility grid is decreased when the net load is increasing.
Furthermore, microgrid’s export to the utility grid in high price
hours is changed to support the ramping limits. For instance the
microgrid power export to the utility grid in hours 12-14, which
was based on economic considerations, is now changed to
power import from the utility grid. Fig. 10 shows the obtained
results of Fig. 9 between hours 12 and 20, which better
8
demonstrates the viable application of the microgrid in reducing
the net load variability and sharp ramping.
Fig. 9. Distribution feeder net load, and microgrid net load for 2 MW/10-min
inter-hour and 0 MW/10-min intra-hour utility ramping.
is due to two main reasons: (i) the need to commit more units
and dispatch them at uneconomical operation points, in a way
that they can provide the required flexibility, and (ii) the
possibility of load curtailment in the microgrid. The ramping
limits are added as constraints to the problem, while the load
curtailment is added as a penalty to the objective function. It
results in prioritizing the flexibility limit (i.e., problem
feasibility) on the load curtailment (i.e., problem optimality).
There of course should be additional measures to consider in
order to prevent load curtailment in the microgrid which are
currently under investigation by the authors. The utility grid
incentive in each case must at least cover the microgrid’s lost
revenue. According to Table III, if the utility grid decides to
eliminate the intra-hour ramping, it should pay at least
$24,557.3 and $697.3 to the microgrid for Δ2 values equal to
0.5 MW/10-min and 5 MW/10-min, respectively. Whereas, in
the case of 2 MW/10-min as desired intra-hour ramping, at least
$77 and $21.8 should be paid to the microgrid for Δ2 equal to
0.5 MW/10-min and 5 MW/10-min, respectively. These results
advocate for the importance of a cost-benefit analysis from the
grid operator to determine the most suitable inter-hour and
intra-hour ramping limits.
V. DISCUSSIONS
Fig. 10. Distribution feeder net load, and microgrid net load for 2 MW/10-min
inter-hour and 0 MW/10-min intra-hour utility ramping, during net load peak
hours.
The results of flexibility-oriented microgrid optimal scheduling
for different amounts of inter-hour (changing between 0.5 and
5) and intra-hour (changing between 0 and 2) ramping limits are
provided in Table III. It should be noted that all obtained results
are near-optimal, mainly due to nonlinearity of the original
problem and presence of uncertainties.
TABLE III
MICROGRID OPERATION COST ($) FOR VARIOUS RAMPING LIMITS
Intra-hour ramping limit Δ1
Inter-hour
(MW/10 min)
ramping limit
Δ2 (MW/10min)
0
0.5
1
2
0.5
36305.6
12077
11886.9
11825.3
1
19799.1
12011.5
11860.1
11804
1.5
14930.4
11977.5
11845.2
11796.5
2
13329.1
11951.4
11834.1
11790.1
2.5
12790.2
11936.4
11826.4
11786
3
12607.5
11925.8
11819.7
11782.1
3.5
12532.5
11916.8
11813.7
11778.6
4
12485.9
11906.8
11808.6
11775.6
4.5
12460.1
11898.3
11804.3
11772.7
5
12445.6
11891
11800.1
11770.1
The obtained results show that the microgrid operation cost
is increased by decreasing the inter-hour and intra-hour
ramping limits, however these changes are not linear. For
example, the microgrid operation cost when the intra-hour
ramping limit is 0 is considerably higher than other cases. This
Microgrids can potentially be utilized in distribution
networks as a solution for mitigating net load ramping and
variability. According to the studied cases in this paper, the
following features of the proposed microgrid optimal
scheduling model with multi-period islanding and flexibility
constraints, could be concluded:
Flexibility consideration: The inter-hour and intra-hour
ramping constraints have been considered in the proposed
model to ensure that the utility grid desired power is
obtained for different time resolutions.
Economic and reliable operation: The proposed model
determines the least-cost schedule of microgrid loads and
DERs while supporting the utility grid in addressing net
load ramping. In addition, the consideration of Θ-k
islanding criterion ensures the microgrid reliability in
supplying local loads during the islanded mode.
High resolution scheduling: 10-minute time interval
scheduling was considered in studied cases, which offers a
high resolution scheduling and is efficient for capturing net
load variabilities. The proposed model offers the capability
to consider various intra-hour time resolutions.
Localized and low-cost solution: Using microgrids as local
solutions for addressing distribution net load ramping can
significantly reduce the utility grid investments in
upgrading the generation, transmission, and distribution
facilities. This significant cost saving would be made
possible at the small expense of incentivizing microgrids to
offer flexibility services.
VI. CONCLUSION
In this paper, the microgrid was utilized to reduce the
distribution network net load variabilities, resulted primarily
due to simultaneous decrease in solar generation and increase
9
in consumers’ loads. A flexibility-oriented microgrid optimal
scheduling model was proposed to efficiently schedule
microgrid resources for supporting the distribution grid
flexibility requirements. These flexibility requirements were
considered in terms of net load ramping limits. The model was
studied for intra-hour and inter-hour time intervals during the
24-hour day-ahead operation. The Θ-k islanding criterion was
further taken into account to ensure that the microgrid has the
capability to switch to the islanded mode, if needed, while
supporting the utility grid during the grid-connected operation.
Numerical simulations were carried out for various amounts of
utility grid’s desired inter-hour and intra-hour ramping to show
the merits and the effectiveness of the proposed model. The
results showed that the grid operator can efficiently leverage the
flexibility of existing microgrids in distribution networks to
address some of the most pressing flexibility-associated
challenges, while removing the need for costly investments in
the generation and distribution facilities.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
U.S. Energy Information Administration, "Annual Energy Review,"
[Online].
Available:
https://www.eia.gov/totalenergy/data/annual/previous.cfm.
G. Barbose, N. Darghouth, and S. Weaver, Lawrence Barkley National
Laboratory (LBNL), “Tracking the Sun VII: An Historical Summary of
the Installed Price of Photovoltaics in the United States from 1998 to
2013,” September 2014.
California ISO, “What the duck curve tells us about managing a green
grid,”.
[Online].
Available:
https://www.caiso.com/Documents/FlexibleResourcesHelpRenewables_
FastFacts.pdf.
A. Costa, A. Crespo, J. Navarro, G. Lizcano, H. Madsen, and E. Feitosa,
“A review on the young history of the wind power short-term prediction,”
Renew. Sustain. Energy Rev., vol. 12, no. 6, pp. 1725–1744, 2008.
A. Ulbig, G. Andersson, "On operational flexibility in power systems," in
Proc. 2012 IEEE Power and Energy Society General Meeting, pp. 1-8.
E. Lannoye, D. Flynn, and M. O’Malley, “Evaluation of power system
flexibility,” IEEE Trans. Power Syst., vol. 27, no. 2, pp. 922–931, May
2012.
D.S. Kirschen, A. Rosso, J. Ma, L.F. Ochoa, "Flexibility from the demand
side," in Proc. IEEE Power and Energy Society General Meeting, San
Diego, CA, 22-26 Jul. 2012, pp. 1-6.
M. Ghamkhari and H. Mohsenian-Rad, “Optimal integration of renewable
energy resources in data centers with behind-the-meter renewable
generator,” in Proc. IEEE International Conference on Communications
(ICC), Ottawa, ON, 10-15 Jun. 2012, pp. 3340-3344.
A. Asadinejad and K. Tomsovic “Impat of incetive based demand
response on large scale renewable integration,” 2016 IEEE PES
Innovative Smart Grid Technologies (ISGT), Minneapolis, MN, 6-9 Sep.
2016.
S. C. Chan, K. M. Tsui, H. C. Wu, Y. Hou, Y. Wu, and F. F. Wu,
“Load/price forecasting and managing demand response for smart grids:
methodologies and challenges,” IEEE Signal Process. Mag., vol. 29, no.
5, pp. 68–85, Sep. 2012.
O. Ma, N. Alkadi, P. Cappers, P. Denholm, J. Dudley, S. Goli, M.
Hummon, S. Kiliccote, J. MacDonald, N. Matson, D. Olsen, C. Rose, M.
D. Sohn, M. Starke, B. Kirby, and M. O’Malley, “Demand Response for
Ancillary Services,” IEEE Trans. Smart Grid, vol. 4, no. 4, pp. 1988–
1995, Dec. 2013.
J. Aghaei and M. I. Alizadeh, “Demand response in smart electricity grids
equipped with renewable energy sources: A review,” Renew. Sustain.
Energy Rev., vol. 18, pp. 64–72, 2013.
P. Denholm, E. Ela, B. Kirby, and M. Milligan, “The role of energy
storage with renewable electricity generation,” National Renewable
Energy Laboratory, Tech. Rep. NREL/TP-6A2-47187, Jan. 2010,
Available: http://www.nrel.gov/docs/fy10osti/47187.pdf
J. Eyer and G. Corey, “Energy Storage for the Electricity Grid : Benefits
and Market Potential Assessment Guide; A Study for the DOE Energy
Storage Systems Program,” Sandia National Laboratories, 2010.
[15] M. Beaudin, H. Zareipour, A. Schellenberglabe, and W. Rosehart,
“Energy storage for mitigating the variability of renewable electricity
sources: An updated review,” Energy Sustain. Dev., vol. 14, no. 4, pp.
302–314, 2010.
[16] P. Denholm and M. Hand, “Grid flexibility and storage required to
achieve very high penetration of variable renewable electricity,” Energy
Policy, vol. 39, no. 3, pp. 1817–1830, 2011.
[17] K. Bradbury, L. Pratson, and D. Patino-Echeverri, “Economic viability of
energy storage systems based on price arbitrage potential in real-time U.S.
electricity markets,” Appl. Energy, vol. 114, pp. 512–519, 2014.
[18] S. Gottwalt, A. Schuller, C. Flath, H. Schmeck, and C. Weinhardt,
“Assessing load flexibility in smart grids: Electric vehicles for renewable
energy integration,” in Proc. 2013 IEEE Power and Energy Society
General Meeting, Vancouver, BC, 21-25 Jul. 2013, pp. 1-5.
[19] S.-I. Inage. (2010). “Modelling load shifting using electric vehicles in a
smart grid environment,” IEA/OECD Working Paper. [Online].
Available:
http://www.iea.org/publications/freepublications/
publication/load_shifting.pdf
[20] A. Majzoobi and A. Khodaei, “Application of Microgrids in Addressing
Distribution Network Net-Load Ramping,” IEEE PES Innovative Smart
Grid Technologies Conference (ISGT), Minneapolis, MN, 6-9 Sep. 2016.
[21] Department of Energy Office of Electricity Delivery and Energy
Reliability, “Summary Report: 2012 DOE Microgrid Workshop,” 2012.
[Online]. Available: http://energy.gov/sites/prod/files/2012 Microgrid
Workshop Report 09102012.pdf.
[22] A. Majzoobi and A. Khodaei “Leveraging microgrids for capturing
uncertain distribution network net load ramping,” in North American
Power Symposium (NAPS), Denver, CO, 18-20 Sep. 2016.
[23] A. Khodaei, “Microgrid optimal scheduling with multi-period islanding
constraints,” IEEE Trans. Power Syst., vol. 29, no. 3, pp. 1383–1392, May
2014.
[24] A. Khodaei, “Resiliency-oriented microgrid optimal scheduling,” IEEE
Trans. Smart Grid, vol. 5, no. 4, pp. 1584–1591, July 2014.
[25] A. Majzoobi, A. Khodaei, S. Bahramirad and M. H. J. Bollen, "Capturing
the Variabilities of Distribution Network Net-Load via Available
Flexibility of Microgrids,” CIGRE Grid of the Future Symposium,
Philadelphia, PA, 30 Oct.-1 Nov. 2016.
[26] A. Khodaei, “Provisional Microgrid Planning,” IEEE Trans. Smart Grid,
In press, 2015.
[27] F. Wu, X. Li, F. Feng and H. B. Gooi, “Modified Cascaded Multilevel
Gird-Connected Inverter to Enhance European Efficiency,” IEEE Trans.
Ind. Informatics, vol. 11, no. 6, pp.1358-1365, Dec. 2015.
[28] M. J. Sanjari, A. H. Yatim and G. B. Gharehpetian, “Online Dynamic
Security Assessment of Microgrids before Intentional Islanding
Occurrence,” Neural Comput. & Applic., vol. 26, no. 3, pp.659-668, Oct.
2014.
[29] M. Shahidehpour and J. F. Clair, “A Functional Microgrid for Enhancing
Reliability, Sustainability, and Energy Efficiency,” The Electr. J., vol. 25,
no. 8, pp. 21–28, 2012.
[30] Microgrid Exchange Group, “DOE Microgrid Workshop Report,” 2011.
[Online].
Available:
http://energy.gov/oe/downloads/microgridworkshop-report-august-2011.
[31] Navigant Research, “Microgrid Deployment Tracker 4Q13;
Commercial/Industrial,
Community/Utility,
Institutional/Campus,
Military, and Remote Microgrids: Operating, Planned, and Proposed
Projects by World Region,” 2013.
[32] S. Parhizi, H. Lotfi, A. Khodaei, and S. Bahramirad, “State of the art in
research on microgrids: A review,” IEEE Access, vol. 3, pp. 890–925, Jul.
2015.
[33] W. Pei, Y. Du, W. Deng, K. Sheng, H. Xiao, and H. Qu, “Optimal Bidding
Strategy and Intramarket Mechanism of Microgrid Aggregator in RealTime Balancing Market,” IEEE Trans. Ind. Informatics, vol. 12, no. 2, pp.
587–596, Apr. 2016.
[34] S. I. Vagropoulos, S. Member, A. G. Bakirtzis, and S. Member, “Optimal
Bidding Strategy for Electric Vehicle Aggregators in Electricity,” vol. 28,
no. 4, pp. 4031–4041, Nov. 2013.
[35] H. Kim and M. Thottan, “A two-stage market model for microgrid power
transactions via aggregators,” Bell Labs Tech. J., vol. 16, pp. 101–107,
2011.
[36] D. T. Nguyen and L. B. Le, “Risk-constrained profit maximization for
microgrid aggregators with demand response,” IEEE Trans. Smart Grid,
vol. 6, no. 1, pp. 135–146, Jan. 2015.
[37] M. J. Sanjari, H. Karami, and H. B. Gooi, “Micro-generation dispatch in
a smart residential multi-carrier energy system considering demand
forecast error,” Energy Convers. Manag., vol. 120, pp. 90–99, 2016.
10
[38] L. Shi, Y. Luo, and G. Y. Tu, “Bidding strategy of microgrid with
consideration of uncertainty for participating in power market,” Electr.
Power Energy Syst., vol. 59, pp. 1–13, 2014.
[39] Y. Gu and L. Xie, “Stochastic Look-Ahead Economic Dispatch With
Variable Generation Resources,” IEEE Trans. Power Syst., In press, Apr.
2016.
[40] B. A. Frew, S. Becker, M. J. Dvorak, G. B. Andersen, and M. Z. Jacobson,
“Flexibility mechanisms and pathways to a highly renewable US
electricity future,” Energy, vol. 101, pp. 65–78, 2016.
[41] Y. Dvorkin, M. A. Ortega-Vazquez, and D. S. Kirschen, “Assessing
flexibility requirements in power systems,” IET Gener., Transm., Distrib,
vol. 8, no. 11, pp. 1820-1830, Nov. 2014.
[42] J. Ma, “Evaluating and planning flexibility in a sustainable power system
with large wind penetration,” Ph.D. dissertation, Dept. Electrical and
Electronic Eng., Univ. Manchester, Manchester, 2012
[43] Estimating the Value of Lost Load [Online]. Available
http://www.ercot.com/content/gridinfo/resource/2015/mktanalysis/ERC
OT_ValueofLostLoad_LiteratureReviewandMacroeconomic.pdf
Alireza Majzoobi (S’15), received the B.Sc. degree
from Sharif University of Technology, Tehran, Iran, in
2007, the M.Sc. degree from the University of Tehran,
Tehran, Iran, in 2011, and the second M.Sc. degree
from Old Dominion University, Norfolk, VA, USA, in
2015, all in electrical engineering. He is currently
pursuing the Ph.D. degree in the Department of
Electrical and Computer Engineering at University of
Denver, Denver, CO, USA. His current research
interests include power system operation and economics, microgrids, and
integrating renewable and distributed resources.
Amin Khodaei (SM’14) received the Ph.D. degree in
electrical engineering from the Illinois Institute of
Technology, Chicago, in 2010. He was a visiting
faculty (2010–2012) in the Robert W. Galvin Center
for Electricity Innovation at Illinois Institute of
Technology. He is currently an Associate Professor
in the Department of Electrical and Computer
Engineering at University of Denver, Denver, CO,
USA. His research interests include power system
operation, planning, computational economics, microgrids, smart electricity
grids, and artificial intelligence.
| 3cs.SY
|
arXiv:1711.03842v1 [cs.PL] 9 Nov 2017
Refinement Reflection: Complete Verification with SMT
NIKI VAZOU, University of Maryland, USA
ANISH TONDWALKAR, University of California, San Diego, USA
VIKRAMAN CHOUDHURY, Indiana University, USA
RYAN G. SCOTT, Indiana University, USA
RYAN R. NEWTON, Indiana University, USA
PHILIP WADLER, University of Edinburgh and Input Output HK, UK
RANJIT JHALA, University of California, San Diego, USA
We introduce Refinement Reflection, a new framework for building SMT-based deductive verifiers. The key idea
is to reflect the code implementing a user-defined function into the function’s (output) refinement type. As a
consequence, at uses of the function, the function definition is instantiated in the SMT logic in a precise fashion
that permits decidable verification. Reflection allows the user to write equational proofs of programs just by
writing other programs e.g. using pattern-matching and recursion to perform case-splitting and induction.
Thus, via the propositions-as-types principle, we show that reflection permits the specification of arbitrary
functional correctness properties. Finally, we introduce a proof-search algorithm called Proof by Logical
Evaluation that uses techniques from model checking and abstract interpretation, to completely automate
equational reasoning. We have implemented reflection in Liqid Haskell and used it to verify that the widely
used instances of the Monoid, Applicative, Functor, and Monad typeclasses actually satisfy key algebraic
laws required to make the clients safe, and have used reflection to build the first library that actually verifies
assumptions about associativity and ordering that are crucial for safe deterministic parallelism.
CCS Concepts: • Software and its engineering → General programming languages; General programming languages;
Additional Key Words and Phrases: refinement types, theorem proving, Haskell, automatic verification
ACM Reference Format:
Niki Vazou, Anish Tondwalkar, Vikraman Choudhury, Ryan G. Scott, Ryan R. Newton, Philip Wadler, and Ranjit
Jhala. 2018. Refinement Reflection: Complete Verification with SMT. Proc. ACM Program. Lang. 2, POPL,
Article 53 (January 2018), 59 pages. https://doi.org/10.1145/3158141
1
INTRODUCTION
Deductive verifiers fall roughly into two camps. Satisfiability Modulo Theory (SMT) based verifiers
(e.g. Dafny and F*) use fast decision procedures to automate the verification of programs that
only require reasoning over a fixed set of theories like linear arithmetic, string, set and bitvector operations. These verifiers, however, encode the semantics of user-defined functions with
universally-quantified axioms and use incomplete (albeit effective) heuristics to instantiate those
Authors’ addresses: Niki Vazou, University of Maryland, USA, [email protected]; Anish Tondwalkar, University of
California, San Diego, USA, [email protected]; Vikraman Choudhury, Indiana University, USA; Ryan G. Scott, Indiana
University, USA; Ryan R. Newton, Indiana University, USA; Philip Wadler, University of Edinburgh and Input Output HK,
UK, [email protected]; Ranjit Jhala, University of California, San Diego, USA, [email protected].
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee
provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and
the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored.
Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires
prior specific permission and/or a fee. Request permissions from [email protected].
© 2018 Association for Computing Machinery.
2475-1421/2018/1-ART53
https://doi.org/10.1145/3158141
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53
53:2
N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
axioms. These heuristics make it difficult to characterize the kinds of proofs that can be automated,
and hence, explain why a given proof attempt fails [30]. At the other end, we have Type-Theory
(TT) based theorem provers (e.g. Coq and Agda) that use type-level computation (normalization)
to facilitate principled reasoning about terminating user-defined functions, but which require the
user to supply lemmas or rewrite hints to discharge proofs over decidable theories.
We introduce Refinement Reflection, a new framework for building SMT-based deductive verifiers,
which permits the specification of arbitrary properties and yet enables complete, automated SMTbased reasoning about user-defined functions. In previous work, refinement types [17, 44] — which
decorate basic types (e.g. Integer) with SMT-decidable predicates (e.g. {v:Integer | 0 ≤ v &&
v < 100}) — were used to retrofit so-called shallow verification, such as array bounds checking,
into several languages: ML [6, 42, 63], C [16, 43], Haskell [56], TypeScript [59], and Racket [26]. In
this work, we extend refinement types with refinement reflection, leading to the following three
contributions.
1. Refinement Reflection Our first contribution is the notion of refinement reflection. To reason
about user-defined functions, the function’s implementation is reflected into its (output) refinementtype specification, thus converting the function’s type signature into a precise description of the
function’s behavior. This simple idea has a profound consequence: at uses of the function, the
standard rule for (dependent) function application yields a precise means of reasoning about the
function (§ 4).
2. Complete Specification Our second contribution is a library of combinators that lets programmers compose sophisticated proofs from basic refinements and function definitions. Our proof
combinators let programmers use existing language mechanisms, like branches (to encode case
splits), recursion (to encode induction), and functions (to encode auxiliary lemmas), to write proofs
that look very much like their pencil-and-paper analogues (§ 2). Furthermore, since proofs are literally just programs, we use the principle of propositions-as-types [62] (known as the Curry-Howard
isomorphism [24]) to show that SMT-based verifiers can express any natural deduction proof,
thus providing a pleasant implementation of natural deduction that can be used for pedagogical
purposes (§ 3).
3. Complete Verification While equational proofs can be very easy and expressive, writing
them out can quickly get exhausting. Our third contribution is Proof by Logical Evaluation (PLE)
a new proof-search algorithm that automates equational reasoning. The key idea in PLE is to
mimic type-level computation within SMT-logics by representing functions in a guarded form [19]
and repeatedly unfolding function application terms by instantiating them with their definition
corresponding to an enabled guard. We formalize a notion of equational proof and show that the
above strategy is complete: i.e. it is guaranteed to find an equational proof if one exists. Furthermore,
using techniques from the literature on Abstract Interpretation [18] and Model Checking [14], we
show that the above proof search corresponds to a universal (or must) abstraction of the concrete
semantics of the user-defined functions. Thus, since those functions are total, we obtain the pleasing
guarantee that proof search terminates (§ 6).
We evaluate our approach by implementing refinement reflection and PLE in Liqid Haskell [56],
thereby turning Haskell into a theorem prover. Repurposing an existing programming language
allows us to take advantage of a mature compiler and an ecosystem of libraries, while keeping
proofs and programs in the same language. We demonstrate the benefits of this conversion by
proving typeclass laws. Haskell’s typeclass machinery has led to a suite of expressive abstractions
and optimizations which, for correctness, crucially require typeclass instances to obey key algebraic
laws. We show how reflection and PLE can be used to verify that widely used instances of the
Monoid, Applicative, Functor, and Monad typeclasses satisfy the respective laws. Finally, we use
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:3
reflection to create the first deterministic parallelism library that actually verifies assumptions
about associativity and ordering that ensure determinism (§ 7).
Thus, our results demonstrate that Refinement Reflection and Proof by Logical Evaluation identify
a new design for deductive verifiers which, by combining the complementary strengths of SMTand TT- based approaches, enables complete verification of expressive specifications spanning
decidable theories and user defined functions.
2
OVERVIEW
We start with an overview of how SMT-based refinement reflection lets us write proofs as plain
functions and how PLE automates equational reasoning.
2.1 Refinement Types
First, we recall some preliminaries about specification and verification with refinement types.
Refinement types are the source program’s (here Haskell’s) types refined with logical predicates
drawn from an SMT-decidable logic [17, 44]. For example, we define Nat as the set of Integer
values v that satisfy the predicate 0 ≤ v from the quantifier-free logic of linear arithmetic and
uninterpreted functions (QF-UFLIA [5]):
type Nat = { v:Integer | 0 ≤ v }
Specification & Verification Throughout this section, to demonstrate the proof features we add
to Liqid Haskell, we will use the textbook Fibonacci function which we type as follows.
fib
fib
fib
fib
:: Nat → Nat
0 = 0
1 = 1
n = fib (n-1) + fib (n-2)
To ensure termination, the input type’s refinement specifies a pre-condition that the parameter
must be Nat. The output type’s refinement specifies a post-condition that the result is also a Nat.
Refinement type checking automatically verifies that if fib is invoked with a non-negative Integer,
then it terminates and yields a non-negative Integer.
Propositions We define a data type representing propositions as an alias for unit:
type Prop = ()
which can be refined with propositions about the code, e.g. that 2 + 2 equals 4
type Plus_2_2 = { v: Prop | 2 + 2 = 4 }
For simplicity, in Liqid Haskell, we abbreviate the above to type Plus_2_2 = { 2 + 2 = 4 }.
Universal & Existential Propositions Using the standard encoding of Howard [24], known as the
Curry-Howard isomorphism, refinements encode universally-quantified propositions as dependent
function types of the form:
type Plus_comm = x:Integer → y:Integer → { x + y = y + x }
As x and y refer to arbitrary inputs, any inhabitant of the above type is a proof that Integer addition
commutes.
Refinements encode existential quantification via dependent pairs of the form:
type Int_up = n:Integer → (m::Integer, {n < m})
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:4
N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
The notation (m :: t, t') describes dependent pairs where the name m of the first element can
appear inside refinements of the second element. Thus, Int_up states the proposition that for every
integer n, there exists one that is larger than n.
While quantifiers cannot appear directly inside the refinements, dependent functions and pairs
allow us to specify quantified propositions. One limitation of this encoding is that quantifiers
cannot exist inside refinement’s logical connectives (like ∧ and ∨). In § 3, we describe how to
encode logical connectives using data types, e.g. conjunction as a product and disjunction as a
union, how to specify arbitrary, quantified propositions using refinement types, i.e. have complete
specifications, and how to verify those propositions using refinement type checking.
Proofs We prove the above propositions by writing Haskell programs, for example
plus_2_2 :: Plus_2_2
plus_2_2 = ()
plus_comm :: Plus_comm
plus_comm = \x y → ()
int_up :: Int_up
int_up = \n → (n+1,())
Standard refinement typing reduces the above to the respective verification conditions (VCs)
true ⇒ 2 + 2 = 4
∀ x, y . true ⇒ x + y = y + x
∀ n .n < n + 1
which are easily deemed valid by the SMT solver, allowing us to prove the respective propositions.
Soundness and Lazy Evaluation Readers familiar with Haskell’s lazy semantics may be concerned that “bottom”, which inhabits all types, makes our proofs suspect. Fortunately, as described
in Vazou et al. [56], Liqid Haskell, by default, checks that user defined functions provably terminate and are total (i.e. return non-bottom values, and do not throw any exceptions), which makes
our proofs sound. Liqid Haskell checks that each function is terminating using a termination
metric i.e. a natural number that decreases at each recursive call. For instance, if we generalize the
signature of fib to Integers, as shown below, then Liqid Haskell reports a termination error:
fib :: n:Integer → Integer
Liqid Haskell generates a type error in the definition of fib since in the recursive calls fib (n-1)
and fib (n-2) the arguments n-1 and n-2 cannot be proved to be non-negative and less than n.
Both these proof obligations are satisfied when the domain of fib is restricted to natural numbers.
fib :: n:Nat → Nat / [n]
The above type signature is explicitly annotated with the user specified termination metric / [n]
declaring that n is the decreasing natural number. Liqid Haskell heuristically assumes that the
termination metric is always the first argument of the function (that can be mapped to natural
numbers), thus, the above explicit termination metric can be omitted.
Not all Haskell functions terminate. The lazy annotation deactivates termination checking, e.g.
for the the diverge function shown below. Haskell terms marked as lazy could unsoundly be used
as proof terms, much like Coq’s unsoundness with respect to Admitted. For example, the following
is accepted by Liqid Haskell
lazy diverge
diverge :: x:Integer → { x = 0 }
diverge x = diverge x
Totality Checking Liqid Haskell further checks that all user-specified functions are totally
defined. For instance the below definition
fibPartial 0 = 0
fibPartial 1 = 1
generates a totality error. The above definition is completed, by GHC, with an error invocation
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:5
fibPartial 0 = 0
fibPartial 1 = 1
fibPartial _ = error "undefined"
Thus, to check totality, Liqid Haskell simply ascribes the precondition false to error
error :: { v:String | false } → a
Thus, to typecheck fibPartial, Liqid Haskell needs to prove that the call to error is dead-code,
i.e. happens under an inconsistent environment. As this check fails, Liqid Haskell generates a
totality error. As with termination checking, Liqid Haskell provides an unsound error function
unsoundError with no false precondition.
In the sequel (and in our evaluation) we assume that proof terms are generated without any
unsound uses of lazy or unsafeError. However, note that lazy, diverging functions can be soundly
verified, and diverging code can soundly co-exist with terminating proof terms: we refer the reader
to Vazou et al. [56] for details.
2.2
Refinement Reflection
Suppose we wish to prove properties about the fib function, e.g. that {fib 2 = 1}. Standard
refinement type checking runs into two problems. First, for decidability and soundness, arbitrary
user-defined functions cannot belong in the refinement logic, i.e. we cannot refer to fib in a
refinement. Second, the only specification that a refinement type checker has about fib is its type
Nat → Nat which is too weak to verify {fib 2 = 1}. To address both problems, we reflect fib
into the logic which sets the three steps of refinement reflection in motion.
Step 1: Definition The annotation creates an uninterpreted function fib :: Integer → Integer in
the refinement logic. By uninterpreted, we mean that the logical fib is not connected to the program
function fib; in the logic, fib only satisfies the congruence axiom ∀n, m. n = m ⇒ fib n = fib m.
On its own, the uninterpreted function is not terribly useful: we cannot check {fib 2 = 1} as the
SMT solver cannot prove the VC true ⇒ fib 2 = 1 which requires reasoning about fib’s definition.
Step 2: Reflection In the next key step, we reflect the definition of fib into its refinement type by
automatically strengthening the user defined type for fib to:
fib :: n:Nat → { v:Nat | v = fib n && fibP n }
where fibP is an alias for a refinement automatically derived from the function’s definition:
fibP n n == 0 ⇒ fib n = 0
∧ n == 1 ⇒ fib n = 1
∧ n > 1 ⇒ fib n = fib (n-1) + fib (n-2)
Step 3: Application With the reflected refinement type, each application of fib in the code
automatically unfolds the definition of fib once in the logic. We prove {fib 2 = 1} by:
pf_fib2 :: { fib 2 = 1 }
pf_fib2 = let { t0 = fib 0; t1 = fib 1; t2 = fib 2 } in ()
We write in bold red, f, to highlight places where the unfolding of f’s definition is important. Via
refinement typing, the above yields the following VC that is discharged by SMT, even though fib
is uninterpreted:
(fibP 0 ∧ fibP 1 ∧ fibP 2) ⇒ (fib 2 = 1)
The verification of pf_fib2 relies merely on the fact that fib is applied to (i.e. unfolded at) 0, 1, and
2. The SMT solver automatically combines the facts, once they are in the antecedent. Thus, the
following is also verified:
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:6
N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
pf_fib2' :: {v:[Nat] | fib 2 = 1 }
pf_fib2' = [ fib 0, fib 1, fib 2 ]
In the next subsection, we will continue to use explicit, step-by-step proofs as above, but we will
introduce tools for proof composition. Then, in § 2.4 we will show how to eliminate unnecessary
details from such proofs, using Proof by Logical Evaluation (PLE).
2.3
Equational Proofs
We can structure proofs to follow the style of calculational or equational reasoning popularized in
classic texts [8, 20] and implemented in Agda [34] and Dafny [31]. To this end, we have developed
a library of proof combinators that permits reasoning about equalities and linear arithmetic.
“Equation” Combinators We equip Liqid Haskell with a family of equation combinators, ⊙, for
logical operators in the theory QF-UFLIA, ⊙ ∈ {=, ,, ≤, <, ≥, >}. In Haskell code, to avoid collisions
with existing operators, we further append a colon “:” to these operators, so that “=” becomes the
Haskell operator (=:). The refinement type of ⊙ requires that x ⊙ y holds and then ensures that the
returned value is equal to x. For example, we define (=:) as:
(=:) :: x:a → y:{ a | x = y } → { v:a | v = x }
x =: _ = x
and use it to write the following equational proof:
fib2_1 :: { fib 2 = 1 }
fib2_1 = fib 2 =: fib 1 + fib 0 =: 1 ** QED
where ** QED constructs proof terms by casting expressions to Prop in a post-fix fashion.
data QED = QED
(**) :: a → QED → Prop
_ ** QED = ()
Proof Arguments Often, we need to compose lemmas into larger theorems. For example, to prove
fib 3 = 2 we may wish to reuse fib2_1 as a lemma. We do so by defining a variant of (=:), written
as (=?), that takes an explicit proof argument:
(=?) :: x:a → y:a → { Prop | x = y } → { v:a | v = x }
x =? _ _ = x
We use the (=?) combinator to prove that fib 3 = 2.
fib3_2 :: { fib 3 = 2 }
fib3_2 = (fib 3 =: fib 2 + fib 1 =? 2 $ fib2_1) ** QED
Here fib 2 is not important to unfold, because fib2_1 already provides the same information.
“Because” Combinators Observe that the proof term fib3_2 needs parentheses, since Haskell’s
($) operator has the lowest (i.e. 0) precedence. To omit parentheses in proof terms, we define a
“because” combinator that operates exactly like Haskell’s ($) operator, but has the same precedence
as the proof combinators (=:) and (=?).
(∵) :: (Prop → a) → Prop → a
f ∵ y = f y
We use the “because” combinator to remove the parentheses from the proof term fib3_2.
fib3_2 :: { fib 3 = 2 }
fib3_2 = fib 3 =: fib 2 + fib 1 =? 2 ∵ fib2_1 ** QED
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:7
Optional Proof Arguments Finally, we unify both combinators (=:) and (=?) using type classes to
define the class method (=.) that takes an optional proof argument, and generalize such definitions
for each operator in ⊙. We define the class OptEq to have one method (=.) that takes two arguments
of type a, to be compared for equality, and returns a value of type r.
class OptEq a r where
(=.) :: a → a → r
When instantiating the result type to be the same as the argument type a, (=.) behaves as (=:).
instance (a~b) ⇒ OptEq a b where
(=.) :: x:a → { y:a | x = y } → { v:b | v = x }
When the result type is instantiated to be a function that given a proof term Prop returns the
argument type a, (=.) behaves exactly as (=?).
instance (a~b) ⇒ OptEq a (Prop → b) where
(=.) :: x:a → y:a → { Prop | x = y } → { v:a | v = x }
Thus, (=.) takes two arguments to be compared for equality and, optionally, a proof term
argument. With this, the proof term fib3_2 is simplified to
fib3_2 :: { fib 3 = 2 }
fib3_2 = fib 3 =. fib 2 + fib 1 =. 2 ∵ fib2_1 ** QED
Arithmetic and Ordering Next, lets see how we can use arithmetic and ordering to prove that
fib is (locally) increasing, i.e. for all n, fib n ≤ fib (n + 1).
type Up f = n:Nat → { f n ≤ f (n + 1) }
fibUp
:: Up fib
fibUp 0 = fib 0 <. fib 1
** QED
fibUp 1 = fib 1 ≤. fib 1 + fib 0
=. fib 2
** QED
fibUp n = fib n ≤. fib n + fib (n-1) =. fib (n+1) ** QED
Case Splitting The proof fibUp works by splitting cases on the value of n. In the cases 0 and 1, we
simply assert the relevant inequalities. These are verified as the reflected refinement unfolds the
definition of fib at those inputs. The derived VCs are (automatically) proved as the SMT solver
concludes 0 < 1 and 1 + 0 ≤ 1 respectively. When n is greater than one, fib n is unfolded to
fib (n-1) + fib (n-2), which, as fib (n-2) is non-negative, completes the proof.
Induction & Higher Order Reasoning Refinement reflection smoothly accommodates induction
and higher-order reasoning. For example, let’s prove that every function f that increases locally
(i.e. f z ≤ f (z+1) for all z) also increases globally (i.e. f x ≤ f y for all x < y)
type Mono = f:(Nat → Integer) → Up f → x:Nat → y:{ x < y } → { f x ≤ f y }
fMono :: Mono / [y]
fMono f up x y
| x+1 == y = f x ≤. f (x+1) ∵ up x ≤. f y
** QED
| x+1 < y = f x ≤. f (y-1) ∵ fMono f up x (y-1) ≤. f y ∵ up (y-1) ** QED
We prove the theorem by induction on y as specified by the annotation / [y] which states that y is
a well-founded termination metric that decreases at each recursive call [56]. If x+1 == y, then we
call the up x proof argument. Otherwise, x+1 < y, and we use the induction hypothesis i.e. apply
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:8
N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
fMono at y-1, after which transitivity of the less-than ordering finishes the proof. We can apply the
general fMono theorem to prove that fib increases monotonically:
fibMono :: n:Nat → m:{ n < m } → { fib n ≤ fib m }
fibMono = fMono fib fibUp
2.4
Complete Verification: Automating Equational Reasoning
While equational proofs can be very easy, writing them out can quickly get exhausting. Lets face
it: fib3_2 is doing rather a lot of work just to prove that fib 3 equals 2! Happily, the calculational
nature of such proofs allows us to develop the following proof search algorithm PLE that is inspired
by model checking [14]
• Guard Normal Form: First, as shown in the definition of fibP in § 2.2, each reflected function
is transformed into a guard normal form ∧i (pi ⇒ f (x) = bi ) i.e. a collection of guards pi and
their corresponding definition bi .
• Unfolding: Second, given a VC of the form Φ ⇒ p, we iteratively unfold function application
terms in Φ and p by instantiating them with the definition corresponding to an enabled guard,
where we check enabled-ness by querying the SMT solver. For example, given the VC true ⇒
fib 3 = 2, the guard 3 > 1 of the application fib 3 is trivially enabled, i.e. is true, hence we
strengthen the hypothesis Φ with the equality fib 3 = fib (3 − 1) + fib (3 − 2) corresponding
to unfolding the definition of fib at 3.
• Fixpoint: We repeat the unfolding process until either the VC is proved or we have reached
a fixpoint, i.e. no further unfolding is enabled. For example, the fixpoint computation of fib 3
unfolds the definition of fib at 3, 2, 1, and 0 and then stops as no further guards are enabled.
Automatic Equational Reasoning In § 6 we formalize a notion of equational proof and show
that the proof search procedure PLE enjoys two key properties. First, that it is guaranteed to find
an equational proof if one can be constructed from unfoldings of function definitions. (The user
must still provide instantiations of lemmas and induction hypotheses.) Second, that under certain
conditions readily met in practice, it is guaranteed to terminate. These two properties allow us
to use PLE to predictably automate proofs: the programmer needs only to supply the relevant
induction hypotheses or helper lemma applications. The remaining long chains of calculations are
performed automatically via SMT-based PLE. That is, the user must provide case statements and
the recursive structure, but can elide the long chains of =. applications. To wit, with complete proof
search, the proofs of § 2.3 shrink to:
fib3_2 :: {fib 3 = 2}
fib3_2 = ()
fMono :: Mono / [y]
fMono f up x y
| x+1 == y = up x
| x+1 < y = up (y-1) &&& fMono up x (y-1)
where the combinator p &&& q = () inserts the propositions p and q to the VC hypothesis.
PLE vs. Axiomatization Existing SMT based verifiers like Dafny [29] and F* [52] use the classical
axiomatic approach to verify assertions over user-defined functions like fib. In these systems, the
function is encoded in the logic as a universally quantified formula (or axiom): ∀n. fibP n after
which the SMT solver may instantiate the above axiom at 3, 2, 1 and 0 in order to automatically
prove {fib 3 = 2}.
The automation offered by axioms is a bit of a devil’s bargain, as axioms render VC checking
undecidable, and in practice automatic axiom instantiation can easily lead to infinite “matching
loops”. For example, the existence of a term fib n in a VC can trigger the above axiom, which may
then produce the terms fib (n − 1) and fib (n − 2), which may then recursively give rise to further
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:9
app_assoc :: AppendAssoc
app_assoc [] ys zs
= ([] ++ ys) ++ zs
=. ys ++ zs
=. [] ++ (ys ++ zs)
** QED
app_assoc (x:xs) ys zs
= ((x : xs) ++ ys) ++ zs
=. (x : (xs ++ ys)) ++ zs
=. x :((xs ++ ys) ++ zs)
∵ app_assoc xs ys zs
=. x : (xs ++ (ys
++ zs))
=. (x : xs) ++ (ys
++ zs) ** QED
app_assoc
:: AppendAssoc
app_assoc []
ys zs = ()
app_assoc (x:xs) ys zs = app_assoc xs ys zs
app_right_id
app_right_id []
app_right_id (x:xs)
:: AppendNilId
= ()
= app_right_id xs
map_fusion
:: MapFusion
map_fusion f g []
= ()
map_fusion f g (x:xs) = map_fusion f g xs
Fig. 1. (L) Equational proof of append associativity.
(R) PLE proof, also of append-id and map-fusion.
instantiations ad infinitum. To prevent matching loops an expert must carefully craft “triggers”
or alternatively, provide a “fuel” parameter that bounds the depth of instantiation [1]. Both these
approaches ensure termination, but can cause the axiom to not be instantiated at the right places,
thereby rendering the VC checking incomplete. The incompleteness is illustrated by the following
example from the Dafny benchmark suite [32]
pos n | n < 0
= 0
| otherwise = 1 + pos (n-1)
test
:: y:{y > 5} → {pos n = 3 + pos (n-3)}
test _ = ()
Dafny (and F*’s) fuel-based approach fails to check the above, when the fuel value is less than 3.
One could simply raise-the-fuel-and-try-again but at what point does the user know when to stop?
In contrast, PLE (1) does not require any fuel parameter, (2) is able to automatically perform the
required unfolding to verify this example, and (3) is guaranteed to terminate.
2.5
Case Study: Laws for Lists
Reflection and PLE are not limited to integers. We end the overview by showing how they verify
textbook properties of lists equipped with append (++) and map functions:
reflect (++) :: [a] → [a] → [a]
[]
++ ys = ys
(x:xs) ++ ys = x : (xs ++ ys)
reflect map :: (a → b) → [a] → [b]
map f []
= []
map f (x:xs) = f x : map f xs
In § 5.1 we will describe how the reflection mechanism illustrated via fibP is extended to account
for ADTs using SMT-decidable selection and projection operations, which reflect the definition
of xs ++ ys into the refinement as: if isNil xs then ys else sel1 xs : (sel2 xs ++ ys). We
require an explicit reflect annotation as not all Haskell functions can be reflected into logic, either
because it is unsound to do so (e.g. due to divergence) or because of limitations of our current
implementation. Recall that Liqid Haskell verifies that all reflected functions, like (++) and map
here, are total [56] and rejects the code otherwise.
Laws We can specify various laws about lists with refinement types. For example, the below
laws state that (1) appending to the right is an identity operation, (2) appending is an associative
operation, and (3) map distributes over function composition:
type AppendNilId = xs:_ → { xs ++ [] = xs }
type AppendAssoc = xs:_ → ys:_ → zs:_ → { xs ++ (ys ++ zs) = (xs ++ ys) ++ zs }
type MapFusion
= f:_ → g:_ → xs:_ → { map (f . g) xs = map f (map g xs) }
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:10 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
Fig. 2. Proofs that swap is idempotent with Coq, Agda, Dafny and PLE.
Proofs On the right in Figure 1 we show the proofs of these laws using PLE, which should be
compared to the classical equational proof e.g. by Wadler [61], shown on the left. With PLE, the
user need only provide the high-level structure — the case splits and invocations of the induction
hypotheses — after which PLE automatically completes the rest of the equational proof. Thus using
SMT-based PLE, app_assoc shrinks down to its essence: an induction over the list xs. The difference
is even more stark with map_fusion whose full equational proof is omitted, as it is twice as long.
PLE vs. Normalization The proofs in Figure 1 may remind readers familiar with Type-Theory
based proof assistants (e.g. Coq or Agda) of the notions of type-level normalization and rewriting
that permit similar proofs in those systems. While our approach of PLE is inspired by the idea of
type level computation, it differs from it in two significant ways. First, from a theoretical point of
view, SMT logics are not equipped with any notion of computation, normalization, canonicity or
rewriting. Instead, our PLE algorithm shows how to emulate those ideas by asserting equalities
corresponding to function definitions (Theorem 6.10). Second, from a practical perspective, the
combination of (decidable) SMT-based theory reasoning and PLE’s proof search can greatly simplify
verification. For example, consider the swap function from Appel [2]’s Coq textbook:
swap :: [Integer] → [Integer]
swap (x1:x2:xs) = if x1 > x2 then x2:x1:x2 else x1:x2:xs
swap xs
= xs
In Figure 2 we show four proofs that swap is idempotent: Appel’s proof using Coq (simplified by
the use of a hint database and the arithmetic tactic omega), its variant in Agda (for any Decidable
Partial Order), the PLE proof, and a proof using the Dafny verifier. It is readily apparent that PLE’s
proof search, working hand-in-glove with SMT-based theory reasoning, makes proving the result
trivial in comparison to Coq or Agda. Of course, proof assistants like Agda, Coq, and Isabelle
emit easily checkable certificates and have decades-worth of tactics, libraries, and proof scripts that
enable large scale proof engineering. On the other hand, Dafny’s fuel-based axiom instantiation
automatically unfolds the definition of swap twice, thereby completing the proof without any user
input. These heuristics are orthogonal to PLE and can be combined with it, if the user wishes to
trade off predictability for even more automation.
Summary We saw an overview of an SMT-automated refinement type checker that achieves SMTdecidable checking by restricting verification conditions to be quantifier-free and hence, decidable.
In existing SMT-based verifiers (e.g. Dafny) there are two main reasons to introduce quantifiers,
namely (1) to express quantified specifications and (2) to encode the semantics of user-defined
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:11
Native Terms
Implication
Negation
Conjunction
Disjunction
Forall
Exists
Logical Formula
e
ϕ1 ⇒ ϕ2
¬ϕ
ϕ1 ∧ ϕ2
ϕ1 ∨ ϕ2
∀x .ϕ
∃x .ϕ
Refinement Type
{e }
ϕ1 → ϕ2
ϕ → { False }
(ϕ 1 , ϕ 2 )
Either ϕ 1 ϕ 2
x :τ →ϕ
(x :: τ , ϕ)
Fig. 3. Mapping from logical predicates to quantifier-free refinement types. { e } abbreviates { v : Prop | e }.
Function binders are not relevant for negation and implication, and hence, elided.
functions. Next, we use propositions-as-types to encode quantified specifications and in § 4 we
show how to encode the semantics of user-defined functions via refinement reflection.
3
EMBEDDING NATURAL DEDUCTION WITH REFINEMENT TYPES
In this section we show how user-provided quantified specifications can be naturally encoded using
λ-abstractions and dependent pairs to encode universal and existential quantification, respectively.
Proof terms can be generated using the standard natural deduction derivation rules, following
Propositions as Types [62] (also known as the Curry-Howard isomorphism [24]). What is new is that
we exploit this encoding to show for the first time that a refinement type system can represent any
proof in Gentzen’s natural deduction [23] while still taking advantage of SMT decision procedures
to automate the quantifier-free portion of natural deduction proofs. For simplicity, in this section
we assume all terms are total.
3.1
Propositions: Refinement Types
Figure 3 maps logical predicates to types constructed over quantifier-free refinements.
Native terms Native terms consist of all of the (quantifier-free) expressions of the refinement
languages. In § 4 we formalize refinement typing in a core calculus λ R where refinements include
(quantifier-free) terminating expressions.
Boolean connectives Implication ϕ 1 ⇒ ϕ 2 is encoded as a function from the proof of ϕ 1 to the
proof of ϕ 2 . Negation is encoded as an implication where the consequent is False. Conjunction
ϕ 1 ∧ ϕ 2 is encoded as the pair (ϕ 1 , ϕ 2 ) that contains the proofs of both conjuncts and disjunction
ϕ 1 ∨ ϕ 2 is encoded as the sum type Either that contains the proofs of one of the disjuncts, i.e. where
data Either a b = Left a | Right b.
Quantifiers Universal quantification ∀x .ϕ is encoded as lambda abstraction x : τ → ϕ and
eliminated by function application. Existential quantification ∃x .ϕ is encoded as a dependent pair
(x::τ , ϕ) that contains the term x and a proof of a formula that depends on x. Even though
refinement type systems do not traditionally come with explicit syntax for dependent pairs, one
can encode dependent pairs in refinements using abstract refinement types [55] which do not add
extra complexity to the system. Consequently, we add the syntax for dependent pairs in Figure 3 as
syntactic sugar for abstract refinements.
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:12 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
3.2
Proofs: Natural Deduction
We overload ϕ to be both a proposition and a refinement type. We connect these two meanings of
ϕ by using the Propositions as Types [62], to prove that if there exists an expression (or proof term)
with refinement type ϕ, then the proposition ϕ is valid.
We construct proofs terms using Gentzen’s natural deduction system [23], whose rules map
directly to refinement type derivations. The rules for natural deduction arise from the propositionsas-types reading of the standard refinement type checking rule (to be defined in § 4) Γ ⊢ e : ϕ as
“ϕ is provable under the assumptions of Γ”. We write Γ ⊢N D ϕ for Gentzen’s natural deduction
judgment “under assumption Γ, proposition ϕ holds”. Then, each of Gentzen’s logical rules can be
recovered from the rules in Figure 5 by rewriting each judgment Γ ⊢ e : ϕ of λ R as Γ ⊢N D ϕ. For
example, conjunction and universal elimination can be derived as:
Γ ⊢N D ϕ 1 ∨ ϕ 2
Γ, ϕ 1 ⊢N D ϕ
Γ, ϕ 2 ⊢N D ϕ
Γ ⊢N D e x term
Γ ⊢N D ∀x .ϕ
∨ -E
∀-E
Γ ⊢N D ϕ[x/e x ]
Γ ⊢N D ϕ
Programs as Proofs As Figure 5 directly maps natural deduction rules to derivations that are
accepted by refinement typing, we conclude that if there exists a natural deduction derivation for a
proposition ϕ, then there exists an expression that has the refinement type ϕ.
Theorem 3.1. If Γ ⊢N D ϕ, then we can construct an e such that Γ ⊢ e : ϕ.
Note that our embedding is not an isomorphism, since the converse of Theorem 3.1 does not
hold. As a counterexample, the law of the excluded middle can be proved in our system (i.e. we
can construct an Either term e, so that p : { Bool | True } ⊢ e : p ∨ ¬p), but cannot be proved using
natural deduction (i.e. { True } ⊬N D p ∨ ¬p). The reason for that is that our system is using the
classical logic of the SMTs, which includes the law of the excluded middle. On the contrary, in
intuitionistic systems that also encode natural deduction (e.g. Coq, Idris, NuPRL) the law of the
excluded middle should be axiomatized.
3.3
Examples
Next, we illustrate our encoding with examples of proofs for quantified propositions ranging from
textbook logical tautologies, properties of datatypes like lists, and induction on natural numbers.
Natural Deduction as Type Derivation We illustrate the mapping from natural deduction rules to
typing rules in Figure 4 which uses typing judgments to express Gentzen’s proof of the proposition
ϕ ≡ (∃x .∀y.(p x y)) ⇒ (∀y.∃x .(p x y))
Read bottom-up, the derivation provides a proof of ϕ. Read top-down, it constructs a proof of the
formula as the term λe y.case e of {(x, e x ) → (x, e x y)}. This proof term corresponds directly to
the following Haskell expression that typechecks with type ϕ.
exAll :: p:(a → a → Bool) → (x::a, y:a → {p x y}) → y:a → (x::a, {p x y})
exAll _ = \e y → case e of {(x, ex) → (x, ex y)}
SMT-aided proofs The great benefit of using refinement types to encode natural deduction is
that the quantifier-free portions of the proof can be automated via SMTs. For every quantifier-free
proposition ϕ, you can convert between {ϕ}, where ϕ is treated as an SMT-proposition and ϕ, where
ϕ is treated as a type; and this conversion goes both ways. For example, let ϕ ≡ p ∧ (q ∨ r ). Then
flatten converts from ϕ to {ϕ} and expand the other way, while this conversion is SMT-aided.
flatten :: p:_ → q:_ → r:_→ ({p}, Either {q} {r}) → {p && (q || r)}
flatten (pf, Left qf) = pf &&& qf
flatten (pf, Right rf) = pf &&& rf
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:13
e :ϕ e , y :τy ⊢ e : ϕ e
e :ϕ e , y :τy , x :t x , e x :ϕ x ⊢ e x : ϕ x
e :ϕ e , y :τy , x :t x , e x :ϕ x ⊢ y : τy
e :ϕ e , y :τy , x :t x , e x :ϕ x ⊢ e x y : p x y
∀-E
e :ϕ e , y :τy ⊢ case e of {(x, e x ) → (x, e x y)} : ∃x .(p x y)
∃-E
e :ϕ e ⊢ λy.case e of {(x, e x ) → (x, e x y)} : ∀y.∃x .(p x y)
∀-I
∅ ⊢ λe y.case e of {(x, e x ) → (x, e x y)} : (∃x .∀y.(p x y)) ⇒ (∀y.∃x .(p x y))
⇒ -I
Fig. 4. Proof of (∃x .∀y.(p x y)) ⇒ (∀y.∃x .(p x y)) where ϕ e ≡ ∃x .∀y.(p x y), ϕ x ≡ ∀y.(p x y).
expand :: p:_ → q:_ → r:_ → {p && (q || r)} → ({p}, Either {q} {r})
expand proof | q = (proof, Left proof)
expand proof | r = (proof, Right proof)
Distributing Quantifiers Next, we construct the proof terms needed to prove two logical properties: that existentials distribute over disjunctions and foralls over conjunctions, i.e.
ϕ ∃ ≡ (∃x .p x ∨ q x) ⇒ ((∃x .p x) ∨ (∃x .q x))
(1)
ϕ ∀ ≡ (∀x .p x ∧ q x) ⇒ ((∀x .p x) ∧ (∀x .q x))
(2)
The specification of these properties requires nesting quantifiers inside connectives and vice versa.
The proof of ϕ ∃ (1) proceeds by existential case splitting and introduction:
exDistOr :: p:_ → q:_ → (x ::a , Either {p x} {q x })
→ Either (x ::a , {p x }) (x ::a , {q x })
exDistOr _ _ (x , Left px ) = Left (x , px )
exDistOr _ _ (x , Right qx ) = Right (x , qx )
Dually, we prove ϕ ∀ (2) via a λ-abstraction and case spitting inside the conjunction pair:
allDistAnd :: p:_ → q:_ →
→
allDistAnd _ _ andx = ( (\ x
, (\ x
(x:a → ({ p x}, {q x }) )
(( x:a → {p x }) , (x:a → {q x }) )
→ case andx x of (px , _) → px )
→ case andx x of (_ , qx ) → qx ) )
The above proof term exactly corresponds to its natural deduction proof derivation but using
SMT-aided verification can get simplified to the following
allDistAnd _ _ andx = (pf , pf )
where pf x = case andx x of (px , py ) → px &&& py
Properties of User Defined Datatypes As ϕ can describe properties of data types like lists, we
can prove properties of such types, e.g. that for every list xs, if there exists a list ys such that
xs == ys ++ ys ,then xs has even length.
ϕ ≡ ∀xs.((∃ys. xs = ys ++ ys) ⇒ (∃n.len xs = n + n))
The proof (evenLen) proceeds by existential elimination and introduction and uses the lenAppend
lemma, which in turn uses induction on the input list and PLE to automate equational reasoning.
evenLen :: xs :[ a] → ( ys ::[ a ] ,{ xs = ys ++ ys }) → (n :: Int ,{ len xs = n+n })
evenLen xs (ys , pf ) = ( len ys , lenAppend ys ys &&& pf )
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:14 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
Γ ⊢ fst e : ϕ 1
Γ ⊢ snd e : ϕ 2
Γ ⊢ e : (ϕ 1 , ϕ 2 )
Γ ⊢ e : (ϕ 1 , ϕ 2 )
Γ ⊢ fst e : ϕ 1
Γ ⊢ e : (ϕ 1 , ϕ 2 )
∧ -I
Γ ⊢ e1 : ϕ1
Γ ⊢ Left e 1 : Either ϕ 1 ϕ 2
Γ ⊢ snd e : ϕ 2
Γ ⊢ case e of { Left x 1 → e 1 ;
Right x 2 → e 2 } : ϕ
Γ ⊢ e1 : ϕ2
Γ ⊢ e : (x :: τ , ϕ[x/fst e])
∧ -R-E
Γ ⊢ e : Either ϕ 1 ϕ 2
Γ, x 1 :ϕ 1 ⊢ e 1 : ϕ
Γ, x 2 :ϕ 2 ⊢ e 2 : ϕ
∨ -L-I
∨ -R-I
Γ ⊢ Right e 2 : Either ϕ 1 ϕ 2
Γ, x :ϕ x ⊢ e : ϕ
⇒ -I
Γ ⊢ λx .e : ϕ x → ϕ
Γ, x :τ ⊢ e : ϕ
∀-I
Γ ⊢ λx .e : (x : τ → ϕ)
Γ ⊢ fst e : τ
Γ, x :τ ⊢ snd e : ϕ
∧ -L-E
Γ ⊢ e : ϕx → ϕ
Γ ⊢ ex : ϕ x
Γ ⊢ e ex : ϕ
Γ ⊢ ex : τ
Γ ⊢ e : (x : τ → ϕ)
∃-I
∨ -E
⇒ -E
∀-E
Γ ⊢ e e x : ϕ[x/e x ]
Γ ⊢ e : (x :: τ , ϕ x )
Γ, x :τ , y :ϕ x ⊢ e ′ : ϕ
Γ ⊢ case e of {(x, y) → e ′ } : ϕ
∃-E
Fig. 5. Natural deduction rules for refinement types. With [fst|snd] e ≡ case e of {(x 1 , x 2 ) → [x 1 |x 2 ]} .
lenAppend :: xs :_ → ys :_ → { len ( xs ++ ys ) = len xs + len ys }
lenAppend []
_ = ()
lenAppend ( x: xs ) ys = lenAppend xs ys
Induction on Natural Numbers Finally, we specify and verify induction on natural numbers:
ϕ ind ≡ (p 0 ∧ (∀n.p (n − 1) ⇒ p n) ⇒ ∀n.p n)
The proof proceeds by induction (e.g. case splitting). In the base case, n == 0, the proof calls the
left conjunct, which contains a proof of the base case. Otherwise, 0 < n, and the proof applies the
induction hypothesis to the right conjunct instantiated at n-1.
ind :: p : _ → ({ p 0} ,( n: Nat → {p (n -1) } → {p n }) ) → n: Nat → {p n}
ind p ( p0 , pn ) 0 = p0
ind p ( p0 , pn ) n = pn n ( ind p (p0 , pn ) (n -1) )
3.4
Consequences
To summarize, we use the propositions-as-types principle to make two important contributions.
First, we show that natural deduction reasoning can smoothly co-exist with SMT-based verification
to automate the decidable, quantifier-free portions of the proof.
Second, we show for the first time how natural deduction proofs can be encoded in refinement
type systems like Liqid Haskell and we expect this encoding to extend, in a straightforward
manner to other SMT-based deductive verifiers (e.g. Dafny and F*). This encoding shows that
refinement type systems are expressive enough to encode any intuitionistic natural deduction
proof, gives a guideline for encoding proofs with nested quantifiers, and provides a pleasant
implementation of natural deduction that is pedagogically useful.
4
REFINEMENT REFLECTION: λ R
Refinement reflection encodes recursive functions in the quantifier-free, SMT logic and it is formalized in three steps. First, we develop a core calculus λ R with an undecidable type system based
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:15
Ops ⊙ ::=
Consts c ::=
|
Vals w ::=
Exprs e ::=
|
Binds
Progs
Bas. Types
Ref. Types
b
p
B
τ
::=
::=
::=
::=
=|<
∧ | ! | ⊙ | +, −, . . .
True | False | 0, −1, 1, . . .
c | λx .e | D w
w | x | ee
case x = e of {D x → e}
e | let rec x : τ = b in b
b | reflect x : τ = e in p
Int | Bool | T
{v : B [⇓] | e} | x :τx → τ
Fig. 6. (Left) Syntax of λ R : Denotational Typing.
Terms p ::= p ▷◁ p | ⊕1p | D
| n | b | x | xp
| if p then p else p
Ints
Bools
Bin-Ops
Un-Ops
Args
n
b
▷◁
⊕1
sa
::=
::=
::=
::=
::=
|
0, −1, 1, . . .
True | False
= | < | ∧ | +, −, . . .
! | ...
Int | Bool | U
Fun sa sa
Sorts s ::= sa | sa → s
(Right) Syntax of λS : Algorithmic Typing.
on denotational semantics and prove it sound. Next, in § 5 we define a language λS that soundly
approximates λ R while enabling decidable, SMT-based type checking. Finally, in § 6 we develop a
complete proof search algorithm to automate equational reasoning.
4.1
Syntax
Figure 6 summarizes the syntax of λ R , which is essentially the calculus λU [56] with explicit
recursion and a special reflect binding to denote terms that are reflected into the refinement
logic. The elements of λ R are constants, values, expressions, binders, and programs.
Constants The constants of λ R include primitive relations ⊙, here, the set {=, <}. Moreover, they
include the booleans True, False, integers −1, 0, 1, etc., and logical operators ∧, !, etc..
Data Constructors Data constructors are special constants. For example, the data type [Int],
which represents finite lists of integers, has two data constructors: [] (nil) and : (cons).
Values & Expressions The values of λ R include constants, λ-abstractions λx .e, and fully applied
data constructors D that wrap values. The expressions of λ R include values, variables x, applications
e e, and case expressions.
Binders & Programs A binder b is a series of possibly recursive let definitions, followed by an
expression. A program p is a series of reflect definitions, each of which names a function that is
reflected into the refinement logic, followed by a binder. The stratification of programs via binders
is required so that arbitrary recursive definitions are allowed in the program but cannot be inserted
into the logic via refinements or reflection. (We can allow non-recursive let binders in expressions
e, but omit them for simplicity.)
4.2
Operational Semantics
We define ,→ to be the small step, call-by-name β-reduction semantics for λ R . We evaluate reflected
terms as recursive let bindings, with termination constraints imposed by the type system:
reflect x : τ = e in p ,→ let rec x : τ = e in p
,→⋆
We define
to be the reflexive, transitive closure of ,→ . Moreover, we define ≈β to be the
reflexive, symmetric, and transitive closure of ,→ .
Constants Application of a constant requires the argument be reduced to a value; in a single step,
the expression is reduced to the output of the primitive constant operation, i.e. c v ,→ δ (c, v). For
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:16 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
example, consider =, the primitive equality operator on integers. We have δ (=, n) =n where
δ (=n , m) equals True iff m is the same as n.
Equality We assume that the equality operator is defined for all values, and, for functions, is
defined as extensional equality. That is, for all f and f ′, (f = f ′) ,→ True iff ∀v. f v ≈β f ′ v. We
assume source terms only contain implementable equalities over non-function types; while function
extensional equality only appears in refinements.
4.3
Types
λ R types include basic types, which are refined with predicates, and dependent function types. Basic
types B comprise integers, booleans, and a family of data-types T (representing lists, trees etc.).
For example, the data type [Int] represents lists of integers. We refine basic types with predicates
(boolean-valued expressions e) to obtain basic refinement types {v : B | e}. We use ⇓ to mark
provably terminating computations and use refinements to ensure that if e:{v : B ⇓ | e ′ }, then e
terminates. As discussed by Vazou et al. [56] termination labels are checked using refinement types
and are used to ensure that refinements cannot diverge as required for soundness of type checking
under lazy evaluation. Termination checking is crucial for this work, as combined with checks for
exhaustive definitions (§ 2.1), it ensures totality (well-formedness) of expressions as required by
propositions-as-types (§ 3) and termination of PLE (§ 6). Finally, we have dependent function types
x :τx → τ where the input x has the type τx and the output τ may refer to the input binder x. We
write B to abbreviate {v : B | True} and τx → τ to abbreviate x :τx → τ , if x does not appear in τ .
Denotations Each type τ denotes a set of expressions [[τ ]], that is defined via the operational
semantics [27]. Let shape(τ ) be the type we get if we erase all refinements from τ and e : shape(τ )
be the standard typing relation for the typed lambda calculus. Then, we define the denotation of
types as:
[[{x : B | r }]] {e | e : B, if e ,→⋆ w then r [x/w] ,→⋆ True}
[[{x : B ⇓ | r }]] [[{x : B | r }]] ∩ {e | ∃ w.e ,→⋆ w }
[[x :τx → τ ]] {e | e : shape(τx → τ ), ∀e x ∈ [[τx ]]. (e e x ) ∈ [[τ [x/e x ]]]}
Constants For each constant c we define its type prim(c) such that c ∈ [[prim(c)]]. For example,
prim(3)
prim(+)
prim(≤)
4.4
{v : Int⇓ | v = 3}
x : Int⇓ → y : Int⇓ → {v : Int⇓ | v = x + y}
x : Int⇓ → y : Int⇓ → {v : Bool⇓ | v ⇔ x ≤ y}
Refinement Reflection
Reflection strengthens function output types with a refinement that reflects the definition of the
function in the logic. We do this by treating each reflect-binder (reflect f : τ = e in p) as a
let rec-binder (let rec f : Reflect(τ , e) = e in p) during type checking (rule T-Refl in Figure 7).
Reflection We write Reflect(τ , e) for the reflection of the term e into the type τ , defined as
Reflect({v : B ⇓ | r }, e)
Reflect(x :τx → τ , λx .e)
{v : B ⇓ | r ∧ v = e}
x :τx → Reflect(τ , e)
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:17
As an example, recall from § 2 that the reflect fib strengthens the type of fib with the refinement
fibP. That is, let the user specified type of fib be t fib and the its definition be definition λn.e fib .
t fib {v : Int⇓ | 0 ≤ v} → {v : Int⇓ | 0 ≤ v}
e fib case x = n ≤ 1 of {True → n; False → fib(n − 1) + fib(n − 2)}
Then, the reflected type of fib will be:
Reflect(t fib , e fib ) = n : {v : Int⇓ | 0 ≤ v} → {v : Int⇓ | 0 ≤ v ∧ v = e fib }
Termination Checking We defined Reflect(·, ·) to be a partial function that only reflects provably
terminating expressions, i.e. expressions whose result type is marked with ⇓. If a non-provably
terminating function is reflected in an λ R expression then type checking will fail (with a reflection
type error in the implementation). This restriction is crucial for soundness, as diverging expressions
can lead to inconsistencies. For example, reflecting the diverging f x = 1 + f x into the logic leads
to an inconsistent system that is able to prove 0 = 1.
Automatic Reflection Reflection of λ R expressions into the refinements happens automatically
by the type system, not manually by the user. The user simply annotates a function f as reflect f .
Then, the rule T-Refl in Figure 7 is used to type check the reflected function by strengthening f ’s
result via Reflect(·, ·). Finally, the rule T-Let is used to check that the automatically strengthened
type of f satisfies f ’s implementation.
4.5
Typing Rules
Next, we present the type-checking rules of λ R , as found in Figure 7.
Environments and Closing Substitutions A type environment Γ is a sequence of type bindings
x 1 : τ1 , . . . , x n : τn . An environment denotes a set of closing substitutions θ which are sequences of
expression bindings: x 1 7→ e 1 , . . . , x n 7→ en such that:
[[Γ]] {θ | ∀x : τ ∈ Γ.θ (x) ∈ [[θ · τ ]]}
where θ · τ applies a substitution to a type (and likewise θ · p, to a program).
A reflection environment R is a sequence that binds the names of the reflected functions with
their definitions f 1 7→ e 1 , . . . , fn 7→ en . A reflection environment respects a type environment
when all reflected functions satisfy their types:
Γ |= R ∀(f 7→ e) ∈ R. ∃τ . (f : τ ) ∈ Γ ∧ (Γ; R ⊢ e : τ )
Typing A judgment Γ; R ⊢ p : τ states that the program p has the type τ in the type environment
Γ and the reflection environment R. That is, when the free variables in p are bound to expressions
described by Γ, the program p will evaluate to a value described by τ .
Rules All but two of the rules are standard [27, 56] except for the addition of the reflection
environment R at each rule. First, rule T-Refl is used to extend the reflection environment with
the binding of the function name with its definition (f 7→ e) and moreover to strengthen the
type of each reflected binder with its definition, as described previously in § 4.4. Second, rule
T-Exact strengthens the expression with a singleton type equating the value and the expression
(i.e. reflecting the expression in the type). This is a generalization of the “selfification” rules from
[27, 39] and is required to equate the reflected functions with their definitions. For example, the
application fib 1 is typed as {v : Int⇓ | fibP 1 ∧ v = fib 1} where the first conjunct comes from
the (reflection-strengthened) output refinement of fib (§ 2) and the second comes from rule T-Exact.
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:18 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
Γ; R ⊢ p : τ
Typing
x :τ ∈ Γ
Γ; R ⊢ x : τ
Γ; R ⊢ p : τ ′
T-Var
Γ; R ⊢ c : prim(c)
Γ; R ⊢ e : {v : B | er }
T-Exact
Γ; R ⊢ e : {v : B | er ∧ v = e}
Γ; R ⊢ e : (x :τx → τ )
T-Con
Γ; R ⊢ e x : τx
T-App
Γ; R ⊢ e e x : τ [x/e x ]
Γ; R ⊢ e : {v : T | er }
∀i.prim(D i ) = y j : τ j → {v : T | er i }
Γ; R ⊢ τ ′ ⪯ τ
T-Sub
Γ; R ⊢ p : τ
Γ, x : τx ; R ⊢ e : τ
Γ; R ⊢ λx .e : (x :τx → τ )
T-Fun
Γ, x : τx ; R ⊢ bx : τx
Γ, x : τx ⊢ τx
Γ, x : τx ; R ⊢ b : τ
Γ ⊢τ
Γ; R ⊢ let rec x : τx = bx in b : τ
Γ ⊢τ
Γ, y j : τ j , x : {v : T | er ∧ er i }; R ⊢ ei : τ
Γ; R ⊢ case x = e of {D i yi → ei } : τ
Γ; R, f 7→ e ⊢ let rec f : Reflect(τf , e) = e in p : τ
Γ; R ⊢ reflect f : τf = e in p : τ
T-Let
T-Case
T-Refl
Γ ⊢τ
Well Formedness
Γ, v : B; ∅ ⊢ e : Bool⇓
Γ ⊢ {v : B | e}
WF-Base
Γ ⊢ τx
Γ, x : τx ⊢ τ
Γ ⊢ x :τx → τ
WF-Fun
Γ; R ⊢ τ1 ⪯ τ2
Subtyping
∀θ ∈ [[Γ]].[[θ · {v : B | e 1 }]] ⊆ [[θ · {v : B | e 2 }]]
Γ; R ⊢ {v : B | e 1 } ⪯ {v : B | e 2 }
Γ; R ⊢ τx′ ⪯ τx
Γ, x : τx′ ; R ⊢ τ ⪯ τ ′
Γ; R ⊢ x :τx → τ ⪯ x :τx′ → τ ′
⪯-Base-λ R
⪯-Fun
Fig. 7. Typing of λ R .
Well-formedness A judgment Γ ⊢ τ states that the refinement type τ is well-formed in the
environment Γ. Following Vazou et al. [56], τ is well-formed if all the refinements in τ are Booltyped, provably terminating expressions in Γ.
Subtyping A judgment Γ; R ⊢ τ1 ⪯ τ2 states that the type τ1 is a subtype of τ2 in the environments
Γ and R. Informally, τ1 is a subtype of τ2 if, when the free variables of τ1 and τ2 are bound to
expressions described by Γ, the denotation of τ1 is contained in the denotation of τ2 . Subtyping of
basic types reduces to denotational containment checking, shown in rule ⪯-Base-λ R . That is, τ1 is
a subtype of τ2 under Γ if for any closing substitution θ in [[Γ]], [[θ · τ1 ]] is contained in [[θ · τ2 ]].
Soundness We prove that typing implies denotational inclusion and evaluation preserves typing.
Theorem 4.1. [Soundness of λ R ]
• Denotations If Γ; R ⊢ p : τ , then ∀θ ∈ [[Γ]].θ · p ∈ [[θ · τ ]].
• Preservation If ∅; ∅ ⊢ p : τ and p ,→⋆ w, then ∅; ∅ ⊢ w : τ .
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:19
The proofs can be found in [58]. Theorem 4.1 lets us prove that if ϕ is a λ R type interpreted as
a proposition (using the mapping of Figure 3) and if there exists a p so that ∅; ∅ ⊢ p : ϕ, the ϕ is
valid. For example, in § 2 we verified that the term fibUp proves n : Nat → {fib n ≤ fib (n + 1)}.
Via soundness of λ R , we get that for each valid input n, the result refinement is valid.
∀n.0 ≤ n ,→⋆ True ⇒ fib n ≤ fib (n + 1) ,→⋆ True
5
ALGORITHMIC CHECKING: λS
λS is a first order approximation of λ R where higher-order features are approximated with uninterpreted functions and the undecidable type subsumption rule ⪯-Base-λ R is replaced with a
decidable one (i.e., ⪯-Base-PLE), yielding an sound and decidable SMT-based algorithmic type
system. Figure 6 summarizes the syntax of λS , the sorted (SMT-) decidable logic of quantifier-free
equality, uninterpreted functions and linear arithmetic (QF-EUFLIA) [5, 36]. The terms of λS include
integers n, booleans b, variables x, data constructors D (encoded as constants), fully applied unary
⊕1 and binary ▷◁ operators, and application x p of an uninterpreted function x. The sorts of λS
include the built-in Int and Bool to represent integers and booleans and the uninterpreted U to
represent data types. The interpreted functions of λS , e.g. the logical constants = and <, have the
function sort s → s. Other functional values in λ R , e.g. reflected λ R functions and λ-expressions,
are represented as first-order values with the uninterpreted sort Fun s s.
5.1
Transforming λ R into λS
The judgment Γ ⊢ e ⇝ p states that a λ R term e is transformed, under an environment Γ, into a λS
term p. If Γ ⊢ e ⇝ p and Γ is clear from the context we write ⌊e⌋ and ⌈p⌉ to denote the translation
from λ R to λS and back. Most of the transformation rules are identity and can be found in [58].
Here we discuss the non-identity ones.
Embedding Types We embed λ R types into λS sorts as:
⌊Int⌋ Int
⌊Bool⌋ Bool
⌊{v : B [⇓] | e}⌋ ⌊B⌋
⌊x :τx → τ ⌋
Fun ⌊τx ⌋ ⌊τ ⌋
⌊T ⌋ U
Embedding Constants Elements shared on both λ R and λS translate to themselves. These elements
include booleans, integers, variables, binary and unary operators. SMT solvers do not support
currying, and so in λS , all function symbols must be fully applied. Thus, we assume that all
applications to primitive constants and data constructors are fully applied, e.g. by converting source
terms like (+ 1) to (\z → z + 1).
Embedding Functions As λS is first-order, we embed λs using the uninterpreted function lam.
Γ, x : τx ⊢ e ⇝ p
Γ; ∅ ⊢ (λx .e) : (x :τx → τ )
x⌋
Γ ⊢ λx .e ⇝ lam ⌊τ
xp
⌊τ ⌋
The term λx .e of type τx → τ is transformed to lamssx x p of sort Fun s x s, where s x and s are
respectively ⌊τx ⌋ and ⌊τ ⌋, lamssx is a special uninterpreted function of sort s x → s → Fun s x s, and
x of sort s x and p of sort s are the embeddings of the binder and body, respectively. As lam is an
SMT-function, it does not create a binding for x. Instead, x is renamed to a fresh SMT name.
Embedding Applications We use defunctionalization [40] to embed applications.
Γ ⊢ e x ⇝ px
Γ ⊢e ⇝p
Γ ⊢ e ex ⇝
Γ; ∅ ⊢ e : τx → τ
x⌋
app ⌊τ
⌊τ ⌋
p px
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:20 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
The term e e x , where e and e x have types τx → τ and τx , is transformed to appssx p px : s where s
and s x are ⌊τ ⌋ and ⌊τx ⌋, the appssx is a special uninterpreted function of sort Fun s x s → s x → s,
and p and px are the respective translations of e and e x .
Embedding Data Types We embed data constructors to a predefined λS constant sD of sort
⌊prim(D)⌋: Γ ⊢ D ⇝ sD . For each datatype, we create reflected functions that check the top-level
constructor and select their individual fields. For example, for lists, we create the functions
isNil []
= True
isNil (x:xs) = False
isCons (x:xs) = True
isCons []
= False
sel1 (x:xs) = x
sel2 (x:xs) = xs
The above selectors can be modeled precisely in the refinement logic via SMT support for ADTs [36].
To generalize, let D i be a data constructor such that prim(D i ) τi,1 → · · · → τi,n → τ . Then
check isD i has the sort Fun ⌊τ ⌋ Bool and select selD i, j has the sort Fun ⌊τ ⌋ ⌊τi, j ⌋.
Embedding Case Expressions We translate case-expressions of λ R into nested if terms in λS , by
using the check functions in the guards and the select functions for the binders of each case.
Γ ⊢e ⇝p
Γ ⊢ ei [yi /selD i x][x/e] ⇝ pi
Γ ⊢ case x = e of {D i yi → ei } ⇝ if app isD 1 p then p1 else . . . else pn
The above translation yields the reflected definition for append (++) from (§ 2.5).
Semantic Preservation The translation preserves the semantics of the expressions. Informally, if
Γ ⊢ e ⇝ p, then for every substitution θ and every logical model σ that respects the environment Γ
if θ · e ,→⋆ v then σ |= p = ⌊v⌋.
5.2
Algorithmic Type Checking
We make the type checking from Figure 7 algorithmic by checking subtyping via our novel, SMTbased Proof by Logical Evaluation(PLE). Next, we formalize how PLE makes checking algorithmic
and in § 6 we describe the PLE procedure in detail.
Verification Conditions Recall that in § 5.1 we defined ⌊·⌋ as the translation from λ R to λS .
Informally, the implication or verification condition (VC) ⌊Γ⌋ ⇒ p1 ⇒ p2 is valid only if the set of
values described by p1 is subsumed by the set of values described by p2 under the assumptions of Γ.
Γ is embedded into logic by conjoining the refinements of terminating binders [56]:
(
Ø
⌊e⌋
if Γ(x) = {x : B ⇓ | e}
⌊Γ⌋
⌊Γ, x⌋ where we embed each binder as ⌊Γ, x⌋
True otherwise.
x ∈Γ
Validity Checking Instead of directly using the VCs to check validity of programs, we use
the procedure PLE that strengthens the assumption environment ⌊Γ⌋ with equational properties.
Concretely, given a reflection environment R, type environment Γ, and expression e, the procedure
PLE(⌊R⌋, ⌊Γ⌋, ⌊e⌋) — we will define ⌊R⌋ in § 6.1 — returns true only when the expression e evaluates
to True under the reflection and type environments R and Γ.
Subtyping via VC Validity Checking We make subtyping, and hence, typing decidable, by
replacing the denotational base subtyping rule ⪯-Base-λ R with the conservative, algorithmic
version ⪯-Base-PLE that uses PLE to check the validity of the subtyping.
PLE(⌊R⌋, ⌊Γ, v : {v : B ⇓ | e 1 }⌋, ⌊e 2 ⌋)
Γ; R ⊢P LE {v : B | e 1 } ⪯ {v : B | e 2 }
⪯-Base-PLE
This typing rule is sound as functions reflected in R always respect the typing environment Γ
(by construction) and because PLE is sound (Theorem 6.2).
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:21
Terms p, t, b
::= λS if-free terms from Figure 6
Functions
F
::= λx .⟨p ⇒ b⟩
Definitional Environment
Ψ
::=
∅ | f 7→ F , Ψ
Logical Environment
Φ
::=
∅ | p, Φ
Fig. 8. Syntax of Predicates, Terms and Reflected Functions.
Lemma 5.1. If Γ; R ⊢P LE {v : B | e 1 } ⪯ {v : B | e 2 } then Γ; R ⊢ {v : B | e 1 } ⪯ {v : B | e 2 }.
Soundness of λS We write Γ; R ⊢PLE e : τ for the judgments that can be derived by the algorithmic
subtyping rule ⪯-Base-PLE (instead of ⪯-Base-λ R .) Lemma 5.1 implies the soundness of λS .
Theorem 5.2 (Soundness of λS ). If Γ; R ⊢PLE e : τ , then Γ; R ⊢ e : τ .
6
COMPLETE VERIFICATION: PROOF BY LOGICAL EVALUATION
Next, we formalize our Proof By Logical Evaluation algorithm PLE and show that it is sound (§ 6.1),
that it is complete with respect to equational proofs (§ 6.2), and that it terminates (§ 6.3).
6.1
Algorithm
Figure 8 describes the input environments for PLE. The logical environment Φ contains a set of
hypotheses p, described in Figure 6. The definitional environment Ψ maps function symbols f to
their definitions λx .⟨p ⇒ b⟩, written as λ-abstractions over guarded bodies. Moreover, the body b
and the guard p contain neither λ nor if. These restrictions do not impact expressiveness: λs can
be named and reflected, and if-expressions can be pulled out into top-level guards using DeIf(·),
defined in [58]. A definitional environment Ψ can be constructed from R as
⌊R⌋ { f 7→ λx .DeIf(⌊e⌋)|(f 7→ λx .e) ∈ R}
Notation We write f (t 1 , . . . , tn ) ≺ Φ if the λS term (app . . . (app f t 1 ) . . . tn ) is a syntactic subterm
of some t ∈ Φ. We abuse notation to write f (t) ≺ t ′ for f (t) ≺ {t ′ }. We write SmtValid(Φ, p) for
SMT validity of the implication Φ ⇒ p.
Instantiation & Unfolding A term q is a (Ψ, Φ)-instance if there exists f (t) ≺ Φ such that:
• Ψ(f ) ≡ λx .⟨pi ⇒
bi⟩,
• SmtValid(Φ, pi t/x ), and
• q ≡ (f (x) = bi ) t/x .
A set of terms Q is a (Ψ, Φ)-instance if every q ∈ Q is an (Ψ, Φ)-instance. The unfolding of Ψ, Φ is the
(finite) set of all (Ψ, Φ)-instances. Procedure Unfold(Ψ, Φ) shown in Figure 9 computes and returns
the conjunction of Φ and the unfolding of Ψ, Φ. The following properties relate (Ψ, Φ)-instances
to the semantics of λ R and SMT validity. Let R[e] denote the evaluation of e under the reflection
environment R, i.e. ∅[e] e and (R, f :e f )[e] R[let rec f = e f in e].
Lemma 6.1. For every Γ |= R and θ ∈ (|Γ|),
• Sat-Inst If ⌊e⌋ is a (⌊R⌋, ⌊Γ⌋)-instance, then θ · R[e] ,→⋆ True.
• SMT-Approx If SmtValid(⌊Γ⌋, ⌊e⌋), then θ · R[e] ,→⋆ True.
• SMT-Inst If q is a (⌊R⌋, ⌊Γ⌋)-instance and SmtValid(⌊Γ⌋ ∪ {q}, ⌊e⌋), then θ · R[e] ,→⋆ True.
The Algorithm Figure 9 shows our proof search algorithm PLE(Ψ, Φ, p) which takes as input a
set of reflected definitions Ψ, an hypothesis Φ, and a goal p. The PLE procedure recursively unfolds
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:22 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
Unfold
:
=
(Ψ, Φ) → Φ
Ð
Φ ∪ f (t )≺Φ Instantiate Ψ, Φ, f , t
n
o
(pi ⇒ bi ) ∈ d, SmtValid(Φ, pi t/x )
(⌊ f (x)⌋ = bi ) t/x
Unfold(Ψ, Φ)
=
=
Ψ(f )
PLE
:
PLE(Ψ, Φ, p)
where
loop (i, Φi )
| SmtValid(Φi , p)
| Φi+1 ⊆ Φi
| otherwise
where
Φi+1
=
(Ψ, Φ, p) → Bool
Ð
loop 0, Φ ∪ f (t )≺p Instantiate Ψ, Φ, f , t
=
=
=
true
false
loop (i + 1, Φi+1 )
=
Φ ∪ Unfold(Ψ, Φi )
Instantiate Ψ, Φ, f , t
where
λx .⟨d⟩
Fig. 9. Algorithm PLE: Proof by Logical Evaluation.
function application terms by invoking Unfold until either the goal can be proved using the unfolded
instances (in which case the search returns true) or no new instances are generated by the unfolding
(in which case the search returns false).
Soundness First, we prove the soundness of PLE.
Theorem 6.2 (Soundness). If PLE(⌊R⌋, ⌊Γ⌋, ⌊e⌋) then ∀θ ∈ (|Γ|), θ · R[e] ,→⋆ True.
We prove Theorem 6.2 using the Lemma 6.1 that relates instantiation, SMT validity, and the exact
semantics. Intuitively, PLE is sound as it reasons about a finite set of instances by conservatively
treating all function applications as uninterpreted [36].
6.2
Completeness
Next, we show that our proof search is complete with respect to equational reasoning. We define
a notion of equational proof Ψ, Φ ⊢ t ↠ t ′ and prove that if there exists such a proof, then
PLE(Ψ, Φ, t = t ′) is guaranteed to return true. To prove this theorem, we introduce the notion of
bounded unfolding which corresponds to unfolding definitions n times. We show that unfolding
preserves congruences, and hence, that an equational proof exists iff the goal can be proved with
some bounded unfolding. Thus, completeness follows by showing that the proof search procedure
computes the limit (i.e. fixpoint) of the bounded unfolding. In § 6.3 we show that the fixpoint is
computable: there exists an unfolding depth at which PLE reaches a fixpoint and hence terminates.
Bounded Unfolding For every Ψ, Φ, and 0 ≤ n, the bounded unfolding of depth n is defined by:
Unfold∗ (Ψ, Φ, 0)
Unfold∗ (Ψ, Φ, n + 1)
Φ
Φn ∪ Unfold(Ψ, Φn ) where Φn = Unfold∗ (Ψ, Φ, n)
That is, the unfolding at depth n essentially performs Unfold upto n times. The bounded-unfoldings
yield a monotonically non-decreasing sequence of formulas such that if two consecutive bounded
unfoldings coincide, then all subsequent unfoldings are the same.
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:23
Ψ, Φ ⊢ t ↠ t
Ψ, Φ ⊢ t ↠ t ′′
Eq-Refl
Φ′ = Unfold(Ψ, Φ ∪ {ν = t ′′ })
SmtValid(Φ′, ν = t ′)
Ψ, Φ ⊢ t ↠ t ′
Ψ, Φ ⊢ t 1 ↠ t 1′
Ψ, Φ ⊢ t 2 ↠ t 2′
SmtValid(Φ, t 1′ ▷◁ t 2′ )
Ψ, Φ ⊢ t 1 ▷◁ t 2
Eq-Trans
Eq-Proof
Fig. 10. Equational Proofs: rules for equational reasoning.
Lemma 6.3 (Monotonicity). ∀0 ≤ n. Unfold∗ (Ψ, Φ, n) ⊆ Unfold∗ (Ψ, Φ, n + 1).
Lemma 6.4 (Fixpoint). Let Φi Unfold∗ (Ψ, Φ, i). If Φn = Φn+1 , then ∀n < m. Φm = Φn .
Uncovering Next, we prove that every function application term that is uncovered by unfolding to
depth n is congruent to a term in the n-depth unfolding.
Lemma 6.5 (Uncovering). Let Φn ≡ Unfold∗ (Ψ, Φ ∪ {ν = t }, n). If SmtValid(Φn , ν = t ′), then
for every f (t ′) ≺ t ′ there exists f (t) ≺ Φn such that SmtValid(Φn , ti = ti′).
We prove the above lemma by induction on n where the inductive step uses the following
property of congruence closure, which itself is proved by induction on the structure of t ′:
Lemma 6.6 (Congruence). If SmtValid(Φ ∪ {ν = t }, ν = t ′) and ν < Φ, t, t ′, then for every
f (t ′) ≺ t ′ there exists f (t) ≺ Φ, t such that SmtValid(Φ, ti = ti′).
Unfolding Preserves Equational Links We use the uncovering Lemma 6.5 and congruence to
show that every instantiation that is valid after n steps is subsumed by the n + 1 depth unfolding.
That is, we show that every possible link in any equational chain can be proved equal to the source
expression via bounded unfolding.
Lemma 6.7 (Link). If SmtValid(Unfold∗ (Ψ, Φ∪{ν = t }, n), ν = t ′), then SmtValid(Unfold∗ (Ψ, Φ∪
{ν = t }, n + 1), Unfold(Ψ, Φ ∪ {ν = t ′ })).
Equational Proof Figure 10 formalizes our rules for equational reasoning. Intuitively, there is
an equational proof that t 1 ▷◁ t 2 under Ψ, Φ, written by the judgment Ψ, Φ ⊢ t 1 ▷◁ t 2 , if by some
sequence of repeated function unfoldings, we can prove that t 1 and t 2 are respectively equal to
t 1′ and t 2′ such that, SmtValid(Φ, t 1′ ▷◁ t 2′ ) holds. Our notion of equational proofs adapts the idea of
type level computation used in TT-based proof assistants to the setting of SMT-based reasoning,
via the directional unfolding judgment Ψ, Φ ⊢ t ↠ t ′. In the SMT-realm, the explicit notion of a
normal or canonical form is converted to the implicit notion of the equivalence classes of the SMT
solver’s congruence closure procedure (post-unfolding).
Completeness of Bounded Unfolding We use the fact that unfolding preserves equational links
to show that bounded unfolding is complete for equational proofs. That is, we prove by induction
on the structure of the equational proof that whenever there is an equational proof of t = t ′, there
exists some bounded unfolding that suffices to prove the equality.
Lemma 6.8. If Ψ, Φ ⊢ t ↠ t ′, then ∃0 ≤ n. SmtValid(Unfold∗ (Ψ, Φ ∪ {ν = t }, n), ν = t ′).
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:24 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
PLE is a Fixpoint of Bounded Unfolding We show that the proof search procedure PLE computes
the least-fixpoint of the bounded unfolding and hence, returns true iff there exists some unfolding
depth n at which the goal can be proved.
Lemma 6.9 (Fixpoint). PLE(Ψ, Φ, t = t ′) iff ∃n. SmtValid(Unfold∗ (Ψ, Φ ∪ {ν = t }, n), ν = t ′).
The proof follows by observing that PLE(Ψ, Φ, t = t ′) computes the least-fixpoint of the sequence
Φi Unfold∗ (Ψ, Φ, i). Specifically, we prove by induction on i that at each invocation of loop (i, Φi )
in Figure 9, Φi is equal to Unfold∗ (Ψ, Φ ∪ {ν = t }, i), which then yields the result.
Completeness of PLE Finally, we combine Lemmas 6.9 and 6.7 to show that PLE is complete, i.e. if
there is an equational proof that t ▷◁ t ′ under Ψ, Φ, then PLE(Ψ, Φ, t ▷◁ t ′) returns true.
Theorem 6.10 (Completeness). If Ψ, Φ ⊢ t ▷◁ t ′ then PLE(Ψ, Φ, t ▷◁ t ′) = true.
6.3
PLE Terminates
So far, we have shown that our proof search procedure PLE is both sound and complete. Both of
these are easy to achieve simply by enumerating all possible instances and repeatedly querying
the SMT solver. Such a monkeys-with-typewriters approach is rather impractical: it may never
terminate. Fortunately, we show that in addition to being sound and complete with respect to
equational proofs, if the hypotheses are transparent, then our proof search procedure always
terminates. Next, we describe transparency and explain intuitively why PLE terminates. We then
develop the formalism needed to prove the termination Theorem 6.16.
Transparency An environment Γ is inconsistent if SmtValid(⌊Γ⌋, false). An environment Γ is
inhabited if there exists some θ ∈ (|Γ|). We say Γ is transparent if it is either inhabited or inconsistent.
As an example of a non-transparent Φ0 consider the predicate lenA xs = 1 + lenB xs, where lenA
and lenB are both identical definitions of the list length function. Clearly there is no θ that causes
the above predicate to evaluate to true. At the same time, the SMT solver cannot (using the
decidable, quantifier-free theories) prove a contradiction as that requires induction over xs. Thus,
non-transparent environments are somewhat pathological and, in practice, we only invoke PLE
on transparent environments. Either the environment is inconsistent, e.g. when doing a proof-bycontradiction, or e.g. when doing a proof-by-case-analysis we can easily find suitable concrete
values via random [13] or SMT-guided generation [45].
Challenge: Connect Concrete and Logical Semantics As suggested by its name, the PLE algorithm aims to lift the notion of evaluation or computations into the level of the refinement logic.
Thus, to prove termination, we must connect the two different notions of evaluation, the concrete
(operational) semantics and the logical semantics being used by PLE. This connection is trickier
than appears at first glance. In the concrete realm totality ensures that every reflected function f
will terminate when run on any individual value v. However, in the logical realm, we are working
with infinite sets of values, compactly represented via logical constraints. In other words, the logical
realm can be viewed (informally) as an abstract interpretation of the concrete semantics. We must
carefully argue that despite the approximation introduced by the logical abstraction, the abstract
interpretation will also terminate.
Solution: Universal Abstract Interpretation We make this termination argument in three steps.
First, we formalize how PLE performs computation at the logical level via logical steps and logical
traces. We show (Lemma 6.13) that the logical steps form a so-called universal (or must) abstraction
of the concrete semantics [14, 18]. Second, we show that if PLE diverges, it is because it creates a
strictly increasing infinite chain, Unfold∗ (Ψ, Φ, 0) ⊂ Unfold∗ (Ψ, Φ, 1) . . . which corresponds to an
infinite logical trace. Third, as the logical computation is a universal abstraction we use inhabitation
to connect the two realms, i.e. to show that an infinite logical trace corresponds to an infinite
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:25
concrete trace. The impossibility of the latter must imply the impossibility of the former, i.e. PLE
terminates. Next, we formalize the above to obtain Theorem 6.16.
Totality A function is total when its evaluation reduces to exactly one value. The totality of R can
and is checked by refinement types (§ 4). Hence, for brevity, in the sequel we implicitly assume that
R is total under Γ.
Definition 6.11 (Total). Let b ≡ λx .⟨⌊p⌋ ⇒ ⌊e⌋⟩. b is total under Γ and R, if forall θ ∈ (|Γ|):
(1) if θ · R[pi ] ,→⋆ True, then ∃v. θ · R[ei ] ,→⋆ v,
(2) if θ · R[pi ] ,→⋆ True and θ · Ψ[p j ] ,→⋆ True, then i = j, and
(3) there exists an i so that θ · R[pi ] ,→⋆ True.
R is total under Γ, if every b ∈ ⌊R⌋ is total under Γ and R.
Subterm Evaluation As the reflected functions are total, the Church-Rosser theorem implies that
evaluation order is not important. To prove termination, we require an evaluation strategy, e.g.
CBV, in which if a reflected function’s guard is satisfied, then the evaluation of the corresponding
function body requires evaluating every subterm inside the body. As DeIf(·) hoists if-expressions
out of the body and into the top-level guards, the below fact follows from the properties of CBV:
Lemma 6.12. Let b ≡ λx .⟨⌊p⌋ ⇒ ⌊e⌋⟩ and f ∈ R. For every Γ, R, and θ ∈ (|Γ|), if θ · R[pi ] ,→⋆ True
and f (⌊e ′⌋) ≺ ⌊ei ⌋, then θ · R[ei ] ,→⋆ C[f (θ · R[e ′])].
Logical Step A pair f (t) ⇝ f ′(t ′) is a Ψ, Φ-logical step (abbrev. step), if
• Ψ(f ) ≡ λx .⟨p ⇒ b⟩,
• SmtValid(Φ ∧ Q, pi ) for some (Ψ, Φ)-instance Q, and
• f ′(t ′) ≺ bi t/x .
Steps and Reductions Next, using Lemmas 6.12, 6.1, and the definition of logical steps, we show
that every logical step corresponds to a sequence of steps in the concrete semantics:
Lemma 6.13 (Step-Reductions). If f (⌊e⌋) ⇝ f ′(⌊e ′⌋) is a logical step under ⌊R⌋, ⌊Γ⌋ and
θ ∈ (|Γ|), then f (θ · R[e]) ,→⋆ C[f (θ · R[e ′])] for some context C.
Logical Trace A sequence f 0 (t 0 ), f 1 (t 1 ), f 2 (t 2 ), . . . is a Ψ, Φ-logical trace (abbrev. trace), if fi (ti ) ⇝
fi+1 (ti+1 ) is a Ψ, Φ-step, for each i. Our termination proof hinges upon the following key result:
inhabited environments only have finite logical traces. We prove this result by contradiction.
Specifically, we show by Lemma 6.13 that an infinite (⌊R⌋, ⌊Γ⌋)-trace combined with the fact that
Γ is inhabited yields at least one infinite concrete trace, which contradicts the totality assumption.
Hence, all the (⌊R⌋, ⌊Γ⌋) logical traces must be finite.
Theorem 6.14 (Finite-Trace). If Γ is inhabited, then every (⌊R⌋, ⌊Γ⌋)-trace is finite.
Ascending Chains and Traces If unfolding Ψ, Φ yields an infinite chain Φ0 ⊂ . . . ⊂ Φn . . .,
then Ψ, Φ has an infinite logical trace. We construct the trace by selecting at level i (i.e. in Φi ) an
application term fi (ti ) that was created by unfolding an application term at level i − 1 (i.e. in Φi−1 ).
Lemma 6.15 (Ascending Chains). Let Φi Unfold∗ (Ψ, Φ, i). If there exists an (infinite) ascending chain Φ0 ⊂ . . . ⊂ Φn . . ., then there exists an (infinite) logical trace f 0 (t 0 ), . . . , fn (tn ), . . ..
Logical Evaluation Terminates Finally, we prove that the proof search procedure PLE terminates.
If PLE loops forever, there must be an infinite strictly ascending chain of unfoldings Φi , and hence,
by Lemma 6.15, an infinite logical trace, which, by Theorem 6.14, is impossible.
Theorem 6.16 (Termination). If Γ is transparent, then PLE(⌊R⌋, ⌊Γ⌋, p) terminates.
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:26 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
Benchmark
Common
Without PLE Search
With PLE Search
Impl (l) Spec (l) Proof (l) Time (s) SMT (q) Proof (l) Time (s) SMT (q)
Arithmetic
Fibonacci
7
Ackermann
20
Class Laws Fig 11
Monoid
33
Functor
48
Applicative
62
Monad
63
Higher-Order Properties
Logical Properties
0
Fold Universal
10
Functional Correctness
SAT-solver
92
Unification
51
Deterministic Parallelism
Conc. Sets
597
n-body
163
Par. Reducers
30
Total
1176
10
73
38
196
2.74
5.40
129
566
16
119
1.92
13.80
79
846
50
44
110
42
109
93
241
122
4.47
4.97
12.00
5.39
34
26
69
49
33
14
74
39
4.22
3.68
10.00
4.89
209
68
1090
250
20
44
33
43
2.71
2.17
32
24
33
14
2.74
1.46
32
48
34
60
0
85
50.00
4.77
50
195
0
21
50.00
5.64
50
422
329
251
212
1279
339
101
124
1524
40.10
7.41
6.63
148.76
339
61
52
1626
229
21
25
638
40.70
6.27
5.56
150.88
861
61
52
4068
Table 1. We report verification Time (in seconds, on a 2.3GHz Intel® Xeon® CPU E5-2699 v3 with 18 physical
cores and 64GiB RAM.), the number of SMT queries and size of Proofs (in lines). The Common columns
show sizes of common Implementations and Specifications . We separately consider proofs Without and
With PLE Search.
7
EVALUATION
We have implemented reflection and PLE in Liqid Haskell [56]. Table 1 summarizes our evaluation
which aims to determine (1) the kinds of programs and properties that can be verified, (2) how PLE
simplifies writing proofs, and (3) how PLE affects the verification time.
Benchmarks We summarize our benchmarks below, see [58] for details.
• Arithmetic We proved arithmetic properties for the textbook Fibonacci function (c.f. § 2) and
the 12 properties of the Ackermann function from [53].
• Class Laws We proved the monoid laws for the Peano, Maybe and List data types and the Functor,
Applicative, and Monad laws, summarized in Figure 11, for the Maybe, List and Identity monads.
• Higher Order Properties We used natural deduction to prove textbook logical properties as
in § 3. We combined natural deduction principles with PLE-search to prove universality of
right-folds, as described in [25] and formalized in Agda [34].
• Functional Correctness We proved correctness of a SAT solver and a unification algorithm
as implemented in Zombie [11]. We proved that the SAT solver takes as input a formula f and
either returns Nothing or an assignment that satisfies f, by reflecting the notion of satisfaction.
Then, we proved that if the unification unify s t of two terms s and t returns a substitution
su, then applying su to s and t yields identical terms. Note that, while the unification function
can itself diverge, and hence cannot be reflected, our method allows terminating and diverging
functions to soundly coexist.
• Deterministic Parallelism Retrofitting verification onto an existing language with a mature
parallel run-time allows us to create three deterministic parallelism libraries that, for the first time,
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:27
verify implicit assumptions about associativity and ordering that are critical for determinism
(c.f. [58] for extended description). First, we proved that the ordering laws hold for keys inserted
into LVar-style concurrent sets [28]. Second, we used monad-par [33] to implement an n-body
simulation, whose correctness relied upon proving that a triple of Real (implementing) 3-d
acceleration was a Monoid. Third, we built a DPJ-style [10] parallel-reducers library whose
correctness relied upon verifying that the reduced arguments form a CommutativeMonoid, and
which was the basis of a parallel array sum.
Proof Effort We split the total lines of code of our benchmarks into three categories: Spec
represents the refinement types that encode theorems, lemmas, and function specifications; Impl
represents the rest of the Haskell code that defines executable functions; Proofs represent the sizes
of the Haskell proof terms (i.e. functions returning Prop). Reflection and PLE are optionally enabled
using pragmas; the latter is enabled either for a whole file/module or per top-level function.
Runtime Overhead Proof terms have no runtime overhead as they will never be evaluated. When
verification of an executable term depends on theorems, we use the below withTheorem function
withTheorem :: x:a → Prop → { v:a | v = x }
withTheorem x _ = x
that inserts the proof argument into the static verification environment, relying upon laziness to
not actually evaluate the proof. For example, when verification depends on the associativity of
append on the lists xs, ys, and zs, the invocation withTheorem xs (app_assoc xs ys zs) extends
the (static) SMT verification environment with the instantiation of the associativity theorem of
Figure 1. This invocation adds no runtime overhead, since even though app_assoc xs ys zs is an
expensive recursive function, it will never actually get evaluated. To ensure that proof terms are
not evaluated in runtime, without using laziness, one can add one rewrite rule for each proof term
that replaces the term with unit. For example, the rewrite rule for app_assoc is
RULES " assoc / runtime "
forall xs ys zs . app_assoc xs ys zs = ()
Such rules are sound, since each proof term is total, thus provably reduces to unit.
Results The highlights of our evaluation are the following. (1) Reflection allows for the specification
and verification of a wide variety of important properties of programs. (2) PLE drastically reduces
the proof effort: by a factor of 2−5× — shrinking the total lines of proof from 1524 to 638— making it
quite modest, about the size of the specifications of the theorems. Since PLE searches for equational
properties, there are some proofs, that rarely occur in practice, that PLE cannot simplify, e.g. the
logical properties from § 3. (3) PLE does not impose a performance penalty: even though proof
search can make an order of magnitude many more SMT queries — increasing the total SMT queries
from 1626 without PLE to 4068 with PLE— most of these queries are simple and it is typically faster
to type-check the compact proofs enabled by PLE than it is to type-check the 2 − 5× longer explicit
proofs written by a human.
8
RELATED WORK
SMT-Based Verification SMT-solvers have been extensively used to automate program verification via Floyd-Hoare logics [36]. Leon introduces an SMT-based algorithm that is complete for
catamorphisms (folds) over ADTs [50] and a semi-decision procedure that is guaranteed to find
satisfying assignments (models) for queries over arbitrary recursive functions, if they exist [51].
Our work is inspired by Dafny’s Verified Calculations [31] but differs in (1) our use of reflection
instead of axiomatization, (2) our use of refinements to compose proofs, and (3) our use of PLE to
automate reasoning about user-defined functions. Dafny (and F* [52]) encode user-functions as
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:28 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
Monoid (for Peano, Maybe, List)
Left Id. mempty x ♦ ≡ x
Right Id. x ♦ mempty ≡ x
Assoc. (x ♦ y) ♦ z ≡ x ♦ (y ♦ z)
Functor (for Maybe, List, Id)
Id. fmap id xs ≡ id xs
Distr. fmap (д ◦ h) xs ≡ (fmap д ◦ fmap h) xs
Applicative (for Maybe, List, Id)
Monad (for Maybe, List, Id)
Id. pure id ⊛ v ≡ v
Left Id. return a ≫= f ≡ f a
Comp. pure (◦) ⊛ u ⊛ v ⊛ w ≡ u ⊛ (v ⊛ w) Right Id. m ≫= return ≡ m
Hom. pure f ⊛ pure x ≡ pure (f x)
Assoc. (m ≫= f ) ≫= д ≡ m ≫= (λx → f x ≫= д)
Inter. u ⊛ pure y ≡ pure ($ y) ⊛ u
Ord (for Int, Double, Either, (, ))
Refl. x ≤ x
Antisym. x ≤ y ∧ y ≤ x =⇒ x ≡ y
Trans. x ≤ y ∧ y ≤ z =⇒ x ≤ z
Total. x ≤ y ∨ y ≤ x
Commutative Monoid (for Int, Double, (, ))
Comm. x ♦ y ≡ y ♦ x
(including Monoid laws)
Fig. 11. Summary of Verified Typeclass Laws.
axioms and use a fixed fuel to instantiate functions upto some fixed unfolding depth [1]. While the
fuel-based approach is incomplete, even for equational or calculational reasoning, it may, although
rare in practice, quickly time out after a fixed, small number of instantiations rather than perform
an exhaustive proof search like PLE. Nevertheless, PLE demonstrates that it is possible to develop
complete and practical algorithms for reasoning about user-defined functions.
Proving Equational Properties Several authors have proposed tools for proving (equational)
properties of (functional) programs. Systems of Sousa and Dillig [48] and Asada et al. [3] extend
classical safety verification algorithms, respectively based on Floyd-Hoare logic and refinement
types, to the setting of relational or k-safety properties that are assertions over k-traces of a program.
Thus, these methods can automatically prove that certain functions are associative, commutative etc.
but are restricted to first-order properties and are not programmer-extensible. Zeno [47] generates
proofs by term rewriting and Halo [60] uses an axiomatic encoding to verify contracts. Both the
above are automatic, but unpredictable and not programmer-extensible, hence, have been limited
to far simpler properties than the ones checked here. HERMIT [22] proves equalities by rewriting
the GHC core language, guided by user specified scripts. Our proofs are Haskell programs, SMT
solvers automate reasoning, and importantly, we connect the validity of proofs with the semantics
of the programs.
Dependent Types in Programming Integration of dependent types into Haskell has been a long
standing goal [21] that dates back to Cayenne [4], a Haskell-like, fully dependent type language
with undecidable type checking. Our approach differs significantly in that reflection and PLE use
SMT-solvers to drastically simplify proofs over decidable theories. Zombie [46] investigates the
design of a dependently typed language where SMT-style congruence closure is used to reason
about the equality of terms. However, Zombie explicitly eschews type-level computation, as the
authors write “equalities that follow from β-reduction” are “incompatible with congruence closure”.
Due to this incompleteness, the programmer must use explicit @join@ terms to indicate where
normalization should be triggered, even so, equality checking is based on fuel, hence, is incomplete.
Theorem Provers Reflection shows how to retrofit deep specification and verification in the style of
Agda [38], Coq [7] and Isabelle [37] into existing languages via refinement typing and PLE shows
how type-level computation can be made compatible with SMT solvers’ native theory reasoning
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:29
yielding a powerful new way to automate proofs (§ 2.5). An extensive comparison [54] between
our approach and mature theorem provers like Coq, Agda, and Isabelle reveals that these provers
have two clear advantages over our approach: they emit certificates, so they rely on a small trusted
computing base, and they have decades-worth of tactics, libraries, and proof scripts that enable large
scale proof engineering. Some tactics even enable embedding of SMT-based proof search heuristics,
e.g. Sledgehammer [9], that is widely used in Isabelle. However, this search does not have the
completeness guarantees of PLE. The issue of extracting checkable certificates from SMT solvers
is well understood [12, 35] and easy to extend to our setting. However, the question of extending
SMT-based verifiers with tactics and scriptable proof search, and more generally, incorporating
interactivity in the style of proof-assistants, perhaps enhanced by proof-completion hints, remains
an interesting direction for future work.
9
CONCLUSIONS AND FUTURE WORK
Our results identify a new design for deductive verifiers wherein: (1) via Refinement Reflection,
we can encode natural deduction proofs as SMT-checkable refinement typed programs and (2) via
Proof by Logical Evaluation we can combine the complementary strengths of SMT- (i.e., decision
procedures) and TT- based approaches (i.e., type-level computation) to obtain completeness guarantees when verifying properties of user-defined functions. However, the increased automation of
SMT and proof-search can sometimes make it harder for a user to debug failed proofs. In future
work, it would be interesting to investigate how to add interactivity to SMT based verifiers, in the
form of tactics and scripts or algorithms for synthesizing proof hints, and to design new ways to
explain and fix refinement type errors.
ACKNOWLEDGMENTS
We thank the anonymous referees, our shepherd Koen Claessen, and Rustan Leino for their feedback
on earlier versions of this paper. Special thanks to George Karachalias and Valentin Robert for their
input on proof combinators, and to Joachim Breitner and Eric Seidel for their suggestion on run-time
proof elimination. This work was supported by the EPSRC programme grant EP/K034413/1, the
National Science Foundation under Grant Numbers CCF-1618756, CNS-1518765, CCF-1518844,
CCF-1422471, CCF-1223850, and CCF-1218344, and a generous gift from Microsoft Research.
REFERENCES
[1] N. Amin, K. R. M. L., and T. Rompf. 2014. Computing with an SMT Solver. In TAP.
[2] A. Appel. 2016. Verified Functional Algorithms. https://www.cs.princeton.edu/~appel/vfa/Perm.html.
[3] K. Asada, R. Sato, and N. Kobayashi. 2015. Verifying Relational Properties of Functional Programs by First-Order
Refinement. In PEPM.
[4] L. Augustsson. 1998. Cayenne - a Language with Dependent Types.. In ICFP.
[5] C. Barrett, A. Stump, and C. Tinelli. 2010. The SMT-LIB Standard: Version 2.0.
[6] J. Bengtson, K. Bhargavan, C. Fournet, A.D. Gordon, and S. Maffeis. 2008. Refinement Types for Secure Implementations.
In CSF.
[7] Y. Bertot and P. Castéran. 2004. Coq’Art: The Calculus of Inductive Constructions. Springer Verlag.
[8] R. S. Bird. 1989. Algebraic Identities for Program Calculation. In The Computer Journal.
[9] J. C. Blanchette, S. Böhme, and L. C. Paulson. 2011. Extending Sledgehammer with SMT Solvers. In CADE.
[10] Jr. Bocchino, L. Robert, V. S. Adve, S. V. Adve, and M. Snir. 2009. Parallel Programming Must Be Deterministic by
Default. In HotPar.
[11] C. Casinghino, V. Sjöberg, and S. Weirich. 2014. Combining proofs and programs in a dependently typed language. In
POPL.
[12] J. Chen, R. Chugh, and N. Swamy. 2010. Type-preserving Compilation of End-to-end Verification of Security Enforcement. In PLDI.
[13] K. Claessen and J. Hughes. 2000. QuickCheck: A Lightweight Tool for Random Testing of Haskell Programs. In ICFP.
[14] E. M. Clarke, O. Grumberg, and D.E. Long. 1992. Model checking and abstraction. In POPL.
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:30 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
[15] S. Collange, D. Defour, S. Graillat, and R. Iakymchuk. 2014. Full-Speed Deterministic Bit-Accurate Parallel Floating-Point
Summation on Multi- and Many-Core Architectures. (2014). https://hal.archives-ouvertes.fr/hal-00949355 working
paper or preprint.
[16] J. Condit, M. Harren, Z. R. Anderson, D. Gay, and G. C. Necula. 2007. Dependent Types for Low-Level Programming.
In ESOP.
[17] R. L. Constable and S. F. Smith. 1987. Partial Objects In Constructive Type Theory. In LICS.
[18] P. Cousot and R. Cousot. 1977. Abstract Interpretation: A Unified Lattice Model for the Static Analysis of Programs. In
POPL.
[19] E. W. Dijkstra. 1975. Guarded Commands, Nondeterminacy and Formal Derivation of Programs. In Communications of
the ACM.
[20] E. W. Dijkstra. 1976. A Discipline of Programming. Prentice-Hall.
[21] R. A. Eisenberg and J. Stolarek. 2014. Promoting functions to type families in Haskell. In Haskell.
[22] A. Farmer, N. Sculthorpe, and A. Gill. 2015. Reasoning with the HERMIT: Tool Support for Equational Reasoning on
GHC Core Programs. In Haskell.
[23] G. Gentzen. 1935. Inverstigations into Logical Deduction. In American Philosophical Quarterly.
[24] W. A. Howard. 1980. The formulae-as-types notion of construction. In Essays on Combinatory Logic, Lambda Calculus
and Formalism.
[25] G. Hutton. 1999. A tutorial on the universality and expressiveness of fold. J. Functional Programming (1999).
[26] A. M. Kent, D. Kempe, and S. Tobin-Hochstadt. 2016. Occurrence typing modulo theories. In PLDI.
[27] K. W. Knowles and C. Flanagan. 2010. Hybrid type checking. In TOPLAS.
[28] L. Kuper, A. Turon, N. R Krishnaswami, and R. R Newton. 2014. Freeze after writing: quasi-deterministic parallel
programming with LVars. In POPL.
[29] K. R. M. Leino. 2010. Dafny: An Automatic Program Verifier for Functional Correctness. In LPAR.
[30] K. R. M. Leino and C. Pit-Claudel. 2016. Trigger selection strategies to stabilize program verifiers. In CAV.
[31] K. R. M. Leino and N. Polikarpova. 2016. Verified Calculations. In VSTTE.
[32] R. Leino. 2016. Dafny. (2016). https://github.com/Microsoft/dafny/blob/master/Test/dafny0/Fuel.dfy.
[33] S. Marlow, R. Newton, and S. Peyton-Jones. 2011. A Monad for Deterministic Parallelism. In Haskell.
[34] S. C. Mu, H. S. Ko, and P. Jansson. 2009. Algebra of Programming in Agda: Dependent Types for Relational Program
Derivation. In J. Funct. Program.
[35] G. C. Necula. 1997. Proof carrying code. In POPL.
[36] C. G. Nelson. 1980. Techniques for Program Verification. Ph.D. Dissertation. Stanford University.
[37] T. Nipkow, L.C. Paulson, and M. Wenzel. 2002. Isabelle/HOL — A Proof Assistant for Higher-Order Logic.
[38] U. Norell. 2007. Towards a practical programming language based on dependent type theory. Ph.D. Dissertation.
Chalmers.
[39] X. Ou, G. Tan, Y. Mandelbaum, and D. Walker. 2004. Dynamic Typing with Dependent Types. In IFIP TCS.
[40] J. C. Reynolds. 1972. Definitional interpreters for higher-order programming languages. In ACM National Conference.
[41] S. R. Della Rocca and L. Paolini. 2004. The Parametric Lambda Calculus, A Metamodel for Computation.
[42] P. Rondon, M. Kawaguchi, and R. Jhala. 2008. Liquid Types. In PLDI.
[43] P. Rondon, M. Kawaguchi, and R. Jhala. 2010. Low-Level Liquid Types. In POPL.
[44] J. Rushby, S. Owre, and N. Shankar. 1998. Subtypes for Specifications: Predicate Subtyping in PVS. TSE.
[45] E. L. Seidel, N. Vazou, and R. Jhala. 2015. Type Targeted Testing. In ESOP.
[46] V. Sjöberg and S. Weirich. 2015. Programming Up to Congruence. In POPL.
[47] W. Sonnex, S. Drossopoulou, and S. Eisenbach. 2012. Zeno: An Automated Prover for Properties of Recursive Data
Structures. In TACAS.
[48] M. Sousa and I. Dillig. 2016. Cartesian hoare logic for verifying k-safety properties. In PLDI.
[49] Martin Sulzmann, Gregory J. Duck, Simon P. Jones, and Peter J. Stuckey. 2006. Understanding Functional Dependencies
via Constraint Handling Rules. Journal of Functional Programming (2006).
[50] P. Suter, M. Dotta, and V. Kuncak. 2010. Decision Procedures for Algebraic Data Types with Abstractions. In POPL.
[51] P. Suter, A. Sinan Köksal, and V. Kuncak. 2011. Satisfiability Modulo Recursive Programs. In SAS.
[52] N. Swamy, C. Hriţcu, C. Keller, A. Rastogi, A. Delignat-Lavaud, S. Forest, K. Bhargavan, C. Fournet, P. Y. Strub, M.
Kohlweiss, J. K. Zinzindohoue, and S. Zanella-Béguelin. 2016. Dependent Types and Multi-Monadic Effects in F*. In
POPL.
[53] G. Tourlakis. 2008. Ackermann’s Function. (2008). http://www.cs.yorku.ca/~gt/papers/Ackermann-function.pdf.
[54] N. Vazou, L. Lampropoulos, and J. Polakow. 2017. A Tale of Two Provers: Verifying Monoidal String Matching in
Liquid Haskell and Coq. In Haskell.
[55] N. Vazou, P. Rondon, and R. Jhala. 2013. Abstract Refinement Types. In ESOP.
[56] N. Vazou, E. L. Seidel, R. Jhala, D. Vytiniotis, and S. L. Peyton-Jones. 2014. Refinement Types for Haskell. In ICFP.
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:31
[57] N. Vazou, E. L. Seidel, R. Jhala, D. Vytiniotis, and S. L. Peyton-Jones. 2014. Technical Report: Refinement Types for
Haskell. (2014). http://goto.ucsd.edu/~nvazou/icfp14/haskell-refinements-techrep.pdf.
[58] N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala. 2017. Extended Version: Refinement Reflection: Complete Verification with SMT. (2017). https://nikivazou.github.io/static/popl18/
extended-refinement-reflection.pdf
[59] P. Vekris, B. Cosman, and R. Jhala. 2016. Refinement types for TypeScript. In PLDI.
[60] D. Vytiniotis, S. L. Peyton-Jones, K. Claessen, and D. Rosén. 2013. HALO: Haskell to Logic through Denotational
Semantics. In POPL.
[61] P. Wadler. 1987. A Critique of Abelson and Sussman or Why Calculating is Better Than Scheming. In SIGPLAN Not.
[62] P. Wadler. 2015. Propositions As Types. In Communications of the ACM.
[63] H. Xi and F. Pfenning. 1998. Eliminating Array Bound Checking Through Dependent Types. In PLDI.
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:32 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
A
PROOF OF MAPFUSION WITHOUT PLE
map_fusion f g []
= (map f . map g) []
=. map f (map g [])
=. map f []
=. map (f . g) []
** QED
map_fusion f g (C x xs)
= map (f . g) (x : xs)
=. ((f . g) x) : (map (f . g) xs)
=. ((f . g) x) : (((map f) . (map g)) xs)
? map_fusion f g xs
=. ((f . g) x) : (map f (map g xs))
=. (f (g x)) : (map f (map g xs))
=. map f ((g x) : (map g xs))
=. map f ((g x) : (map g xs))
=. map f (map g (x : xs))
=. map f ((map g) (x : xs))
=. ((map f) . (map g)) (x : xs)
** QED
B
PROOFS FOR PLE
Proofs for § 6
Proof. (Of lemma 6.5) We prove the result by induction on n.
Case n = 0: Immediate as t ≡ t ′.
Case n = k + 1: Consider any t ′ such that SmtValid(Φk +1 , ν = t ′). By definition Φk +1 =
Unfold(Ψ, Φk ), hence SmtValid(Unfold(Ψ, Φk ), v = t, ν = t ′). Consider any f (t ′) ≺ t ′; Lemma 6.6
completes the proof.
□
Proof. (Of lemma 6.6) Proof by induction on the structure of t ′: Case t ′ ≡ x: There are no
subterms, hence immediate. Case t ′ ≡ c: There are no subterms, hence immediate. Case t ′ ≡ f (t ′):
Consider the last link in Φ connecting the equivalence class of ν (and t) to t ′. Suppose the last link
is a congruence link of the form t = t ′ where t ≡ f (t) and SmtValid(Φ, t = t ′). Then f (t) ≺ Φ, t and
we are done. Suppose instead, the last link is an equality link in Φ of the form z = f (t ′). In this case,
f (t ′) ≺ Φ and again, we are done.
□
Proof. (Of lemma 6.7)
Let G k Unfold∗ (Ψ, Φ, k). Let us assume that
SmtValid(Unfold∗ (Ψ, Φ, ν = t, n), ν = t ′)
(3)
q ≡ f (t ′) = bi t ′/x in Unfold(Ψ, Φ ∧ ν = t ′)
(4)
Consider any instance
By the definition of Unfold, we have
f (t ′) ≺ Φ ∧ ν = t ′ such that SmtValid(Φ, pi t ′/x )
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
(5)
53:33
By (3) and Lemma 6.5 there exists f (t) ≺ Φn such that
SmtValid(Φn , t = t ′)
As Φ ⊆ Φn and (5), by congruence
SmtValid(Φn , pi t/x )
Hence, the instance
f (t) = bi t/x is inΦn+1
That is
SmtValid(Φn+1 , t = t ′ ∧ f (t) = bi [t/x])
And so by congruence closure
SmtValid(Φn+1 , q)
As the above holds for every instance, we have
SmtValid(Φn+1 , Unfold(Ψ, Φ ∧ ν = t ′))
□
Proof. (Of lemma 6.8) The proof follows by induction on the structure of Ψ, Φ ⊢ t ↠ t ′.
Base Case Eq-Refl: Follows immediately as t ≡ t ′.
Inductive Case Eq-Trans
In this case, there exists t ′′ such that
Ψ, Φ ⊢ t ↠ t ′′
(6)
SmtValid(Unfold(Ψ, Φ ∧ ν = t ), ν = t )
′′
′
(7)
By the induction hypothesis (6) implies there exists 0 ≤ n such that
SmtValid(Unfold∗ (Ψ, Φ ∧ ν = t, n), ν = t ′′)
By Lemma 6.7 we have
SmtValid(Unfold∗ (Ψ, Φ, ν = t, n + 1), Unfold(Ψ, Φ ∧ ν = t ′′))
Thus, by (7) and modus ponens we get
SmtValid(Unfold∗ (Ψ, Φ, ν = t, n + 1), ν = t ′)
□
Proof. (Of Lemma 6.9) Let Φ′ = Φ ∧ ν = t.
Case ⇒: Assume that PLE(Ψ, Φ, t = t ′). That is, at some iteration i we have SmtValid(Φi , ν = t ′),
i.e. by (6.2) we have SmtValid(Unfold∗ (Ψ, Φ′, i), ν = t ′).
Case ⇐: Pick the smallest n such that SmtValid(Unfold∗ (Ψ, Φ′, n), ν = t ′). Using Lemmas 6.3 and 6.4
we can then show that forall 0 ≤ k < n, we have
Unfold∗ (Ψ, Φ′, k) ⊂ Unfold∗ (Ψ, Φ′, k + 1)
and
Unfold∗ (Ψ, Φ′, k) ⊬ ν = t ′
Hence, after n iterations of the recursive loop, PLE(Ψ, Φ, t = t ′), returns true.
□
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:34 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
Steps and Values Next, we show that if f (y) ⇝ t ′ is a logical step under an Γ that is inhabited by
θ then f (y) reduces to a value under θ . The proof follows by observing that if Γ is inhabited by θ ,
and a particular step is possible, then the guard corresponding to that step must also be true under
θ and hence, by totality, the function must reduce to a value under the given store.
Lemma B.1 (Step-Value). If θ ∈ (|Γ|) and f (y) ⇝ t ′ is a ⌊R⌋, ⌊Γ⌋ step then R[θ · f (y)] ,→⋆ v.
Proof. (Of Lemma B.1)
Assume that θ ∈ (|Γ|)
(8)
Let θ θ [x 7→ θ · y]
∗
(9)
Ψ(f ) λx .⟨p ⇒ b⟩
As f (y) ⇝
t′
(10)
is a ⌊R⌋Γ step, for some i, ⌊R⌋-instance Q we have
SmtValid(⌊Γ⌋ ∧ Q, pi [y/x])
Hence, by (8) and Lemma 6.1 θ · R[pi [y/x]] ,→⋆ True
As θ · pi ≡ θ · pi [y/x]
∗
(11)
(12)
⋆
The fact (11) yields θ · R[pi ] ,→ True
∗
By the Totality Assumption 6.11 R[θ ∗ · f (x)] ,→⋆ v
That is R[θ · f (y)] ,→⋆ v
(13)
(14)
□
Divergence A closed term t diverges under R if there is no v such that R[t] ,→⋆ v .
Lemma B.2 (Divergence). If ∀0 ≤ i we have R[ti ] ,→⋆ C[ti+1 ] then t 0 diverges under Ψ.
Proof. (Of Theorem 6.14)
Assume that θ ∈ (|Γ|)
and assume an infinite ⌊R⌋, ⌊Γ⌋ trace:
(15)
f 0 (t 0 ), f 1 (t 1 ), . . .
(16)
Where additionally t 0 ≡ x 0
Define
By Lemma 6.13, for every i ∈ N
Hence, by Lemma B.2
ti∗
(17)
≡ θ · fi (ti )
(18)
∗
R[ti∗ ] ,→⋆ Ci [ti+1
]
∗
t 0 diverges under Ψ
i.e., by (17, 18) θ · f 0 (x 0 ) diverges under Ψ
⋆
But by (15) and Lemma B.1 R[θ · f 0 (x 0 )] ,→ v
contradicting (19)
Hence, assumption (16) cannot hold, i.e. all all the Ψ, Φ symbolic traces must be finite.
Proof. (Of Theorem 6.16) As Φ is transparent, there are two cases.
Case: Γ is inconsistent.
By definition of inconsistency
(19)
SmtValid(⌊Γ⌋, false)
Hence
SmtValid(⌊Γ⌋, p)
That is
PLE(⌊R⌋, ⌊Γ⌋, p) terminates immediately.
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
□
53:35
Case: Γ is inhabited.
That is, exists θ s.t. θ ∈ (|Γ|)
(20)
Suppose that PLE(⌊R⌋, ⌊Γ⌋, p) does not terminate.
That is, there is an infinitely increasing chain:
By Lemma 6.15
Φ0 ⊂ . . . ⊂ Φn . . .
(21)
⌊R⌋, ⌊Γ⌋ has an infinite trace
Which, by (20) contradicts Theorem 6.14. Thus, (21) is impossible, i.e. PLE(⌊R⌋, ⌊Γ⌋, p) terminates.
□
C
PROOF OF SECTION: EMBEDDING NATURAL DEDUCTION WITH REFINEMENT
TYPES
Lemma C.1 (Validity). If there exists e ∈ (|ϕ |) then ϕ is valid.
Proof. We prove the lemma by case analysis in the shape of ϕ.
• ϕ ≡ { p }. Since the set (|{ p }|) = {e |p ,→⋆ True} is not empty, then p ,→⋆ True.
• ϕ ≡ ϕ 1 → ϕ 2 . By assumption, there exists an expressions f so that ∀e x ∈ (|ϕ 1 |), f e x ∈ (|ϕ 2 |).
So, if there exists an expression e 1 ∈ (|ϕ 1 |) that makes ϕ 1 valid then f e 1 makes ϕ 2 valid.
• ϕ ≡ ϕ → False. By assumption, there exists an expressions f so that ∀e x ∈ (|ϕ |), f e x ∈
(|{ False }|). So, if there exists an expression e 1 ∈ (|ϕ |) that makes ϕ valid then f e 1 makes
{ False } valid, which is impossible, thus ϕ cannot be valid.
• ϕ ≡ (ϕ 1 , ϕ 2 ). If there exists a total expression e ∈ (|ϕ |) then e evaluates to (e 1 , e 2 ).
• ϕ ≡ Either ϕ 1 ϕ 2 . If there exists a total expression e ∈ (|ϕ |) then e evaluates to either Left e ′
or Right e ′.
• ϕ ≡ x : τ → ϕ. By assumption, there exists an expressions f so that ∀e x ∈ (|τ |), f e x ∈
(|ϕ[x/e x ]|). So, if there exists an expression e 1 ∈ (|τ |) then f e 1 makes ϕ[x/e 1 ] valid.
• ϕ ≡ (x :: τ , ϕ). By assumption, there exists an expressions p that evaluates to a pair (e x , ey )
so that e x ∈ (|τ |) and ey ∈ (|ϕ[x/e x ]|).
□
Theorem C.2 (Validity). If ∅; ∅ ⊢ e : ϕ then ϕ is valid.
Proof. By direct implication of Lemma C.1 and soundness of λ R (Theorem 4.1).
D
□
REFINEMENT REFLECTION: λ R : EXTENDED VERSION WITH PROOFS
Our first step towards formalizing refinement reflection is a core calculus λ R with an undecidable
type system based on denotational semantics. We show how the soundness of the type system
allows us to prove theorems using λ R .
D.1
Syntax
Figure 12 summarizes the syntax of λ R , which is essentially the calculus λU [56] with explicit
recursion and a special reflect binding form to denote terms that are reflected into the refinement
logic. In λ R refinements r are arbitrary expressions e (hence r ::= e in Figure 12). This choice allows
us to prove preservation and progress, but renders typechecking undecidable. In § F we will see
how to recover decidability by soundly approximating refinements.
The syntactic elements of λ R are layered into primitive constants, values, expressions, binders
and programs.
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:36 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
Operators
Constants
⊙
c
Values w
Expressions e
Binders
Program
Basic Types
Refined Types
b
p
B
τ
::=
::=
|
::=
::=
|
::=
::=
::=
::=
=|<
∧ | ! | ⊙ | +, −, . . .
True | False | 0, 1, −1, . . .
c | λx .e | D w
w | x | ee
case x = e of {D x → e}
e | let rec x : τ = b in b
b | reflect x : τ = e in p
Int | Bool | T
{v : B | {} | | }x : τ → τ
Fig. 12. Syntax of λ R
Contexts
C
C
::= •
| Ce | cC | DeCe
| case y = C of {D i x → ei }
p ,→ p ′
Reductions
C[p]
cv
(λx .e) e ′
case y = D j e of {D i x i → ei }
reflect x : τ = e in p
let rec x : τ = bx in b
fix p
,→
,→
,→
,→
,→
,→
,→
C[p ′], if p ,→ p ′
δ (c, v)
e [e ′/x]
e j D j e/y [e/x i ]
p [fix (λx .e)/x]
b [fix (λx .bx )/x]
p (fix p)
Fig. 13. Operational Semantics of λ R
Constants The primitive constants of λ R include all the primitive logical operators ⊕, here, the
set {=, <}. Moreover, they include the primitive booleans True, False, integers −1, 0, 1, etc., and
logical operators ∧, ∨, !, etc..
Data Constructors We encode data constructors as special constants. For example the data type
[Int], which represents finite lists of integers, has two data constructors: [] (“nil”) and : (“cons”).
Values & Expressions The values of λ R include constants, λ-abstractions λx .e, and fully applied data constructors D that wrap values. The expressions of λ R include values and variables x,
applications e e, and case expressions.
Binders & Programs A binder b is a series of possibly recursive let definitions, followed by an
expression. A program p is a series of reflect definitions, each of which names a function that
can be reflected into the refinement logic, followed by a binder. The stratification of programs via
binders is required so that arbitrary recursive definitions are allowed but cannot be inserted into
the logic via refinements or reflection. (We can allow non-recursive let binders in e, but omit them
for simplicity.)
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:37
D.2
Operational Semantics
Figure 12 summarizes the small step contextual β-reduction semantics for λ R . We write e ,→j e ′
if there exist e 1 , . . . , e j such that e is e 1 , e ′ is e j and ∀i, j, 1 ≤ i < j, we have ei ,→ ei+1 . We write
e ,→⋆ e ′ if there exists some finite j such that e ,→j e ′. We define ≈β to be the reflexive, symmetric,
transitive closure of ,→ .
Constants Application of a constant requires the argument be reduced to a value; in a single
step the expression is reduced to the output of the primitive constant operation. For example,
consider =, the primitive equality operator on integers. We have δ (=, n) =n where δ (=n , m)
equals True iff m is the same as n. We assume that the equality operator is defined for all values,
and, for functions, is defined as extensional equality. That is, for all f and f ′ we have (f = f ′) ,→
True iff ∀v. f v ≈β f ′ v. We assume source terms only contain implementable equalities over
non-function types; the above only appears in refinements and allows us to state and prove facts
about extensional equality § H.2.
D.3
Types
λR
types include basic types, which are refined with predicates, and dependent function types. Basic
types B comprise integers, booleans, and a family of data-types T (representing lists, trees etc..)
For example the data type [Int] represents lists of integers. We refine basic types with predicates
(boolean valued expressions e) to obtain basic refinement types {v : B | e}. Finally, we have dependent
function types x : τx → τ where the input x has the type τx and the output τ may refer to the input
binder x. We write B to abbreviate {v : B | True}, and τx → τ to abbreviate x : τx → τ if x does
not appear in τ . We use r to refer to refinements.
Denotations Each type τ denotes a set of expressions [[τ ]], that are defined via the dynamic
semantics [27]. Let shape(τ ) be the type we get if we erase all refinements from τ and e : shape(τ )
be the standard typing relation for the typed lambda calculus. Then, we define the denotation of
types as:
[[{x : B | r }]] {e | e : B, if e ,→⋆ w then r [x/w] ,→⋆ True}
[[x : τx → τ ]] {e | e : shape(τx → τ ), ∀e x ∈ [[τx ]]. e e x ∈ [[τ [x/e x ]]]}
Constants For each constant c we define its type prim(c) such that c ∈ [[prim(c)]]. For example,
prim(3)
prim(+)
prim(≤)
{v : Int | v = 3}
x : Int → y : Int → {v : Int | v = x + y}
x : Int → y : Int → {v : Bool | v ⇔ x ≤ y}
So, by definition we get the constant typing lemma
Lemma D.1. [Constant Typing] Every constant c ∈ [[prim(c)]].
Thus, if prim(c) x : τx → τ , then for every value w ∈ [[τx ]], we require δ (c, w) ∈ [[τ [x/w]]].
D.4
Refinement Reflection
The simple, but key idea in our work is to strengthen the output type of functions with a refinement
that reflects the definition of the function in the logic. We do this by treating each reflect-binder:
reflect f : τ = e in p as a let rec-binder: let rec f : Reflect(τ , e) = e in p during type checking
(rule T-Refl in Figure 7).
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:38 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
Reflection We write Reflect(τ , e) for the reflection of term e into the type τ , defined by strengthening τ as:
Reflect({v : B | r }, e)
{v : B | r ∧ v = e}
Reflect(x : τx → τ , λy.e) x : τx → Reflect(τ , e[y/x])
As an example, recall from § 2 that the reflect fib strengthens the type of fib with the reflected
refinement fibP.
Consequences for Verification Reflection has two consequences for verification. First, the reflected refinement is not trusted; it is itself verified (as a valid output type) during type checking.
Second, instead of being tethered to quantifier instantiation heuristics or having to program “triggers” as in Dafny [29] or F* [52] the programmer can predictably “unfold” the definition of the
function during a proof simply by “calling” the function, which we have found to be a very natural
way of structuring proofs § 7.
D.5
Refining & Reflecting Data Constructors with Measures
We assume that each data type is equipped with a set of measures which are unary functions whose
(1) domain is the data type, and (2) body is a single case-expression over the datatype [56]:
measure f : τ = λx .case y = x of {D i z → ei }
For example, len measures the size of an [Int]:
measure len :: [ Int ] → Nat
len = \ x → case x of
[]
→ 0
(x: xs ) → 1 + len xs
Checking and Projection We assume the existence of measures that check the top-level constructor, and project their individual fields. In § F.2 we show how to use these measures to reflect
functions over datatypes. For example, for lists, we assume the existence of measures:
isNil []
isNil ( x : xs )
= True
= False
isCons ( x : xs ) = True
isCons []
= False
sel1 ( x : xs )
sel2 ( x : xs )
= x
= xs
Refining Data Constructors with Measures We use measures to strengthen the types of data
constructors, and we use these strengthened types during construction and destruction (patternmatching). Let: (1) D be a data constructor, with unrefined type x : τ → T (2) the i-th measure
definition with domain T is:
measure fi : τ = λx .case y = x of {D z → ei }
Then, the refined type of D is defined:
prim(D) x : τ → {v : T | ∧i fi v = ei [x/z]}
Thus, each data constructor’s output type is refined to reflect the definition of each of its measures.
For example, we use the measures len, isNil, isCons, sel1, and sel2 to strengthen the types of []
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:39
and : to:
prim([]) {v : [Int] | r [] }
prim(:) x : Int → xs : [Int] → {v : [Int] | r : }
where the output refinements are
r [] len v = 0 ∧ isNil v ∧ !isCons v
r : len v = 1 + len xs ∧ !isNil v ∧ isCons v
∧ sel1 v = x ∧ sel2 v = xs
It is easy to prove that Lemma D.1 holds for data constructors, by construction. For example,
len [] = 0 evaluates to true.
D.6
Typing Rules
Next, we present the type-checking judgments and rules of λ R .
Environments and Closing Substitutions A type environment Γ is a sequence of type bindings
x 1 : τ1 , . . . , x n : τn . An environment denotes a set of closing substitutions θ which are sequences of
expression bindings: x 1 7→ e 1 , . . . , x n 7→ en such that:
[[Γ]] {θ | ∀x : τ ∈ Γ.θ (x) ∈ [[θ · τ ]]}
Judgments We use environments to define three kinds of rules: Well-formedness, Subtyping,
and Typing [27, 56]. A judgment Γ ⊢ τ states that the refinement type τ is well-formed in the
environment Γ. Intuitively, the type τ is well-formed if all the refinements in τ are Bool-typed in Γ.
A judgment Γ ⊢ τ1 ⪯ τ2 states that the type τ1 is a subtype of τ2 in the environment Γ. Informally, τ1
is a subtype of τ2 if, when the free variables of τ1 and τ2 are bound tomeasures expressions described
by Γ, the denotation of τ1 is contained in the denotation of τ2 . Subtyping of basic types reduces
to denotational containment checking. That is, for any closing substitution θ in the denotation
of Γ, for every expression e, if e ∈ [[θ · τ1 ]] then e ∈ [[θ · τ2 ]]. A judgment Γ ⊢ p : τ states that the
program p has the type τ in the environment Γ. That is, when the free variables in p are bound to
expressions described by Γ, the program p will evaluate to a value described by τ .
Rules All but three of the rules are standard [27, 56]. First, rule T-Refl is used to strengthen the
type of each reflected binder with its definition, as described previously in § D.4. Second, rule
T-Exact strengthens the expression with a singleton type equating the value and the expression
(i.e. reflecting the expression in the type). This is a generalization of the “selfification” rules from
[27, 39], and is required to equate the reflected functions with their definitions. For example, the
application (fib 1) is typed as {v : Int | fibP v 1 ∧ v = fib 1} where the first conjunct comes
from the (reflection-strengthened) output refinement of fib § 2, and the second conjunct comes
from rule T-Exact. Finally, rule T-Fix is used to type the intermediate fix expressions that appear,
not in the surface language but as intermediate terms in the operational semantics.
Soundness Following λU [56], we can show that evaluation preserves typing and that typing
implies denotational inclusion.
Theorem D.2. [Soundness of λ R ]
• Denotations If Γ ⊢ p : τ then ∀θ ∈ [[Γ]].θ · p ∈ [[θ · τ ]].
• Preservation If ∅ ⊢ p : τ and p ,→⋆ w then ∅ ⊢ w : τ .
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:40 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
Γ ⊢p :τ
Typing
x :τ ∈ Γ
Γ ⊢x :τ
T-Var
Γ ⊢ p : τ′
T-Con
Γ ⊢ c : prim(c)
Γ ⊢ τ′ ⪯ τ
Γ ⊢p :τ
Γ ⊢ e : {v : B | { | r }}
T-Sub
T-Exact
Γ ⊢ e : {v : B | { | r } ∧ v = e}
Γ, x : τx ⊢ e : τ
T-Fun
Γ ⊢ λx .e : x : τx → τ
Γ ⊢ e 1 : (x : τx → τ )
Γ ⊢ e 2 : τx
Γ ⊢ e1 e2 : τ
T-App
Γ, x : τx ⊢ bx : τx
Γ, x : τx ⊢ τx
Γ, x : τx ⊢ b : τ
Γ ⊢τ
Γ ⊢ let rec x : τx = bx in b : τ
T-Let
Γ ⊢ let rec f : Reflect(τf , e) = e in p : τ
Γ ⊢ reflect f : τf = e in p : τ
Γ ⊢ e : {v : T | er }
Γ ⊢τ
∀i.prim(D i ) = y j : τ j → {v : T | er i }
Γ, y j : τ j , x : {v : T | er ∧ er i } ⊢ ei : τ
Γ ⊢ case x = e of {D i yi → ei } : τ
T-Refl
T-Case
Γ ⊢τ
Well Formedness
Γ, v : B ⊢ e : Bool⇓
WF-Base
Γ ⊢ {v : B | e}
Γ ⊢ τx
Γ, x : τx ⊢ τ
WF-Fun
Γ ⊢ x : τx → τ
Γ ⊢ τ1 ⪯ τ2
Subtyping
Γ ′ Γ, v : {B ⇓ |e}
Γ′ ⊢ e ′ ⇝ p′
SmtValid(⌊Γ ′⌋ ⇒ p ′)
Γ ⊢ {v : B | e} ⪯ {v : B | e ′ }
Γ ⊢ τx′ ⪯ τx
Γ, x : τx′ ⊢ τ ⪯ τ ′
Γ ⊢ x : τx → τ ⪯ x : τx′ → τ ′
⪯-Base-λ S
⪯-Fun
Fig. 14. Typing of λ R
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:41
D.7
From Programs & Types to Propositions & Proofs
The denotational soundness Theorem D.2 lets us interpret well typed programs as proofs of
propositions.
“Definitions” A definition d is a sequence of reflected binders:
d ::= • | reflect x : τ = e in d
A definition’s environment Γ(d) comprises its binders and their reflected types:
Γ(•) ∅
Γ(reflect f : τ = e in d) (f , Reflect(τ , e)), Γ(d)
A definition’s substitution θ (d) maps each binder to its definition:
θ (•) []
θ (reflect f : τ = e in d) [[f /fix f e], θ (d)]
“Propositions” A proposition is a type
x 1 : τ1 → . . . → x n : τn → {v : Unit | prop}
For brevity, we abbreviate propositions like the above to x : τ → {prop} and we call prop the
proposition’s refinement. For simplicity we assume that fv(τi ) = ∅.
“Validity”
A proposition x : τ → {prop} is valid under d if
∀w ∈ [[τ ]]. θ (d) · prop [w/x] ,→⋆ True
That is, the proposition is valid if its refinement evaluates to True for every (well typed) interpretation for its parameters x under d.
“Proofs” A binder b proves a proposition τ under d if
∅ ⊢ d[let rec x : τ = b in unit] : Unit
That is, if the binder b has the proposition’s type τ under the definition d’s environment.
Theorem D.3. [Proofs] If b proves τ under d then τ is valid under d.
Proof. As b proves τ under d, we have
∅ ⊢ d[let rec x : τ = b in unit] : Unit
(22)
By Theorem D.2 on 22 we get
θ (d) ∈ [[Γ(d)]]
(23)
Furthermore, by the typing rules 22 implies Γ(d) ⊢ b : τ and hence, via Theorem D.2
∀θ ∈ [[Γ(d)]]. θ · b ∈ [[θ · τ ]]
(24)
θ (d) · b ∈ [[θ (d) · τ ]]
(25)
[[θ (d) · τ ]] { f | τ is valid under d }
(26)
Together, 23 and 24 imply
By the definition of type denotations, we have
By 25, the above set is not empty, and hence τ is valid under d.
□
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:42 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
Example: Fibonacci is increasing In § 2 we verified that under a definition d that includes fib,
the term fibUp proves
n : Nat → {fib n ≤ fib (n + 1)}
Thus, by Theorem D.3 we get
∀n.0 ≤ n ,→⋆ True ⇒ fib n ≤ fib (n + 1) ,→⋆ True
E
PROOF OF SOUNDNESS
We prove Theorem 4.1 of § D by reduction to Soundness of λU [56].
Theorem E.1. [Denotations] If Γ ⊢ p : τ then ∀θ ∈ [[Γ]].θ · p ∈ [[θ · τ ]].
Proof. We use the proof from [57] and specifically Lemma 4 that is identical to the statement
we need to prove. Since the proof proceeds by induction in the type derivation, we need to ensure
that all the modified rules satisfy the statement.
• T-Exact Assume Γ ⊢ e : {v : B | { | r } ∧ v = e}. By inversion Γ ⊢ e : {v : B | { | r }}(1). By
(1) and IH we get ∀θ ∈ [[Γ]].θ · e ∈ [[θ · {v : B | { | r }}]]. We fix a θ ∈ [[Γ]] We get that if
θ · e ,→⋆ w, then θ · { | r }[v/w] ,→⋆ True. By the Definition of = we get that w = w ,→⋆
True. Since θ · (v = e)[v/w] ,→⋆ w = w, then θ · ({ | r } ∧ v = e)[v/w] ,→⋆ True. Thus θ · e ∈
[[θ · {v : B | { | r } ∧ v = e}]] and since this holds for any fixed θ , ∀θ ∈ [[Γ]].θ · e ∈ [[θ ·
{v : B | { | r } ∧ v = e}]].
• T-Let Assume Γ ⊢ let rec x : τx = e x in p : τ . By inversion Γ, x : τx ⊢ e x : τx (1), Γ, x : τx ⊢
p : τ (2), and Γ ⊢ τ (3). By IH ∀θ ∈ [[Γ, x : τx ]].θ ·e x ∈ [[θ ·τx ]] (1’) ∀θ ∈ [[Γ, x : τx ]].θ ·p ∈ [[θ ·τ ]]
(2’). By (1’) and by the type of fix ∀θ ∈ [[Γ, x : τx ]].θ · fix x e x ∈ [[θ · τx ]]. By which, (2’) and
(3) ∀θ ∈ [[Γ]].θ · p [fix x e x /x] ∈ [[θ · τ ]].
• T-Refl Assume Γ ⊢ reflect f : τf = e in p : τ . By inversion, Γ ⊢ let rec f : Reflect(τf , e) =
e in p : τ . By IH, ∀θ ∈ [[Γ]].θ · let rec f : Reflect(τf , e) = e in p ∈ [[θ · τ ]]. Since denotations
are closed under evaluation, ∀θ ∈ [[Γ]].θ · reflect f : Reflect(τf , e) = e in p ∈ [[θ · τ ]].
• T-Fix In Theorem 8.3 from [57] (and using the textbook proofs from [41]) we proved that for
each type τ , fix τ ∈ [[(τ → τ ) → τ ]].
□
Theorem E.2. [Preservation] If ∅ ⊢ p : τ and p ,→⋆ w then ∅ ⊢ w : τ .
Proof. In [57] proof proceeds by iterative application of Type Preservation Lemma 7. Thus, it
suffices to ensure Type Preservation in λ R , which it true by the following Lemma.
□
Lemma E.3. If ∅ ⊢ p : τ and p ,→ p ′ then ∅ ⊢ p ′ : τ .
Proof. Since Type Preservation in λU is proved by induction on the type derivation tree, we
need to ensure that all the modified rules satisfy the statement.
• T-Exact Assume ∅ ⊢ p : {v : B | { | r }∧v = p}. By inversion ∅ ⊢ p : {v : B | { | r }}. By IH we get
∅ ⊢ p ′ : {v : B | { | r }}. By rule T-Exact we get ∅ ⊢ p ′ : {v : B | { | r } ∧v = p ′ }. Since subtyping
is closed under evaluation, we get ∅ ⊢ {v : B | { | r } ∧ v = p ′ } ⪯ {v : B | { | r } ∧ v = p}. By rule
T-Sub we get ∅ ⊢ p ′ : {v : B | { | r } ∧ v = p}.
• T-Let Assume ∅ ⊢ let rec x : τx = e x in p : τ . By inversion, x : τx ⊢ e x : τx (1), x : τx ⊢ p : τ
(2), and Γ ⊢ τ (3). By rule T-Fix x : τx ⊢ fix x e x : τx (1’). By (1’), (2) and Lemma 6 of [57],
we get ⊢ p [fix x e x /x] : τ [fix x e x /x]. By (3) τ [fix x e x /x] ≡ τ . Since p ′ ≡ p [fix x e x /x],
we have ∅ ⊢ p ′ : τ .
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:43
Predicates
p
Integers
Booleans
Bin Operators
Un Operators
Model
Sort Arguments
Sorts
n
b
▷◁
⊕1
σ
sa
s
::=
|
|
::=
::=
::=
::=
::=
::=
::=
p ▷◁ p | ⊕1p
n | b | x | D | xp
if p then p else p
0, −1, 1, . . .
True | False
= | < | ∧ | + | − | ...
! | ...
σ , (x : p) | ∅
Int | Bool | U | Fun sa sa
sa → s
Fig. 15. Syntax of λS
• T-Refl Assume ∅ ⊢ reflect x : τx = e x in p : τ . By double inversion, with τx′ ≡ Reflect(τx , e x );
x : τx′ ⊢ e x : τx′ (1), x : τx′ ⊢ p : τ (2), and Γ ⊢ τ (3). By rule T-Fix x : τx′ ⊢ fix x e x : τx′ (1’). By (1’),
(2) and Lemma 6 of [57], we get ⊢ p [fix x e x /x] : τ [fix x e x /x]. By (3) τ [fix x e x /x] ≡ τ .
Since p ′ ≡ p [fix x e x /x], we have ∅ ⊢ p ′ : τ .
• T-Fix This case cannot occur, as fix does not evaluate to any program.
□
F
ALGORITHMIC CHECKING λS : EXTENDED VERSION
Next, we describe λS , a conservative approximation of λ R where the undecidable type subsumption
rule is replaced with a decidable one, yielding an SMT-based algorithmic type system that enjoys
the same soundness guarantees.
F.1
The SMT logic λS
Syntax: Terms & Sorts Figure 15 summarizes the syntax of λS , the sorted (SMT-) decidable logic
of quantifier-free equality, uninterpreted functions and linear arithmetic (QF-EUFLIA) [5, 36]. The
terms of λS include integers n, booleans b, variables x, data constructors D (encoded as constants),
fully applied unary ⊕1 and binary ▷◁ operators, and application x p of an uninterpreted function x.
The sorts of λS include built-in integer Int and Bool for representing integers and booleans. The
interpreted functions of λS , e.g. the logical constants = and <, have the function sort s → s. Other
functional values in λ R , e.g. reflected λ R functions and λ-expressions, are represented as first-order
values with uninterpreted sort Fun s s. The universal sort U represents all other values.
Semantics: Satisfaction & Validity An assignment σ is a mapping from variables to terms
σ {x 1 7→ p1 , . . . , x n 7→ pn }. We write σ |= p if the assignment σ is a model of p, intuitively if
σ p “is true” [36]. A predicate p is satisfiable if there exists σ |= p. A predicate p is valid if for all
assignments σ |= p.
F.2
Transforming λ R into λS
The judgment Γ ⊢ e ⇝ p states that a λ R term e is transformed, under an environment Γ, into a λS
term p. The transformation rules are summarized in Figure 16.
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:44 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
Γ ⊢e ⇝p
Transformation
− |= B ,→∗ ool
Γ ⊢b ⇝b
Γ ⊢ e 1 ⇝ p1
Γ ⊢ e 2 ⇝ p2
Γ ⊢ e 1 ▷◁ e 2 ⇝ p1 ▷◁ p2
Γ ⊢e ⇝p
− |= U ,→∗ n
Γ ⊢ ⊕1e ⇝ ⊕1p
− |= O ,→∗ p
Γ ⊢ c ⇝ sc
Γ, x : τx ⊢ e ⇝ p
− |= I ,→∗ nt
Γ ⊢n⇝n
− |= B ,→∗ in
Γ ⊢ (λx .e) : (x : τx → τ )
Γ ⊢e ⇝p
Γ ⊢ e e′ ⇝
Γ ⊢e ⇝p
x |)
app(|τ
(|τ |)
− |= D ,→∗ C
Γ ⊢ D ⇝ sD
− |= F ,→∗ un
x |)
Γ ⊢ λx .e ⇝ lam(|τ
xp
(|τ |)
Γ ⊢ e′ ⇝ p′
− |= V ,→∗ ar
Γ⊢x ⇝x
Γ ⊢ e : τx → τ
− |= A ,→∗ pp
p p′
Γ ⊢ ei [x/e] ⇝ pi
Γ ⊢ case x = e of {True → e 1 ; False → e 2 }
⇝ if p then p1 else p2
Γ ⊢e ⇝p
Γ ⊢ ei [yi /selD i x][x/e] ⇝ pi
Γ ⊢ case x = e of {D i yi → ei }
⇝ if app isD 1 p then p1 else . . . else pn
− |= I ,→∗ f
− |= C ,→∗ ase
Fig. 16. Transforming λ R terms into λS .
Embedding Types We embed λ R types into λS sorts as:
(|Int|)
(|Bool|)
Int
Bool
(|T |)
(|x : τx → τ |)
U
Fun (|τx |) (|τ |)
Embedding Constants Elements shared on both λ R and λS translate to themselves. These elements include booleans (− |= B ,→∗ ool), integers (− |= I ,→∗ nt), variables (− |= V ,→∗ ar), binary
(− |= B ,→∗ in) and unary (− |= U ,→∗ n) operators. SMT solvers do not support currying, and so
in λS , all function symbols must be fully applied. Thus, we assume that all applications to primitive
constants and data constructors are saturated, i.e. fully applied, e.g. by converting source level terms
like (+ 1) to (\z → z + 1).
Embedding Functions As λS is a first-order logic, we embed λ-abstraction and application using the uninterpreted functions lam and app. We embed λ-abstractions using lam as shown in
rule − |= F ,→∗ un. The term λx .e of type τx → τ is transformed to lamssx x p of sort Fun s x s,
where s x and s are respectively (|τx |) and (|τ |), lamssx is a special uninterpreted function of sort
s x → s → Fun s x s, and x of sort s x and r of sort s are the embedding of the binder and body,
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:45
respectively. As lam is just an SMT-function, it does not create a binding for x. Instead, the binder x
is renamed to a fresh name pre-declared in the SMT environment.
Embedding Applications Dually, we embed applications via defunctionalization [40] using an
uninterpreted apply function app as shown in rule − |= A ,→∗ pp. The term e e ′, where e and e ′
have types τx → τ and τx , is transformed to appssx p p ′ : s where s and s x are respectively (|τ |) and
(|τx |), the appssx is a special uninterpreted function of sort Fun s x s → s x → s, and p and p ′ are the
respective translations of e and e ′.
Embedding Data Types Rule − |= D ,→∗ C translates each data constructor to a predefined λS
constant sD of sort (|prim(D)|). Let D i be a non-boolean data constructor such that
prim(D i ) τi,1 → · · · → τi,n → τ
Then the check function isD i has the sort Fun (|τ |) Bool, and the select function selD i, j has the sort
Fun (|τ |) (|τi, j |). Rule − |= C ,→∗ ase translates case-expressions of λ R into nested if terms in λS , by
using the check functions in the guards, and the select functions for the binders of each case. For
example, following the above, the body of the list append function
[]
++ ys = ys
(x : xs ) ++ ys = x : ( xs ++ ys )
is reflected into the λS refinement:
if isNil xs then ys else sel1 xs : (sel2 xs ++ ys)
We favor selectors to the axiomatic translation of HALO [60] and F* [52] to avoid universally
quantified formulas and the resulting instantiation unpredictability.
F.3
Correctness of Translation
Informally, the translation relation Γ ⊢ e ⇝ p is correct in the sense that if e is a terminating boolean
expression then e reduces to True iff p is SMT-satisfiable by a model that respects β-equivalence.
Definition F.1 (β-Model). A β−model σ β is an extension of a model σ where lam and app satisfy
the axioms of β-equivalence:
∀x y e.lam x e
∀x e x e.(app (lam x e) e x
= lam y (e[x/y])
= e[x/e x ]
Semantics Preservation We define the translation of a λ R term into λS under the empty environment as (|e |) p if ∅ ⊢ e ⇝ p. A lifted substitution θ ⊥ is a set of models σ where each “bottom” in
the substitution θ is mapped to an arbitrary logical value of the respective sort [56]. We connect
the semantics of λ R and translated λS via the following theorems:
Theorem F.2. If Γ ⊢ e ⇝ p, then for every θ ∈ [[Γ]] and every σ ∈ θ ⊥ , if θ ⊥ · e ,→⋆ v then σ β |=
p = (|v |).
Corollary F.3. If Γ ⊢ e : Bool, e reduces to a value and Γ ⊢ e ⇝ p, then for every θ ∈ [[Γ]] and
every σ ∈ θ ⊥ , θ ⊥ · e ,→⋆ True iff σ β |= p.
F.4
Decidable Type Checking
Figure 17 summarizes the modifications required to obtain decidable type checking. Namely, basic
types are extended with labels that track termination and subtyping is checked via an SMT solver.
Termination Under arbitrary beta-reduction semantics (which includes lazy evaluation), soundness of refinement type checking requires checking termination, for two reasons: (1) to ensure that
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:46 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
Refined Types τ
::=
{v : B [⇓] | {} | | }x : τ → τ
Γ ⊢S τ
Well Formedness
Γ, v : B ⊢S e : Bool⇓
Γ ⊢S {v : B | e}
WF-Base
Γ ⊢S τ ⪯ τ ′
Subtyping
Γ ′ Γ, v : {B ⇓ |e}
Γ′ ⊢ e ′ ⇝ p′
SmtValid(⌊Γ ′⌋ ⇒ p ′)
Γ ⊢S {v : B | e} ⪯ {v : B | e ′ }
⪯-Base-λ S
Fig. 17. Algorithmic Typing (other rules in Figs 12 and 14.)
refinements cannot diverge, and (2) to account for the environment during subtyping [56]. We use
⇓ to mark provably terminating computations, and extend the rules to use refinements to ensure
that if Γ ⊢S e : {v : B ⇓ | r }, then e terminates [56].
Verification Conditions The verification condition (VC) ⌊Γ⌋ ⇒ p is valid only if the set of values
described by Γ, is subsumed by the set of values described by p. Γ is embedded into logic by
conjoining (the embeddings of) the refinements of provably terminating binders [56]:
(|Γ|)
Û
(|Γ, x |)
x ∈Γ
where we embed each binder as
(
p
(|Γ, x |)
True
if Γ(x) = {v : B ⇓ | e}, Γ ⊢ e[v/x] ⇝ p
otherwise.
Subtyping via SMT Validity We make subtyping, and hence, typing decidable, by replacing the
denotational base subtyping rule ⪯ -Base-λS with a conservative, algorithmic version that uses an
SMT solver to check the validity of the subtyping VC. We use Corollary F.3 to prove soundness of
subtyping.
Lemma F.4. If Γ ⊢S {v : B | e 1 } ⪯ {v : B | e 2 } then Γ ⊢ {v : B | e 1 } ⪯ {v : B | e 2 }.
Soundness of λS Lemma F.4 directly implies the soundness of λS .
Theorem F.5 (Soundness of λS ). If Γ ⊢S e : τ then Γ ⊢ e : τ .
G
SOUNDNESS OF ALGORITHMIC VERIFICATION
In this section we prove soundness of Algorithmic verification, by proving the theorems of § 5 by
referring to the proofs in [57].
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:47
G.1
Transformation
Definition G.1 (Initial Environment). We define the initial SMT environment ∆0 to include
sc
lamssx
appssx
sD
isD
selD i
x is
Where
x is
: (|prim(c)|)
: s x → s → Fun s x s
: Fun s x s → s x → s
: (|prim(D)|)
: (|T → Bool|)
: (|T → τi |)
: s
∀c ∈ λ R
∀s x , s ∈ λS
∀s x , s ∈ λS
∀D ∈ λ R
∀D ∈ λ R of data type T
∀D ∈ λ R of data type T
and i-th argument τi
∀s ∈ λS and1 ≤ i ≤ M λ
are M λ global names that only appear as lambda arguments.
We modify the − |= F ,→∗ un rule to ensure that logical abstraction is performed using the
minimum globally defined lambda argument that is not already abstracted. We do so, using the
helper function MaxLam(s, p):
MaxLam(s, lamss ′ x is p) =max(i, MaxLam(s, p))
MaxLam(s, r r ) =max(MaxLam(s, p, p))
MaxLam(s, p1 ▷◁ p2 ) =max(MaxLam(s, p1 , p2 ))
MaxLam(s, ⊕1p) =MaxLam(s, p)
MaxLam(s, if p then p1 else p2 ) =max(MaxLam(s, p, p1 , p2 ))
MaxLam(s, p) =0
i = MaxLam((|τx |), p)
Γ, y : τx ⊢ e[x/y] ⇝ p
Γ ⊢ λx .e ⇝
(|τx |)
i < Mλ
y = x i+1
Γ ⊢ (λx .e) : (x : τx → τ )
x |)
lam(|τ
(|τ |)
yp
− |= F ,→∗ un
Lemma G.2 (Type Transformation). If Γ ⊢ e ⇝ p, and Γ ⊢ e : τ , then ∆0 , (|Γ|) ⊢S p:(|τ |).
Proof. We proceed by induction on the translation
• − |= B ,→∗ ool: Since (|Bool|) = Bool, If Γ ⊢ b : Bool, then ∆0 , (|Γ|) ⊢S b:(|Bool|).
• − |= I ,→∗ nt: Since (|Int|) = Int, If Γ ⊢ n : Int, then ∆0 , (|Γ|) ⊢S n:(|Int|).
• − |= U ,→∗ n: Since Γ ⊢ ! e : τ , then it should be Γ ⊢ e : Bool and τ ≡ Bool. By IH,
∆0 , (|Γ|) ⊢S p:(|Bool|), thus ∆0 , (|Γ|) ⊢S !p:(|Bool|).
• − |= B ,→∗ in Assume Γ ⊢ e 1 ▷◁ e 2 ⇝ p1 ▷◁ p2 . By inversion Γ ⊢ e 1 ⇝ p1 , and Γ ⊢ e 2 ⇝ p2 .
Since Γ ⊢ e 1 ▷◁ e 2 : τ , then Γ ⊢ e 1 : τ1 and Γ ⊢ e 1 : τ2 . By IH, ∆0 , (|Γ|) ⊢S p1 :(|τ1 |) and
∆0 , (|Γ|) ⊢S p2 :(|τ2 |). We split cases on ▷◁
– If ▷◁≡=, then τ1 = τ2 , thus (|τ1 |) = (|τ2 |) and (|τ |) = τ = Bool.
– If ▷◁≡<, then τ1 = τ2 = Int, thus (|τ1 |) = (|τ2 |) = Int and (|τ |) = τ = Bool.
– If ▷◁≡ ∧, then τ1 = τ2 = Bool, thus (|τ1 |) = (|τ2 |) = Bool and (|τ |) = τ = Bool.
– If ▷◁≡ + or ▷◁≡ −, then τ1 = τ2 = Int, thus (|τ1 |) = (|τ2 |) = Int and (|τ |) = τ = Int.
• − |= V ,→∗ ar: Assume Γ ⊢ x ⇝ x Then Γ ⊢ x : Γ(x) and ∆0 , (|Γ|) ⊢S x:(|Γ|)(x). But by
definition ((|Γ|))(x) = (|Γ(x)|).
• − |= O ,→∗ p: Assume Γ ⊢ c ⇝ sc Also, Γ ⊢ c : prim(c) and ∆0 , (|Γ|) ⊢S sc :∆0 (sc ). But by
Definition G.1 ∆0 (sc ) = (|prim(c)|).
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:48 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
• − |= D ,→∗ C: Assume Γ ⊢ D ⇝ sD Also, Γ ⊢ D : prim(D) and ∆0 , (|Γ|) ⊢S sD :∆0 (sD ). But by
Definition G.1 ∆0 (sD ) = (|prim(c)|).
x |)
• − |= F ,→∗ un: Assume Γ ⊢ λx .e ⇝ lam(|τ
x i(|τx |) p. By inversion i ≤ M λ Γ, x i(|τx |) : τx ⊢
(|τ |)
e[x/x i(|τx |) ] ⇝ p, and Γ ⊢ (λx .e) : (x : τx → τ ). By the Definition G.1 on lam, x is and induction,
x |)
we get ∆0 , (|Γ|) ⊢S lam(|τ
x i(|τx |) p:Fun (|τx |) (|τ |). By the definition of the type embeddings we
(|τ |)
have (|x → τx τ |) Fun (|τx |) (|τ |).
x |)
• − |= A ,→∗ pp: Assume Γ ⊢ e e ′ ⇝ app(|τ
p p ′. By inversion, Γ ⊢ e ′ ⇝ p ′, Γ ⊢ e ⇝ p,
(|τ |)
x |)
Γ ⊢ e : x : τx → τ . By IH and the type of app we get that ∆0 , (|Γ|) ⊢S app(|τ
p p ′:(|τ |).
(|τ |)
• − |= I ,→∗ f : Assume Γ ⊢ case x = e of {True → e 1 ; False → e 2 } ⇝ if p then p1 else p2
Since Γ ⊢ case x = e of {True → e 1 ; False → e 2 } : τ , then Γ ⊢ e : Bool, Γ ⊢ e 1 : τ , and
Γ ⊢ e 2 : τ . By inversion and IH, ∆0 , (|Γ|) ⊢S p:Bool, ∆0 , (|Γ|) ⊢S p1 :(|τ |), and ∆0 , (|Γ|) ⊢S p2 :(|τ |).
Thus, ∆0 , (|Γ|) ⊢S if p then p1 else p2 :(|τ |).
• − |= C ,→∗ ase: Assume Γ ⊢ case x = e of {D i yi → ei } ⇝ if isD 1 p then p1 else . . . else pn
and Γ ⊢ case x = e of {D i yi → ei } : τ . By inversion we get Γ ⊢ e ⇝ p and Γ ⊢
ei [yi /selD i x][x/e] ⇝ pi . By IH and the Definition G.1 on the checkers and selectors, we
get ∆0 , (|Γ|) ⊢S if isD 1 p then p1 else . . . else pn :(|τ |).
□
Theorem G.3. If Γ ⊢ e ⇝ p, then for every substitution θ ∈ [[Γ]] and every model σ ∈ [[θ ⊥ ]], if
· e ,→⋆ v then σ β |= p = [[v]].
θ⊥
Proof. We proceed using the notion of tracking substitutions from Figure 8 of [57]. Since
θ ⊥ · e ,→⋆ v, there exists a sequence of evaluations via tracked substitutions,
⟨θ 1⊥ ; e 1 ⟩ ,→ . . . ⟨θ i⊥ ; ei ⟩ · · · ,→ ⟨θ n⊥ ; en ⟩
with θ 1⊥ ≡ θ ⊥ , e 1 ≡ e, and en ≡ v. Moreover, each ei+1 is well formed under Γ, thus it has a
translation Γ ⊢ ei+1 ⇝ pi+1 . Thus we can iteratively apply Lemma G.5 n − 1 times and since v is a
value the extra variables in θ n⊥ are irrelevant, thus we get the required σ β |= p = [[v]].
□
For Boolean expressions we specialize the above to
Corollary G.4. If Γ ⊢ e : Bool↓ and Γ ⊢ e ⇝ p, then for every substitution θ ∈ [[Γ]] and every
model σ ∈ [[θ ⊥ ]], θ ⊥ · e ,→⋆ True ⇐⇒ σ β |= p
Proof. We prove the left and right implication separately:
• ⇒ By direct application of Theorem G.3 for v ≡ True.
• ⇐ Since e is terminating, θ ⊥ · e ,→⋆ v. with either v ≡ True or v ≡ False. Assume
v ≡ False, then by Theorem G.3, σ β |= !p, which is a contradiction. Thus, v ≡ True.
□
Lemma G.5 (Eqivalence Preservation). If Γ ⊢ e ⇝ p, then for every substitution θ ∈ [[Γ]]
β
and every model σ ∈ [[θ ⊥ ]], if ⟨θ ⊥ ; e⟩ ,→ ⟨θ 2⊥ ; e 2 ⟩ and for Γ ⊆ Γ2 so that θ 2⊥ ∈ [[Γ2 ]] and σ2 ∈ [[θ 2⊥ ]],
β
Γ2 ⊢ e 2 ⇝ p2 then σ β ∪ (σ2 \ σ β ) |= p = p2 .
Proof. We proceed by case analysis on the derivation ⟨θ ⊥ ; e⟩ ,→ ⟨θ 2⊥ ; e 2 ⟩.
• Assume ⟨θ ⊥ ; e 1 e 2 ⟩ ,→ ⟨θ 2⊥ ; e 1′ e 2 ⟩. By inversion ⟨θ ⊥ ; e 1 ⟩ ,→ ⟨θ 2⊥ ; e 1′ ⟩. Assume Γ ⊢ e 1 ⇝ p1 ,
β
β
Γ ⊢ e 2 ⇝ p2 , Γ2 ⊢ e 1′ ⇝ p1′ . By IH σ β ∪ (σ2 \ σ β ) |= p1 = p1′ , thus σ β ∪ (σ2 \ σ β ) |= app p1 p2 =
app p1′ p2 .
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:49
• Assume ⟨θ ⊥ ; c e⟩ ,→ ⟨θ 2⊥ ; c e ′⟩. By inversion ⟨θ ⊥ ; e⟩ ,→ ⟨θ 2⊥ ; e ′⟩. Assume Γ ⊢ e ⇝ p,
β
β
Γ ⊢ e ′ ⇝ p ′. By IH σ β ∪ (σ2 \ σ β ) |= p = p ′, thus σ β ∪ (σ2 \ σ β ) |= app c p = app c p ′.
• Assume ⟨θ ⊥ ; case x = e of {D i yi → ei }⟩ ,→ ⟨θ 2⊥ ; case x = e ′ of {D i yi → ei }⟩. By
inversion ⟨θ ⊥ ; e⟩ ,→ ⟨θ 2⊥ ; e ′⟩. Assume Γ ⊢ e ⇝ p, Γ ⊢ e ′ ⇝ p ′. Γ ⊢ ei [yi /selD i x][x/e] ⇝ pi .
β
β
By IH σ β ∪(σ2 \σ β ) |= p = p ′, thus σ β ∪(σ2 \σ β ) |= if isD 1 p then p1 else . . . else pn (|τ |)
= if isD 1 p ′ then p1 else . . . else pn (|τ |).
• Assume ⟨θ ⊥ ; D ei e e j ⟩ ,→ ⟨θ 2⊥ ; D ei e ′ e j ⟩. By inversion ⟨θ ⊥ ; e⟩ ,→ ⟨θ 2⊥ ; e ′⟩. Assume Γ ⊢
β
β
e ⇝ p, Γ ⊢ ei ⇝ pi , Γ ⊢ e ′ ⇝ p ′. By IH σ β ∪ (σ2 \ σ β ) |= p = p ′, thus σ β ∪ (σ2 \ σ β ) |=
′
app D pi p p j = app D pi p p j .
• Assume ⟨θ ⊥ ; c w⟩ ,→ ⟨θ ⊥ ; δ (c, w)⟩. By the definition of the syntax, c w is a fully applied
β
logical operator, thus σ β ∪ (σ2 \ σ β ) |= c w = [[δ (c, w)]]
• Assume ⟨θ ⊥ ; (λx .e)e x ⟩ ,→ ⟨θ ⊥ ; e[x/e x ]⟩. Assume Γ ⊢ e ⇝ p, Γ ⊢ e x ⇝ px . Since σ β is defined
β
to satisfy the β-reduction axiom, σ β ∪ (σ2 \ σ β ) |= app (lam x e) px = p[x/px ].
• Assume ⟨θ ⊥ ; case x = D j e of {D i yi → ei }⟩ ,→ ⟨θ ⊥ ; e j [x/D j e][yi /e]⟩. Also, let Γ ⊢ e ⇝ p,
Γ ⊢ ei [x/D j e][yi /ei ] ⇝ pi . By the axiomatic behavior of the measure selector isD j p , we get
σ β |= isD j p . Thus, σ β if isD 1 p then p1 else . . . else pn = p j .
• Assume ⟨(x, e x )θ ⊥ ; x⟩ ,→ ⟨(x, e x′ )θ 2⊥ ; x⟩. By inversion ⟨θ ⊥ ; e x ⟩ ,→ ⟨θ 2⊥ ; e x′ ⟩. By identity of
β
equality, (x, px )σ β ∪ (σ2 \ σ β ) |= x = x.
⊥
• Assume ⟨(y, ey )θ ; x⟩ ,→ ⟨(y, ey )θ 2⊥ ; e x ⟩. By inversion ⟨θ ⊥ ; x⟩ ,→ ⟨θ 2⊥ ; e x ⟩. Assume Γ ⊢ e x ⇝
β
β
px . By IH σ β ∪ (σ2 \ σ β ) |= x = px . Thus (y, py )σ β ∪ (σ2 \ σ β ) |= x = px .
• Assume ⟨(x, w)θ ⊥ ; x⟩ ,→ ⟨(x, w)θ ⊥ ; w⟩. Thus (x, [[w]])σ β |= x = [[w]].
• Assume ⟨(x, D y)θ ⊥ ; x⟩ ,→ ⟨(x, D y)θ ⊥ ; D y⟩. Thus (x, app D y)σ β |= x = app D y.
• Assume ⟨(x, D e)θ ⊥ ; x⟩ ,→ ⟨(x, D y), (yi , ei )θ ⊥ ; D y⟩. Assume Γ ⊢ ei ⇝ pi . Thus (x, app D y), (yi , pi )σ β |=
x = app D y.
□
G.2
Soundness of Approximation
Theorem G.6 (Soundness of Algorithmic). If Γ ⊢S e : τ then Γ ⊢ e : τ .
Proof. To prove soundness it suffices to prove that subtyping is appropriately approximated, as
stated by the following lemma.
□
Lemma G.7. If Γ ⊢S {v : B | e 1 } ⪯ {v : B | e 2 } then Γ ⊢ {v : B | e 1 } ⪯ {v : B | e 2 }.
Proof. By rule ⪯-Base-λS , we need to show that ∀θ ∈ [[Γ]].[[θ · {v : B | e 1 }]] ⊆ [[θ · {v : B | e 2 }]].
We fix a θ ∈ [[Γ]]. and get that forall bindings (x i : {v : B ↓ | ei }) ∈ Γ, θ · ei [v/x i ] ,→⋆ True.
Then need to show that for each e, if e ∈ [[θ · {v : B | e 1 }]], then e ∈ [[θ · {v : B | e 2 }]].
If e diverges then the statement trivially holds. Assume e ,→⋆ w. We need to show that if
θ · e 1 [v/w] ,→⋆ True then θ · e 2 [v/w] ,→⋆ True.
Let θ ⊥ the lifted substitution that satisfies the above. Then by Lemma G.4 for each model
β
σ ∈ [[θ ⊥ ]], σ β |= pi , and σ β |= q 1 for Γ ⊢ ei [v/x i ] ⇝ pi Γ ⊢ ei [v/w] ⇝ qi . Since Γ ⊢S {v : B |
e 1 } ⪯ {v : B | e 2 } we get
Û
pi ⇒ q 1 ⇒ q 2
i
thus
σβ
|= q 2 . By Theorem F.3 we get θ · e 2 [v/w] ,→⋆ True.
□
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:50 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
H
REASONING ABOUT LAMBDAS
Encoding of λ-abstractions and applications via uninterpreted functions, while sound, is imprecise
as it makes it hard to prove theorems that require α- and β-equivalence or extensional equality.
Using the universally quantified α- and β- equivalence axioms would let the type checker accept
more programs, but would render validity, and hence, type checking undecidable. Next, we identify
a middle ground by describing an not provably complete, but sound and decidable approach to
increase the precision of type checking by strengthening the VCs with instances of the α- and
β- equivalence axioms § H.1 and by introducing a combinator for safely asserting extensional
equality § H.2. In the sequel, we omit app when it is clear from the context.
H.1
Equivalence
As soundness relies on satisfiability under a σ β (see Definition F.1), we can safely instantiate the
axioms of α- and β-equivalence on any set of terms of our choosing and still preserve soundness
(Theorem 5.2). That is, instead of checking the validity of a VC p ⇒ q, we check the validity of a
strengthened VC, a ⇒ p ⇒ q, where a is a (finite) conjunction of equivalence instances derived from
p and q, as discussed below.
Representation Invariant The lambda binders, for each SMT sort, are drawn from a pool of
names x i where the index i = 1, 2, . . .. When representing λ terms we enforce a normalization
invariant that for each lambda term lam x i e, the index i is greater than any lambda argument
appearing in e.
α-instances For each syntactic term lam x i e and λ-binder x j such that i < j appearing in the VC,
we generate an α-equivalence instance predicate (or α-instance):
lam x i e = lam x j e[x i /x j ]
The conjunction of α-instances can be more precise than De Bruijn representation, as they let
the SMT solver deduce more equalities via congruence. For example, this VC is needed to prove the
applicative laws for Reader:
d = lam x 1 (x x 1 )
⇒
lam x 2 ((lam x 1 (x x 1 )) x 2 ) = lam x 1 (d x 1 )
The α instance lam x 1 (d x 1 ) = lam x 2 (d x 2 ) derived from the VC’s hypothesis, combined with
congruence immediately yields the VC’s consequence.
β-instances For each syntactic term app (lam x e) e x , with e x not containing any λ-abstractions,
appearing in the VC, we generate a β-equivalence instance predicate (or β-instance):
app (lam x i e) e x = e [e x /x i ] , s.t. e x is λ-free
The λ-free restriction is a simple way to enforce that the reduced term e [e ′/x i ] enjoys the representation invariant. For example, consider the following VC needed to prove that the bind operator
for lists satisfies the monadic associativity law.
(f x ≫= д) = app (lam y (f y ≫= д)) x
The right-hand side of the above VC generates a β-instance that corresponds directly to the equality,
allowing the SMT solver to prove the (strengthened) VC.
Normalization The combination of α- and β-instances is often required to discharge proof
obligations. For example, when proving that the bind operator for the Reader monad is associative,
we need to prove the VC:
lam x 2 (lam x 1 w) = lam x 3 (app (lam x 2 (lam x 1 w)) w)
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:51
The SMT solver proves the VC via the equalities corresponding to an α and then β-instance:
lam x 2 (lam x 1 w) =α lam x 3 (lam x 1 w) =β lam x 3 (app (lam x 2 (lam x 1 w)) w)
H.2
Extensionality
Often, we need to prove that two functions are equal, given the definitions of reflected binders.
Consider
reflect id
id x = x
Liqid Haskell accepts the proof that id x = x for all x:
id_x_eq_x :: x:a → { id x = x}
id_x_eq_x = \x → id x =. x ** QED
as “calling” id unfolds its definition, completing the proof. However, consider this η-expanded
variant of the above proposition:
type Id_eq_id = {(\ x → id x) = (\ y → y)}
Liqid Haskell rejects the proof:
fails :: Id_eq_id
fails = (\ x → id x) =. (\ y → y) ** QED
The invocation of id unfolds the definition, but the resulting equality refinement {id x = x} is
trapped under the λ-abstraction. That is, the equality is absent from the typing environment at
the top level, where the left-hand side term is compared to \y → y. Note that the above equality
requires the definition of id and hence is outside the scope of purely the α- and β-instances.
An Exensionality Operator To allow function equality via extensionality, we provide the user
with a (family of) function comparison operator(s) that transform an explanation p which is a proof
that f x = g x for every argument x, into a proof that f = g.
=* :: f:(a → b) → g:(a → b) → exp:(x:a → {f x = g x}) → {f = g}
Of course, =* cannot be implemented; its type is assumed. We can use =* to prove Id_eq_id by
providing a suitable explanation:
pf_id_id :: Id_eq_id
pf_id_id = (\y → y) =* (\x → id x) ∵ expl ** QED where expl = (\x → id x =. x
** QED)
The explanation is the second argument to ∵ which has the following type that syntactically fires
β-instances:
x : a → {(\ x → id x) x = ((\ x → x) x}
I
IMPLEMENTATION
Refinement reflection and PLE are implemented in Liquid Haskell. The implementation can be
found in the Liquid Haskell GitHub repository, all the benchmarks of § 2 and § 7 are included in
the nople and ple test directories. The benchmarks for deterministic parallelism can be found at
class-laws and detpar-laws.
Next, we describe the file ProofCombinators.hs, the library of proof combinators used by our
benchmarks and discuss known limitations of our implementation.
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:52 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
I.1
ProofCombinators: The Proof Combinators Library
In this section we present ProofCombinators, a Haskell library used to structure proof terms.
ProofCombinators is inspired by Equational Reasoning Data Types in Adga [34], providing operators to construct proofs for equality and linear arithmetic in Haskell. The constructed proofs are
checked by an SMT-solver via Liquid Types.
Proof terms are defined in ProofCombinators as a type alias for unit, a data type that curries no
run-time information
type Proof = ()
Proof types are refined to express theorems about program functions. For example, the following
Proof type expresses that fib 2 == 1
fib2 :: () → {v: Proof | fib 2 == 1}
We simplify the above type by omitting the irrelevant basic type Proof and variable v
fib2 :: () → { fib 2 == 1 }
ProofCombinators provides primitives to construct proof terms by casting expressions to proofs.
To resemble mathematical proofs, we make this casting post-fix. We write p *** QED to cast p to a
proof term, by defining two operators QED and *** as
data QED = QED
(***) :: a → QED → Proof
_ *** _ = ()
Proof construction. To construct proof terms, ProofCombinators provides a proof constructor ⊙.
for logical operators of the theory of linear arithmetic and equality: {=, ,, ≤, <, ≥, >} ∈ ⊙. ⊙. x y
ensures that x ⊙ y holds, and returns x
⊙.:: x : a → y :{ a| x ⊙ y} → {v:a| v == x}
⊙. x _ = x
-- for example
==.:: x : a → y :{ a| x == y} → {v:a| v == x}
For instance, using ==. we construct a proof, in terms of Haskell code, that fib 2 == 1:
fib2 _
=
fib 2
==. fib 1 + fib 0
==. 1
*** QED
Reusing proofs: Proofs as optional arguments. Often, proofs require reusing existing proof
terms. For example, to prove fib 3 == 2 we can reuse the above fib2 proof. We extend the proof
combinators, to receive an optional third argument of Proof type.
⊙.:: x : a → y :a → {x ⊙ y} → {v:a|v == x}
⊙. x _ _ = x
⊙. x y p returns x while the third argument p explicitly proves x ⊙ y.
Optional Arguments. The proof term argument is optional. To implement optional arguments in
Haskell we use the standard technique where for each operator ⊙ we define a type class Opt⊙ that
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:53
takes as input two expressions a and returns a result r, which will be instantiated with either the
result value r:=a or a function form a proof to the result r:=Proof → a.
class Opt⊙ a r where
(⊙.) :: a → a → r
When no explicit proof argument is required, the result type is just an y:a that curries the proof
x ⊙ y
instance Opt⊙ a a where
(⊙.) :: x : a → y :{ a| x ⊙ y} → {v:a | v == x }
(⊙.) x _ = x
Note that Haskell’s type inference [49] requires both type class parameters a and r to be constrainted
at class instance matching time. In most our examples, the result type parameter r is not constrained
at instance matching time, thus due to the Open World Assumption the matching instance could not
be determined. To address the above, we used another common Haskell trick, of generalizing the
instance to type arguments a and b and then constraint a and b to be equal a~b. This generalization
allows the instance to always match and imposed the equality constraint after matching.
instance ( a ~ b ) ⇒ Opt⊙ a b where
(⊙.) :: x : a → y :{ x ⊙ y} → {v:b | v == x }
(⊙.) x _ = x
To explicitly provide a proof argument, the result type r is instantiated to r:= Proof → a. For
the same instance matching restrictions as above, the type is further generalized to return some b
that is constraint to be equal to a.
instance ( a ~ b ) ⇒ Opt⊙ a ( Proof → b) where
(⊙.) :: x : a → y:a → {x ⊙ y} → {v:b | v == x }
(⊙.) x _ _ = x
As a concrete example, we define the equality operator ==. via the type class OptEq as
class OptEq a r where
(==.) :: a → a → r
instance ( a ~ b ) ⇒ OptEq a b where
(==.) :: x : a → y :{ a|x == y} → {v:b|v == x}
(==.) x _ = x
instance ( a ~ b ) ⇒ OptEq a ( Proof → b) where
(==.) :: x : a → y:a → {x == y} → {v:b|v == x}
(==.) x _ _ = x
Explanation Operator. The “explanation operator” (?), or (∵), is used to better structure the
proofs. (?) is an infix operator with same fixity as (⊙.) that allows for the equivalence x ⊙. y ?
p == (⊙.) x y p
(? ) :: ( Proof → a) → Proof → a
f ? y = f y
Putting it all together Using the above operators, we prove that fib 3 == 2, reusing the previous
proof of fib 2 == 1, in a Haskell term that resembles mathematical proofs
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:54 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
fib3 :: () → { fib 3 == 2}
fib3 _
=
fib 3
==. fib 2 + fib 1
==. 2
? fib2 ()
*** QED
Unverified Operators All operators in ProofCombinators, but two are implemented in Haskell
with implementations verified by Liquid Haskell. The ”unsound“ operators are the assume (1). (==?)
that eases proof construction by assuming equalities, to be proven later and (2). (=*) extentional
proof equality.
Assume Operator (==?) eases proof construction by assuming equalities while the proof is in
process. It is not implemented in that its body is undefined. Thus, if we run proof terms including
assume operator, the proof will merely crash (instead of returning ()). Proofs including the assume
operator are not considered complete, as via assume operator any statement can be proven,
Function Extensional Equality Unlike the assume operator that is undefined and included in
unfinished thus unsound proofs, the functions extensionality is included in valid proofs that assume
function extensionality, an axioms that is assumed, as it cannot be proven by our logic.
To allow function equality via extensionality, we provide the user with a function comparison
operator that for each function f and g it transforms a proof that for every argument x, f x = g x
to a proof on function equality f = g.
(=*) :: Arg a ⇒ f :( a → b) → g :( a → b)
→ p :( x : a → {f x = g x })
→ {f = g}
The function (=*) is not implemented in the library: it returns () and its type is assumed. But
soundness of its usage requires the argument type variable a to be constrained by a type class
constraint Arg a, for both operational and type theoretic reasons.
From operational point of view, an implementation of (=*) would require checking equality
of f x = g x forall arguments x of type a. This equality would hold due to the proof argument p.
The only missing point is a way to enumerate all the argument a, but this could be provided by a
method of the type clas Arg a. Yet, we have not implement (=*) because we do not know how to
provide such an implementation that can provably satisfy (=*)’s type.
From type theoretic point of view, the type variable argument a appears only on negative positions.
Liquid type inference is smart enough to infer that since a appears only negative (=*) cannot use
any a and thus will not call any of its argument arguments f, g, nor the p. Thus, at each call site of
(=*) the type variable ‘a‘ is instantiated with the refinement type {v:a | false} indicating deadcode (since as will not be used by the callee.) Refining the argument x:a with false at each call-site
though leads to unsoundness, as each proof argument p is a valid proof under the false assumption.
What Liquid inference cannot predict is our intention to call f, g and p at every possible argument.
This information is capture by the type class constraint Arg a that (as discussed before [55]) states
that methods of the type class Arg a may create values of type a, thus, due to lack of information
on the values that are created by the methods of Arg a, a can only be refined with True.
With extensional equality, we can prove that \x → x is equal to \x → id x, by providing an
explicit explanation that if we call both these functions with the same argument x, they return the
same result, for each x.
safe :: Arg a ⇒ a
→ {(\x → id x) = (\x → x)}
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:55
safe _
= (\x → x)
=*(\x → id x) ∵ (exp ())
exp :: Arg a ⇒ a → x:a
→ {(\x → id x) x = (\x → x) x}
exp _ x = id x
==. x
*** QED
Note that the result of exp is an equality of the redexes (\x → id x) x and ((\x → x) x. Extentional function equality requires as argument an equality on such redexes. Via β equality
instantiations, both such redexes will automatically reduce, requiring exp to prove id x = x, with
is direct.
Admittedly, proving function equality via extensionality is requires a cumbersome indirect proof.
For each function equality in the main proof one needs to define an explanation function that
proves the equality for every argument.
I.2
Engineering Limitations
The theory of refinement reflection is fully implemented in Liquid Haskell. Yet, to make this
extension usable in real world applications there are four known engineering limitations that need
to be addressed. All these limitations seem straightforward to address and we plan to fix them soon.
The language of refinements is typed lambda calculus. That is the types of the lambda arguments
are explicitly specified instead of being inferred. As another minor limitation, the refinement
language parser requires the argument to be enclosed in parenthesis in applications where the
function is not a variable. Thus the Haskell expression (\x → x) e should be written as (\x:a →
x) (e) in the refinement logic,
Class instances methods can not be reflected. Instead, the methods we want to use in the
theorems/propositions should be defined as Haskell functions. This restriction has two major
implications. Firstly, we can not verify correctness of library provided instances but we need to
redifine them ourselves. Secondly, we cannot really verify class instances with class preconditions.
For example, during verification of monoid associativity of the Maybe instance
instance ( Monoid a) ⇒ Monoid ( Maybe a)
there is this Monoid a class constraint assumption we needed to raise to proceed verification.
Only user defined data types can currently used in verification. The reason for this limitation is
that reflection of case expressions requires checker and projector measures for each data type used
in reflected functions. Thus, not only should these data types be defined in the verified module, but
also should be be injected in the logic by providing a refined version of the definition that can (or
may not) be trivially refined.
For example, to reflect a function that uses Peano numbers, the Haskell and the refined Peano
definitions should be provided
data Peano = Z | S Peano
{-@ data Peano [ toInt ]
= Z
| S { prev :: Peano }
@ -}
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:56 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
Note that the termination function toInt that maps Peano numbers to natural numbers is also
crucial for soundness of reflection.
There is no module support. All reflected definitions, including, measures (automatically generated checkers and selector, but also the classic lifted Haskell functions to measures) and the
reflected types of the reflected functions, are not exposed outside of the module they are defined.
Thus all definitions and propositions should exist in the same module.
J
VERIFIED DETERMINISTIC PARALLELISM
Finally, we evaluate our deterministic parallelism prototypes. Aside from the lines of proof code
added, we evaluate the impact on runtime performance. Were we using a proof tool external to
Haskell, this would not be necessary. But our proofs are Haskell programs—they are necessarily
visible to the compiler. In particular, this means a proliferation of unit values and functions returning
unit values. Also, typeclass instances are witnessed at runtime by “dictionary” data structures passed
between functions. Layering proof methods on top of existing classes like Ord could potentially
add indirection or change the code generated, depending on the details of the optimizer. In our
experiments we find little or no effect on runtime performance. Benchmarks were run on a singlesocket Intel® Xeon® CPU E5-2699 v3 with 18 physical cores and 64GiB RAM.
J.1
LVish: Concurrent Sets
First, we use the verifiedInsert operation to observe the runtime slowdown imposed by the extra
proof methods of VerifiedOrd. We benchmark concurrent sets storing 64-bit integers. Figure 18
compares the parallel speedups for a fixed number of parallel insert operations against parallel
verifiedInsert operations, varying the number of concurrent threads. There is a slight observable
difference between the two lines because the extra proof methods do exist at runtime. We repeat
the experiment for two set implementations: a concurrent skiplist (SLSet) and a purely functional
set inside an atomic reference (PureSet) as described in Kuper et al. [28].
J.2 monad-par: n-body simulation
Next, we verify deterministic behavior of an n-body simulation program that leverages monad-par,
a Haskell library which provides deterministic parallelism for pure code [33].
Each simulated particle is represented by a type Body that stores its position, velocity, and mass.
The function accel computes the relative acceleration between two bodies:
accel :: Body → Body → Accel
where Accel represents the three-dimensional acceleration
data Accel = Accel Real Real Real
To compute the total acceleration of a body b we (1) compute the relative acceleration between b
and each body of the system (Vec Body) and (2) we add each acceleration component. For efficiency,
we use a parallel mapReduce for the above computation that first maps each vector body to get the
acceleration relative to b (accel b) and then adds each Accel value by pointwise addition. mapReduce
is only deterministic if the element is a VerifiedMonoid.
mapReduce :: VerifiedMonoid b ⇒ (a → b) → Vec a → b
To enforce the determinism of an n-body simulation, we need to provide a VerifiedMonoid instance
for Accel. We can prove that (Real, +, 0.0) is a monoid. By product proof composition, we get a
verified monoid instance for
type Accel' = (Real, (Real, Real))
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
2
8
7
1.5
6
5
1
4
3
0.5
PureSet
Verified PureSet
SLSet
Verified SLSet
0
2
1
Parallel speedup (relative to SLSet)
Parallel speedup (relative to PureSet)
53:57
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Threads
Fig. 18. Parallel speedup for doing 1 million parallel inserts over 10 iterations, verified and unverified, relative
to the unverified version, for PureSet and SLSet.
which is isomorphic to Accel (i.e. Iso Accel' Accel).
Figure 19 shows the results of running two versions of the n-body simulation with 2,048 bodies
over 5 iterations, with and without verification, using floating point doubles for Real1 . Notably,
the two programs have almost identical runtime performance. This demonstrates that even when
verifying code that is run in a tight loop (like accel), we can expect that our programs will not be
slowed down by an unacceptable amount.
J.3
DPJ: Parallel Reducers
The Deterministic Parallel Java (DPJ) project provides a deterministic-by-default semantics for the
Java programming language [10]. In DPJ, one can declare a method as commutative and thus assert
that racing instances of that method result in a deterministic outcome. For example:
commutative void updateSum ( int n) writes R { sum += n; }
1 Floating
point numbers notoriously violate associativity, but we use this approximation because Haskell does net yet have
an implementation of superaccumulators [15].
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
Parallel speedup (relative to unverified)
53:58 N. Vazou, A. Tondwalkar, V. Choudhury, R. G. Scott, R. R. Newton, P. Wadler, and R. Jhala
7
6
5
4
3
2
1
0
Verified n-body
Unverified n-body
Verified reducer
Unverified reducer
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
Threads
Fig. 19. Parallel speedup for doing a parallel n-body simulation and parallel array reduction. The speedup is
relative to the unverified version of each respective class of program.
But, DPJ provides no means to formally prove commutativity and thus determinism of parallel
reduction. In Liquid Haskell, we specified commutativity as an extra proof method that extends the
VerifiedMonoid class.
class VerifiedMonoid a ⇒ VerifiedCommutativeMonoid a where
commutes :: x:a → y:a → { x <> y = y <> x }
Provably commutative appends can be used to deterministically update a reducer variable, since
the result is the same regardless of the order of appends. We used LVish [28] to encode a reducer
variable with a value a and a region s as RVar s a.
newtype RVar s a
We specify that safe (i.e. deterministic) parallel updates require provably commutative appending.
updateRVar :: VerifiedCommutativeMonoid a ⇒ a → RVar s a → Par s ()
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
53:59
Following the DPJ program, we used updateRVar’s provably deterministic interface to compute, in
parallel, the sum of an array with 3x109 elements by updating a single, global reduction variable
using a varying number of threads. Each thread sums segments of an array, sequentially, and
updates the variable with these partial sums. In Figure 19, we compare the verified and unverified
versions of our implementation to observe no appreciable difference in performance.
Proceedings of the ACM on Programming Languages, Vol. 2, No. POPL, Article 53. Publication date: January 2018.
| 6cs.PL
|
arXiv:1605.06416v3 [math.ST] 12 Feb 2017
Statistical Inference for Cluster Trees
Jisu Kim
Department of Statistics
Carnegie Mellon University
Pittsburgh, USA
[email protected]
Yen-Chi Chen
Department of Statistics
University of Washington
Seattle, USA
[email protected]
Alessandro Rinaldo
Department of Statistics
Carnegie Mellon University
Pittsburgh, USA
[email protected]
Sivaraman Balakrishnan
Department of Statistics
Carnegie Mellon University
Pittsburgh, USA
[email protected]
Larry Wasserman
Department of Statistics
Carnegie Mellon University
Pittsburgh, USA
[email protected]
Abstract
A cluster tree provides a highly-interpretable summary of a density function by
representing the hierarchy of its high-density clusters. It is estimated using the
empirical tree, which is the cluster tree constructed from a density estimator. This
paper addresses the basic question of quantifying our uncertainty by assessing the
statistical significance of topological features of an empirical cluster tree. We first
study a variety of metrics that can be used to compare different trees, analyze their
properties and assess their suitability for inference. We then propose methods to
construct and summarize confidence sets for the unknown true cluster tree. We
introduce a partial ordering on cluster trees which we use to prune some of the
statistically insignificant features of the empirical tree, yielding interpretable and
parsimonious cluster trees. Finally, we illustrate the proposed methods on a variety
of synthetic examples and furthermore demonstrate their utility in the analysis of a
Graft-versus-Host Disease (GvHD) data set.
1
Introduction
Clustering is a central problem in the analysis and exploration of data. It is a broad topic, with several
existing distinct formulations, objectives, and methods. Despite the extensive literature on the topic,
a common aspect of the clustering methodologies that has hindered its widespread scientific adoption
is the dearth of methods for statistical inference in the context of clustering. Methods for inference
broadly allow us to quantify our uncertainty, to discern “true” clusters from finite-sample artifacts, as
well as to rigorously test hypotheses related to the estimated cluster structure.
In this paper, we study statistical inference for the cluster tree of an unknown density. We assume that
we observe an i.i.d. sample {X1 , . . . , Xn } from a distribution P0 with unknown density p0 . Here,
Xi ∈ X ⊂ Rd . The connected components C(λ), of the upper level set {x : p0 (x) ≥ λ}, are called
high-density clusters. The set of high-density clusters forms a nested hierarchy which is referred to
as the cluster tree1 of p0 , which we denote as Tp0 .
Methods for density clustering fall broadly in the space of hierarchical clustering algorithms, and
inherit several of their advantages: they allow for extremely general cluster shapes and sizes, and
in general do not require the pre-specification of the number of clusters. Furthermore, unlike flat
1
It is also referred to as the density tree or the level-set tree.
29th Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain.
clustering methods, hierarchical methods are able to provide a multi-resolution summary of the
underlying density. The cluster tree, irrespective of the dimensionality of the input random variable, is
displayed as a two-dimensional object and this makes it an ideal tool to visualize data. In the context
of statistical inference, density clustering has another important advantage over other clustering
methods: the object of inference, the cluster tree of the unknown density p0 , is clearly specified.
In practice, the cluster tree is estimated from a finite sample, {X1 , . . . , Xn } ∼ p0 . In a scientific
application, we are often most interested in reliably distinguishing topological features genuinely
present in the cluster tree of the unknown p0 , from topological features that arise due to random
fluctuations in the finite sample {X1 , . . . , Xn }. In this paper, we focus our inference on the cluster
tree of the kernel density estimator, Tpbh , where pbh is the kernel density estimator,
n
1 X
kx − Xi k
pbh (x) =
K
,
(1)
nhd i=1
h
where K is a kernel and h is an appropriately chosen bandwidth 2 .
To develop methods for statistical inference on cluster trees, we construct a confidence set for Tp0 ,
i.e. a collection of trees that will include Tp0 with some (pre-specified) probability. A confidence
set can be converted to a hypothesis test, and a confidence set shows both statistical and scientific
significances while a hypothesis test can only show statistical significances [23, p.155].
To construct and understand the confidence set, we need to solve a few technical and conceptual
issues. The first issue is that we need a metric on trees, in order to quantify the collection of trees
that are in some sense “close enough” to Tpbh to be statistically indistinguishable from it. We use the
bootstrap to construct tight data-driven confidence sets. However, only some metrics are sufficiently
“regular” to be amenable to bootstrap inference, which guides our choice of a suitable metric on trees.
On the basis of a finite sample, the true density is indistinguishable from a density with additional
infinitesimal perturbations. This leads to the second technical issue which is that our confidence
set invariably contains infinitely complex trees. Inspired by the idea of one-sided inference [9],
we propose a partial ordering on the set of all density trees to define simple trees. To find simple
representative trees in the confidence set, we prune the empirical cluster tree by removing statistically
insignificant features. These pruned trees are valid with statistical guarantees that are simpler than
the empirical cluster tree in the proposed partial ordering.
Our contributions: We begin by considering a variety of metrics on trees, studying their properties
and discussing their suitability for inference. We then propose a method of constructing confidence
sets and for visualizing trees in this set. This distinguishes aspects of the estimated tree correspond
to real features (those present in the cluster tree Tp0 ) from noise features. Finally, we apply our
methods to several simulations, and a Graft-versus-Host Disease (GvHD) data set to demonstrate the
usefulness of our techniques and the role of statistical inference in clustering problems.
Related work: There is a vast literature on density trees (see for instance the book by Klemelä [16]),
and we focus our review on works most closely aligned with our paper. The formal definition of
the cluster tree, and notions of consistency in estimation of the cluster tree date back to the work of
Hartigan [15]. Hartigan studied the efficacy of single-linkage in estimating the cluster tree and showed
that single-linkage is inconsistent when the input dimension d > 1. Several fixes to single-linkage
have since been proposed (see for instance [21]). The paper of Chaudhuri and Dasgupta [4] provided
the first rigorous minimax analysis of the density clustering and provided a computationally tractable,
consistent estimator of the cluster tree. The papers [1, 5, 12, 17] propose various modifications and
analyses of estimators for the cluster tree. While the question of estimation has been extensively
addressed, to our knowledge our paper is the first concerning inference for the cluster tree.
There is a literature on inference for phylogenetic trees (see the papers [13, 10]), but the object of
inference and the hypothesized generative models are typically quite different. Finally, in our paper,
we also consider various metrics on trees. There are several recent works, in the computational
topology literature, that have considered different metrics on trees. The most relevant to our own
work, are the papers [2, 18] that propose the functional distortion metric and the interleaving distance
on trees. These metrics, however, are NP-hard to compute in general. In Section 3, we consider a
variety of computationally tractable metrics and assess their suitability for inference.
2
We address computing the tree Tpbh , and the choice of bandwidth in more detail in what follows.
2
p(x)
p(x)
x
x
Figure 1: Examples of density trees. Black curves are the original density functions and the red trees
are the associated density trees.
2
Background and Definitions
We work with densities defined on a subset X ⊂ Rd , and denote by k.k the Euclidean norm on X .
Throughout this paper we restrict our attention to cluster tree estimators that are specified in terms of
a function f : X 7→ [0, ∞), i.e. we have the following definition:
Definition 1. For any f : X 7→ [0, ∞) the cluster tree of f is a function Tf : R 7→ 2X , where 2X is
the set of all subsets of X , and Tf (λ) is the set of the connected components of the upper-level
set
S
{x ∈ X : f (x) ≥ λ}. We define the collection of connected components {Tf }, as {Tf } = Tf (λ).
λ
As will be clearer in what follows, working only with cluster trees defined via a function f simplifies
our search for metrics on trees, allowing us to use metrics specified in terms of the function f . With a
slight abuse of notation, we will use Tf to denote also {Tf }, and write C ∈ Tf to signify C ∈ {Tf }.
The cluster tree Tf indeed has a tree structure, since for every pair C1 , C2 ∈ Tf , either C1 ⊂ C2 ,
C2 ⊂ C1 , or C1 ∩ C2 = ∅ holds. See Figure 1 for a graphical illustration of a cluster tree. The formal
definition of the tree requires some topological theory; these details are in Appendix B.
In the context of hierarchical clustering, we are often interested in the “height” at which two points or
two clusters merge in the clustering. We introduce the merge height from [12, Definition 6]:
Definition 2. For any two points x, y ∈ X , any f : X 7→ [0, ∞), and its tree Tf , their merge height
mf (x, y) is defined as the largest λ such that x and y are in the same density cluster at level λ, i.e.
mf (x, y) = sup {λ ∈ R : there exists C ∈ Tf (λ) such that x, y ∈ C} .
We refer to the function mf : X × X 7→ R as the merge height function. For any two clusters
C1 , C2 ∈ {Tf }, their merge height mf (C1 , C2 ) is defined analogously,
mf (C1 , C2 ) = sup {λ ∈ R : there exists C ∈ Tf (λ) such that C1 , C2 ⊂ C} .
One of the contributions of this paper is to construct valid confidence sets for the unknown true
tree and to develop methods for visualizing the trees contained in this confidence set. Formally, we
assume that we have samples {X1 , . . . , Xn } from a distribution P0 with density p0 .
Definition 3. An asymptotic (1 − α) confidence set, Cα , is a collection of trees with the property that
P0 (Tp0 ∈ Cα ) = 1 − α + o(1).
We also provide non-asymptotic upper bounds on the o(1) term in the above definition. Additionally,
we provide methods to summarize the confidence set above. In order to summarize the confidence
set, we define a partial ordering on trees.
Definition 4. For any f, g : X 7→ [0, ∞) and their trees Tf , Tg , we say Tf Tg if there exists a map
Φ : {Tf } → {Tg } such that for any C1 , C2 ∈ Tf , we have C1 ⊂ C2 if and only if Φ(C1 ) ⊂ Φ(C2 ).
With Definition 3 and 4, we describe the confidence set succinctly via some of the simplest trees in
the confidence set in Section 4. Intuitively, these are trees without statistically insignificant splits.
It is easy to check that the partial order in Definition 4 is reflexive (i.e. Tf Tf ) and transitive (i.e.
that Tf1 Tf2 and Tf2 Tf3 implies Tf1 Tf3 ). However, to argue that is a partial order, we
need to show the antisymmetry, i.e. Tf Tg and Tg Tf implies that Tf and Tg are equivalent in
some sense. In Appendices A and B, we show an important result: for an appropriate topology on
trees, Tf Tg and Tg Tf implies that Tf and Tf are topologically equivalent.
3
Tp
Tp
x
Tp
x
(a)
(b)
Tq
x
(c)
Tq
x
Tq
x
(d)
(e)
x
(f)
Figure 2: Three illustrations of the partial order in Definition 4. In each case, in agreement with
our intuitive notion of simplicity, the tree on the top ((a), (b), and (c)) is lower than the corresponding
tree on the bottom((d), (e), and (f)) in the partial order, i.e. for each example Tp Tq .
The partial order in Definition 4 matches intuitive notions of the complexity of the tree for several
reasons (see Figure 2). Firstly, Tf Tg implies (number of edges of Tf ) ≤ (number of edges of Tg )
(compare Figure 2(a) and (d), and see Lemma 6 in Appendix B). Secondly, if Tg is obtained from
Tf by adding edges, then Tf Tg (compare Figure 2(b) and (e), and see Lemma 7 in Appendix B).
Finally, the existence of a topology preserving embedding from {Tf } to {Tg } implies the relationship
Tf Tg (compare Figure 2(c) and (f), and see Lemma 8 in Appendix B).
3
Tree Metrics
In this section, we introduce some natural metrics on cluster trees and study some of their properties
that determine their suitability for statistical inference. We let p, q : X → [0, ∞) be nonnegative
functions and let Tp and Tq be the corresponding trees.
3.1
Metrics
We consider three metrics on cluster trees, the first is the standard `∞ metric, while the second and
third are metrics that appear in the work of Eldridge et al. [12].
`∞ metric: The simplest metric is d∞ (Tp , Tq ) = kp − qk∞ = supx∈X |p(x) − q(x)|. We will show
in what follows that, in the context of statistical inference, this metric has several advantages over
other metrics.
Merge distortion metric: The merge distortion metric intuitively measures the discrepancy in the
merge height functions of two trees in Definition 2. We consider the merge distortion metric [12,
Definition 11] defined by
dM (Tp , Tq ) = sup |mp (x, y) − mq (x, y)|.
x,y∈X
The merge distortion metric we consider is a special case of the metric introduced by Eldridge et al.
[12]3 . The merge distortion metric was introduced by Eldridge et al. [12] to study the convergence of
cluster tree estimators. They establish several interesting properties of the merge distortion metric:
in particular, the metric is stable to perturbations in `∞ , and further, that convergence in the merge
distortion metric strengthens previous notions of convergence of the cluster trees.
Modified merge distortion metric: We also consider the modified merge distortion metric given by
dMM (Tp , Tq ) = sup |dTp (x, y) − dTq (x, y)|,
x,y∈X
where dTp (x, y) = p(x) + p(y) − 2mp (x, y), which corresponds to the (pseudo)-distance between x
and y along the tree. The metric dMM is used in various proofs in the work of Eldridge et al. [12].
3
They further allow flexibility in taking a sup over a subset of X .
4
It is sensitive to both distortions of the merge heights in Definition 2, as well as of the underlying
densities. Since the metric captures the distortion of distances between points along the tree, it is
in some sense most closely aligned with the cluster tree. Finally, it is worth noting that unlike the
interleaving distance and the functional distortion metric [2, 18], the three metrics we consider in this
paper are quite simple to approximate to a high-precision.
3.2
Properties of the Metrics
The following Lemma gives some basic relationships between the three metrics d∞ , dM and dMM . We
define pinf = inf x∈X p(x), and qinf analogously, and a = inf x∈X {p(x) + q(x)} − 2 min{pinf , qinf }.
Note that when the Lebesgue measure µ(X ) is infinite, then pinf = qinf = a = 0.
Lemma 1. For any densities p and q, the following relationships hold: (i) When p and q are
continuous, then d∞ (Tp , Tq ) = dM (Tp , Tq ). (ii) dMM (Tp , Tq ) ≤ 4d∞ (Tp , Tq ). (iii) dMM (Tp , Tq ) ≥
d∞ (Tp , Tq ) − a, where a is defined as above. Additionally when µ(X ) = ∞, then dMM (Tp , Tq ) ≥
d∞ (Tp , Tq ).
The proof is in Appendix F. From Lemma 1, we can see that under a mild assumption (continuity of
the densities), d∞ and dM are equivalent. We note again that the work of Eldridge et al. [12] actually
defines a family of merge distortion metrics, while we restrict our attention to a canonical one. We
can also see from Lemma 1 that while the modified merge metric is not equivalent to d∞ , it is usually
multiplicatively sandwiched by d∞ .
Our next line of investigation is aimed at assessing the suitability of the three metrics for the task
of statistical inference. Given the strong equivalence of d∞ and dM we focus our attention on d∞
and dMM . Based on prior work (see [7, 8]), the large sample behavior of d∞ is well understood. In
particular, d∞ (Tpbh , Tp0 ) converges to the supremum of an appropriate Gaussian process, on the basis
of which we can construct confidence intervals for the d∞ metric.
The situation for the metric dMM is substantially more subtle. One of our eventual goals is to use
the non-parametric bootstrap to construct valid estimates of the confidence set. In general, a way to
assess the amenability of a functional to the bootstrap is via Hadamard differentiability [24]. Roughly
speaking, Hadamard-differentiability is a type of statistical stability, that ensures that the functional
under consideration is stable to perturbations in the input distribution. In Appendix C, we formally
define Hadamard differentiability and prove that dMM is not point-wise Hadamard differentiable.
This does not completely rule out the possibility of finding a way to construct confidence sets based
on dMM , but doing so would be difficult and so far we know of no way to do it.
In summary, based on computational considerations we eliminate the interleaving distance and
the functional distortion metric [2, 18], we eliminate the dMM metric based on its unsuitability for
statistical inference and focus the rest of our paper on the d∞ (or equivalently dM ) metric which is
both computationally tractable and has well understood statistical behavior.
4
Confidence Sets
In this section, we consider the construction of valid confidence intervals centered around the kernel
density estimator, defined in Equation (1). We first observe that a fixed bandwidth for the KDE
gives a dimension-free rate of convergence for estimating a cluster tree. For estimating a density
in high dimensions, the KDE has a poor rate of convergence, due to a decreasing bandwidth for
simultaneously optimizing the bias and the variance of the KDE.
When estimating a cluster tree, the bias of the KDE does not affect its cluster tree. Intuitively, the
cluster tree is a shape characteristic of a function, which is not affected by the bias. Defining the
biased density, ph (x) = E[b
ph (x)], two cluster trees from ph and the true density p0 are equivalent
with respect to the topology in Appendix A, if h is small enough and p0 is regular enough:
Lemma 2. Suppose that the true unknown density p0 , has no non-degenerate critical points 4 , then
there exists a constant h0 > 0 such that for all 0 < h ≤ h0 , the two cluster trees, Tp0 and Tph have
the same topology in Appendix A.
4
The Hessian of p0 at every critical point is non-degenerate. Such functions are known as Morse functions.
5
From Lemma 2, proved in Appendix G, a fixed bandwidth for the KDE can be applied to give a
dimension-free rate of convergence for estimating the cluster tree. Instead of decreasing bandwidth h
and inferring the cluster tree of the true density Tp0 at rate OP (n−2/(4+d) ), Lemma 2 implies that we
can fix h > 0 and infer the cluster tree of the biased density Tph at rate OP (n−1/2 ) independently of
the dimension. Hence a fixed bandwidth crucially enhances the convergence rate of the proposed
methods in high-dimensional settings.
4.1
A data-driven confidence set
We recall that we base our inference on the d∞ metric, and we recall the definition of a valid
confidence set (see Definition 3). As a conceptual first step, suppose that for a specified value α we
could compute the 1 − α quantile of the distribution of d∞ (Tpbh , Tph ), and denote this value tα . Then
a valid confidence set for the unknown Tph is Cα = {T : d∞ (T, Tpbh ) ≤ tα }. To estimate tα , we use
e 1, · · · , X
en1 }, . . . , {X
eB, · · · , X
enB },
the bootstrap. Specifically, we generate B bootstrap samples, {X
1
1
by sampling with replacement from the original sample. On each bootstrap sample, we compute
the KDE, and the associated cluster tree. We denote the cluster trees {Tep1h , . . . , TepBh }. Finally, we
estimate tα by
n
1 X
−1
b
b
b
tα = F (1 − α), where F (s) =
I(d∞ (Tepih , Tpbh ) < s).
B i=1
bα = {T : d∞ (T, Tbh ) ≤ b
tα }. Using techniques from [8, 7],
Then the data-driven confidence set is C
the following can be shown (proof omitted):
Theorem 3. Under mild regularity conditions on the kernel5 , we have that the constructed confidence
set is asymptotically valid and satisfies,
7 1/6
bα = 1 − α + O log n
P Th ∈ C
.
nhd
Hence our data-driven confidence set is consistent at dimension independent rate. When h is a fixed
small constant, Lemma 2 implies that Tp0 and Tph have the same topology, and Theorem 3 guarantees
that the non-parametric bootstrap is consistent at a dimension independent O(((log n)7 /n)1/6 ) rate.
For reasons explained in [8], this rate is believed to be optimal.
4.2
Probing the Confidence Set
bα is an infinite set with a complex structure. Infinitesimal perturbations of the
The confidence set C
density estimate are in our confidence set and so this set contains very complex trees. One way to
understand the structure of the confidence set is to focus attention on simple trees in the confidence
set. Intuitively, these trees only contain topological features (splits and branches) that are sufficiently
strongly supported by the data.
We propose two pruning schemes to find trees, that are simpler than the empirical tree Tpbh that are in
the confidence set. Pruning the empirical tree aids visualization as well as de-noises the empirical
tree by eliminating some features that arise solely due to the stochastic variability of the finite-sample.
The algorithms are (see Figure 3):
1. Pruning only leaves: Remove all leaves of length less than 2b
tα (Figure 3(b)).
2. Pruning leaves and internal branches: In this case, we first prune the leaves as above. This
yields a new tree. Now we again prune (using cumulative length) any leaf of length less than 2b
tα . We
continue iteratively until all remaining leaves are of cumulative length larger than 2b
tα (Figure 3(c)).
In Appendix D.2 we formally define the pruning operation and show the following. The remaining
tree Te after either of the above pruning operations satisfies: (i) Te Tpbh , (ii) there exists a function f
bα (see Lemma 10 in Appendix D.2). In other words, we identified a
whose tree is Te, and (iii) Te ∈ C
valid tree with a statistical guarantee that is simpler than the original estimate Tpbh . Intuitively, some
of the statistically insignificant features have been removed from Tpbh . We should point out, however,
5
See Appendix D.1 for details.
6
L1
L2
L6
L5
L3 L4
E5
E4
E1
E3
E2
(a) The empirical tree.
(b) Pruning only leaves.
(c) Pruning leaves and branches.
Figure 3: Illustrations of our two pruning strategies. (a) shows the empirical tree. In (b), leaves that
are insignificant are pruned, while in (c), insignificant internal branches are further pruned top-down.
(a)
(b)
(c)
−
−
0.2
0.4
0.6
(d)
0.8
1.0
lambda
0
lambda
0
0.208 0.272
0
lambda
−
0.0
−
0.07
0.291
−
Yingyang data, alpha = 0.05
0.035 0.044 0.052
−
0.255
Mickey mouse data, alpha = 0.05
0.529
Ring data, alpha = 0.05
0.0
0.2
0.4
0.6
(e)
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
(f)
Figure 4: Simulation examples. (a) and (d) are the ring data; (b) and (e) are the mickey mouse data;
(c) and (f) are the yingyang data. The solid lines are the pruned trees; the dashed lines are leaves (and
edges) removed by the pruning procedure. A bar of length 2b
tα is at the top right corner. The pruned
trees recover the actual structure of connected components.
bα . Ideally, we would like to
that there may exist other trees that are simpler than Tpbh that are in C
have an algorithm that identifies all trees in the confidence set that are minimal with respect to the
partial order in Definition 4. This is an open question that we will address in future work.
5
Experiments
In this section, we demonstrate the techniques we have developed for inference on synthetic data, as
well as on a real dataset.
5.1
Simulated data
We consider three simulations: the ring data (Figure 4(a) and (d)), the Mickey Mouse data (Figure 4(b)
and (e)), and the yingyang data (Figure 4(c) and (f)). The smoothing bandwidth is chosen by the
Silverman reference rule [20] and we pick the significance level α = 0.05.
7
4e−10
−
−
0e+00
0e+00
1e−10
2e−10
2e−10
4e−10
3e−10
6e−10
8e−10
−
−
0.0
0.2
0.4
0.6
0.8
1.0
0.0
(a) The positive treatment data.
0.2
0.4
0.6
0.8
1.0
(b) The control data.
Figure 5: The GvHD data. The solid brown lines are the remaining branches after pruning; the blue
dashed lines are the pruned leaves (or edges). A bar of length 2b
tα is at the top right corner.
Example 1: The ring data. (Figure 4(a) and (d)) The ring data consists of two structures: an outer
ring and a center node. The outer circle consists of 1000 points and the central node contains 200
points. To construct the tree, we used h = 0.202.
Example 2: The Mickey Mouse data. (Figure 4(b) and (e)) The Mickey Mouse data has three
components: the top left and right uniform circle (400 points each) and the center circle (1200 points).
In this case, we select h = 0.200.
Example 3: The yingyang data. (Figure 4(c) and (f)) This data has 5 connected components: outer
ring (2000 points), the two moon-shape regions (400 points each), and the two nodes (200 points
each). We choose h = 0.385.
Figure 4 shows those data ((a), (b), and (c)) along with the pruned density trees (solid parts in (d), (e),
and (f)). Before pruning the tree (both solid and dashed parts), there are more leaves than the actual
number of connected components. But after pruning (only the solid parts), every leaf corresponds to
an actual connected component. This demonstrates the power of a good pruning procedure.
5.2
GvHD dataset
Now we apply our method to the GvHD (Graft-versus-Host Disease) dataset [3]. GvHD is a
complication that may occur when transplanting bone marrow or stem cells from one subject to
another [3]. We obtained the GvHD dataset from R package ‘mclust’. There are two subsamples: the
control sample and the positive (treatment) sample. The control sample consists of 9083 observations
and the positive sample contains 6809 observations on 4 biomarker measurements (d = 4). By the
normal reference rule [20], we pick h = 39.1 for the positive sample and h = 42.2 for the control
sample. We set the significance level α = 0.05.
Figure 5 shows the density trees in both samples. The solid brown parts are the remaining components
of density trees after pruning and the dashed blue parts are the branches removed by pruning. As can
be seen, the pruned density tree of the positive sample (Figure 5(a)) is quite different from the pruned
tree of the control sample (Figure 5(b)). The density function of the positive sample has fewer bumps
(2 significant leaves) than the control sample (3 significant leaves). By comparing the pruned trees,
we can see how the two distributions differ from each other.
6
Discussion
There are several open questions that we will address in future work. First, it would be useful to have
an algorithm that can find all trees in the confidence set that are minimal with respect to the partial
order . These are the simplest trees consistent with the data. Second, we would like to find a way
to derive valid confidence sets using the metric dMM which we view as an appealing metric for tree
inference. Finally, we have used the Silverman reference rule [20] for choosing the bandwidth but we
would like to find a bandwidth selection method that is more targeted to tree inference.
8
References
[1] S. Balakrishnan, S. Narayanan, A. Rinaldo, A. Singh, and L. Wasserman. Cluster trees on manifolds. In
Advances in Neural Information Processing Systems, 2012.
[2] U. Bauer, E. Munch, and Y. Wang. Strong equivalence of the interleaving and functional distortion metrics
for reeb graphs. In 31st International Symposium on Computational Geometry (SoCG 2015), volume 34,
pages 461–475. Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik, 2015.
[3] R. R. Brinkman, M. Gasparetto, S.-J. J. Lee, A. J. Ribickas, J. Perkins, W. Janssen, R. Smiley, and
C. Smith. High-content flow cytometry and temporal data analysis for defining a cellular signature of
graft-versus-host disease. Biology of Blood and Marrow Transplantation, 13(6):691–700, 2007.
[4] K. Chaudhuri and S. Dasgupta. Rates of convergence for the cluster tree. In Advances in Neural Information
Processing Systems, pages 343–351, 2010.
[5] K. Chaudhuri, S. Dasgupta, S. Kpotufe, and U. von Luxburg. Consistent procedures for cluster tree
estimation and pruning. IEEE Transactions on Information Theory, 2014.
[6] F. Chazal, B. T. Fasy, F. Lecci, B. Michel, A. Rinaldo, and L. Wasserman. Robust topological inference:
Distance to a measure and kernel distance. arXiv preprint arXiv:1412.7197, 2014.
[7] Y.-C. Chen, C. R. Genovese, and L. Wasserman. Density level sets: Asymptotics, inference, and visualization. arXiv:1504.05438, 2015.
[8] V. Chernozhukov, D. Chetverikov, and K. Kato. Central limit theorems and bootstrap in high dimensions.
Annals of Probability, 2016.
[9] D. Donoho. One-sided inference about functionals of a density. The Annals of Statistics, 16(4):1390–1420,
1988.
[10] B. Efron, E. Halloran, and S. Holmes. Bootstrap confidence levels for phylogenetic trees. Proceedings of
the National Academy of Sciences, 93(23), 1996.
[11] U. Einmahl and D. M. Mason. Uniform in bandwidth consistency of kernel-type function estimators. The
Annals of Statistics, 33(3):1380–1403, 2005.
[12] J. Eldridge, M. Belkin, and Y. Wang. Beyond hartigan consistency: Merge distortion metric for hierarchical
clustering. In Proceedings of The 28th Conference on Learning Theory, pages 588–606, 2015.
[13] J. Felsenstein. Confidence limits on phylogenies, a justification. Evolution, 39, 1985.
[14] C. R. Genovese, M. Perone-Pacifico, I. Verdinelli, and L. Wasserman. Nonparametric ridge estimation.
The Annals of Statistics, 42(4):1511–1545, 2014.
[15] J. A. Hartigan. Consistency of single linkage for high-density clusters. Journal of the American Statistical
Association, 1981.
[16] J. Klemelä. Smoothing of multivariate data: density estimation and visualization, volume 737. John Wiley
& Sons, 2009.
[17] S. Kpotufe and U. V. Luxburg. Pruning nearest neighbor cluster trees. In Proceedings of the 28th
International Conference on Machine Learning (ICML-11), pages 225–232, 2011.
[18] D. Morozov, K. Beketayev, and G. Weber. Interleaving distance between merge trees. Discrete and
Computational Geometry, 49:22–45, 2013.
[19] D. W. Scott. Multivariate density estimation: theory, practice, and visualization. John Wiley & Sons,
2015.
[20] B. W. Silverman. Density estimation for statistics and data analysis, volume 26. CRC press, 1986.
[21] W. Stuetzle and R. Nugent. A generalized single linkage method for estimating the cluster tree of a density.
Journal of Computational and Graphical Statistics, 19(2), 2010.
[22] L. Wasserman. All of nonparametric statistics. Springer Science & Business Media, 2006.
[23] L. Wasserman. All of Statistics: A Concise Course in Statistical Inference. Springer Science & Business
Media, 2010. ISBN 1441923225, 9781441923226.
[24] J. Wellner. Weak Convergence and Empirical Processes: With Applications to Statistics. Springer Science
& Business Media, 2013.
9
A
Topological Preliminaries
The goal of this section is to define an appropriate topology on the cluster tree Tf in Definition
1. Defining an appropriate topology for the cluster tree Tf is important in this paper for several
reasons: (1) the topology gives geometric insight for the cluster tree, (2) homeomorphism (topological
equivalence) is connected to equivalence in the partial order in Definition 4, and (3) the topology
bα as in Lemma 2
gives a justification for using a fixed bandwidth h for constructing confidence set C
to obtain faster rates of convergence.
We construct the topology of the cluster tree Tf by imposing a topology on the corresponding
collection of connected components {Tf } in Definition 1. For defining a topology on {Tf }, we define
the tree distance function dTf in Definition 5, and impose the metric topology induced from the tree
distance function. Using a distance function for topology not only eases formulating topology but
also enables us to inherit all the good properties of the metric topology.
The desired tree distance function dTf : {Tf }×{Tf } → [0, ∞) is based on the merge height function
mf in Definition 2. For later use in the proof, we define the tree distance function dTf on both X and
{Tf } as follows:
Definition 5. Let f : X → [0, ∞) be a function, and Tf be its cluster tree in Definition 1. For any
two points x, y ∈ X , the tree distance function dTf : X × X → [0, ∞) of Tf on X is defined as
dTf (x, y) = f (x) + f (y) − 2mf (x, y).
Similarly, for any two clusters C1 , C2 ∈ {Tf }, we first define λ1 = sup{λ : C1 ∈ Tf (λ)}, and λ2
analogously. We then define the tree distance function dTf : {Tf } × {Tf } → [0, ∞) of Tf on X as:
dTf (C1 , C2 ) = λ1 + λ2 − 2mf (C1 , C2 ).
The tree distance function dTf in Definition 2 is a pseudometric on X and is a metric on {Tf } as
desired, proven in Lemma 4. The proof is given later in Appendix E.
Lemma 4. Let f : X → [0, ∞) be a function, Tf be its cluster tree in Definition 1, and dTf be its
tree distance function in Definition 5. Then dTf on X is a pseudometric and dTf on {Tf } is a metric.
From the metric dTf on {Tf } in Definition 5, we impose the induced metric topology on {Tf }.
We say Tf is homeomorphic to Tg , or Tf ∼
= Tg , when their corresponding collection of connected
components are homeomorphic, i.e. {Tf } ∼
= {Tg }. (Two spaces are homeomorphic if there exists a
bijective continuous function between them, with a continuous inverse.)
To get some geometric understanding of the cluster tree in Definition 1, we identify edges that
constitute the cluster tree. Intuitively, edges correspond to either leaves or internal branches. An edge
is roughly defined as a set of clusters whose inclusion relationship with respect to clusters outside
an edge are equivalent, so that when the collection of connected components is divided into edges,
we observe the same inclusion relationship between representative clusters whenever any cluster is
selected as a representative for each edge.
For formally defining edges, we define an interval in the cluster tree and the equivalence relation in
the cluster tree. For any two clusters A, B ∈ {Tf }, the interval [A, B] ⊂ {Tf } is defined as a set
clusters that contain A and are contained in B, i.e.
[A, B] := {C ∈ {Tf } : A ⊂ C ⊂ B} ,
The equivalence relation ∼ is defined as A ∼ B if and only if their inclusion relationship with respect
to clusters outside [A, B] and [B, A], i.e.
A ∼ B if and only if
for all C ∈ {Tf } such that C ∈
/ [A, B] ∪ [B, A], C ⊂ A iff C ⊂ B and A ⊂ C iff B ⊂ C.
Then it is easy to see that the relation ∼ is reflexive(A ∼ A), symmetric(A ∼ B implies B ∼ A),
and transitive (A ∼ B and B ∼ C implies A ∼ C). Hence the relation ∼ is indeed an equivalence
relation, and we can consider the set of equivalence classes {Tf }/∼ . We define the edge set E(Tf )
as E(Tf ) := {Tf }/∼ .
For later use, we define the partial order on the edge set E(Tf )) as follows: [C1 ] ≤ [C2 ] if and only if
for all A ∈ [C1 ] and B ∈ [C2 ], A ⊂ B. We say that a tree Tf is finite if its edge E(Tf ) is a finite set.
10
B
The Partial Order
As discussed in Section 2, to see that the partial order in Definition 4 is indeed a partial order, we
need to check the reflexivity, the transitivity, and the antisymmetry. The reflexivity and the transitivity
are easier to check, but to show antisymmetric, we need to show that if two trees Tf and Tg satisfies
Tf Tg and Tg Tf , then Tf and Tg are equivalent in some sense. And we give the equivalence
relation as the topology on the cluster tree defined in Appendix A. The argument is formally stated in
Lemma 5. The proof is done later in Appendix E.
Lemma 5. Let f, g : X → [0, ∞) be functions, and Tf , Tg be their cluster trees in Definition 1.
Then if f, g are continuous and Tf , Tg are finite, Tf Tg and Tg Tf implies that there exists a
homeomorphism Φ : {Tf } → {Tg } that preserves the root, i.e. Φ(X ) = X . Conversely, if there
exists a homeomorphism Φ : {Tf } → {Tg } that preserves the root, Tf Tg and Tg Tf hold.
The partial order in Definition 4 gives a formal definition of simplicity of trees, and it is used
to justify pruning schemes in Section 4.2. Hence it is important to match the partial order with
the intuitive notions of the complexity of the tree. We provided three arguments in Section 2: (1)
if Tf Tg holds then it must be the case that (number of edges of Tf ) ≤ (number of edges of Tg ),
(2) if Tg can be obtained from Tf by adding edges, then Tf Tg holds, and (3) the existence of a
topology preserving embedding from {Tf } to {Tg } implies the relationship Tf Tg . We formally
state each item in Lemma 6, 7, and 8. Proofs of these lemmas are done later in Appendix E.
Lemma 6. Let f, g : X → [0, ∞) be functions, and Tf , Tg be their cluster trees in Definition
1. Suppose Tf Tg via Φ : {Tf } → {Tg }. Define Φ̄ : E(Tf ) → E(Tg ) by for [C] ∈ E(Tf )
choosing any C ∈ [C] and defining as Φ̄([C]) = [Φ(C)]. Then Φ̄ is injective, and as a consequence,
|E(Tf )| ≤ |E(Tg )|.
Lemma 7. Let f, g : X → [0, ∞) be functions, and Tf , Tg be their cluster trees in Definition 1. If
Tg can be obtained from Tf by adding edges, then Tf Tg holds.
Lemma 8. Let f, g : X → [0, ∞) be functions, and Tf , Tg be their cluster trees in Definition 1.
If there exists a one-to-one map Φ : {Tf } → {Tg } that is a homeomorphism between {Tf } and
Φ({Tf }) and preserves the root, i.e. Φ(X ) = X , then Tf Tg holds.
C
Hadamard Differentiability
Definition 6 (see page 281 of [24]). Let D and E be normed spaces and let φ : Dφ → E be a map
defined on a subset Dφ ⊂ D. Then φ is Hadamard differentiable at θ if there exists a continuous,
linear map φ0θ : D → E such that
φ(θ + tqt ) − φ(θ)
− φ0θ (h)
t
→0
E
as t → 0, for every qt → q.
Hadamard differentiability is a key property for bootstrap inference since it is a sufficient condition
for the delta method; for more details, see section 3.1 of [24]. Recall that dMM is based on the
function dTp (x, y) = p(x) + p(y) − 2mp (x, y). The following theorem shows that the function dTp
is not Hadamard differentiable for some pairs (x, y). In our case D is the set of continuous functions
on the sample space, E is the real line, θ = p, φ(p) is dTp (x, y) and the norm on E is the usual
Euclidean norm.
Theorem 9. Let B(x) be the smallest set B ∈ Tp such that x ∈ B. dTp (x, y) is not Hadamard
differentiable for x 6= y when one of the following two scenarios occurs:
(i) min{p(x), p(y)} = p(c) for some critical point c.
(ii) B(x) = B(y) and p(x) = p(y).
The merge distortion metric dM is also not Hadamard differentiable.
11
D
Confidence Sets Constructions
D.1
Regularity conditions on the kernel
To apply the results in [8] which imply that the bootstrap confidence set is consistent, we consider the
following two assumptions.
(K1) The kernel function K has the bounded second derivative and is symmetric, non-negative, and
Z
Z
x2 K(x)dx < ∞,
K(x)2 dx < ∞.
(K2) The kernel function K satisfies
x−y
: x ∈ Rd , h > 0 .
K = y 7→ K
h
(2)
We require that K satisfies
supN K, L2 (P ), kF kL2 (P )
P
v
A
≤
(3)
for some positive numbers A and v, where N (T, d, ) denotes the -covering number of
the metric space (T, d), F is the envelope function of K, and the supremum is taken over
the whole RdR. The A and v are usually called the VC characteristics of K. The norm
kF k2L2 (P ) = |F (x)|2 dP (x).
Assumption (K1) is to ensure that the variance of the KDE is bounded and ph has the bounded second
derivative. This assumption is very common in statistical literature, see e.g. [22, 19]. Assumption
(K2) is to regularize the complexity of the kernel function so that the supremum norm for kernel
functions and their derivatives can be bounded in probability. A similar assumption appears in [11]
and [14]. The Gaussian kernel and most compactly supported kernels satisfy both assumptions.
D.2
Pruning
The goal of this section is to formally define the pruning scheme in Section 4.2. Note that when
pruning leaves and internal branches, when the cumulative length is computed for each leaf and
internal branch, then the pruning process can be done at once. We provide two pruning schemes in
Section 4.2 in a unifying framework by defining an appropriate notion of lifetime for each edge, and
deleting all insignificant edges with small lifetimes. To follow the pruning schemes in Section 4.2,
we require that the lifetime of a child edge is shorter than the lifetime of a parent edge, so that we
can delete edges from the top. We evaluate the lifetime of each edge by an appropriate nonnegative
(possibly infinite) function life. We formally define the pruned tree P runedlife,btα (Tbh ) as follows:
Definition 7. Suppose the function life : E(Tbh ) → [0, +∞] satisfies that [C1 ] ≤ [C2 ] =⇒
life([C1 ]) ⊂ life([C2 ]). We define the pruned tree P runedlife,btα (Tbh ) : R → 2X as
n
o
P runedlife,btα (Tbh )(λ) = C ∈ Tbh (λ − b
tα ) : life([C]) > b
tα .
We suggest two life functions corresponding to two pruning schemes in Section 4.2. We first need
several definitions. For any [C] ∈ E(Tbh ), define its level as
n
o
level([C]) := λ : there exists A ∈ [C] ∩ Tbh (λ) ,
and define its cumulative level as
n
o
cumlevel([C]) := λ : there exists A ∈ Tbh (λ), B ∈ [C] such that A ⊂ B .
Then lifeleaf corresponds to first pruning scheme in Section 4.2, which is to prune out only insignificant leaves.
sup{level([C])} − inf{level([C])} if inf{level([C])} =
6 inf {cumlevel([C])}
lifeleaf ([C]) =
.
+∞
otherwise.
12
And lifetop corresponds to second pruning scheme in Section 4.2, which is to prune out insignificant
edges from the top.
lifetop ([C]) = sup{cumlevel([C])} − inf {cumlevel([C])} .
Note that lifeleaf is lower bounded by lifetop . In fact, for any life function that is lower bounded
bα that is simpler than
by lifetop , the pruned tree P runedlife,btα is a valid tree in the confidence set C
the original estimate Tbh , so that the pruned tree is the desired tree as discussed in Section 4.2. We
formally state as follows. The proof is given in Appendix G
Lemma 10. Suppose that the life function satisfies: for all [C] ∈ E(Tbh ), lifetop ([C]) ≤ life([C]).
Then
(i) P runedlife,btα (Tbh ) Tpbh .
(ii) there exists a function pe such that Tpe = P runedlife,btα (Tbh ).
bα .
(iii) pe in (ii) satisfies pe ∈ C
Remark: It can be shown that complete pruning — simultaneously removing all leaves and branches
with length less than 2b
tα — can in general yield a tree that is outside the confidence set. For example,
see Figure 3. If we do complete pruning to this tree, we will get the trivial tree.
E
E.1
Proofs for Appendix A and B
Proof of Lemma 4
Lemma 4. Let f : X → [0, ∞) be a function, Tf be its cluster tree in Definition 1, and dTf be its
tree distance function in Definition 5. Then dTf on X is a pseudometric and dTf on {Tf } is a metric.
Proof.
First, we show that dTf on X is a pseudometric. To do this, we need to show nonnegativity(dTf (x, y) ≥ 0), x = y implying dTf (x, y) = 0, symmetry(dTf (x, y) = dTf (y, x)),
and subadditivity(dTf (x, y) + dTf (y, z) ≤ dTf (x, z)).
For non-negativity, note that for all x, y ∈ X , mf (x, y) ≤ min {f (x), f (y)}, so
dTf (x, y) = f (x) + f (y) − 2mf (x, y) ≥ 0.
(4)
For x = y implying dTf (x, y) = 0, x = y implies mf (x, y) = f (x) = f (y), so
x = y =⇒ dTf (x, y) = 0.
(5)
For symmetry, since mf (x, y) = mf (y, x),
dTf (x, y) = dTf (y, x).
(6)
For subadditivity, note first that mf (x, y) ≤ f (y) and mf (y, z) ≤ f (y) holds, so
max {mf (x, y), mf (y, z)} ≤ f (y).
(7)
And also note that there exists Cxy , Cyz ∈ Tf (min {mf (x, y), mf (y, z)}) that satisfies x, y ∈ Cxy
and y, z ⊂ Cyz . Then y ∈ Cxy ∩ Cyz 6= ∅, so x, z ∈ Cxy = Cyz . Then from definition of mf (x, z),
this implies that
min {mf (x, y), mf (y, z)} ≤ mf (x, z).
(8)
And by applying (7) and (8), dTf (x, y) + dTf (y, z) is upper bounded by dTf (x, z) as
dTf (x, y) + dTf (y, z)
= f (x) + f (y) − 2mf (x, y) + f (y) + f (z) − 2mf (y, z)
= f (x) + f (z) − 2 (min {mf (x, y), mf (y, z)} + max {mf (x, y), mf (y, z)} − f (y))
≥ f (x) + f (z) − 2mf (x, z)
= dTf (x, z).
13
(9)
Hence (4), (5), (6), and (9) implies that dTf on X is a pseudometric.
Second, we show that dTf on {Tf } is a metric. To do this, we need to show nonnegativity(dTf (x, y) ≥ 0), identity of indiscernibles(x = y
⇐⇒
dTf (x, y) = 0),
symmetry(dTf (x, y) = dTf (y, x)), and subadditivity(dTf (x, y) + dTf (y, z) ≤ dTf (x, z)).
For nonnegativity, note that if C1 ∈ Tf (λ1 ) and C2 ∈ Tf (λ2 ), then mf (C1 , C2 ) ≤ min{λ1 , λ2 }, so
dTf (C1 , C2 ) = λ1 + λ2 − 2mf (C1 , C2 ) ≥ 0.
(10)
For identity of indiscernibles, C1 = C2 implies mf (C1 , C2 ) = λ1 = λ2 , so
C1 = C2 =⇒ dTf (C1 , C2 ) = 0.
(11)
And conversely, dTf (C1 , C2 ) = 0 implies λ1 = λ2 = mf (C1 , C2 ), so there exists C ∈ Tf (λ1 ) such
that C1 ⊂ C and C2 ⊂ C. Then since C1 , C2 , C ∈ Tf (λ1 ), so C1 ∩ C 6= ∅ implies C1 = C and
similarly C2 = C, so
dTf (C1 , C2 ) = 0 =⇒ C1 = C2 .
(12)
Hence (11) and (12) implies identity of indiscernibles as
C1 = C2 ⇐⇒ dTf (C1 , C2 ) = 0.
(13)
For symmetry, since mf (C1 , C2 ) = mf (C2 , C1 ),
dTf (C1 , C2 ) = dTf (C2 , C1 ).
(14)
For subadditivity, note that mf (C1 , C2 ) ≤ λ2 and mf (C2 , C3 ) ≤ λ2 holds, so
max {mf (C1 , C2 ), mf (C2 , C3 )} ≤ λ2 .
(15)
And also note that there exists C12 , C23 ∈ Tf (min {mf (C1 , C2 ), mf (C2 , C3 )}) that satisfies
C1 , C2 ⊂ C12 and C2 , C3 ⊂ C23 . Then C2 ⊂ C12 ∩ C23 6= ∅, so C1 , C3 ∈ C12 = C23 . Then from
definition of mf (C1 , C3 ), this implies that
min {mf (C1 , C2 ), mf (C2 , C3 )} ≤ mf (C1 , C3 ).
(16)
And by applying (15) and (16), dTf (C1 , C2 ) + dTf (C2 , C3 ) is upper bounded by dTf (C1 , C3 ) as
dTf (C1 , C2 ) + dTf (C2 , C3 )
= λ1 + λ2 − 2mf (C1 , C2 ) + λ2 + λ3 − 2mf (C2 , C3 )
= λ1 + λ3 − 2 (min {mf (C1 , C2 ), mf (C2 , C3 )} + max {mf (C1 , C2 ), mf (C2 , C3 )} − λ2 )
≥ λ1 + λ3 − 2mf (C1 , C3 )
= dTf (C1 , C3 ).
(17)
Hence (10), (13), (14), and (17) dTf on {Tf } is a metric.
E.2
Proof of Lemma 5
Lemma 5. Let f, g : X → [0, ∞) be functions, and Tf , Tg be their cluster trees in Definition 1.
Then if f, g are continuous and Tf , Tg are finite, Tf Tg and Tg Tf implies that there exists a
homeomorphism Φ : {Tf } → {Tg } that preserves the root, i.e. Φ(X ) = X . Conversely, if there
exists a homeomorphism Φ : {Tf } → {Tg } that preserves the root, Tf Tg and Tg Tf hold.
Proof.
First, we show that Tf Tg and Tg Tf implies homeomorphism. Let Φ : {Tf } → {Tg } be the
map that gives the partial order Tf Tg in Definition 4. Then from Lemma 6, Φ̄ : E(Tf ) → E(Tg )
is injective and |E(Tf )| ≤ |E(Tg )|. With a similar argument, |E(Tg )| ≤ |E(Tf )| holds, so
|E(Tf )| = |E(Tg )|.
Since we assumed that Tf and Tg are finite, i.e. |E(Tf )| and |E(Tg )| are finite, Φ̄ becomes a bijection.
Now, let [C1 ] and [C2 ] be adjacent edges in E(Tf ), and without loss of generality, assume C1 ⊂ C2 .
We argue below that Φ̄([C1 ]) and Φ̄([C2 ]) are also adjacent edges. Then Φ(C1 ) ⊂ Φ(C2 ) holds from
14
Definition 4, and since Φ̄ is bijective, [Φ(C1 )] = Φ̄([C1 ]) and [Φ(C2 )] = Φ̄([C2 ]) holds. Suppose
e3 ∈ {Tg } such that [C
e3 ] ∈
e3 ⊂ Φ(C2 ). Then
there exists C
/ {Φ̄([C1 ]), Φ̄([C2 ])} and Φ(C1 ) ⊂ C
e
e3 ⊂ Φ(C2 )
since Φ̄ is bijective, there exists C3 ∈ {Tf } such that [Φ(C3 )] = [C3 ]. Then Φ(C1 ) ⊂ C
implies that C1 ⊂ C3 ⊂ C2 , and Φ̄ being a bijection implies that [C3 ] ∈
/ {[C1 ], [C3 ]}. This is a
e3 , and Φ̄([C1 ]) and
contradiction since [C1 ] and [C2 ] are adjacent edges. Hence there is no such C
Φ̄([C2 ]) are adjacent edges. Therefore, Φ̄ : E(Tf ) → E(Tg ) is a bijective map that sends adjacent
edges to adjacent edges, and also sends root edge to root edge.
Then combining Φ̄ : E(Tf ) → E(Tg ) being bijective sending adjacent edges to adjacent edges and
root edge to root edge, and f, g being continuous, the map Φ̄ : E(Tf ) → E(Tg ) can be extended to a
homeomorphism {Tg } → {Tf } that preserves the root.
Second, the part that homeomorphism implies Tf Tg and Tg Tf follows by Lemma 8.
E.3
Proof of Lemma 6
Lemma 6. Let f, g : X → [0, ∞) be functions, and Tf , Tg be their cluster trees in Definition
1. Suppose Tf Tg via Φ : {Tf } → {Tg }. Define Φ̄ : E(Tf ) → E(Tg ) by for [C] ∈ E(Tf )
choosing any C ∈ [C] and defining as Φ̄([C]) = [Φ(C)]. Then Φ̄ is injective, and as a consequence,
|E(Tf )| ≤ |E(Tg )|.
Proof.
We will first show that equivalence relation on {Tg } implies equivalence relation on {Tf }, i.e.
Φ(C1 ) ∼ Φ(C2 ) =⇒ C1 ∼ C2 .
(18)
Suppose Φ(C1 ) ∼ Φ(C2 ) in {Tg }. Then from Definition 4 of Φ, for any C ∈ {Tf } such that
C∈
/ [C1 , C2 ] ∪ [C2 , C1 ], Φ(C) ∈
/ [Φ(C1 ), Φ(C2 )] ∪ [Φ(C2 ), Φ(C1 )] holds. Then from definition of
Φ(C1 ) ∼ Φ(C2 ),
Φ(C) ⊂ Φ(C1 ) iff Φ(C) ⊂ Φ(C2 ) and Φ(C1 ) ⊂ Φ(C) iff Φ(C2 ) ⊂ Φ(C).
Then again from Definition 4 of Φ, equivalence relation holds for C1 and C2 holds as well, i.e.
C ⊂ C1 iff C ⊂ C2 and C1 ⊂ C iff C2 ⊂ C.
Hence (18) is shown, and this implies that
Φ̄([C1 ]) = Φ̄([C2 ]) =⇒ [Φ(C1 )] = [Φ(C2 )]
=⇒ Φ(C1 ) ∼ Φ(C2 )
=⇒ C1 ∼ C2
=⇒ [C1 ] = [C2 ],
so Φ̄ is injective.
E.4
Proof of Lemma 7
Lemma 7. Let f, g : X → [0, ∞) be functions, and Tf , Tg be their cluster trees in Definition 1. If
Tg can be obtained from Tf by adding edges, then Tf Tg holds.
Proof. Since Tg can be obtained from Tf by adding edges, there is a map Φ : Tf → Tg which
preserves order, i.e. C1 ⊂ C2 if and only if Φ(C1 ) ⊂ Φ(C2 ). Hence Tf Tg holds.
E.5
Proof of Lemma 8
Lemma 8. Let f, g : X → [0, ∞) be functions, and Tf , Tg be their cluster trees in Definition 1.
If there exists a one-to-one map Φ : {Tf } → {Tg } that is a homeomorphism between {Tf } and
Φ({Tf }) and preserves root, i.e. Φ(X ) = X , then Tf Tg holds.
Proof. For any C ∈ {Tf }, note that [C, X ] ⊂ {Tf } is homeomorphic to an interval, hence
Φ([C, X ]) ⊂ {Tg } is also homeomorphic to an interval. Since {Tg } is topologically a tree, an
interval in a tree with fixed boundary points is uniquely determined, i.e.
Φ([C, X ]) = [Φ(C), Φ(X )] = [Φ(C), X ].
15
(19)
For showing Tf Tg , we need to argue that for all C1 , C2 ∈ {Tf }, C1 ⊂ C2 holds if and only if
Φ(C1 ) ⊂ Φ(C2 ). For only if direction, suppose C1 ⊂ C2 . Then C2 ∈ [C1 , X ], so Definition 4 and
(19) implies
Φ(C2 ) ⊂ Φ([C1 , X ]) = [Φ(C1 ), X ].
And this implies
Φ(C1 ) ⊂ Φ(C2 ).
(20)
For if direction, suppose Φ(C1 ) ⊂ Φ(C2 ). Then since Φ−1 : Φ({Tf }) → {Tf } is also an homeomorphism with Φ−1 (X ) = X , hence by repeating above argument, we have
C1 = Φ−1 (Φ(C1 )) ⊂ Φ−1 (Φ(C2 )) = C2 .
(21)
Hence (20) and (21) implies Tf Tg .
F
F.1
Proofs for Section 3 and Appendix C
Proof of Lemma 1 and extreme cases
Lemma 1. For any densities p and q, the following relationships hold:
(i) When p and q are continuous, then d∞ (Tp , Tq ) = dM (Tp , Tq ).
(ii) dMM (Tp , Tq ) ≤ 4d∞ (Tp , Tq ).
(iii) dMM (Tp , Tq ) ≥ d∞ (Tp , Tq )−a, where a is defined as above. Additionally when µ(X ) = ∞,
then dMM (Tp , Tq ) ≥ d∞ (Tp , Tq ).
Proof.
(i)
First, we show dM (Tp , Tq ) ≤ d∞ (Tp , Tq ). Note that this part is implicitly shown in Eldridge et al.
[12, Proof of Theorem 6]. For all > 0 and for any x, y ∈ X , let C0 ∈ Tp (mp (x, y) − ) with
x, y ∈ C0 . Then for all z ∈ C0 , q(z) is lower bounded as
q(z) > p(z) − d∞ (Tp , Tq )
≥ mp (x, y) − − d∞ (Tp , Tq ),
so C0 ⊂ q −1 (mp (x, y) − − d∞ (Tp , Tq ), ∞) and C0 is connected, so x and y are in the same
connected component of q −1 (mp (x, y) − − d∞ (Tp , Tq ), ∞), which implies
mq (x, y) ≤ mp (x, y) − − d∞ (Tp , Tq ).
(22)
A similar argument holds for other direction as
mp (x, y) ≤ mq (x, y) − − d∞ (Tp , Tq ),
(23)
so (22) and (23) being held for all > 0 implies
|mp (x, y) − mq (x, y)| ≤ d∞ (Tp , Tq ).
(24)
And taking sup over all x, y ∈ X in (24) dM (Tp , Tq ) is upper bounded by d∞ (Tp , Tq ), i.e.
dM (Tp , Tq ) ≤ d∞ (Tp , Tq ).
(25)
Second, we show dM (Tp , Tq ) ≥ d∞ (Tp , Tq ). For all > 0, Let x be such that |p(x) − q(x)| >
d∞ (Tp , Tq ) − 2 . Then since p and q are continuous, there exists δ > 0 such that
B(x, δ) ⊂ p−1 p(x) − , ∞ ∩ q −1 q(x) − , ∞ .
2
2
Then for any y ∈ B(x, δ), since B(x, δ) is connected, p(x) − 2 ≤ mp (x, y) ≤ p(x) holds and
q(x) − 2 ≤ mq (x, y) ≤ q(x), so
|mp (x, y) − mq (x, y)| ≥ |p(x) − q(x)| −
> d∞ (Tp , Tq ) − .
16
2
Since this holds for any > 0, dM (Tp , Tq ) is lower bounded by d∞ (Tp , Tq ), i.e.
dM (Tp , Tq ) ≥ d∞ (Tp , Tq ).
(26)
(25) and (26) implies d∞ (Tp , Tq ) = dM (Tp , Tq ).
(ii)
We have already seen that for all x, y ∈ X , |mp (x, y) − mq (x, y)| ≤ d∞ (Tp , Tq ) in (24). Hence for
all x, y ∈ X ,
|[p(x) + p(y) − 2mp (x, y)] − [q(x) + q(y) − 2mq (x, y)]|
≤ |p(x) − q(x)| + |p(y) − q(y)| + 2|mp (x, y) − mq (x, y)|
≤ 4d∞ (Tp , Tq ).
Since this holds for all x, y ∈ X , so
dMM (Tp , Tq ) ≤ 4d∞ (Tp , Tq ).
(iii)
For all > 0, Let x be such that |p(x) − q(x)| > d∞ (Tp , Tq ) − 2 , and without loss of generality
assume that p(x) > q(x). Let y be such that p(y) + q(y) < inf (p(x) + q(x)) + 2 . Then mp (x, y) ≤
x
p(y) holds, and since X is connected, qinf ≤ mq (x, y) holds. Hence
[p(x) + p(y) − 2mp (x, y)] − [q(x) + q(y) − 2mq (x, y)]
≥ [p(x) + p(y) − 2p(y)] − [q(x) + q(y) − 2qinf ]
= p(x) − q(x) − (p(y) + q(y) − 2qinf )
> d∞ (Tp , Tq ) − inf (p(x) + q(x)) − 2qinf −
x
≥ d∞ (Tp , Tq ) − a − ,
where a = inf (p(x) + q(x)) − 2 min {pinf , qinf }. Since this holds for all > 0, we have
x∈X
dMM (Tp , Tq ) ≥ d∞ (Tp , Tq ) − a.
Hence 0 ≤ dMM (Tp , Tq ) ≤ 4d∞ (Tp , Tq ) holds. And both extreme cases can happen, i.e.
dMM (Tp , Tq ) = 4d∞ (Tp , Tq ) > 0 and dMM (Tp , Tq ) = 0, d∞ (Tp , Tq ) > 0 can happens.
Lemma 11. There exists densities p, q for both dMM (Tp , Tq ) = 4d∞ (Tp , Tq ) > 0 and
dMM (Tp , Tq ) = 0, d∞ (Tp , Tq ) > 0.
Proof. Let X = R, p(x) = I(x ∈ [0, 1]) and q(x) = 2I x ∈ 0, 14 + 2I x ∈ 34 , 1 . Then
d∞ (Tp , Tq ) = 1. And with x = 18 and y = 78 ,
|[p(x) + p(y) − 2mp (x, y)] − [q(x) + q(y) − 2mq (x, y)]| = |[1 + 1 − 2] − [2 + 2 − 0]|
= 4,
hence dMM (Tp , Tq ) = 4d∞ (Tp , Tq ).
1
1
Let X = [0,1), p(x)
=
2I
x
∈
0,
and
q(x)
=
2I
x
∈
2
2 , 1 . Then d∞ (Tp , Tq ) = 2. And
1
1
for any x ∈ 0, 2 and y ∈ 2 , 1 ,
|[p(x) + p(y) − 2mp (x, y)] − [q(x) + q(y) − 2mq (x, y)]| = |(2 + 0 − 0) + (0 + 2 − 0)|
= 0.
1
1
A similar case holds for x ∈ 2 , 1 and y ∈ 0, 2 . And for any x, y ∈ 0, 12 ,
|[p(x) + p(y) − 2mp (x, y)] − [q(x) + q(y) − 2mq (x, y)]| = |(2 + 2 − 4) + (0 + 0 − 0)|
= 0.
1
and a similar case holds for x, y ∈ 2 , 1 . Hence dMM (Tp , Tq ) = 0.
17
y0
c
x0
Figure 6: The example used in the proof of Theorem 9.
F.2
Proof of Theorem 9
Theorem 9. Let B(x) be the smallest set B ∈ Tp such that x ∈ B. dTp (x, y) is not Hadamard
differentiable for x 6= y when one of the following two scenarios occurs:
(i) min{p(x), p(y)} = p(c) for some critical point c.
(ii) B(x) = B(y) and p(x) = p(y).
Proof. For x, y ∈ K, note that the merge height satisfies
mp (x, y) = min{t : (x, y) are in the same connected component of L(t)}.
Recall that
dTp (x, y) = p(x) + p(y) − 2mp (x, y).
Note that the modified merge distortion metric is dMM (p, q) = supx,y |dTp (x, y) − dTq (x, y)|.
A feature of the merge height is that
mp (x, y) = p(x) ⇒ B(y) ⊂ B(x)
mp (x, y) = p(y) ⇒ B(x) ⊂ B(y)
mp (x, y) 6= p(y) or p(x) ⇒ ∃c(x, y) ∈ C s.t. mp (x, y) = p(c(x, y)).
where C is the collection of all critical points. Thus, we have
p(x) − p(y)
dTp (x, y) = p(y) − p(x)
p(x) + p(y) − 2p(c(x, y))
if B(y) ⊂ B(x)
if B(x) ⊂ B(y) .
otherwise
Case 1:
We pick a pair of x0 , y0 as in Figure 6. Now we consider a smooth symmetric function g(x) > 0 such
that it peaks at 0 and monotonically decay and has support [−δ, δ] for some small δ > 0. We pick δ
small enough such that p (x0 ) = p(x0 ), p (y0 ) = p(y0 ). For simplicity, let g(0) = maxx g(x) = 1.
Now consider perturbing p(x) along g(x − c) with amount . Namely, we define
p (x) = p(x) + · g(x − c).
For notational convenience, define ξp, = dTp (x0 , y0 ). When || is sufficiently small, define
ξp, (x0 , y0 ) = dTp (x0 , y0 ) if > 0,
ξp, (x0 , y0 ) = dTp (x0 , y0 ) − 2
if < 0.
This is because when > 0, the p (c) > p(c), so the merge height for x0 , y0 using p is still the same
as p(y0 ), which implies ξp, (x0 , y0 ) = dTp (x0 , y0 ). On the other hand, when < 0, p (c) < p(c), so
the merge height is no longer p(y0 ) but p (c). Then using the fact that || = p(c) − p (c) we obtain
the result.
18
Now we show that dTp (x0 , y0 ) is not Hadamard differentiable. In this case, φ(p) = ξp (x0 , y0 ). First,
we pick a sequence of n such that n → 0 and n > 0 if n is even and n < 0 if n is odd. Plugging
t ≡ n and qt = g into the definition of Hadamard differentiability, we have
φ0 (p) ≡
ξp,n (x0 , y0 ) − dTp (x0 , y0 )
n
is alternating between 0 and 2, so it does not converge. This shows that the function dTp (x, y) at such
a pair of (x0 , y0 ) is non-Hadamard differentiable.
Case 2:
The proof of this case uses the similar idea as the proof of case 1. We pick the pair (x0 , y0 ) satisfying
the desire conditions. We consider the same function g but now we perturb p by
p (x) = p(x) + · g(x − x0 ),
and as long as δ is small, we will have p (y0 ) = p(y0 ). Since B(x0 ) = B(y0 ) and p(x0 ) = p(y0 ),
dTp (x0 , y0 ) = 0. When > 0, ξp, (x0 , y0 ) = , and on the other hand, when < 0, δ (x0 , y0 ) = −.
In this case, again, φ(p) = ξp (x0 , y0 ). Now we use the similar trick as case 1: picking a sequence of
n such that n → 0 and n > 0 if n is even and n < 0 if n is odd. Under this sequence of n , the
‘derivative’ along g
ξp,n (x0 , y0 ) − dTp (x0 , y0 )
φ0 (p) ≡
n
is alternating between 1 and −1, so it does not converge. Thus, dTp (x, y) at such a pair of (x0 , y0 ) is
non-Hadamard differentiable.
G
G.1
Proofs for Section 4 and Appendix D
Proof of Lemma 2
Lemma 2. Let ph = E[b
ph ] where pbh is the kernel estimator with bandwidth h. We assume that p is
a Morse function supported on a compact set with finitely many, distinct, critical values. There exists
h0 > 0 such that for all 0 < h < h0 , Tp and Tph have the same topology in Appendix A.
Proof. Let S be the compact support of p. By the classical stability properties of the Morse
function, there exists a constant C0 > 0 such that for any other smooth function q : S → R with
kq − pk∞ , k∇q − ∇pk∞ , k∇2 q − ∇2 pk∞ < C0 , q is a Morse function. Moreover, there exist
two diffeomorphisms h : R → R and φ : S → S such that q = h ◦ p ◦ φ See e.g., proof of [6,
Lemma 16]. Further, h should be nondecreasing if C0 is small enough. Hence for any C ∈ Tp (λ),
since q ◦ φ−1 (C) = h ◦ p(C), so φ−1 (C) is a connected component of Tq (h(λ)). Now define
Φ : {Tp } → {Tq } as Φ(C) = φ−1 (C). Then since φ is a diffeomorphism, C1 ⊂ C2 if and only if
Φ(C1 ) = φ−1 (C1 ) ⊂ φ−1 (C2 ) = Φ(C2 ), hence Tp Tq holds. And from p ◦ φ = h−1 ◦ q, we
can similarly show Tq Tp as well. Hence from Lemma 5, two trees Tp and Tq are topologically
equivalent according to the topology in Appendix A.
Now by the nonparametric theory (see e.g. page 144-145 of [19], and [22]),
qthere is a constant
2
0
C1 > 0 such that kph − pk2,max ≤ C1 h when h < 1. Thus, when 0 ≤ h ≤ C
C1 , Th = Tph and
T = Tp have the same topology.
G.2
Proof of Lemma 10
Lemma 10. Suppose that the life function satisfies: for all [C] ∈ E(Tbh ), lifetop ([C]) ≤ life([C]).
Then
(i) P runedlife,btα (Tbh ) Tpbh .
(ii) there exists a function pe such that Tpe = P runedlife,btα (Tbh ).
bα .
(iii) pe in (ii) satisfies pe ∈ C
19
Proof.
(i)
This is implied by Lemma 7.
(ii)
Note that P runedlife,btα (Tbh ) is generated by function pe defined as
n
o
pe(x) = sup λ : there exists C ∈ Tbh (λ) such that x ∈ C and life([C]) > 2b
tα + b
tα .
(iii)
Let C0 :=
S
{C : life([C]) ≤ 2b
tα }. Then note that
n
o
pb(x) = sup λ : there exists C ∈ Tbh (λ) such that x ∈ C ,
so for all x, pe(x) ≤ pb(x) + b
tα , and if x ∈
/ C0 , pe(x) = pb(x) + b
tα . Then note that
n
o
λ : there exists C ∈ Tbh (λ) such that x ∈ C
n
o
\ λ : there exists C ∈ Tbh (λ) such that x ∈ C and life([C]) > 2b
tα
n
o
⊂ λ : there exists C ∈ Tbh (λ) such that x ∈ C and life([C]) ≤ 2b
tα
Let ex := max e : x ∈ ∪e, life(e) ≤ 2b
tα . Then note that x ∈ C and life([C]) ≤ 2b
tα implies that
we can find some B ∈ ex such that C ⊂ B, so
n
o
λ : there exists C ∈ Tbh (λ) such that x ∈ C and life([C]) ≤ 2b
tα ⊂ cumlevel(ex ).
Hence
pb(x) + b
tα − pe(x) ≤ sup{cumlevel(ex )} − inf {cumlevel(ex )}
= lifetop (ex )
≤ life(ex ) ≤ 2b
tα ,
and hence
pb(x) − b
tα ≤ pe(x) ≤ pb(x) + b
tα .
20
| 10math.ST
|
Comments on: “Lyapunov matrices for a class of time delay systems” by V. L. Kharitonov
Murad Abu-Khalaf
MIT CSAIL, MIT, Cambridge, MA
Suat Gumussoy
IEEE Member
[email protected]
[email protected]
February 19, 2018
Abstract
We prove that an auxiliary two-point boundary value problem presented in V. L. Kharitonov, Lyapunov
matrices for a class of time delay systems, Systems & Control Letters 55 (2006) 610-617 has linearly
dependent boundary conditions, and consequently a unique solution does not exist. Therefore, the twopoint boundary value problem presented therein fails to be a basis for constructing Lyapunov matrices for
the class of time delay systems investigated.
1
Introduction
In [1], the author considers a stable linear time-delay system of the form
0
x(=
t ) A0 x(t ) + A1 x(t − h) + ∫ G(θ )x(t + θ )dθ ,
t ≥0
(1)
−h
and is interested in
∞
U(τ ) ∫ ΦT (t )W Φ(t + τ )dt ,
0
∀τ ∈
(2)
where Φ(t ) is the fundamental matrix and U(⋅) is well-defined and is referred to as a Lyapunov matrix. It
was shown in both [1] and [2], and references therein, that U(⋅) is characterized by a dynamic, a
symmetric, and an algebraic property
0
U (τ=) U(τ )A0 + U(τ − h)A1 + ∫ U(τ + θ )G(θ )dθ ,
τ ≥0
−h
U(−τ ) =
U(τ )T
(3)
0
0
−h
−h
W U(0)A0 + U(−h)A1 + ∫ U(θ )G(θ )dθ + A0T U(0) + A1T U(h) + ∫ G(θ )T U(−θ )dθ
−=
Note that x(t ) is an n ×1 vector function of time; G(⋅) and U(⋅) are n × n function matrices; A0 , A1 , and
W = W T are n × n constant matrices; and h a nonnegative scalar.
To solve for U(⋅) , [1] proposed solving for (3) indirectly by first solving for an auxiliary two-point boundary
value problem and use the solution of this auxiliary system to construct a solution to (3). The following
concrete example was provided in Section 4 of [1]
0
x=
(t ) A0 x(t ) + A1 x(t − 1) + ∫ G(θ )x(t + θ )dθ
−1
where
0.3
−1 0
0 1
0.3 0
0
A0 =
, A1 =
, B0 =
, B1
=
0 −1
−1 0
0 0.3
−0.3 0
=
G(θ ) sin(πθ )B0 + cos(πθ )B1
and [1] derived based on its methodology the following auxiliary two-point boundary value problem
Z (τ ) = Z (τ )A0 + V (τ )A1 + X 0 (τ )B0 + X1 (τ )B1
V (τ ) =
− A1T Z (τ ) − A0T V (τ ) − B0T X 0 (τ ) + B1T X1 (τ )
X 0 (τ ) = −π X1 (τ )
X 1 (τ ) = Z (τ ) + V (τ ) + π X 0 (τ )
(4)
with boundary conditions
W Z (0)A0 + A0T Z (0) + V (0)A1 + A1T V T (0) + X 0 (0)B0 + B0T X 0T (0) + X1 (0)B1 + B1T X1T (0)
−=
=
0 Z (0) − V (1)
0 X 0 (0) − X 0T (1)
=
(5)
0 X1 (0) + X1T (1)
=
Once Z (τ ) is computed, then U(τ ) can be constructed
=
U(τ ) Z (τ ),
τ ≥0
=
U(−τ ) Z (τ ),
τ ≥0
T
(6)
The work [1], and similarly [2], does not examine existence and uniqueness of solutions conditions for the
auxiliary two-point boundary value problem (4)-(5). Following our experience in not being able to
reproduce the numerical results of this example of Section 4 of [1], we provide in what follows a proof
showing that the boundary conditions of system (4)-(5) are in fact linearly dependent. Therefore, (4)-(5)
does not yield a well-defined solution, and cannot generate a unique Z (τ ) that can be used to construct
a unique U(τ ) . In particular, we show that the following set of boundary conditions is linearly dependent
=
0 Z (0) − V (1)
=
0 X 0 (0) − X 0T (1)
(7)
T
1
=
0 X1 (0) + X (1)
2
Linear Dependence of Boundary Conditions
Rewriting the auxiliary two-point boundary value problem in Kronecker product form [3], equation (4)
becomes
H
T
T
A1 ⊗ In
B0T ⊗ In
B1T ⊗ In z
z A0 ⊗ In
T
T
T
In ⊗ B1T v
v = −In ⊗ A1 −In ⊗ A0 −In ⊗ B0
x 0
0
0
0
−π In ⊗ In x0
In ⊗ In
0
π In ⊗ In
x1 In ⊗ In
x1
(8)
where z = vec(Z ) , v = vec(V ) , x0 = vec( X 0 ) , and x1 = vec( X1 ) are n2 × 1 vectors; vec(A) generates a vector
by stacking the columns A into a single column; In is an identity matrix of size n × n ; 0 is a n2 × n2 matrix
of zeros. Note the following properties follow from [3]
=
=
(A), T T T
vec(AT ) Tvec
=) (C T ⊗ A)vec(B)
vec(ABC
T (A ⊗ C T )Tvec(B) =
T (A ⊗ C T )vec(BT )
= Tvec(C T BT AT )
= vec(ABC )
= (C T ⊗ A)vec(B)
Moreover, equation (7) becomes
0
=
0
In ⊗ In
0
0
0
0
0 In ⊗ In
0
0
z(0)
0
0 −In ⊗ In
v
(0)
+ 0
0
0
x0 (0)
In ⊗ In
0
0
x1 (0)
z(1)
0
v(1)
−T 0
x (1)
0 T 0
x1 (1)
0
(9)
Evaluating the following at τ = 1 and plugging in (9)
z(τ )
v(τ )
= eHτ
x0 (τ )
x1 (τ )
z(0)
v(0)
x0 (0)
x1 (0)
(10)
results in
0
=
0
In ⊗ In
0
0
0
0
0 In ⊗ In
0
0
0 0 −In ⊗ In
0 + 0
0
In ⊗ In 0
0
0 0
−T 0 eHτ
0 T
z(0)
v(0)
x0 (0)
x (0)
1
(11)
which has 3n2 rows. Clearly if the rank of the row space of (11) is less than 3n2 , then the rows are linearly
=
0 Z (1)T − V (0) –
dependent. Without loss of generality, we add to (9) an additional boundary condition
one we initially guess to be in the row space of (9) – and investigate instead the squared matrix
0
0
In ⊗ In
0
0
0
0
In ⊗ In
0
0
0
0
In ⊗ In
0
0 z(0) 0
0 v(0) T
−
0 x0 (0) 0
In ⊗ In x1 (0) 0
which after evaluating (10) at τ = 1 and plugging in (12) we get
T 0
0 0
0 z(1)
0 v(1)
0 T 0 x0 (1)
0 0 −T x1 (1)
(12)
0
=
0
I
J
In ⊗ In
0
0
0 0 T 0 0 z(0)
0
0 T 0 0 0 H v(0)
In ⊗ In
0
e
−
0
0
0 0 0 T 0 x0 (0)
In ⊗ In
0
0
In ⊗ In 0 0 0 −T x1 (0)
0
(13)
z(0)
0
v(0)
H
= (I − Je ) x (0)
0
0
x
1 (0)
(14)
T
Note that from T = T T , it follows that J = J T and that=I JJ=
J T J . Moreover, system (13) has 4n2 rows.
If the rank of (13) is less than 3n2 , then this implies at least n2 + 1 rows of (13), and similarly (12), can be
eliminated. This implies that (11), and similarly (9), must be linearly dependent. This is further investigated
in the subsequent Theorem 1 and Corollary 1 which are influenced by an analysis appearing in [4].
Lemma 1: J T HJ = −H , where H is given by (8) and J by (13).
Proof:
0
T
JT HJ =
0
0
T
0
0
0
0
0
0
0
0
A0T ⊗ In
T
−In ⊗ A1
T
0
0 −T In ⊗ In
A1T ⊗ In
−In ⊗ A0T
0
B0T ⊗ In
−In ⊗ B0T
0
In ⊗ In
π In ⊗ In
B1T ⊗ In 0
In ⊗ B1T T
−π In ⊗ In 0
0
0
−T (In ⊗ A1T ) −T (In ⊗ A0T ) −T (In ⊗ B0T ) T (In ⊗ B1T ) 0
T (A0T ⊗ In ) T (A1T ⊗ In ) T (B0T ⊗ In ) T (B1T ⊗ In ) T
=
0
0
0
−π T 0
0
−T
−T
−π T
0
T
0
0
0
T
0
0
0
0
0
0
0
0
T
0 −T
0
0
0
0
T 0
0 −T
−T (In ⊗ A0T )T −T (In ⊗ A1T )T −T (In ⊗ B0T )T −T (In ⊗ B1T )T
T (A1T ⊗ In )T
T (A0T ⊗ In )T
T (B0T ⊗ In )T −T (B1T ⊗ In )T
=
0
0
0
π TT
0
−TT
−TT
−π TT
− A0T ⊗ In − A1T ⊗ In −B0T ⊗ In −B1T ⊗ In
In ⊗ A1T
In ⊗ A0T
In ⊗ B0T −In ⊗ B1T
=
0
0
0
π TT
−TT
−π TT
0
−TT
(15)
= −H
2n2 .
Theorem 1: Dimension of the nullspace of I − JeH in (14) is 2n2 , i.e. dim Null(I − JeH ) =
∎
Proof: Let λi be an eigenvalue of H with algebraic multiplicity mi – geometric multiplicity is µi with
µi ≤ mi and represents the dimension of associated eigenspace – whose Jordan block form is Σ i ∈ m ×m
i
i
2
and its corresponding generalized eigenvectors Vi ∈ 4 n ×mi . If λi is an eigenvalue, then so is −λi
HVi = Vi Σ i
T
J HJJ T=
Vi J T Vi Σ i
(16)
−HJ T Vi =J T V Σ i
H(=
J T Vi ) ( J T Vi )(−Σ i )
Thus, all the distinct eigenvalues of H are { ± λ1 , ± λ2 , , ± λs } where Re λi ≤ 0 and s ≤ 2n2 each repeated
per its algebraic multiplicity. Let the Jordan form corresponding to { ± λ1 , ± λ2 , , ± λs } be
2
2
Σ
∈ 4 n ×4 n
−Σ
where
Σ1
2 n2 ×2 n2
=
Σ
∈
Σ s
2
corresponds to {λ1 , λ2 , , λs } and their respective generalized eigenvectors Vi ∈ 4 n ×mi forming
2
2
=
V [V1 Vs ]∈ 4 n ×2n
Note that from (16) it follows that
H V
J T V = V
Σ 0
J T V
0 −Σ
By diagonalizing the exponential matrix – in Jordan Block Form – as follows
eHτ = V
e Στ
J T V
0
0
V
e
−Στ
J T V
−1
the state trajectories are written in terms of its eigenmotions
z(τ )
v(τ )
= V
x0 (τ )
x1 (τ )
Letting
Στ
e
J T V
0
0
V
e
−Στ
z(0)
−1 v (0)
J T V
x0 (0)
x1 (0)
(17)
c1
c
2 = V
c3
c4
z(0)
−1 v (0)
T
J V
x0 (0)
x1 (0)
this simplifies to
z(τ )
v(τ )
= V
x0 (τ )
x1 (τ )
Στ
e
J T V
0
c1
0 c2
e −Στ c3
c4
(18)
To understand how arbitrary coefficients c1 , , c4 are, evaluate (18) at τ = 0 and plug in (14) to get
0
H
= (I − Je ) V
0
c1
c
T
J V 2
c3
c4
(19)
Evaluating (17) at τ = 1 and recalling that J = J T , we have
0
=
0
I − J V
e
J T V
0
Σ
0
V
e −Σ
−1
J T V V
c1
e Σ 0 c2
T
T
= V J V − J V J V
−Σ
0 e c3
c4
c
c
c
c
=V 1 − e −Σ 3 + J T V 3 − JVe Σ 1
c2
c4
c4
c2
= V
Since V
c1 −Σ c3
− e
c2
c4
T
J V
c
c
3 − eΣ 1
c2
c4
J T V is linearly independent, it follows that
c3 Σ c1
c = e c
4
2
This implies that
c1
c
T
J V 2
c3
c4
z(0)
v(0)
= V
x0 (0)
x1 (0)
c1
c
T
J V 2
c3
c4
I 2 2 c
J T V 2 n ×Σ2n 1
e c2
= V
(20)
and that c1 and c2 are arbitrary thus
dim Range V
I
J T V Σ = 2n2
e
Plugging (20) into (14), it can be verified that
(I − JeH ) V
I c
I c
V J T V Σ 1 − J V
J T V Σ 1 =
e c2
e c2
= V
= V
e Σ 0 I c1
J T V
−Σ Σ
0 e e c2
eΣ c
J T V 1
I c2
I c
J T V Σ 1 − J V
e c2
I c
I c
J T V Σ 1 − V J T V Σ 1
e c2
e c2
(21)
0
=
0
Since (20) spans a subspace of the nullspace of I − JeH , Null(I − JeH ) , this implies dim Null(I − JeH ) ≥ 2n2 .
Equation (14) shows that 1 is an eigenvalue of JeH and thus dim Null(I − JeH ) equals the dimension of the
generalized eigenspace associated with 1. Similarly, -1 is an eigenvalue of JeH by noting that
(−I − Je ) V
H
0
I c1
J V Σ =
−e c2 0
T
and dim Null(−I − JeH ) equals the dimension of the generalized eigenspace associated with -1. Thus
dim Range V
I
J T V Σ = 2n2
−e
which spans a subspace of Null(−I − JeH ) . This implies dim Null(−I − JeH ) ≥ 2n2 . Since the dimensions of the
generalized eigenspaces of JeH is 4n2 , it follows that
dim Null(I − JeH ) + dim Null(−I − JeH ) ≤ 4n2
dim Null(I − JeH ) ≤ 4n2 − dim Null(−I − JeH )
≤ 2n2
2n2 .
Both dim Null(I − JeH ) ≥ 2n2 and dim Null(I − JeH ) ≤ 2n2 imply that dim Null(I − JeH ) =
∎
Corollary 1: The null space of the subsystem (11) has a dimension of at least n2 , thus the subsystem (9) is
linear dependent.
2n2 from Theorem 1, and given that (11) is a 3n2 -equations subsystem of
Proof: Since dim Null(I − JeH ) =
the 4n2 -equations system (13), then dimension of the null space of the subsystem (11) is at most 2n2 ,
and at least n2 . This implies at least n2 + 1 rows of (11), and similarly of (9), can be eliminated and thus
implies that (9) is linearly dependent.
3
Conclusion
∎
In [5], we propose an alternative structure for an auxiliary two-point boundary value problem whose
solution is well-defined, and provide a necessary and sufficient condition for the existence and uniqueness
of solutions.
References
[1]
[2]
[3]
[4]
[5]
V. L. Kharitonov, "Lyapunov matrices for a class of time delay systems," Systems & Control Letters, vol. 55, no. 7, pp. 610–617, Jul. 2006.
V. L. Kharitonov, Time-delay systems: Lyapunov Functionals and Matrices. New York: Birkhäuser, 2013.
J. Brewer, "Kronecker Products and Matrix Calculus In System Theory". IEEE Transactions on Circuits and Systems 25.9 (1978): 772-781.
Young Soo Suh, Seiichi Shin, “Stability of State Delay Systems Based on Finite Characterization of a Lyapunov Functional”, Transactions
of the Society of Instrument and Control Engineers, Volume 35, Issue 9, Pages 1170-1175, 1999.
S. Gumussoy, M. Abu-Khalaf, “Analytic Solution of a Delay Differential Equation Arising in Cost Functionals for Systems with Distributed
Delays”, 1st version reviewed by IEEE TAC, 2nd version to be submitted to IEEE TAC.
| 3cs.SY
|
Fast and Flexible Software Polar List Decoders
Mathieu Léonardon∗† , Adrien Cassagne∗ , Camille Leroux∗ , Christophe Jégo∗ ,
Louis-Philippe Hamelin‡ and Yvon Savaria†
arXiv:1710.08314v1 [cs.IT] 23 Oct 2017
∗ IMS
Laboratory, UMR CNRS 5218, Bordeaux INP, University of Bordeaux, Talence, France
† École Polytechnique de Montréal, QC, Canada
‡ Huawei Technologies Canada Co. LTD, Ottawa, ON, Canada
Abstract—Flexibility is one mandatory aspect of channel coding
in modern wireless communication systems. Among other things,
the channel decoder has to support several code lengths and code
rates. This need for flexibility applies to polar codes that are
considered for control channels in the future 5G standard. This
paper presents a new generic and flexible implementation of a
software Successive Cancellation List (SCL) decoder. A large
set of parameters can be fine-tuned dynamically without recompiling the software source code: the code length, the code
rate, the frozen bits set, the puncturing patterns, the cyclic
redundancy check, the list size, the type of decoding algorithm, the tree-pruning strategy and the data quantization. This
generic and flexible SCL decoder enables to explore tradeoffs
between throughput, latency and decoding performance. Several
optimizations are proposed to achieve a competitive decoding
speed despite the constraints induced by the genericity and
the flexibility. The resulting polar list decoder is about 4 times
faster than a generic software decoder and only 2 times slower
than a non-flexible unrolled decoder. Thanks to the flexibility
of the decoder, the fully adaptive SCL algorithm can be easily
implemented and achieves higher throughput than any other
similar decoder in the literature (up to 425 Mb/s on a single
processor core for N = 2048 and K = 1723 at 4.5 dB).
Index Terms—Polar Codes, Adaptive Successive Cancellation
List decoder, Software Implementation, 5G Standard, Generic
Decoder, Flexible Decoder.
I. I NTRODUCTION
Polar codes [1] are the first provably capacity achieving channel codes, for an infinite code length. The decoding performance of the original Successive Cancellation (SC) decoding
algorithm is however not satisfactory for short polar codes.
The Successive Cancellation List (SCL) decoding algorithm
has been proposed in [2] to counter this fact along with
the concatenation of a Cyclic Redundancy Check (CRC).
The decoding performance of SCL decoding is such that
polar codes is included in the fifth generation (5G) mobile
communications standard [3].
Cloud radio access network (Cloud-RAN) is foreseen by both
academic [4], [5] and industrial [6], [7] actors as one of the
key technologies of the 5G standard. In the Cloud-RAN the
virtualization of the physical layer (PHY) would allow for deep
cooperative multipoint processing and computational diversity
[4]. PHY-layer cooperation enables interference mitigation,
while computational diversity lets the network balance the
computational load accross multiple users. But the virtualization of the FEC decoder is a challenge as it is one of the most
computationally intensive tasks of the signal processing chain
in a Cloud-RAN context [8], [9]. Therefore, efficient, flexible
and parallel software implementations of FEC decoders are
needed to enable some of the expected features of Cloud-RAN.
To date, the fastest software implementations of SCL polar
decoders have been proposed in [10]. The high decoding speed
is achieved at the price of flexibility, because the software
decoder is only dedicated to a specific polar code. In a wireless
communication context, the source code of this fast software
polar decoder would have to be recompiled every time the
Modulation and Coding Scheme (MCS) changes, which may
happen every millisecond.
In this work, we propose a software SCL polar decoder able
to switch between different channel coding contexts (block
length, code rate, frozen bits sets, puncturing patterns and
CRC code). This property is denoted as genericity. Moreover,
the proposed decoder supports different list-based decoding
algorithms, several list sizes (L), quantization formats and treepruning techniques during a real time execution. Again, this
is done dynamically without having to recompile the software
description. We denote this feature as flexibility. The genericity
and the flexibility of the decoder are achieved without sacrificing the decoding throughput and latency thanks to several
implementation optimizations. Actually, the proposed software
SCL decoder is only 2 times slower than a polar code specific
decoder [10] and 4 times faster than a generic decoder [11].
Unlike these fast decoders, the proposed decoder supports a
fully adaptive version of SCL. It reaches 425 Mb/s on a single
processor core for N = 2048 and K = 1723 at 4.5 dB.
The genericity of our decoder makes it compliant with a wireless communication context: one can change the polar code parameters dynamically. Thanks to the decoder flexibility, some
new tradeoffs between throughput and error rate performance
are now possible. Finally, besides the genericity/flexibilitydriven improvements, some specific optimizations were implemented in order to match the state-of-the-art throughputs
of software SCL decoders. Among other optimizations, a new
sorting technique is applied to different parts of the algorithm
which is faster than any other according to our experimentations. New methods to speed up the CRC processing are also
applied. The polar functions library described in [12], [13] is
used in order to benefit from a portable implementation of
SIMD instructions.
The rest of the paper is organized as follows: Section II
describes the SCL decoding algorithm and the improved
versions. The genericity and the flexibility of the proposed
decoder are highlighted in Section III. Section IV details the
Function f
Layer
Function g
16 (LLR, ŝ)
is R = 1/2 contains 1024 informations bits. Such a code is
denoted as (2048,1024).
0
Function h
B. Polar Decoding Algorithms
1) SC decoding algorithm: The SC decoding process can be
seen as the pre-order traversal of a binary tree as shown in
Figure 1. The tree contains log2 N + 1 layers. Each layer
contains 2d nodes, where d is the depth of the layer in the
tree. Each node contains a set of 2n−d Log-Likelihood Ratios
(LLRs) λ and partial sums ŝ. The partial sums correspond to
the propagation towards the top of the tree of hard decisions
made in the update paths() function. As shown in Figure 1,
LLRs, which take real values, and partial sums, which take
binary values, are the two types of data contained in the
decoding tree, and three functions, f , g and h are necessary
for updating the nodes:
= sign(λa .λb ). min(|λa |, |λb |)
f (λa , λb )
g(λa , λb , ŝa ) = (1 − 2ŝa )λa + λb
h(ŝa , ŝb )
= (ŝa ⊕ ŝb , ŝb )
1
8 (LLR, ŝ)
2
4 (LLR, ŝ)
3
2 (LLR, ŝ)
4
1 (LLR, ŝ)
Fig. 1. Full SC decoding tree (N = 16).
speed-oriented optimizations. Finally, Section V provides the
throughput and latency performance.
II. P OLAR C ODES
In this section, we first present the polar encoding process.
Then the SC and SC-List based decoding algorithms are
reviewed. Finally we discuss the tradeoffs between speed and
decoding performance of different decoding algorithms.
A. Polar Encoding Process
In the polar encoding process, an information sequence b of
length K is transformed into a codeword x of length N . The
first step is to build a vector u in which the information bits
b are mapped on a subset uA where A ⊂ {0, ..., N − 1}.
The remaining bits uAc = (ai : i 6∈ A) called frozen bits
1
are usually set to zero. The selection of the frozen bits is 2
critical for the effectiveness of the polar codes. Two of the 3
main techniques to date for constructing polar codes are based 4
on the Density Evolution approach [14] and on the Gaussian 5
Approximation [15]. These techniques sort the polar channels 6
according to their reliability in order to choose the frozen bits
set for a given code length. Then, an intermediate vector u0 7
is generated thanks to an encoding matrix1 : u0 = uF ⊗n . 8
9
Finally the bits in the subset u0Ac are set to zero and the
10
0 ⊗n
output codeword is x = u F . This encoding method
is called systematic because the information sequence b is
present in the codeword (xA = b). In this paper, systematic 11
encoding schemes are considered. A CRC of length c may be 12
concatenated to the information sequence b in order to improve 13
the decoding performance of SCL decoding algorithms. In this 14
case, |A| = K + c and the CRC is included in uA . In this
paper, the code rate is defined as R = K/N and the c bits of 15
16
the CRC are not considered as information bits. For instance,
a polar code whose block length is N = 2048 and code rate 17
18
⊗n−1
1 0
F
0n−1
and ∀n > 1, F ⊗n =
, where
⊗n−1
⊗n−1
1 1
F
F
n = log2 (N ), N is the codeword length, and 0n is a 2n -by-2n matrix of
zeros.
1 F ⊗1
In comparison with the SCL algorithm and its derivatives,
the computational complexity of the SC algorithm is low:
O(N log2 N ). Therefore, both software [16] and hardware
[17] implementations achieve multi-Gb/s throughputs with low
latencies. The drawback of the SC decoding algorithm is its
decoding performance especially for short polar codes. This
is an issue for the future 5G wireless standard in which polar
codes are targeted for control channels, with code lengths
shorter than 2048 [3].
Algorithm 1: SCL decoding algorithm
Data: λ is a 2D buffer ([L][2N ]) to store the LLRs.
Data: ŝ is a 2D buffer ([L][N ]) to store the bits.
Function SCL decode (N, oλ , oŝ )
N 1 = N/2
2
if N > 1 then // not a leaf node
for p = 0 to L − 1 do // loop over the paths
for i = 0 to N 1 − 1 do // apply the f function
2
λ[p][oλ +N +i] = f (λ[p][oλ +i], λ[p][oλ +N 1 +i])
2
SCL decode (N 1 , oλ + N, oŝ )
2
for p = 0 to L − 1 do
for i = 0 to N 1 − 1 do // apply the g function
2
λ[p][oλ + N + i] =
g(λ[p][oλ + i], λ[p][oλ + N 1 + i], ŝ[p][oŝ + i])
2
SCL decode (N 1 , oλ + N, oŝ + N 1 )
2
2
for p = 0 to L − 1 do
for i = 0 to N 1 − 1 do // update the partial sums
2
ŝ[p][oŝ + i] = h(ŝ[p][oŝ + i], ŝ[p][oŝ + N 1 + i])
2
else // a leaf node
update paths () // update, create and delete paths
SCL decode (N, 0, 0) // launch the decoder
select best path ()
=
2) SCL decoding algorithm: The SCL algorithm is summarized in Algorithm 1. Unlike the SC algorithm, the SCL de-
coder builds a list of candidate codewords along the decoding.
At each call of the update paths() sub-routine (Alg. 1, l.16),
2L candidates are generated. A path metric is then evaluated
to keep only the L best candidates among the 2L paths. The
path metrics are calculated as in [18]. At the end of the
decoding process, the candidate codeword with the best path
metric is selected in the select best path() sub-routine (Alg. 1,
l.18). The decoding complexity of the SCL algorithm grows
as O(LN log2 N ). This linear increase in complexity with L
leads to significant improvements in BER/FER performances,
especially for small code lengths.
3) Simplified SC and SCL decoding algorithms: All aforementioned polar decoding algorithms have in common that
they can be seen as a pre-order tree traversal algorithm. In [19],
a tree pruning technique called the Simplified SC (SSC) was
applied to SC decoding. An improved version was proposed
in [17]. This technique relies on the fact that, depending on
the frozen bits location in the leaves of the tree, the definition
of dedicated nodes enables to prune the decoding tree: Rate0 nodes (R0) correspond to a sub-tree whose all leaves are
frozen bits, Rate-1 nodes (R1) correspond to a sub-tree in
which all leaves are information bits, REPetition (REP) and
Single Parity Check (SPC) nodes correspond to repetition and
SPC codes sub-trees. These special nodes, originally defined
for SC decoding, can be employed in the case of SCL decoding
as long as some modifications are made in the path metric
calculation [10]. This tree-pruned version of the algorithm is
called Simplified SCL (SSCL). The tree pruning technique can
drastically reduce the amount of computation in the decoding
process. Moreover, it increases the available parallelism by
replacing small nodes by larger ones. As will be discussed
in Section III, the tree pruning may have a small impact on
decoding performance.
4) CRC concatenation scheme: The authors in [2] observed
that when a decoding error occurs, the right codeword is
often in the final list, but not with the best path metric.
They proposed to concatenate a CRC to the codeword in
order to discriminate the candidate codewords at the final
stage of the SCL decoding. Indeed, this technique drastically
improves the FER performance of the decoder. We denote this
algorithm CA-SCL and its simplified version CA-SSCL. In
terms of computational complexity, the overhead consists in
the computation of L CRC at the end of each decoding.
5) Adaptive SCL decoding algorithm: The presence of the
CRC can be further used to reduce the decoding time by gradually increasing L. This variation of SCL is called Adaptive
SCL (A-SCL) [20]. The first step of the A-SCL algorithm is
to decode the received frame with the SC algorithm. Then, the
decoded polar codeword is checked with a CRC. If the CRC
is not valid, the SCL algorithm is applied with L = 2. If no
candidate in the list satisfies the CRC, L is gradually doubled
until it reaches the value Lmax . In this paper, we call this
version of the A-SCL decoding the Fully Adaptive SCL (FASCL) as opposed to the Partially Adaptive SCL (PA-SCL),
in which the L value is not gradually doubled but directly
increased from 1 (SC) to Lmax . The simplified versions of
TABLE I
T HROUGHPUT AND LATENCY COMPARISON OF POLAR DECODING
ALGORITHMS .
Decoding
Algorithm
SC
SSC
SCL
SSCL
CA-SSCL
PA-SSCL
FA-SSCL
BER & FER
Performances
poor
poor
good
good
very good
very good
very good
Throughput
(T )
medium
high
low
low
low
high
high
Max. Latency
(Lworst )
medium
low
high
medium
medium
medium
high
these algorithms are denoted PA-SSCL and FA-SSCL. In
order to simplify the algorithmic range, in the remainder of
the paper, only the simplified versions are considered. The
use of either FA-SSCL or PA-SSCL algorithmic improvement
introduces no BER or FER performance degradation as long as
the CRC length is adapted to the polar code length. If the CRC
length is too short, the decoding performance may be degraded
because of false detections. These adaptive versions of SSCL
can achieve higher throughputs. Indeed, a large proportion of
frames can be decoded with a single SC decoding. This is
especially true when the SNR is high. This will be further
discussed in Section III.
C. Algorithmic Comparison
In order to better distinguish all the algorithmic variations,
we compare their main features in Table I. Each algorithm is
characterized in terms of decoding performance, throughput,
and worst case latency for a software implementation. The
non-simplified versions of the adaptive SCL algorithms are
not included in the Table for readability.
The SC and especially the SSC algorithms achieve very
high throughput and low latency with poor BER and FER
performances. The SCL algorithm improves the decoding performance compared to the SC algorithm, but its computational
complexity leads to an increased latency and a lower throughput. The SSCL algorithm improves the decoding throughput
and latency without any impact in terms of BER and FER
performances, as long as the tree pruning is not too deep,
as will be discussed in Section III. Therefore, tree pruning is
applied to all the following algorithms, namely CA-SSCL, FASSCL and PA-SSCL. By applying CRC to the SCL algorithm,
one can achieve better BER and FER performances at the
cost of computational complexity overhead. The Adaptive
SCL algorithms reduce the decoding time with no impact on
BER and FER performances. Furthermore, a tradeoff between
throughput and worst case latency is possible with the use of
either PA-SSCL or FA-SSCL decoding algorithms.
To the best of our knowledge, SC and CA-SCL decoding
performances have never been investigated in the literature for
large code lengths (N > 214 ). This is probably due to the long
simulation durations. In this work the AFF3CT2 tool enables
multi-threaded and multi-nodes simulations. All the presented
2 AFF3CT is an Open-source software (MIT license) for fast forward error
correction simulations, see http://aff3ct.github.io
10−4
flexibility of a decoder includes all the customizations that can
be applied to the decoding algorithm for a given polar code:
variant of the decoding algorithm, data representation format,
list size L, tree pruning strategy, ... These customizations are
not enforced by a standard. The flexibility gives some degrees
of freedom to the decoder in order to find the best tradeoff
between decoding performance, throughput or latency for a
given polar code.
10−5
A. Genericity
100
N
10−1
212
FER
10
SC
28
216
10−2
10
SCL32
220
N SCL128
−3
220
−6
0.5
1
1.5
2
2.5
Eb /N0 (dB)
3
3.5
4
4.5
Fig. 2. Decoding performance comparison between CA-SCL and SC decoders. Code rate R = 1/2, and 32-bit CRC (GZip).
simulations use the Monte Carlo method with a Binary PhaseShift Keying (BPSK) modulation. The communication channel
is an Additive White Gaussian Noise (AWGN) channel based
on the Mersenne Twister pseudo-random number generator
(MT19937) [21] and the Box-Muller transform [22]. Figure 2
compares the BER/FER performances of CA-SCL with SC
decoding for a large range of code lengths. As expected, it
appears that the coding gain brought by the SCL algorithm
decreases for larger N values. In the case of N = 216 , the
improvement caused by the use of the CA-SCL algorithm with
L = 32 and a 32-bit GZip CRC (0x04C11DB7 polynomial)
instead of SC is about 0.75 dB compared to 1.2 dB with a
polar code of size N = 212 . For larger polar codes, N = 220 ,
the gain is reduced to 0.5 dB, even with a list depth of 128
that is very costly in terms of computational complexity.
The tradeoffs between speed and decoding performance show
some general trends. However, the efficiency of each decoding
algorithm is strongly dependent on the polar code length, code
rate, list depth and code construction. It is expected that the
best tradeoff is not always obtained with a single algorithm and
parameter set combination. It is consequently highly relevant
to use a generic and flexible decoder, that supports all variants
of the decoding algorithms. Thus, it is possible to switch from
one to another as shown in the following section.
III. G ENERIC AND F LEXIBLE P OLAR D ECODER
The main contribution of this work lies in the flexibility and
the genericity of the proposed software decoder. These terms
need to be clearly defined in order to circumvent possible
ambiguity. In the remainder of the paper, the genericity of the
decoder concerns all the parameters that define the supported
polar code such as the codeword length, the code rate, the
frozen bits set, the puncturing patterns and the concatenated
CRC. These parameters are imposed by the telecommunication standard or the communication context. In the wireless
communications context, these are constantly adapted by AMC
methods [23]. In this work, a decoder is considered generic if it
is able to support any combination of these parameters that can
be changed during a real time execution. On the other hand, the
In the context of wireless communications, the standards
enforce several different code lengths N that have to be
supported to share bandwidth between different users. This is
also the case for the code rate R that needs to be adapted to
the quality of the transmission channel. Therefore, a practical
implementation should be adapted to both N and R in realtime in order to limit latency.
A polar code is completely defined by N and the frozen bits
set uAc . Several methods exist to generate some ”good” sets
of frozen bits [14], [15]. The code rate R depends on the size
of uAc . In their original form, polar code lengths are only
powers of two. The puncturing and shortening techniques in
[24]–[26] enable to construct polar codes of any length at the
cost of slightly degraded decoding performance. The coding
scheme can be completed with the specification of a CRC.
In [10], the unrolling method is used: a specific description
of the decoder has to be generated for a specific polar code
parameter set of N , K, R, frozen bits set, puncturing pattern,
CRC. This approach leads to very fast software decoders at
the price of the genericity, since a new source code should
be generated and compiled every time the modulation and
coding scheme (MCS) changes. This method is not adapted to
wireless communication standards, in which these parameters
have to be adapted not only over time, but also for the different
users.
The proposed decoder does not use the unrolling method and is
completely generic regarding the code dimension K, the code
length N , the frozen bits set uAc and the puncturing patterns.
All of them are dynamic parameters of the decoder and can
be defined in input files. All CRC listed in [27] are available
along with the possibility to define others. It is shown in [28]
that custom CRCs for polar codes can have a very good impact
on the decoding performance.
Relying on an unique software description also implies that
the tree pruning technique also has to be dynamically defined.
Indeed, this technique depends on the frozen bits set uAc .
Not sacrificing throughput or latency while maintaining the
genericity imposed by wireless communication standards is at
the core of the proposed implementation. Flexibility in terms
of decoding algorithms, described in the following, along with
improvements presented in Section IV, is necessary to deal
with this challenge.
B. Flexibility
On one hand, the reason for the decoder genericity is the
compliance to the telecommunication standards. On the other
L
1
2
4
8
16
32
64
128
10−1
FER
10−2
10−3
10−4
7
Ti (Mb/s)
ref
215.15
26.67
12.54
6.00
2.96
1.40
0.66
0.29
R0
R1
REP
SPC4
SPC4+
6
Coded throughput (Mb/s)
100
10−5
5
4
3
2
10−6
1
1
1.5
2
3
3.5
0
10
9/
10
10
8/
7/
10
10
10
10
6/
5/
4/
3/
hand, the flexibility of the decoder regroups several algorithmic variations that are discussed in the following. These
variations allow several tradeoffs of multiple sorts, whatever
the standard. They are all included in a single source code.
In the proposed decoders the following parameters can be
changed dynamically without re-compilation: the list size L,
the tree pruning strategy, the quantization of the LLRs and
the different SCL variants. Each of these adjustments can be
applied to access to different tradeoffs between throughput,
latency, and error rate performance. As a consequence, one
can easily fine-tune the configuration of the software decoder
for any given polar code.
1) List size: As mentioned earlier, the list size L impacts both
speed and decoding performance. In Figure 3, the throughput
as well as BER and FER performances of the CA-SSCL
algorithm are shown for different L values. A (2048,1024)
polar code with a 32-bit CRC is considered. The computational
complexity increases linearly with L: the throughput is approximately halved when L is doubled, except for the case of the
SC algorithm (L = 1) which is much faster. Indeed, there is no
overhead due to the management of different candidate paths
during the decoding. For L ≥ 4 and Eb /N0 = 2, the FER is
also approximately halved when the list size L is doubled.
2) Tree pruning strategy: A second degree of flexibility is
the customization of the SCL tree pruning. The authors in
[10], [19] defined dedicated nodes to prune the decoding
tree and therefore to reduce the computational complexity. In
this proposed decoder, each dedicated node can be activated
separately. The ability to activate dedicated nodes at will is
useful in order to explore the contribution of each node type on
the throughput. Figure 4 shows the impact of the different tree
pruning optimizations on the CA-SSCL decoder throughput
depending on the code rate. The performance improvements
are cumulative. Coded throughput, in which the redundant
bits are taken in account, is shown instead of information
throughput, for which only information bits are considered
in order to illustrate the computational effort without the
influence of the fact that higher rate codes involve higher
10
10
Fig. 3. Tradeoffs between CA-SSCL decoding and throughput performances
depending on L. N = 2048, R = 0.5, and 32-bit CRC (GZip). For L = 1,
the SSC decoder is used with a (2048,1024) polar code.
2/
1/
2.5
Eb /N0 (dB)
Code rate (R = K/N )
Fig. 4. Dedicated nodes impact on CA-SSCL. N = 2048 and L = 32.
information throughput.
Without pruning, the coded throughput decreases as the code
rate increases. Indeed, frozen bit leaf nodes are faster to
process than information bit leaf nodes, in which a threshold
detection is necessary. As there are more R0 and REP nodes
in low code rates, the tree pruning is more efficient in the case
of low code rates. The same explanation can be given for R1
nodes in high code rates. R1 node pruning is more efficient
than R0 node pruning on average. Indeed, a higher amount of
computations is saved in R1 nodes than in R0 nodes.
It has also been observed in [10] that when the SPC node
size is not limited to 4, the decoding performance may be
degraded. Consequently the size is limited to 4 in SPC4. In
SPC4+ nodes, there is no size limit. The two node types are
considered in Figure 4. Therefore, the depth at which dedicated
nodes are activated in the proposed decoder can be adjusted,
in order to offer a tradeoff between throughput and decoding
performance.
According to our experiments, the aforementioned statement
about performance degradation caused by SPC4+ nodes is
not always accurate depending on the code and decoder
parameters. The impact of switching on or off SPC4+ nodes
on decoding performance and throughput at a FER of 10−5
is detailed in Table II. It shows that SPC4+ nodes have only
a small effect on the decoding performance. With L = 8,
TABLE II
E FFECTS OF THE SPC4+ NODES ON THE CA-SSCL @ 10−5 FER
L
8
32
R
1/3
1/2
2/3
1/3
1/2
2/3
N = 256
Eb /N0 Thr.
loss
gain
%
dB
0.15
09.7
0.09
08.6
0.03
20.5
0.52
11.8
0.30
10.3
0.27
22.6
N = 1024
Eb /N0 Thr.
loss
gain
%
dB
0.03
12.6
0.04
16.4
0.04
11.3
0.19
12.9
0.24
16.5
0.22
15.2
N = 4096
Eb /N0 Thr.
loss
gain
%
dB
0.02
09.5
0.07
20.2
0.09
14.3
0.22
12.5
0.26
19.9
0.25
17.1
512
10−2
10−3
10−3
10−6
10−5
10−6
REP2+
16-bit REP2+
8-bit REP2+
8-bit REP8float
−7
10
10−8
10−9
10−2
FER
10−5
3.5
4
Eb /N0 (dB)
10
(a): Impact of the REP node size
on fixed-point SSCL decoding
3.5
4
Eb /N0 (dB)
(b): Impact of the CRC size
on SSCL and A-SSCL decoding
3) LLR Quantization: Another important parameter in both
software and hardware implementations is the quantization
of data in the decoder. More specifically, the representations
of LLRs and partial sums in the decoder have an impact
on decoding performance. Quantized implementations of the
SC algorithm have already been proposed in [30] but to the
best of our knowledge, the proposed decoder is the first SCL
software implementation that can benefit from the 8-bit and
16-bit fixed-point representations of LLRs and internal path
metrics. In the 8-bit mode LLRs and path metrics are saturated
between −127 and +127 after each operation. Moreover, to
avoid overflows, the path metrics are normalized after each
update paths() call (cf. Alg. 1) by subtracting the smallest
metric to each one of them. Figure 5a shows the BER and FER
performances of the CA-SSCL decoder for 32-bit floatingpoint, 16-bit and 8-bit fixed-point representations. One can
TABLE III
T HROUGHPUT AND LATENCY COMPARISONS BETWEEN FLOATING - POINT
(32- BIT ) AND FIXED - POINT (16- BIT AND 8- BIT ) A DAPTIVE SSCL
DECODERS . C ODE (2048,1723), L = 32 AND 32- BIT CRC (G ZIP ).
PA-SSCL
FA-SSCL
Prec.
Lworst
635
622
651
1201
1198
1259
3.5 dB
Lavg
Ti
232.3
7.6
219.6
8.0
232.4
7.6
67.2
26.1
68.7
25.6
71.8
24.4
4.0
Lavg
41.7
40.1
41.2
8.5
7.7
7.7
dB
Ti
42.1
43.8
42.6
207.8
225.7
227.3
4.5
Lavg
7.4
6.6
6.5
5.1
4.3
4.1
FA
PA
8
32
128
64
32
16
8
4
2
1
4.5
an SNR degradation lower than 0.1 dB is observed, except
for one particular configuration Throughput improvements of
8 to 20 percents are observed. If L = 32, the SNR losses
are more substantial (up to 0.5 dB), whereas throughput
improvements are approximately the same. Besides this observation, Table II shows how the proposed decoder flexibility
in the AFF3CT environment enables to optimize easily the
decoder tree pruning, both for software implementations or
for hardware implementations in which tree pruning can also
be applied [29].
32-bit
16-bit
8-bit
32-bit
16-bit
8-bit
L
10−7
Fig. 5. Decoding performance of the SSCL and the A-SSCL decoders. Code
(2048,1723), L = 32.
Decoder
10−4
10−6
CRC SSCL A-SSCL
8-bit
16-bit
32-bit
10−8
10−9
10−3
10−5
−7
4.5
256
10−1
10−4
FER
FER
10−4
100
Info. throughput (Mb/s)
10−2
dB
Ti
237.6
267.5
268.3
345.5
408.7
425.9
2.5
3
3.5
4
Eb /N0 (dB)
4.5
2.5
3
3.5
4
Eb /N0 (dB)
4.5
Fig. 6. Frame Error Rate (FER) performance and throughput of the Fully
and Partially Adaptive SSCL decoders (FA and PA). Code (2048,1723) and
32-bit CRC (GZip). 32-bit floating-point representation.
observe that the REP nodes degrade the decoding performance
in a 8-bit representation because of accumulation (red triangles
curve). Indeed, it is necessary to add all the LLRs of a
REP node together in order to process it, which may lead
to an overflow in the case of fixed-point representation. It
can happen when the size of the repetition nodes is not
limited (REP2+ ). However, thhe size limitation of the repetition
nodes to 8 (REP8- ) fixes this issue. In Table III, maximum
latency (Lworst in µs), average latency (Lavg in µs) and
information throughput (Ti in Mb/s) are given. Note that
in 8-bit configuration only the REP8- nodes are used. The
fixed-point implementation reduces, on average, the latency.
In the high SNR region, the frame errors are less frequent.
Therefore, the SCL algorithm is less necessary than in low
SNR regions for Adaptive SCL algorithms. As the gain of
fixed-point implementation benefits more to the SC algorithm
than to the SCL algorithm, the throughput is higher in high
SNR regions. For instance, up to 425.9 Mb/s is achieved
in 8-bit representation with the FA-SSCL decoder. Note that
the improvements described in Section IV are applied to the
decoders that are given in Table III.
4) Supporting different variants of the decoding algorithms:
Besides the L values, the tree pruning and quantization aspects, the proposed software polar decoder supports different
variants of the SCL algorithm: CA-SSCL, PA-SSCL, FASSCL.
As shown in [10], the adaptive version of the SCL algorithm
yields significant speedups, specially for high SNR. The
original adaptive SCL described in [20], denoted as Fully
Adaptive SCL (FA-SSCL) in this paper, gradually doubles
the list depth L of the SCL decoder when the CRC is not
valid for any of the generated codewords at a given stage
until the value Lmax . By contrast, the adaptive decoding
algorithm implemented in [10], called in this paper Partially
Adaptive SCL (PA-SSCL), directly increases the list depth
from 1 (SC) to Lmax . In Figure 6, the two versions (FASSCL and PA-SSCL) are compared on a (2048,1723) polar
code and 32-bit CRC (GZip). The LLRs values are based on
a 32-bit floating point representation. Note that as the FER
performance of PA-SSCL and FA-SSCL are exactly the same,
the related error performance plots completely overlap. The
throughput of the FA-SSCL algorithm is higher than that of the
PA-SSCL algorithm for some SNR values, depending on the
code parameters. Considering typical FER values for wireless
communication standards (10−3 to 10−5 ), in the case of a
(2048,1723) polar code, the throughput of FA-SSCL is double
that of PA-SSCL with L = 8, while it is multiplied by a factor
of 7 with L = 32. The drawback of FA-SSCL is that although
the average latency decreases, the worst case latency increases.
The adaptive versions of the algorithm achieve better throughputs, but CA-SCL may also be chosen depending on the
CRC. One may observe in Figure 5b that an adaptive decoder
dedicated to an 8-bit CRC with a (2048,1723) polar code and
L = 32 leads to a loss of 0.5 dB for a FER of 10−5 compared
to its non adaptive counterpart.
Both polar code genericity and decoding algorithm flexibility are helpful to support the recommendations of wireless
communications in an SDR or cloud RAN context. The code
and decoder parameters can be dynamically changed in the
proposed decoder, while maintaining competitive throughput
and latency. The following section introduces algorithmic
and implementation improvements applied in the proposed
decoders to keep a low decoding time.
1 class API_polar
2{
3
template <typename R>
4
mipp::Reg<R> f_simd(const mipp::Reg<R> &la,
5
const mipp::Reg<R> &lb)
6
{
7
auto abs_la = mipp::abs(la);
8
auto abs_lb = mipp::abs(lb);
9
auto abs_min = mipp::min(abs_la, abs_lb);
10
auto sign
= mipp::sign(la, lb);
11
auto lc
= mipp::neg(abs_min, sign);
12
13
return lc;
14
}
15
16
template <typename B, typename R>
17
mipp::Reg<R> g_simd(const mipp::Reg<R> &la,
18
const mipp::Reg<R> &lb,
19
const mipp::Reg<B> &sa)
20
{
21
auto neg_la = mipp::neg(la, sa);
22
auto lc
= neg_la + lb;
23
24
return lc;
25 }
26
27 template <typename B>
28 mipp::Reg<B> h_simd(const mipp::Reg<B>& sa,
29
const mipp::Reg<B>& sb)
30 {
31
return sa ˆ sb;
32 }
33 };
IV. S OFTWARE IMPLEMENTATION OPTIMIZATIONS
Fig. 7. C++ SIMD implementation of the f , g and h functions.
The genericity and flexibility of the formerly described decoder prevent from using some optimizations. Unrolling the
description as in [10] is not possible at runtime, although
code generation could be used to produce an unrolled version
of any decoder as in [12]. Moreover, in the case of large
code lengths, the unrolling strategy can generate very large
compiled binary files. This can cause instruction cache misses
that would dramatically impact the decoder throughput. As
this unrolling method is not applied, some implementation
improvements are necessary in order to be competitive with
specific decoders of the literature. The software library for
polar codes from [12], [13] enables to benefit from the SIMD
instructions for various target architectures. Optimizations of
CRC checking benefit to both the non-adaptive and adaptive
versions of the CA-SCL algorithms. The new sorting technique
presented in Section IV-C can be applied to each variation
of the SCL algorithm. Finally, an efficient implementation
of the partial sums memory management is proposed. It is
particularly effective for short polar codes.
implementation is based on MIPP, a SIMD wrapper for the
intrinsic functions (assembly code), and the template metaprogramming technique. Consequently, the description is clear,
portable, multi-format (32-bit floating-point, 16-bit and 8-bit
fixed-points) and as fast as an architecture specific code. The
mipp::Reg<B> and mipp::Reg<R> types correspond to
SIMD registers. B and R define the type of the elements that
are contained in this register. B for bit could be int, short
or char. R for real could be float, short or char. In
Figure 7, each operation is made on multiple elements at the
same time. For instance, line 22, the addition between all the
elements of the neg_la and lb registers is executed in one
CPU cycle.
In the context of software decoders, there are two well-known
strategies to exploit SIMD instructions: use the elements of a
register to compute 1 )many frames in parallel (INTER frame)
or 2) multiple elements from a single frame (INTRA frame).
In this paper, only the INTRA frame strategy is considered.
The advantage of this strategy is the latency reduction by
comparison to the INTER frame strategy. However, due to
the nature of the polar codes, there are sometimes not enough
elements to fill the SIMD registers completely. This is especially true in the nodes near the leaves. For this reason,
SIMD instructions in the lower layers of the tree do not
bring any speedup. In this context, the building blocks of
the Polar API automatically switch from SIMD to sequential
implementations. In the case of the CA-SSCL algorithm, using
SIMD instructions for decoding a (2048, 1723) polar code
A. Polar Application Programming Interface
Reducing the decoding time with SIMD instructions is a
classical technique in former software polar decoder implementations. The proposed list decoders are based on specific
building blocks included from the Polar API [12], [13]. These
blocks are fast and optimized implementations of the f ,
g, h (and their variants) polar intrinsic functions. Figure 7
details the SIMD implementation of these functions. This
leads to an improvement of 20% of the decoding throughput
on average for different values of the list depth L.
By profiling the Adaptive SCL decoder, one may observe that
a significant amount of time is spent to process the cyclic
redundancy checks. Its computational complexity is O(LN )
versus the computational complexity of the SCL decoding,
O(LN log N ). The first is not negligible compared to the
second.
In the adaptive decoder, the CRC verification is performed a
first time after the SC decoding. In the following, we show
how to reduce the computational complexity of these CRC
verifications.
First, an efficient CRC checking code has been implemented.
Whenever the decoder needs to check the CRC, the bits are
packed and then computed 32 by 32. In order to further
speed up the implementation, a lookup table used to store
pre-computed CRC sub-sequences, and thus reduce the computational complexity.
After a regular SC decoding, a decision vector of size N is produced. Then, the K information bits must be extracted to apply
cyclic redundancy check. The profiling of our decoder description shows that this extraction takes a significant amount of
time compared to the check operation itself. Consequently, a
specific extraction function was implemented. This function
takes advantage of the leaf node type knowledge to perform
efficient multi-element copies.
Concerning SCL decoding, it is possible to sort the candidates
according to their respective metrics and then to check the
CRC of each candidate from the best to the worst. Once a
candidate with a valid CRC is found, it is chosen as the
decision. This method is strictly equivalent to do the cyclic
redundancy check of each candidate and then to select the one
with the best metric. With the adopted order, decoding time is
saved by reducing the average number of checked candidates.
C. LLR and Metric Sorting
Metric sorting is involved in the aforementioned path selection
step, but also in the update paths() sub-routine (Alg. 1,
L16) and consequently in each leaf. Sorting the LLRs is
also necessary in R1 and SPC nodes. Because of a lack of
information about the sorting technique presented in [10], its
reproduction is not possible. In the following of the paragraph
the sorting algorithm used in the SCL decoder is described.
In R1 nodes, a Chase-2 [31] algorithm is applied. The two
maximum absolute values of the LLRs have to be identified.
The way to do the minimum number of comparisons to
identify the 2 largest of n ≥ 2 elements was originally
described by Schreier in [32] and reported in [33]. The lower
stages of this algorithm can be parallelized thanks to SIMD
instructions in the way described in [34]. According to our
experimentations, Schreier’s algorithm is the most efficient
compared to parallelized Batcher’s merge exchange, partial
quick-sort or heap-sort implemented in the C++ standard
library in the case of R1 nodes. At the end, we chose not
Info. throughput (Mb/s)
B. Improving Cyclic Redundancy Checking
6
5
4
3
2
SSCLcpy SSCLptr
8-bit
16-bit
32-bit
1
0
27
28
29 210 211 212 213 214 215 216 217 218 219 220 221 222
Codeword size (N )
Fig. 8. Information throughput of the SSCL decoder depending on the
codeword size (N ) and the partial sums management. R = 1/2, L = 8.
to apply the SIMD implementation of the Schreier’s algorithm
because: 1) the speedup was negligible, 2) in 8-bit fixed-point,
only N ≤ 256 codewords can be considered.
Concerning path metrics, partial quick-sort appeared to yield
no gains in terms of throughput by comparison with the algorithm in [32], neither did heap-sort or parallelized Batcher’s
merge exchange. For a matter of consistency, only Schreier’s
algorithm is used in the proposed decoder, for both LLR
sorting in R1 and SPC nodes and for path metrics sorting.
The sorting of path metrics is applied to choose the paths to
be removed, kept or duplicated.
D. Partial Sum Memory Management
An SCL decoder can be seen as L replications of an SC
decoder. The first possible memory layout is the one given
in Figure 1. In this layout, the partial sums ŝ of each node is
stored in a dedicated array. Therefore, a memory of size 2N −1
bits is necessary in the SC decoder, or L × (2N − 1) bits in
the SCL decoder. This memory layout is described in [2] and
applied in previous software implementations [10], [11], [35].
A possible improvement is to change the memory layout to
reduce its footprint. Due to the order of operations in both SC
and SCL algorithms, the partial sums on a given layer are only
used once by the h function and can then be overwritten. Thus,
a dedicated memory allocation is not necessary at each layer
of the tree. The memory can be shared between the stages.
Therefore the memory footprint can be reduced from 2N − 1
to N in the SC decoder as shown in [36]. A reduction from
L × (2N − 1) to LN can be obtained in the SCL decoder.
In the case of the SCL algorithm, L paths have to be assigned
to L partial sum memory arrays. In [2], this assignment is
made with pointers. The advantage of pointers is that when a
path is duplicated, in the update paths() sub-routine of Alg. 1,
the partial sums are not copied. Actually, they can be shared
between paths thanks to the use of pointers. This method limits
the number of memory transactions. Unfortunately, it is not
possible to take advantage of the memory space reduction: the
partial sums have to be stored on L × (2N − 1) bits. There is
an alternative to this mechanism. If a logical path is statically
assigned to a memory array, no pointers are necessary at the
cost that partial sums must be copied when a path is duplicated
(only LN bits are required). This method is called SSCLcpy
whereas the former is called SSCLptr .
Our experiments have proved that the overhead of handling
pointers plus the extra memory space requirement cause the
SSCLcpy to be more efficient than the SSCLptr for short and
medium code lengths, as shown in Figure 8. The 32-bit version
uses floating-point LLRs, whereas 16-bit and 8-bit versions
are in fixed-point. Notice that in this work, each bit of the
partial sums is stored on an 8-bit, 16-bit or 32-bit number
accordingly to the LLR data type. The code rate R is equal
to 1/2. The throughput of the SSCLcpy version is higher for
N ≤ 8192 whereas the SSCLptr version is more efficient for
higher values of N . Although it does not appear in Figure 8,
experiments showed that the lower L is, the more efficient
SSCLcpy is compared to SSCLptr . Figure 8 also illustrates
the impact of the representation of partial sums. For very high
values of N , 8-bit fixed point representation takes advantage
of fewer cache misses. According to the results presented in
Figure 2, as the decoding performance improvements of the
SCL algorithm are not very significant compared to the SC
algorithm for long polar codes, SSCLcpy is the appropriate
solution in most practical cases.
In our decoder description, LLRs are managed with pointers,
as it is the case in other software implementations of the
literature [10], [11], [35]. We tried to remove the pointer
handling as for the partial sums, but it appeared that it was
not beneficial in any use case.
V. E XPERIMENTS AND M EASUREMENTS
Throughput and latency measurements are detailed in this
section. The proposed decoder implementation is compared
with the previous software decoders. Despite the additional
levels of genericity and flexibility, the proposed implementation is very competitive with its counterparts. Note that all the
results presented in the following can be reproduced with the
AFF3CT tool.
During our investigations, all the throughput and latency
measurements have been obtained on a single core of an
Intel i5-6600K CPU (Skylake architecture with AVX2 SIMD)
with a base clock frequency of 3.6 GHz and a maximum
turbo frequency of 3.9 GHz. The description has been compiled on Linux with the C++ GNU compiler (version 5.4.0)
and with the following options: -Ofast -march=native
-funroll-loops.
A. Fully Adaptive SCL
Being able to easily change the list size of the SCL decoders
enables the use of the FA-SSCL algorithm. With an unrolled
decoder as proposed in [10], the fully adaptive decoder would
imply to generate a fully unrolled decoder for each value
of the list depth. In our work, only one source code gives
the designer the possibility to run each variation of the SCL
decoders. FA-SSCL algorithm is the key to achieve the highest
possible throughput. As shown in Table III, with an 8-bit fixed
TABLE IV
T HROUGHPUT AND LATENCY COMPARISON WITH STATE - OF - THE - ART SCL
DECODERS . 32- BIT FLOATING - POINT REPRESENTATION .
C ODE (2048,1723), L = 32, 32- BIT CRC.
Target
Decoder
i7-4790K
CA-SCL [35]
CA-SCL [11]
CA-SSCL [11]
PA-SSCL [11]
CA-SCL [10]
CA-SSCL [10]
PA-SSCL [10]
This CA-SCL
This CA-SSCL
This PA-SSCL
This FA-SSCL
This CA-SCL
This CA-SSCL
This PA-SSCL
This FA-SSCL
i7-2600
i7-2600
i7-2600
i5-6600K
Lworst
(µs)
1572
23000
3300
≈ 3300
2294
433
≈ 433
4819
770
847
1602
3635
577
635
1201
3.5 dB
1.10
0.07
0.52
0.9
0.76
4.0
8.6
0.37
2.3
5.5
19.4
0.48
3.0
7.6
26.1
T i (Mb/s)
4.0 dB
1.10
0.07
0.52
4.90
0.76
4.0
33.0
0.37
2.3
31.1
149.0
0.48
3.0
42.1
207.8
4.5 dB
1.10
0.07
0.52
54.0
0.76
4.0
196.0
0.37
2.3
168.4
244.3
0.48
3.0
237.6
345.5
point representation of the decoder inner values, the achieved
throughput in the case of the (2048,1723) polar code is about
425 Mb/s on the i5-6600K for an Eb /N0 value of 4.5 dB.
It corresponds to a FER of 5 × 10−8 . This throughput is
almost 2 times higher than the throughput of the PA-SSCL
algorithm. The highest throughput increase from PA-SSCL
to FA-SSCL, of about 380%, is in the domain where the
FER is between 10−3 and 10−5 . It is the targeted domain for
wireless communications like LTE or 5G. In these conditions,
the throughput of FA-SSCL algorithm is about 227 Mb/s
compared to 42 Mb/s for the PA-SSCL algorithm.
In Adaptive SCL algorithms, the worst case latency is the sum
of the latency of each triggered algorithm. In the case of PASSCL with Lmax = 32, it is just the sum of the latency of
the SC algorithm, plus the latency of the SCL algorithm with
L = 32. In the case of the FA-SSCL algorithm, it is the sum of
the decoding latency of the SC algorithm and all the decoding
latencies of the SCL algorithm for L = 2, 4, 8, 16, 32. This is
the reason why the worst latency of the PA-SSCL algorithm is
lower while the average latency and consequently the average
throughput is better with the FA-SSCL algorithm.
B. Comparison With State-Of-The-Art SCL Decoders.
The throughput and latency of the proposed decoder compared
to other reported implementations are detailed in Table IV.
For all the decoders, all the available tree pruning optimizations are applied excluding the SPC4+ nodes because of the
performance degradation. Each decoder is based on a 32-bit
floating-point representation. The polar code parameters are
N = 2048, K = 1723 and the 32-bit GZip CRC is used. The
list size is L = 32.
The latency given in Table IV is the worst case latency
and the throughput is the average information throughput.
The first version, CA-SCL, is the implementation of the
CA-SCL algorithm without any tree pruning. As mentioned
before the throughput of the proposed CA-SSCL decoder (2.3
Mb/s) is only halved compared to the specific unrolled CASSCL decoder described in [10] (4.0 Mb/s). The proposed
CA-SSCL decoder is approximately 4 times faster than the
generic implementation in [11] (0.52 Mb/s) and 2 times faster
than the CA-SCL implementation in [35] (1.1 Mb/s) thanks
to the implementation improvements detailed in Section IV.
Furthermore, the proposed decoder exhibits a much deeper
level of genericity and flexibility than the ones proposed in
[11], [35]. Indeed, the following features were not enabled: the
customization of the tree pruning, the 8-bit and 16-bit fixedpoint representations of the LLRs, the puncturing patterns and
the FA-SSCL algorithm.
When implemented on the same target (i7-2600), the proposed
PA-SSCL is competitive with the unrolled PA-SSCL in [10],
being only two times slower. This can be explained by
the improvements concerning the CRC that are described in
Section IV-B, especially the information bits extraction in the
SC decoder. Finally, as mentioned before, the throughput of
the proposed FA-SSCL significantly outperforms all the other
SCL decoders (up to 345.5 Mb/s at 4.5 dB in 32-bit floatingpoint).
VI. C ONCLUSION
The trend towards Cloud RAN networks in the context of mobile communications and the upcoming 5G standard motivated
an investigation of the possibility of implementing generic and
flexible software polar decoders. Means of implementing such
flexible decoders are reported in this paper. A single source
code is necessary to address any code lengths, code rates,
frozen bits sets, puncturing patterns and cyclic redundancy
check polynomials.
This genericity is obtained without sacrificing the throughput
of the decoders, thanks to the possibility to adjust the decoding
algorithm and the possibility to apply multiple implementation
related and algorithmic optimizations. In fact, to the best of our
knowledge, the proposed adaptive SCL decoder is the fastest
to be found in the literature, with a throughput of 425 Mb/s
on a single core for N = 2048 and K = 1723 at 4.5 dB.
Being included in the open-source AFF3CT tool, all the results
presented in this paper can be easily reproduced. Moreover,
this tool can be used for polar codes exploration, which is of
interest for the definition of digital communication standards
and for practical implementations in an SDR environment.
ACKNOWLEDGMENTS
The authors would like to thank the Natural Sciences and
Engineering Research Council of Canada, Prompt, and Huawei
Technologies Canada Co. Ltd. for financial support to this
project. This work was also supported by a grant overseen by
the French National Research Agency (ANR), ANR-15-CE250006-01.
R EFERENCES
[1] E. Arikan, “Channel polarization: a method for constructing capacityachieving codes for symmetric binary-input memoryless channels,” IEEE
Transactions on Information Theory (TIT), vol. 55, no. 7, pp. 3051–3073,
2009.
[2] I. Tal and A. Vardy, “List decoding of polar codes,” in Proceedings of
the IEEE International Symposium on Information Theory (ISIT), 2011,
pp. 1–5.
[3] “3GPP TSG RAN WG1 meeting #87, Chairmans notes of agenda item
7.1.5 Channel coding and modulation,” 2016.
[4] D. Wübben, P. Rost, J. S. Bartelt, M. Lalam, V. Savin, M. Gorgoglione,
A. Dekorsy, and G. Fettweis, “Benefits and impact of cloud computing
on 5G signal processing: flexible centralization through cloud-ran,” IEEE
Signal Processing Magazine, vol. 31, no. 6, pp. 35–44, 2014.
[5] P. Rost, C. J. Bernardos, A. De Domenico, M. Di Girolamo, M. Lalam,
A. Maeder, D. Sabella, and D. Wübben, “Cloud technologies for flexible
5G radio access networks,” IEEE Communications Magazine, vol. 52,
no. 5, pp. 68–76, 2014.
[6] Ericsson, “Cloud ran - the benefits of cirtualization, centralisation
and coordination,” Tech. Rep., 2015. [Online]. Available: https://www.
ericsson.com/assets/local/publications/white-papers/wp-cloud-ran.pdf
[7] Huawei, “5G: A technology vision,” Tech. Rep., 2013. [Online].
Available: https://www.huawei.com/ilink/en/download/HW 314849
[8] V. Q. Rodriguez and F. Guillemin, “Towards the deployment of a fully
centralized cloud-ran architecture,” in Proceedings of the IEEE International Wireless Communications and Mobile Computing Conference
(IWCMC), 2017, pp. 1055–1060.
[9] N. Nikaein, “Processing radio access network functions in the cloud:
critical issues and modeling,” in Proceedings of the ACM International
Workshop on Mobile Cloud Computing and Services (MCS), 2015, pp.
36–43.
[10] G. Sarkis, P. Giard, A. Vardy, C. Thibeault, and W. J. Gross, “Fast
list decoders for polar codes,” IEEE Journal on Selected Areas in
Communications (JSAC), vol. 34, no. 2, pp. 318–328, 2016.
[11] ——, “Increasing the speed of polar list decoders,” in Proceedings of
the IEEE International Workshop on Signal Processing Systems (SiPS),
2014, pp. 1–6.
[12] A. Cassagne, B. Le Gal, C. Leroux, O. Aumage, and D. Barthou,
“An efficient, portable and generic library for successive cancellation
decoding of polar codes,” in Proceedings of the Springer International
Workshop on Languages and Compilers for Parallel Computing (LCPC),
2015, pp. 303–317.
[13] A. Cassagne, O. Aumage, C. Leroux, D. Barthou, and B. Le Gal,
“Energy consumption analysis of software polar decoders on low power
processors,” in Proceedings of the IEEE European Signal Processing
Conference (EUSIPCO), 2016, pp. 642–646.
[14] I. Tal and A. Vardy, “How to construct polar codes,” IEEE Transactions
on Information Theory (TIT), vol. 59, no. 10, pp. 6562–6582, Oct 2013.
[15] P. Trifonov, “Efficient design and decoding of polar codes,” IEEE
Transactions on Communications, vol. 60, no. 11, pp. 3221–3227,
November 2012.
[16] B. Le Gal, C. Leroux, and C. Jégo, “Multi-Gb/s software decoding of
polar codes,” IEEE Transactions on Signal Processing (TSP), vol. 63,
no. 2, pp. 349–359, 2015.
[17] G. Sarkis, P. Giard, A. Vardy, C. Thibeault, and W. J. Gross, “Fast polar
decoders: algorithm and implementation,” IEEE Journal on Selected
Areas in Communications (JSAC), vol. 32, no. 5, pp. 946–957, 2014.
[18] A. Balatsoukas-Stimming, M. B. Parizi, and A. Burg, “LLR-based
successive cancellation list decoding of polar codes,” IEEE Transactions
on Signal Processing (TSP), vol. 63, no. 19, pp. 5165–5179, 2015.
[19] A. Alamdar-Yazdi and F. Kschischang, “A simplified successivecancellation decoder for polar codes,” IEEE Communications Letters,
vol. 15, no. 12, pp. 1378–1380, 2011.
[20] B. Li, H. Shen, and D. Tse, “An adaptive successive cancellation
list decoder for polar codes with cyclic redundancy check,” IEEE
Communications Letters, vol. 16, no. 12, pp. 2044–2047, December
2012.
[21] M. Matsumoto and T. Nishimura, “Mersenne twister: a 623dimensionally equidistributed uniform pseudo-random number generator,” ACM Transactions on Modeling and Computer Simulation
(TOMACS), vol. 8, no. 1, pp. 3–30, 1998.
[22] G. E. P. Box, M. E. Muller et al., “A note on the generation of random
normal deviates,” The annals of mathematical statistics, vol. 29, no. 2,
pp. 610–611, 1958.
[23] E. Dahlman, S. Parkvall, and J. Skold, 4G: LTE/LTE-advanced for
mobile broadband. Academic press, 2013.
[24] R. Wang and R. Liu, “A novel puncturing scheme for polar codes,” IEEE
Communications Letters, vol. 18, no. 12, pp. 2081–2084, Dec 2014.
[25] K. Niu, K. Chen, and J. R. Lin, “Beyond turbo codes: rate-compatible
punctured polar codes,” in Proceedings of the IEEE International
Conference on Communications (ICC), June 2013, pp. 3423–3427.
[26] V. Miloslavskaya, “Shortened polar codes,” IEEE Transactions on Information Theory (TIT), vol. 61, no. 9, pp. 4852–4865, Sept 2015.
[27] “Cyclic redundancy check,” https://en.wikipedia.org/wiki/Cyclic
redundancy check, accessed: 2017-03-13.
[28] Q. Zhang, A. Liu, X. Pan, and K. Pan, “CRC code design for list
decoding of polar codes,” IEEE Communications Letters, vol. 21, no. 6,
pp. 1229–1232, 2017.
[29] J. Lin, C. Xiong, and Z. Yan, “A reduced latency list decoding algorithm
for polar codes,” in Proceedings of the IEEE International Workshop on
Signal Processing Systems (SiPS), 2014, pp. 1–6.
[30] P. Giard, G. Sarkis, C. Leroux, C. Thibeault, and W. J. Gross, “Lowlatency software polar decoders,” Springer Journal of Signal Processing
Systems (JSPS), pp. 31–53, Jul 2016.
[31] D. Chase, “Class of algorithms for decoding block codes with channel
measurement information,” IEEE Transactions on Information Theory
(TIT), vol. 18, no. 1, pp. 170–182, 1972.
[32] J. Schreier, “On tournament elimination systems,” Mathesis Polska,
vol. 7, pp. 154–160, 1932.
[33] D. Knuth, The art of computer programming. Addison-Wesley, 1973,
no. 3.
[34] T. Furtak, J. N. Amaral, and R. Niewiadomski, “Using SIMD registers
and instructions to enable instruction-level parallelism in sorting algorithms,” in Proceedings of the ACM Symposium on Parallel Algorithms
and Architectures, 2007, pp. 348–357.
[35] Y. Shen, C. Zhang, J. Yang, S. Zhang, and X. You, “Low-latency
software successive cancellation list polar decoder using stage-located
copy,” in Proceedings of the IEEE International Conference on Digital
Signal Processing (DSP), 2016.
[36] C. Leroux, A. J. Raymond, G. Sarkis, and W. J. Gross, “A semi-parallel
successive-cancellation decoder for polar codes,” IEEE Transactions on
Signal Processing (TSP), vol. 61, no. 2, pp. 289–299, 2013.
| 7cs.IT
|
arXiv:1712.06281v2 [math.OC] 22 Dec 2017
Model Reduction in Chemical Reaction Networks: A Data-Driven
Sparse-Learning Approach
Omar A. Khalila,∗, Farshad Harirchia,∗ , Doohyun Kimc , Sijia Liua , Paolo Elvatic ,
Angela Violic,d , Alfred O. Heroa,b
a
Department of Electrical Engineering and Computer Science,
b
Department of Statistics,
c
Department of Mechanical Engineering,
d
Departments of Chemical Engineering, Biomedical Engineering,
Macromolecular Science and Engineering, Biophysics Program,
University of Michigan, Ann Arbor, MI 48109 †
{oakhalil,harirchi,gogum,lsjxjtu,elvati,avioli,hero}@umich.edu
December 27, 2017
Abstract
The reduction of large kinetic mechanisms is a crucial step for fluid dynamics simulations of combustion systems. In this paper, we introduce a novel approach for mechanism reduction that presents
unique features. We propose an unbiased reaction-based method that exploits an optimization-based
sparse-learning approach to identify the set of most influential reactions in a chemical reaction network.
The problem is first formulated as a mixed-integer linear program, and then a relaxation method is
leveraged to reduce its computational complexity. Not only this method calculates the minimal set of
reactions subject to the user-specified error tolerance bounds, but it also incorporates a bound on the
propagation of error over a time horizon caused by reducing the mechanism. The method is unbiased
toward the optimization of any characteristic of the system, such as ignition delay, since it is assembled
based on the identification of a reduced mechanism that fits the species concentrations and reaction
rate generated by the full mechanisms. Qualitative and quantitative validations of the sparse encoding
approach demonstrate that the reduced model captures important network structural properties.
1
Introduction
Combustion is a ubiquitous process with applications that vary widely from heating and transportation
to mass production of metallic and ceramic nanoparticles. Despite its importance and several decades of
intense scientific work, the chemical reaction networks that drive the processes are still elusive; they involve
mechanisms that can span a wide range of length and time scales, highly reactive species, and are controlled
by the delicate interplay between several chemical and physical phenomena. For all these reasons modeling
has been extensively used to interpret and predict data on combustion. However, despite the massive increase
in computational resources, we are still far from being able to simulate but the simplest systems once detailed
fluid dynamic description is required. Specifically, using a detailed chemical description for the combustion of
large molecules is still unrealistic, and instead reduced versions of the chemical kinetics have been developed.
Various reduction methods that exploit different mathematical model reduction algorithms have been
proposed, including Principal Component Analysis (PCA) [1], Intrinsic Low Dimensional Manifolds [2, 3],
Computational Singular Perturbation (CSP) method [4, 5], isomer lumping [6], and an optimization-based
method [7]. Also, the Directed Relation Graph (DRG) method and its derivatives [8, 9] have shown their
∗ Similar
† This
contributions from these authors.
research has been funded by the US Army Research Office grants W911NF-15-1-0241 and W911NF-14-1-0359.
1
capabilities to generate a skeletal version of very large detailed chemical mechanisms in a computationally efficient manner. However, since DRG-type approaches are species-based methods (they identify unimportant
species and remove all reactions associated with them), they are not as effective as reaction-based methods
for reducing chemical reaction networks. On the other hand, reaction-based mechanism reduction approaches
(they identify unimportant reactions and remove them) are able to calculate more compact reduced mechanisms. Many reaction-based mechanism reduction approaches such as the ones based on PCA, CSP and
brute-force sensitivity analysis require generating data for many perturbations of each rate parameter, which
is computationally demanding [10, 11].
Optimization-based reaction elimination approaches also received some attention from the combustion
community. For example, a nonlinear mixed-integer programming approach was proposed in [12, 13]. Because
of the non-convex nature of these problems, they are computationally intractable, and their optimality is
not guaranteed. In [7], an approach is proposed to obtain a group of reduced mechanisms that are only valid
around one reference point (i.e., temperature, pressure, molar concentrations). This enables the approach to
be formulated as a mixed-integer linear programming problem, which is achieved by linearizing the nonlinear
ODE around a given reference point. This approach delivers a group of different reduced mechanisms that are
valid locally around a reference point. Furthermore, it requires Jacobian evaluation that is computationally
demanding.
In this paper, we propose a novel approach based on sparse-learning method to model reduction in
chemical reaction networks. Our data-driven approach is part of the class of reaction-based methods, and
hence carries the advantages associated with this class that are illustrated in figure 1. On the other hand,
it does not require extensive number of simulations to capture sensitivity to different parameters. We pose
the reaction-elimination problem as a mixed integer linear program, which can be solved to optimality in
a relatively efficient manner, thanks to state-of-the-art MIP solvers [14]. As opposed to [7], the proposed
approach aims to find a global reduced mechanism that is valid in the whole dynamical state space range.
Our optimization-based approach identifies and removes unimportant reactions, by taking advantage of
simulation data from the chemical process, and finds a global reduced mechanism with the minimal set of
reactions, for which, the fitting error induced by removing the rest of the reactions remains in a user-specified
bounded set. That is, the method finds the smallest set of reactions that still satisfy the bound condition
on the error. Furthermore, it constraints the propagation of error caused by removing reactions in time. As
the method is constructed based on finding a reduced mechanism that fits the species concentrations and
reaction rates generated from the full mechanism, it is unbiased toward optimizing any specific characteristics
such as ignition delay or temperature profile. In addition, the proposed method only requires tuning one
parameter that has a physical meaning, which simply indicates the upper-bound on the tolerance of error
caused by mechanism reduction. A preliminary version of this paper is presented in [15].
2
Modeling Framework
Chemical systems involve multiple species and reactions that describe how a set of chemical species interact
through reactions and how their species concentrations evolve over time. The dynamics of chemical reactions
are commonly modeled using mass-action kinetic equations [16, 17, 18, 19, 20]. Under Euler discretization
method [21], the continuous dynamics can then be discretized as the following linear system
Xt+1 = Xt + Mrt ∆t + ω t , t = 1, 2, . . . , T − 1,
(1)
where Xt is an Ns × 1 vector of molar concentrations of all species at time t, Ns is the number of species,
Nr is the number of reactions, M is an Ns × Nr matrix that contains the stoichiometric coefficients of all
reactions, ∆t is the sampling time used for discretization, ω t denotes the discretization error and process
noise. rt is an Nr × 1 vector of the reaction rates of all reactions at time t, which is a polynomial function
of concentration values as follows:
Nr
Y
rt (i) = di
Xt (j)αi,j ,
(2)
j=1
where di is time varying reaction constant for the reaction i, Xt (j) denotes the j th component of Xt , and
αi,j is the stoichiometric reaction coefficient relating reaction i to species j. The values of αi,j are contained
in M.
2
Full mechanism
A
A
A
D
B
D
B
C
G
F
H
I
D
G
Reduction by
species-based
method
F
H
J
B
E
C
C
G
Reduction by
this work
F
H
I
J
I
J
Figure 1: Schematic representation of a chemical reaction network and its reduction. The circular nodes
are chemical species and the arrows are chemical reactions with the thickness representing their relative
importance within the system. This illustration highlights that the reduction method proposed in this work
can generate smaller reaction network compared to the one by typical species-based method by eliminating
all less-influential reactions.
Given the discrete-time mass-action kinetic equations in Eq. (1), the focus of this paper is to infer a
reduced mechanism, in terms of a reduced number of reactions, based only on observations such as species
concentrations and reaction rates of the chemical systems.
3
Mechanism reduction via reaction selection
In this section, we propose a data-driven sparse-learning approach towards identifying a reduced set of
reactions, which approximately replicates the behavior of the detailed mechanism.
For each of Nr reactions, we introduce a binary selection variable wi,t ∈ {0, 1} to encode whether or
not the ith reaction is selected at time t. If reaction i is chosen for the reduced mechanism at time t, then
wi,t = 1; otherwise, wi,t = 0. With the aid of the selection variables wt = [w1,t , w2,t , . . . , wNr ,t ]| , we
formally state the problem of mechanism reduction.
Problem Statement Given the measurements of species concentrations Xt and reaction rates rt at all
times, determine the smallest subset of reactions (in terms of w) such that the estimation error induced by
the reduced mechanism remains in a user-specified tolerance range () at all time instances.
3.1
Sparse-Learning Mechanism Reduction (SLMR)
Let us first define the error induced by the reduced mechanism characterized by wt at time t on molar
concentration of ith species as follows:
Ei,t (wt ) = |Xt+1 (i) − Xt (i) − Mi (wt
where Mi denotes the ith row of matrix M and
rt )∆t|,
(3)
denotes the element-wise product.
Error tolerance on individual concentrations We can enforce constraints on the error in the change
of individual species concentrations for all time instances, i.e.,
Ei,t (wt ) ≤ Nt (i), ∀t, ∀i ∈ {1, . . . , Ns },
(4)
where Nt is the normalization factor at time t, which is defined to be the summation of absolute changes
of all concentrations at time t as described in Eq. (5), and is a tuning parameter that indicates the error
tolerance that is acceptable in terms of the percentage of Nt , i.e., = 0.05 indicates 5% error tolerance.
3
Nt (i) = |Mi |rt ∆t , i ∈ {1, . . . , Ns }.
(5)
Error propagation Constraint Eq. (6) is added to avoid the propagation of error through time. The
size of time horizon can be chosen based on the speed of changes in concentrations. We indicate such time
horizon with H. For the concentration change of each species i in time horizon [t, t + H − 1], we have that
|Xt+H−1 (i) − Xt (i) −
t+H−2
X
Mi wk
rk ∆k | ≤ β
t+H−2
X
k=t
Nk (i),
(6)
k=t
where β is a tuning parameter that indicates the tightness of the bound on propagated error, and t =
0, H, 2H, . . ..
Incremental reaction inclusion At each time instance t, we seek the minimum number of reactions that
are required to be added to the previously selected reactions up to time t such that Eq. (4) and Eq. (6) are
satisfied, i.e.,
wt+1 ≥ wt .
(7)
Sparse-Learning Mechanism Reduction Now that the constraints that we consider for the optimization problem are described, we can formulate our data-driven sparse-learning mechanism reduction approach.
For each time batch of size H, we solve the following mixed-integer linear programming problem:
{wk∗ }t+H−1
= minimize
k=t
{wk }t+H−1
k=t
Nr
t+H−1
X X
k=t
wi,k
(PSLM R )
i=1
subject to wi,k ∈ {0, 1}, i = 1, 2, ..., Nr ,
∀k ∈ {t, t + 1, . . . , t + H − 1}
Eqs. (4),(6),(7) hold.
where {wk }t+H−1
are the optimization variables. Note that solving problem (PSLM R ) delivers the minimum
k=t
number of reactions that are required to be added to the previously selected reactions, i.e., incremental
reaction inclusion Eq. (7) such that the error tolerance constraint on individual concentrations Eq. (4)
and error propagation Eq. (6) are satisfied. The objective function minimizes the total number of selected
reactions in the time horizon [t, t + H − 1].
Effect of different initial conditions In order to have a global reduced mechanism that mimics the
dynamical behavior of full mechanism, we need to use data from simulations over a sparse set of initial
conditions that encompass the whole range of temperature T , equivalence ratio φ and pressure p. We refer
to each initial condition as a point in the parameter space, namely θ = [T, φ, p].
Reduced mechanism for entire simulation time The next step is to find the reduced mechanism
not just for the time horizon [t, t + H − 1], but for the entire simulation time [0, tf ], where tf is the time
that simulation ends. In order to do this, we solve problem (PSLM R ) for time intervals [0, H − 1], [H, 2H −
1], ..., [KH, tf ], where KH is the largest integer multiple of H that is smaller or equal to tf , as described in
Algorithm 1.
Global mechanism Assume that data is available for a sparse set of initial conditions that is denoted by
Θ. In order to obtain a global reduced mechanism we calculate the set union of all selected reactions for all
the conditions in Θ. In other words:
[
w∗ =
R(w(θ)),
(8)
θ∈Θ
where R(w(θ)) denotes the set of reactions corresponding to the output of Algorithm 1 for initial condition
θ.
4
Algorithm 1 Calculating w(θ)
t
t
t
f
f
f
Input: {Xt }t=0
, {rt }t=0
, {∆t }t=1
, , β, θ = [φ, T, p].
Initialize: t = 0, w−1 = 0.
1. While (t ≤ tf )
• W = min{H, tf − t + 1}
• set the time horizon to [t, t + W − 1].
∗
• solve problem (PSLM R ) to obtain wt+W
−1
• t + W → t.
∗
• wt−1 = wt−1
.
End while
2. Return wt∗f as w(θ).
3.2
Computational Complexity
Problem (PSLM R ) is a standard mixed-integer linear programming (MILP) problem, and thus can be efficiently solved using state-of-the-art MILP solvers such as Gurobi [14] and CPLEX [22]. Even though these
solvers can solve problems with large number of integer variables in a relatively fast manner by employing
advanced branch and bounding algorithms, the worst-case complexity of MILP is exponential. In order to
further reduce the complexity of our data-driven sparse-learning approach, we can utilize convex relaxation
methods [23] by replacing the Boolean variable constraint with its convex compartment 0 ≤ wi,t ≤ 1 for all
i. The latter yields a linear programming problem, which can be solved efficiently in polynomial time, and
therefore makes our approach promising for extremely large-scale chemical reaction networks at the cost of
losing optimality guarantees.
4
Application
We applied the reduction algorithm presented in the previous section to the chemical mechanism of H2
oxidation by Hong et al. [24]. We selected this mechanism for two reasons: first is a very small mechanism
that is suitable for testing of different approaches and as such optimal to develop new methodologies; second,
due to the small number of reactions and species, is very difficult to reduce without affecting its accuracy. We
want to stress, however, that the mechanism reduction of H2 is intended to show the algorithm capabilities
and not a real world application.
We first performed 48 simulations using the full mechanism in a H2 /air 0-D homogeneous reactor for
thermodynamic regimes that relevant to practical combustion devices by using using Chemkin [25]. The
species concentrations and reaction rates generated for initial conditions encompassing 5 − 20 atm, 800 −
1100 K, and equivalence ratios between 0.5 − 2 were used to generate the reduced mechanism. We first fix
the time horizon for error propagation to be at H = 10. Then, by setting the tuning parameters β to 100
and , i.e., the upper bound on the fitting error tolerance, to 0.02, 0.05, and 0.1 results in three reduced
mechanisms with 47, 40 and 37 reactions (forward and reverse counted separately). Compared to the full
mechanism, which includes 58 reactions excluding the reactions associated with Ar, that is a reduction of
about 19%, 31%, and 36%, respectively.
Figure 2 shows, in the form of a directed bipartite graph, the comparison between the full mechanism
and the 37-reactions reduced mechanism. In this representation, species are connected to the reactions in
which they appear as reactants by red dashed lines, while reactions are connected to their products by blue
dotted lines. Finally, gray lines show the reactions that were removed by the mechanism reduction.
To test the quality of the reduced mechanisms , we compared calculated ignition characteristics and
the time evolutions of species concentrations in a range of conditions slightly wider than the one used to
generate them (i.e., 5 − 20 atm, 700 − 1300 K, equivalence ratio of 0.5 − 2) for the same homogeneous
5
O2
HO
2
HO2
HO
2
O
H
O
H2
->
2
H2
O
H
5:
H
O
R3
0
R5
R3
6:
HO
7:
R3
H
:2
->
O
HH
8: H
O2
-
2O
>H
2 ->
2O
2
HH
-> O
HO2
:O
R39
R3
2 H -> H
O2 H2
2 -> O
OH O
R40:
R42: HO2 H2 -> H2O2 H
R41: H2O
R43: H2O2 H -> H2O OH
H2
2H
->
2
2H
9:
H2
->
R4
2H
8:
4
2H
R
O
2 ->
2O2
7: H
-> H
R4
O2
HH
O
:
O2
R46
OH H
O ->
H2O2
R45:
2O2 H
OH -> H
R44: H2O
2
2H
->
H2
2
HO
:2
34
1
R5
R
H
->
H
:
33
:H
2
R5
R
H
2:
->
2H
2:
2
H
H2
->
2-
H
>2
:H
R5
3
R3
H
H
H
H2
O
>H
H2
2-
H
OH
->
2H
4:
1:
R5
R3
:H
R30
: 2O
H2
-> O
2
R55
: O2
R56
R29
H2
-> O
OH
2 ->
:OH
R57: O
R28: H
H
HO
OH ->
O H2
-> 2O
H -> O
R58: OH ->
H
R27: O H2 ->
OH
R26: H2O O -> 2OH
R25: 2OH ->
H
OH
H
R24: O
O
H2O O
HHH
H2
:H
R2
1:
O
R2
2
HO
Species
9:
R1
8:
O2
2
H2
O
H2
O2
2H
O
7:
2
O
2
->
> OH
OH
-> H
2H2O
H H2
O
2O
O-
>H
H2
O2
-
->
OH
>2
O2
HO
2
H2
O2
2H
O
2
O
2
H2
OH
2
O
->
2
O2
2
H2O2
H2O
2O
HH
>O
2O
HO
>H
O-
O2
HO2 H2O
-> OH
-> H
H2O
H2O2 ->
2
H2O
H2
2-
O2
OH
HO2
HO
O2
OH
->
5:
O
H2
R1
6:
2H
R1
4: H
:
R13
R1
R12:
-> H O2
R10: 2OH -> H2O2
R11: OH
2
O
->
O
O
H
O2
R9: H2O2 -> 2OH
H
R8: HO2
->
2
HO
R7: H
O2 ->
HO2
R6:
HO2
-> H
R5
O2
:
H
O2
-> H
O2
:
->
O
2
HO
2
H
O2
->
R4
:H
O
H
R3
:O
R1
R2
:H
H2
0:
R1
O2
Reaction
H2O
-
R22
OH
removed edges
2O ->
R23
:2
O2
R1
Reaction->Species
O
->
2O
H2
Species->Reaction
2O
->
2
HO
Figure 2: The directed bipartite graph for the reduced mechanism with 37 reactions ( = 0.1). Red dashed
lines indicate the edges from species to reactions, blue dotted lines represent the edges from reactions to
species, and light grey solid lines show the removed edges from the detailed mechanism.
reactor. Figure 3 shows the ignition delay times, relative deviations in ignition delay time from the full
mechanism, and maximum temperature increase at 10 atm, and equivalence ratios of 0.5 and 1. All three
reduced mechanisms are in excellent agreement with the full mechanism, with, as expected, larger reduced
mechanisms showing better agreements. The relative deviation in the ignition delay times are, however,
even in the worst case scenario within 1% under all the tested conditions. Also the maximum temperature
increase (bottom panels of Fig. 3) is well-maintained in the reduced mechanisms. This agreements are not
limited to 10 atm and φ =0.5 and 1, as very similar observations can be made in all the condition tested
(not shown).
Figure 4 shows the time evolutions of selected species for the simulations at 1000 K, 10 atm, equivalence
ratios of 0.5 and 1. All reduced mechanisms generated very similar H2 O and H profiles throughout the whole
ignition event compared to those from the full mechanism. However, OH profiles show that the correct
concentrations after ignition is not completely matched by the 40- and 37-reactions mechanism. Considering
the small size of the full mechanism this is not entirely unexpected and is exacerbated by the equilibrium
reached in the concentrations after ignition. Overall, however, the above comparisons show that the reduced
mechanisms generated by our new reduction approach are effective in reproducing key combustion properties
such as ignition characteristics and heat release, as well as species profiles during pre-ignition and ignition.
6
Tmax - T init (K)
10atm, phi=1
10 -1
10 -3
10 -5
ID deviation (%)
Ignition Delay (s)
10atm, phi=0.5
10 1
1
0
-1
2500
full mechanism
47-reactions
40-reactions
37-reactions
2000
1500
0.8
0.9
1
1.1
1.2
1.3
1.4
0.8
1000/T (1/K)
0.9
1
1.1
1.2
1.3
1.4
1000/T (1/K)
Figure 3: Computed ignition delay times (top), percent deviations in ignition delay from the full mechanism
(middle), and maximum temperature increase (bottom) for 10 atm, equivalence ratios of 0.5 and 1 cases.
OH Mole Frac (-)
H Mole Frac (-)
H 2O Mole Frac (-)
10atm, phi=0.5, T init=1000K
10atm, phi=1, T init=1000K
0.2
0.3
0.2
0.1
0.1
0
0
0.06
full mechanism
47-reactions
40-reactions
37-reactions
0.02
0.01
0.04
0.02
0
0
0.02
0.03
0.02
0.01
0.01
0
0
0
0.005
0.01
0
Time (sec)
0.004
0.008
Time (sec)
Figure 4: Time evolutions of H2 O, H, and OH in mole fractions from homogeneous reactor simulations with
initial conditions of 1000 K, 10 atm, and equivalence ratios of 0.5 and 1.
5
Conclusions
In this work, we present a new method to reduce chemical reaction network by employing data-driven sparse
learning approach to obtain a reduced mechanism that is valid for all process conditions (temperature,
pressure and equivalence ration) and all ranges of species molar concentrations. In addition, our method
incorporates a user-specified bound on the propagation of error cause by eliminating reactions. That is, the
proposed approach ensures that the discrepancy between the original and the reduced mechanisms remains
in the user-specified acceptable range in all time instances and in all time horizons of specified length. The
reduction is performed without introducing any bias towards specific properties or species therefore providing
a mechanism with optimal performance, which is independent of the chemical reaction network that is being
considered. Finally, the approach has a low computational cost and scales gracefully with the size of the full
mechanism, making it a viable option for the reduction of large mechanisms.
The performance of the method is evaluated on a H2 mechanism, which due to its small initial size was
expected to be difficult to reduce. Surprisingly, however, our method was able to reduced the number of
7
reactions by more than one third without any notable discrepancy in ignition delay or heat release.
References
[1] S. Vajda, P. Valko, and T. Turányi. Principal component analysis of kinetic models. International
Journal of Chemical Kinetics, 17(1):55–81, 1985.
[2] U. Maas and S. B. Pope. Implementation of simplified chemical kinetics based on intrinsic lowdimensional manifolds. Symposium (International) on Combustion, 24(1):103–112, 1992.
[3] U. Maas and S. B. Pope. Simplifying chemical kinetics: Intrinsic low-dimensional manifolds in composition space. Combustion and Flame, 88(3):239–264, 1992.
[4] S. H. Lam and D. A. Goussis. The CSP method for simplifying kinetics. International Journal of
Chemical Kinetics, 26(4):461–486, 1994.
[5] S. H. Lam. Using CSP to Understand Complex Chemical Kinetics. Combustion Science and Technology,
89(5-6):375–404, 1993.
[6] Syed Sayeed Ahmed, Fabian Mauß, Gladys Moréac, and Thomas Zeuch. A comprehensive and compact
n-heptane oxidation model derived using chemical lumping. Physical Chemistry Chemical Physics,
9(9):1107–1126, 2007.
[7] Binita Bhattacharjee, Douglas A. Schwer, Paul I. Barton, and William H. Green. Optimally-reduced
kinetic models: Reaction elimination in large-scale kinetic mechanisms. Combustion and Flame,
135(3):191–208, 2003.
[8] Tianfeng Lu and Chung K. Law. A directed relation graph method for mechanism reduction. Proceedings
of the Combustion Institute, 30, 2005.
[9] P Pepiotdesjardins and H Pitsch. An efficient error-propagation-based reduction method for large
chemical kinetic mechanisms. Combustion and Flame, 154(1-2):67–81, 2008.
[10] T Lu and C Law. Strategies for mechanism reduction for large hydrocarbons: N-heptane. Combustion
and Flame, 154(1-2):153–163, 2008.
[11] Kyle E. Niemeyer, Chih-Jen Sung, and Mandhapati P. Raju. Skeletal mechanism generation for surrogate fuels using directed relation graph with error propagation and sensitivity analysis. Combustion
and Flame, 157(9):1760–1770, 2010.
[12] Linda Petzold and Wenjie Zhu. Model reduction for chemical kinetics: An optimization approach.
AIChE journal, 45(4):869–886, 1999.
[13] Keith Edwards, TF Edgar, and VI Manousiouthakis. Reaction mechanism simplification using mixedinteger nonlinear programming. Computers & Chemical Engineering, 24(1):67–79, 2000.
[14] Inc. Gurobi Optimization. Gurobi optimizer reference manual, 2016.
[15] Farshad Harirchi, Omar A. Khalil, Sijia Liu, Paolo Elvati, Angela Violi, and Alfred O. Hero. A datadriven sparse-learning approach to model reduction in chemical reaction networks. In Proceedings of
NIPS 2017 Workshop on Advances in Modeling and Learning Interactions from Complex Data, 2017.
[16] Martin Feinberg. Chemical reaction network structure and the stability of complex isothermal reactors–
i. the deficiency zero and deficiency one theorems. Chemical Engineering Science, 42(10):2229–2268,
1987.
[17] Jeremy Gunawardena. Chemical reaction network theory for in-silico biologists. Notes available for
download at http: // vcp. med. harvard. edu/ papers/ crnt. pdf , 2003.
[18] Daniel T Gillespie. Stochastic simulation of chemical kinetics. Annu. Rev. Phys. Chem., 58:35–55, 2007.
8
[19] Vijaysekhar Chellaboina, Sanjay P Bhat, Wassim M Haddad, and Dennis S Bernstein. Modeling and
analysis of mass-action kinetics. IEEE Control Systems, 29(4):60–78, 2009.
[20] David F Anderson and Thomas G Kurtz. Continuous time markov chain models for chemical reaction
networks. In Design and analysis of biomolecular circuits, pages 3–42. Springer, 2011.
[21] Uri M Ascher and Linda R Petzold. Computer methods for ordinary differential equations and
differential-algebraic equations, volume 61. Siam, 1998.
[22] IBM ILOG CPLEX. User’s manual for CPLEX. Int. Bus. Mach. Corp., 46(53):157, 2009.
[23] Jorge Nocedal and Stephen J. Wright. Numerical Optimization. Springer, New York, NY, USA, second
edition, 2006.
[24] Zekai Hong, David F. Davidson, and Ronald K. Hanson. An improved H2/O2 mechanism based on
recent shock tube/laser absorption measurements. Combustion and Flame, 158(4):633–644, 2011.
[25] CHEMKIN version 18.0, 2016.
9
| 3cs.SY
|
A Density Matrix-based Algorithm for Solving Eigenvalue Problems
Eric Polizzi∗
Department of Electrical and Computer Engineering, University of Massachusetts, Amherst
(Dated: January 17, 2009)
arXiv:0901.2665v1 [cs.CE] 17 Jan 2009
A new numerical algorithm for solving the symmetric eigenvalue problem is presented. The
technique deviates fundamentally from the traditional Krylov subspace iteration based techniques
(Arnoldi and Lanczos algorithms) or other Davidson-Jacobi techniques, and takes its inspiration
from the contour integration and density matrix representation in quantum mechanics. It will be
shown that this new algorithm - named FEAST - exhibits high efficiency, robustness, accuracy
and scalability on parallel architectures. Examples from electronic structure calculations of Carbon
nanotubes (CNT) are presented, and numerical performances and capabilities are discussed.
PACS numbers: 02.60.-x,02.70.Hm,02.70.-c,02.10.Ud,31.15.-p,71.15.Dx
Keywords: diagonalization technique, contour integration, electronic structure calculations
I.
INTRODUCTION
The generalized eigenvalue problem, that is the determination of nontrivial solutions (λ, x) of Ax = λBx
with A and B square matrices, is a central topic in numerical linear algebra and arises from a wide range of
applications in sciences and engineering. In electronic
structure calculations, in particular, the eigenvalue problem is one of the most challenging applied numerical
process - also called diagonalization procedure or spectral decomposition. In these calculations, the electron
density can be formally calculated by summation of the
amplitude square of the wave functions Ψm solution of
the Schrödinger-like eigenvalue problem HΨm = Em SΨm
with different discrete energies Em (where H represents
the Hamiltonian Hermitian matrix and S is a symmetric positive matrix obtained using non-orthogonal basis
functions). This procedure can be quite computationally
challenging for large-scale simulations of systems containing more than a hundred of atoms and/or where a large
number of eigenpairs (Em , Ψm ) are needed. Progress in
electronic structure calculations as for other large-scale
modern applications, are then much likely dependent on
advances in diagonalization methods.
In the past decades, the eigenvalue problem has led
to many challenging numerical questions and a central
problem [1]: how can we compute eigenvalues and eigenvectors in an efficient manner and how accurate are they?
Powerful tools have then been developed from Jacobi
method and power iterations, to iterative Krylov subspace techniques (including Arnoldi, and Lanczos methods), or other Davidson-Jacobi techniques [2]. Traditional numerical algorithms and library packages are yet
facing new challenges for addressing the current largescale simulations needs for ever higher level of efficiency,
accuracy and scalability in modern parallel architectures.
This article presents a new robust and scalable algorithm design for solving the eigenvalue problem- named
∗ URL:
http://www.ecs.umass.edu/ece/polizzi
FEAST- which deviates fundamentally from the traditional techniques above, and takes its inspiration from
the density matrix representation and contour integration in quantum mechanics. Section II summarizes the
electronic structure and contour integration problems
which have motivated the development of the new algorithm. The FEAST algorithm is then described in detail in Section III, and numerical examples and performance results are presented in Section IV. Finally, Section V presents some discussions regarding the efficiency,
robustness and scalability of the algorithm.
II. THE CONTOUR INTEGRATION
TECHNIQUE IN ELECTRONIC STRUCTURE
CALCULATIONS
Although new fast sparse solvers have allowed considerable time saving for obtaining the eigenpairs (Em , Ψm )
in electronic structure calculations, such as the Rayleighquotient multigrid [3] developed for the MIKA package,
or the parallel Chebyshev subspace iteration technique
developed for the PARSEC package [4, 5], these calculations are still considered computationally extremely challenging and linear scalability is not easily achievable.
An alternative approach to the Schrödinger picture
for obtaining the electron density consists in performing a contour integration of the diagonal elements of
the Green’s function matrix G(Z) = (ZS − H)−1 over the
complex energy space [6]. At zero temperature, the resulting expression for the electron density in real-space
is:
Z
X
1
n(r) = −
dZ G(r, r, Z) =
|Ψm (r)|2 , (1)
2πı C
m
where the complex contour C includes all the eigenvalues Em below the Fermi level EF , and where the spin
factor is not considered. It should be noted that at non
zero temperatures, this expression would also include the
contribution of the residues of all poles of the FermiDirac distribution function on the imaginary axis at the
position of the Fermi level [7]. For transport problems
2
and open systems, in turn, the contour integration is often used to compute the equilibrium part of the electron
density [8] where self-energy boundary conditions need to
be included in the Hamiltonian matrix H. The contour
integration technique represents a priori an attractive alternative approach to the traditional eigenvalue problem
for computing the electron density since the number of
Green’s function to be calculated -typically ∼ O(10) using a Gauss quadrature procedure- is independent of the
size of the system. In particular, an efficient linear scaling strategy CMB (CMB stands for Contour integration
- Mode approach - Banded system solver), has been proposed in [9, 10] for simulating nanowire-type structures
within a real-space mesh framework while overcoming the
impossible numerical task of inverting a large size matrix
at each point of the contour. For arbitrary systems (i.e.
beyond nanowire structures), however, there are no numerical advantages of abandoning the traditional eigenvalue problem in favor of the contour integration technique for computing the electron density. In addition, it
is clear from equation (1) that the contour integration
technique does not provide a natural route for obtaining
the individual eigenvectors but rather the summation of
their amplitudes square. In the following section, a new
numerical algorithm design FEAST is proposed for obtaining directly the eigenpairs solutions using the density
matrix representation and a numerically efficient contour
integration technique.
III.
A.
FEAST
Introduction
In this section, a new algorithm is presented for solving
generalized eigenvalue problems of this form
Ax = λBx,
in quantum mechanics. One can show that this factorization can be expressed in terms of the eigenvectors present
inside the contour as follows:
ρ=−
Z
dZ G(Z) ==
C
M
X
|xm ihxm |.
(3)
m=1
In matrix notations the second term of the equation reads
XXT where XN×M = {x1 , x2 , ..xM } (M being the number of eigenvalue inside the contour and N the size of
G). It should be noted that the diagonal elements of ρ
represent the electron density in quantum mechanics (1)
discussed in Section II.
Postmultiplying ρ by a set of M linearly independent random vectors YN×M = {y1 , y2 , ..yM }, the first expression in (3) leads to a new set of M independent vectors
QN×M = {q1 , q2 , ..qM } obtained by solving linear systems along the contour
Z
1
QN×M = −
dZ G(Z)YN×M ,
(4)
2πı C
while the second expression in (3), implies these vectors
Q can be formally generated by the eigenfunctions X
inside the contour
QN×M = XN×M WM×M
with
Wi,j = xT
i yj .
(5)
In other words, each Q column vector obtained in (4)
represents a different linear combination of unknown basis functions X in (5). Using a Rayleigh-Ritz procedure, the problem (2) is now equivalent to computing the
eigenpairs (ǫm , Φm ) of the following reduced generalized
eigenvalue problem of size M:
AQ Φ = ǫBQ Φ
(2)
within a given interval [λmin , λmax ], where A is real symmetric or Hermitian and B is a symmetric positive definite (s.p.d). One common way to accelerate the convergence rate of traditional iterative techniques consists
in performing a factorization of the Green’s function
G(σ) = (σB − A)−1 for some reasonable shift σ close
to the eigenvalues in the search interval and which leads
to solving linear systems (i.e. shifting strategy). More
recently, Sakurai et al. [11, 12] have proposed a root
finding technique which consists of a contour integration
of a projected Laurent series-type decomposition of the
Green’s function. In principle, a set of complex moments
can be obtained by solving few linear systems along the
contour, which can generate an identical subspace to the
one spanned by the eigenvectors present inside the contour. In practice, however, robustness and accuracy are
not easily achievable. In our approach, we avoid decomposing directly the Green’s function and perform instead
an exact mathematical factorization of its contour integration - which represents the reduced density matrix ρ
1
2πı
T
with AQ = Q AQ
T
and BQ = Q BQ.
(6)
(7)
The Ritz values and vectors are then given by:
λm = ǫm , m = 1, . . . , M
XN×M = QN×M ΦM×M
(8)
(9)
where ΦM×M = {Φ1 , Φ2 , ..ΦM }. One can show that the
obtained eigenvectors X are naturally B-orthonormal i.e.
xT
i Bxj = δi,j , if the eigenvectors of the reduced problem
(6) are BQ -orthonormal i.e. ΦT
i BQ Φj = δi,j .
B.
Practical Considerations and Pseudocode
In practice, the vectors Q are computed by performing
a numerical integration of each vectors G(Z)Y (4) along
the complex contour C. Let us consider a circle centered
in the middle of the search interval [λmin , λmax ], it should
be noted that the expression of the contour integration
3
(ii) Postmultiplying the density matrix (3) by M0 random vectors (rather than M) where M0 is greater than
M. The reduced dense generalized eigenvalue problem
(6) of size M0 can be solved using standard eigenvalue
LAPACK routines [13]. Since we do not perform the orthogonalization of the vectors Q, one has to make sure
that BQM0 ,M0 is symmetric positive definite i.e. M0 does
not exceed an upper limit which can easily be obtained
a posteriori.
C+
λmin
λ
λmax
FIG. 1: Schematic representation of the complex contour integral defined by the positive half circle C + . In practice, the vectors Q are computed via a numerical integration (e.g. GaussLegendre quadrature) where only few linear systems G(Z)Y
needs to be solved at specific points Ze along the contour.
1- Select M0 > M random vectors YN×M0 ∈ RN×M0
2- Set Q = 0 with Q ∈ RN×M0 ; r = (λmax − λmin )/2;
For e = 1, . . . Ne
can be further simplified since G(Z̄) = G† (Z). Denoting C + the positive half circle of the complex contour, it
comes if A is Hermitian:
Z
1
dZ {G(Z) − G† (Z)},
(10)
ρ=−
2πı C +
and if A is real symmetric:
Z
1
ρ=−
dZ ℑ{G(Z)},
π C+
compute θe = −(π/2)(xe − 1),
compute Ze = (λmax + λmin )/2 + r exp(ıθe ),
solve (Ze B − A)Qe = Y to obtain Qe ∈ CN×M0
compute Q = Q − (ωe /2)ℜ{r exp(ıθe ) Qe }
End
3- Form AQ M0 ×M0 = QT AQ and BQ M0 ×M0 = QT BQ
4- Solve AQ Φ = ǫBQ Φ to obtain the M0 eigenvalue ǫm ,
(11)
where ℑ{} stands for the imaginary part. Using a Ne point Gauss-Legendre quadrature on the positive half circle C + (see Figure 1), with xe the eth Gauss node associated with the weight ωe , one obtains if A is Hermitian
and Y, Q ∈ CN×M :
and eigenvectors ΦM0 ×M0 ∈ RM0 ×M0
5- Set λm = ǫm and compute XN×M0 = QN×M0 ΦM0 ×M0
If λm ∈ [λmin , λmax ], λm is an eigenvalue solution
and its eigenvector is Xm (the mth column of X).
6- Check convergence for the trace of the eigenvalues λm
If iterative refinement is needed, compute Y = BX
Q=−
Ne
X
1
e=1
4
ωe r exp(ıθe ) G(Ze ) + exp(−ıθe ) G† (Ze ) Y,
(12)
and go back to step 2
(15)
FIG. 2: FEAST pseudocode (sequential version) for solving
the generalized eigenvalue problem Ax = λBx, where A is
real symmetric and B is s.p.d., and obtaining all the M eigenpairs within a given interval [λmin , λmax ]. The numerical integration is performed using Ne -point Gauss-Legendre quadrature with xe the eth Gauss node associated with the weight
ωe . For the case Ne = 8, one can use:
(x1 , ω1 ) = (0.183434642495649, 0.362683783378361),
(x3 , ω3 ) = (0.525532409916328, 0.313706645877887),
(x5 , ω5 ) = (0.796666477413626, 0.222381034453374),
(x7 , ω7 ) = (0.960289856497536, 0.101228536290376),
and (x2i , ω2i )i=1,...,4 = (−x2i−1 , ω2i−1 )
where ℜ{} stands for the real part.
In order to reduce the numerical quadrature error of
the contour integral, one may consider the two following
improvements:
(i) Performing outer-iterative refinement steps. Once the
eigenvectors X are obtained (9), a new set of initial guess
vectors Y = BX can be used. Postmultiplying the density matrix (3) by Y, one now obtains from (5) that Q
converges to X since XT BX = I (i.e. Wi,j = δi,j and
then ρBX = X). A fast test for convergence can be
obtained by checking the trace of the eigenvalues (8).
The performances of the basic FEAST algorithm will
then depend on a trade off between the choices of the
number of Gauss quadrature points Ne , the size of the
subspace M0 , and the number of outer refinement loops.
So far, using M0 ≥ 1.5M, Ne = 8, and with at most 2
refinement loops, we have consistently obtained a relative residual equal or smaller than 10−10 seeking up to
1000 eigenpairs for a variety of problems. The basic
pseudocode for the FEAST algorithm is given in Figure 2 in the case of A real symmetric. In the case of
A complex Hermitian, we note the following changes:
with
r=
λmax − λmin
, θe = −(π/2)(xe − 1),
2
λmax + λmin
+ r exp(ıθe ).
Ze =
2
(13)
(14)
If A is real symmetric, Y, Q ∈ RN×M and one can use:
Q=−
Ne
X
1
e=1
2
ωe ℜ{r exp(ıθe ) G(Ze )Y},
4
Y, Q ∈ CN×M0 , Φ ∈ CM0 ×M0 , and the construction of
the vectors Q in step-2 of the pseudocode must be modified to satisfy (12).
IV.
NUMERICAL EXPERIMENTS
In this section, we propose to demonstrate the numerical stability, robustness and scalability of the FEAST
algorithm using three examples derived from electronic
structure calculations of Carbon nanotube (CNT).
A.
Example I
Let us first consider a family of eigenvalue problems,
Test-CNT, obtained using a 2D FEM discretization of the
DFT/Kohn-Sham equations at a given cross-section of a
(13,0) CNT – the 2D atomistic potential is derived from
the mode approach used in the CMB strategy for solving
the full 3D problem presented in [9]. In Test-CNT, A is
real symmetric, and B is s.p.d., the size of both matrices
is N = 12, 450 and their sparsity pattern is identical with
a number of non-zero elements nnz = 86, 808.
In Table I, we report the times and relative residual
obtained by the public domain eigenvalue solver package ARPACK [14] (using the shift-invert strategy) and
the FEAST algorithm presented in Figure 2 for solving
the Test-CNT example seeking up to M = 800 (lowest) eigenpairs. The inner linear systems in ARPACK
and FEAST are solved using the shared-memory parallel
direct solver PARDISO [15]. It should be noted, however, that FEAST benefits more than ARPACK from the
PARDISO solver, as the inner linear systems have multiple right-hand sides. Although both algorithms could
benefit from a parallel distributed implementation (e.g.
using the P ARPACK package), the simulation runs are
here restricted to a given node of a 8-cores Intel Clovertown system (16Gb,2.66GHz) where the linear systems
in FEAST are factorized and solved one after another.
The performances of ARPACK and FEAST can also depend on fine tunings parameters such as the choices of the
size of the subspace M0 (M0 = 1.5M here for both algorithms), the inner systems solvers, the number of contour
points Ne for FEAST, or the stopping criteria for obtaining the residual. The simulation results in this section
are then not intended to compare quantitatively the two
solvers but rather to point out the potentialities of the
FEAST algorithm.
In our experiments, the convergence criteria on the relative residual for FEAST is obtained
P when the relative
error on the trace of the eigenvalues m λm in the search
interval is smaller or equal to 10−13 . Table II shows the
variation of the relative error on the trace with the number of outer-iterative refinement for FEAST. These results demonstrate that only 2 to 3 refinement loops are
necessary to obtain the small relative residuals for the
different cases reported in Table I. It should be noted
TEST-CNT
ARPACK
N = 12, 450 Time(s)
Resid.
M = 100
12.2
2.0 ∗ 10−11
M = 200
31
2.0 ∗ 10−11
M = 400
86
1.4 ∗ 10−11
M = 800
213
4.5 ∗ 10−9
FEAST
Time(s)
Resid.
7.8
4.5 ∗ 10−10
14
5.5 ∗ 10−10
21
1.8 ∗ 10−10
58
3.4 ∗ 10−11
TABLE I: Simulations times and relative residual
maxi (||Axi − λi Bxi ||1 /||Axi ||1 ), obtained by the solver
ARPACK and FEAST on the TEST-CNT system seeking
M (lowest) eigenpairs for different search intervals. The
simulations are performed on a Intel Clovertown (8cores, 1
node, 2.66GHz, 16Gb). The shift-strategy has been used in
ARPACK to accelerate the convergence (the regular mode
would give ∼ 300s for M = 100). The inner linear systems
in ARPACK and FEAST are both solved using the direct
parallel solver PARDISO [15] on 8 cores. Finally, the size
of the subspace has been chosen to be M0 = 1.5M for both
algorithms, and the number of contour points for FEAST is
fixed at Ne = 8.
that only one loop is necessary to obtain the eigenvalues
with an accuracy of ∼ 10−5 or below.
TEST-CNT
N = 12, 450
M = 100
M = 200
M = 400
M = 800
Relative error on the Trace
1st loop
2nd loop
3rd loop
3.0 ∗ 10−6 2.9 ∗ 10−12 1.0 ∗ 10−15
1.8 ∗ 10−5 4.8 ∗ 10−12 2.1 ∗ 10−14
2.4 ∗ 10−8 3.2 ∗ 10−16
1.8 ∗ 10−9 4.3 ∗ 10−16
TABLE II: Variation
of the relative error on the trace of
P
the eigenvalues m λm for different search intervals with the
number of iterative refinement loops. The convergence criteria is set to 10−13 where the final relative residual on the
eigenpairs is reported in Table I.
The simulation results in Table I demonstrate very
good scalability for FEAST while the search interval
keeps increasing but the number of contour points Ne
stays identical (i.e. the number of numerical operations
stays the same for a given loop of FEAST with a fixed
Ne = 8 linear systems to solve). In addition, from Table III, one can see how the robustness of the FEAST
algorithm is affected while the number of contour points
Ne changes. In particular, Ne = 4 points along the contour did suffice to capture M = 100 eigenpairs with a
relatively small residual (decreasing the simulation time
reported in Table I for this case), while the case Ne = 16
points generated a residual smaller than the one obtained
by ARPACK (using M0 = 1.5M).
B.
Example II
In another set of numerical experiments, we intend to
demonstrate the robustness of FEAST in capturing the
5
TEST-CNT
ARPACK
N = 12, 450
M = 100
Time(s)
Resid.
(N, M)
12.2
2.0 ∗ 10−11
2(N, M)
85
3.5 ∗ 10−11
4(N, M)
668
4.6 ∗ 10−11
8(N, M)
5492 6.2 ∗ 10−11
FEAST
Time(s)
7.8
27
109
523
Resid.
4.5 ∗ 10−10
7.7 ∗ 10−10
8.8 ∗ 10−10
6.5 ∗ 10−10
TABLE IV: Simulations times and relative residual
maxi (||Axi − λi Bxi ||1 /||Axi ||1 ), obtained by the solver
ARPACK and FEAST on the k(N, M) TEST-CNT systems
which artificially reproduce k times the original TEST-CNT
system. The kM (lowest) eigenpairs are found where each
eigenvalue has a multiplicity of k.
0
-2
E (eV)
multiplicity of the eigenvalues. We propose to create artificially new TEST-CNT systems called k(N, M) where
the matrices A and B are repeated k times along the
main diagonal (the new system matrix is block diagonal
with k blocks). Physically, these systems can describe
the cross section of a bundle composed by k CNTs, where
we do not consider the interactions between the different
tubes such that each eigenvalue is now k times degenerate. If we keep the same search interval used to obtain
M = 100 eigenpairs for k = 1 (where the size of the matrices A and B is N), 100k eigenpairs must now be found
for k ≥ 1, where each one of them have the multiplicity k.
In Table IV, we report the simulation times and relative
residuals obtained using ARPACK and FEAST on these
k(N, M) TEST-CNT systems. For the case 8(N, M), for
example, the size of the new system matrix is 99, 600
and the first 100 eigenvalues have all the multiplicity 8
(so 800 eigenpairs are found in total). The simulation results show linear scalability performances with the size of
the system and the number of eigenpairs. In contrast to
ARPACK where the number of matrix-vector multiplications and linear system solves would keep increasing with
k, the number of operations in FEAST stays the same for
all these cases. The scalability of the algorithm depends
then mainly on the scalability of the linear system solver.
Example III
We have shown that FEAST can re-use the computed
subspace as suitable initial guess for performing iterative refinements. This capability can also be of benefit
to modern applications in science and engineering where
it is often necessary to solve a series of eigenvalue problems that are close one another. In bandstructure calculations, in particular, many eigenvalue problems of the
form (A + Sk )xk = λk Bxk need to be solved at different
locations in the k-space (i.e. for different values of k and
where S is Hermitian with S0 = I). Let us consider the
eigenvalue sparse system of size N = 492, 982 obtained for
a (5,5) metallic CNT using our in-house DFT/real-space
mesh technique framework for bandstructure calculations
of nanowires-type structure [16]. In Figure 3, we propose
to solve this eigenvalue problem using the same search
interval for the eigenvalues λ for different locations of k
where the subspace computed by FEAST at the point
k − 1 is successively used as initial guess for the neighboring point k. In addition, the inner linear systems in
-4
-6
-8
-10
#Eigenvalues
TABLE III: Performance results obtained by FEAST seeking
M = 100 eigenpairs for different values of Ne . The convergence is obtained when the error on the trace is equal or
smaller to 10−13 .
C.
#FEAST loop
TEST-CNT
FEAST
M = 100 Time(s)
Resid.
# loops
Ne = 4
7.0
8.3 ∗ 10−8
6
Ne = 8
7.8
4.5 ∗ 10−10
4
Ne = 16
10.2
3.4 ∗ 10−12
3
-12
20
20
18
18
16
16
14
14
12
4
12
4
3
3
2
2
1
1
0
Γ
k
0
X
FIG. 3: Bandstructure calculations of a (5,5) metallic CNT.
The eigenvalue problems are solved successively for all the k
points (from Γ to X), while the computed subspace of size
M0 = 25 at the point k is used as initial guess for the point
k + 1. The number of eigenvalues found ranges from 13 to 20,
and by the third k point, the FEAST convergence is obtained
using only one refinement loop. The convergence is obtained
with the relative error on trace of the eigenvalues smaller or
equal to 10−8 , while the inner linear systems are solved using an iterative method with an accuracy of 10−3 . The final
relative residuals on the eigenpairs range from 10−3 to 10−5 .
FEAST are solved using an iterative method with preconditioner where a modest relative residual of 10−3 is used
(e.g. a suitable banded preconditioner can be obtained
using a mode approach [9]). It should be noted that the
convergence criteria for the relative error on the trace
of the eigenvalues is chosen much smaller at 10−8 , while
the eigenvectors are expected to be obtained within the
same (or a smaller) order of accuracy that the one used
6
for the solutions of the inner systems. Figure 3 shows
that 13 to 20 eigenvalues (i.e. energies) are found within
the selected search interval along the different k points
(from the Γ to the X point in the graph). Although the
size of the subspace stays identical at M0 = 25, after the
first initial point at k = 0 (Γ point in the graph) FEAST
converges within only one refinement loop for almost all
the other k points.
V.
DISCUSSIONS
In comparison to iterative Krylov subspace techniques,
FEAST can be cast as a ”direct” technique which is
based on an exact mathematical derivation (3). FEAST
does naturally then capture all the multiplicity and noorthogonalization procedure is necessary (such as GramSchmidt orthogonalization process). As described above,
the main computational tasks in FEAST consist of solving Ne independent linear systems along the contour
with M0 right-hand-sides and a reduced dense generalized eigenvalue problem of size M0 . Since FEAST has
the ability to re-use the basis from the previously computed subspace, an outer-iterative refinement procedure
is proposed to improve the accuracy of the solutions. The
capability to take advantage of suitable initial guess can
also be of benefit to modern applications in sciences and
engineering where it is often necessary to solve a series
of eigenvalue problems that are close one another (e.g.
Bandstructure calculations in Example III Section IV).
In one sense, the difficulty of solving an eigenvalue
problem has been replaced by the difficulty of solving
a linear system with multiple right-hand sides. For large
sparse systems, this latter can be solved using either a
direct system solver such as PARDISO [15] (as proposed
in Section IV), or an iterative system solver with preconditioner. In turn, for banded systems or banded preconditioner, FEAST can be seen as an outer-layer for the
author’s SPIKE parallel system solver [17]. It should be
noted that the inner linear systems arising from standard eigenvalue solvers (using the shift-strategy), need
often to be solved highly accurately via direct methods.
Direct system solvers, however, are not always suited
for addressing large-scale modern applications because of
memory requirements. In Example III of Section IV we
have shown that FEAST can take advantage of iterative
solvers for solving the inner linear systems with modest
[1] G. Golub and H. A. van der Vorst, J. of Comput. and
Appl. Math. 123, 35 (2000).
[2] Z. Bai, J. Demmel, J. Dongarra, A. Ruhe, and H. van der
Vorst, Templates for the solution of Algebraic Eigenvalue
Problems: A Practical Guide (Society for Industrial and
Applied Mathematics, Philadelphia, PA, 2000).
[3] T. Torsti, M. Heiskanen, M. J. Puska, and R. M. Nieminen, Int. J. Quantum Chem 91, 171 (2003).
relative residual and obtaining the eigenvectors solution
within the same order of accuracy. The resulting subspace could also be used as a very good initial guess for a
one step more accurate refinement procedure (i.e. using
more accurate relative residual for the inner systems).
FEAST exhibits important potentialities for parallelism at three different levels: (i) many search interval
[λmin , λmax ] can be run independently, (ii) each linear
systems can be solved simultaneously (e.g. on each node
of parallel architecture where the factorization of the linear system can be done only once for all the refinement
loops), (iii) the linear system solver can be parallel (e.g.
within a given node as in Section IV). Depending on
the parallel architecture at hand, the local memory of
a given node and the properties of the matrices of the
eigenvalue problems, one may preferably select one parallel option among the others, or just take advantage of
a combination of those. In particular, there will be a
trade off between how many search intervals to consider
and how many eigenpairs FEAST can handle by intervals. For example if M0 is more than few thousands, one
could either (i) solve the obtained reduced system of size
M0 using efficient dense parallel symmetric eigenvalue
solvers [18], or (ii) propose to divide the initial search
interval into two or more to be processed in parallel. In
addition, it should be noted that the orthonormalization
step is absent from FEAST which will drastically reduce
the communication overhead for performing scalar products on high-end parallel architectures (the scalar product in step-3 in Fig. 2 has to be done only once per
iterative refinement). Given the recent advances in parallel architectures and parallel linear system solvers, it
is reasonable to envision using FEAST in a near future
for obtaining up to millions of eigenpairs of large sparse
symmetric eigenvalue problems. Finally the capabilities
of FEAST could potentially be enhanced for addressing
non-symmetric eigenvalue problems where the contour
integration would then be performed in a given region of
the complex space.
Acknowledgments
The author wishes to acknowledge helpful discussions
with Dr. Ahmed Sameh and Dr. Massimo Fischetti.
This material is supported by NSF under Grant #CCF0635196.
[4] Y. Saad, Y. Zhou, C. Bekas, M. L. Tiago, and J. Chelikowsky, physica status solidi (b) 243, 2188 (2006).
[5] Y. Zhou, Y. Saad, M. L. Tiago, and J. Chelikowsky, Phys.
Rev. E. 74, 066704 (2006).
[6] R.Zeller, J. Deutz, and P. Dederichs, Solid state communications 44, 993 (1982).
[7] J. Taylor, H. Guo, and J. Wang, Phys. Rev. B. 63, 245407
(2001).
7
[8] M. Brandbyge, J.-L. Mozos, P. Ordejon, J. Taylor, and
K. Stokbro, Phys. Rev. B. 65, 165401 (2002).
[9] D. Zhang and E. Polizzi, J. Comput. Elec. 7, 427 (2008).
[10] D. Zhang and E. Polizzi, 2008 NSTI Nanotechnology
Conference and Trade Show. Technical Proceedings. 1,
12 (2008).
[11] T. Sakurai and H. Sugiura, J. of Comput. and Appl.
Math. 159, 119 (2003).
[12] T. Sakurai, Y. Kodaki, H. Tadano, D. Takahashi,
M. Sato, and U. Nagashima, Future Generation Computer Systems 24, 613 (2008).
[13] E. Anderson, Z. Bai, C. Bischof, S. Blackford, J. Demmel, J. D. C. J. Dongarra, A. Greenbaum, S. Hammarling, A. McKenney, and D. Sorensen, LAPACK Users
Guide (Society for Industrial and Applied Mathematics,
Philadelphia, PA, 1999), third ed. ed.
[14] R. B. Lehoucq, D. C. Sorensen, and C. Yang., ARPACK
Users Guide: Solution of Large Scale Eigenvalue Problems with Implicitly Restarted Arnoldi Methods. (Society
for Industrial and Applied Mathematics, Philadelphia,
PA, 1998).
[15] O. Schenk and K. Gärtner, Journal of Future Generation
Computer Systems 20, 475 (2004).
[16] D. Zhang and E. Polizzi, in preparation (2009).
[17] E. Polizzi and A. Sameh, Parallel Computing 32, 177
(2006).
[18] L. Blackford, J. Choi, A. Cleary, E. D’Azevedo, J. Demmel, I. Dhillon, J. Dongarra, S. Hammarling, G. Henry,
A. Petitet, et al., ScaLAPACK Users Guide (Society for
Industrial and Applied Mathematics, Philadelphia, PA,
1997).
| 5cs.CE
|
Statistics and Computing manuscript No.
(will be inserted by the editor)
A critical analysis of resampling strategies for the
regularized particle filter
arXiv:1705.04219v1 [stat.CO] 11 May 2017
Pierre Carmier · Olexiy Kyrgyzov · Paul-Henry Cournède
March 16, 2018
Abstract We analyze the performance of different resampling strategies for the regularized particle filter regarding
parameter estimation. We show in particular, building on
analytical insight obtained in the linear Gaussian case, that
resampling systematically can prevent the filtered density
from converging towards the true posterior distribution. We
discuss several means to overcome this limitation, including
kernel bandwidth modulation, and provide evidence that the
resulting particle filter clearly outperforms traditional bootstrap particle filters. Our results are supported by numerical
simulations on a linear textbook example, the logistic map
and a non-linear plant growth model.
Keywords Particle filtering · Bayesian inference · Hidden
Markov model · Kernel density estimation
1 Introduction
Sequential Monte Carlo (SMC) methods [13, 16] are powerful algorithms which allow performing Bayesian inference
in dynamic systems. These methods, also known under the
heading of particle filters in the context of dynamic systems,
have become extremely popular way beyond the borders of
the statistics community and are nowadays widely used by
engineers and scientists from other fields faced with nontrivial inference tasks which naturally abound with the emergence of data science. The purpose of this paper is to provide
a critical analysis of the performance of regularized particle filter (RPF) algorithms [26, 30] which represent an interesting variation of traditional particle filters. They rely on
kernel-based [31] posterior density estimation at each filtering step, with particle resampling performed from the [email protected]
Biomathematics Team, MICS Lab,
CentraleSupélec, Université Paris-Saclay, France
structed density. In many situations, this method was proven
to help prevent sample impoverishment [16] and thereby enhance the robustness of particle filtering. Our main achievement will be to demonstrate in this paper that the approach
which consists in resampling systematically can lead to seriously over-estimate the posterior mean squared error, motivating us to exhibit alternative resampling strategies which
allow fully resolving this issue.
Before going any further, let us introduce the general
framework we shall be working with in the rest of this paper. To begin with, we shall assume that the processes from
which we seek to infer some information belong to the family of hidden Markov models (HMM) [4], which describe
stochastic processes characterizing the evolution in discrete
time n of two sets of random variables {xn }n≥0 and {yn }n≥1 ,
respectively defined on some space X ⊆ Rdx and Y ⊆ Rdy .
The former are hidden random variables referred to as states,
while the latter are noisy observations of the former. Starting from a prior probability measure π0 (dx0 ) on the initial
state, hidden variables follow a Markovian evolution characterized by a sequence of Markov kernels fn (dxn |xn−1 )
also named transition functions. Observations are considered conditionally independent and described by a conditional probability density function gn (yn |xn ), the so-called
measurement function, with respect to the Lebesgue measure λ . Parameters {Θ }, with a compact support set, can
easily be accounted for in that framework by augmenting the
size of the state space, as can the presence of environmental
variables {un }n≥0 defined on some space U ⊆ Rdu .
A problem of broad interest in this context is to determine the posterior probability measure πk (dxk |y1:n ) given a
sequence of observations y1:n = {y1 , ..., yn }. If k < n, this is
called a smoothing problem, while if k > n this is a prediction problem. In the following we shall be interested in the
situation where k = n, which is called the filtering problem.
R
Thus, given any Borel set A ⊂ X, πn (A) = A πn (dxn |y1:n )
2
Pierre Carmier et al.
shall be the posterior probability of event A. More generally, for any test function φ defined on X, applying Bayes
rule recursively allows expressing the integral of φ with respect to the posterior probability measure as a function of
past measurement and transition functions [4]
Z
X
=
φ (xn )πn (dxn |y1:n )
1
Zn
n
Z
Xn+1
(1)
φ (xn )π0 (dx0 ) ∏ g j (y j |x j ) f j (dx j |x j−1 ) ,
j=1
introducing the marginal likelihood
Z
Zn =
Xn+1
n
π0 (dx0 ) ∏ g j (y j |x j ) f j (dx j |x j−1 ) .
(2)
modified particle filter clearly outperforms that of the original bootstrap particle filter [18]. The rest of the paper is organized as follows. In section 2, we recall basic principles of
SMC methods and introduce the bootstrap particle filter and
the RPF, before illustrating what can potentially go wrong
when resampling is performed on a systematic basis. We
then provide analytical insight regarding the effect of different resampling strategies on the behavior of the RPF in
section 3, using a very simple linear textbook example for
illustrative purposes. That our results qualitatively hold for
more general, non-linear, models is discussed in section 4.
Conclusions are drawn in section 5, while a series of technical details are gathered in appendices.
j=1
The purpose of SMC methods [28, 14, 16] is to provide an
estimation of such probability measures in the general case
of nonlinear state-space models, where analytical attempts
are essentially hopeless. The idea of such methods is to represent the probability distributions by particles, i.e. independent identically distributed samples of the distributions, and
to propagate and update them using the HMM equations,
hence the equivalent nomenclature of particle filter methods.
The estimation of parameters in this context has received
acute attention from the community these last years (see [24]
for a review). One of the main challenges is to circumvent
the celebrated particle degeneracy problem [18] and to ensure that parameter space is sufficiently explored to enable
accurate estimation. Several methods have been proposed to
address this challenge, among which iterated filtering [20]
and particle Markov chain Monte Carlo [1] in the specific
context of joint parameter / hidden state estimation. They
constitute batch or offline methods, however, in that they are
unable to take into account new observations dynamically,
as opposed to online algorithms such as SMC2 [10, 21].
A different idea, introduced some time ago in [36] and
discussed in more detail in [27, 30], is to allow freedom in
parameter space by regularizing the posterior density using a mixture. In this paper, we focus on the RPF introduced in Ref. [26] and further developed and discussed in
Refs [30, 2, 7, 17]. We shall highlight some shortcomings of
this approach when resampling is performed systematically
(as routinely done), which can be made explicit by computing the posterior mean squared error. In particular, building
upon some well-known results obtained for linear Gaussian
state-space models, we prove that the mean squared error
is lower bounded by a positive constant when the number
of particles is finite, independently of the number of observations. We shall provide a detailed understanding of this
phenomenon on a very simple linear textbook example, offering evidence that it is related to the particle filter losing
memory of past observations. We will then show that longranged memory can be restored by appropriately modulating
the kernel bandwidth and that the efficiency of this suitably
2 Sequential Monte Carlo methods
Let us begin by recalling the basic principles of particle filter
methods. Starting from the initial prior probability measure
(i)
from which N particles are sampled, x0 ∼ π0 (dx0 ), particles are propagated until the next time step using a Markov
(i)
(i)
kernel generally chosen [18, 25] as x1|0 ∼ f1 (dx1 |x0 ) in
(i)
nonlinear state space models for which f1 (dx1 |x0 , y1 ) is not
readily available [28]. The resulting empirical estimator for
the predicted probability measure can be expressed as
N
π1|0
(dx) =
1 N
∑ δx(i)1|0 (dx)
N i=1
(3)
with δ the Dirac measure. The estimator for the updated
posterior probability measure is obtained by weighting each
particle according to the observation y1 :
N
(i)
π1N (dx1 |y1 ) = ∑ w1 δ
i=1
(i)
x1|0
(dx1 )
(4)
with weights
(i)
(i)
w1
=
g1 (y1 |x1|0 )
(5)
( j)
∑Nj=1 g1 (y1 |x1|0 )
suitably normalized to ensure that total weight is conserved.
Given any bounded function φ defined on X, its integral with
respect to the empirical posterior probability measure is then
given by
Z
X
N
(i)
(i)
φ (x1 )π1N (dx1 |y1 ) = ∑ w1 φ (x1|0 ) ,
(6)
i=1
and the estimator can be shown to converge asymptotically,
for an infinite number of particles, towards the true posterior
probability measure given by Bayes rule [9]:
Z
lim
N→∞ X
φ (x1 )π1N (dx1 |y1 ) =
Z
X
φ (x1 )π1 (dx1 |y1 ) .
(7)
A critical analysis of resampling strategies for the regularized particle filter
Practically speaking, however, one always works with a finite number of particles and life is not as simple. A wellknown problem which can occur with the above estimation
procedure is the so-called weight degeneracy issue [16]. If
the likelihood is relatively peaked around the observation
value, or in other words if the observation noise is rather
small, particle weights will differ by orders of magnitude
depending on the distance between the particle and the observation. Actually, this will happen almost regardless of the
magnitude of the observation noise for a sufficient number
of observations, since iterating the above procedure yields
(i)
(i)
wn ∝ ∏nj=1 g j (y j |x j| j−1 ), using the recursive formulation
(i)
(i)
x j| j−1 ∼ f j (dx j |x j−1 ). This implies that, eventually, most of
the weight will be carried by a small number of particles,
with a dramatic effect on the precision of the resulting estimation as can be seen by looking at the unbiased estimator
of the marginal likelihood at step n:
3
2.1 Regularized particle filter
In Ref. [26], LeGland, Musso and Oujdane proposed to address these shortcomings by regularizing the probability measure using a Parzen-Rosenblatt [31] kernel density estimator.
The idea is to replace the Dirac measure by a measure able
to smoothly interpolate between all weighted particles. In
other words, observing that the coarse-grained density on
Borel set Aε ⊂ X
N
1
1
(i)
πnN (Aε ) =
wn δ (i) (Aε ) ,
∑
xn|n−1
λ (Aε )
λ (Aε ) i=1
R
where λ (Aε ) = Aε λ (dxn ), the regularization procedure consists in replacing δx (Aε )/λ (Aε ) as λ (Aε ) → 0 by a kernel
density function with respect to the Lebesgue measure [35],
thanks to which the regularized estimator for the posterior
probability density function reads
N
ZnN
1 N n
(i)
= ∑ ∏ g j (y j |x j| j−1 ) .
N i=1 j=1
(8)
(12)
(i)
(i)
pNn (xn |y1:n ) = ∑ wn Khn (xn − xn|n−1 ) ,
(13)
i=1
with
Its variance can be estimated using the standard empirical
estimator
1 N
∑
N 2 i=1
!2
n
(i)
∏ g j (y j |x j| j−1 ) − ZnN
j=1
where we introduced the effective sample size
1
(i)
∑Ni=1 (wn )2
(10)
which is a measure of the effective number of particles [25,
6]: if all particles carry the same weight, ESSn = N, while if
a single particle carries all the weight, ESSn = 1. This result
therefore illustrates that the accuracy of the estimation by
the empirical measure relies on a relatively homogeneneous
distribution of particle weights.
A natural strategy to counter the effect of particle degeneracy is to ’unweight’ particles by resampling from the em(i)
pirical posterior probability measure, xn ∼ πnN (dxn |y1:n ),
such that the new approximation to the posterior is
π̃nN (dxn |y1:n ) =
1 N
∑ δx(i)n (dxn ) .
N i=1
(i)
( j)
(i)
(i)
• xn|n−1 ∼ fn (dxn |xn−1 )
(ZnN )2
ESSn
=
,
1−
ESSn
N
(9)
ESSn =
(i)
• wn = gn (yn |xn|n−1 )/ ∑Nj=1 gn (yn |xn|n−1 )
(11)
This is the idea behind the bootstrap particle filter [18], also
known as the sequential importance resampling (SIR) algorithm. Its main drawback is that resampling according
to a probability measure such as Eq. (4) essentially duplicates particles (particles are selected following a multinomial law), leading to so-called sample impoverishment [16].
(i)
• xn−1 ∼ pNn−1 (xn−1 |y1:n−1 ) .
In the following, we shall consider a Gaussian kernel Khn =
N(0, h2n ): Eq. (13) then effectively describes a Gaussian mixture and resampling according to this distribution now consists in a two-step procedure, with the particle selection (performed using the systematic resampling algorithm [19]) complemented by a perturbation controlled by the kernel bandwidth hn . Each particle generated in the process is thus unique.
Note that, if the likelihood function gn is unknown, it can itself be emulated by a kernel density estimator [2, 7].
There is a whole literature on how the kernel bandwidth
hn can be chosen. For now, we will follow Ref. [30] and pick
h2n = αh Σn , with
αh =
4
N(dx + 2)
2
dx +4
.
(14)
This
minimizes
the mean integrated squared error
choice
N
2
E ||pn − pn ||L2 , provided the true probability density function pn is Gaussian (see Appendix A for the proof in one dimension with uniform weights). In practice, the true covariance matrix Σn is unknown and is therefore replaced by an
(i)
(i)
estimator ΣnN based on the weighted samples (xn|n−1 , wn ):
ΣnN =
N N (i) (i)
∑ wn (xn|n−1 − µnN )2
N − 1 i=1
(i) (i)
(15)
with µnN = ∑Ni=1 wn xn|n−1 . As an important corollary, because limN→∞ hn = 0, we retrieve asymptotically that, just
4
Pierre Carmier et al.
as for the empirical probability measure, the regularized estimator converges towards the true posterior density:
2
E ||pNn − pn ||L2 = O N − dx +4 .
(16)
N→∞
2.2 Problem statement
While the general strategy underlying the RPF seems very
principled, we proceed to highlight a problem which can be
encountered when resampling is performed on a systematic
basis, i.e. at every step, as generally done. Recall that, given
a prior probability measure π0 (dx0 ) and a series of observations {y1:n }, we seek to determine the true posterior distribution x̂n ∼ πn (dxn |y1:n ). In particular, we wish to estimate
how the RPF estimator x̂Nn ∼ pNn (xn |y1:n ) fares with respect
to x̂n . In what follows, we shall provide evidence that, for
a finite number of particles, the RPF estimator features a
systematic deviation from the true posterior density which
increases with the number of observations. This deviation
can be made explicit by comparing the root mean squared
error of the RPF with that of the true posterior density:
1/2
RMSE(x̂n , xn ) = E ||x̂n − xn ||2
q
= ||µn − xn ||2 + Tr[Σn ] ,
(17)
where ||.|| indicates the usual Euclidean norm, and introducing the mean value µn := E[x̂n |y1:n ] and the covariance matrix Σn := E[(x̂n − µn )(x̂n − µn )T ]. For illustrative purposes,
consider the very simple case of a one-dimensional stationary state xn+1 = xn with time-independent additive observation noise η ∼ N(0, R): yn = xn + η. This is of course an
elementary textbook setting, which is in fact a special case
of the Kalman filter framework which shall be reviewed in
the following section. The advantage of this setup is to provide us with the exact result with respect to which particle
filter expectations can be benchmarked. In the left frame
of Fig. 1, we plot the analytical solution and compare it
with the outcome of the RPF when resampling is performed
at each step. As can be clearly seen, while the analytical
expectation steadily decreases with each new observation,
the RMSE computed using the RPF rather quickly saturates
and remains lower bounded by a positive constant no matter how many observations are considered. The explanation
for this behavior essentially stems from the fact that resampling from the regularized estimator, while allowing to explore particle space, also acts as a perturbation which increases the variance of the estimator. This perturbation becomes detrimental for the estimation process in the long run,
as is manifest in the left frame of Fig. 1. While this problem
was already acknowledged in a slightly different context in
the early days of particle filters [36], some of its implications may not have been fully grasped at the time. In partic-
ular, we claim that the over-estimation of the variance prevents the estimator from remaining asymptotically efficient
as the number of observations increases, in a sense which
shall be defined shortly. Mathematical arguments supporting this claim for linear state-space models will be provided
in the following section, while numerical simulations indicating that our results extend to non-linear models will be
presented in section 4.
3 Theoretical results in the linear Gaussian case
We now address the origin of the saturation effect in the estimated RMSE in more detail and claim that it can be understood using results which we obtain in the setting of linear
Gaussian state-space models. Before providing mathematical evidence supporting our claim, we begin by recalling
standard results for the derivation of the filtered probability
density function in the linear Gaussian case, which is the
framework of the famous Kalman filter.
3.1 Kalman filter framework
The framework of the Kalman filter [22, 23] is obtained by
considering the special case where transition and measurement functions are linear, and assuming that noise and prior
probability density functions are Gaussian distributed with
respect to the Lebesgue measure. The HMM equations then
read
xn = An xn−1 + εn
(18)
yn = Bn xn + ηn
where {An }n≥1 and {Bn }n≥1 are referred to as transition and
measurement matrices, and with εn ∼ N(0, Qn ) and ηn ∼
N(0, Rn ). In this situation, given a prior x̂0 ∼ N(µ0 , Σ0 ), the
filtering problem can be solved exactly: analytical expressions can be obtained for the posterior x̂n ∼ N(µn , Σn ) which
is fully characterized by its mean value µn and (symmetric
positive definite) covariance matrix Σn . Their expressions
are provided by the recursive formulae
−1
Σn−1 = Σn|n−1
+ BTn R−1
n Bn
−1
µn|n−1 + BTn R−1
Σn−1 µn = Σn|n−1
n yn
,
(19)
with µn|n−1 = An µn−1 and Σn|n−1 = An Σn−1 ATn + Qn (see
Appendix B for a derivation). A simple way of obtaining
these results is to make use of the Bayesian conjugacy property of Gaussian distributions: in other words, given a Gaussian prior and a Gaussian likelihood, the posterior is also
Gaussian. Applying Bayes rule to the predicted and posterior probability measures, respectively as
πn|n−1 (dxn |y1:n−1 ) =
Z
X
fn (dxn |xn−1 )
× πn−1 (dxn−1 |y1:n−1 ) ,
(20)
A critical analysis of resampling strategies for the regularized particle filter
5
1
RMSE
RMSE
1
0,1
0,01
1
10
n
0,1
0,01
1
100
10
100
n
Fig. 1 Left frame: Normalized root mean squared error (RMSE) of the estimated state as a function of the number of observations in logarithmic
scale for N = 1000 particles and normalized observation noise variance R/Σ0 = 0.25. The black solid line is the expectation based on the Kalman
filter, while the red solid line is the outcome of the RPF with resampling at each step. The latter clearly shows a saturation effect at a value given
by the red dotted line, which we compute analytically in section 3. Blue lines are for N = 10000 particles and green lines for R/Σ0 = 1. The black
dashed line is obtained by taking into account the bandwidth perturbation inside the√Kalman filter, with a quench in normalized observation noise
variance from 0.25 to 1 at n = 300. Other parameters: prior mean value µ0 = x0 + Σ0 . Right frame: Likewise, for different bandwidth selection
criteria: Silverman’s rule of thumb h2n = (4/N)2/3 Σn (dark green), direct plugin method by Sheather and Jones [33] (orange), West shrinkage [36]
(cyan) and our proposal of kernel bandwidth modulation αn = αh /(1 + nαh ) (violet).
and
gn (yn |xn )πn|n−1 (dxn |y1:n−1 )
,
X gn (yn |xn )πn|n−1 (dxn |y1:n−1 )
πn (dxn |y1:n ) = R
(21)
thus allows for a straightforward derivation of Eq. (19). We
next focus on the special case where all states are assumed
stationary. This corresponds to taking Ak = I and Qk = 0
for k ≥ 1 in Eq. (18). If we further fix the measurement covariance matrices to unity Bk = I and assume that observation noise covariance matrices are time-independent, one
can show using the matrix inversion lemma that
Σn = R(R + nΣ0 )−1 Σ0
µn = Σn Σ0−1 µ0 + (I − Σn Σ0−1 )hyin
(22)
with hyin := ∑nj=1 y j /n the maximum likelihood estimator.
This situation corresponds to a multi-dimensional generalization of the stationary model we considered earlier. Note
that Σn reaches the Cramer-Rao bound R/n as n → ∞, while
limn→∞ µn = x0 : we thus retrieve that, in accordance with
the Bernstein-von Mises theorem, the Bayes estimator x̂n
is asymptotically efficient and coincides in that limit with
the maximum likelihood estimator, as the dependence on
the choice of prior vanishes. The corresponding RMSE =
O(n−1/2 ) also vanishes asymptotically. This behavior might
seem like a natural byproduct in the limit of a large number
of observations, since the latter might be expected to eventually provide sufficient information for an arbitrarily accurate estimation of the true state. There are, however, elementary counterexamples to this fact which can be very easily
proven using the Kalman filter equations we have derived.
For example, if we restore a simple time-independent transition matrix A = aI to the above estimation paradigm, one
can easily show that the asymptotic behavior of Σn depends
on the magnitude of a: while if a < 1, the covariance matrix decays exponentially fast, if a > 1 however, then the covariance matrix remains lower bounded as Σn ≥ (1 − a−2 )R.
Note that if a = 1, we recover the algebraic dependence as
stated in Eq. (22). Another enlightening case is when state
noise Q 6= 0 is included. Given time-independent state and
observation noises, one can show that Σn necessarily remains lower bounded by a positive matrix [4].
3.2 Regularized RMSE asymptotics
Let us now investigate how the above results can be transposed in the framework of the RPF. For a Gaussian regularizing kernel, Bayesian conjugacy ensures that the posterior
distribution Eq. (13) is asymptotically Gaussian:
pNn (xn |y1:n ) =
1
N
(i)
(i)
∑Ni=1 gn (yn |xn|n−1 )Khn (xn − xn|n−1 )
1
N
( j)
∑Nj=1 gn (yn |xn|n−1 )
0
N
0
0
0
X λ (dxn )pn|n−1 (xn |y1:n−1 )gn (yn |xn )Khn (xn − xn )
R
N
0
0
0
X λ (dxn )pn|n−1 (xn |y1:n−1 )gn (yn |xn )
R
≈
N→∞
,
(23)
with mean
N
N
N
−1 N
T −1
(Σn|n−1
)−1 + BTn R−1
n Bn µn = (Σ n|n−1 ) µn|n−1 +Bn Rn yn
(24)
6
Pierre Carmier et al.
and covariance matrix
−1
N
ΣnN = h2n + (Σn|n−1
)−1 + BTn R−1
n Bn
−1
N
= (1 + αh ) (Σn|n−1
)−1 + BTn R−1
,
n Bn
(25)
Lemma 1 Consider the recursively defined set of equations
for n ≥ 1 on the couple of variables (µn , Σn ),
(26)
Σn = (1 + αn )R (R + Σn−1 )−1 Σn−1
where {αn }n≥1 is a sequence of positive real numbers, and
R and {y}n≥1 are defined as in Eq. (22). Starting from the
initial condition (µ0 , Σ0 ), the solution to the above set of
equations can be expressed as
−1
j−1
Σn = ∏nj=1 (1 + α j )R R + ∑nj=1 ∏k=0
(1 + αk )Σ0
Σ0
µn =
1
Σn Σ0−1 µ0
n
(1
∏ j=1 + α j )
1
Σn Σ0−1
+ I− n
(1
∏ j=1 + α j )
!
1
Σ N Σ −1 µ0
(1 + αh )n n 0
n
(1 + αh ) j−1 y j
1
N −1 ∑ j=1
Σ
Σ
+ I−
n
n
0
(1 + αh )n
∑ j=1 (1 + αh ) j−1
µnN =
N
N
N
N AT + Q , and
with µn|n−1
= An µn−1
and Σn|n−1
= An Σn−1
n
n
where αh is defined in Eq. (14). Note that, as N → ∞, we recover the standard Kalman result from Eq. (19) as we should,
since the kernel acts as a Dirac delta-function in Eq. (23)
as αh → 0. Additionnally, while we work in the limit of
N → ∞ and choose to conserve the N-dependence of the kernel bandwidth, we explicitly assume the Monte Carlo error
∝ N −1/2 is negligible when replacing the sum by the integral in Eq. (23). We now proceed to illustrate the consequences of Eqs. (24,25) by returning for simplicity to the
case of stationary states, where exact expressions for the
posterior moments can be obtained by solving Eqs. (24,25)
recursively. For reasons which will be clear later, we shall
allow the possibility for αh to be time-dependent. Note that
our analysis should carry over to any dynamical system described by Eqs. (18) for which the Bayes estimator is asymptotically efficient: this includes systems with noiseless and
compact supported state dynamics, and for which the observation model is regular in some sense [5].
−1
−1
Σn−1
+ R−1 µn = Σn−1
µn−1 + R−1 yn
indicated in Eq. (14): αn = αh ∀n ≥ 1. Expressions for the
posterior moments simplify as
−1
ΣnN = (1 + αh )n R R + ∑nj=1 (1 + αh ) j−1 Σ0
Σ0
j−1
∑nj=1 ∏k=0 (1 + αk )y j
j−1
∑nj=1 ∏k=0 (1 + αk )
(27)
with α0 = 0.
Proof This can be straightforwardly proven by recurrence.
t
u
As a first application of lemma 1, consider the case where
the sequence {αn }n≥1 is time-independent. This is our original motivation, corresponding to the choice of bandwidth
(28)
We obtain as a direct consequence that, if αh > 0 (for finite
N), the covariance matrix does not converge to zero as the
number of observations increases.
Theorem 1 (RPF fixed point) The RPF estimator for the
posterior covariance matrix converges exponentially fast towards a non-vanishing fixed point:
lim Σ N
n→∞ n
= αh R(1 + O(e−nαh )) .
(29)
Likewise, the estimated mean value converges exponentially
fast towards a random variable
µnN ≈
n→∞
∑nj=1 (1 + αh ) j−1 y j
1 + O(e−nαh )
n
j−1
∑ j=1 (1 + αh )
(30)
centered around the observation mean value xn = x0 , with
residual covariance
αh
lim E[(µnN − xn )(µnN − xn )T ] =
R.
(31)
n→∞
2 + αh
√
As a result, the asymptotic RMSE = O( αh ) remains lower
bounded by a positive constant as soon as the number of
particles is finite, no matter how many observations.
As we have already mentioned, this situation can also occur in the original Kalman filter (retrieved in the limit of
an infinite number of particles) as soon as process noise
Q 6= 0 is included. In the present case, the reason for this
phenomenon is that the kernel bandwidth used for resampling acts as a perturbation which increases the variance and
competes with the variance reduction due to each new observation. The problem is that the variance reduction rate with
ΣnN is faster (quadratic) than that of the bandwidth (linear):
N − Σ N = (α R − Σ N )(R + Σ N )−1 Σ N . Arrives therefore
Σn+1
h
n
n
n
n
a moment when the ever smaller piece of information provided by the new observation is exactly compensated by the
bandwidth perturbation.
A complementary and enlightening interpretation of our
results can be put forward by noting that each observation
in Eq. (30) contributes differently to the estimated posterior
mean value, i.e. each y j carries a different weight. In particular, the weighting ratio between first (oldest) and last (most
recent) observations roughly scales like e−nαh , which signifies that the RPF estimator loses memory of previous observations exponentially fast. This loss of memory is directly
A critical analysis of resampling strategies for the regularized particle filter
responsible for the inability of the estimator to process the
information provided by the full history of observations as
exhaustively as the Kalman estimator, since it acts as an effective temporal resolution. The lower bound on the RMSE
obtained in theorem 1 is thus simply the translation of this
temporal resolution in state space.
Returning to Fig. 1, we check the validity of our results by comparing them with numerical simulations of the
one-dimensional stationary HMM introduced in section 2.2.
Interestingly, although conclusions of theorem 1 based on
Eqs. (24,25) were derived in the asymptotic limit N → ∞,
numerical simulations support their well-foundedness already
for N = 1000 particles. The late time fluctuations in the
RMSE of the RPF correspond to those of mean value in
Eq. (31). Next, because limN→∞ αh = 0, we expect according to theorem 1 the lower bound to decrease as N increases.
This is clearly illustrated by the blue curve in the left frame
of Fig. 1. Increasing the observation noise variance obviously leads to the opposite effect (green curve). We plotted as a dashed black line the analytical prediction based on
the Kalman filter from Eq. (28), where the observation noise
variance R/Σ0 was quenched from the value 0.25 to 1 at
n = 300. As one can see, the computed posterior variance on
the state reacts to the quench by increasing and converging
towards the new asymptotic fixed point. Notice that increasing dx is also detrimental (not shown), since limdx →∞ αh = 1
according to Eq. (14). Next, we show in the right frame
that other bandwidth selection criteria such as Silverman
and Scott’s rule of thumbs or the direct plug-in method by
Sheather and Jones [33] lead to the same problem when resampling is performed at each step. Another natural strategy is to perform resampling only occasionally. However,
resampling periodically every p step only allows reducing
the location of the fixed point roughly a factor 1/p, as is
shown in the following corollary.
Corollary 1 Given a resampling period p ∈ N∗ and an integer 0 ≤ q < p, the RPF estimator for the posterior covariN
ance matrix Σnp+q
converges exponentially fast towards the
non-vanishing q-dependent solution:
lim Σ N
n→∞ np+q
=
αh R
(1 + O(e−nαh )) .
p + αh q
(32)
As a result, the asymptotic covariance matrix fluctuates between the upper bound αh R/p and the lower bound αh R/(p+
(p − 1)αh ). Likewise, the estimated mean value converges
exponentially fast towards a random variable centered around
the observation mean value xn , with a q-dependent residual
covariance
N
N
lim E[(µnp+q
− xn )(µnp+q
− xn )T ]
n→∞
=
αh p + αh (2 + αh )q
R.
2 + αh (p + αh q)2
p
This translates in an asymptotic RMSE = O( αh /p).
7
Proof This can be proven by deriving from Eq. (27) the
new expression for the posterior covariance matrix when
αnp+q = αh δq,0 :
!−1
np+q
N
Σnp+q
n
= (1 + αh ) R R +
[ j−1] p
∑ (1 + αh )
Σ0
Σ0 ,
j=1
(34)
where we defined [ j − 1] p as the integer part of ( j − 1)/p .
Similar calculations allow obtaining the corresponding expression for the mean value:
1
Σ N Σ −1 µ0
(1 + αh )n np+q 0
np+q
(1 + αh )[ j−1] p y j
1
N
−1 ∑ j=1
Σ
Σ
+ I−
.
np+q
(1 + αh )n np+q 0
∑ j=1 (1 + αh )[ j−1] p
N
µnp+q
=
(35)
t
u
That the RMSE remains lower bounded by a positive constant when resampling is performed periodically is illustrated numerically in the left frame of Fig. 2, where the analytical lower bound (32) is indicated by the dotted line.
3.3 Kernel bandwidth modulation
One is thus left with two possibilities in order to allow the
estimator to keep up with the Kalman filter prediction: either resample less and less often with time (which seems
compatible with the idea of resampling only if the ESS falls
lower than a critical threshold, see subsection 3.4), or choose
the bandwidth parameter differently so that consistency of
the estimator is maintained despite performing resampling
at each time. To begin with, it is clear from theorem 1 that
choosing {αn }n≥1 such that limn→∞ αn = 0 enforces that
limn→∞ RMSE(x̂n , xn ) = 0 (this can easily be proven by contradiction). This idea was actually hinted at in [7]. We proceed one step further by characterizing in theorem 2 which
sequences ensure that the rate of convergence of the RPF
estimator is optimal in the sense defined below.
Definition 1 (Optimal convergence rate) The rate of convergence of an estimator x̂n towards the true state xn as the
number of observations n → ∞ will be called optimal if it
is comparable with that given by the inverse of the Fisher
information matrix: RMSE(x̂n , xn ) = O(n−1/2 ).
(33)
Since the Kalman filter estimator is asymptotically efficient,
this implies in particular that the optimal rate of convergence
is that given by the Kalman filter.
8
Pierre Carmier et al.
1
1
RPF, ESScrit = 1
SIR, ESScrit = 1
RPF, ESScrit = 0.9
RPF, ESScrit = 0.1
SIR, ESScrit = 0.1
SIS
RMSE
RMSE
p=2
SIR, ESScrit = 1
RPF, ESScrit = 0.9
SIR, ESScrit = 0.1
SIS
0,1
0,1
0,01
1
10
n
100
1
10
100
n
Fig. 2 Left frame: Comparison of RPF and SIR simulations for different values of the resampling ratio. Sequential importance sampling (SIS)
is a special case of the SIR algorithm when no resampling is performed (ESScrit = 0). The Kalman filter expectation is provided for reference
(solid black line). Estimation based on the RPF with p = 2 periodic resampling is also displayed (solid maroon) and shown to saturate around
the analytically predicted solution (dotted line). Right frame: Same but for a prior location µ0 three standard deviations away from the mean
observation. This time, the SIR approximation breaks down, as opposed to the RPF which remains well-behaved.
Theorem 2 Let {αn }n≥1 be a sequence of positive numbers
as in Eq. (27). If αn = O(n−1 ), the RPF estimator converges
at an optimal rate. In contrast, given any ε ∈]0, 1/2[, if αn ≈
nε−1 , then ∃ c ∈ R+ ∀n ≥ 1 RMSE ≥ cn(ε−1)/2 .
2R/n, and for the mean value
Proof Suppose that ∃ c ∈ R+ ∀n ≥ 1 nαn ≤ c. Then, using that x − x2 /2 < log (1 + x) < x for x > 0, one has that
∏nj=1 (1 + α j ) = O(nc ), and thus also that ∑nj=1 ∏k−1
j=0 (1 +
c+1
α j ) = O(n ). Plugging this in Eq. (27) implies that ΣnN =
O(R/n) and therefore that the RPF estimator converges at
an optimal rate. In contrast, if αn ≈ nε−1 , then the same reasoning implies that ∏nj=1 (1 + α j ) = O(exp (nε /ε)). Making
use of the following integral inequality
centered around the observation mean value xn and with
residual covariance E[(µnN − xn )(µnN − xn )T ] ≈ 4R/(3n).
n→∞
Thus, optimal convergence rate is recovered, as expected
from theorem 2. Interestingly, as compared to Eq. (30), the
dependence of Eq. (37) on the observations is long-ranged:
the weighting ratio between the last and the first observations scales linearly with n instead of exponentially as in
Eq. (30). This illustrates that consistency of the regularized
estimator requires that memory of past observations should
be kept long enough. Note that such a requirement, while
understandable for parameter estimation, may be irrelevant
for tracking applications where memory of the past might be
considered relatively useless.
n−1
nε−1
n−1
jε
jε
∑ exp ( ε ) ≤ ∑ ( j + 1)ε−1 exp ( ε )
j=0
j=0
≤
Z n
0
(36)
nε
xε
dx xε−1 exp ( ) ≤ exp ( ) ,
ε
ε
we deduce that ∃ c ∈ R+ ∀n ≥ 1 ΣnN ≥ cnε−1 R. Finally, notε
ε
ing that ∑n−1
j=0 exp (2 j /ε) ≥ exp (2(n − 1)/ε), we conclude
+
(ε−1)/2
that ∃ c ∈ R ∀n ≥ 1 RMSE ≥ cn
.
t
u
This result still leaves us with a wide range of sequences at
our disposal. Finding a suitable statistical criterion both justifying this type of kernel bandwidth modulation (and therefore different from the mean integrated squared error used to
derive Eq. (14)) and allowing to identify a sequence which
is optimal in some sense remains an open question. Let us
look at two examples for the sequence {αn }n≥1 .
• Example 1: αn = (n + αh−1 )−1 .
For this choice, products and sums in Eq. (27) can be computed exactly and we get for the covariance matrix ΣnN ≈
n→∞
µnN ≈
n→∞
2
n2
n
∑ jy j (1 + O(n−2 ))
(37)
j=1
• Example 2: αn = αh e−nαh .
In this case, one cannot compute sums and products in Eq. (27)
exactly but one can easily show that ∏nj=1 (1+αh e− jαh ) converges to a non-zero limit using that it is a strictly increasing
bounded sequence:
n
1 ≤ ∏ (1 + αh e− jαh ) ≤ exp (
j=1
αh
).
−1
e αh
(38)
j−1
This result allows immediately deducing that ∑nj=1 ∏k=0
(1+
αh e−kαh ) = O(n), and thus that ΣnN = O(R/n) and that the
mean value
j−1
µnN ≈
n→∞
∑nj=1 ∏k=0 (1 + αh e−kαh )y j
j−1
∑nj=1 ∏k=0 (1 + αh e−kαh )
1 + O(n−1 )
(39)
A critical analysis of resampling strategies for the regularized particle filter
is centered around the observation mean value xn with residual covariance E[(µnN −xn )(µnN −xn )T ] = O(R/n), using that
j−1
∏k=0 (1 + αh e−kαh )2 is also bounded. Optimal convergence
rate is thus once more recovered, this time with a weighting
ratio between different observations bounded by ∏∞j=1 (1 +
αh e− jαh ).
This strategy of modulating the bandwidth by enforcing
it to decrease fast enough is positively supported by numerical simulations in the right frame of Fig. 1. It thus constitutes a functional method to overcome the limitations in the
RPF we discussed earlier. Note that a different strategy discussed in Ref. [36] consists in counterbalancing the variance
originating from the bandwidth perturbation by rescaling the
particle values towards their center of mass. Centering the
(i)
kernel densities around the shrunken values axn|n−1 + (1 −
√
a)µnN , and taking a = 1 − αh , one can indeed verify that
the posterior density in the RPF asymptotically coincides
with that expected from the Kalman filter approach.
3.4 Strategies based on the resampling ratio
We now examine another possibility, which is to perform
resampling only if the ESS introduced in Eq. (10) becomes
lower than a determined threshold which we shall refer to as
the resampling ratio ESScrit [15]. In the left frame of Fig. 2,
we compare the posterior RMSE for different values of this
threshold, both for the RPF and the SIR algorithm. As one
can clearly see, as soon as the resampling ratio is smaller
than 1, the RPF yields results in agreement with the exact
solution. The SIR algorithm provides comparable results,
whatever the value of the resampling ratio. However, if the
location µ0 of the prior density is taken farther away from
the correct value, the situation changes since estimating the
posterior mean value µn accurately will increasingly depend
on the ability of particles to explore phase space well enough
to diffuse towards the correct value. This is illustrated
√ in the
right frame of Fig. 2, where µ0 was taken √
to be 3 Σ0 away
from the observed mean value, instead of Σ0 until now. In
this case, the SIR algorithm almost always fails to correctly
estimate the posterior variance due to sample impoverishment arising as a direct consequence of severe weight degeneracy. In fact, the estimated variance often goes to zero,
reflecting the fact that only a single particle eventually survives the resampling process. On the other hand, the RPF
proves very robust in this situation and provides correct results as before, despite the fact that its RMSE has a tendency
to underestimate the optimal Kalman expectation. Note that
the curves displayed on Fig. 2 are outputs of single particle
filter simulations (no averaging is performed).
Resampling conditional on a sub-unit value of ESScrit
thus also appears to resolve the RPF shortcomings discussed
earlier. One can actually show that the resampling frequency
9
for a sub-unit resampling ratio decreases exponentially fast
with the number of observations, thereby justifying that the
posterior RMSE should no longer be lower bounded by a
positive constant. We provide below a formal proof of this
statement in the case of the stationary HMM.
Lemma 2 Assume that ESScrit < 1. Then the spacing between consecutive resampling times increases exponentially
fast with the number of observations for a stationary HMM.
Proof Our starting point is the observation that
2
(i)
1 N
n
g
(y
|x
)
∏
∑
N i=1 j=1 j j j| j−1
ESSn
lim
= lim
(i)
n
N→∞ N
N→∞ 1 N
2
N ∑i=1 ∏ j=1 g j (y j |x j| j−1 )
R
2
n
π
(dx
)
g
(y
|x
)
f
(dx
|x
)
n+1
∏
j
j
j
j
j
0
0
j−1
j=1
X
.
= R
n
2
π
(dx
)
n+1
∏
0
0
j=1 g j (y j |x j ) f j (dx j |x j−1 )
X
(40)
For simplicity, we consider the case of the one-dimensional
stationary state, where integration with respect to the Markov
kernels reduces to ∏nj=1 g j (y j |x0 ). It is then a simple task to
compute the effective sample size:
p
γn (2 + γn )
ESSn
hµ0 − yi2n
lim
=
exp −
N→∞ N
1 + γn
Σ0 (1 + γn )(2 + γn )
p
hµ0 − yi2n
.
2γn exp −
≈
n→∞
2Σ0
(41)
We see that ESSn depends sensitively on how close the ML
hyin is from the typical particle value µ0 and on the quantity γn := R/(nΣ0 ). The number of observations essentially
reduces the effective observation noise variance. If resampling is performed at step n, then one can express the ESS at
step n + m asymptotically as
p
n(2m + n)
ESSn+m
≈
.
(42)
n,m→∞
N
m+n
The value of m required for the ESS to fall below a critical
threshold ESScrit N, with ESScrit ∈ [0, 1], can then easily be
calculated and we get
m
ESSn+m
< ESScrit ⇔ > βcrit
N
n
with
q
1 − ESS2crit
q
βcrit =
.
1 − 1 − ESS2crit
(43)
(44)
Iterating, we thus see that the number of steps mk asymptotically required to perform resampling for the kth time grows
exponentially fast:
k−1
mk = βcrit (n + ∑ m j ) = (1 + βcrit )mk−1
j=1
k−1
= βcrit (1 + βcrit )
n.
(45)
10
Pierre Carmier et al.
1
1
0,8
ESS
RMSE
0,6
1
0,4
0,1
ESS
0,8
0,2
0,6
0,4
0,2
0
1
1
10
10
n
ESScrit=1
ESScrit=0.5
ESScrit=0.9
ESScrit=1(SIR)
ESScrit=1(oracle)
ESScrit=0.9(oracle)
αh/(1+nαh)
5
ESScrit=0.5, N=10
100
100
1
Fig. 3 Main: Normalized effective sample size (ESS) for N = 1000
particles as a function of the number of noiseless observations n in logarithmic scale for the stationary HMM. The exponential growth of the
characteristic time required for the ESS to fall below a given threshold (here ESScrit = 0.5 indicated by the dashed line) is manifest. Inset:
Same but for noisy observations. This time the exponential growth is
somewhat less apparent due to the random fluctuations.
10
n
Fig. 4 Normalized root mean squared error of the estimated parameter
a in the logistic map as a function of the number of observations in
logarithmic scale for N = 1000 particles and normalized observation
noise variance R/Σ0 = 0.33. The saturation in the RMSE (solid red)
testifies that the effect predicted for linear models remains valid in nonlinear cases. Other strategies comparably improve the estimation.
(46)
t
u
Note that similar conclusions were reached in [8]. We provide numerical simulations supporting this claim in Fig. 3.
Interestingly, we compared the results we obtain for noisy
observations (inset) with that for noiseless observations (symbols in main frame). The latter can be interpreted as representing the most favourable scenario, in which observations
are always equal to the noiseless signal, and shall therefore
be referred to from now on as the oracle limit, by analogy
with the influential decision-theoretic concept [3]. As one
can notice in the inset, the exponential growth of the resampling step sizes can be occasionally violated if observations
cause large fluctuations in the exponential term of Eq. (41).
These fluctuations also explain why the ESS is not necessarily monotonous between consecutive resampling times for
noisy observations.
Equipped with lemma 2, we can now conclude our demonstration that the RPF estimator is well-behaved as soon as
ESScrit < 1.
Theorem 3 Assume that ESScrit < 1. Then the rate of convergence of the RPF estimator is optimal in the sense of definition 1.
Proof According to the previous lemma, we know that the
spacing between consecutive resampling times increases exponentially fast, with rate 1 + βcrit . As a consequence,
log (1+α )
n
log n
h
log (1+βcrit )
= O n log (1+βcrit ) ,
∏ (1 + α j ) = O (1 + αh )
j=1
100
and thus ∃ c ∈ R+ ∀n ≥ 1 ∏nj=1 (1 + α j ) = O(nc ). The same
reasoning as that followed in the proof of theorem 2 allows
us to conclude.
t
u
4 Extension to non-linear models
We now provide numerical evidence that the above results
qualitatively hold in the case of non-linear models. We shall
consider two examples: the logistic map, which is notably
famous for possessing regions of parameter space displaying
chaos, and a multi-dimensional plant growth model.
4.1 Logistic map
The logistic map, popularized by Robert May [29], is a discretized version of the logistic equation first introduced by
Verhulst in the 19th century to describe population dynamics
with a finite carrying capacity. It is described by the deceptively simple equation xn+1 = axn (1−xn ), with X = [0, 1] for
a ∈ [0, 4]. The nature of the state dynamics can change drastically depending on the value of a: it converges to a stationary solution for 0 ≤ a ≤ 3, oscillates between multiple values for 3 < a ≤ ac , and displays chaotic behavior for a > ac ,
where ac ' 3.57, with islands of stability found amid. The
effect of the initial condition, on the other hand, is often of
litte relevance and shall be assumed known to us: x0 = 0.5.
In the following, we seek to estimate a given multiplicative
A critical analysis of resampling strategies for the regularized particle filter
11
4
2000
rn
µRUE
3,8
1000
3,6
3,4
0
100
50
3,2
150
100
50
1
0,8
σRUE
ln
300
200
100
0
0,6
0,4
0,2
50
n
100
150
0
50
n
100
Fig. 5 Upper left: Observations of sugarbeet root mass rn (noisy = symbols, noiseless = thick line). Lower left: Likewise for leaf mass ln . Upper
right: Evolution of the mean value of the estimated parameter RUE in the LNAS model after each new observation at time n, for ESScrit = 0.5 and
various particle numbers: N = 1000 (blue), N = 104 (brown), N = 105 (black) and N = 106 (yellow). Convergence of the particle filter estimation
towards the true posterior mean value is apparent, as a comparison with a SIR simulation testifies (N = 106 , dashed red). Lower right: Likewise
for the normalized standard deviation.
noisy observations yn = xn η, with η ∼ log N(1, R). Starting
from a prior distribution a(i) ∼ N(3.0, 0.3), the resulting inference of parameter a given observations generated with the
value a∗ = 3.33 and R = 0.1 is displayed in Fig. 4. Saturation of RMSE(ân , a∗ ) when resampling is performed at each
step is clearly apparent (solid red). This effect disappears using either of the resampling strategies we discussed earlier,
however, as the comparison with a benchmark simulation
for N = 105 particles (solid black) testifies. Interestingly, we
compared the results we obtain for noisy observations (thick
lines) with that in the oracle limit (dotted lines). The RMSE
obtained in the latter case acts as a lower bound with respect
to that obtained for noisy observations, with numerical simulations even indicating that limn→∞ RMSE(ân , a∗ ) = 0 in
this case for ESScrit < 1.
If instead we choose to work in the chaotic parameter
region and take a∗ > ac , our numerical results (not shown)
seemingly point to the fact that correctly inferring the value
of a∗ is always challenging, even in the oracle limit. The
same can be told for the estimation of the initial condition
x0 , for a fixed known value of a. The reason for this behavior
seems to stem from the fact that, in such situations, the information provided by observations on the quantity of interest
is a (rapidly) decreasing function of time. This loss of memory on the initial condition ultimately limits the resolution
with which inference can be carried, as concentration results
recently obtained in a similar context may testify [32].
4.2 LNAS model
Next, we look at a more elaborate model, which describes
the growth of leaf mass ln and root mass rn of a sugarbeet
plant as a function of daily temperature Tn and radiation φn .
The latter constitute environmental variables, while the former are the hidden states. This model is a simplified version
of the LNAS model [11]. In particular, we consider deterministic transition functions. Leaf and root masses are updated as
ln
ln−1
an
=
+
qn ,
(47)
rn
rn−1
1 − an
where qn is the daily produced biomass and an is an allocation function. Biomass production is governed by the BeerLambert law
−1
qn = RUE φn 1 − e−ρ ln ,
(48)
where RUE is the radiation use efficiency and ρ −1 the leaf
mass density. Allocation depends on the thermal time τn as
γ
1
τn
an =
1 − erf[ √
log ( )] ,
(49)
2
µa
2σa
where γ controls early biomass partition, erf is the error
function and (µa , σa ) are parameters governing the shape of
the allocation function. Thermal time increases monotonically as a function of temperature as τn = τn−1 + max[0, Tn ].
The system is initialized as τ0 = r0 = 0, and l0 6= 0 such
that ρ −1 l0 1. For environmental variables, we use real
data obtained from a weather station in France. The filtering problem in the present case then consists in estimating
the parameters (RUE, γ, µa ), given independent prior den(i)
sities RUE(i) ∼ N(3.8, 0.3), γ (i) ∼ N(0.7, 0.05) and µa ∼
N(500, 50), and given multiplicative noisy observations of
the hidden states ln η and rn η with η ∼ log N(1, R). We shall
assume ρ and σa known beforehand.
Because we no longer have access to the optimal posterior density, as opposed to the case of the stationary model,
12
Pierre Carmier et al.
650
0,68
600
3,6
0,66
3,4
0,64
550
500
0,62
3,2
0,6
100
50
450
100
50
1
1
2
0,8
0,8
1,5
0,4
0,2
0
0,4
n
100
0
1
0,5
0,2
50
100
50
a
0,6
σµ
0,6
σγ
σRUE
µµ
µγ
µRUE
4
a
0,7
3,8
50
n
100
0
50
n
100
Fig. 6 Upper left: Evolution of the mean value of the estimated parameter RUE in the LNAS model after each new observation at time n,
for N = 1000 particles and different resampling strategies: ESScrit = 0.5 (blue), αn = αh /(1 + nαh ) (violet), ESScrit = 0 (orange). Lower left:
Likewise for the normalized standard deviation. The bootstrap particle filter expectations (ESScrit = 1.0 (dashed red), ESScrit = 0.5 (dashed blue))
are clearly flawed with respect to the benchmark limit (N = 106 particles, solid yellow), while the RPF provides excellent results already for
N = 1000 particles. Middle: Likewise for γ. Right: Likewise for µa .
we plot the evolution of posterior standard deviations and
mean values of each parameter (instead of the RMSE) and
use results obtained by the RPF for N = 106 particles and
ESScrit = 0.5 as a benchmark. We checked that this corresponds to the optimal posterior density by plotting the results obtained for N = 106 particles using the SIR algorithm
by comparison, and indeed observed that convergence as a
function of N is achieved (see Fig. 5). These simulations
were carried by running our algorithms on the computing
mesocenter Fusion. Given observations generated with values (RUE∗ , γ ∗ , µa∗ ) = (3.56, 0.625, 550) and R = 0.1, the resulting inference for various numbers of particles and resampling strategies is displayed in Fig. 6. We see that the performance of the RPF for N = 1000 particles and ESScrit = 0.5
is already close to that obtained in the benchmark scenario.
Using our kernel bandwidth modulation strategy, we additionally demonstrate that equivalent results can be obtained
even when ESScrit = 1, whereas the traditional choice of
Eq. (14) leads to an unstable exploration of parameter space
(not shown). This holds independently of the chosen parameter. In contrast, expectations based on the bootstrap particle
filter or in the absence of any resampling are evidently unable to capture the correct behavior for the posterior variance. As a consequence, these particle filters will fail to
provide appropriate credibility regions for the parameters
and thereby make any reliable prediction impossible. This
is also illustrated in table 1 where we compared the performances of our various resampling strategies, each time averaged over 10 independent runs. These results thus clearly
demonstrate the ability of the RPF to outperform such traditional particle filters for a moderate number of particles.
5 Conclusions
To conclude, we demonstrated that the choice of resampling
strategy in the regularized particle filter can be paramount,
as far as estimation of the posterior probability density is
concerned. In particular, using analytical arguments inspired
from linear Gaussian state-space models, we proved that resampling on a systematic basis causes the RMSE to be lower
bounded by a constant which depends on the number of
particles and on the dimensionality of the state space. Numerical evidence supporting our claims, including for nonlinear state-space models, was also highlighted. Finally, we
showed that if one introduces a resampling ratio, or if the
kernel bandwidth is appropriately modulated, then the regularized particle filter clearly outperforms traditional bootstrap particle filters.
As a closing remark, we note that the idea we highlighted in this paper of progressively reducing the bandwidth
perturbation seems related to recent approaches known as
tempering [34], which consist in building a probability measure which depends on a tunable parameter allowing one to
smoothly transport particles from the prior π0 to the posterior πn . Similar ideas have also been invoked to perform
Bayesian update progressively [12], and investigating the
ramifications of such connections in more detail would certainly be worthy of interest. We also leave for future research
the question of which statistical criterion (Kullback-Leibler
divergence, Wasserstein metric, etc.) could justify this type
of kernel bandwidth modulation.
Acknowledgements The authors wish to thank Gautier Viaud for useful conversations at various stages of this work.
References
1. Andrieu, C., Doucet, A., Holenstein, R.: Particle Markov chain
Monte Carlo methods. J. Royal Stat. Soc. B 72(3), 269–342 (2010)
2. Campillo, F., Rossi, V.: Convolution particle filter for parameter
estimation in general state-space models. IEEE Transactions on
Aerospace and Electronic Systems 45(3), 1063–1072 (2009)
3. Candes, E.J.: Modern statistical estimation via oracle inequalities.
Acta Numerica 15, 257325 (2006)
A critical analysis of resampling strategies for the regularized particle filter
13
Table 1 Comparison of particle filter predictions for different resampling strategies at n = 100 averaged over 10 runs.
µRUE
σRUE
µγ
σγ
µ µa
σ µa
Benchmark
3.521
0.0325
0.632
0.0070
542.6
8.297
RPF (ESScrit = 0.5)
3.479 (0.0096)
0.0487 (0.0041)
0.639 (0.0026)
0.0146 (0.0016)
537.1 (2.37)
14.43 (1.69)
RPF (αn = (n + αh−1 )−1 )
3.505 (0.0291)
0.0341 (0.0083)
0.636 (0.0090)
0.0074 (0.0032)
540.4 (7.55)
9.996 (1.74)
4. Cappé, O., Moulines, E., Ryden, T.: Inference in Hidden Markov
Models. Springer Verlag (2005)
5. Carmier, P.: Bayesian inference in linear Gaussian hidden Markov
models: a tutorial on the Kalman filter. in preparation
6. Carpenter, J., Clifford, P., Fearnhead, P.: Improved particle filter
for nonlinear problems. IEE Proceedings - Radar, Sonar and Navigation 146(1), 2–7 (1999)
7. Chen, Y.: Bayesian inference in plant growth models for prediction
and uncertainty assessment. Ecole Centrale des Arts et Manufactures (2014)
8. Chopin, N.: A sequential particle filter for static models.
Biometrika 89, 539–552 (2002)
9. Chopin, N.: Central limit theorem for sequential Monte Carlo
methods and its application to Bayesian inference. Annals of
Statistics 32, 2385–2411 (2004)
10. Chopin, N., Jacob, P.E., Papaspiliopoulos, O.: SMC2 : an efficient
algorithm for sequential analysis of state space models. J. Royal
Stat. Soc. B 75(3), 397–426 (2013)
11. Cournede, P.H., Chen, Y., Wu, Q., Baey, C., Bayol, B.: Development and evaluation of plant growth models: Methodology and
implementation in the pygmalion platform. Mathematical Modeling of Natural Phenomena 8, 112–130 (2013)
12. Daum, F., Huang, J.: Particle flow for nonlinear filters with loghomotopy. Proc. SPIE 6969, 696,918 (2008)
13. DelMoral, P.: Feynman-Kac Formulae: Genealogical and Interacting Particle Systems with Applications. Springer (2004)
14. DelMoral, P., Doucet, A., Jasra, A.: Sequential Monte Carlo samplers. J. Roy. Stat. Soc. B 68(3), 411–436 (2006)
15. DelMoral, P., Doucet, A., Jasra, A.: On adaptive resampling strategies for sequential Monte Carlo methods. Bernoulli 18(1), 252–
278 (2012)
16. Doucet, A., de Freitas, N., Gordon, N.: Sequential Monte Carlo
methods in practice. Springer Verlag (2001)
17. Gauchi, J.P., Villa, J.P.: Nonparametric particle filtering approaches for identification and inference in nonlinear state-space
dynamic systems. Statistics and Computing 23, 523–533 (2013)
18. Gordon, N., Salmond, D., Smith, A.: A novel approach to
nonlinear/non-Gaussian Bayesian state estimation. IEE Proc. F
140, 107–113(6) (1993)
19. Hol, J., Schon, T., Gustafsson, F.: On resampling algorithms for
particle filters. In: IEEE Nonlinear Statistical Signal Processing
Workshop, pp. 79–82 (2006)
20. Ionides, E.L., Bhadra, A., Atchadé, Y., King, A.A.: Iterated filtering. Ann. Stat. 39(3), 1776–1802 (2011)
21. Jacob, P.: Sequential Bayesian inference for implicit hidden
Markov models and current limitations. ESAIM: Procs. 51, 24–48
(2015)
22. Kalman, R.: A new approach to linear filtering and prediction
problems. J. Basic Engineering 82, 35–45 (1960)
23. Kalman, R., Bucy, R.: New results in linear filtering and prediction
theory. J. Basic Engineering 83, 95–107 (1961)
24. Kantas, N., Doucet, A., Singh, S.S., Maciejowski, J.M., Chopin,
N.: On particle methods for parameter estimation in state-space
models. Statistical Science 30(3), 328–351 (2015)
SIR (ESScrit = 0.5)
3.554 (0.0431)
0.0126 (0.0161)
0.623 (0.0095)
0.0024 (0.0030)
547.2 (18.08)
4.764 (5.89)
SIS (ESScrit = 0)
3.503 (0.0569)
0.0060 (0.0049)
0.634 (0.0089)
0.0007 (0.0005)
540.4 (10.50)
2.477 (3.09)
25. Kong, A., Liu, J.S., Wong, W.H.: Sequential imputations and
Bayesian missing data problems. J. Americ. Stat. Assoc. 89(425),
278–288 (1994)
26. LeGland, F., Musso, C., Oujdane, N.: An analysis of regularized
interacting particle methods for nonlinear filtering. In: Proceedings of the 3rd IEEE European Workshop on Computer-Intensive
Methods in Control and Signal Processing, pp. 167–174 (1998)
27. Liu, J., West, M.: Combined parameter and state estimation in
simulation-based filtering (Ch. 10), in Sequential Monte Carlo
methods in practice. Springer Verlag (2001)
28. Liu, J.S., Chen, R.: Sequential Monte Carlo methods for dynamic
systems. J. Americ. Stat. Assoc. 93(443), 1032–1044 (1998)
29. May, R.: Simple mathematical models with very complicated dynamics. Nature 261, 459–467 (1976)
30. Musso, C., Oujdane, N., LeGland, F.: Improving regularized particle filters (Ch. 12), in Sequential Monte Carlo methods in practice.
Springer Verlag (2001)
31. Parzen, E.: On estimation of a probability density function and
mode. Ann. Math. Stat. 33(3), 1065–1076 (1962)
32. Paulin, D., Jasra, A., Crisan, D., Beskos, A.: On concentration
properties of partially observed chaotic systems. unpublished
(2016). URL https://arxiv.org/abs/1608.08348
33. Sheather, S.: Density estimation. Statistical Science 19, 588–597
(2004)
34. Svensson, A., Schön, T.B., Lindsten, F.: Learning of state-space
models with highly informative observations: a tempered sequential Monte Carlo solution. unpublished (2017). URL https:
//arxiv.org/abs/1702.01618
35. Tsybakov, A.: Introduction to Nonparametric Estimation.
Springer (2009)
36. West, M.: Approximating posterior distributions by mixtures. J.
Roy. Stat. Soc. B 55(2), 409–422 (1993)
A Optimal bandwidth selection
There are various ways [33] of choosing the bandwidth parameter hn
in Eq. (13). The most popular choice relies on minimizing the mean
integrated squared error (MISE) between the true posterior probability
density function p and the kernel density pNh (x) = ∑Ni=1 Kh (x − x(i) )/N,
with x(i) ∼ p(x):
MISE(h) = E ||pNh − p||2L2
= ||E[pNh ] − p||2L2 +
Z
dz V[pNh (z)] ,
(50)
pR
dz p(z)2 . As
where we introduced the L2 (R) norm p 7→ ||p||L2 =
explicited above, the MISE can allegedly be decomposed into the inteN
Rgrated squared bias and the integrated variance. Observing that E[ph (x)] =
dz p(z)Kh (x − z) and developing p around x as a Taylor series, one
obtains:
E[pNh (x)] − p(x) =
h2 00
p (x)
2
Z
dz z2 K1 (z) + o(h2 ) .
(51)
14
Pierre Carmier et al.
Likewise, the variance can be computed by decomposing E[pNh (x)2 ]
into diagonal and non-diagonal contributions:
E[pNh (x)2 ] =
1
N
Kn = Σn|n−1 BTn (Bn Σn|n−1 BTn + Rn )−1 ,
Z
dz p(z)Kh (x − z)2
Z
2
N(N − 1)
+
dz
p(z)K
(x
−
z)
h
N2
(52)
such that V[pNh (x)] = (||K1 ||2L2 p(x) + o(1))/(Nh). Injecting these results in Eq. (50) and minimizing with respect to h is straightforward
and provides the optimal choice:
h5MISE =
||K1 ||2L2
N||p00 ||2L2 ( dz z2 K1 (z))2
R
.
(53)
Note that this implies that MISE ∝ N −4/5 . The only problem left is
that this choice depends on the second moment of the underlying true
probability density function which is generally unknown. There are
different strategies to estimate this quantity, which we will not delve
into [33]. If we assume that both p and K are Gaussian distributed,
5/2
the expression for h reduces to the simple h5MISE = 4Σ p /(3N) which
coincides with Eq. (14) for dx = 1.
B Kalman filter equations
The Kalman filter equations (19) can be derived recursively in two
steps, starting from the probability density function at time n − 1. The
first step is the prediction step, in which we seek to determine p(x̂n |y1:n−1 ),
in other words how the probability density function is propagated between two time steps. Using Eq. (18), one can easily show that
µn|n−1 := E[x̂n |y1:n−1 ] = An µn−1
(54)
using that state noise has null expectancy, and
Σn|n−1 := E[(x̂n − µn|n−1 )(x̂n − µn|n−1 )T ]
= An Σn−1 ATn + Qn
(55)
using that state noise is uncorrelated with the dynamics. The second
step is the update step, in which we seek to compute the posterior
moments µn and Σn , expressions for which are traditionally obtained
by enforcing unbiasedness of the estimator and minimizing the mean
squared error, as we will now show. We begin by writing the posterior
mean as a linear combination between the prior mean and the observation, that is µn = Kn0 µn−1 + Kn yn . Enforcing E[x̂n − µn |y1:n−1 ] = 0
leads to Kn0 = (I − Kn Bn )An , and thus:
µn = (I − Kn Bn )µn|n−1 + Kn yn
(56)
which expresses the posterior mean as a linear combination between
the predicted mean and the observation, with a tradeoff controlled by
the so-called Kalman gain matrix Kn , an expression for which will be
derived below. We begin by computing Σn , which follows directly from
the previous equation:
Σn = Cov[(I − Kn Bn )(x̂n − µn|n−1 ) + Kn ηn ]
= (I − Kn Bn )Σn|n−1 (I − Kn Bn )T + Kn Rn KnT .
(57)
Here we defined Cov[x̂] = E[x̂x̂T ] as a shorthand and used the fact
that observation noise is uncorrelated with the dynamics. The Kalman
gain is then obtained by minimizing the mean squared error, which
corresponds to the trace of the posterior covariance matrix:
0=
∂ Tr[Σn ]
∂ Kn
= −2Σn|n−1 BTn + 2Kn Bn Σn|n−1 BTn + 2Kn Rn ,
where derivation with respect to a matrix is performed element-wise.
Solving Eq. (58) yields the Kalman gain expression
(58)
(59)
where matrix inversion is justified by positive definiteness. Injecting
this in Eq. (57), we obtain a much simpler expression for the posterior
covariance matrix:
Σn = (I − Kn Bn )Σn|n−1 .
(60)
From Eqs. (59,60), we also derive the useful and synthetic relation
Σn BTn = Kn Rn . Finally, using this and rearranging Eqs. (56,60), we obtain the central result Eq. (19). Notice that we made no real use of the
assumption of Gaussian distributions. In fact, the above results hold
true irrespective of the underlying probability density function, provided noises are centered. The distinctive feature of the Gaussian case
is that the knowledge of the first two moments allows recovering the
entire distribution. Specifically, for Gaussian probability density functions, the Kalman filter equations can actually also be derived without
any particular assumptions using Bayes rule.
| 10math.ST
|
arXiv:1605.03778v2 [math.GR] 16 Nov 2016
Frobenius restricted varieties in numerical
semigroups
Aureliano M. Robles-Pérez∗† and José Carlos Rosales∗‡
Abstract
The common behaviour of many families of numerical semigroups led
up to defining, firstly, the Frobenius varieties and, secondly, the (Frobenius) pseudo-varieties. However, some interesting families are still out
of these definitions. To overcome this situation, here we introduce the
concept of Frobenius restricted variety (or R-variety). We will generalize
most of the results for varieties and pseudo-varieties to R-varieties. In
particular, we will study the tree structure that arise within them.
Keywords: R-varieties; Frobenius restricted number; varieties; pseudo-varieties;
monoids; numerical semigroups; tree (associated to an R-variety).
2010 AMS Classification: 20M14
1
Introduction
In [11], the concept of (Frobenius) variety was introduced in order to unify
several results which have appeared in [1], [3], [16], and [17]. Moreover, the
work made in [11] has allowed to study other notables families of numerical
semigroups, such as those that appear in [7], [9], [12], and [13].
There exist families of numerical semigroups which are not varieties but
have a similar structure. For example, the family of numerical semigroups with
maximal embedding dimension and fixed multiplicity (see [15]). The study of
this family, in [2], led to the concept of m-variety.
In order to generalize the concepts of variety and m-variety, in [8] were
introduced the (Frobenius) pseudo-varieties. Moreover, recently, the results
obtained in [8] allowed us to study several interesting families of numerical
semigroups (for instance, see [10]).
∗ Both authors are supported by the project MTM2014-55367-P, which is funded by Ministerio de Economı́a y Competitividad and Fondo Europeo de Desarrollo Regional FEDER,
and by the Junta de Andalucı́a Grant Number FQM-343. The second author is also partially
supported by Junta de Andalucı́a/Feder Grant Number FQM-5849.
† Departamento de Matemática Aplicada, Universidad de Granada, 18071-Granada, Spain.
E-mail: [email protected]
‡ Departamento de Álgebra, Universidad de Granada, 18071-Granada, Spain.
E-mail: [email protected]
1
In this work, our aim will be to introduce and study the concept of Rvariety (that is, Frobenius restricted variety). We will see how it generalizes the
concept of pseudo-variety and we will show that there exist significant families
of numerical semigroups which are R-varieties but not pseudo-varieties.
Let N be the set of nonnegative integers. A numerical semigroup is a subset
S of N such that it is closed under addition, contains the zero element, and N\ S
is finite.
It is well known (see [14, Lemma 4.5]) that, if S and T are numerical semigroups such that S T , then S ∪ {max(T \ S)} is another numerical semigroup.
We will denote by FT (S) = max(T \ S) and we will call it as the Frobenius
number of S restricted to T .
An R-variety is a non-empty family R of numerical semigroups that fulfills
the following conditions.
1. R has a maximum element with respect to the inclusion order (that we
will denote by ∆(R)).
2. If S, T ∈ R, then S ∩ T ∈ R.
3. If S ∈ R and S 6= ∆(R), then S ∪ {F∆(R) (S)} ∈ R.
In Section 2 we will see that every pseudo-variety is an R-variety. Moreover,
we will show that, if V is a variety and T is a numerical semigroup, then VT =
{S ∩ T | S ∈ V} is an R-variety. In fact, we will prove that every R-variety is
of this form.
Let R be an R-variety and let M be a submonoid of (N, +). We will say that
M is an R-monoid if it can be expressed as intersection of elements of R. It is
clear that the intersection of R-monoids is another R-monoid and, therefore, we
can define the R-monoid generated by a subset of ∆(R). In Section 3 we will
show that every R-monoid admits a unique minimal R-system of generators.
In addition, we will see that, if M is an R-monoid and x ∈ M , then M \ {x} is
another R-monoid if and only if x belongs to the minimal R-system of generators
of M .
In Section 4 we will show that the elements of an R-variety, R, can be
arranged in a tree with root ∆(R). Moreover, we will prove that the set of
children of a vertex S, of such a tree, is equal to {S \ {x} | x is an element
of the minimal R-system of generators of S and x > F∆(R) (S) . This fact
will allow us to show an algorithmic process in order to recurrently build the
elements of an R-variety.
Finally, in Section 5 we will see that, in general and contrary to what happens
with varieties and pseudo-varieties, we cannot define the smallest R-variety that
contains a given family F of numerical semigroups. Nevertheless, we will show
that, if ∆ is a numerical semigroup such that S ⊆ ∆ for all S ∈ F , then there
exists the smallest R-variety (denoted by R(F , ∆)) containing F and having ∆
as maximum (with respect the inclusion order). Moreover, we will prove that
R(F , ∆) is finite if and only if F is finite. In such a case, that fact will allow
us to compute, for a given R(F , ∆)-monoid, its minimal R(F , ∆)-system of
2
generators. In this way, we will obtain an algorithmic process to determine all
the elements of R(F , ∆) by starting from F and ∆.
Let us observe that the proofs, of some results of this work, are similar to the
proofs of the analogous results for varieties and pseudo-varieties. However, in
order to get a self-contained paper, we have not omitted several of such proofs.
2
Varieties, pseudo-varieties, and R-varieties
It is said that M is a submonoid of (N, +) if M is a subset of N which is closed
for the addition and such that 0 ∈ M . It particular, if S is a submonoid of
(N, +) such that N \ S is finite, then S is a numerical semigroup.
Let A be a non-empty subset of N. Then it is denoted by hAi the submonoid
of (N, +) generated by A, that is,
hAi = {λ1 a1 + · · · + λn an | n ∈ N \ {0}, a1 , . . . , an ∈ A, λ1 , . . . , λn ∈ N}.
It is well known (see for instance [14, Lemma 2.1]) that hAi is a numerical
semigroup if and only if gcd(A) = 1.
Let M be a submonoid of (N, +) and let A ⊆ N. If M = hAi, then it
is said that A is a system of generators of M . Moreover, it is said that A
is a minimal system of generators of M if M 6= hBi for all B
A. It is a
classical result that every submonoid M of (N, +) has a unique minimal system
of generators (denoted by msg(M )) which, in addition, is finite (see for instance
[14, Corollary 2.8]).
Let S be a numerical semigroup. Being that N \ S is finite, it is possible to
define several notable invariants of S. One of them is the Frobenius number of
S (denoted by F(S)) which is the greatest integer that does not belong to S (see
[6]). Another one is the genus of S (denoted by g(S)) which is the cardinality
of N \ S.
Let S be a numerical semigroup different from N. Then it is obvious that
S ∪ {F(S)} is also a numerical semigroup. Moreover, from [14, Proposition 7.1],
we have that T is a numerical semigroup with g(T ) = g + 1 if and only if there
exist a numerical semigroup S and x ∈ msg(S) such that g(S) = g, x > F(S),
and T = S \ {x}. This result is the key to build the set of all numerical
semigroups with genus g + 1 when we have the set of all numerical semigroups
with genus g (see [14, Proposition 7.4]).
In [11] it was introduced the concept of (Frobenius) variety in order to generalize the previous situation to some relevant families of numerical semigroups.
It is said that a non-empty family of numerical semigroups V is a (Frobenius)
variety if the following conditions are verified.
1. If S, T ∈ V, then S ∩ T ∈ V.
2. If S ∈ V and S 6= N, then S ∪ {F(S)} ∈ V.
However, there exist families of numerical semigroups that are not varieties,
but have a very similar behavior. By studying these families of numerical semigroups, we introduced in [8] the concept of (Frobenius) pseudo-variety.
3
It is said that a non-empty family of numerical semigroups P is a (Frobenius)
pseudo-variety if the following conditions are verified.
1. P has a maximum element with respect to the inclusion order (that we
will denote by ∆(P)).
2. If S, T ∈ P, then S ∩ T ∈ P.
3. If S ∈ P and S 6= ∆(P), then S ∪ {F(S)} ∈ P.
From the definitions, it is clear that every variety is a pseudo-variety. Moreover, as a consequence of [8, Proposition 1], we have the next result.
Proposition 2.1. Let P be a pseudo-variety. Then P is a variety if and only
if N ∈ P.
The following result asserts that the concept of R-variety generalizes the
concept of pseudo-variety.
Proposition 2.2. Every pseudo-variety is an R-variety.
Proof. Let P be a pseudo-variety. In order to prove that P is an R-variety,
we have to show that, if S ∈ P and S 6= ∆(P), then S ∪ {F∆(P) (S)} ∈ P.
Since P is a pseudo-variety, we know that S ∪ {F(S)} ∈ P. Thus, to finish
the proof, it is enough to see that F(S) = F∆(P) (S). On the one hand, it is
clear that F∆(P) (S) ≤ F(S). On the other hand, since S ∪ {F(S)} ∈ P, then
we have that F(S) ∈ ∆(P). Therefore, F(S) ∈ ∆(P) \ S and, consequently,
F(S) ≤ F∆(P) (S).
In the next example we see that there exist R-varieties that are not pseudovarieties.
Example 2.3. Let R be the set formed by all numerical semigroups which are
contained in the numerical semigroup h5, 7, 9i. It is clear that R is an R-variety.
However, since S = h5, 7, 9i \ {5} ∈ R, S 6= ∆(R) = h5, 7, 9i, F(S) = 13, and
S ∪ {13} ∈
/ R, we have that R is not a pseudo-variety.
Generalizing the above example, we can obtain several R-varieties, most of
which are not pseudo-varieties.
1. Let T be a numerical semigroup. Then LT = {S | S is a numerical
semigroup and S ⊆ T } is an R-variety. Observe that LT is the set formed
by all numerical subsemigroups of T .
2. Let S1 and S2 be two numerical semigroups such that S1 ⊆ S2 . Then
[S1 , S2 ] = {S | S is a numerical semigroup and S1 ⊆ S ⊆ S2 } is an
R-variety.
3. Let T be a numerical semigroup and let A ⊆ T . Then R(A, T ) = {S | S
is a numerical semigroup and A ⊆ S ⊆ T } is an R-variety. Observe that
both of the previous examples are particular cases of this one.
4
Remark 2.4. Let p, q be relatively prime integers such that 1 < p < q. Let us
take the numerical semigroups S1 = hp, qi and S2 = S21 = {s ∈ N | 2s ∈ S1 }. In
[4, 5], Kunz and Waldi study the family of numerical semigroups [S1 , S2 ], which
is an R-variety but not a pseudo-variety.
The next result establishes when an R-variety is a pseudo-variety.
Proposition 2.5. Let R be an R-variety. Then R is a pseudo-variety if and
only if F(S) ∈ ∆(R) for all S ∈ R such that S 6= ∆(R).
Proof. (Necessity.) If R is a pseudo-variety and S ∈ R with S 6= ∆(R), then
S ∪ {F(S)} ∈ R. Therefore, F(S) ∈ ∆(R).
(Sufficiency.) In order to show that R is a pseudo-variety, it will be enough
to see that S ∪ {F(S)} ∈ R for all S ∈ R such that S 6= ∆(R). For that,
since F(S) ∈ ∆(R), then it is clear that F∆(R) (S) = F(S) and, therefore,
S ∪ {F(S)} = S ∪ {F∆(R) (S)} ∈ R.
An immediate consequence of Propositions 2.1 and 2.5 is the following result.
Corollary 2.6. Let R be an R-variety. Then R is a variety if and only if
N ∈ R.
Our next purpose, in this section, will be to show that to give an R-variety
is equivalent to give a pair (V, T ) where V is a variety and T is a numerical
semigroup. Before that we need to introduce some concepts and results.
Let S be a numerical semigroup. Then we define recurrently the following
sequence of numerical semigroups.
• S0 = S,
• if Si 6= N, then Si+1 = Si ∪ {F(Si )}.
Since N \ S is a finite set with cardinality equal to g(S), then we get a finite
chain of numerical semigroups S = S0 S1 · · · Sg(S) = N. We will denote
by C(S) the set {S0 , S1 , . . . , Sg(S) } and will say that it is the chain of numerical
semigroups associated to S. If F is a non-empty
family of numerical semigroups,
S
then we will denote by C(F ) the set S∈F C(S).
Let F be a non-empty family of numerical semigroups. We know that there
exists the smallest variety containing F (see [11]). Moreover, by [11, Thoerem 4],
we have the next result.
Proposition 2.7. Let F be a non-empty family of numerical semigroups. Then
the smallest variety containing F is the set formed by all finite intersections of
elements of C(F ).
Now, let R be an R-variety. By applying repeatedly that, if S ∈ R and
S 6= ∆(R), then S ∪ {F∆(R) (S)} ∈ R, we get the following result.
Lemma 2.8. Let R be an R-variety. If S ∈ R and n ∈ N, then S ∪{x ∈ ∆(R) |
x ≥ n} ∈ R.
5
We are ready to show the announced result.
Theorem 2.9. Let V be a variety and let T be a numerical semigroup. Then
VT = {S ∩ T | S ∈ V} is an R-variety. Moreover, every R-variety is of this
form.
Proof. By Proposition 2.1, we know that, if V is a variety, then N ∈ V and,
therefore, T is the maximum of VT (that is, T = ∆(VT )). On the other hand,
it is clear that, if S1 , S2 ∈ VT , then S1 ∩ S2 ∈ VT .
Now, let S ∈ V such that S ∩ T 6= T and let us have t = FT (S ∩ T ). In
order to conclude that VT is an R-variety, we will see that (S ∩ T ) ∪ {t} ∈ VT .
First, let us observe that t = max(T \ (S ∩ T )) = max(T \ S). Then, because
S ∈ V and V is a variety, we can easily deduce that S̄ = S ∪ {t, →} ∈ V.
Moreover, (S ∩ T ) ∪ {t} ⊆ (S ∩ T ) ∪ ({t, →} ∩ T ) = S̄ ∩ T . Let us see now that
S̄ ∩ T ⊆ (S ∩ T ) ∪ {t}. In other case, there exists t′ > t such that t′ ∈ T and
t′ ∈
/ S, in contradiction with the maximality of t. Therefore, (S∩T )∪{t} = S̄∩T
and S̄ ∈ V. Consequently, (S ∩ T ) ∪ {t} ∈ VT .
Let R be an R-variety and let V be the smallest variety containing R. To
conclude the proof of the theorem, we will see that R = V∆(R) . It is clear that
R ⊆ V∆(R) . Thus, let us see the reverse one. For that, we will prove that, if
S ∈ V, then S ∩ ∆(R) ∈ R. In effect, by Proposition 2.7 we have that, if S ∈ V,
then there exist S1 , . . . , Sk ∈ C(R) such that S = S1 ∩ · · · ∩ Sk . Therefore,
S ∩ ∆(R) = (S1 ∩ ∆(R)) ∩ · · · ∩ (Sk ∩ ∆(R)). Since R is an R-variety, then R is
closed under finite intersections. Thereby, to see that S ∩∆(R) ∈ R, it is enough
to show that Si ∩ ∆(R) ∈ R for all i ∈ {1, . . . , k}. Since Si ∈ C(R), then it is
clear that there exist Si′ ∈ R and ni ∈ N such that Si = Si′ ∪{ni , →}. Therefore,
Si ∩ ∆(R) = Si′ ∪ {x ∈ ∆(R) | x ≥ ni } ∈ R, by applying Lemma 2.8.
The above theorem allows us to give many examples of R-varieties starting
from already known varieties.
1. Let us observe that, if V is a variety and T ∈ V, then VT = {S ∩ T | S ∈
V} = {S ∈ V | S ⊆ T } is an R-variety contained in V. Thus, for instance,
we have that the set formed by all Arf numerical semigroups, which are
contained in a certain Arf numerical semigroup, is an R-variety.
2. Observe also that, if V is a variety and T is a numerical semigroup such
that T ∈
/ V, then VT = {S ∩ T | S ∈ V} is an R-variety not contained in V
(because T ∈ VT and T ∈
/ V). Let us take, for example, the variety V of
all Arf numerical semigroups and T = h5, 8i ∈
/ V. In such a case, VT is the
R-variety formed by the numerical semigroups which are the intersection
of an Arf numerical semigroup and T .
Corollary 2.10. Let R be an R-variety and let U be a numerical semigroup.
Then RU = {S ∩ U | S ∈ R} is an R-variety.
Proof. By applying Theorem 2.9, we have that there exist a variety V and
a numerical semigroup T such that R = VT = {S ∩ T | S ∈ V}. Therefore,
6
RU = {S∩T ∩U | S ∈ V} = VT ∩U , which is clearly an R-variety (by Theorem 2.9
again).
The next result says us that Theorem 2.9 remains true when variety is replaced with pseudo-variety.
Corollary 2.11. Let P be a pseudo-variety and let T be a numerical semigroup.
Then PT = {S ∩ T | S ∈ P} is an R-variety. Moreover, every R-variety is of
this form.
Proof. By Proposition 2.2, we know that, if P is a pseudo-variety, then P is
an R-variety. Thereby, by applying Corollary 2.10, we conclude that PT is an
R-variety.
Now, by Theorem 2.9, we know that, if R is an R-variety, then there exist a
variety V and a numerical semigroup T such that R = VT . To finish the proof,
it is enough to observe that all varieties are pseudo-varieties.
Let us see an illustrative example of the above corollary.
Example 2.12. From [8, Example 7], we have the pseudo-variety
P = {h5, 6, 8, 9i, h5, 6, 9, 13i, h5, 6, 8i, h5, 6, 13, 14i,
h5, 6, 9i, h5, 6, 14i, h5, 6, 13i, h5, 6, 19i, h5, 6i} .
Thereby, we have that PT is an R-variety for each numerical semigroup T .
3
Monoids associated to an R-variety
In this section, R we will be an R-variety. Now, let M be a submonoid of (N, +).
We will say that M is an R-monoid if it is the intersection of elements of R.
The next result is easy to proof.
Lemma 3.1. The intersection of R-monoids is an R-monoid.
From the above lemma we have the following definition: let A ⊆ ∆(R). We
will say that R(A) is the R-monoid generated by A if R(A) is equal to the
intersection of all the R-monoids which contain the set A. Observe that R(A)
is the smallest R-monoid which contains the set A (with respect to the inclusion
order). The next result has an easy proof too.
Lemma 3.2. If A ⊆ ∆(R), then R(A) is equal to the intersection of all the
elements of R which contain the set A.
Let us take A ⊆ ∆(R). If M = R(A), then we will say that A is an Rsystem of generators of M . Moreover,we will say that A is a minimal R-system
of generators of M if M 6= R(B) for all B
A. The next purpose in this
section will be to show that every R-monoid has a unique minimal R-system of
generators. For that, we will give some previous lemmas. We can easily deduced
the first one from Lemma 3.2.
7
Lemma 3.3. Let A, B be two subsets of ∆(R) and let M be an R-monoid. We
have that
1. if A ⊆ B, then R(A) ⊆ R(B);
2. R(A) = R(hAi);
3. R(M ) = M .
If M is an R-monoid, then M is a submonoid of (N, +). Moreover, as we
commented in Section 2, we know that there exists a finite subset A of M such
that M = hAi. Thereby, by applying Lemma 3.3, we have that M = R(M ) =
R(hAi) = R(A). Consequently, A is a finite R-system of generators of M .
Thus, we can establish the next result.
Lemma 3.4. Every R-monoid has a finite R-system of generators.
In the following result, we characterize the minimal R-systems of generators.
Lemma 3.5. Let A ⊆ ∆(R) and M = R(A). Then A is a minimal R-system
of generators of M if and only if a ∈
/ R(A \ {a}) for all a ∈ A.
Proof. (Necessity.) If a ∈ R(A \ {a}), then A ⊆ R(A \ {a}). Thus, by Lemma 3.3, we get that M = R(A) ⊆ R(R(A \ {a})) = R(A \ {a}) ⊆ R(A) = M .
Therefore, M = R(A \ {a}), in contradiction with the minimality of A.
(Sufficiency.) If A is not a minimal R-system of generators of M , then there
exists B
A such that R(B) = M . Then, by Lemma 3.3, if a ∈ A \ B, then
a ∈ M = R(B) ⊆ R(A \ {a}), in contradiction with the hypothesis.
The next result generalizes an evident property of submonoids of (N, +).
More concretely, every element x of a submonoid M of (N, +) is expressible as a
non-negative integer linear combination of the generators of M that are smaller
than or equal to x.
Lemma 3.6. Let A ⊆ ∆(R) and x ∈ R(A). Then x ∈ R({a ∈ A | a ≤ x}).
Proof. Let us suppose that x 6∈ R({a ∈ A | a ≤ x}). Then, from Lemma 3.2,
we know that there exists S ∈ R such that {a ∈ A | a ≤ x} ⊆ S and x 6∈ S. By
applying now Lemma 2.8, we have that S̄ = S ∪ {m ∈ ∆(R) | m ≥ x + 1} ∈ R.
Observe that, obviously, A ⊆ S̄ and x ∈
/ S̄. Therefore, by applying once again
Lemma 3.2, we get that x ∈
/ R(A), in contradiction with the hypothesis.
We are now ready to show the above announced result.
Theorem 3.7. Every R-monoid admits a unique minimal R-system of generators. In addition, such a R-system is finite.
Proof. Let M be an R-monoid and let A, B be two minimal R-systems of generators of M . We are going to see that A = B. For that, let us suppose
that A = {a1 < a2 < · · · } and B = {b1 < b2 < · · · }. If A 6= B, then
8
there exists i = min{k | ak 6= bk }. Let us assume, without loss of generality, that ai < bi . Since ai ∈ M = R(A) = R(B), by Lemma 3.6, we have
that ai ∈ R({b1 , . . . , bi−1 }). Because {b1 , . . . , bi−1 } = {a1 , . . . , ai−1 }, then
ai ∈ R({a1 , . . . , ai−1 }), in contradiction with Lemma 3.5. Finally, by Lemma 3.4, we have that the minimal R-system of generators is finite.
If M is a R-monoid, then the cardinality of the minimal R-system of generators of M will be called the R-range of M .
Example 3.8. Let S, T be two numerical semigroups such that S ⊆ T . We
define recurrently the following sequence of numerical semigroups.
• S0 = S,
• if Si 6= T , then Si+1 = Si ∪ {FT (Si )}.
Since T \ S is a finite set, then we get a finite chain of numerical semigroups
S = S0 S1 · · · Sn = T . We will denote by C(S, T ) the set {S0 , S1 , . . . , Sn }
and will say that it is the chain of S restricted to T . It is clear that C(S, T ) is
an R-variety. Moreover, it is also clear that, for each i ∈ {1, . . . , n}, Si is the
smallest element of C(S, T ) containing FT (Si−1 ). Therefore, {FT (Si−1 )} is the
minimal C(S, T )-system of generators of Si for all i ∈ {1, . . . , n}. Let us also
observe that the empty set, ∅, is the minimal C(S, T )-system of generators of
S0 . Thereby, the C(S, T )-range of Si is equal to 1, if i ∈ {1, . . . , n}, and 0, if
i = 0.
It is well known that, if M is a submonoid of (N, +) and x ∈ M , then M \{x}
is another submonoid of (N, +) if and only if x ∈ msg(M ). In the next result
we generalize this property to R-monoids.
Proposition 3.9. Let M be an R-monoid and let x ∈ M . Then M \ {x} is an
R-monoid if and only if x belongs to the minimal R-system of generators of M .
Proof. Let A be the minimal R-system of generators of M . If x 6∈ A, then
A ⊆ M \{x}. Therefore, M \{x} is a R-monoid containing A and, consequently,
M = R(A) ⊆ M \ {x}, which is a contradiction.
Conversely, by Theorem 3.7, we have that, if x ∈ A, then R(M \ {x}) 6=
R(A) = M . Thereby, R(M \ {x}) = M \ {x}. Consequently, M \ {x} is a
R-monoid.
Let us illustrate the above proposition with an example.
Example 3.10. Let T be a numerical semigroup and let A ⊆ T . Then we
know that R(A, T ) = {S | S is a numerical semigroup and A ⊆ S ⊆ T } is an
R-variety. By applying Proposition 3.9, we easily deduce that, if S ∈ R(A, T ),
then the minimal R(A, T )-system of generators of S is {x ∈ msg | x ∈
/ A}.
From Theorem 2.9 we know that every R-variety is of the form VT = {S ∩T |
S ∈ V}, where V is a variety and T is a numerical semigroup. Now, our purpose
is to study the relation between V-monoids and VT -monoids.
9
Proposition 3.11. Let M be a submonoid of (N, +) and let T be a numerical
semigroup. Then M is a VT -monoid if and only if there exists a V-monoid M ′
such that M = M ′ ∩ T .
Proof. T(Necessity.) If M is a VT -monoid, then there exists F ⊆ VT such that
M = S∈F S. But, if S ∈ F , then S ∈ VT and, consequently, there exists
S ′ ∈ VTsuch that S = S ′ ∩ T . Now, let F ′ = {S ′ ∈ V | S ′ ∩ T ∈ F } and let
M ′ = S ′ ∈F ′ S ′ . Then it is clear that M ′ is a V-monoid and that M = M ′ ∩ T .
(Sufficiency.)
If M ′ is a V-monoid, then there exists F ′ ∈ V such that
T
′
′
M =T S ′ ∈F ′ S . Let F = {S ′ ∩ T | S ′ ∈ F ′ }. Then it is clear that F ⊆ VT and
that S∈F S = M ′ ∩ T . Therefore, M ′ ∩ T is a VT -monoid.
Observe that, as a consequence of the above proposition, we have that the
set of VT -monoids is precisely given by {M ∩ T | M is a V-monoid}.
Corollary 3.12. Let T be a numerical semigroup. If A ⊆ T , then VT (A) =
V(A) ∩ T .
Proof. By Proposition 3.11, we know that V(A) ∩ T is aVT -monoid containing
A. Therefore, VT (A) ⊆ V(A) ∩ T .
Let us see now the opposite inclusion. By applying once more Proposition 3.11, we deduce that there exists a V-monoid M such that VT (A) = M ∩ T .
Thus, it is clear that A ⊆ M and, thereby, V(A) ⊆ M . Consequently, V(A)∩T ⊆
M ∩ T = VT (A).
From Corollary 3.12, we have that the set formed by the VT -monoids is
{V(A) ∩ T | A ⊆ T } = {M ∩ T | M is a V-monoid and its minimal V-system
of generators is including in T }. Moreover, observe that, if T ∈ V, then
VT (A) = V(A) and, therefore, in such a case the set formed by all the VT monoids coincides with the set formed by all the V-monoids that are contained
in T .
For some varieties there exist algorithms that allow us to compute V(A) by
starting from A. Thereby, we can use such results in order to compute VT (A).
Let us see two examples of this fact.
Example 3.13. An LD-semigroup (see [12]) is a numerical semigroup S fulfilling that a + b − 1 ∈ S for all a, b ∈ S \ {0}. Let V the set formed by all
LD-semigroups. In [12] it is shown that V is a variety. Let T = h5, 7, 9i (observe
that T ∈
/ V). By Theorem 2.9, we know that VT = {S ∩ T | S ∈ V} is an
R-variety. Let us suppose that we can compute VT ({5}).
In [12] we have an algorithm to compute V(A) by starting from A. By using
such algorithm, in [12, Example 33] it is shown that V({5}) = h5, 9, 13, 17, 21i.
Therefore, by applying Corollary 3.12, we have that VT ({5}) = h5, 9, 13, 17, 21i∩
h5, 7, 9i = h5, 9, 17, 21i.
Example 3.14. An PL-semigroup (see [7]) is a numerical semigroup S fulfilling
that a+b+1 ∈ S for all a, b ∈ S\{0}. Let V the set formed by all PL-semigroups.
In [7] it is shown that V is a variety and it is given an algorithm to compute V(A)
10
by starting from A. Let T = h4, 7, 13i (observe that T ∈ V). By Theorem 2.9,
we know that VT = {S ∩ T | S ∈ V} is an R-variety. Let us suppose that we
can compute VT ({4, 7}).
From [7, Example 48], we know that V({4, 7}) = h4, 7, 9i. Thus, by applying
Corollary 3.12, we have that VT ({4, 7}) = h4, 7, 9i ∩ h4, 7, 13i = h4, 7, 13i.
Let T be a numerical semigroup. We know that, if M is a VT -monoid, then
there exists a V-monoid, M ′ , with minimal V-system of generators contained
in T , such that M = M ′ ∩ T . The next result says us that, in this situation,
the minimal V-system of generators of M ′ is just the minimal VT -system of
generators of M .
Proposition 3.15. Let A ⊆ T . Then A is the minimal VT -system of generators
of VT (A) if and only if A is the minimal V-system of generators of V(A).
Proof. (Necessity.) Let us suppose that A is not the minimal V-system of
generators of V(A). That is, there exists B A such that V(B) = V(A). Then,
from Corollary 3.12, we have that VT (A) = V(A) ∩ T = V(B) ∩ T = VT (B).
Therefore, A is not the minimal VT -system of generators of VT (A).
(Sufficiency.) Let us suppose that A is not the minimal VT -system of
generators of VT (A). Then VT (B) = VT (A) for some subset B
A. On
the other hand, due to A is the minimal V-system of generators of V(A),
from Lemma 3.5, we have an element a ∈ A such that a ∈
/ V(B). Consequently, a ∈ V(A) ∩ T and a ∈
/ V(B) ∩ T . Finally, from Corollary 3.12,
VT (A) = V(A) ∩ T 6= V(B) ∩ T = VT (B), which is a contradiction.
We finish this section with two examples that illustrate the above proposition.
Example 3.16. Let V be such as in Example 3.13 and let T = h4, 6, 7i. From
[12, Example 26], we know that V({4, 7, 10}) = h4, 7, 10, 13i and, moreover, that
{4} is its minimal V-system of generators. Then, from Proposition 3.15, {4} is
the minimal V-system of generators of VT ({4, 7, 10}) = h4, 7, 10, 13i ∩ h4, 6, 7i.
Example 3.17. Let V be such as in Example 3.14 and let T = h3, 4i. From
[7, Example 44], we know that {3} is the minimal V-system of generators os
S = h3, 7, 11i. Therefore, by Proposition 3.15, {3} is the minimal VT -system of
generators of S ∩ T .
4
The tree associated to an R-variety
Let V be a non-empty set and let E ⊆ {(v, w) ∈ V × V | v 6= w}. It is said that
the pair G = (V, E) is a graph. In addition, the vertices and edges of G are the
elements of V and E, respectively.
Let x, y ∈ V and let us suppose that (v0 , v1 ), (v1 , v2 ), . . . , (vn−1 , vn ) is a
sequence of different edges such that v0 = x and vn = y. Then, it is said that
such a sequence is a path (of length n) connecting x and y.
11
Let G be a graph. Let us suppose that there exists r, vertex of G, such that
it is connected with any other vertex x by a unique path. Then it is said that
G is a tree and that r is its root.
Let x, y be vertices of a tree G and let us suppose that there exists a path
that connects x and y. Then it is said that x is a descendant of y. Specifically,
it is said that x is a child of y when (x, y) is an edge of G.
From now on in this section, let R denote an R-variety. We define the graph
G(R) in the following way,
• R is the set of vertices of G(R);
• (S, S ′ ) ∈ R × R is an edge of G(R) if and only if S ′ = S ∪ {F∆(R) (S)}.
If S ∈ R, then we can define recurrently (such as we did in Example 3.8)
the sequence of elements in R,
• S0 = S,
• if Si 6= ∆(R), then Si+1 = Si ∪ {F∆(R) (Si )}.
Thus, we obtain a chain (of elements in R) S = S0
S1
· · · Sn = ∆(R)
such that (Si , Si+1 ) is an edge of G(R) for all i ∈ {0, . . . , n − 1}. We will denote
by CR (S) the set {S0 , S1 , . . . , Sn } and will say that it is the chain of S in R.
The next result is easy to prove.
Proposition 4.1. G(R) is a tree with root ∆(R).
Observe that, in order to recurrently construct G(R) starting from ∆(R),
it is sufficient to compute the children of each vertex of G(R). Let us also
observe that, if T is a child of S, then S = T ∪ {F∆(R) (T )}. Therefore, T =
S \ {F∆(R) (T )}. Thus, if T is a child of S, then there exists an integer x >
F∆(R) (S) such that T = S \ {x}. As a consequence of Propositions 3.9 and 4.1,
and defining F∆(R) (∆(R)) = −1, we have the following result.
Theorem 4.2. The graph G(R) is a tree with root equal to ∆(R). Moreover,
the set formed by the children of a vertex S ∈ R is {S \ {x} | x is an element
of the minimal R-system of generators of S and x > F∆(R) (S) .
We can reformulate the above theorem in the following way.
Corollary 4.3. The graph G(R) is a tree with root equal to ∆(R). Moreover,
the set formed by the children of a vertex S ∈ R is {S \ {x} | x ∈ msg(S),
x > F∆(R) (S) and S \ {x} ∈ R .
We illustrate the previous results with an example.
Example 4.4. We are going to build the R-variety R = [h5, 6i, h5, 6, 7i] =
{S | S is a numerical semigroup and h5, 6i ⊆ S ⊆ h5, 6, 7i}. Observe that, if
S ∈ R and x ∈ msg(S), then S \{x} ∈ R if and only if x ∈
/ {5, 6}. Moreover, the
maximum of R is ∆ = h5, 6, 7i. By applying Corollary 4.3, we can recurrently
build G(R) in the following way.
12
• h5, 6, 7i has got a unique child, which is h5, 6, 7i \ {7} = h5, 6, 13, 14i.
Moreover, F∆ (h5, 6, 13, 14i) = 7.
• h5, 6, 13, 14i has got two children, which are h5, 6, 13, 14i \ {13} = h5, 6, 14i
and h5, 6, 13, 14i \ {14} = h5, 6, 13i. Moreover, F∆ (h5, 6, 14i) = 13 and
F∆ (h5, 6, 13i) = 14.
• h5, 6, 13i has not got children.
• h5, 6, 14i has got a unique child, which is h5, 6, 14i \ {14} = h5, 6, 19i.
Moreover, F∆ (h5, 6, 19i) = 14.
• h5, 6, 19i has got a unique child, which is h5, 6, 19i \ {19} = h5, 6i. Moreover, F∆ (h5, 6i) = 19.
• h5, 6i has not got children.
Therefore, in this situation, G(R) is given by the next diagram.
h5, 6, 7i
✻
h5, 6, 13, 14i
✯
❨❍
❍
✟✟
✟
❍
h5, 6, 14i
h5, 6, 13i
✻
h5, 6, 19i
✻
h5, 6i
Observe that, if we represent the vertices of G(R) using their minimal R-systems
of generators, then we have that G(R) is given by the following diagram.
R({7})
✻
R({13, 14})
✯
❨❍
❍
✟✟
✟
❍
R({14})
R({13})
✻
R({19})
✻
R(∅)
Let us observe that the R-variety R = [h5, 6i, h5, 6, 7i] depict in the above
example is finite and, therefore, we have been able to build all its elements in
a finite number of steps. If the R-variety is infinite, then it is not possible such
situation. However, as a consequence of Theorem 4.2, we can show an algorithm
in order to compute all the elements of the R-variety when the genus is fixed.
13
Algorithm 4.5. INPUT: A positive integer g.
OUTPUT: {S ∈ R | g(S) = g}.
(1) If g < g(∆(R)), then return ∅.
(2) Set A = {∆(R)} and i = g(∆(R)).
(3) If i = g, then return A.
(4) For each S ∈ A, compute the set BS formed by all elements of the
minimal R-system of generators of S that are greater than F∆(R) (S).
[
(5) If
BS = ∅, then return ∅.
S∈A
(6) Set A =
[
S \ {x} | x ∈ BS , i = i + 1, and go to (3).
S∈A
We illustrate the operation of this algorithm with an example.
Example 4.6. Let ∆ = h4, 6, 7i = {0, 4, 6, 7, 8, 10, →}. It is clear that g(∆) = 5.
Let R = {S | S is a numerical semigroup and {4, 6} ⊆ S ⊆ ∆}. We have that
R is an infinite R-variety because h4, 6, 2k + 1i ∈ R for all k ∈ {5, →}. By using
Algorithm 4.5, we are going to compute the set {S ∈ R | g(S) = 8}.
• A = {∆}, i = 5.
• B∆ = {7}.
• A = {h4, 6, 11, 13i}, i = 6.
• Bh4,6,11,13i = {11, 13}.
• A = {h4, 6, 13, 15i, h4, 6, 11i}, i = 7.
• Bh4,6,13,15i = h13, 15i and Bh4,6,11i = ∅.
• A = {h4, 6, 15, 17i, h4, 6, 13i}, i = 8.
• The algorithm returns {h4, 6, 15, 17i, h4, 6, 13i}.
Our next purpose in this section will be to show that, if R is an R-variety
and T ∈ R, then the set formed by all the descendants of T in the tree G(R)
is also an R-variety. It is clear that, if S, T ∈ R, then S is a descendant of T if
and only if T ∈ CR (S). Therefore, we can establish the following result.
Lemma 4.7. Let R be an R-variety and S, T ∈ R. Then S is a descendant of
T if and only if there exists n ∈ N such that T = S ∪ {x ∈ ∆(R) | x ≥ n}.
As an immediate consequence of the above lemma, we have the next one.
Lemma 4.8. Let R be an R-variety and S, T ∈ R such that S 6= T . If S is a
descendant of T , then F∆(R) (S) = FT (S).
Now we are ready to show the announced result.
14
Theorem 4.9. Let R be an R-variety and T ∈ R. Then
D(T ) = {S ∈ R | S is a descendant of T in the tree G(R)}
is an R-variety.
Proof. Clearly, T is the maximum of D(T ). Let us see that, if S1 , S2 ∈ D(T ),
then S1 ∩ S2 ∈ D(T ). Since, from Lemma 4.7, we know that there exist
n1 , n2 ∈ N such that T = Si ∪ {x ∈ ∆(R) | x ≥ ni }, i = 1, 2, it is sufficient
to show that T = (S1 ∩ S2 ) ∪ {x ∈ ∆(R) | x ≥ min{n1 , n2 }}. It is obvious
that (S1 ∩ S2 ) ∪ {x ∈ ∆(R) | x ≥ min{n1 , n2 }} ⊆ T . Let us see now the opposite inclusion. For that, let t ∈ T such that t ∈
/ S1 ∩ S2 . Then t ∈
/ S1 or
t ∈
/ S2 and, therefore, t ∈ {x ∈ ∆(R) | x ≥ n1 } or t ∈ {x ∈ ∆(R) | x ≥ n2 }.
Thereby, t ∈ {x ∈ ∆(R) | x ≥ min{n1 , n2 }}. Consequently, T ⊆ (S1 ∩ S2 ) ∪
{x ∈ ∆(R) | x ≥ min{n1 , n2 }}. By applying again Lemma 4.7, we can assert
that S1 ∩ S2 ∈ D(T ).
Finally, let S ∈ D(T ) such that S 6= T . From Lemma 4.8, S ∪ {FT (S)} =
S ∪ {F∆(R) (S)} ∈ R and, in consequence, S ∪ {FT (S)} ∈ D(T ).
From the previous comment to [8, Example 7], we know that, if V is a variety
and T ∈ V, then D(T ) is a pseudo-variety and, moreover, every pseudo-variety
can be obtained in this way. Therefore, there exist R-varieties which are not
the set formed by all the descendants of an element belonging to a variety.
The following result shows that an R-variety can be obtained as the set
formed by intersecting all the descendants, of an element belonging to a variety,
with a numerical semigroup.
Corollary 4.10. Let V be a variety, let ∆ ∈ V, and let T be a numerical semigroup. Let D(∆) = {S | S is a descendant of ∆ in G(V)} and let D(∆, T ) =
{S ∩ T | S ∈ D(∆)}. Then D(∆, T ) is an R-variety. Moreover, every R-variety
can be obtained in this way.
Proof. If V is a variety, then V is an R-variety and, by applying Theoem 4.9, we
have that D(∆) is an R-variety. From Corollary 2.10, we conclude that D(∆, T )
is an R-variety.
If R is an R-variety, by Theorem 2.9, we know that there exist a variety V
and a numerical semigroup T such that R = {S ∩ T | S ∈ V}. Now, it is clear
that V = D(N) = {S | S is a descendant of N in G(V)}. Therefore, we have
that R = {S ∩ T | S ∈ D(N)} = D(N, T ).
In the next result we see that the above corollary is also true when we write
pseudo-variety instead of variety.
Corollary 4.11. Let P be a pseudo-variety, let ∆ ∈ P, and let T be a numerical semigroup. Let D(∆) = {S | S is a descendant of ∆ in G(P)} and let
D(∆, T ) = {S ∩ T | S ∈ D(∆)}. Then D(∆, T ) is an R-variety. Moreover,
every R-variety can be obtained in this way.
15
Proof. If P is a pseudo-variety, then P is an R-variety and, by applying Theorem 4.9, we have that D(∆) is an R-variety as well. Now, from Corollary 2.10,
we have that D(∆, T ) is an R-variety.
That every R-variety can be obtained in this way is an immediate consequence of Corollary 4.10 and having in mind that each variety is a pseudovariety.
We conclude this section by illustrating the above corollary with an example.
Example 4.12. Let P the pseudo-variety which appear in Example 2.12. In
[8, Example 7] it is shown that G(P) is given by the next subtree.
h5, 6, 8, 9i
✯
❨❍
❍
✟✟
✟
❍
h5, 6, 9, 13i
h5, 6, 8i
✯
❨❍
❍
✟✟
✟
❍
h5, 6, 13, 14i
h5, 6, 9i
✯
❨❍
❍
✟✟
✟
❍
h5, 6, 14i
h5, 6, 13i
✻
h5, 6, 19i
✻
h5, 6i
By applying Corollary 4.11, we have that, if T is a numerical semigroup,
then R = {S ∩ T | S ∈ D(h5, 6, 13, 14i)} is an R-variety. Let us observe that
D(h5, 6, 13, 14i) = {h5, 6, 13, 14i, h5, 6, 14i, h5, 6, 13i, h5, 6, 19i, h5, 6i}.
5
The smallest R-variety containing a family of
numerical semigroups
In [11, Proposition 2] it is proved that the intersection of varieties is a variety.
As a consequence of this, we have that there exists the smallest variety which
contains a given family of numerical semigroups.
On the other hand, in [8] was shown that, in general, the intersection of
pseudo-varieties is not a pseudo-variety. Nevertheless, in [8, Theorem 4] it
is proved that there exists the smallest pseudo-variety which contains a given
family of numerical semigroups.
Our first objective in this section will be to show that, in general, we cannot
talk about the smallest R-variety which contains a given family of numerical
semigroups.
Lemma 5.1. Let F be a family of numerical semigroups and let ∆ be a numerical semigroup such that S ⊆ ∆ for all S ∈ F . Then there exists an R-variety
R such that F ⊆ R and max(R) = ∆.
16
Proof. Let R = {S | S is a numerical semigroup and S ⊆ ∆}. From Item 1 in
Example 2.3, we have that R is an R-variety. Now, it is trivial that F ⊆ R and
max(R) = ∆.
The proof of the next lemma is straightforward and we can omit it.
Lemma 5.2. Let {R
such that
max(Ri ) = ∆
T i }i∈I be a family of R-varieties T
for all i ∈ I. Then i∈I Ri is an R-variety and max i∈I Ri = ∆.
The following result says us that there exists the smallest R-variety which
contains a given family of numerical semigroups and has a certain maximum.
Proposition 5.3. Let F be a family of numerical semigroups and let ∆ be a
numerical semigroup such that S ⊆ ∆ for all S ∈ F . Then there exists the
smallest R-variety which contains F and with maximum equal to ∆.
Proof. Let R be the intersection of all the R-varieties containing F and with
maximum equal to ∆. From Lemmas 5.1 and 5.2 we have the conclusion.
We will denote by R(F , ∆) the R-variety given by Proposition 5.3. Now we
are interested in describe the elements of such an R-variety.
Lemma 5.4. Let S1 , S2 , . . . , Sn , ∆ be numerical semigroups such that Si ⊆ ∆
for all i ∈ {1, . . . , n}. Then F∆ (S1 ∩ · · · ∩ Sn ) = max {F∆ (S1 ), . . . , F∆ (Sn )}.
Proof. We have that F∆ (S1 ∩ · · · ∩ Sn ) = max (∆ \ (S1 ∩ · · · ∩ Sn )) =
max ((∆ \ S1 ) ∪ · · · ∪ (∆ \ Sn )) = max {max(∆ \ S1 ), . . . , max(∆ \ Sn )} =
max {F∆ (S1 ), . . . , F∆ (Sn )} .
Let us recall that, if S and ∆ are numerical semigroups such that S ⊆ ∆,
then we defined C(S, ∆) in Example 3.8 (that is, the chain of S restricted to
∆). If F is a family of numerical semigroups
such that S ⊆ ∆ for all S ∈ F ,
S
then we will denote by C(F , ∆) the set S∈F C(S, ∆).
Theorem 5.5. Let F be a family of numerical semigroups and let ∆ be a
numerical semigroup such that S ⊆ ∆ for all S ∈ F . Then R(F , ∆) is the set
formed by all the finite intersections of elements in C(F , ∆).
Proof. Let R = {S1 ∩ · · · ∩ Sn | n ∈ N \ {0} and S1 , . . . , Sn ∈ C(F , ∆)}. Having in mind that R(F , ∆) is an R-variety which contains F and with maximum
equal to ∆, we easily deduce that R ⊆ R(F , ∆).
Let us see now that R is an R-variety. On the one hand, it is clear that
∆ = max(R) and that, if S, T ∈ R, then S ∩ T ∈ R. On the other hand, let
S ∈ R such that S 6= ∆. Then S = S1 ∩ · · · ∩ Sn for some S1 , . . . , Sn ∈ C(F , ∆).
Now, from Lemma 5.4, we have that F∆ (S) = max {F∆ (S1 ), . . . , F∆ (Sn )} and,
thus, F∆ (Si ) ≤ F∆ (S) for all i ∈ {1, . . . , n}. Let us observe that, if F∆ (S) >
F∆ (Si ), then Si ∪ {F∆ (S)} = Si . Moreover, if F∆ (S) = F∆ (Si ), then we get
Si ∪ {F∆ (S)} = Si ∪ {F∆ (Si )} ∈ C(F , ∆). Therefore, Si ∪ {F∆ (S)} ∈ C(F , ∆)
for all i ∈ {1, . . . , n}. Since S ∪{F∆ (S)} = (S1 ∪{F∆ (S)})∩· · ·∩(Sn ∪{F∆ (S)}),
then S ∪ {F∆ (S)} ∈ R. Consequently, R is an R-variety.
Finally, since R is an R-variety which contains F and with maximum equal
to ∆, then R(F , ∆) ⊆ R and, thereby, we conclude that R = R(F , ∆).
17
Let us observe that, if F is a finite family, then C(F , ∆) is a finite set and,
therefore, R(F , ∆) is a finite R-variety.
Lemma 5.6. Let R and R′ be two R-varieties. If R ⊆ R′ , then ∆(R) ⊆ ∆(R′ ).
Proof. If R ⊆ R′ , then ∆(R) ∈ R′ . Therefore, ∆(R) ⊆ ∆(R′ ).
The next example shows us that, in general, we cannot talk about the smallest R-variety which contains a given family of numerical semigroups.
Example 5.7. Let F = {h5, 6i, h5, 7i}. As a consequence of Lemma 5.6, the
candidate to be the smallest R-variety which contains F must have as maximum
the numerical semigroup h5, 6, 7i (that is, the smallest numerical semigroup
containing h5, 6i and h5, 7i). Thus, the candidate to be the smallest R-variety
which contains F is R(F , h5, 6, 7i).
Let us see now that R(F , h5, 6, 7i) 6⊆ R(F , h5, 6, 7, 8i) and, in this way,
that there does not exist the smallest R-variety which contains F . In order
to do it, we will show that h5, 6, 7i ∈
/ R(F , h5, 6, 7, 8i). In fact, by applying
Theorem 5.5, if h5, 6, 7i ∈ R(F , h5, 6, 7, 8i), then we deduce that there exist
S1 ∈ C(h5, 6i, h5, 6, 7, 8i) and S2 ∈ C(h5, 7i, h5, 6, 7, 8i) such that S1 ∩ S2 =
h5, 6, 7i. Since S1 ∈ C(h5, 6i, h5, 6, 7, 8i), then there exists n1 ∈ N such that
S1 = h5, 6i ∪ {x ∈ h5, 6, 7, 8i | x ≥ n1 }. Moreover, h5, 6, 7i ⊆ S1 and, thereby,
n1 ≤ 7. Consequently, 8 ∈ S1 . By an analogous reasoning, we have that 8 ∈ S2
too. Consequently, 8 ∈ S1 ∩ S2 = h5, 6, 7i, which is false.
Let R be an R-variety. We will say that F (subset of R) is a system of
generators of R if R = R(F , ∆(R)). Let us observe that, in such a case, R is
the smallest R-variety which contains F and with maximum equal to ∆(R).
We will say that an R-variety, R, is finitely generated if there exists a finite
set F ⊆ R such that R = R(F , ∆(R)) (that is, if R has a finite system of
generators). As a consequence of Theorem 5.5, we have the following result.
Corollary 5.8. An R-variety is finitely generated if and only if it is finite.
From now on, F will denote a family of numerical semigroups and ∆ will
denote a numerical semigroup such that S ⊆ ∆ for all S ∈ F . Our purpose is to
give a method in order to compute the minimal R(F , ∆)-system of generators
of a R(F , ∆)-monoid by starting from F and ∆.
If A ⊆ ∆, then for each S ∈ F we define
(
S, if A ⊆ S,
α(S) =
S ∪ {x ∈ ∆ | x ≥ xS }, if A * S,
where xS = min{a ∈ A | a ∈
/ S}.
As a consequence of Lemma 3.2 and Theorem 5.5, we have the next result.
T
Lemma 5.9. The R(F , ∆)-monoid generated by A is S∈F α(S).
Recalling that R(F , ∆)(A) denotes the R(F , ∆)-monoid generated by A, we
have the following result.
18
Proposition 5.10. If A ⊆ ∆, then B = {xS | S ∈ F and A 6⊆ S} is the minimal R(F , ∆)-system of generators of R(F , ∆)(A).
Proof. Let us observe that, if S ∈ F , then A ⊆ S if and only if B ⊆ S. Moreover,
if A 6⊆ S, then min{a ∈ A | a ∈
/ S} = min{b ∈ B | b ∈
/ S}. Therefore, by applying Lemma 5.9, we have that R(F , ∆)(A) = R(F , ∆)(B). Consequently, in order to prove that B is the minimal R(F , ∆)-system of generators of R(F , ∆)(A),
will be enough to see that, if C B, then R(F , ∆)(C) 6= R(F , ∆)(A).
In effect, if C B, then there exists S ∈ F such that xS ∈
/ C and, thereby,
we have that C ⊆ S or that min{c ∈ C | c ∈
/ S} > xS . Now, by applying
once more time Lemma 5.9, we easily deduce that xS ∈
/ R(F , ∆)(C). Since
xS ∈ B ⊆ A, then we get that A 6⊆ R(F , ∆)(C) and, therefore, R(F , ∆)(C) 6=
R(F , ∆)(A).
As an immediate consequence of the above proposition we have the next
result.
Corollary 5.11. Every R(F , ∆)-monoid has R(F , ∆)-range less than or equal
to the cardinality of F .
We will finish this section by illustrating its content with an example.
Example 5.12. Let F = {h5, 7, 9, 11, 13i, h4, 10, 11, 13i} and ∆ = h4, 5, 7i. We
are going to compute the tree G(R(F , ∆)).
First of all, to compute the minimal R(F , ∆)-system of generators of h4, 5, 7i,
we apply Proposition 5.10 with A = {4, 5, 7}. Since xh5,7,9,11,13i = 4 and
xh4,10,11,13i = 5, then {4, 5} is such a minimal R(F , ∆)-system. Now, because F∆ (h4, 5, 7i) = −1, and by applying Theorem 4.2, we get that h4, 5, 7i
has two children, h4, 5, 7i \ {4} = h5, 7, 8, 9, 11i (with F∆ (h5, 7, 8, 9, 11i) = 4)
and h4, 5, 7i \ {5} = h4, 7, 9, 10i (with F∆ (h4, 7, 9, 10i) = 5).
Now, if we take A = {5, 7, 8, 9, 11} in Proposition 5.10, then we have that
xh5,7,9,11,13i = 8 and xh4,10,11,13i = 5. Thus, we conclude that {5, 8} is the
minimal R(F , ∆)-system of h5, 7, 8, 9, 11i. Moreover, since F∆ (h5, 7, 8, 9, 11i) =
4, then Theorem 4.2 asserts that h5, 7, 8, 9, 11i \ {5} = h7, 8, 9, 10, 11, 12, 13i
(with F∆ (h7, 8, 9, 10, 11, 12, 13i) = 5) and h5, 7, 8, 9, 11i \ {8} = h5, 7, 9, 11, 13i
(with F∆ (h5, 7, 9, 11, 13i) = 8) are the two children of h5, 7, 8, 9, 11i.
With A = {4, 7, 9, 10}, we get that {4, 7} is the minimal R(F , ∆)-system of
h4, 7, 9, 10i. By recalling that F∆ (h4, 7, 9, 10i) = 5, we conclude that h4, 7, 9, 10i
has only one child, that is h4, 7, 9, 10i\{7} = h4, 9, 10, 11i (with F∆ (h4, 9, 10, 11i)
= 7).
By repeating the above process, we get the whole tree G(R(F , ∆)).
19
h4, 5, 7i
✟
✯
❨❍
❍
✟
✟
❍
h5, 7, 8, 9, 11i
h4, 7, 9, 10i
✯
❨❍
❍
✻
✟✟
✟
❍
h7, 8, 9, 10, 11, 12, 13i h5, 7, 9, 11, 13i
h4, 9, 10, 11i
✯
✟
❨
❍
✻
✟
❍
✟
❍
h8, 9, 10, 11, 12, 13, 14, 15i h7, 9, 10, 11, 12, 13, 15i
h4, 10, 11, 13i
✐PP
P
✻
PP
h9, 10, 11, 12, 13, 14, 15, 16, 17i h8, 10, 11, 12, 13, 15, 17i
✻
h10, 11, 12, 13, 14, 15, 16, 17, 18, 19i
Now, we are going to represent the vertices of G(R(F , ∆)) using their minimal R(F , ∆)-systems of generators. Moreover, we add to each vertex the corresponding Frobenius number restricted to ∆. Thus, we clarify all the steps to
build the tree G(R(F , ∆)).
R(F , ∆)({4, 5}), −1
✯
✟
❨
❍
❍❍
✟✟
R(F , ∆)({5, 8}), 4
R(F , ∆)({4, 7}), 5
✯
✟
❨
❍
✻
❍❍
✟✟
R(F , ∆)({7, 8}), 5
R(F , ∆)({5}), 8 R(F , ∆)({4, 9}), 7
✯
✟
❨
❍
✻
✟
❍
✟
❍
R(F , ∆)({8, 9}), 7
R(F , ∆)({7}), 8
R(F , ∆)({4}), 9
✐
P
PP
✻
PP
R(F , ∆)({9}), 8
R(F , ∆)({8}), 9
✻
R(F , ∆)(∅), 9
6
Conclusion
We have been able to give a structure to certain families of numerical semigroups
which are not (Frobenius) varieties or (Frobenius) pseudo-varieties. For that we
have generalized the concept of Frobenius number to the concept of restricted
Frobenius number and, then, we have defined the R-varieties (or (Frobenius)
restricted variety).
After studying relations among varieties, pseudo-varieties, and R-varieties,
we have introduced the concepts of R-monoid and minimal R-system of generators of a R-monoid, which lead to associate a tree with each R-variety and, in
consequence, obtain recurrently all the elements of an R-variety.
Finally, although in general it is not possible to define the smallest R-variety
that contains a given family F of numerical semigroups, we have been able to
20
give an alternative when we fix in advance the maximum of the smallest Rvariety.
References
[1] M. Bras-Amorós, P. A. Garcı́a-Sánchez, Patterns on numerical semigroups,
Linear Algebra Appl. 414 (2006), 652–669.
[2] M. Bras-Amorós, P. A. Garcı́a-Sánchez, and A. Vico-Oton, Nonhomogeneous patterns on numerical semigroups, Int. J. Algebra Comput. 23 (2013),
1469–1483.
[3] M. Delgado, P. A. Garcı́a-Sánchez, J. C. Rosales, J. M. Urbano-Blanco, Systems of proportionally modular Diophantine inequalities, Semigroup Forum
76 (2008), 469–488.
[4] E. Kunz, On the type of certain numerical semigroups and a question of
Wilf, Semigroup Forum 93 (2016), 205–210.
[5] E. Kunz, R. Waldi, Geometrical illustration of numerical semigroups and
of some of their invariants, Semigroup Forum 89 (2014), 664–691.
[6] J. L. Ramı́rez Alfonsı́n, The Diophantine Frobenius problem (Oxford Univ.
Press, 2005).
[7] A. M. Robles-Pérez, J. C. Rosales, The numerical semigroup of phrases’
lengths in a simple alphabet, The Scientific World Journal 2013 (2013),
Article ID 459024, 9 pages.
[8] A. M. Robles-Pérez, J. C. Rosales, Frobenius pseudo-varieties in numerical
semigroups, Ann. Mat. Pura Appl. 194 (2015), 275–287.
[9] A. M. Robles-Pérez, J. C. Rosales, Numerical semigroups in a problem
about cost-effective transport. To appear in Forum Math.
DOI: 10.1515/forum-2015-0123
[10] A. M. Robles-Pérez, J. C. Rosales, Numerical semigroups in a problem about economic incentives for consumers. Preprint (arXiv:1605.03900
[math.GR]).
[11] J. C. Rosales, Families of numerical semigroups closed under finite intersections and for the Frobenius number, Houst. J. Math. 34 (2008), 339–348.
[12] J. C. Rosales, M. B. Branco, D. Torrão, Sets of positive integers closed
under product and the number of decimal digits, J. Number Theory 147
(2015), 1–13.
[13] J. C. Rosales, M. B. Branco, D. Torrão, Bracelet monoids and numerical
semigroups, Appl. Algebr. Eng. Commun. Comput. 27 (2016), 169–183.
21
[14] J. C. Rosales and P. A. Garcı́a-Sánchez, Numerical semigroups (Developments in Mathematics, vol. 20, Springer, New York, 2009).
[15] J. C. Rosales, P. A. Garcı́a-Sánchez, J. I. Garcı́a-Garcı́a and M. B. Branco,
Numerical semigroups with maximal embedding dimension, Int. J. Commutat. Rings 2 (2003), 47–53.
[16] J. C. Rosales, P. A. Garcı́a-Sánchez, J. I. Garcı́a-Garcı́a and M. B. Branco,
Arf numerical semigroups, J. Algebra 276 (2004), 3–12.
[17] J. C. Rosales, P. A. Garcı́a-Sánchez, J. I. Garcı́a-Garcı́a and M. B. Branco,
Saturated numerical semigroups, Houst. J. Math. 30 (2004), 321–330.
22
| 4math.GR
|
1
Sparse Representation for Wireless Communications
Zhijin Qin1 , Jiancun Fan2 , Yuanwei Liu3 , Yue Gao3 , and Geoffrey Ye Li4
1
Lancaster University, Lancaster, UK
Xi’an Jiaotong University, Xi’an, China
3
Queen Mary University of London, London, UK
4
Georgia Institute of Technology, Atlanta, GA, USA
arXiv:1801.08206v1 [cs.IT] 24 Jan 2018
2
Abstract—Sparse representation can efficiently model
signals in different applications to facilitate processing. In
this article, we will discuss various applications of sparse
representation in wireless communications, with focus
on the most recent compressive sensing (CS) enabled
approaches. With the help of the sparsity property, CS
is able to enhance the spectrum efficiency and energy
efficiency for the fifth generation (5G) networks and
Internet of Things (IoT) networks. This article starts
from a comprehensive overview of CS principles and
different sparse domains potentially used in 5G and IoT
networks. Then recent research progress on applying CS
to address the major opportunities and challenges in
5G and IoT networks is introduced, including wideband
spectrum sensing in cognitive radio networks, data
collection in IoT networks, and channel estimation and
feedback in massive MIMO systems. Moreover, other
potential applications and research challenges on sparse
representation for 5G and IoT networks are identified.
This article will provide readers a clear picture of how to
exploit the sparsity properties to process wireless signals
in different applications.
Keywords: Wireless communications, compressive sensing, sparsity property, 5G, Internet of Things.
I. I NTRODUCTION
Sparse representation expresses some signals as a
linear combination of a few atoms from a prespecified
and over-complete dictionary [1]. This form of sparse
(or compressible) structure arises naturally in many
applications [2]. For example, audio signals are sparse in
frequency domain, especially for the sounds representing
tones. Image processing can exploit a sparsity property
in the discrete cosine domain, i.e. many discrete cosine
transform (DCT) coefficients of images are zero or small
enough to be regarded as zero. This type of sparsity
property has enabled intensive research on signal and
data processing, such as dimension reduction in data
science, wideband sensing in cognitive radio networks
(CRNs), data collection in large-scale wireless sensor
networks (WSNs), and channel estimation and feedback
in massive MIMO.
Traditionally, signal acquisition and transmission
adopt the procedure with sampling and compression.
As massive connectivity is expected to be supported
in the fifth generation (5G) networks and Internet of
Things (IoT) networks, the amount of generated data
becomes huge. Therefore, signal processing has been
confronted with challenges on high sampling rates for
data acquisition and large amount of data for storage and
transmission, especially in IoT applications with powerconstrained sensor nodes. Except for developing more
advanced sampling and compression techniques, it is
natural to ask whether there is an approach to achieve
signal sampling and compression simultaneously.
As an appealing approach employing sparse representations, compressive sensing (CS) technique [3] has been
proposed to reduce data acquisition costs by enabling
sub-Nyqusit sampling. Based on the advanced theory [4],
CS has been widely applied in many areas. The key
idea of CS is to enable exact signal reconstruction from
far fewer samples than that is required by the NyquistShannon sampling theorem provided that the signal
admits a sparse representation in a certain domain. In CS,
compressed samples are acquired via a small set of nonadaptive, linear, and usually randomized measurements,
and signal recovery is usually formulated as an l0 norm minimization problem to find the sparsest solution
satisfying the constraints. Since l0 -norm minimization
is an NP-hard problem, most of the exiting research
contributions on CS solve it by either approximating
it to a convex l1 -norm minimization problem [4] or
adopting greedy algorithms, such as orthogonal match
pursuit (OMP).
It is often the case that the sparsifying transformation is unknown or difficult to determine. Therefore,
projecting a signal to its proper sparse domain is quite
essential in many applications that invoke CS. In 5G
and IoT networks, the identified sparse domains mainly
include frequency domain, spatial domain, wavelet domain, DCT domain, etc. CS can be used to improve
spectrum efficiency (SE) and energy efficiency (EE) for
these networks. By enabling the unlicensed usage of
2
spectrum, CRNs exploit spectral opportunities over a
wide frequency range to enhance the network SE. In
wideband spectrum sensing, spectral signals naturally
exploit a sparsity property in frequency domain due
to low utilization of spectrum [5], [6], which enables
sub-Nyquist sampling on cognitive devices. Another
interesting scenario is a small amount of data collection in large-scale WSNs with power-constrained sensor
nodes, such as smart meter monitoring infrastructure in
IoT applications. In particular, the monitoring readings
usually have a sparse representation in DCT domain due
to the temporal and spatial correlations [7]. CS can be
applied to enhance the EE of WSNs and to extend the
lifetime of sensor nodes. Moreover, massive MIMO is a
critical technique for 5G networks. In massive MIMO
systems, channels corresponding to different antennas
are correlated. Furthermore, a huge number of channel
coefficients can be represented by only a few parameters
due to a hidden joint sparsity property caused by the
shared local scatterers in the radio propagation environment. Therefore, CS can be potentially used in massive
MIMO systems to reduce the overhead for channel estimation and feedback and facilitate precoding [8]. Even
though various applications have different characters, it
is worth noting that the signals in different scenarios
share a common sparsity property even though the sparse
domains can be different, which enables CS to enhance
the SE and EE of wireless communications networks.
There have been some interesting surveys on CS [9]
and its applications [10]–[12]. One of the most popular
articles on CS [9] has provided an overview on the theory
of CS as a novel sampling paradigm that goes against
the common wisdom in data acquisition. CS-enabled
sparse channel estimation has been summarized in [10].
In [11], a comprehensive review of the application of
CS in CRNs has been provided. A more specific survey
on compressive covariance sensing has been presented
in [12] that includes the reconstruction of second-order
statistics even in the absence of prior sparsity information. These existing surveys serves different purposes.
Some cover the basic principles for beginners and others
focus on specific aspects of CS. Different from the
existing literature, our article provides a comprehensive
overview of the recent contributions on CS-enabled wireless communications from the perspective of adopting
different sparse domain projections.
In this article, we will first introduce the basic principles of CS briefly. Then we will present the different
sparse domains for signals in wireless communications.
Subsequently, we will provide CS-enabled frameworks
in various wireless communications scenarios, including
wideband spectrum sensing in CRNs, data collection in
large-scale WSNs, and channel estimation and feedback
for massive MIMO, as they have been identified to be
critical to 5G and IoT networks and share the same
spirit by exploiting the sparse domains aforementioned.
Within each identified scenario, we start with projecting
a signal to a sparse domain, then introduce the CSenabled framework, and finally illustrate how to exploit
joint sparsity in the CS-enabled framework. Moreover,
the reweighted CS approaches for each scenario will be
discussed, where the weights are constructed by prior
information depending on specific application scenarios.
The other potential applications and research challenges
on applying CS in wireless networks will also be discussed and followed by conclusions.
This article gives readers a clear picture on the research and development of the applications of CS in different scenarios. By identifying the different sparse domains, this article illustrates the benefits and challenges
on applying CS in wireless communication networks.
II. S PARSE R EPRESENTATION
Sparse representation of signals has received extensive attention due to its capacity for efficient signal
modelling and related applications. The problem solved
by the sparse representation is to search for the most
compact representation of a signal in terms of a linear
combination of the atoms in an overcomplete dictionary.
In the literature, three aspects of research on the sparse
representation have been focused:
1) Pursuit methods for solving the optimization problem, such as matching pursuit and basis pursuit;
2) Design of the dictionary, such as the K-SVD
method;
3) Applications of the sparse representation, such as
wideband spectrum sensing, channel estimation of
massive MIMO, and data collection in WSNs.
General sparse representation methods, such as principal component analysis (PCA) and independent component analysis (ICA), aim to obtain a representation
that enables sufficient reconstruction. It has been demonstrated that PCA and ICA are able to deal with signal
corruption, such as noise, missing data, and outliers.
For sparse signals without measurement noise, CS can
recover the sparse signals exactly with random measurements. Furthermore, the random measurements significantly outperform measurements based on PCA and ICA
for the sparse signals without corruption [13]–[15]. In the
following, we will focus on the principles of CS and the
common sparse domains potentially used in 5G and IoT
scenarios.
3
A. Principles of Standard Compressive Sensing
The principles of standard CS, such as to be performed
at a single node, can be summarized into the following
three parts [3]:
1) Sparse Representation: Generally speaking, sparse
signals contain much less information than their ambient
dimension suggests. Sparsity of a signal is defined as the
number of non-zero elements in the signal under a certain
domain. Let f be an N -dimensional signal of interest,
which is sparse over the orthonormal transformation basis matrix Ψ ∈ RN ×N , and s be the sparse representation
of f over the basis Ψ. Then f can be given by
f = Ψs.
(1)
Apparently, f can be the time or space domain representation of a signal, and s is the equivalent representation
of f in the Ψ domain. For example, if Ψ is the inverse
Fourier transform (FT) matrix, then s can be regarded as
the frequency domain representation of the time domain
signal, f. Signal f is said to be K -sparse in the Ψ domain
if there are only K (K N ) out of the N coefficients
in s that are non-zero. If a signal is able to be sparsely
represented in a certain domain, the CS technique can
be invoked to take only a few linear and non-adaptive
measurements.
2) Projection: When the original signal f arrives at
the receiver, it is processed by the measurement matrix
Φ ∈ RP ×N with P < N , to get the compressed version
of the signal, that is,
x = Φf = ΦΨs=Θs,
(2)
where Θ = ΦΨ is an P × N matrix, called the sensing
matrix. As Φ is independent of signal f, the projection
process is non-adaptive.
Fig. 1 illustrates how the different sensing matrices Θ
influence the projection of a signal from high dimension
to its space, i.e., mapping s ∈ R3 to x ∈ R2 . As shown
in Fig. 1, s = s s 0 is a three-dimensional signal.
When s is mapped into a!two-dimensional space by
1 −1 0
taking Θ1 =
as the sensing matrix, the
0 0 1
original signal s cannot be recorded based on the projection under Θ1 . This is because that the plane spanned
by the two row vectors of Θ1 is orthogonal to signal s
as shown in Fig. 1(a). Therefore, Θ1 corresponds to the
worst projection. As shown in Fig. 1(b), we can also ob!
1 0 0
serve that the projection by taking Θ2 =
0 0 1
is not a good one. It is noted that the plane spanned
by the two row vectors of Θ2 can only contain part
of information of the sparse signal s, and the sparse
component in the direction of s2 is missed when the
signal s is projected into the two-dimensional space.
!
1 1 0
When the sensing matrix is set to Θ3 =
,
0 0 1
as shown in Fig. 1(c), the signal s can be fully recorded
as it falls into the plane spanned by the two row vectors
of Θ3 . Therefore, Θ3 results in a good projection and s
can be exactly recovered by its projection x in the twodimensional space. Then it is natural to ask what type of
projection is good enough to guarantee the exact signal
recovery?
The key of CS theory is to find out a stable basis Ψ or
measurement matrix Φ to achieve exact recovery of the
signal with length N from P measurements. It seems an
undetermined problem as P < N . However, it has been
proved in [4] that exact recovery can be guaranteed under
the following conditions:
• Restricted isometry property (RIP): Measurement
matrix Φ has the RIP of order K if
kΦf k2`2
≤ 1 + δK
(3)
1 − δK ≤
kf k2`2
holds for all K -sparse signal f , where δK is the
restricted isometry constant of a matrix Φ.
• Incoherence property: Incoherence property requires that the rows of measurement matrix Φ
cannot sparsely represent the columns of the sparsifying matrix Ψ and vice verse. More specifically,
a good measurement will pick up a little bit information of each component in s based on the
condition that Φ is incoherent with Ψ. As a result,
the extracted information can be maximized by
using the minimal number of measurements.
It has been pointed out that verifying both the RIP
condition and incoherence property is computationally
complicated but they could be achieved with a high
probability simply by selecting Φ as a random matrix.
The common random matrices include Gaussian matrix,
Bernoulli matrix, or almost all others matrices with
independent and identically distributed (i.i.d.) entries.
Besides, with the properties of the matrix with i.i.d.
entries Φ, the matrix Θ = ΦΨ is also random i.i.d.,
regardless of the choice of Ψ. Therefore, the random
matrices are universal as they are random enough to be
incoherent with any fixed basis. It has been demonstrated
that random measurements can universally capture the
information relevant for many compressive signal processing applications without any prior knowledge of
either the signal class and its sparse domain or the
ultimate signal processing task.
Moreover, for Gaussian matrices the number of measurements required to guarantee the exact signal recovery
4
s2
s2
s2
s
S = s
0
s
S = s
0
s1
s3
s1
s3
1 −1 0
Θ1 =
0 0 1
(a) Worst projection
s
S = s
0
s1
s3
1 0 0
Θ2 =
0 0 1
1 1 0
Θ3 =
0 0 1
(b) Bad projection
(c) Good projection
Fig. 1: Projection of a sparse signal with one non-zero component with different sensing matrices.
3) Signal Reconstruction: After the compressed measurements are collected, the original signal should be
reconstructed. Since most of the basis coefficients in s
are negligible, the original signal can be reconstructed by
finding out the minimal set of coefficients that matches
1
0.8
0.6
Pd
is almost minimal. However, random matrices inherently
have two major drawbacks in practical applications: huge
memory buffering for storage of matrix elements, and
high computational complexity due to their completely
unstructured nature [16]. Compared to the standard
CS that limits its scope to standard discrete-to-discrete
measurement architectures using random measurement
matrices and signal models based on standard sparsity,
more structured sensing architectures, named structured
CS, have been proposed to implement CS on feasible
acquisition hardware. So far, many efforts have been put
on the design of structured CS matrices, i.e., random
demodulator [17], to make CS implementable with expense of performance degradation. Particularly, the main
principle of random demodulator is to multiply the input
signal with a high-rate pseudonoise sequence, which
spreads the signal across the entire spectrum. Then a
low-pass anti-aliasing filter is applied and the signal is
captured by sampling it at a relatively low rate. With the
additional digital processing to reduce the burden on the
analog hardware, random demodulator bypasses the need
for a high-rate analogue-to-digital converter (ADC) [17].
A comparison of Gaussian sampling matrix and random
demodulator is provided in Fig. 2 in terms of detection
probability with different compression ratios P/N . From
the figure, the Gaussian sampling matrix performs better
than the random demodulator.
0.4
0.2
Random demodulator
Gaussian distributed matrix
0 -2
10
-1
10
P/N
0
10
Fig. 2: Detection probability versus compression ratio
with different measurement matrices. In this case, the
signal is one-sparse and the simulation iteration is 1000.
the set of compressed measurements x, that is, by solving
ŝ = arg min ksk`p subject to Θs = x,
s
(4)
where k·k`p is the `p -norm and p = 0 corresponds to
counting the number of non-zero elements in s. However,
the reconstruction problem in (4) is both numerically
unstable and NP-hard [3] when `0 -norm is used.
So far, there are mainly two types of relaxations to
problem (4) to find a sparse solution. The first type is
convex relaxation, where `1 -norm is used to substitute
`0 -norm in (4). Then (4) can be solved by standard
convex solvers, e.g., cvx. It has been proved that `1 norm
results in the same solution as `√
0 norm when RIP is
satisfied with the constant δ2k < 2 − 1 [18]. Another
type of solution is to use a greedy algorithm, such as
OMP [19], to find a local optimum in each iteration. In
comparison with the convex relaxation, the greedy algo-
5
rithm usually requires lower computational complexity
and time cost, which makes it more practical for wireless
communication systems. Furthermore, the recent result
has shown that the recovery accuracy achieved by some
greedy algorithms is comparable to the convex relaxation
but requiring much lower computational cost [20].
B. Reweighted Compressive Sensing
As aforementioned, `1 -norm is a good approximation
for the NP-hard `0 -norm problem when RIP holds. However, the large coefficients are penalized more heavily
than the small ones in `1 -norm minimization, which
leads to performance degradation on signal recovery.
To balance the penalty on the large and the small
coefficients, reweighted CS is introduced by providing
different penalties on those large and small coefficients.
A reweighted `1 -norm minimization framework [3] has
been developed to enhance the signal recovery performance with fewer compressed measurements by solving
ŝ = arg min kWsk`1 subject to Θs = x,
s
(5)
where W is a diagonal matrix with w1 , . . . , wn on the
diagonal and zeros elsewhere.
Moreover, `p -norm, e.g. 0 < p < 1, is utilized to lower
the computational complexity of signal recovery process
caused by the `1 -norm optimization problem. Iterative
reweighted least-square (IRLS) based CS approach has
been proposed in [21] to solve (4) in a non-convex
approach as
ŝ = arg min
s
N
X
wi si subject to Θs = x,
i=1
(l−1) p−2
where wi = si
(6)
where M is the the set of nodes in the network. As
stated in (2), Θm is the sensing matrix deployed at the
m-th node, and sm is a sparse signal of interest. DCS
becomes a standard CS when M = 1.
In the applications of standard CS, the signal received
at the same node has its sparsity property due to its intracorrelation. While for the networks with multiple nodes,
signals received at different nodes exhibit strong intercorrelation. The intra-correlation and inter-correlation of
signals from the multiple nodes lead to a joint sparsity
property. The joint sparsity level is usually smaller than
the aggregate over the individual signal’s sparsity level.
As a result, the number of compressed measurements
required for exact recovery in DCS can be reduced
significantly compared to the case performing standard
CS at each single node independently.
In DCS, there are two closely related concepts: distributed networks and distributed CS solvers. The distributed networks refer to networks that different nodes
perform data acquisition in a distributed way and the
standard CS can be applied at each node individually
to perform signal recovery. While for DCS solver as
proposed in [25], the data acquisition process requires no
collaboration among sensors and the signal recovery process is performed at several computational nodes, which
can be distributed in a network or locally placed within
a multiple core processor. Generally, it is of interest
to minimize both computation cost and communication
overhead in DCS. The most popular application scenario
of DCS is that all signals share the common sparse
support but with different non-zero coefficients.
is computed based on the result
(l−1)
of the last iteration, si
.
It is worth noting that (4) becomes non-convex when
p < 1. The existing algorithms cannot guarantee to
reach a global optimum and may only produce local
minima. However, it has been proved [22], [23] that
under some circumstances the reconstruction in (4) will
reach a unique and global minimizer [24], which is
exactly ŝ = s. Therefore, we can still exactly recover
the signal in practice.
C. Distributed Compressive Sensing
The distributed compressive sensing (DCS) [25] is an
extension of the standard one by considering networks
with M nodes. At the m-th node, measurement xm can
be given by
xm = Θm sm , ∀m ∈ M,
(7)
D. Common Sparse Domains for CS-enabled 5G and
IoT Networks
CS-enabled sub-Nyquist sampling is possible only if
the signal is sparse in a certain domain. The common
sparse domains utilized in CS-enabled 5G and IoT
networks include frequency domain, wavelet domain,
discrete cosine domain, angular domain, to name a few.
1) Frequency domain: due to low spectrum utilization, the wideband spectrum signal shows a sparse
property when it is converted into the frequency
domain.
2) Discrete cosine domain: due to temporal correlation, signals in some applications, such as environment information monitoring, show a sparse
property in the discrete cosine domain as the
readings normally do not change too much within
a short period.
3) Spatial domain: as the number of paths as well
as the angle-of-arrival are much smaller than the
6
Joint sparsity
Unoccupied
1) Spatial-frequency domain: An illustration of the
DCS-enabled cooperative CRN is given in Fig. 3,
where the joint sparsity in the spatial-frequency
domain is utilized. Specifically, each column represents the signal received at each location, which
is sparse in frequency domain as only a few
channels are occupied. At different locations, the
same frequency bands may be occupied, but the
signal powers for each frequency band are various
due to fading and shadowing. Therefore, different
columns of the matrix share the common sparse
support though each node operates without cooperation. With the DCS framework, each node
performs sub-Nyquist sampling individually first
and then the original signals can be recovered
simultaneously. More details on this issue will be
discussed in Section III.
2) Spatial-temporal domain: In WSNs, sensor nodes
are deployed to periodically monitor data and
to send the compressed data to the sink. Then
the sink is responsible for recovering the original
reading by CS algorithms as the readings across
all sensor nodes exhibit both spatial and temporal
correlations, as we can see from the discussion in
Section IV.
3) Angular-time domain: Massive MIMO channels
between some users and the massive base station
(BS) antennas appear the spatial common sparsity
in both the time domain and the angular domain,
as we will discuss in detail in Section V.
Different sparse domains and their applications in 5G
and IoT networks are summarized in Table I, where
‘WT’ is short for wavelet transform. In addition to the
listed sparse domains and their applications, it is worth
noting that the core of applying CS is to identify how
to exploit the sparse property in 5G and IoT networks.
In the following, we discuss three major applications of
CS in 5G and IoT networks.
...
...
...
...
...
For multi-node cases, due to the spatical correlation,
the joint sparsity is exploited to apply DCS in spatialx domains, where ‘x’ can be any of the aforementioned
domains. Here, we give some examples on how the joint
sparsity is utilized in different scenarios in 5G and IoT
networks:
J sensors at different locations
I channels
number of antennas in massive MIMO systems,
the channel conditions can be represented by a
limited number of parameters. In this case, the
spatial domain turns into the angular domain.
Occupied
Fig. 3: Spatial-frequency correlation.
III. C OMPRESSIVE S ENSING ENABLED C OGNITIVE
R ADIO N ETWORKS
In this section, we introduce the applications of CS
in CRNs. Radio frequency (RF) spectrum is a valuable
but tightly regulated resource due to its unique and
important role in wireless communications. The demand
for RF spectrum is increasing due to a rapidly expanding
market of multimedia wireless services while the usable
spectrum is becoming scarce due to the current rigid
spectrum allocation policies. However, according to the
reports from the Federal Communications Commission
(FCC) and the Office of Communications (Ofcom),
localized temporal and geographic spectrum utilization
is extremely low and unbalanced in reality. Cognitive
radio (CR) has become a promising solution to solve the
spectrum scarcity problem, by allowing unlicensed secondary users (SUs) to opportunistically access a licensed
band when the licensed primary user (PU) is absent. In
order to avoid any harmful interference to the PUs, SUs
in CRNs should be aware of the spectrum occupancy
over the spectrum of interest. Spectrum sensing, which
detects the spectrum holes, the frequency bands that are
not utilized by PUs, is one of the most challenging tasks
in CR. As the radio environment changes over time and
space, an efficient spectrum sensing technique should
be capable of tracking these fast changes [26]. A good
approach for detecting PUs is to adopting the traditional
narrowband sensing algorithms, which include energy
detection, matched-filtering, and cyclostationary feature
detection. Here, the term narrowband implies that the
frequency range is sufficiently narrow such that the
channel frequency response can be considered as flat.
In other words, the bandwidth of interest is less than the
coherence bandwidth of the channel [27].
While the present spectrum sensing algorithms have
7
TABLE I: Common sparse domains and their applications in 5G and IoT networks.
Sparse domain
Applications
Why sparse?
Sparsity property
Frequency domain
Sparsifying
transform
FT
Wideband spectrum sensing
Single sparsity
Spatial domain
-
Discrete cosine domain
Wavelet domain
Spatial-frequency
domain
Spatial-discrete
cosine/wavelet
domain
Angular-time domain
DCT
Channel estimation in massive MIMO
Sensor data gathering
Low spectrum utilization in practice
Number of paths is much fewer
than the number of antennas
Temporal correlation
Sensor data gathering
Cooperative wideband spectrum sensing
Active node detection/data
gathering
Temporal correlation
Spatial correlation and low spectrum utilization
Spatial and temporal correlation
Single sparsity
Joint sparsity
Channel estimation in massive MIMO
Number of paths and degrees of
arrival are much fewer than the
number of antennas
Joint sparsity
WT
FT
DCT/WT
-
focused on exploiting spectral opportunities over a narrow frequency band, future CRNs will eventually be
required to exploit spectral opportunities over a wide
frequency range from hundreds of megahertz (MHz) to
several gigahertz (GHz), in order to improve spectrum
efficiency and achieve higher opportunistic throughput.
As driven by the Nyquist-Shannon sampling theory,
a simple approach is to acquire the wideband signal
directly by a high-speed ADC, which is particularly
challenging or even unaffordable, especially for energyconstrained devices, such as smart phones or even
battery-free devices. Therefore, revolutionary wideband
spectrum sensing techniques become more than desired
to release the burden on high-speed ADCs.
A. Standard Compressive Spectrum Sensing
Recent development on CS theory inspires subNyquist sampling, by utilizing the sparse nature of
signals [3]. Driven by the inborn nature of the sparsity
property of signals in wireless communications, e.g., the
sparse utilization of spectrum, CS theory is capable of
enabling sub-Nyquist sampling for wideband spectrum
sensing.
1) Energy Detection based Compressive Spectrum
Sensing: CS theory has been applied to wideband spectrum sensing in [28], where sub-Nyquist sampling is
achieved without loss of any information. A general
framework for compressive spectrum sensing with the
energy detection method is summarized as shown in
Fig. 4, where the analog signal at the receiver, r (t) has
a sparse representation sf in the frequency domain. The
received signals are then sampled at a sub-Nyquist rate.
Due to low spectrum utilization, sf can be recovered
from the under-sampled measurements. Then the energy
Single sparsity
Single sparsity
Joint sparsity
of each channel can be calculated, and therefore the
spectrum occupancy can be determined.
Lately, it has been identified in [29] that the CSenabled system is somewhat sensitive to noise, exhibiting
a 3 dB SNR loss per octave of subsampling, which
parallels the classic noise-folding phenomenon. In order
to improve robustness to noise, a denoised compressive
spectrum sensing algorithm has been proposed in [6].
The sparsity level is required in advance in order to
determine the lower sampling rate at SUs locally without loss of any information. However, sparsity level
is dependent on spectrum occupancy, which is usually
unavailable in dynamic CR networks. In order to solve
this problem, a have proposed a two-step CS scheme
has been proposed in [30] to minimize the sampling rates
when the sparsity level is changing. In this approach, the
actual sparsity level is estimated first and the number
of compressed measurements to be collected is then
adjusted before sampling. However, this algorithm introduces extra computational complexity by performing
the sparsity level estimation. In order to avoid sparsity
level estimation, Sun et al. [31] have proposed to adjust
the number of compressed measurements adaptively by
acquiring compressed measurements step by step in
continuous sensing slots. However, this iterative process
incurs higher computational complexities at the SU as (4)
has to be solved several times until the exact signal
recovery is achieved. A low-complexity compressive
spectrum sensing algorithm has been proposed in [5]
by alleviating the iterative process of signal recovery.
More specifically, geolocation data can provide a rough
estimation of the sparsity level to minimize the sampling
rates. Subsequently, data from geolocation database is
utilized as the prior information for signal recovery. By
doing so, signal recovery performance is improved with
8
significant reduction on the computational complexity
and minimal number of measurements.
2) Compressive Power Spectral Density Estimation:
Different from the aforementioned approaches that concentrate on spectral estimation with perfect reconstruction of the original signals, compressive power spectral
density (PSD) estimation provides another approach for
spectrum detection without requiring complete recovery
of the original signals. Compressive PSD estimation
has been widely applied as the original signals are not
actually required in many signal processing applications. For wideband spectrum sensing applications with
spectrally sparse signals, only the PSD, or equivalently
the autocorrelation function, needs to be recovered as
only the spectrum occupancy status is required for each
channel.
Polo et al. [32] have proposed to reconstruct the
autocorrelation of the compressed signal to provide an
estimate of the signal spectrum by utilizing the sparsity
property of the edge spectrum, in which the CS is
performed directly on the wide-band analog signal. Nevertheless, the compressive measurements are assumed to
be wide-sense stationary in [32], which is not true for
some compressive measurement matrices. Subsequently,
Lexa et al. [33] have proposed a multicoset sampling
based power spectrum estimation method, by exploiting
the fact that a wide-sense stationary signal corresponds
to a diagonal covariance matrix of the frequency domain
representation of the signal. Additionally, Leus et al. [34]
have solved the power spectrum blind sampling problem
based on a periodic sampling procedure and have further proposed a simple least-square (LS) reconstruction
method for power spectrum recovery.
3) Beyond Sparsity: For spectrum blind sampling, the
goal is to perfectly reconstruct the spectrum and subNyquist rate sampling is only possible if the spectrum
is sparse. However, sub-Nyquist rate sampling can be
achieved in [34] without making any constraints on the
power spectrum, but the LS reconstruction requires some
rank conditions to be satisfied. Leus et al. [35] have
further proposed an efficient power spectrum reconstruction and a novel multicoset sampling implementation
by exploiting the spectral correlation properties, without
requiring any sparsity constraints on the power spectrum.
More recently, Cohen and Eldar [36] have developed a
compressive power spectrum estimation framework for
both sparse and non-sparse signals as well as blind and
non-blind detection in the sparse case. For each one
of those scenarios, the minimal sampling rate allowing
perfect reconstruction of the signal’s power spectrum is
derived in a noise-free environment.
B. Cooperative Spectrum Sensing with Joint Sparsity
In spectrum sensing, the performance is degraded by
noise uncertainty, deep fading, shadowing and hidden
nodes. Cooperative spectrum sensing (CSS) has been
proposed to improve sensing performance by exploiting
the collaboration among all the participating nodes. In
CRNs, a CSS network constructs a multi-node network.
As aforementioned, joint sparsity property and lowrank property can be utilized to recover the original
signals simultaneously with fewer measurements and the
DCS is utilized as it fits the CSS model perfectly. In
the existing literature, cooperative compressive spectrum
sensing mainly include two categories: i) centralized
approaches; ii) decentralized approaches.
A centralized approach involves a fusion center (FC)
performing signal recovery by utilizing the compressed
measurements contributed by the spatially distributed
SUs. In [6], a robust wideband spectrum sensing algorithm has been proposed for centralized CSS. Specifically, each SU senses a segment of the spectrum at
sub-Nyquist rate to reduce the sensing burden. With
the collected compressive measurements, CS recovery
algorithms are performed at the FC in order to recover
the original signals by exploiting the sparse nature
of spectrum occupancy. It is worth noting that the
sparse property of signals received at the SUs can be
transformed into the low-rank property of the matrix
constructed at the FC. In the case of CSS with subNyquist sampling, the measurements collected by the
participating SUs are sent to the FC, where the joint
sparsity or low-rank property is exploited to recover the
complete matrix.
A typical decentralized approach has been proposed
in [37], in which the decentralized consensus optimization algorithm can attain high sensing performance at a
reasonable computational cost and power overhead by
utilizing the joint sparsity property. Different from the
centralized approach, signal recovery or matrix completion is performed at each individual node in the
decentralized approaches. Compared with the centralized
approaches, the decentralized ones are more robust as it
adopts a FC-free network structure. Another advantage
of the decentralized approaches is that they allow the
recovery of individual sparse components at each node
as well as the common sparse components shared by all
participating nodes.
Furthermore, the privacy and security issues in CSS
have been investigated in [38] by exploiting joint sparsity
in the frequency and spatial domains. In [38], measurements corrupted by malicious users are removed during
the signal recovery process at the FC so that the recovery
9
r(t)
Compressive
sampling
x = Φr ( t )
−1
rf
Θ=ΦF
x
×
=
Px1
Spectrum
recovery
+
sˆ f
Decide spectrum
occupacy
Energy
detection
min sˆ f
lp
2
s.t. Θ ⋅ h f sˆ f − x ≤ ε
2
PxN
K non-zero entries
K<P<N
P<k<N
sf
wf
Nx1
Nx1
Fig. 4: Framework of compressive spectrum sensing with energy detection.
accuracy and security of the considered networks can be
improved.
C. Compressive Spectrum Sensing with Prior Information
In conventional compressive spectrum sensing, only
the sparsity property is utilized. Certain prior information
is available in some scenarios and can be exploited to
improve performance of wideband spectrum sensing in
CRNs. For example, in the case of spectrum sensing
over TV white space (TVWS), where the PUs are TV
signals and the transmitted waveforms are determined
by the standard, this prior information together with the
specifications dictated by the spectrum regulatory bodies,
i.e., carrier frequencies, bandwidths, can be also utilized
to enhance the signal recovery performance. Thus, it
is reasonable to assume that the PSD of the individual
transmission is known up to a scaling factor.
As discussed in II-B, reweighted CS normally introduces weights to provide different penalties on large and
small coefficients, which naturally inspires the application of reweighted CS in wideband spectrum sensing
with available prior information. In [39], the whole
spectrum is divided into different segments as the bounds
between different types of primary radios are known in
advance. Within each segment, an iteratively reweighted
`1 /`2 formulation has been proposed to recover the original signals. In [5], a low-complexity wideband spectrum
sensing algorithm for the TVWS spectrum has been
proposed to improve the signal recovery performance,
in which the weights are constructed by utilizing the
prior information from the geolocation database. For
example, in the TVWS spectrum, there are 40 TV
channels and each channel spans over 8 MHz that can
be either occupied or not. Hence, the TV signals show
a group sparsity property in the frequency domain as
the non-zero coefficients show up in clusters. A more
efficient approach has been developed in [40] by utilizing
such group sparsity property. Moreover, the signals in
wideband spectrum sensing have the following two characteristics: i) the input signals are stationary so that their
covariance matrices are redundant; ii) most information
in practical signals is concentrated on the first few lags
of the autocorrelation. Inspired by these characteristics, a
spectral prior information assisted structured covariance
estimation algorithm has been proposed in [41] with
low-computational complexity, which especially fits in
application on low-end devices.
D. Potential Research
We have reviewed some research results in CS-enabled
CRNs. There are still many open research issues in
the area, especially when practical constraints are considered. In this section, we will introduce a couple of
important ones.
1) Performance Limitations under Practical Constraints: Although there exist many research contributions in the field of compressive spectrum sensing, most
of them have assumed some ideal operating conditions.
In practice, there may exist various imperfection, such as
noise uncertainty, channel uncertainty, dynamic spectrum
occupancy, and transceiver hardware imperfection [11].
10
For example, the centralized compressive spectrum sensing normally considers ideal reporting channels, which
is not the case in practice. This imperfection may lead to
significant performance degradation in practice. Another
example comes from the measurement matrix design. As
shown in Fig. 2, the Gaussian distributed matrix achieves
better performance but with a higher implementation
cost. Even through some structured measurement matrices, such as random demodulator, with a lower cost
and acceptable recovery performance degradation, have
been proposed to enable the implementation of CS as a
replacement of high-speed ADCs, the nonlinear recovery
process limits its implementation. Therefore, it is a big
challenge to further investigate compressive spectrum
sensing in the presence of practical imperfection and to
develop a common framework to combat their aggregate
effects in CS-enabled CRNs.
2) Generalized Platform for Compressive Spectrum
Sensing: The existing hardware implementation of subNyquits sampling system follows the procedure that
the theoretic algorithm is specifically designed for the
current available hardware devices. However, it is very
difficult or sometimes even impossible to extend the
current hardware architectures to implement other existing compressive spectrum sensing algorithms. Thus,
it is desired to have a generalized hardware platform,
which can be easily adjusted to implement different
compressive spectrum sensing algorithms with different
types of measurement matrices and recovery algorithms.
IV. C OMPRESSIVE S ENSING ENABLED
L ARGE -S CALE W IRELESS S ENSOR N ETWORKS
WSNs provide the ability to monitor diverse physical
characteristics of the real world, such as sound, temperature, and humidity, by incorporating information and
communication technologies (ICT), which are especially
important to various IoT applications. In the typical setup
of WSNs, a large number of inexpensive and maybe
individually unreliable sensor nodes with limited energy
storage and low computational capability are distributed
in the smart environment to perform a variety of data
processing tasks, such as sensing, data collection, classification, modeling, and tracking. Cyber-physical systems
(CPSs) merge wireless communication technologies and
environment dynamics for efficient data acquisition and
smart environments control. Typically, a CPS consists
of a large number of sensor nodes and actuator nodes,
which monitor and control a physical plant, respectively,
by transmitting data to an elaboration node, named local
controller (LC) or FC.
Traditional environment information monitoring approaches take sensing samples at a predefined speed
uniformly at power-constrained sensor nodes and then
report the data to a LC/FC, which is normally powerful
and is capable of handling complex computations. The
data transmitted to the LC/FC usually have redundancies,
which can be exploited to reduce power consumption for
data transmission. A common and efficient method is to
compress at each individual sensor node and then transmit. However, data compression introduces additional
power consumption for individual sensor node although
the power consumption on data transmission is reduced.
Furthermore, this approach is unsuitable for real-time
applications owing to the high latency in data collection
and the high computational complexity to execute a
compression algorithm at the power-constrained sensor
node.
It is noted that most natural signals can be transformed
to a sparse domain, such as discrete cosine domain or
wavelet domain, where a small number of the coefficients
can represent most of the power of the signals that
used to be represented by a large number of samples
in their original domains. In fact, the data collected at
each sensor node show a sparsity property in the discrete
cosine domain or the wavelet domain due to the temporal
correlation. Inspired by this, CS can be applied at each
sensor node to collect the compressed measurements
directly and then send to the LC/FC or the neighbour
nodes. As a result, fewer measurements are sampled and
transmitted and the corresponding power consumption is
reduced significantly.
Power consumption of sensor nodes mainly comes
from sensing, data processing, and communications with
the LC/FC. In a large-scale WSN, the sensor nodes with
low power level would wish to take samples at a lower
speed or even turn themselves into the sleep mode in
order to extend lifetime. As the signal received at each
sensor node shows temporal correlation and neighboring
nodes show spatial correlation, the joint sparsity can be
exploited to recover signals from all sensor nodes even
though samples from part of the participating sensor
nodes are missed. The active sensor nodes can be preselected according to their power levels, therefore, with
invoking of CS in WSNs, lifetime of sensor nodes and
the whole network can be extended.
The existing work on CS-enabled WSNs mainly falls
into the aforementioned two categories: data gathering
and active node selection, which will be introduced
subsequently.
A. Data Gathering
In the traditional setting up of data gathering, there
are a large number of sensor nodes deployed in a WSN
11
to collect monitoring data. Each sensor node generates
a reading periodically and then sends it to the LC/FC,
which is normally powerful and capable of conducting
complex computation. As the sensor nodes are generally
limited in computation and energy storage, the data
gathering process in WSNs should be energy efficient
with low overhead, which becomes very challenging in
IoT scenarios where a huge number of sensor nodes
are deployed. Taking the temperature monitoring as an
example, the adjacent sensors will generate the similar
readings as the temperatures of nearby locations are similar. Furthermore, for each sensor node, the readings from
adjacent snapshots are close to each other. The above
two important observations indicate the temporal-spatial
correlations among temperature readings, which enables
the application of CS to reduce the network overhead
as well as to extend the network lifetime. Moreover,
such a joint sparsity is smaller than the aggregate over
the individual signal sparsity, which results in a further
reduction in the number of required measurements to
exactly recover the original signals.
Instead of applying compression on the data after
it is sampled and buffered, each sensor node collects
the compressed measurements directly by projecting the
signal to its sparse domain. At each individual sensor
node, one can naively obtain separate measurements of
its signal and then recover the signal for each sensor
separately at the LC/FC by utilizing the intra-signal
correlation. Moreover, it is also possible to obtain compressed measurements that each of them is a combination
of all signals from the cooperative sensor nodes in a
WSN. Subsequently, signals can be recovered simultaneously by exploiting both the inter-signal and intra-signal
correlations at the LC/FC.
1) Measurement Matrix Design: When adopting CS
techniques for data gathering in WSNs, sampling at
uniformly distributed random moments satisfies the RIP
if the sparse basis Ψ is orthogonal. For an arbitrary
sensor node i, the P × N measurement matrix can be a
spike one that only has P number of non-zero items, as
shown by
0 1 0 0 ... 0
0 0 0 1 ... 0
.
(8)
Φi =
..
.
0 0 0 0 ... 1
Sensor node will take a sample at the moment when the
corresponding item in Φi is one.
However, random sampling is not proper for practical
WSNs since two samples may be too close to each
other, which becomes very challenging for the cheap
sensor nodes. In order to solve this issue, Chen and
Wassell [7] have proposed a random sampling scheme
by utilizing the temporal correlation of signals received
at a sensor node. In the proposed scheme, the sensor
node will send a pseudorandom generator seed to the
FC and then send out the samples that are obtained at
an affordable highest rate until a sampling rate indicator
(SRI) is received from the FC. Here, the SRI is decided
based on the recovery accuracy calculated at the FC.
Once the recovery accuracy goes out the required range,
the sensor node will gradually increase its sampling
rate until the recovery error becomes acceptable. By
adopting such a scheme, the sensor node can adjust its
sampling rate adaptively without the knowledge of the
sparsity level. In order to further reduce the sampling
rate at sensor nodes, spatial correlation is exploited in
combination with temporal correlation. Therefore, joint
sparsity property can be exploited at the FC to reduce
the number of required measurements.
More recently, a big data enabled WSNs framework
has been proposed in [42], which invokes CS for data
completion with a minimal number of samples. The
proposed data collection framework consists of two core
components: i) at the cloud, an online learning component predicts the minimal amount of data to be collected
to reduce the amount of data for transmission, and these
data are considered as the principal data and their amount
is constrained by CS; ii) at each individual node, a
local control component tunes the collection strategy
according to the dynamics and unexpected environment
variation. Combining these two components, this framework can reduce power consumption and guarantee data
quality simultaneously.
2) Abnormal Sensor Detection: In the CS-enabled
data gathering processing, abnormal sensor readings may
still lead to severe degradation on signal recovery at the
LC/FC even if CS shows robustness to abnormal sensor
readings as it does not rely on the statistical distribution
of data to be preserved during runtime. This is because
that the abnormal readings will damage the sparsity
property of signals, as shown in Fig. 5. Therefore, it is
critical to find out those abnormal sensors to guarantee
the security of WSNs and make it abnormality-free.
Generally, abnormal readings are caused by either internal errors or external events according to their specific
patterns. Abnormal readings due to internal errors fail to
represent the sensed physical data, thus they should be
removed at least. But the abnormal readings caused by
the external events should be preserved as they reflect
the actual scenarios of WSNs.
Inspired by recovering data from over-complete dictionary, an abnormal detection mechanism has been
12
(a) Original signal
(c) Abnormal signal
(b) Discrete cosine transform of original signal
(d) Discrete cosine transform of abnormal signal
Fig. 5: Effect of abnormal readings on the sparsity level of temperature readings in discrete cosine domain.
proposed to enhance the compressibility of the signals.
First, the abnormal values are detected with the help
of recovering signal from an over-complete dictionary.
Second, the failing sensor nodes are categorized into
different types according to their patterns. Thirdly, the
failing nodes caused by the internal errors are removed
and then the data recovery is carried out to obtain the
ordinal data.
B. Active Node Selection
In large-scale WSNs, the events are relatively sparse in
comparison with the number of sensor nodes. Due to the
power constraint, it is unnecessary to activate all sensor
nodes at all the time. By utilizing the sparsity property
constructed by the spatial correlation, the number of
active sensor nodes in each time slot can be significantly reduced without scarifying performance. Taking
the smart monitoring system as an example as shown in
Fig. 6, the number of source nodes is N , and there are
K (K N ) sparse events that are generated by the N
source nodes. By invoking CS, only M (M ≤ N ) active
sensor nodes are required to capture the K sparse events.
1) Centralized Node Selection: A centralized node
selection approach has been proposed in [43] by applying CS and matrix completion at the LC/FC with
the purpose of optimizing the network throughput and
extending the lifetime of sensors. As a node is either
active or sleeping, the state index for a node becomes
binary, i.e., X ∈ {0, 1}. While conventional node selection in WSNs only exploits the spatial correlation
of sensor nodes, Chen and Wassell [44] have exploited
the temporal correlation by using the support of the
data reconstructed in the previous recovery period to
select the active nodes. Specifically, the FC performs
an optimized node selection, which is formulated as the
design of a specialized measurement matrix, where the
sensing matrix, Φ, consists of selected rows of an identity
matrix as shown in (8).
The sensing costs of taking samples from different
sensor nodes are assumed to be equal in most of the node
selection approaches. However, in WSNs with power-
13
Active node
Link in centralized network
Inactive node
Link in decentralized network
Fig. 6: Node selection in compressive sensing based wireless sensor networks.
constrained sensor nodes, this assumption does not hold
due to the different physical conditions at different sensor
nodes. For example, it is preferred to activate sensor
nodes with adequate energy rather than those almost
running out of energy to extend the lifetime of WSNs.
Therefore, a cost-aware node selection approach has
been proposed in [45] to minimize the sampling cost of
the whole WSNs with constraints on the reconstruction
accuracy.
2) Decentralized Node Selection: Different from the
DCS, which normally conducts signal recovery at the
FC by utilizing the data collected from the distributed
sensing sources via exploiting the joint sparsity, decentralized CS-enabled WSNs approach aims to achieve innetwork processing for node selection. A decentralized
approach has been proposed in [46] to perform node
selection by allowing each active sensor node to monitor and recover its local data only, by collaborating
with its neighboring active sensor nodes through onehop communication, and iteratively improving the local
estimation until reaching the global optimum. It should
be noted that an active sensor node not only optimizes
for itself, but also for its inactive neighbors. Moreover,
in order to extend the network lifetime, an in-network
CS framework has been developed in [47] by enabling
each sensor node to make an autonomous decision on the
data compression and forward strategy with the purpose
of minimizing the number of packets to be transmitted.
Generally speaking, the drawbacks for distributed
node selection approach come from the following two
aspects: i) the optimized node selection requires an
iterative process, which may require a long period;
ii) the flexibility to vary the number of active sensor
nodes is limited, especially according to the dynamic
sparsity levels or the channel conditions, which could be
time-varying. While for the centralized approach, extra
bandwidth resource and power consumption are required
to coordinate active sensor nodes.
C. Potential Research
Even though extensive research has been carried out to
investigate the application of CS in WSNs, most of them
have focused on reducing power consumption at sensor
nodes and extending the network lifetime. However, in
large-scale WSNs for different IoT applications, big data
should be exploited to enhance CS recovery accuracy in
addition to further reduce the power consumption.
1) Machine Learning aided Adaptive Measurement
Matrix Design: The core concept for sparse representation is the same even though different applications
bring different constraints. Therefore, it is straightforward to ask if there is a general framework for the
sparse representation of the data in different 5G and
IoT applications for urban scenarios. In order to take
the minimal number of samples from the set of sensor
nodes with the best capability, i.e., highest power levels,
the measurement matrix should be properly designed.
It has been demonstrated that machine learning can be
an efficient tool to aid the measurement matrix design
so that the lifetime of the whole network as well each
14
individual sensor node can be extended to the most.
Furthermore, it should be noted that when designing a
measurement matrix, the possibility of implementation
in real network is one of the most critical factors to be
considered. We believe that extensive research work in
this direction is highly desired.
2) Data Privacy in CS-enabled IoT Networks: The
sensing data collected from a variety of featured sensors
in IoT networks, our daily activities, surroundings, and
even the physical information, can be recorded and
analyzed, which at the same time greatly intensifies the
risk of privacy exposure. There are few effective privacypreserving mechanisms in mobile sensing systems. In
order to provide privacy protection, privacy-preserving
noise has been proposed to be added to the original
data to guarantee effective privacy. The added noise will
dominate the data when the original sparse data points
are zero or near zero, thus reducing the data sparsity.
However, CS aims to achieve kind of efficiency for
sparse data processing. Then it comes into a conflict
countable between privacy and efficiency during big
data processing in CS-enabled IoT networks. Therefore,
extensive research work is expected in this area.
V. C HANNEL ACQUISITION AND P RECODING IN
M ASSIVE MIMO
To satisfy high data rate requirements caused by
increasing mobile applications, a lot of efforts have been
made to improve transmission SE. An effective way is to
exploit the spatial degrees of freedom (DoF) provided by
large-scale antennas at the transmitter and the receiver to
form massive MIMO systems [48]. It has been shown in
[48] that the spatial resolution of a large-scale antenna
array will be very high and the channels corresponding
to different users are approximately orthogonal when
the number of the antennas at the BS are very large.
Consequently, linear processing is good enough to make
the system performance to approach optimum if the CSI
is known at the BS.
Accurate CSI at the BS is essential for massive MIMO
to obtain the above advantage. Due to the large channel
dimension, downlink CSI acquisition in massive MIMO
systems sometimes becomes challenging even if uplink
CSI estimation is relatively simple. In time-division duplexing (TDD) systems, the downlink CSI can be easily
obtained by exploiting channel reciprocity. However,
most of the practical deployed systems mainly employ
frequency-division duplex (FDD), where the channel
reciprocity does not hold any more. In this situation, the
downlink channel has to be estimated directly and then
fed back to the BS, which will result in the extremely
high overhead.
To address CSI estimation and feedback issue in FDD
systems, sparsity of massive MIMO channels must be
exploited. Some CS-enabled CSI acquisition methods
for FDD massive MIMO have been proposed, where
the correlation in massive MIMO channels has been
successfully exploited to reduce the amount of training
symbols and feedback overhead.
In this section, we focus on the CS-enabled channel
acquisition and its related applications. We will first
introduce the channel sparsity feature and then discuss
channel estimation and feedback, and precoding and
detection subsequently. It should be noted that mmWave
communications are often used with massive MIMO
techniques since short wavelength here makes it very
easy to pack a large number of antennas in a small
area. The channel acquisition and precoding based on
CS in mmWave massive MIMO will be also included in
the following discussion even if mmWave channels are
slightly different from the traditional wireless channels.
A. Sparsity of Channels
In the channel acquisition and precoding schemes
based on CS, the key idea is to use the channel sparsity.
Although the channel sparsity in massive MIMO generally exists in the time domain, the frequency domain,
and the spatial domain, we mainly focus on the spatial
domain channel sparsity in this section.
In conventional MIMO systems, a rich-scattering multipath channel model is often assumed so that the channel
coefficients can be modelled as independent random
variables. However, this assumption is not true any more
in massive MIMO systems. It has been shown that the
massive MIMO channel is spatial correlated and has a
sparse structure in the spatial domain. This correlation
and sparsity is due to two reasons: the exploitation of
high radio frequency and the deployment of large-scale
antenna arrays in future wireless communications. In
high frequency band, the channels have fewer propagation paths while more transmit and receive antennas will
make the distinguished paths to be much fewer than the
number of channel coefficients, which makes the rich
scatters to become limited or sparse.
As shown in Fig. 7, a classical channel model with
limited scatterers at the BS is often used in the literature [49]. In this model, different user channels have a
partially common sparsity support due to the shared scatterers and an independent sparsity support caused by the
individual scatterers in the propagation environments. By
using this sparsity structure, many CS-enabled channel
acquisition and precoding schemes have been proposed,
as we will discuss in the following.
15
Scatters for user 1
at BS side
Local scatters for
user 1
Initial channel,H k
Sparsity representation, Hk
H1 AH1
Common scatters for
all users
BS with large
scale antennas
Individual parameters
Common parameters
Local scatters for
user K
Scatters for user
K at BS side
H K AHK
Non-zero elements
Zero elements
Fig. 7: Channel model with limited scatters, where A is the spatial correlation matrix.
B. Compressive Pilot Design
In order to obtain a good channel estimation, the
length of the orthogonal training sequence must be at
least same as the number of transmit antenna elements.
Due to a huge number of antennas at the BS in massive
MIMO systems, the downlink pilots occupy a high
proportion of the resources. Consequently, the traditional
pilot design is not applicable here. It is necessary to
design specific pilots to reduce the training overhead
in FDD massive MIMO systems. It has been shown
that channel spatial correlation or sparsity can be used
to shrink the original channel to an effective one with
a much lower dimension so that low-overhead training
is enough in massive MIMO systems. Based on this
principle, CS-enabled pilot design schemes have been
proposed in [50].
When exploiting the CS theory to acquire the CSI in
the correlated FDD MIMO systems, how much training should be sent is a very important question. Once
the amount of the training is determined, the training
symbols can be designed by using of the channel common or/and individual sparsity in the correlated massive
MIMO systems. Besides using the common support, a
new pilot structure with joint common and dedicated
support has been proposed in [50], where the common
one is used to estimate the common channel parameters
of the related users in multi-user massive MIMO systems
while the dedicated one is used to estimate the individual
channel parameters of each user.
Besides the general massive MIMO systems, a compressed hierarchical multiresolution codebook has also
been designed specially for mmWave massive MIMO
systems to construct training beamforming vectors [51].
Based on this idea, a lot of different hierarchical multiresolution codebooks have been proposed from different angles. For example, in [52], a compressive beacon
codebook with different sets of pseudorandom phases
has been designed. In [53], a common codebook satisfying the conflicting design requirements as well as
validating practical mmWave systems has been proposed through utilizing the strong directivity of mmWave
channels. In [54], a multi-resolution uniform-weighting
based codebook, with similar to an normalized DFT
matrix, has been proposed to reduce the implementation
complexity, where the estimation of the angle of arrival
(AoA) and angle of departure (AoD) will have a much
lower training overhead. Therefore, the codebook based
beamforming training procedure can achieve a good
balance between complexity and high performance for
the practical systems.
C. Compressive Channel Estimation and Feedback
CS can be used to reduce the cost of the channel estimation and feedback by exploiting the channel
sparsity. The existing CS based channel estimation and
feedback schemes can be divided into the following three
categories according to exploiting sparsity in different
domains.
16
1) With Time Domain Sparsity: It has been shown that
the channel is slowly-varying in various applications so
that the prior channel estimation result can be still used
to reconstruct the subsequent channels, as illustrated
in Fig. 8. By using the structure in the figure, several
CS algorithms have been proposed in [8], [55], [56] to
recover massive MIMO channels.
Exploiting the common time-domain support, a CSenabled differential CSI estimation and feedback scheme
has been proposed in [55]. The scheme in [56] further
combines LS and CS techniques to improve estimation
performance. In this scheme, the LS and CS techniques
can be used to estimate a dense vector obtained by
projecting into the previous support and a sparse vector
obtained by projecting into the null space of the previous
support, respectively. Since the current channel vector
has only a small number of non-zero elements outside
the support of the previous channel, the proposed scheme
can reduce the pilot overhead and improve the tracking
performance of the channel subspace. Since the quality
of the prior support information will affect the estimation
accuracy, a greedy pursuit-based approach with the prior
support information and its quality information has been
developed in [8], where the prior support information is
adaptively exploited based on its quality information to
further improve the channel estimation performance.
Besides channel estimation, several channel feedback
schemes have been also proposed in [55] and [49]. To exploit the common support, a CS-enabled differential CSI
feedback scheme has been developed in [55] by using the
temporal correlation of MIMO channels. The proposed
scheme can reduce the feedback overhead by about 20%
compared with the direct CS-enabled channel feedback.
In [49], a robust closed-loop pilot and channel state
information at the transmitter (CSIT) feedback resource
adaptation framework has been developed by using the
temporal correlation of multiuser massive MIMO channels. In this framework, the pilot and feedback resources
can be adaptively adjusted for successful CSIT recovery
under unknown and time-varying channel sparsity levels.
2) With Spatial Domain Sparsity: In practice, the
antenna spacing in massive MIMO is usually set to be
half-wave length to keep the array aperture compact.
Furthermore, the BS with a large-scale antenna array is
generally deployed at the top of high buildings such that
there are only limited local scatters [57]. In this case,
the large-scale MIMO channels exhibit strong angulardomain sparsity or spatial sparsity. This channel sparsity
property can be used to reduce the channel estimation
and feedback overhead in FDD massive MIMO systems.
In [58], a block optimization algorithm is employed
to extract the common angular support information from
the channel matrices. The extracted common support
information is then used to form the weighted factors
and design a weighted block optimization algorithm to
estimate the channel matrix. In [59], a spatial sparsitybased compression mechanism has been proposed to
reduce the load of the channel feedback. In this mechanism, random projection with unknown sparsity basis
and direct compression based on known sparsity basis
are used. Since the spatial sparsity will reduce channel
rank, a dictionary learning method, which captures the
communication environment and the antennas property,
has been proposed to obtain the compressed channel
estimation.
After the compressed hierarchical multiresolution
codebook is designed in [51]–[54], the corresponding
channel estimation schemes can be developed. In these
schemes, the hierarchical multi-resolution codebook can
be capable of generating variable beam width radiation
patterns to facilate the usage of robust adaptive multipath
channel estimation algorithms. Meanwhile, the exploitation of adaptive compressive sensing algorithm will also
reduce implementation complexity and estimation error.
3) With Spatial-Temporal Sparsity: In the above
schemes, the temporal-domain sparsity and the spatialdomain one are independently exploited. In practice, they
can be jointly used to further reduce the cost of channel
estimation and feedback.
In [60], a structured-CS enabled differential joint
channel training and feedback scheme has been proposed, where a structured compressive sampling matching pursuit (S-CoSaMP) algorithm uses the structured
spatial-time sparsity of wireless MIMO channels to
reduce the training and feedback overhead. In [61],
a Bayesian CS based feedback mechanism has been
proposed for time-varying spatially and temporally correlated vector auto-regression wireless channels, where
the feedback rate can be adaptively adjusted.
4) With Spatial-Frequency Sparsity: Besides the
spatial-temporal sparsity, the spatial-frequency one can
also be exploited to reduce the cost of channel estimation and feedback. In [62], the sparsity in the spatialfrequency domain is first exploited and an adaptive CSenabled feedback scheme is correspondingly proposed to
reduce the feedback overhead. In this scheme, the feedback can be dynamically configured based on the channel
conditions to improve the efficiency. Due to sharing
sparse common support for the adjacent subcarriers in
orthogonal frequency division modulation (OFDM), an
approximate message passing with the nearest neighbor
sparsity pattern learning (AMP-NNSPL) algorithm has
been developed to adaptively learn the underlying structure to obtain better performance.
17
Zero entry
Non-zero entry
Measurement matrix
NM L
Sparse basis of channel
LG
Received signals
Sparse representation
y g w
Estimate ĝ based on
pursuit algorithms
h g
Support from Support from
prior signals current signals
(a) Support from prior and current signals
Channel reconstruction
h gˆ
Received signals
corresponding to N
received symbols
y r0T
T
rLT
NM 1
(b) Compressive channel estimation by using time-domain correlation
Fig. 8: Illustration of compressive channel estimation with time-domain correlation signal support.
D. Precoding and Detection
The precoder design based on the estimated CSI is
a very important problem in massive MIMO systems,
especially in the mmWave wideband systems. Since
wideband mmWave massive MIMO channel generally
exhibits frequency selective fading, the precoder design
based on the estimated CSI will become challenging.
Generally, the sparse structure of mmWave massive
MIMO channels in angle domain or beam space can
be used to simplify the precoder design [63], [64]. In
[63], compressive subspace estimation is used to get the
full channel information and then design the precoder
to maximize the system SE. In order to reduce the CSI
acquisition cost and address the mismatch between few
radio frequency chains and many antennas, the hybrid
precoding design with baseband and radio frequency
precoders has been used in mmWave massive MIMO
systems. However, this hybrid precoding will introduce
performance loss. In order to mitigate performance loss,
an iterative OMP has been utilized to refine the quality
of hybrid precoders [64]. Meanwhile, a limited feedback
technique has also been proposed for hybrid precoding
to reduce the feedback cost.
CS can be also used in signal detection of massive
MIMO with spatial modulation (SM). In massive SMMIMO, the maximum likelihood (ML) detector has a
prohibitively high complexity. Thanks to the structured
sparsity of multiple SM signals, a low-complexity signal
detector based on CS has been introduced to improve
signal detection performance. In [65], a joint SM transmission scheme for user equipment and a structured
CS-enabled multi-user detector for the BS have been
developed, the proposed detector can reliably detect the
resultant SM signals with low complexity by exploiting
the intrinsical sparse features.
In the above, we mainly discuss CSI acquisition and
precoding based on CS theory in FDD massive MIMO
systems. In practice, the CS theory can be also applied to
TDD massive MIMO systems. A channel estimation approach based on block-structured CS has been developed
in [66], where the common support in sparse channels
and the channel reciprocity in TDD mode are used
simultaneously so that the computational complexity and
pilot overhead can be reduced significantly.
E. Potential Research
As we can see from the above discussion, CS has
been successfully used in massive MIMO to improve
the performance of channel estimation and precoding.
However, there are still many open topics before implementing CS in massive MIMO systems. We will discuss
some of them in this section.
1) Effect of Antenna Deployment: Due to space limitation, large-scale antenna may be deployed at various
topologies, i.e., centralized or distributed. Since the
different antenna topologies corresponding to different
channel sparsity, the effect of antenna configuration on
the performance of CS-enabled channel estimation is still
open.
2) Measure of Channel Sparsity: As mentioned before, the channel sparsity is very important to CS-enabled
channel estimation. In the current research works, various sparsity models have been assumed and exploited
in channel acquisition. Many of these assumed sparsity
models lack of verification by measurement results. It is
desired that either the sparsity model or the CS-enabled
approach can be confirmed by some measurement results
under different propagation conditions.
3) Channel Estimation and Feedback with Joint Support: In the above channel estimation and feedback
schemes, one or two of the three types of channel
sparsity are used to improve the channel estimation
performance and reduce the feedback overhead. If all
three types of channel sparsity are used, can we further
improve performance? How much performance gain can
be achieved if possible?
18
4) Channel Estimation and Feedback with Unknown
Support: Channel sparsity directly affects CS-enabled
channel estimation performance. However, it is still a
challenging problem how to get the information on
channel sparsity support. At the same time, it is also
worth to study how channel estimation and feedback
schemes should be designed without channel sparsity
support.
VI. OTHER C OMPRESSIVE S ENSING A PPLICATIONS
In addition to the aforementioned applications, CS
has been applied in various other areas in wireless
communications. Some of them are introduced in this
section.
A. Compressive Sensing Aided Localization
In a multiple target localization network, the multiple
target locations can be formulated as a sparse matrix
in the discrete spatial domain. By exploiting the spatial
sparsity, instead of recording all received signal strengths
(RSSs) over the spatial grid to construct a radio map
from targets, only far fewer number of RSS measurements need to be collected at runtime. Subsequently, the
target locations can be recovered from the collected measurements through solving an `1 minimization problem.
As a result, the multiple target locations can be recovered
accordingly.
B. Compressive Sensing Aided Impulse Noise Cancellation
In certain applications, such as OFDM, impulsive
noise will degrade the system performance severely.
OFDM signal is often processed in the frequency domain. Even if impulse noise only lasts a short period
of time, it affects a wide frequency range. By regarding
impulse noise as a sparse vector, CS technique has been
exploited to mitigate such type of impulse noise [67].
Such a group sparsity property inspires us to apply CS
to active RRH selection in green C-RANs to minimize
the network power consumption [68], [69]. Additionally,
in the uplink of C-RANs, the channel estimation from the
active users to the RRHs is the key to achieve the spatial
multiplexing gain. Generally, the number of active users
is low in C-RANs, which makes it possible to apply
CS to reduce the uplink training overhead for channel
estimation. Moreover, the correlation among active users
at different RRHs exhibit a joint sparsity property, which
can further facilitate the active user detection and channel
estimation in C-RANs [70].
VII. C ONCLUDING R EMARKS
This article has provided a comprehensive overview
of sparse representation with applications in wireless
communications. Specifically, after introducing the basic
principles of CS, the common sparse domains in 5G
and IoT networks have been identified. Subsequently,
three CS-enabled networks, including wideband spectrum sensing in CRNs, data collection in IoT networks,
and channel estimation and feedback in massive MIMO
systems, have been discussed by exploiting different
sparsity properties. From the above discussions, it has
been concluded that by invoking of CS, the SE and EE
of 5G and IoT networks can be enhanced from different
perspectives. Furthermore, potential research challenges
have been identified to provide a guide for researchers
interested in the sparse representation in 5G and IoT
networks.
ACKNOWLEDGMENT
Jiancun Fan is supported by the National Natural
Science Foundation of China under Grant 61671367,
the China Postdoctoral Science Foundations under Grant
2014M560780 and Grant 2015T81031. Yue Gao is
supported by funding from Physical Sciences Research Council (EPSRC) in the U.K. with Grant No.
EP/R00711X/1.
C. Compressive Sensing Aided Cloud Radio Access Networks
R EFERENCES
Cloud radio access networks (C-RANs) has been
proposed as a promising technology to support massive
connectivity in 5G networks. In C-RAN, the BSs are
replaced by remote radio heads (RRHs) and connected
to a central processor via digital backhaul links. Thanks
to the spatial and temporal variation of the mobile traffic,
it is feasible to switch off some RRHs with guarantee on the quality of service in green C-RANs. More
specifically, one RRH will be switched off only when
all the coefficients in its beamformer are set to zeros.
[1] M. Zibulevsky and M. Elad, “L1-L2 optimization in signal and
image processing,” IEEE Signal. Proc. Mag., vol. 27, no. 3, pp.
76–88, May 2010.
[2] A. M. Bruckstein, D. L. Donoho, and M. Elad, “From sparse
solutions of systems of equations to sparse modeling of signals
and images,” SIAM review, vol. 51, no. 1, pp. 34–81, Feb. 2009.
[3] E. Candes, “Compressive sampling,” in Proc. Int. Congress
Math., vol. 3, Madrid, Spain, Oct. 2006, pp. 1433–1452.
[4] E. J. Candes, J. Romberg, and T. Tao, “Robust uncertainty
principles: Exact signal reconstruction from highly incomplete
frequency information,” IEEE Trans. Inf. Theory, vol. 52, no. 2,
pp. 489–509, Feb. 2006.
19
[5] Z. Qin, Y. Gao, and C. G. Parini, “Data-assisted low complexity
compressive spectrum sensing on real-time signals under subNyquist rate,” IEEE Trans. Wireless Commun., vol. 15, no. 2,
pp. 1174–1185, Feb. 2016.
[6] Z. Qin, Y. Gao, M. D. Plumbley, and C. G. Parini, “Wideband
spectrum sensing on real-time signals at sub-Nyquist sampling
rates in single and cooperative multiple nodes,” IEEE Trans.
Signal Process., vol. 64, no. 12, pp. 3106–3117, Jun. 2016.
[7] W. Chen and I. J. Wassell, “Energy-efficient signal acquisition
in wireless sensor networks: a compressive sensing framework,”
IET Wireless Sensor Systems, vol. 2, no. 1, pp. 1–8, Mar. 2012.
[8] X. Rao and V. K. Lau, “Compressive sensing with prior support
quality information and application to massive MIMO channel
estimation with temporal correlation,” IEEE Trans. Signal Process., vol. 63, no. 18, pp. 4914–4924, Sep. 2015.
[9] E. J. Candes and M. B. Wakin, “An introduction to compressive
sampling,” IEEE Signal. Proc. Mag., vol. 25, no. 2, pp. 21–30,
Mar. 2008.
[10] C. R. Berger, Z. Wang, J. Huang, and S. Zhou, “Application
of compressive sensing to sparse channel estimation,” IEEE
Commun. Mag., vol. 48, no. 11, pp. 164–174, Nov. 2010.
[11] S. K. Sharma, E. Lagunas, S. Chatzinotas, and B. Ottersten,
“Application of compressive sensing in cognitive radio communications: A survey,” IEEE Commun. Surveys Tuts., vol. 18,
no. 3, pp. 1838–1860, thirdquarter 2016.
[12] D. Romero, D. D. Ariananda, Z. Tian, and G. Leus, “Compressive covariance sensing: Structure-based compressive sensing
beyond sparsity,” IEEE Signal. Proc. Mag., vol. 33, no. 1, pp.
78–93, Jan. 2016.
[13] H. S. Chang, Y. Weiss, and W. T. Freeman, “Informative sensing
of natural images,” in IEEE International Conference on Image
Processing (ICIP), Cairo, Egypt, Nov. 2009, pp. 3025–3028.
[14] J. Wright, Y. Ma, J. Mairal, G. Sapiro, T. S. Huang, and
S. Yan, “Sparse representation for computer vision and pattern
recognition,” Proc. IEEE, vol. 98, no. 6, pp. 1031–1044, Jun.
2010.
[15] H. S. Chang, Y. Weiss, and W. T. Freeman, “Informative
sensing,” CoRR, vol. abs/0901.4275, 2009. [Online]. Available:
http://arxiv.org/abs/0901.4275
[16] E. Candes and J. Romberg, “Sparsity and incoherence in
compressive sampling,” Inverse problems, vol. 23, no. 3, p. 969,
Apr. 2007.
[17] J. Tropp, J. Laska, M. Duarte, J. Romberg, and R. Baraniuk,
“Beyond nyquist: Efficient sampling of sparse bandlimited
signals,” IEEE Trans. Inf. Theory, vol. 56, no. 1, pp. 520–544,
Jan. 2010.
[18] E. J. Cands, “The restricted isometry property and its
implications for compressed sensing,” Comptes Rendus
Mathematique, vol. 346, no. 9, pp. 589 – 592,
2008. [Online]. Available: http://www.sciencedirect.com/
science/article/pii/S1631073X08000964
[19] J. A. Tropp and A. C. Gilbert, “Signal recovery from random
measurements via orthogonal matching pursuit,” IEEE Trans.
Inf. Theory, vol. 53, no. 12, pp. 4655–4666, Dec. 2007.
[20] J. W. Choi, B. Shim, Y. Ding, B. Rao, and D. I. Kim,
“Compressed sensing for wireless communications: Useful tips
and tricks,” IEEE Commun. Surveys Tuts., vol. 19, no. 3, pp.
1527–1550, thirdquarter 2017.
[21] B. D. Rao and K. Kreutz-Delgado, “An affine scaling methodology for best basis selection,” IEEE Trans. Signal Process.,
vol. 47, no. 1, pp. 187–200, Jan. 1999.
[22] R. Chartrand, “Exact reconstruction of sparse signals via nonconvex minimization,” IEEE Signal Process. Lett., vol. 14,
no. 10, pp. 707–710, Oct. 2007.
[23] R. Chartrand and V. Staneva, “Restricted isometry properties
and nonconvex compressive sensing,” Inverse Problems, vol. 24,
no. 3, p. 035020, May. 2008.
[24] R. Chartrand and W. Yin, “Iteratively reweighted algorithms for
compressive sensing,” in Proc. IEEE Int. Conf. Acoust. Speech
Signal Process.(ICASSP), Las Vegas, NV, Mar. 2008, pp. 3869–
3872.
[25] D. Baron, M. F. Duarte, M. B. Wakin, S. Sarvotham, and R. G.
Baraniuk, “Distributed compressive sensing,” 2009.
[26] Y.-C. Liang, Y. Zeng, E. Peh, and A. T. Hoang, “Sensingthroughput tradeoff for cognitive radio networks,” IEEE Trans.
Wireless Commun., vol. 7, no. 4, pp. 1326–1337, Apr. 2008.
[27] H. Sun, A. Nallanathan, C.-X. Wang, and Y. Chen, “Wideband
spectrum sensing for cognitive radio networks: a survey,” IEEE
Wireless Commun., vol. 20, no. 2, pp. 74–81, Mar. 2013.
[28] Z. Tian and G. Giannakis, “Compressed sensing for wideband
cognitive radios,” in Proc. IEEE Int. Conf. Acoust. Speech
Signal Process.(ICASSP), Honolulu, HI, Apr. 2007, pp. 1357–
1360.
[29] M. A. Davenport, J. N. Laska, J. R. Treichler, and R. G.
Baraniuk, “The pros and cons of compressive sensing for
wideband signal acquisition: Noise folding versus dynamic
range,” IEEE Trans. Signal Process., vol. 60, no. 9, pp. 4628–
4642, Sep. 2012.
[30] Y. Wang, Z. Tian, and C. Feng, “Sparsity order estimation and
its application in compressive spectrum sensing for cognitive
radios,” IEEE Trans. Wireless Commun., vol. 11, no. 6, pp.
2116–2125, Jun. 2012.
[31] H. Sun, W.-Y. Chiu, and A. Nallanathan, “Adaptive compressive
spectrum sensing for wideband cognitive radios,” IEEE Commun. Lett., vol. 16, no. 11, pp. 1812–1815, Nov. 2012.
[32] Y. L. Polo, Y. Wang, A. Pandharipande, and G. Leus, “Compressive wide-band spectrum sensing,” in Proc. IEEE Int. Conf.
Acoust. Speech Signal Process.(ICASSP), Taipei, Taiwan, Apr.
2009, pp. 2337–2340.
[33] M. A. Lexa, M. E. Davies, J. S. Thompson, and J. Nikolic,
“Compressive power spectral density estimation,” in Proc. IEEE
Int. Conf. Acoust. Speech Signal Process.(ICASSP), Prague,
Czech Republic, May 2011, pp. 3884–3887.
[34] G. Leus and D. D. Ariananda, “Power spectrum blind sampling,” IEEE Signal Process. Lett., vol. 18, no. 8, pp. 443–446,
Aug. 2011.
[35] D. D. Ariananda and G. Leus, “Compressive wideband power
spectrum estimation,” IEEE Trans. Signal Process., vol. 60,
no. 9, pp. 4775–4789, Sep. 2012.
[36] D. Cohen and Y. C. Eldar, “Sub-Nyquist sampling for power
spectrum sensing in cognitive radios: A unified approach,” IEEE
Trans. Signal Process., vol. 62, no. 15, pp. 3897–3910, Aug.
2014.
[37] F. Zeng, C. Li, and Z. Tian, “Distributed compressive spectrum
sensing in cooperative multihop cognitive networks,” IEEE J.
Sel. Areas Signal. Proc., vol. 5, no. 1, pp. 37 –48, Feb. 2011.
[38] Z. Qin, Y. Gao, and M. D. Plumbley, “Malicious user detection
based on low-rank matrix completion in wideband spectrum
sensing,” IEEE Trans. Signal Process., vol. 66, no. 1, pp. 5–17,
Jan. 2018.
[39] Y. Liu and Q. Wan, “Enhanced compressive wideband frequency spectrum sensing for dynamic spectrum access,”
EURASIP J. Adv. Signal Process., vol. 2012, no. 1, p. 177,
Dec. 2012.
[40] Y. C. Eldar, P. Kuppinger, and H. Bolcskei, “Block-sparse
signals: Uncertainty relations and efficient recovery,” IEEE
Trans. Signal Process., vol. 58, no. 6, pp. 3042–3054, Jun. 2010.
[41] D. Romero and G. Leus, “Wideband spectrum sensing from
compressed measurements using spectral prior information,”
20
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
[56]
[57]
[58]
[59]
IEEE Trans. Signal Process., vol. 61, no. 24, pp. 6232–6246,
Dec. 2013.
L. Kong, D. Zhang, Z. He, Q. Xiang, J. Wan, and M. Tao, “Embracing big data with compressive sensing: A green approach
in industrial wireless networks,” IEEE Commun. Mag., vol. 54,
no. 10, pp. 53–59, Oct. 2016.
Z. Qin, Y. Liu, Y. Gao, M. Elkashlan, and A. Nallanathan,
“Wireless powered cognitive radio networks with compressive
sensing and matrix completion,” IEEE Trans. Commun., vol. 65,
no. 4, pp. 1464–1476, Apr. 2017.
W. Chen and I. J. Wassell, “Optimized node selection for
compressive sleeping wireless sensor networks,” IEEE Trans.
Veh. Technol., vol. 65, no. 2, pp. 827–836, Feb. 2016.
——, “Cost-aware activity scheduling for compressive sleeping
wireless sensor networks,” IEEE Trans. Signal Process., vol. 64,
no. 9, pp. 2314–2323, May 2016.
Q. Ling and Z. Tian, “Decentralized sparse signal recovery for
compressive sleeping wireless sensor networks,” IEEE Trans.
Signal Process., vol. 58, no. 7, pp. 3816–3827, Jul. 2010.
C. Caione, D. Brunelli, and L. Benini, “Distributed compressive
sampling for lifetime optimization in dense wireless sensor
networks,” IEEE Trans. Ind. Informat., vol. 8, no. 1, pp. 30–40,
Feb. 2012.
T. L. Marzetta, “Noncooperative cellular wireless with unlimited numbers of base station antennas,” IEEE Trans. Wireless
Commun., vol. 9, no. 11, pp. 3590–3600, Nov. 2010.
A. Liu, F. Zhu, and V. K. Lau, “Closed-loop autonomous
pilot and compressive CSIT feedback resource adaptation in
multi-user FDD massive MIMO systems,” IEEE Trans. Signal
Process., vol. 65, no. 1, pp. 173–183, Jan. 2017.
V. K. Lau, S. Cai, and A. Liu, “Closed-loop compressive
CSIT estimation in FDD massive MIMO systems with 1 bit
feedback,” IEEE Trans. Signal Process., vol. 64, no. 8, pp.
2146–2155, Apr. 2016.
A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, “Channel
estimation and hybrid precoding for millimeter wave cellular
systems,” IEEE J. Sel. Areas Signal. Proc., vol. 8, no. 5, pp.
831–846, 2014.
Z. Marzi, D. Ramasamy, and U. Madhow, “Compressive channel estimation and tracking for large arrays in mm-wave picocells,” IEEE J. Sel. Areas Signal. Proc., vol. 10, no. 3, pp.
514–527, 2016.
J. Song, J. Choi, and D. J. Love, “Common codebook millimeter wave beam design: Designing beams for both sounding
and communication with uniform planar arrays,” IEEE Trans.
Commun., vol. 65, no. 4, pp. 1859–1872, 2017.
G. Wang, S. A. Naeimi, and G. Ascheid, “Low complexity channel estimation based on dft for short range communication,” in
Proc. IEEE Int. Conf. Commun. (ICC). IEEE, 2017, pp. 1–7.
W. Shen, L. Dai, Y. Shi, X. Zhu, and Z. Wang, “Compressive sensing-based differential channel feedback for massive
MIMO,” Electr. Lett., vol. 51, no. 22, pp. 1824–1826, Oct. 2015.
Y. Han, J. Lee, and D. J. Love, “Compressed sensing-aided
downlink channel training for FDD massive MIMO systems,”
IEEE Trans. Commun., vol. 65, no. 7, pp. 2852–2862, Jul. 2017.
L. Lu, G. Li, A. Swindlehurst, A. Ashikhmin, and R. Zhang,
“An overview of massive MIMO: Benefits and challenges,”
IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 742–758,
Oct. 2014.
C. C. Tseng, J. Y. Wu, and T. S. Lee, “Enhanced compressive
downlink CSI recovery for FDD massive MIMO systems using
weighted block-minimization,” IEEE Trans. Commun., vol. 64,
no. 3, pp. 1055–1067, Mar. 2016.
M. S. Sim, J. Park, C.-B. Chae, and R. W. Heath, “Compressed
channel feedback for correlated massive MIMO systems,” J. of
Commun. and Netw., vol. 18, no. 1, pp. 95–104, Feb. 2016.
[60] W. Shen, L. Dai, Y. Shi, B. Shim, and Z. Wang, “Joint channel
training and feedback for FDD massive MIMO systems,” IEEE
Trans. Veh. Technol., vol. 65, no. 10, pp. 8762–8767, Oct. 2016.
[61] X.-L. Huang, J. Wu, Y. Wen, F. Hu, Y. Wang, and T. Jiang,
“Rate-adaptive feedback with Bayesian compressive sensing
in multiuser MIMO beamforming systems,” IEEE Trans. on
Wireless Commun., vol. 15, no. 7, pp. 4839–4851, Jul. 2016.
[62] P.-H. Kuo, H. Kung, and P.-A. Ting, “Compressive sensing
based channel feedback protocols for spatially-correlated massive antenna arrays,” in Proc. Wireless Commun. Netw. Conf.
(WCNC), Paris, France, Apr. 2012, pp. 492–497.
[63] K. Venugopal, N. Gonzlez-Prelcic, and R. W. Heath, “Optimality of frequency flat precoding in frequency selective millimeter
wave channels,” IEEE Wireless Commun. Lett., vol. 6, no. 3, pp.
330–333, Jun. 2017.
[64] J. Mirza, B. Ali, S. S. Naqvi, and S. Saleem, “Hybrid precoding
via successive refinement for millimeter wave MIMO communication systems,” IEEE Commun. Lett., vol. 21, no. 5, May
2017.
[65] Z. Gao, L. Dai, Z. Wang, S. Chen, and L. Hanzo, “Compressivesensing-based multiuser detector for the large-scale SM-MIMO
uplink,” IEEE Trans. Veh. Technol., vol. 65, no. 10, pp. 8725–
8730, Oct. 2016.
[66] Y. Nan, L. Zhang, and X. Sun, “Efficient downlink channel estimation scheme based on block-structured compressive sensing
for tdd massive MU-MIMO systems,” IEEE Wireless Commun.
Lett., vol. 4, no. 4, pp. 345–348, Aug. 2015.
[67] A. B. Ramirez, R. E. Carrillo, G. Arce, K. E. Barner, and
B. Sadler, “An overview of robust compressive sensing of sparse
signals in impulsive noise,” in European Signal Process. Conf.
(EUSIPCO), Aug. 2015, pp. 2859–2863.
[68] Y. Shi, J. Zhang, and K. B. Letaief, “Group sparse beamforming
for green cloud-RAN,” vol. 13, no. 5, pp. 2809–2823, May
2014.
[69] B. Dai and W. Yu, “Sparse beamforming and user-centric
clustering for downlink cloud radio access network,” IEEE
Access, vol. 2, pp. 1326–1339, Oct. 2014.
[70] X. Xu, X. Rao, and V. K. N. Lau, “Active user detection and
channel estimation in uplink cran systems,” in Proc. IEEE Int.
Conf. Commun. (ICC), London, UK, Jun. 2015, pp. 2727–2732.
| 7cs.IT
|
THE GENERIC CHARACTER TABLE OF A SYLOW p-SUBGROUP OF
A FINITE CHEVALLEY GROUP OF TYPE D4
arXiv:1508.06937v2 [math.RT] 10 Feb 2017
SIMON M. GOODWIN, TUNG LE AND KAY MAGAARD
Abstract. Let U be a Sylow p-subgroup of the finite Chevalley group of type D4 over
the field of q elements, where q is a power of a prime p. We describe a construction of the
generic character table of U .
1. Introduction
Let p be a prime and let q be a power of p. Let G be a finite reductive group over the
field of q elements, and let U be a Sylow p-subgroup of G. The aim of this paper is to
describe a construction of the generic character table of U for the case where G is split and
of type D4 . This generic character table can be determined immediately from the sequence
of propositions and lemmas in Section 5. The construction is “uniform” over primes p > 2,
but we observe differences for the bad prime p = 2.
There has been much research into constructing generic character tables of G. For G of
small rank these have been programmed into the computing system CHEVIE, [10]. More
recently there has been interest in generic character tables for Borel and parabolic subgroups
of G, see for example [17], [19], [21], [22] and [25], where many cases for G of Fq -rank 3 or less
are considered. These papers and similar methods have led to applications in determining
decomposition numbers in the modular character theory of G, see for example, [1], [18], [20],
[23], [26], [34] and [41].
There has been a great deal of research interest in the character theory of U , with motivation coming from understanding how this character theory varies with q.
Interest in properties of U for G = GLn (q) goes back to G. Higman [16], where the
conjugacy classes of U are considered. In [31], G. Lehrer determined how discrete series
characters decompose when restricted to U for G = GLn (q). This led to a study of the
irreducible characters of U and in particular it was conjectured that the degrees of the
irreducible characters are always a power of q, and that the number of characters of degree
q d , for d ∈ Z≥0 is given by a polynomial in q with integer coefficients; this refines a well
known conjecture attributed to Higman. There has been further work in this direction, for
G = GLn (q), by I. M. Isaacs, and more generally on the character theory of algebra groups,
see [27] and [28]. In particular, Isaacs verified that the degree of any irreducible character
of an algebra group over Fq is a power of q. For other recent developments we refer, for
example, to the work of A. Evseev [8], E. Marberg [32], and I. Pak and A. Soffer [35]; in
particular, we remark that the results of [35] suggest that the aforementioned conjecture on
the number of characters being given by a polynomial in q is false. An interesting aspect of
the character theory of U is the theory of supercharacters, which was first studied by André,
see for example [2]; this theory was fully developed by Diaconis and Isaacs in [6].
1
There has been much interest on the complex irreducible characters of Sylow p-subgroups
of finite reductive groups of other types. For instance, it has been proved that for G of type
B, C or D the degrees of all irreducible characters of U (q) are powers of q if and only if
p 6= 2, see [36], [37] and [40]. There has been much further interest in determining certain
families of irreducible character, see for example [15], [33], [38] and [39]. Recently, a notion
of supercharacters of U for G of types B, C and D was introduced by André and Neto in
[3]. For p larger than the Coxeter number of G, the Kirillov orbit method gives a bijection
between the coadjoint orbits of U on its Lie algebra and the irreducible characters of U ; in
[13] an algorithm is given, which is used to calculate a parametrization of these coadjoint
orbits when G is split and of rank less than or equal to 8, excluding type E8 . We refer the
reader also to [30] for other results on the parametrization of irreducible characters of U .
From now on we suppose G is a finite Chevalley group of type D4 . The two key ingredients
that enable us to calculate the generic table of U are the papers [24] and [14], which provide
a parametrization of the irreducible characters and the conjugacy classes of U ; we also use
[4] for the conjugacy classes for the case p = 2. Our construction of this generic character
table is given in such a way that it is straightforward to convert it to a computer file that
can be used in future calculations.
Our parametrization of the characters and conjugacy classes of U partitions them into
families that are independent of q, as can be observed in Tables 3 and 4. We have refined
the parameterizations from [14] and [24], so that they are stable under the automorphisms
of G coming from the graph automorphisms of the Dynkin diagram; this is likely to be of
significant benefit for applications. Our calculations are based on developing the methods
from [24]. A key additional ingredient is that we modify those constructions making them
more explicit by realizing the irreducible characters of U as induced characters of linear
characters of certain subgroups.
By the generic character table we mean the table with rows labelled by the families of the
characters, and columns labelled by the families of representatives of conjugacy classes. The
values in the table are given by expressions depending on the parameters for the families
and on q. This can be viewed as an extension of the ideas discussed above regarding how
the character theory of U varies with q.
We observe that the characters and their values are uniform for p > 2, whereas for p =
2 there are differences. More explicitly the parametrization of the family of irreducible
characters F8,9,10 is different for p = 2, and although the parametrization of the characters
in the families F11 and F12 is not different for p = 2, the values of these characters do differ.
These differences give some explanation of why p = 2 is a bad prime for D4 .
We summarize our results in the following theorem.
Theorem 1.1. Let G be a a finite Chevalley group of type D4 over Fq and let U be a Sylow
p-subgroup of G. The generic character table of U is determined from: the parametrization
of the conjugacy classes given in Section 3; the parametrization of the irreducible characters
given in Section 4; and the character values given in Section 5.
The methods in this paper serve as a model, which can be applied to determine generic
character tables of Sylow p-subgroups of finite groups of Lie type, where p is the defining
characteristic.
We outline the structure of paper. First we recall the required theory and notation regarding the finite Chevalley group of type D4 in Section 2. Then in Sections 3 and 4, we
2
recall the results from [4], [12] and [24] giving the parametrization of the conjugacy classes
and irreducible characters of U . The majority of the work is done in Section 5, where we
determine the character values.
Acknowledgements: We would like to thank the referee for many helpful suggestions that
led to improvements in the paper. We also thank A. Paolini for pointing out some corrections.
Part of this research was completed during a visit of the second author to the University of
Birmingham; we thank the LMS for a grant to cover his expenses.
2. Notation and Setup
Let p be a prime, let q be a power of p, and denote the finite field with q elements by
Fq . We let G denote a simply connected simple algebraic group of type D4 defined and split
over Fp ; in other words G is Spin8 defined over Fp . Let B be a Borel subgroup of G defined
over Fp , let U be the unipotent radical of B, and T ⊆ B be a maximal torus defined over
Fp . For a subgroup H of G defined over Fp , we write H = H(q) for the group of Fq -rational
points of H. So G = G(q) is a finite Chevalley group of type D4 and U = U(q) is a Sylow
p-subgroup of G. We cite [7, Sec. 3], as a reference for the theory of algebraic groups over
finite fields, and for the terminology used here.
+
∼
We note that G = Spin+
8 (q) is not a simple group, but that G/Z(G) = P Ω8 (q) is a simple
group. As we are mainly interested in U in this paper, which is isomorphic to its image in
G/Z(G), this distinction causes us no harm.
Let Φ be a root system of G with respect to T and let Φ+ be the system of positive
roots determined by B. We write {α1 , α2 , α3 , α4 } for the corresponding base of simple roots
ordered so that the Dynkin diagram of Φ is
t
α1@
@
@t
α2
α3
t
α4
t
For α =
1
2 1
1
P4
i=1
ai αi ∈ Φ we write ht(α) =
P4
i=1
ai for the height of α. We use the notation
for the root α1 + α2 + 2α3 + α4 ∈ Φ+ and we use a similar notation for the remaining
positive roots. We enumerate the roots in Φ+ as shown in Table 1 below.
The triality automorphism of the Dynkin diagram can be lifted to an automorphism of G
that stabilizes T and B, and restricts to an automorphism of G. We refer the reader to [5,
Section 4.2] for more information and let τ denote this automorphism of G. In a similar way
we can define three automorphisms of G corresponding to the involutory automorphisms of
the Dynkin diagram given by switching two of the outer vertices. We denote these by σ12 ,
σ14 and σ24 in the obvious way.
The group G is generated by the root elements xα (t) for α ∈ Φ and t ∈ Fq . We
may choose these root elements so that, for α, β ∈ Φ such that α + β ∈ Φ, we have
[xα (t), xβ (u)] = xα+β (±tu) and the signs are invariant under τ and the σij . For positive roots,
we use the abbreviation xi (t) := xαi (t) and let Xi = {xi (t) | t ∈ Fq }, for i = 1, 2, . . . , 12.
We have τ (xi (t)) = xiρ (t) where ρ is the permutation ρ = (1, 4, 2)(5, 7, 6)(8, 9, 10); and
3
Height Roots
1
α1
2
3
4
5
α2
1
α5 := 1 0
0
1
α8 := 1 0
1
1
α11 := 1 1
1
1
α12 := 2 1
1
α3
α6 :=
α9 :=
0
1 0
1
1
1 1
0
α4
0
α7 := 1 1
0
0
α10 := 1 1
1
Table 1. Positive roots of the root system Φ of type D4 .
have a similar description for the automorphisms σij . The commutators [xi (t), xj (u)] =
xi (t)−1 xj (u)−1 xi (t)xj (u) are given in Table 2. All [xi (t), xj (u)] not listed in this table are
equal to 1. We refer the reader to [5, Chapter 4] for details on root elements and commutator
relations.
[x1 (t), x3 (u)]
[x1 (t), x7 (u)]
[x2 (t), x3 (u)]
[x2 (t), x7 (u)]
[x3 (t), x4 (u)]
[x4 (t), x5 (u)]
[x4 (t), x8 (u)]
[x6 (t), x9 (u)]
=
=
=
=
=
=
=
=
x5 (tu),
x9 (tu),
x6 (tu),
x10 (tu),
x7 (−tu),
x9 (tu),
x11 (tu),
x12 (−tu),
[x1 (t), x6 (u)]
[x1 (t), x10 (u)]
[x2 (t), x5 (u)]
[x2 (t), x9 (u)]
[x3 (t), x11 (u)]
[x4 (t), x6 (u)]
[x5 (t), x10 (u)]
[x7 (t), x8 (u)]
=
=
=
=
=
=
=
=
x8 (tu),
x11 (tu),
x8 (tu),
x11 (tu),
x12 (tu),
x10 (tu),
x12 (−tu),
x12 (−tu).
Table 2. Commutator relations for type D4 .
We have that U is the subgroup of GQ
generated by the elements xi (ti ) for i = 1, 2, . . . , 12
and ti ∈ Fq . In fact we have that U = 12
i=1 Xi . It turns out that it is convenient for us to
write our elements of U with an element from X3 at the front, and we use the notation
(2.1)
x(t) = x3 (t3 )x1 (t1 )x2 (t2 )x4 (t4 )x5 (t5 ) · · · x12 (t12 ),
where t = (t1 , . . . , t12 ).
There is a further piece of notation that we use frequently in the sequel, we define the
normal subgroups Mi of U for i = 1, . . . , 13 by
(2.2)
M1 = U,
M2 = X1 X2
12
Y
Xj ,
j=4
M3 = X2
12
Y
j=4
4
Xj and
(2.3)
Mi =
12
Y
Xj
for i = 4, . . . , 13.
j=i
3. Conjugacy Classes and their Representatives
In [14], an algorithm for calculating the conjugacy classes of Sylow p-subgroups of finite
Chevalley groups in good characteristic is described; so we assume p > 2 for the first part
of this section. This algorithm was implemented in GAP4 [9] and, in particular, used to
determine the conjugacy classes of U = U (q). The conjugacy class representatives are given
as minimal representatives as defined in [11] and discussed below.
(For ease of exposition we temporarily let x1 = xα3 , x2 = xα1 and x3 = xα2 in the next
paragraph.)
The idea behind the calculation of minimal representatives is to recursively determine the
orbits of U on U/Mi+1 for i = 1, . . . , 12; recall that Mi+1 is defined in (2.2). To explain
this we consider the set A(x, i) = {xxi (t)Mi+1 | t ∈ Fq } ⊆ U/Mi+1 for x ∈ U . The
centralizer CU (xMi ) acts on A(x, i) by conjugation. The key results for the theory of minimal
representatives are [11, Lem. 5.1 and Prop. 6.2], which imply the following dichotomy:
(I) all elements of A(x, i) are conjugate under CU (xMi ); or
(R) no two elements of A(x, i) are conjugate under CU (xMi ).
We say that i is an inert point of x if (I) holds and i is a ramification point of x if (R) holds.
We say x(t) as defined in (2.1) is the minimal representative of its conjugacy class if ti = 0
whenever i is an inert point of x(t). Thanks to [11, Prop. 5.4 and Prop. 6.2], the minimal
representatives of conjugacy classes give a complete set of representatives of the conjugacy
classes in U .
In Table 3, we present the minimal representatives of the conjugacy classes in U (q). Starting with the minimal representatives calculated as in [14], we have parameterized the conjugacy classes giving representatives partitioned into families. This has been done to reduce
the number of families required to what appears to be an optimal number.
The first column of Table 3 gives a label for the family, then in the second column a
parametrization of the elements of the family is given, where the notational convention is
that an ai is an element of F×
q = Fq \ {0}, a bi is any element of Fq , and for a sequence of the
form ci1 , . . . , cir the cij are elements of Fq , which are not all zero. These parameters have to
satisfy the conditions given in the third column. In the fourth column we give the number
of elements in the family. The families are constructed so that the size of the centralizer
of an element in the family does not depend on the choice, and the final column gives this
centralizer size.
The families of representatives have been chosen, so that they are permuted by the automorphisms τ , σ12 , σ14 and σ24 . Thus we can read off the action of these automorphisms
on the conjugacy classes. However, for the family, C1,2,4,q6 , it is not possible for p = 3 to get
this stability under the triality automorphism and we have to distinguish the cases p = 3
and p > 3; we note that the family given for p = 3 is actually also valid for p > 3.
To end this section we explain the differences in the parametrization of the conjugacy
classes for p = 2, where the theory of minimal representatives fails. First we note that there
are a number of families, where there is a relation between parameters involving a sum of two
5
Label
C1,2,3,4
Representative
x3 (a3 )x1 (a1 )x2 (a2 )x4 (a4 )
Conditions
Number
(q − 1)4
Centralizer
q4
C1,2,3
C1,3,4
C2,3,4
x3 (a3 )x1 (a1 )x2 (a2 )x9 (b9 )x10 (b10 )
x3 (a3 )x1 (a1 )x4 (a4 )x8 (b8 )x10 (b10 )
x3 (a3 )x2 (a1 )x4 (a4 )x8 (b8 )x9 (b9 )
b9 /a1 + b10 /a2 = 0
b8 /a1 + b10 /a4 = 0
b8 /a2 + b9 /a3 = 0
(q − 1)3 q
(q − 1)3 q
(q − 1)3 q
q5
q5
q5
C1,3
C2,3
C3,4
x3 (a3 )x1 (a1 )x10 (b10 )
x3 (a3 )x2 (a2 )x9 (b9 )
x3 (a3 )x4 (a4 )x8 (b8 )
(q − 1)2 q
(q − 1)2 q
(q − 1)2 q
q5
q5
q5
C1,2,4,q6 (p > 3)
C1,2,4,q6 (p = 3)
C1,2,4,q7
x1 (a1 )x2 (a2 )x4 (a4 )x5 (c5 )x6 (c6 )x7 (c7 )
x1 (a1 )x2 (a2 )x4 (a4 )x6 (c6 )x7 (c7 )
x1 (a1 )x2 (a2 )x4 (a4 )x12 (b12 )
c5 /a1 + c6 /a2 + c7 /a4 = 0
(q − 1)3 (q 2 − 1)
(q − 1)3 (q 2 − 1)
(q − 1)3 q
q6
q6
q7
C1,2,q6
C1,2,q7
C1,2,q8
C1,4,q6
C1,4,q7
C1,4,q8
C2,4,q6
C2,4,q7
C2,4,q8
x1 (a1 )x2 (a2 )x5 (c5 )x6 (c6 )x7 (c7 )
x1 (a1 )x2 (a2 )x9 (a9 )x10 (a10 )
x1 (a1 )x2 (a2 )x12 (b12 )
x1 (a1 )x4 (a4 )x5 (c5 )x6 (c6 )x7 (c7 )
x1 (a1 )x4 (a4 )x8 (a8 )x10 (a10 )
x1 (a1 )x4 (a4 )x12 (b12 )
x2 (a2 )x4 (a4 )x5 (c5 )x6 (c6 )x7 (c7 )
x2 (a2 )x4 (a4 )x8 (a8 )x9 (a9 )
x2 (a2 )x4 (a2 )x12 (b12 )
c5 /a1 + c6 /a2 = 0
a9 /a1 + a10 /a2 = 0
(q − 1)2 (q 2 − 1)
(q − 1)3
(q − 1)2 q
(q − 1)2 (q 2 − 1)
(q − 1)3
(q − 1)2 q
(q − 1)2 (q 2 − 1)
(q − 1)3
(q − 1)2 q
q6
q7
q8
q6
q7
q8
q6
q7
q8
C1,q6
C1,q7
C1,q8
C2,q6
C2,q7
C2,q8
C4,q6
C4,q7
C4,q8
x1 (a1 )x6 (c6 )x7 (c7 )
x1 (a1 )x10 (a10 )
x1 (a1 )x12 (b12 )
x2 (a2 )x5 (c5 )x7 (c7 )
x2 (a2 )x9 (a9 )
x2 (a2 )x12 (b12 )
x4 (a4 )x5 (c5 )x6 (c6 )
x4 (a4 )x8 (a8 )
x4 (a4 )x12 (b12 )
(q − 1)(q 2 − 1)
(q − 1)2
(q − 1)q
(q − 1)(q 2 − 1)
(q − 1)2
(q − 1)q
(q − 1)(q 2 − 1)
(q − 1)2
(q − 1)q
q6
q7
q8
q6
q7
q8
q6
q7
q8
C3
x3 (a3 )x8 (b8 )x9 (b9 )x10 (b10 )x11 (b11 )
(q − 1)q 4
q8
1)3 q
q8
1)3
(q −
(q − 1)2 q
(q − 1)3
(q − 1)2 q
(q − 1)3
(q − 1)2 q
q8
q9
q8
q9
q8
q9
C5,6,7
c5 /a1 + c7 /a4 = 0
a8 /a1 + a10 /a4 = 0
c6 /a2 + c7 /a4 = 0
a8 /a2 + a9 /a4 = 0
(q −
x5 (a5 )x6 (a6 )x7 (a7 )x11 (b11 )
C5,6,q8
C5,6,q9
C5,7,q8
C5,7,q9
C6,7,q8
C6,7,q9
x5 (a5 )x6 (a6 )x9 (a9 )x10 (a10 )
x5 (a5 )x6 (a6 )x11 (b11 )
x5 (a5 )x7 (a7 )x8 (a8 )x10 (a10 )
x5 (a5 )x7 (a7 )x11 (b11 )
x6 (a6 )x7 (a7 )x8 (a8 )x9 (a9 )
x6 (a6 )x7 (a7 )x11 (b11 )
C5,q8
C5,q9
C6,q8
C6,q9
C7,q8
C7,q9
x5 (a5 )x10 (a10 )
x5 (a5 )x11 (b11 )
x6 (a6 )x9 (a9 )
x6 (a6 )x11 (b11 )
x7 (a7 )x8 (a8 )
x7 (a7 )x11 (b11 )
(q − 1)2
(q − 1)q
(q − 1)2
(q − 1)q
(q − 1)2
(q − 1)q
q8
q9
q8
q9
q8
q9
C8,9,10
x8 (c8 )x9 (c9 )x10 (c10 )
q3 − 1
q 10
C11
x11 (a11 )
q−1
q 11
C12
x12 (b12 )
q
q 12
a9 /a5 + a10 /a6 = 0
a8 /a5 + a10 /a7 = 0
a8 /a6 + a9 /a7 = 0
Table 3. Representatives of conjugacy classes in U (q) for p > 2
terms and this has to be replaced by just setting one of the parameters equal to zero. For example in C1,2,3 , we have to replace the representatives simply with x3 (a3 )x1 (a1 )x2 (a2 )x10 (b10 )
and there are no conditions. This does lead to the families of representatives not being stable
under the automorphisms, but this cannot be helped.
The more significant differences are as follows.
6
p=2
• The family C1,2,3,4 is replaced by the family C1,2,3,4
. The representatives are of the
form x3 (a3 )x1 (a1 )x2 (a2 )x4 (a4 )x10 (d10 ), where d10 is either 0, or an element of Fq not
in the image of the map t 7→ a3 a2 a4 (t2 + t). So there are 2(q − 1)4 representatives in
this family. Their centralizers have order 2q 4 .
• The family C1,2,4,q6 should be included as in the p = 3 case in the table above.
• The family C1,2,4,q7 is replaced by two families.
p=2
– C1,2,4,2q
7 . The representatives are of the form x1 (a1 )x2 (a2 )x4 (a4 )x10 (a10 )x12 (d12 ),
where d12 is either 0, or an element of Fq not in the image of the map t 7→
a1 a2 a4 t2 + a1 a10 t. So there are 2(q − 1)4 representatives in this family. Their
centralizers have order 2q 7 .
p=2
3
– C1,2,4,q
7 . The representatives are of the form x1 (a1 )x2 (a2 )x4 (a4 ). There are (q−1)
representatives in this family and their centralizers have order q 7 .
• The family C5,6,7 is replaced by two new families namely.
p=2
– C5,6,7,2q
8 . The representatives are of the form x5 (a5 )x6 (a6 )x7 (a7 )x10 (a10 )x11 (d11 ),
where d11 is either 0, or an element of Fq not in the image of the map t 7→
a5 a6 a7 t2 + a5 a10 t. There are 2(q − 1)4 representatives in this family and their
centralizers have order 2q 8 .
p=2
– C5,6,7,q
8 . The representatives are of the form x5 (a5 )x6 (a6 )x7 (a7 ). The number of
representatives in this family is (q − 1)3 and their centralizers have order q 8 .
4. The irreducible characters of U
In [24], the irreducible characters of U were determined giving a partition into families,
which are given in Table 4 for the case p > 2. We have modified these constructions to
make them more explicit, so that we can realize them all as induced characters from linear
characters of certain subgroups of U . In the next section we calculate the values of the
characters, and in doing so we explain their construction.
We explain the contents of Table 4. The first column gives the name of the family and the
second column gives the names of the characters in that family. The notational convention
is that parameters denoted by ai range over F×
q = Fq \ {0}, and parameters denoted by bi
range over Fq . The subscripts in the notation give an idea of “where these parameters act”,
as can be understood from the descriptions of how the characters are constructed given in
the next section. For some families there is a relation between some of the parameters and
this is given in the third column. The fourth column records the number of characters in
that family. The families are constructed in such a way that the degree of the characters in
a family is constant, and the fifth column gives this degree.
The triality automorphism τ of G, permutes the families of characters by acting on the
subscripts of the families by the permutation ρ; similarly the automorphisms σ12 , σ14 and
σ24 act on the families. As mentioned earlier, we have updated the parametrization from [24,
§4], so that we can read off the action of these automorphisms on the characters. Though it
is not possible to do this for the family F5,6,7 for p = 3.
We end this section by noting that for p = 2, the parametrization of irreducible characters
is slightly different. As explained in [24], the family denoted F8,9,10 is more complicated
p=2
and has to be replaced by the family F8,9,10
, which consists of two types of characters as
explained below and in more detail at the end of §5.4.
7
Family
Notation
F12
Number
Degree
χa1212 ,b1 ,b2 ,b4
q 3 (q − 1)
q4
F11
χa1111 ,b5 ,b6 ,b7 ,b3
q 4 (q − 1)
q3
F8,9,10
8 ,a9 ,a10 ,b3
χa8,9,10
q(q − 1)3
q3
F8,9
8 ,a9 ,a6 ,a7
χa8,9,q
3
a6 /a8 + a7 /a9 = 0
(q − 1)3
q3
8 ,a9 ,b2 ,b3 ,b4
χa8,9,q
2
b2 /a8 + b4 /a9 = 0
q 2 (q − 1)2 q 2
8 ,a10 ,a5 ,a7
χa8,10,q
3
a5 /a8 + a7 /a10 = 0
(q − 1)3
8 ,a10 ,b1 ,b3 ,b4
χa8,10,q
2
b1 /a8 + b4 /a10 = 0
q 2 (q − 1)2 q 2
9 ,a10 ,a5 ,a6
χa9,10,q
3
a5 /a9 + a6 /a10 = 0
(q − 1)3
9 ,a10 ,b1 ,b2 ,b3
χa9,10,q
2
b1 /a9 + b2 /a10 = 0
q 2 (q − 1)2 q 2
F8,10
F9,10
F8
F9
F10
Conditions
q3
q3
χa8,q8 ,a3 7
(q − 1)2
q3
χa8,q8 ,b23 ,b4
q 2 (q − 1)
q2
χa9,q9 ,a3 6
(q − 1)2
q3
χa9,q9 ,b22 ,b3
q 2 (q − 1)
q2
10 ,a5
χa10,q
3
(q − 1)2
q3
10 ,b1 ,b2
χa10,q
2
q 2 (q − 1)
q2
5 ,a6 ,a7 ,b1 ,b2 ,b4
F5,6,7 (p > 3) χa5,6,7
b1 /a5 + b2 /a6 + b4 /a7 = 0 q 2 (q − 1)3 q
5 ,a6 ,a7 ,b2 ,b4
F5,6,7 (p = 3) χa5,6,7
q 2 (q − 1)3 q
F5,6
5 ,a6 ,b1 ,b2 ,b4
χa5,6
b1 /a5 + b2 /a6 = 0
q 2 (q − 1)2 q
F5,7
5 ,a7 ,b1 ,b2 ,b4
χa5,7
b1 /a5 + b4 /a7 = 0
q 2 (q − 1)2 q
F6,7
6 ,a7 ,b1 ,b2 ,b4
χa6,7
b2 /a6 + b4 /a7 = 0
q 2 (q − 1)2 q
F5
χa55 ,b2 ,b4
q 2 (q − 1)
q
F6
χa66 ,b1 ,b4
q 2 (q − 1)
q
F7
χa77 ,b1 ,b2
q 2 (q − 1)
q
Flin
χblin1 ,b2 ,b3 ,b4
q4
1
Table 4. The irreducible characters of U for p > 2.
8 ,a9 ,a10
3
3
• The characters χa8,9,10,q
3 . There are (q − 1) of these and they have degree q .
8
• The characters χ
q3
.
2
a8 ,a9 ,a10 ,a5,6,7 ,d1,2,4 ,d3
3
8,9,10, q2
. There are 4(q − 1)4 of these and their degree is
The parameters d1,2,4 and d3 can each take one of two values as explained at the
end of §5.4.
Also for p = 2 the parametrization of characters where there is a relation between parameters involving the sum of two terms has to be updated in a similar way to the corresponding
situation for conjugacy class representatives. For example, in the family F5,6 , we just have
5 ,a6 ,b2
(so we have set b1 = 0 and b2 ranges over Fq ).
characters χa5,6
5. Determining the character values
In the following subsections we give the construction of the irreducible characters and
determine the character values. We split up our determination of the character values according to the families given in Table 4, and in each of these subsections we calculate the
values of the irreducible characters on elements of U . For the families considered in §5.1, §5.2
and §5.3 we determine the character values on general elements of U , whereas for the other
characters we take advantage of the representatives of the conjugacy classes from Table 3.
Before we embark on calculating the values of the irreducible characters of U , we give
some notation that we use.
i2π Tr(x)
Denote by Tr : Fq → Fp the trace map, and define φ : Fq → C× by φ(x) = e p , so that
φ is a nontrivial character from the additive group of Fq to the multiplicative group C× .
We record two important elementary observations about φ that we require frequently in
the sequel. First we note
P that, for t ∈ Fq , we have φ(st) = 1 for all s ∈ Fq if and only t = 0,
second we note that s∈Fq φ(s) = 0.
Let H be a finite group. We write Z(H) for the centre of H. Let K be a subgroup
of H, let x ∈ H, h ∈ H and let ψ : K → C. We write ψ̇ : H → C for the function
defined by ψ̇(x) = ψ(x) if x ∈ K and ψ̇(x) = 0 if x ∈
/ K. We denote conjugation by
h
−1
h
h
x = h xh and ψ : H → C is defined by ψ(x) = ψ̇(xh ). If ψ is a class function on
K, then we write ψ HPfor the induced class function on H, which is given by the standard
1
h
formula ψ H (h) = |K|
x∈H ψ̇(x ).
We use the notation x(t) from (2.1), where we recall that x3 (t3 ) is on the left. Frequently,
we consider quotients of U by the normal subgroups Mi defined in (2.2). When we do this
we often want to identify the root subgroup Xj with its image in U/Mi for j < i, and we
will
Q do this without further explanation. We consider subgroups V of U/Mi of the form
where I is a subset of {1, . . . , i}. For such a subgroup V we use the notation
j∈I Xj , Q
xV (t) = j∈I xj (tj ) to denote a general element of V , where the factors are ordered as in
x(t), so that x3 (t3 ) is on the left (if 3 ∈ I).
We often use the notation δs,0 , where s ∈ Fm
q for some m, which is defined as usual by
δs,0 = 1 if s = 0 and δs,0 = 0 if s 6= 0.
As mentioned earlier, we have adapted the construction of the characters from [24] so
that they are more explicit. In particular, we realise each of the irreducible characters as
an induced character of a certain subgroup, which leads to a more precise parametrization.
Given these differences, we briefly explain why our constructions line up with those in [24].
For each irreducible character χ that we construct, we give a subgroup V of a quotient
U/M of U and a linear character λ of V . Then χ is equal to the induced character λU/M
9
inflated to U possibly tensored with a linear character of U . In each case there is a subgroup
X of U/M , which is a transversal of V in U/M . In each of subsequent subsections, where we
calculate the characters values, we state what V and X are. We also justify that the induced
characters that we obtain are indeed irreducible, and that the characters as given in Table 4
are distinct. The main tool for achieving this is the Clifford theory for characters of special
groups of type q 1+2m , as is also the case in [24]. Now we can count the number of characters
that we have constructed in each family, and verify that this gives the same numbers in [24].
Therefore, we can deduce that we have covered all of the irreducible characters of U .
5.1. The family Flin . Here we state the value of χblin1 ,b2 ,b3 ,b4 on all elements of U . We note
that M5 is the commutator subgroup of U , and the characters χblin1 ,b2 ,b3 ,b4 are precisely the
characters of U/M5 inflated to U ; they are all the linear characters of U . The following
formula, therefore, defines these characters.
χblin1 ,b2 ,b3 ,b4 (x(t)) = φ(b1 t1 + b2 t2 + b3 t3 + b4 t4 ).
∗
. The characters in the families F5 , F6 , F7 , F5,6 , F5,7 , F6,7 and F5,6,7
5.2. The family F5,6,7
can all be constructed in essentially the same way. Thus we combine these families and
∗
.
denote their union by F5,6,7
5 ,c6 ,c7 ,b1 ,b2 ,b4
We begin by giving the construction of the characters χc5,6,7
, where c5 , c6 , c7 ∈ Fq
are not all zero and b1 , b2 , b4 ∈ Fq . From these characters we obtain all of the characters in
c5 ,c6 ,c7 ,b1 ,b2 ,b4
∗
F5,6,7
. First we note that M8 is in the kernel of χ5,6,7
. As explained at the start of
the section we identify Xi with its image in U/M8 for i = 1, 2, . . . , 7.
We have that U/M8 = X3 X1 X2 X4 X5 X6 X7 , and that V = X1 X2 X4 X5 X6 X7 ∼
= F6q is an
6
elementary abelian subgroup of U/M8 of order q , and X = X3 is a transversal of U in U/M8 .
We define the linear character λc5 ,c6 ,c7 of V by λc5 ,c6 ,c7 (xV (t)) = φ(c5 t5 + c6 t6 + c7 t7 ).
5 ,c6 ,c7
Let χc5,6,7
be the character of U/M8 we obtain by inducing λc5 ,c6 ,c7 ; by a mild abuse
5 ,c6 ,c7
of notation we also write χc5,6,7
for the character of U given by inflation. The character
c5 ,c6 ,c7 ,b1 ,b2 ,b4
5 ,c6 ,c7
χ5,6,7
is defined by tensoring χc5,6,7
with the linear character χblin1 ,b2 ,0,b4 .
5 ,c6 ,c7 ,b1 ,b2 ,b4
The following proposition gives the character values of χc5,6,7
. It is clear from the
x c5 ,c6 ,c7
c5 ,c6 ,c7
calculations in the proof below that λ
6= λ
for all x ∈ X with x 6= 1, so that
χc5 ,c6 ,c7 ,b1 ,b2 ,b4 is irreducible by an application of Clifford theory.
Proposition 5.1.
c5 ,c6 ,c7 ,b1 ,b2 ,b4
χ5,6,7
(x(t)) = qδ(t3 ,c5 t1 +c6 t2 +c7 t4 ),0 φ(b1 t1 + b2 t2 + b4 t4 + c5 t5 + c6 t6 + c7 t7 )).
5 ,c6 ,c7
Proof. We write λ for λc5 ,c6 ,c7 and χ for χc5,6,7
. We have that X = X3 is a transversal of V
in U , and also that Z(U/M8 ) = X5 X6 X7 . From the induction formula we easily obtain that
χ(x(t)) = 0 if t3 6= 0, and also that χ(x(t)) = δt3 ,0 χ(x1 (t1 )x2 (t2 )x4 (t4 ))λ(x5 (t5 )x6 (t6 )x7 (t7 )).
Thus it suffices to calculate χ(x1 (t1 )x2 (t2 )x4 (t4 )). We have
X
x3 (s)
χ(x1 (t1 )x2 (t2 )x4 (t4 )) =
λ(x1 (t1 )x2 (t2 )x4 (t4 ))
s∈Fq
=
X
λ((x1 (t1 )x2 (t2 )x4 (t4 ))x3 (s) )
s∈Fq
10
=
X
λ([x1 (t1 ), x3 (s)][x2 (t2 ), x3 (s)][x4 (t4 ), x3 (s)])
s∈Fq
=
X
λ(x5 (st1 )x6 (st2 )x7 (st4 ))
s∈Fq
=
X
φ(s(c5 t1 + c6 t2 + c7 t4 ))
s∈Fq
= qδc5 t1 +c6 t2 +c7 t4 ,0 .
From this we can deduce the proposition.
5 ,c6 ,c7 ,b1 ,b2 ,b4
We note that the characters χc5,6,7
can be equal for different values of bi . A fairly
c ,c ,c ,b0 ,b0 ,b0
5 6 7 1 2 4
5 ,c6 ,c7 ,b1 ,b2 ,b4
= χ5,6,7
easy calculation shows that χc5,6,7
if and only if (b1 −b01 , b2 −b02 , b4 −b04 )
is a multiple of (c5 , c6 , c7 ). Thus we choose representatives as follows for p 6= 2, 3.
• If c5 6= 0 and c6 , c7 = 0, then we take b1 = 0.
• If c5 , c6 6= 0 and c7 = 0, then we take b1 , b2 such that c6 b1 + c5 b2 = 0.
• If c5 , c6 , c7 6= 0, then we take b1 , b2 , b4 such that c6 c7 b1 + c5 c7 b2 + c5 c6 b4 = 0.
We deal with the other cases symmetrically with respect to the triality. This gives the
characters as in Table 4, and we can give their character values from Proposition 5.1.
We note that for p = 2 or p = 3 we have to choose our representatives slightly differently.
For p = 3 the only difference is when c5 , c6 , c7 6= 0 and this is shown in Table 4. In the case
p = 2, and we have c5 , c6 6= 0 and c7 = 0, then we take b1 = 0, and we deal with the other
cases symmetrically with respect to the triality.
5.3. The Families F8 , F9 , F10 , F8,9 , F8,10 and F9,10 . The characters in the families F8 , F9 ,
F10 , F8,9 , F8,10 and F9,10 are constructed in a similar way, so we deal with them together. We
include the details for the characters in the family F8,9 and then remark that the calculations
for the family F8 are entirely similar. The character in the families F9 , F10 , F8,10 and F9,10
are defined from characters in F8 or F8,9 through the triality τ . Thus the values of these
characters can be immediately deduced.
From Table 4 we see that the characters in F8,9 fall into two subfamilies determined
by degree. We start with the characters of degree q 3 and explain the construction of the
characters in this family.
First we note that M10 lies in the kernel of the characters in F8,9 . We have that U/M10 =
Q9
i=1 Xi and we define the normal subgroup
Y
V =
Xi
i6=1,2,4
of U/M10 . We have that V is isomorphic to F6q and so is elementary abelian, and X =
X1 X2 X4 is a transversal of V in U/M10 . Therefore, we have linear characters λa8 ,a9 ,a6 ,a7 of
V defined by
λa8 ,a9 ,a6 ,a7 (xV (t)) = φ(a6 t6 + a7 t7 + a8 t8 + a9 t9 ),
where a8 , a9 , a6 , a7 ∈ F×
q and we assume that (a6 , a7 ) is not a multiple of (a8 , a9 ). Then we
a8 ,a9 ,a6 ,a7
define χ8,9,q3
to be the character we obtain by inducing λa8 ,a9 ,a6 ,a7 to U/M10 .
11
From the calculations of the character values in Proposition 5.2 below, we observe that
λ
6= λa8 ,a9 ,a6 ,a7 for all x ∈ X with x 6= 1.. Therefore, an application of Clifford
8 ,a9 ,a6 ,a7
is irreducible.
theory implies that each χa8,9,q
3
We proceed to consider the characters in F8,9 of degree q 2 . Again we have that M10 is in
the kernel of these characters. We consider the subgroup
Y
W =
Xi
x a8 ,a9 ,a6 ,a7
i6=1,5
of U/M10 . We note that M = X6 X7 is a normal subgroup of W , that W/M is isomorphic to
F5q and that X = X1 X5 is a transversal of W in U/M10 . Therefore, we have linear characters
µa8 ,a9 of W defined by
µa8 ,a9 (xW (t)) = φ(a8 t8 + a9 t9 ),
a8 ,a9
a8 ,a9
where a8 , a9 ∈ F×
q . Then we define χ8,9,q 2 to be the character we obtain by inducing µ
0,b2 ,b3 ,b4
8 ,a9
8 ,a9 ,b2 ,b3 ,b4
are given by tensoring χa8,9,q
to U/M10 . The characters χa8,9,q
.
2 with χlin
2
a8 ,a9
0
a8 ,a9
To see that χ8,9,q2 is irreducible we can consider the restriction µ of µ
to the subgroup
0
V 0 = X2 X6 X8 X9 of V . Then we let χ0 = µ0U , where U 0 is the subgroup X1 X2 X5 X6 X8 X9 of
U/M10 . Now V 0 is normal in U 0 and it is an easy calculation to check that x µ0 6= µ0 for all
x ∈ X with x 6= 1. Now we deduce that χ0 is irreducible by an application of Clifford theory.
a8 ,a9 ,b2 ,b3 ,b4
8 ,a9
From this it follows that χa8,9,q
is irreducible.
2 is irreducible and thus also χ8,9,q 2
We now state our proposition giving the values of the characters in F8,9 .
Proposition 5.2. (a)
8 ,a9 ,a6 ,a7
χa8,9,q
(x(t)) = q 3 δ(t1 ,t2 ,t3 ,t4 ,t5 ,a8 t6 +a9 t7 ),0 φ(a6 t6 + a7 t7 + a8 t8 + a9 t9 ).
3
(b) If t3 = 0, then
8 ,a9 ,b2 ,b3 ,b4
χa8,9,q
(x(t)) = q 2 δ(t1 ,t5 ,a8 t2 +a9 t4 ,a8 t6 +a9 t7 ),0 φ(b2 t2 + b4 t4 + a8 t8 + a9 t9 ).
2
If t3 6= 0, then
a8 ,a9 ,b2 ,b3 ,b4
χ8,9,q
(x(t)) =
2
qδ(t1 ,a8 t2 +a9 t4 ),0 φ(b2 t2 + b3 t3 + b4 t4 + a8 t8 + a9 t9 + t5 (a8 t2 + a9 t4 − (a8 t6 + a9 t7 )/t3 )).
8 ,a9 ,a6 ,a7
Proof. (a) We write λ = λa8 ,a9 ,a6 ,a7 and χ = χa8,9,q
. We consider the series of normal
2
subgroups of U/M10
Y
Y
V0 = U/M10 , V1 =
Xi , V2 =
Xi , V3 = V.
i6=1
i6=1,2
First we calculate the values of λV2 . We have Z(V2 ) = X6 X7 X8 X9 , so that
λV2 (xV2 (t)) = λV2 (x3 (t3 )x4 (t4 )x5 (t5 ))λ(x6 (t6 )x7 (t7 )x8 (t8 )x9 (t9 )).
Also λV2 (xV2 (t)) = 0 if t4 6= 0. Thus it suffices to determine λV2 (x3 (t3 )x5 (t5 )). Now for
s ∈ Fq , we have (x3 (t3 )x5 (t5 ))x4 (s) = x3 (t3 )x5 (t5 )x7 (−st3 )x9 (−st5 ) in U/M10 . Therefore,
X
λV2 (x3 (t3 )x5 (t5 )) =
λ(x3 (t3 )x5 (t5 )x7 (−st3 )x9 (−st5 ))
s∈Fq
12
X
=
φ(−a7 st3 − a9 st5 )
s∈Fq
X
=
φ(−s(a7 t3 + a9 t5 ))
s∈Fq
= qδa7 t3 +a9 t5 ,0 .
and thus
λV2 (xV2 (t)) = qδ(t4 ,a7 t3 +a9 t5 ),0 φ(a6 t6 + a7 t7 + a8 t8 + a9 t9 ).
Next we induce λV2 to V1 ; this requires a very similar calculation, which we omit. We obtain
λV1 (xV1 (t)) = q 2 δ(t2 ,t4 ,a7 t3 +a9 t5 ,a6 t3 +a8 t5 ),0 φ(a6 t6 + a7 t7 + a8 t8 + a9 t9 ).
Our assumption that (a6 , a7 ) is not a multiple of (a8 , a9 ) implies that for the above to be
nonzero we require t3 , t5 = 0. Thus we deduce that
λV1 (xV1 (t)) = q 2 δ(t2 ,t3 ,t4 ,t5 ),0 φ(a6 t6 + a7 t7 + a8 t8 + a9 t9 ).
Now we do the final induction up to U/M10 to obtain χ. First we observe that Z(V0 ) = X8 X9
and it is easy to check that χ(x(t)) = 0 if any of t1 , t2 , t3 , t4 or t5 is equal to 0. Thus it
suffices to calculate χ(x6 (t6 )x7 (t7 )), and we obtain
X
λV1 (x6 (t6 )x7 (t7 )x8 (−st6 )x9 (−st7 ))
χ(x6 (t6 )x7 (t7 )) =
s∈Fq
= q2
X
φ(a6 t7 + a7 t7 − a8 st6 − a9 st7 )
s∈Fq
= q 3 δ0,a8 t6 +a9 t7 φ(a6 t6 + a7 t7 ).
8 ,a9 ,a6 ,a7
Putting this all together we obtain the stated value of χa8,9,q
.
3
a8 ,a9
a8 ,a9
(b) We write µ = µ
and χ = χ8,9,q2 . In order to determine the value of χ we first note
that Z(U/M10 ) = X8 X9 and that we have χ(x(t)) = 0 if t1 6= 0. Thus it suffices to calculate
χ(x3 (t3 )x2 (t2 )x4 (t4 )x5 (t5 )x6 (t6 )x7 (t7 )). Next we calculate that
(x3 (t3 )x2 (t2 )x4 (t4 )x5 (t5 )x6 (t6 )x7 (t7 ))x5 (s5 )x1 (s1 ) =
x3 (t3 )x2 (t2 )x4 (t4 )x5 (t5 − s1 t3 )x6 (t6 )x7 (t7 )x8 (−s1 t6 + s5 t2 + s1 t2 t3 )x9 (−s1 t7 + s5 t4 + s1 t3 t4 ).
Thus, χ(x3 (t3 )x2 (t2 )x4 (t4 )x5 (t5 )x6 (t6 )x7 (t7 )) is equal to
X
µ(x3 (t3 )x2 (t2 )x4 (t4 )x5 (t5 )x6 (t6 )x7 (t7 ))x5 (s5 )x1 (s1 ) )
s1 ,s5 ∈Fq
=
X
δt5 −s1 t3 ,0 φ(a8 (−s1 t6 + s5 t2 + s1 t2 t3 ) + a9 (−s1 t7 + s5 t4 + s1 t3 t4 )).
s1 ,s5 ∈Fq
If t3 = 0, then we see that this is equal to
X
δt5 ,0
φ(a8 (−s1 t6 + s5 t2 ) + a9 (−s1 t7 + s5 t4 )) = q 2 δ(t5 ,a8 t2 +a9 t4 ,a8 t6 +a9 t7 ),0 .
s1 ,s5 ∈Fq
Whereas for t3 6= 0, in the sum over s1 we only get a nonzero contribution for s1 = t5 /t3 ,
thus we get
13
X
φ(a8 (−t5 t6 /t3 + s5 t2 + t2 t5 ) + a9 (−t5 t7 /t3 + s5 t4 + t4 t5 )) =
s5 ∈Fq
qδa8 t2 +a9 t4 ,0 φ(t5 (a8 t2 + a9 t4 − (a8 t6 + a9 t7 )/t3 )).
8 ,a9 ,b2 ,b3 ,b4
From the formulas above we deduce the stated values of χa8,9,q
.
2
8 ,a9 ,a6 ,a7
can be equal for different values of ai . From the character
We note the characters χa8,9,q
3
a0 ,a0 ,a0 ,a0
8 ,a9 ,a6 ,a7
8 9 6 7
= χ8,9,q
if and only if (a08 , a09 , a06 , a07 ) is obtained from
values we observe that χa8,9,q
3
3
(a8 , a9 , a6 , a7 ) by adding a multiple of (a8 , a9 ) to (a6 , a7 ). So we choose (a6 , a7 ) such that
a6 a9 + a7 a8 = 0, except when p = 2 where we just put a6 = 0.
a08 ,a09 ,b02 ,b03 ,b04
8 ,a9 ,b2 ,b3 ,b4
= χ8,9,q
if and only if (a08 , a09 , b02 , b03 , b04 ) is obtained
Also we observe that χa8,9,q
3
3
from (a8 , a9 , b2 , b3 , b4 ) by adding a multiple of (a8 , a9 ) to (b2 , b4 ). So we choose a coset
representative such that b2 a9 + b4 a8 = 0 (except in the case p = 2 where we just take b2 = 0).
We move on to briefly discuss the characters in the family F8 . These can be defined in
essentially the same way as those in F8,9 . The characters of degree q 3 are denoted χa8,q8 ,a3 7
8 ,a9 ,a6 ,a7
and are defined in exactly the same way as χa8,9,q
except we set a9 = 0 and a6 = 0.
3
a8 ,b3 ,b4
8 ,a9 ,b2 ,b3 ,b4
The characters χ8,q2
are defined in the same way as χa8,9,q
except we set a9 = 0
2
and b2 = 0. Thus we can immediately deduce the values of these characters given in the
proposition below.
Proposition 5.3. (a)
χa8,q8 ,a3 7 (x(t)) = q 3 δ(t1 ,t2 ,t3 ,t4 ,t5 ,t6 ),0 φ(a7 t7 + a8 t8 ).
(b) If t3 = 0, then
a8 ,b3 ,b4
χ8,q
(x(t)) = q 2 δ(t1 ,t2 ,t5 ,t6 ),0 φ(b4 t4 + a8 t8 ).
2
If t3 6= 0, then
χa8,q8 ,b23 ,b4 (x(t)) = qδ(t1 ,t2 ),0 φ(b3 t3 + b4 t4 + a8 (t8 − t5 t6 /t3 )).
5.4. The Family F8,9,10 . We begin with the assumption that p > 2, as the situation for
p = 2 is slightly different and is discussed at the end of the section.
First we note that M11 lies in the kernel of the characters in F8,9,10 . We have that U/M11 =
Q10
i=1 Xi and we define the normal subgroup
Y
V =
Xi
i6=1,2,4
of U/M11 . We have that V is isomorphic to F7q and that X = X1 X2 X4 is a transversal of V
in U/M11 . We define the character λa8 ,a9 ,a10 of V by
λa8 ,a9 ,a10 (xV (t)) = φ(a8 t8 + a9 t9 + a10 t10 )
a8 ,a9 ,a10
for a8 , a9 , a10 ∈ F×
to be the character obtained by inducing
q . Then we define χ8,9,10
8 ,a9 ,a10 ,b3
a8 ,a9 ,a10
λ
to U/M11 . The irreducible characters χa8,9,10
in the family F8,9,10 are then
a8 ,a9 ,a10
0,0,b3 ,0
obtained by tensoring χ8,9,10 with the linear character χlin
. The construction of these
characters is the same as in [24, §4], where they are shown to be irreducible and distinct.
In order to obtain the values of χ we take advantage of the conjugacy class representatives
in Table 3. First note that since χ(x(t)) = 0 for x(t) 6∈ V , so we do not need to consider
14
any conjugacy classes in Table 3 above the family C3 . We split up our calculations for the
remaining elements by considering elements in the family C3 , and those lying in any of the
other classes.
Proposition 5.4. Suppose p > 2.
8 ,a9 ,a10 ,b3
(x(t)) = 0.
(a) Suppose that at least one of t1 , t2 or t4 is nonzero. Then χa8,9,10
(b) Let x(t) = x3 (t3 )x8 (t8 )x9 (t9 )x10 (t10 ), where t3 6= 0. Then
X
8 ,a9 ,a10 ,b3
(x(t)) = qφ(b3 t3 + a8 t8 + a9 t9 + a10 t10 )
φ(−a8 a9 a10 t3 s2 ).
χa8,9,10
s∈Fq
(c) Let x(t) = x5 (t5 )x6 (t6 )x7 (t7 )x8 (t8 )x9 (t9 )x10 (t10 ). Then
8 ,a9 ,a10 ,b3
(x(t)) = q 3 δ(t5 ,t6 ,t7 ),0 φ(a8 t8 + a9 t9 + a10 t10 ).
χa8,9,10
P
Remark 5.5. Before we proceed to the proof, we remark on the term t∈Fq φ(−a8 a9 a10 t3 t2 )
P
in (b) (for the case t3 6= 0). Clearly this can be deduced from the value of t∈Fq φ(t2 ): if
P
P
−a8 a9 a10 t3 ∈ F×
φ(−a8 a9 a10 t3 t2 ) = t∈Fq φ(t2 ) and if −a8 a9 a10 t3 ∈
q is a square, then
t∈F
q
P
P
P
2
2
2
F×
q is not a square, then
t∈Fq φ(−a8 a9 a10 t3 t ) = −
t∈Fq φ(t ). The sum
t∈Fq φ(t ) is
referred to as a quadratic
For q = p the value of this GaussP
sum is well known:
P Gauss sum. √
√
if p ≡ 1 mod 4, then t∈Fp φ(t2 ) = p, and if p ≡ −1 mod 4, then t∈Fp φ(t2 ) = i p.
However, for general q the situation is more complicated: it is known that the absolute value
√
is q, see for example [29, Proposition 11.5], but a formula giving the exact value appears
not to be known in general.
Proof. We begin by noting that it suffices to consider the case b3 = 0. We write λ = λa8 ,a9 ,a10
8 ,a9 ,a10
and χ = χa8,9,10
.
(a) We dealt with this case in the discussion before the statement of the proposition.
(b) The first step is to calculate in U/M11 that
x3 (t3 )x1 (s1 )x2 (s2 )x4 (s4 ) = x3 (t3 )x5 (−s1 t3 )x6 (−s2 t3 )x7 (−s4 t3 )x8 (s1 s2 t3 )x9 (s1 s4 t3 )x10 (s2 s4 t3 ).
Therefore, we have
X
χ(x3 (t3 )) =
φ(a8 s1 s2 t3 + a9 s1 s4 t3 + a10 s2 s4 t3 )
s1 ,s2 ,s4 ∈Fq
X
=
φ(a10 s2 s4 t3 )
s2 ,s4 ∈Fq
X
=
X
φ(s1 (a8 s2 + a9 s4 )t3 )
s1 ∈Fq
φ(a10 t3 s2 s4 )qδ0,a8 s2 +a9 s4
s2 ,s4 ∈Fq
=q
X
φ(−a10 a9 s24 t3 /a8 )
s4 ∈Fq
=q
X
φ(−a8 a9 a10 t3 s2 ).
s∈Fq
For the last equality we just substitute s4 = a8 s. Now using that Z(U/M11 ) = X8 X9 X10 ,
the value of χ given in (b) can be deduced.
(c) In this case the starting point is to calculate in U/M11 that
15
(x5 (t5 )x6 (t6 )x7 (t7 ))x1 (s1 )x2 (s2 )x4 (s4 ) =
x5 (t5 )x6 (t6 )x7 (t7 )x8 (−s1 t6 − s2 t5 )x9 (−s1 t7 − s4 t5 )x10 (−s2 t7 − s4 t6 ).
Therefore, χ(x5 (t5 )x6 (t6 )x7 (t7 )) is equal to
X
X
X
φ(−s1 (a8 t6 + a9 t7 ))
φ(−s2 (a8 t5 + a10 t7 ))
φ(−s4 (a9 t5 + a10 t6 ))
s1 ∈Fq
s4 ∈Fq
s2 ∈Fq
3
= q δ(a8 t6 +a9 t7 ,a8 t5 +a10 t7 ,a9 t5 +a10 t6 ),0 .
Now the linear system
a8 t6 + a9 t7 = 0
a8 t5
+ a10 t7 = 0
a9 t5 + a10 t6
= 0
has only the trivial solution t5 = t6 = t7 = 0, as we are assuming p > 2. Thus we get
χ(x5 (t5 )x6 (t6 )x7 (t7 )) = q 3 δ(t5 ,t6 ,t7 ),0 . Now using that Z(U/M11 ) = X8 X9 X10 , the value of
a8 ,a9 ,a10 ,b3
χ8,9,10
given in (c) can be deduced.
To end this section we consider the case p = 2, where the situation is more complicated.
a8 ,a9 ,a10
8 ,a9 ,a10
The characters χa8,9,10,q
for p > 3.
3 are defined in the same way as the characters χ8,9,10
a ,a ,a ,a
,d
,d
8 9 10 5,6,7 1,2,4 3
There are also characters denoted χ8,9,10,q
, which are defined as follows.
3 /2
Q
Again we let V = i6=1,2,4 Xi and let a = a5,6,7 , a8 , a9 , a10 ∈ F×
q . We define the linear
a8 ,a9 ,a10 ,a
character λ = λ
of V by
λ(xV (t)) = φ(a8 a9 at5 + a8 a10 at6 + a9 a10 at7 + a8 t8 + a9 t9 + a10 t10 ).
Now let Y124 = {1, x1 (a10 a)x2 (a9 a)x4 (a8 a)} and V̄ = Y124 V . We can extend λ to V̄ in two
ways, namely λ̄d1,2,4 , where λ̄d1,2,4 (x1 (a10 a)x2 (a9 a)x4 (a8 a)) = (−1)d1,2,4 for d1,2,4 ∈ F2 . We
then induce λ̄d1,2,4 to U/M11 to obtain a character ψ d1,2,4 , which is irreducible as shown in
[24, §4]. Now let d3 be either 0 or an element of Fq , which is not in the image of the map
a8 ,a9 ,a10 ,a5,6,7 ,d1,2,4 ,d3
s 7→ aa8 a9 a10 s + a8 a9 a10 s2 . Then we define χ8,9,10,q
by tensoring ψ d1,2,4 with
3 /2
0,0,d3 ,0
χlin
.
The proposition below states the values of these characters, these can be deduced from
[30, Thm. 2.3]. Note that with our specific choice of φ, we have ker(φ) = {tp − t : t ∈ Fq }.
−p
p
p−1
Thus for each a ∈ F×
t | t ∈ Fq } as
q , we have that ker(φ) = a Ta , where Ta = {t − a
defined in [30, Definition 1.2]. In turn this implies that aφ as defined in [30, Definition 1.4]
8 ,a9 ,a10
is equal to a−p . We note that the values for χa8,9,10,q
3 look different to those for p > 2, which
is explained by the fact that the quadratic Gauss sums for p = 2 are clearly zero.
Proposition 5.6. Let p = 2.
(a)
3
8 ,a9 ,a10
χa8,9,10,q
3 (x(t)) = δ(t1 ,t2 ,t3 ,t4 ,a8 t5 +a10 t7 ,a8 t6 +a9 t7 ),0 q φ(a8 t8 + a9 t9 + a10 t10 ).
(b) Let a = a5,6,7 . If (t1 , t2 , t4 ) 6∈ {(0, 0, 0), (a10 a, a9 a, a8 a)}, then
a ,a ,a ,a
8 9 10 5,6,7
χ8,9,10,q
3 /2
,d1,2,4 ,d3
16
(x(t)) = 0.
If x(t) = x1 (a10 r)x2 (a9 r)x4 (a8 r)x5 (a10 t)x6 (a9 t)x7 (a8 t)x8 (t8 )x9 (t9 )x10 (t10 ), where r ∈ {0, a}
and t, t8 , t9 , t10 ∈ Fq , then
q3
φ(d1,2,4 r + a8 a9 a10 at + a8 t8 + a9 t9 + a10 t10 ).
2
For r ∈ {0, a} and x(t) = x3 (t3 )x1 (a10 r)x2 (a9 r)x4 (a8 r)x5 (t5 )x6 (t6 )x7 (t7 )x8 (t8 )x9 (t9 )x10 (t10 ),
a ,a ,a ,a
8 9 10 5,6,7
χ8,9,10,q
3 /2
a ,a ,a ,a
8 9 10 5,6,7
χ8,9,10,q
3 /2
,d1,2,4 ,d3
,d1,2,4 ,d3
(x(t)) =
(x(t)) =
q2
t7
δt3 , a−p φ(d1,2,4 r + d3 t3 + a8 a9 a10 a +
a8 a9 a10
2
a8
2
(a8 a9 a10 ) t5
t7 t6
t7
+ )( + ) + a8 t8 + a9 t9 + a10 t10 ).
(
−p
a
a10 a8 a9 a8
5.5. The Family F11 . Q
First we note that M12 lies in the kernel of the characters in F11 .
We have that U/M12 = 11
i=1 Xi and we define the normal subgroup
Y
V =
Xi
i6=1,2,4
of U/M12 . We have that V is isomorphic to F8q and X = X1 X2 X4 is a transversal of V in
U/M12 . We define the character λa11 ,b5 ,b6 ,b7 of V by
λa11 ,b5 ,b6 ,b7 (xV (t)) = φ(b5 t5 + b6 t6 + b7 t7 + a11 t11 )
for a11 ∈ F×
q and b5 , b6 , b7 ∈ Fq .
a11 ,b5 ,b6 ,b7
We define χ11
to be the character obtained by inducing λa11 ,b5 ,b6 ,b7 to U/M12 . The
a11 ,b5 ,b6 ,b7 ,b3
a11 ,b5 ,b6 ,b7
irreducible characters χ11
in the family F11 are then obtained by tensoring χ11
3 ,0
with the linear character χ0,0,b
.
lin
Using [30, Lemma 1.5] with Z = X11 , Y = X8 X9 X10 , X = X1 X2 X4 and M = X3 X5 X6 X7 ,
a0 ,b0 ,b0 ,b0 ,b0
a11 ,b5 ,b6 ,b7 ,b3
we obtain that each χ11
is irreducible, and that χa1111 ,b5 ,b6 ,b7 ,b3 = χ1111 5 6 7 3 if and
only if (a11 , b5 , b6 , b7 , b3 ) = (a011 , b05 , b06 , b07 , b03 ). The fact that the hypothesis of that lemma
holds follows from the fact that XY Z forms a subgroup of V which is special of type q 1+6 .
In order to obtain the values of χa1111 ,b5 ,b6 ,b7 ,b3 in Proposition 5.7 below, we take advantage of
the conjugacy class representatives in Table 3. We assume p > 2 at first, as the situation is
slightly different for p = 2, which we discuss after. First note that since χa1111 ,b5 ,b6 ,b7 ,b3 (x(t)) =
0 for x(t) 6∈ V , we can deal with any conjugacy classes in Table 3 above the family C3
quickly in (a). Dealing with elements in the family C3 in (b) is the most complicated part of
the proposition. Part (c) covers all of those in the families C5,6,7 , C5,6 and C5 , and then we
can then use the triality automorphism τ to determine the values on representatives in the
classes C5,7 , C6,7 , C6 and C7 . We finish by considering the remaining classes together in (d),
a11 ,b5 ,b6 ,b7 ,b3
.
so that we have determined the value of χ11
Proposition 5.7. Suppose p > 2.
(a) Suppose that at least one of t1 , t2 or t4 is nonzero. Then
χa1111 ,b5 ,b6 ,b7 ,b3 (x(t)) = 0.
(b) Let x(t) = x3 (t3 )x8 (t8 )x9 (t9 )x10 (t10 )x11 (t11 ), where t3 6= 0. Then
a11 ,b5 ,b6 ,b7 ,b3
χ11
(x(t)) = qφ(b3 t3 + a11 t11 )(qδ(b6 t3 +a11 t9 ,b7 t3 +a11 t8 ),0 )
17
X
φ(−(b5 t3 + a11 t10 )s + (b6 t3 + a11 t9 )(b7 t3 + a11 t8 )(a11 t3 )−1 s−1 )).
s∈F×
q
(c)
(i) Let x(t) = x5 (t5 )x6 (t6 )x7 (t7 )x11 (t11 ), where t5 , t6 , t7 6= 0. Then
X
χa1111 ,b5 ,b6 ,b7 ,b3 (x(t)) = qφ(b5 t5 + b6 t6 + b7 t7 + a11 t11 )
φ(−a11 t5 t6 t7 s2 ).
s∈Fq
(ii) Let x(t) = x5 (t5 )x6 (t6 )x9 (t9 )x10 (t10 )x11 (t11 ), where t5 6= 0. Then
χa1111 ,b5 ,b6 ,b7 ,b3 (x(t)) = q 2 δt6 t9 ,t5 t10 φ(b5 t5 + b6 t6 + a11 t11 ).
(d) Let x(t) = x8 (t8 )x9 (t9 )x10 (t10 )x11 (t11 ). Then
χa1111 ,b5 ,b6 ,b7 ,b3 = q 3 δ(t8 ,t9 ,t10 ),0 φ(a11 t11 ).
P
Remark 5.8. Before the proof, we comment on the term of the form s∈Fq φ(As + Bs ), where
A, B ∈ F×
q in (b). It seems that a general formula for this is not known.
Proof. We note that it suffices to deal with the case b3 = 0 and we let λ = λa11 ,b5 ,b6 ,b7 and
χ = χa11 ,b5 ,b6 ,b7 .
(a) We dealt with this case in the discussion before the statement of the proposition.
(b) First we note that Z(U/M12 ) = X11 . Therefore, it suffices to calculate the values of
χ(x3 (t3 )x8 (t8 )x9 (t9 )x10 (t10 )). We calculate
(x3 (t3 )x8 (t8 )x9 (t9 )x10 (t10 ))x1 (s1 )x2 (s2 )x4 (s4 ) =
x3 (t3 )x5 (−t3 s1 )x6 (−t3 s2 )x7 (−t3 s4 )x8 (t8 + t3 s1 s2 )
x9 (t9 + t3 s1 s4 )x10 (t10 + t3 s2 s4 )x11 (−t9 s2 − t10 s1 − t8 s4 − t3 s1 s2 s4 )).
Thus we see that χ(x3 (t3 )x8 (t8 )x9 (t9 )x10 (t10 )) is equal to
X
φ(−b5 t3 s1 − b6 t3 s2 − b7 t3 s4 + a11 (−t9 s2 − t10 s1 − t8 s4 − t3 s1 s2 s4 ))
s1 ,s2 ,s4 ∈Fq
X
=
φ(−b5 t3 s1 − b6 t3 s2 + a11 (−t9 s2 − t10 s1 ))
s1 ,s2 ∈Fq
X
=
X
φ(−(b7 t3 + a11 t8 + a11 t3 s1 s2 )s4 )
s4 ∈Fq
φ(−b5 t3 s1 − b6 t3 s2 + a11 (−t9 s2 − t10 s1 ))qδb7 t3 +a11 t8 +a11 t3 s1 s2 ,0
s1 ,s2 ∈Fq
=q
X
φ((−b6 t3 − a11 t9 )s2 )δb7 t3 +a11 t8 ,0
s2 ∈Fq
+
X
φ(−b5 t3 s1 − a11 t10 s1 )
s1 ∈F×
q
=q
X
φ((−b6 t3 − a11 t9 )s2 )δb7 t3 +a11 t8 +a11 t3 s1 s2 ,0
s2 ∈Fq
qδ(b7 t3 +a11 t8 ,b6 t3 +a11 t9 ),0
18
+
X
φ(−(b5 t3 + a11 t10 )s1 )φ((b6 t3 + a11 t9 )(b7 t3 + a11 t8 )(a11 t3 )−1 s−1
1 )
s1 ∈F×
q
= q(qδ(b7 t3 +a11 t8 ,b6 t3 +a11 t9 ),0
X
+
φ(−(b5 t3 + a11 t10 )s + (b6 t3 + a11 t9 )(b7 t3 + a11 t8 )(a11 t3 )−1 s−1 )).
s∈F×
q
This gives the formula in (b).
(c) We moveQon to consider the case where t3 = 0 and t5 6= 0. Consider the normal
subgroup W = i6=1 Xi of U/M12 . A calculation shows that
λW (x5 (t5 )x6 (t6 )x7 (t7 )x8 (t8 )x9 (t9 )x10 (t10 )x11 (t11 )) =
qφ(b5 t5 + b6 t6 + b7 t7 + a11 (−t−1
5 t8 t9 + t11 )).
To obtain χ(x5 (t5 )x6 (t6 )x7 (t7 )x8 (t8 )x9 (t9 )x10 (t10 )x11 (t11 )) we just have to induce this formula
over X1 to U/M12 . First we consider elements of the conjugacy class C5,6,7 , and we get
χ(x5 (t5 )x6 (t6 )x7 (t7 )x11 (t11 ))
X
= qφ(b5 t5 + b6 t6 + b7 t7 + a11 t11 )
φ(−a11 t−1
5 (−s1 t6 )(−s1 t7 ))
s1 ∈Fq
= qφ(b5 t5 + b6 t6 + b7 t7 + a11 t11 )
X
φ(−a11 t5 t6 t7 s2 )
s∈Fq
as stated in (i).
Next we consider elements in the class C5,6 or C5 and we get
χ(x5 (t5 )x6 (t6 )x9 (t9 )x10 (t10 )x11 (t11 ))
X
= qφ(b5 t5 + b6 t6 + a11 t11 )
φ(−a11 t−1
5 ((−s1 t6 )t9 − s1 t10 ))
s1 ∈Fq
= qφ(b5 t5 + b6 t6 + b7 t7 + a11 t11 )
X
φ(a11 t−1
5 s1 (t6 t9 + t10 ))
s∈Fq
2
= q δt6 t9 +t10 ,0 φ(b5 t5 + b6 t6 + b7 t7 + a11 t11 )
as stated in (ii).
(d) This is an easy calculation, which we omit.
In the case p = 2, the only difference is that we now have to consider elements of the form
x5 (t5 )x6 (t6 )x7 (t7 )x10 (t10 )x11 (t11 ) even when t5 , t6 , t7 6= 0, to cover the family of conjugacy
p=2
classes C5,6,7,2q
8 ; the value given below is in a form where we do not have a closed formula.
p=2
Also the value for the family C5,6,7,q
8 is simplified as the quadratic Gauss sum is equal to 0.
Lemma 5.9. Let p = 2.
(a) Let x(t) = x5 (t5 )x6 (t6 )x7 (t7 )x10 (t10 )x11 (t11 ), where t5 , t6 , t7 , t10 ∈ F×
q and t11 ∈ Fq . Then
X
χa1111 ,b5 ,b6 ,b7 ,b3 (x(t)) = qφ(b5 t5 + b6 t6 + b7 t7 + a11 t11 )
φ(a11 (t5 t6 t7 s2 + t5 t10 s)).
s∈Fq
19
(b) Let x(t) = x5 (t5 )x6 (t6 )x7 (t7 ), where t5 , t6 , t7 ∈ F×
q . Then
χa1111 ,b5 ,b6 ,b7 ,b3 (x(t)) = 0.
Proof. (a) We can proceed as in the case p > 2 and we obtain the formula
λW (x5 (t5 )x6 (t6 )x7 (t7 )x8 (t8 )x9 (t9 )x10 (t10 )x11 (t11 )) =
qφ(b5 t5 + b6 t6 + b7 t7 + a11 (t−1
5 t8 t9 + t11 )).
Now we induce over X1 to obtain
χ(x5 (t5 )x6 (t6 )x7 (t7 )x10 (t10 )x11 (t11 )) =
X
λW (x5 (t5 )x6 (t6 )x7 (t7 )x8 (s1 t6 )x9 (s1 t7 )x10 (t10 )x11 (s1 t10 + t11 ))
s1 ∈Fq
=
X
2
qφ(b5 t5 + b6 t6 + b7 t7 + a11 (t−1
5 s1 t6 t7 + s1 t10 + t11 )
s1 ∈Fq
= qφ(b5 t5 + b6 t6 + b7 t7 + a11 t11 )
X
φ(a11 (t5 t6 t7 s2 + t5 t10 s)).
s∈Fq
(b) This is obtained directly from the calculation for p > 2 by noting that the sum there
is equal to 0.
Q
5.6. The Family F12 . Let V = i6=3,5,6,7 Xi and W = X8 X9 X10 X11 . Then W is a normal
subgroup of V and the quotient V /W is elementary abelian isomorphic to F4q . Also X =
X3 X5 X6 X7 is a transversal of V in U . Given a12 ∈ F×
q , we define the linear character
a12
a12
λ : V → C by λ (xV (t)) = φ(a12 t12 ). Then we define χa1212 = (λa12 )U and χa1212 ,b1 ,b2 ,b4 =
χa1212 χblin1 ,b2 ,0,b4 .
Using [30, Lemma 1.5] with Z = X12 , Y = X8 X9 X10 X11 , X = X3 X5 X6 X7 and M =
a0 ,b0 ,b0 ,b0
a12 ,b1 ,b2 ,b4
X1 X2 X4 , we obtain that each χa1112 ,b1 ,b2 ,b4 is irreducible, and that χ11
= χ1112 1 2 4 if
and only if (a12 , b1 , b2 , b4 ) = (a012 , b01 , b02 , b04 ). The hypothesis of that lemma holds due to the
fact that XY Z forms a subgroup of V which is special of type q 1+8 .
In the next proposition we state the values of χa1212 ,b1 ,b2 ,b4 by considering a number of cases,
which cover all of the conjugacy class representatives in Table 3. We note that in (c) we
only deal with conjugacy classes up to the action of τ . We first restrict to the case p > 2,
and then afterwards we describe some differences that occur for the case p = 2.
Proposition 5.10. Suppose that p > 2.
(a) Suppose that t3 6= 0. Then
χa1212 ,b1 ,b2 ,b4 (x(t)) = 0.
(b) Suppose x(t) lies in one of the families C1,2,4,q6 , C1,2,q6 , C1,4,q6 , C2,4,q6 , C1,q6 , C2,q6 or C4,q6 .
Then
χa1212 ,b1 ,b2 ,b4 (x(t)) = 0.
(c) The values of χa1212 ,b1 ,b2 ,b4 on the families of conjugacy classes C1,2,4,q7 , C1,2,q7 , C1,2,q8 , C1,q7
and C1,q8 are given as follows.
20
C1,2,4,q7 : χa1212 ,b1 ,b2 ,b4 (x1 (t1 )x2 (t2 )x4 (t4 )x12 (t
12 ))
X
= qφ(b1 t1 + b2 t2 + b4 t4 + a12 t12 )
φ(−a12 t1 t2 t4 s2 ), for t1 , t2 , t4 ∈ F×
q and t12 ∈ Fq .
C1,2,q7 :
C1,2,q8 :
s∈Fq
a12 ,b1 ,b2 ,b4
(x1 (t1 )x2 (t2 )x9 (t9 )x10 (t10 )) = q 2 δ0,t10 t1 −t2 t9 φ(b1 t1 + b2 t2 ),
χ12
for t1 , t2 , t9 , t10 ∈ F×
q .
a12 ,b1 ,b2 ,b4
(x1 (t1 )x2 (t2 )x12 (t12 )) = q 2 φ(b1 t1 + b2 t2 + a12 t12 ),
χ12
for t1 , t2 ∈ F×
q and t12 ∈ Fq .
a12 ,b1 ,b2 ,b4
(x1 (t1 )x10 (t10 )) = 0, for t1 , t10 ∈ F×
χ12
q .
a12 ,b1 ,b2 ,b4
2
(x1 (t1 )x12 (t12 )) = q φ(b1 t1 + a12 t12 ), for t1 ∈ F×
χ12
q and t12
C1,q7 :
C1,q8 :
Q
(d) Let x(t) = i≥5 xi (ti ). Then
∈ Fq .
a12 ,b1 ,b2 ,b4
(x(t)) = q 4 δ0,(t5 ,t6 ,t7 ,t8 ,t9 ,t10 ,t11 ) φ(a12 t12 ).
χ12
Proof. It is clear that it suffices to calculate the values of χa1212 . We write λ = λa12 and
χ = χa1212 .
In order to explain our calculations, we introduce some intermediate subgroups. We let
V1 = V X7 , V2 = V X6 X7 and V3 = V X5 X6 X7 . To abbreviate notation we write λi = λVi for
each i = 1, 2, 3. By direct calculation we find
λ1 (xV1 (t)) = qδ(t7 ,t8 ),0 φ(a12 t12 ),
and
(5.1)
λ2 (xV2 (t)) =
qδ0,(t6 ,t7 ) φ(a12 (t12 − t9t1t8 )) if t1 =
6 0
2
q δ0,(t6 ,t7 ,t8 ,t9 ) φ(a12 t12 )
if t1 = 0.
We do not calculate λ3 fully here, but do note that
(5.2)
λ3 (xV3 (t)) = δ0,(t5 ,t6 ,t7 ) λ3 (xV3 (t)).
(a) It is clear that χ(x(t)) = 0, if t3 6= 0.
(b) Let x = x1 (c1 )x2 (c2 )x4 (c4 )x5 (e5 )x6 (e6 )x7 (e7 ) lie in the family C1,2,4,q6 and consider
y = xx3 (s3 ) for s3 ∈ Fq . The theory of minimal representatives outlined in Section 3 implies that the coefficient of one of x5 , x6 or x7 in y must be nonzero. Therefore, χ(x) =
P
x3 (s)
) = 0 by (5.2).
s∈Fq ψ̇3 (x
Similar arguments deal with the other families in the statement.
(c) First we show that if one of t1 , t2 or t4 is nonzero, then
!
!
Y
Y
(5.3)
χ
xi (ti ) = λ3
xi (ti ) .
i6=3,5,6,7
i6=3,5,6,7
We just
Q deal with the case t1 6= 0, as the other cases can be dealt with similarly. Consider
y = i6=3,5,6,7 xi (ti )x3 (s3 ) , for s3 ∈ Fq . We calculate the coefficient of x5 in y to be t1 s3 .
Therefore, by (5.2), the only nonzero term in the right side of
!
!
Y
X
Y
λ̇3
xi (ti )x3 (s3 )
χ
xi (ti ) =
i6=3,5,6,7
s3 ∈Fq
is when s3 = 0.
21
i6=3,5,6,7
We now give the calculation for C1,2,4,q7
χ(x1 (t1 )x2 (t2 )x4 (t4 )x12 (t12 ))
= λ3 (x1 (t1 )x2 (t2 )x4 (t4 )x12 (t12 ))
X
=
λ̇2 ((x1 (t1 )x2 (t2 )x4 (t4 )x12 (t12 ))x5 (s5 ) )
s5 ∈Fq
=
X
λ̇2 (x1 (t1 )x2 (t2 )x8 (t2 s5 )x4 (t4 )x9 (t4 s5 )x12 (t12 ))
s5 ∈Fq
=
X
λ̇2 (x1 (t1 )x2 (t2 )x4 (t4 )x8 (t2 s5 )x9 (t4 s5 )x11 (−t2 t4 s5 )x12 (t12 ))
s5 ∈Fq
= qφ(a12 t12 )
X
φ(−a12 t1 t2 t4 s2 ).
s∈Fq
The first equality is given by (5.3), and the last uses (5.1) and substitution s = t1 s5 .
We omit the calculations for the other conjugacy classes in (c), as they are very similar.
(d) This is a straightforward calculation, which we omit.
p=2
p=2
For the case p = 2, we just have to deal with the classes C1,2,4,2q
7 and C1,2,4,q 7 , which we do
in the following lemma. The calculations involved are a minor modification to those for the
class C1,2,4 for p > 2, so we omit the proof.
p=2
p=2
Lemma 5.11. Let p = 2. The values of on the classes C1,2,4,2q
7 and C1,2,4,q 7 are given as
follows.
p=2
C1,2,4,2q
χa1212 ,b1 ,b2 ,b4 (x1 (t1 )x2 (t2 )x4 (t4 )x10 (t10 )x12 (t12 ))
7:
P
= qφ(b1 t1 + b2 t2 + b4 t4 + a12 t12 ) s∈Fq φ(a12 (st1 t10 + t1 t2 t4 s2 )), for t1 , t2 , t4 , t10 ∈ F×
q
and t12 ∈ Fq .
p=2
C1,2,4,q
χa1212 ,b1 ,b2 ,b4 (x1 (t1 )x2 (t2 )x4 (t4 )x12 (t12 )) = 0, for t1 , t2 , t4 ∈ F×
7:
q and t12 ∈ Fq .
References
[1] J. An and G. Hiss, Restricting unipotent characters in finite symplectic groups, Comm. Algebra 39
(2011), no. 3, 1104–1130.
[2] C. A. M. André, The basic character table of the unitriangular group J. Algebra 241 (2001), no. 1,
437–471.
[3] C. André and A. Neto, A supercharacter theory for the Sylow p-subgroups of the finite symplectic and
orthogonal groups, J. Algebra 322 (2009), 1273–1294.
[4] J. D. Bradley and S. M. Goodwin, Conjugacy classes in Sylow p-subgroups of finite Chevalley groups in
bad characteristic, Comm. Algebra 42 (2014), no. 8, 3245–3258.
[5] R. W. Carter, Simple groups of Lie type, A Wiley-Interscience publication, London, 1972.
[6] P. Diaconis, I.M. Isaacs, Supercharacters and superclasses for algebra groups, Trans. Amer. Math. Soc.
360 (2008), no. 5, 2359–2392.
[7] F. Digne and J. Michel, Representations of finite groups of Lie type, London Mathematical Society
Student Texts 21, Cambridge University Press, Cambridge, 1991.
[8] A. Evseev, Reduction for characters of finite algebra groups, J. Algebra 325 (2011), 321–351.
[9] The GAP Group, GAP – Groups, Algorithms, and Programming, Version 4.4.12 (2008),
http://www.gap-system.org.
22
[10] M. Geck, G. Hiss, F. Lübeck, G. Malle and G. Pfeiffer, CHEVIE A system for computing and processing
generic character tables for finite groups of Lie type, Weyl groups and Hecke algebras, Appl. Algebra
Engrg. Comm. Comput. 7 (1996), 175–210.
[11] S. M. Goodwin, On the conjugacy classes in maximal unipotent subgroups of simple algebraic groups,
Transform. Groups 11 (2006), no. 1, 51–76.
[12] S. M. Goodwin, P. Mosch and G. Röhrle, Calculating conjugacy classes in Sylow p-subgroups of finite
Chevalley groups of rank six and seven, LMS J. Comput. Math. 17 (2014), no. 1, 109–122.
, On the coadjoint orbits of maximal unipotent subgroups of reductive groups, Transform. Groups
[13]
21 (2016), no. 2, 399–426.
[14] S. M. Goodwin and G. Röhrle, Calculating conjugacy classes in Sylow p-subgroups of finite Chevalley
groups, J. Algebra 321 (2009), 3321–3334.
[15] R. Gow, M. Marjoram and A. Previtali On the irreducible characters of a Sylow 2-subgroup of the finite
symplectic group in characteristic 2, J. Algebra 241 (2001), 393–409
[16] G. Higman, Enumerating p-groups. I: Inequalities, Proc. London Math. Soc. 10 (1960), 24–30.
[17] F. Himstedt, Character tables of parabolic subgroups of Steinberg’s triality groups, J. Algebra 281 (2004),
774–822.
, On the 2-decomposition numbers of Steinberg’s triality groups 3 D4 (q), q odd, J. Algebra 309
[18]
(2007), 569–593.
, Character tables of parabolic subgroups of Steinberg’s triality groups 3 D4 (2n ), J. Algebra 316
[19]
(2007), 254–283.
[20]
On the decomposition numbers of the Ree groups 2 F4 (q 2 ) in non-defining characteristic, J. Algebra 325 (2011), 364–403.
[21] F. Himstedt and S. Huang, Character table of a Borel subgroup of the Ree groups 2 F4 (q 2 ), LMS J.
Comput. Math. 12 (2009), 1–53.
[22]
, Character tables of the maximal parabolic subgroups of the Ree groups 2 F4 (q 2 ), LMS J. Comput.
Math. 13 (2010), 90–110.
[23]
, On the decomposition numbers of Steinberg’s triality groups 3 D4 (2n ) in odd characteristics,
Comm. Algebra 41 (2013), no. 4, 1484–1498.
[24] F. Himstedt, T. Le and K. Magaard Characters of the Sylow p-subgroups of the Chevalley groups D4 (pn ),
J. Algebra 332 (2011), 414–427.
[25] F. Himstedt and F. Noeske, Decomposition numbers of SO7 (q) and Sp6 (q), J. Algebra 413 (2014),
15–40.
[26]
, Restricting unipotent characters in special orthogonal groups, LMS J. Comput. Math. 18 (2015),
456–488.
[27] I. M. Isaacs, Characters of groups associated with finite algebras, J. Algebra, 177 (1995), 708–730.
[28]
, Counting characters of upper triangular groups, J. Algebra 315 (2007), no. 2, 698–719.
[29] H. Iwaniec and E. Kowalski, Analytic number theory, American Mathematical Society Colloquium Publications 53. American Mathematical Society, Providence, RI, 2004.
[30] T. Le and K. Magaard, On the character degrees of Sylow p-subgroups of Chevalley groups G(pf ) of type
E, Forum Math. 27 (2015), no. 1, 1–55.
[31] G. Lehrer, Discrete series and the unipotent subgroup, Compos. Math., 28 (1974), 9–19.
[32] E. Marberg, Combinatorial methods for character enumeration for the unitriangular groups, J. Algebra
345 (2011), 295–323.
[33] M. Marjoram, Irreducible characters of a Sylow p-subgroup of the orthogonal group, Comm. Algebra
27 (1999), 1171–1195.
[34] T. Okuyama and K. Waki, Decomposition numbers of Sp(4, q), J. Algebra 199 (1998), 544–555.
[35] I. Pak and A. Soffer, On Higman’s k(Un (Fq )) conjecture, preprint, arxiv:1507.00411.
[36] A. Previtali, On a conjecture concerning character degrees of some p-groups, Arch. Math. 65 (1995)
375–378.
[37] J. Sangroniz, Character degrees of the Sylow p-subgroups of classical groups, Groups St. Andrews 2001 in
Oxford. Vol. II, 487–493, London Math. Soc. Lecture Note Ser., 305, Cambridge Univ. Press, Cambridge,
2003.
23
[38]
, Irreducible characters of large degree of Sylow p-subgroups of classical groups, J. Algebra 321
(2009), 1480–1496.
[39] F. Szechtman, Irreducible characters of Sylow subgroups of symplectic and unitary groups, J. Algebra
303 (2006), 722–730.
[40] B. Szegedy, Characters of the Borel and Sylow subgroups of classical groups, J. Algebra 267 (2003),
130–136
[41] K. Waki, A note on decomposition numbers of G2 (2n ), J. Algebra 274 (2004), 602–606.
School of Mathematics, University of Birmingham, Birmingham, B15 2TT, U.K.
E-mail address: [email protected]
E-mail address: [email protected]
Department of Mathematics and Applied Mathematics, University of Pretoria, Pretoria
0002, South Africa
E-mail address: [email protected]
24
| 4math.GR
|
A Pseudo Knockoff Filter for Correlated Features
Jiajie Chen ∗, Anthony Hou †, Thomas Y. Hou
‡
arXiv:1708.09305v3 [stat.ME] 19 Sep 2017
September 21, 2017
Abstract
In [3], the authors introduce a new variable selection procedure called the knockoff filter to
control the false discovery rate (FDR) and prove that this method achieves exact FDR control.
Inspired by the work of [3], we propose and analyze a pseudo-knockoff filter that inherits some
advantages of the original knockoff filter and has more flexibility in constructing its knockoff
matrix. Although we have not been able to obtain exact FDR control of the pseudo knockoff
filter, we show that it satisfies an expectation inequality that offers some insight into FDR
control. Moreover, we provide some partial analysis of the pseudo knockoff filter for the half
Lasso and the least squares statistics. Our analysis indicates that the inverse of the covariance
matrix of the feature matrix plays an important role in designing and analyzing the pseudo
knockoff filter. Our preliminary numerical experiments show that the pseudo knockoff filter
with the half Lasso statistic has FDR control. Moreover, our numerical experiments show that
the pseudo-knockoff filter could offer more power than the original knockoff filter with the OMP
or Lasso Path statistic when the features are correlated and non-sparse.
1
Introduction
In many applications, we need to study a statistical model that consists of a response variable and a
large number of potential explanatory variables and determine which variables are truly associated
with the response. In [3], Barber and Candès introduce the knockoff filter to control the FDR in
a statistical linear model. More specifically, the knockoff filter constructs knockoff variables that
mimic the correlation structure of the true feature variables to obtain exact FDR control in finite
sample settings. It has been demonstrated that this method has more power than existing selection
rules when the proportion of null variables is high.
1.1
A brief review of the knockoff filter
Consider the following linear regression model
y = Xβ + ǫ,
where the feature matrix X is a n × p ( n ≥ p) matrix with full rank, its columns have been
normalized to be unit vectors in the l2 norm, and ǫ is a Gaussian noise N (0, σ 2 In ). The knockoff
filter begins with the construction of a knockoff matrix X̃ that obeys
X̃ T X̃ = X T X,
X̃ T X = X T X − diag(s),
(1)
where si ∈ [0, 1], i = 1, 2, ..., p. The positive definiteness of the Gram matrix [X X̃]T [X X̃] requires
diag(s) 2X T X.
∗
Applied and Computational Mathematics, Caltech; Email: [email protected]
Department of Statistics, Harvard University; Email: [email protected]
‡
Applied and Computational Mathematics, Caltech; Email: [email protected]
†
1
(2)
The first condition in (1) ensures that X̃ has the same covariance structure as the original feature
matrix X. The second condition in (1) guarantees that the correlations between distinct original
and knockoff variables are the same as those between the original variables. To ensure that the
method has good statistical power to detect signals, we should choose sj as large as possible
to maximize the difference between Xj and its knockoff X̃j . These two conditions are crucial
in establishing the exchangeability condition [3], which in turn gives exact FDR control. After
constructing the knockoff matrix, one needs to calculate a statistic, Wj , for each pair of Xj and
X̃j . One of the knockoff statistics considered in [3] is the Lasso path statistic, which is defined as
Wj = max(Zj , Z̃j ) · sign(Zj − Z̃j ), where Zj and Z̃j are the solutions of the Lasso path problem
given below:
1
2
||y − Xb − X̃ b̃||2 + λ(||b||1 + ||b̃||1 ) ,
(β̂(λ), β̃(λ)) = argmin
2
(b,b̃)
Zj = sup{λ : β̂j (λ) 6= 0}, Z̃j = sup{λ : β̃j (λ) 6= 0}.
The final step is to run the knockoff+ selection procedure at level q
1 + #{j : Wj ≤ −t}
T , min t > 0 :
≤ q , Ŝ , {j : Wj ≥ T } .
#{j : Wj ≥ t} ∨ 1
(3)
The main result in [3] is that such procedure controls exact FDR
"
#
#{j ∈ Ŝ : βj = 0}
F DR , E
≤ q, ∀q ∈ (0, 1) .
#{j ∈ Ŝ} ∨ 1
A knockoff filter for high-dimensional selective inference and model free knockoffs have been recently
established in [4, 5]. This research has inspired a number of follow-up works, such as [6, 7, 9, 14–16].
The power (the proportion of true discoveries) of the knockoff filter depends critically on the
value of si . In [6], we perform some analysis and numerical experiments on the knockoff filter
to understand how (1) and (2) impose a constraint on si when the features are correlated. Our
analysis shows that si could be small for strongly correlated features. On the other hand, for
strongly correlated features with a finite sample size, we simply do not have enough resolution in
the data to tell which variables are responsible for the response. In this case, it is more appropriate
to define clusters of these correlated variables and ask whether there is any signal in those clusters
rather than trying to decide which variable within the cluster is truly significant.
We note that a prototype knockoff filter based on group clustering of correlated data has been
proposed in [14], and a group knockoff filter has been also proposed by Dai and Barber in [7] with
exact FDR control. In [6], we also propose a PCA prototype group selection filter that achieves
exact FDR control. Our PCA prototype group selection filter has the advantage of being more
efficient and has higher power for correlated features. A localized knockoff filter has been proposed
by Xu et al [16]. There are several other feature selection methods that offer some level of FDR
control (e.g. [1, 2, 8, 10–13]). We refer to [3] for a thorough comparison between the knockoff filter
and these other approaches.
1.2
Three classes of pseudo knockoff filters
In this paper, we propose several pseudo-knockoff filters that inherit some advantages of the original
knockoff filter and have greater flexibility in constructing their pseudo-knockoff matrix. In the
pseudo knockoff filter, we seek to establish a weaker version of the exchangeability condition by
relaxing the second condition in (1). The first condition that we impose on the pseudo knockoff
filter is the following orthogonality condition:
(X + X̃)T (X − X̃) = 0.
2
(4)
It can be shown that this condition is equivalent to X T X = X̃ T X̃, X T X̃ = X̃ T X. In a linear
model y = Xβ + ǫ, ǫ ∼ N (0, σ 2 In ), this orthogonality condition implies that the projection of ǫ
onto span(X + X̃) and span(X − X̃) are independent.
In this paper, we consider three classes of pseudo knockoffs. For the first class of pseudo knockoff
filters, the pseudo knockoff matrix X̃ is chosen to be orthogonal to X, i.e. X T X̃ = X̃ T X = 0. We
call this pseudo knockoff the orthogonal pseudo knockoff. It maximizes the difference between the
pseudo knockoff matrix X̃ and its original design matrix X. Due to this orthogonality condition, we
can always distinguish the original Xj and its pseudo knockoff X̃j , independent of the correlation
structure of X.
The second class of pseudo knockoff filters is called the block diagonal pseudo knockoff. We
begin by constructing a block diagonal matrix B that satisfies the property B Σ−1 . We can then
solve for X̃ from the relationship B = 4[(X − X̃)T (X − X̃)]−1 . More specifically, we consider a
−1
−1
−1
block diagonal matrix B = 2diag(S11
, S22
, ..., Skk
), where Sii ’s are invertible matrices. It can be
shown that the condition (4) and the constraint B Σ−1 imply that
X T X = X̃ T X̃,
X T X − X T X̃ = diag(S11 , S22 , ..., Skk ).
(5)
It is interesting to note that the pseudo knockoff matrix, X̃, in (5) is exactly a group knockoff
matrix of X in [7]. Moreover, if we further impose Sii to be a diagonal matrix for i = 1, 2, ..., k,
then X̃ is a knockoff matrix of X [3]. In this sense, we can consider the block diagonal pseudo
knockoff filter as a generalization of the knockoff filter. We note that the group knockoff filter is
designed for group selection while the block diagonal pseudo knockoff filter is designed for feature
selection.
The third class of the pseudo knockoff filter is called the banded knockoff filter by imposing B
to be a banded matrix. The construction is similar to the case when B is a block diagonal matrix.
1.3
A half Lasso statistic
We propose to use a half penalized method to construct our pseudo knockoff filter. More specifically,
the pseudo knockoff statistic is based on the solution of the following half penalized optimization
problem
1
(6)
min ||y − X β̂ − X̃ β̃||22 + P (β̂ + β̃),
β̂,β̃ 2
where P (x) is an even non-negative and non-decreasing function in each coordinate of x. An important consequence of the orthogonality condition (4) is that we can reformulate the half penalized
problem into two sub-problems, i.e the minimization problem (6) is equivalent to the following
minimization problem:
)
(
)
(
1
X
−
X̃
1 X + X̃ ls
||
(β̂ + β̃ ls − β̂ − β̃)||22 + P (β̂ + β̃) + min
||
(β̂ ls − β̃ ls − (β̂ − β̃))||22 ,
min
2
2
2
2
β̂−β̃
β̂+β̃
where β ls and β̃ ls are the least squares coefficients by regressing y on the augmented feature matrix
[X, X̃]. If we choose P ≡ 0, we recover the least squares method. If we choose P = λ|| · ||l1 , we
obtain a half Lasso method, which we introduce in [6]. Once we solve the half penalized problem,
we can construct the pseudo knockoff statistic as follows
Wj , (β̂j + β̃j ) · sign(β̂j − β̃j ) or Wj = max (|β̂j |, |β̃j |) · sign(|β̂j | − |β̃j |).
(7)
We can show the following symmetry property of the pseudo knockoff filter. Conditional on some
σ- field, we have for any threshold t that
d
#{j : βj = 0 and Wj ≥ t} = #{j : βj = 0 and Wj ≤ −t}.
3
(8)
1.4
Partial analysis of the pseudo knockoff filter
The construction of our pseudo knockoff filter does not satisfy the exchangeability property, which
is crucial for the knockoff filter in establishing exact FDR control by using a supermartingale
argument [3]. In this paper, we provide some partial analysis for the pseudo knockoff filter by
trying to establish the following expectation inequality:
#{j : Wj ≥ t, βj = 0}
≤1
(9)
E
#{j : Wj ≤ −t, βj = 0} + m
for any fixed threshold t > 0 and some m ∈ N+ that depends on the pseudo knockoff matrix X̃.
This inequality can be considered as an approximation to the corresponding expectation inequality
for the knockoff filter:
#{j : Wj ≥ T, βj = 0}
E
≤ 1,
(10)
#{j : Wj ≤ −T, βj = 0} + 1
for an adaptive threshold T .
In this paper, we establish the expectation inequality (9) for the block diagonal and the banded
pseudo knockoff filters. For the orthogonal pseudo knockoff filter, we use a probabilistic argument
to obtain a relatively tight upper bound to the expectation in (9) for t = 0 when Σ−1 is diagonally
dominated or when Σ−1 has some special structure. Interestingly, our analysis reveals that the
decaying property of Σ−1 plays an important role in determining the performance of the pseudo
knockoff filter and the knockoff filter.
We also carry out a number of numerical experiments to test the performance of the three
classes of pseudo knockoff filters and compare their performance with that of the knockoff filter. In
the examples that we consider in this paper, we find that the three classes of pseudo knockoff filters
with the half Lasso statistic have FDR control. The orthogonal pseudo knockoff filter seems to offer
the most power among all other pseudo knockoff filters. In the case when the features are highly
correlated and non-sparse (e.g. about 20% non-null features), the orthogonal pseudo knockoff filter
offers more power than the knockoff filter with the OMP or Lasso Path statistic. In the extreme
case when Σ−1 decays very slowly away from its main diagonal, the knockoff filter with the OMP
or the Lasso Path statistic tends to lose considerable power. The orthogonal pseudo knockoff filter
still offers reasonably high power in this case.
The rest of the paper is organized as follows. In Section 2, we introduce the three classes
of pseudo-knockoff filters and discuss some essential properties of the pseudo knockoff filters. In
Section 3, we present a number of numerical experiments to demonstrate the effectiveness of the
proposed methods. In Section 4, we provide some partial analysis of the pseudo knockoff filters
and present additional numerical experiments to compare the performance of the pseudo knockoff
filters with the knockoff filter using the OMP and the Lasso Path statistics.
2
2.1
A pseudo knockoff filter
An expectation inequality
In the knockoff filter, the exchangeability property is essential [3]. This important property allows
one to apply a supermartingale argument to obtain the following expectation inequality
#{j : Wj ≥ T, βj = 0}
≤ 1,
(11)
E
#{j : Wj ≤ −T, βj = 0} + 1
for an adaptive threshold T . The above estimate plays a crucial role in establishing exact FDR
control of the knockoff filter [3]. Our motivation is to relax the second constraint in (1) in a way
such that (11) is still approximately valid. In particular, we consider a modified version of (11)
4
#{j : Wj ≥ t, βj = 0}
≤1
E
#{j : Wj ≤ −t, βj = 0} + m
(12)
for any fixed threshold t > 0 and some m ∈ N+ that depends on the pseudo knockoff matrix X̃.
To see how (12) relates to FDR control, we first define the pseudo knockoff+ adaptive threshold T
as follows
#{j : Wj ≤ −t} + m
≤q .
(13)
Tm , min t :
#{j : Wj ≥ t} ∨ 1
By the definition of Tm , we have
#{j ∈ S0 : Wj ≥ T }
#{j ∈ S0 : Wj ≥ T }
#{j ∈ S0 : Wj ≤ −T } + m
F DR , E
·
=E
#{j : Wj ≥ T } ∨ 1
#{j ∈ S0 : Wj ≤ −T } + m
#{j : Wj ≥ T } ∨ 1
#{j : Wj ≥ T, βj = 0}
≤q·E
,
#{j : Wj ≤ −T, βj = 0} + m
(14)
where S0 = {j : βj = 0}. For m = 1, (11) and (14) establish FDR control of the knockoff filter.
One may consider (12) as an approximation of (11). Our numerical study shows that when (12) is
valid, (11) is approximately valid and we have FDR control for the pseudo knockoff filter. However,
(12) alone does not give exact FDR control since Tm is a random variable.
2.2
The Basic Constraint and a Symmetry Property
The basic constraint of the pseudo knockoff matrix is given by
X̃ T X̃ = X T X,
X T X̃ = X̃ T X.
(15)
The correlation structure and the commutativity imply
(X + X̃)T (X − X̃) = X T X + X̃ T X − X T X̃ − X̃ T X̃ = 0.
(16)
We can prove that (15) and (16) are equivalent. If (16) holds, we have X T X−X̃ T X̃ = X T X̃ −X̃ T X.
Note that the right hand side is a symmetric matrix, while the left hand side is a skew-symmetric
matrix. It follows that X T X − X̃ T X̃ is symmetric and skew-symmetric. Thus we must have
X T X − X̃ T X̃ = 0, which further implies X T X̃ − X̃ T X = 0. These two equations establish (15).
The orthogonality condition (16) is the foundation of the pseudo knockoff filter and leads to a
symmetry property of the pseudo knockoff filter.
Least squares coefficients
mented design matrix [X X̃]
Consider the least squares coefficients of regressing y on the aug-
y ∼ X β̂ ls + X̃ β̃ ls =
X − X̃ ls
X + X̃ ls
(β̂ + β̃ ls ) +
(β̂ − β̃ ls ).
2
2
Using y = Xβ + ǫ and the orthogonality condition (16), we have a simple expression of the least
squares coefficients,
"
#T "
#−1
!
ls
X+X̃ T
ls
X
+
X̃
X
−
X̃
X
−
X̃
X
+
X̃
(
)
ǫ
β̂ + β̃ − β
2
,
,
=
2
2
2
2
β̂ ls − β̃ ls − β
( X−2 X̃ )T ǫ
(17)
h
i−1
(1)
( X+2 X̃ )T X+2 X̃
( X+2 X̃ )T ǫ
, ǫ
i−1
= h
.
ǫ(2)
( X−2 X̃ )T X−2 X̃
( X−2 X̃ )T ǫ
5
The above relationship will be used repeatedly throughout the paper. Denote
η , β̂ ls + β̃ ls = β + ǫ(1) , ξ , β̂ ls − β̃ ls = β + ǫ(2) .
(18)
From the orthogonality property (16), we know that ( X+2 X̃ )T ǫ and ( X−2 X̃ )T ǫ have independent
multivariate normal distributions. Using (17), we know that ǫ(1) and ǫ(2) , η = β̂ ls + β̃ ls and
ξ = β̂ ls − β̃ ls are also independent.
The Pseudo Knockoff Statistic Similar to the knockoff filter process, we construct a pseudo
knockoff statistic for each pair of the original feature Xj and its pseudo knockoff X̃j .
Consider a half penalized optimization problem
1
min ||y − X β̂ − X̃ β̃||22 + P (β̂ + β̃).
β̂,β̃ 2
(19)
Denote by r the residue of regressing y onto the augmented design matrix [X X̃], i.e. r = y − X β̂ ls −
X̃ β̃ ls . The geometric property of the least squares method implies r ⊥ X, X̃, which leads to
||y − X β̂ − X̃ β̃||22 = ||r + X β̂ ls + X̃ β̃ ls − X β̂ − X̃ β̃||22 = ||r||22 + ||X β̂ ls + X̃ β̃ ls − X β̂ − X̃ β̃||22 .
Using the orthogonality condition (X + X̃)T (X − X̃) = 0 and the fact that r is independent of β̂, β̃,
we can exclude the residue r from the optimization problem (19) and reformulate it as follows
1
1
1
min ||y − X β̂ − X̃ β̃||22 + P (β̂ + β̃) = min ||r||22 + ||X β̂ ls + X̃ β̃ ls − X β̂ − X̃ β̃||22 + P (β̂ + β̃)
2
β̂,β̃ 2
β̂,β̃ 2
1
⇐⇒ min ||X β̂ ls + X̃ β̃ ls − X β̂ − X̃ β̃||22 + P (β̂ + β̃)
β̂,β̃ 2
)
) (
(
1 X − X̃ ls
1 X + X̃ ls
ls
2
ls
2
||
(β̂ + β̃ − β̂ − β̃)||2 + P (β̂ + β̃) +
||
(β̂ − β̃ − (β̂ − β̃))||2
= min
2
2
2
2
β̂,β̃
(
)
(
)
1 X + X̃ ls
1 X − X̃ ls
ls
2
ls
2
⇐⇒ min
||
(β̂ + β̃ − β̂ − β̃)||2 + P (β̂ + β̃) + min
||
(β̂ − β̃ − (β̂ − β̃))||2 ,
2
2
2
2
β̂+β̃
β̂−β̃
(20)
ls
ls
where β̂ and β̃ are the least squares coefficients given in (17). It is easy to derive that the
solution of these problems can be expressed as
β̂ + β̃ = f (β̂ ls + β̃ ls ) = f (η),
β̂ − β̃ = β̂ ls − β̃ ls = ξ ,
(21)
for some function f : Rp → Rp . We construct the pseudo knockoff statistic as follows
Wj , (β̂j + β̃j ) · sign(β̂j − β̃j ) or Wj = max (|β̂j |, |β̃j |) · sign(|β̂j | − |β̃j |).
(22)
The pseudo knockoff statistic satisfies the following two properties.
Amplitude Property The amplitude of W is decided by β̂ + β̃ = f (η) and |β̂ − β̃| = |ξ|.
In fact, using the definition of W and (21), we have
|W | = |β̂ + β̃| = |f (η)|
or
1
|W | = |β̂| ∨ |β̃| = (|β̂ + β̃ + |β̂ − β̃|| ∨ |β̂ + β̃ − |β̂ − β̃||).
2
Sign Property The sign of W is determined by sign(β̂ + β̃) and sign(β̂ − β̃). Since sign(|β̂|−
|β̃|) = sign(|β̂|2 − |β̃|2 ), we have
sign(W ) = sign(β̂ + β̃) · sign(β̂ − β̃) = sign(f (η)) · sign(ξ),
for both definitions of W .
Now we show that the pseudo knockoff statistic satisfies a symmetry property.
6
Proposition 2.1. (The Symmetry Property of the Pseudo Knockoff Statistic) Conditional on η,
we have for any threshold t that
d
#{j : βj = 0 and Wj ≥ t} = #{j : βj = 0 and Wj ≤ −t},
(23)
where the pseudo knockoff statistic Wj is defined in (22).
Proof. According to (18) and (21), the solution of the half penalized problem can be expressed as
β̂ + β̃ = f (η) = f (β + ǫ(1) ),
(ǫ(1) , ǫ(2) )
(ǫ(1) , −ǫ(2) )
β̂ − β̃ = ξ = β + ǫ(2) .
Next, we replace
by
to generate a new pair of solutions
(2)
(2)
From (18), changing ǫ to −ǫ does not change η. Thus, we obtain
β̂ new + β̃ new = f (η) = β̂ + β̃,
From (24) and (25), we have
(24)
(β̂ new , β̃ new ).
β̂ new − β̃ new = β − ǫ(2) .
(β̂ new + β̃ new )S0 = f (η)S0 = (β̂ + β̃)S0 ,
(2)
(β̂ new − β̃ new )S0 = (β − ǫ(2) )S0 = −ǫS0 = −(β̂ − β̃)S0 ,
(25)
(26)
where S0 , {j : βj = 0}. The amplitude and sign property of W imply |WSnew
| = |WS0 | and
0
sign(WSnew
) = sign((f (η))S0 · (−ξ)S0 ) = −sign((f (η))S0 · (ξ)S0 ) = −sign(WS0 ).
0
Hence
WSnew
= −WS0 .
0
Recall that WS0 is generated by ǫ(1) , ǫ(2) and that ǫ(1) , ǫ(2) have independent multivariate normal
distributions with zero mean. Conditional on η (or equivalently ǫ(1) ), we have
d
d
(ǫ(1) , ǫ(2) ) = (ǫ(1) , −ǫ(2) ) =⇒ WS0 = WSnew
= −WS0 .
0
(27)
As a result, conditional on η, we get
d
#{j : βj = 0 and Wj ≥ t} = #{j : βj = 0 and Wj ≤ −t},
for any threshold t > 0.
Pseudo knockoff least squares and half Lasso statistics We give two examples of the pseudo
knockoff statistics.
The least squares statistic Let P ≡ 0 in (20). The half penalized problem becomes the
least squares problem of the response y and the augmented design matrix [X X̃].
A half Lasso statistic In [6], we introduce a half penalized method for the knockoff filter.
The derivation in (20) implies that we can also apply this idea to the pseudo knockoff filter.
Specifically, we choose P (x) = λ||x||1 in (20) to obtain a half Lasso statistic:
1
min ||y − X β̂ − X̃ β̃||22 + λ||β̂ + β̃||1 .
β̂,β̃ 2
Let β̂, β̃ be the solution of the least squares or the half Lasso problem and we define the pseudo
knockoff statistic Wj = |β̂j + β̃j |sign(β̂ − β̃) or Wj , max(|β̂j |, |β̃j |)·sign(|β̂j |−|β̃j |). The symmetry
property (23) of these statistics is guaranteed by Proposition 2.1. The tuning parameter λ can be
determined by a projection
1
λ=µ· √
||U T y||2 ,
n − 2p
where U ∈ Rn×(n−2p) is an orthonormal matrix such that [X X̃]T U = 0 and µ is a parameter that
we can choose empirically. In fact, U T y is exactly the residue of regressing y onto [X X̃]. One can
verify the symmetry property of the pseudo knockoff statistic with this tuning parameter using a
similar argument.
7
2.3
Three Classes of the Pseudo Knockoff Matrix
We have proposed the basic constraint (15) for the pseudo knockoff matrix in the last section. In this
section, we impose an additional constraint on X̃ so that we can establish an additional property
on the pseudo knockoff statistic. In particular, we are interested in three classes of pseudo knockoff
matrices, namely the orthogonal, the block diagonal, and the banded pseudo knockoff matrices.
2.3.1
An Orthogonal Construction
In this special pseudo knockoff filter, we impose an orthogonal constraint on the second condition
of (15)
X T X = X̃ T X̃, X T X̃ = X̃ T X = 0
(28)
to yield the orthogonal pseudo knockoff matrix.
To construct an orthogonal pseudo knockoff matrix X̃, we first find the SVD of X ∈ Rn×p : X =
U DV T , U ∈ Orthn×p , D = diag{σ1 , ..., σp } and V ∈ Orthp×p . We then choose any orthonormal
matrix W ∈ Rn×p , whose column space is orthogonal to that of X (i.e. X T W = 0), and construct
the pseudo knockoff matrix X̃ as X̃ = W DV T . It is easy to verify that X̃ satisfies the pseudo
matrix condition (28).
2.3.2
A Block Diagonal Construction
It follows from (17) and (18) that the covariance matrix of ǫ(2) , or equivalently ξ, is given by
B , 4[(X − X̃)T (X − X̃)]−1 .
(29)
We can design B to yield a special correlation structure on ξ. Due to the existing constraint (15) or
(16), the covariance matrix B cannot be chosen arbitrarily. Below we give a necessary and sufficient
condition on B to find X̃ that satisfies (16) and (29).
Necessary Condition on B Assume that there exists some X̃ that satisfies (16) and (29)
and X − X̃ has full rank. Performing SVD on (X − X̃)/2, we have (X − X̃)/2 = PM−1 for
some orthonormal matrix P ∈ Rn×p and some invertible matrix M ∈ Rp×p . As a result, we get
B = [(PM−1 )T (PM−1 )]−1 = MMT and X̃ = X − 2PM−1 . Substituting the last equation into
the orthogonal condition (X + X̃)T (X − X̃) = 0 (16), we obtain
4(X − PM−1 )T PM−1 = 0
Since P ∈
Rn×p
⇐⇒
M−1 = P T X
is orthonormal, we have
⇐⇒
=⇒
X T P P T X X T IX = X T X = Σ
=⇒
M−T M−1 = M−T PT X
B = (X T P P T X)−1 .
B = (X T P P T X)−1 Σ−1 .
(30)
Sufficiency We prove that the condition (30) on B is sufficient to find a X̃ that satisfies (16) (or
(15)) and (29). To see this, we construct X̃ as follows
X̃ = X(I − 2Σ−1 B−1 ) + 2UCB−1
(31)
where C ∈ Rp×p satisfies CT C = B − Σ−1 and U ∈ Rn×p is an orthonormal matrix with UT X = 0.
We can verify that X̃ constructed from (31) satisfies (16) and (29) in the Appendix.
−1
−1
−1
),
A Block Diagonal Construction Consider a block diagonal matrix B = 2diag(S11
, S22
, ..., Skk
where Sii ’s are invertible matrices. The constraint on B is equivalent to
2B−1 = diag(S11 , S22 , ..., Skk ) 2Σ.
(32)
Hence (X − X̃)T (X − X̃) = 4B−1 = 2diag(S11 , S22 , ..., Skk ). Using this relationship together with
the basic constraint (15), i.e. X T X = X̃ T X̃, X T X̃ = X̃ T X, we obtain
X T X = X̃ T X̃,
X T X − X T X̃ = diag(S11 , S22 , ..., Skk ).
8
(33)
Construction Assume that X can be clustered into (XG1 , XG2 , ..., XGk ), |Gi | = gi . Inspired by
the group knockoff construction in [7], we first choose Sii as
T
Sii , γΣGi ,Gi = γXG
XGi , i = 1, 2, ..., k.
i
The constraint (32) implies γ · diag(ΣG1 ,G1 , ΣG2 ,G2 , ..., ΣGk ,Gk ) 2Σ. In order to maximize the
difference between X and X̃, γ should be chosen as large as possible: γ ≤ min{1, 2 · λmin (DΣD)},
−1/2
−1/2
−1/2
where D = diag(ΣG1 ,G1 , ΣG2 ,G2 , ..., ΣGk ,Gk ). To ensure that the matrix (X + X̃)T (X + X̃) is
1
min{1, 2 · λmin (DΣD)} in our numerical experiments. Once we
nonsingular, we choose γ = 1.2
construct B, we can generate the pseudo knockoff matrix via the procedure described earlier.
Connection between the pseudo knockoff filter and the knockoff filter. By comparing
our block diagonal pseudo knockoff construction with the group knockoff filter in [7], we can see that
the pseudo knockoff matrix, X̃, in (33) is actually a group knockoff matrix of X. Moreover, if Sii
is a diagonal matrix for i = 1, 2, .., k, then X̃ is a knockoff matrix of X [3]. Recall that the second
constraint in (1) in the original knockoff filter requires that X T X − X T X̃ be a diagonal matrix. In
the block diagonal pseudo knockoff filter, we only requite that X T X − X T X̃ be a block diagonal
matrix. In this sense, we can consider the block diagonal construction of the pseudo knockoff filter
as a generalization of the group knockoff matrix proposed by Dai-Barber in [7]. The group knockoff
filter is originally designed for group selection with group FDR control while our block diagonal
pseudo knockoff filter is designed for feature selection. We will show that the pseudo knockoff filter
with the block diagonal construction satisfies (12) and provide some partial analysis of this method
later on. Our numerical experiments seem to suggest that this method works reasonably well.
2.3.3
A Banded Matrix Construction
Consider a banded matrix B with band width m, i.e. Bij = 0 if |i − j| > m. We are mostly
interested in m = 2. The other cases can be handled similarly. There are many approaches to
construct B that satisfy B Σ−1 . We propose one of these approaches and outline the main steps
as follows.
Extract the tri-diagonal elements of Σ−1 , i.e.
(Σ−1 )11 (Σ−1 )12
..
..
(Σ−1 )21
.
.
D,
.
.
..
..
(Σ−1 )p−1,p
−1
(Σ )p,p−1
(Σ−1 )pp
Step 1: Extraction
.
In general, D is not a positive definite matrix. We need to modify the diagonal elements of D to
get a positive definite matrix.
Step 2: Modification of D We consider the following modification of D:
e ii = Dii +
D
D2i,i+1
, i = 1, 2, ..., p − 1,
Di+1,i+1
e pp = Dpp ,
D
e ij = Dij , ∀i 6= j.
D
e is positive definite, we consider vT Dv
e for an arbitrary vector v 6= 0 ∈ Rp .
To see that D
!
p−1
p−1
X
X
D2i,i+1
T e
Di,i+1 vi vi+1 + Dpp vp2
vi2 + 2
Dii +
v Dv =
Di+1,i+1
i=1
i=1
!
p−1
2
X
D
i,i+1
2
Di+1,i+1 vi+1
+
=
v2 + 2Di,i+1 vi vi+1 + D11 v12
Di+1,i+1 i
i=1
9
=
p−1
X
i=1
1
Di+1,i+1
(Di,i+1 vi + Di+1,i+1 vi+1 )2 + D11 v12 > 0,
∀ v 6= 0.
e is positive definite, we take
Step 3: Construction of B Since D
e −1/2 Σ−1 D
e −1/2 ) · D,
e
B , τ λmax (D
e −1/2 Σ−1 D
e −1/2 )I D
e −1/2 Σ−1 D
e −1/2 , we have
where τ ≥ 1 is a parameter. Since λmax (D
e 1/2 · λmax (D
e −1/2 Σ−1 D
e −1/2 )I · D
e 1/2 D
e 1/2 · D
e −1/2 Σ−1 D
e −1/2 · D
e 1/2 = Σ−1 .
B = τD
e T (X + X)
e is nonsingular.
In our numerical experiments, we choose τ = 1.2 to ensure that (X + X)
Remark 1. For some design matrices with a special correlation structure, we can use a different
construction to increase the difference between X and X̃. If X can be clustered into a number
of groups such that the between-group correlation is weak, we can utilize the block diagonal construction. If the correlation between Xi , Xj decreases rapidly as |i − j| increases, we can adopt the
banded construction to increase the difference between X and X̃.
Independence Let m be the largest block size of B in the block diagonal construction or the
band width in the banded construction. Since the covariance matrix of ξ = β̂ ls − β̃ ls is B, we can
divide ξi , 1 ≤ i ≤ p into m groups C1 , C2 , ..., Cm such that the random variables in the same group
are mutually independent.
For the block diagonal construction, Ci consists of the i-th element in each block if there exists
such an element. For the banded construction, Ci , {j : j ≡ i(mod m), 1 ≤ j ≤ p}. It is easy to
verify that the random variables within each group are independent.
2.4
A theoretical result for the expectation inequality
In this section, we prove (12) for the pseudo knockoff statistic that is constructed using either a
block diagonal or a banded matrix construction.
Theorem 2.2. Assume that the pseudo knockoff matrix is generated via a block diagonal or a
banded construction and the pseudo knockoff statistic is defined as W , (β̂ + β̃) · sign(β̂ − β̃). For
any t > 0, we have
#{j ∈ S0 : Wj ≥ t}
≤1,
(34)
E
#{j ∈ S0 : Wj ≤ −t} + m
where m is the largest group size in the block diagonal construction or the band width in the
banded construction. When m = 1, (34) replicates a result similar to (11).
Proof. Let F be the σ-field generated by η. Since |W | = |f (η)| by our choice of W , conditional
on η, we can determine |W | and Nt , {j ∈ S0 : |Wj | ≥ t}. We will show that the conditional
expectation satisfies
#{j ∈ S0 : Wj ≥ t}
E
F ≤ 1.
(35)
#{j ∈ S0 : Wj ≤ −t} + m
Once we obtain (35), we can integrate both sides to obtain (34). From our previous discussion, we
can divide Nt into m groups C1P
, C2 , ..., Cm (Ci ⊂ S0 ) such that the elements of ξCi are mutually
independent. Obviously, |Nt | = ki=1 |Ci |. Using the following Cauchy-Schwarz inequality
m
m
X
a2 X
i
i=1
bi
i=1
m
X
1
ai )2 ⇐⇒ Pm
bi ≥ (
i=1
i=1 ai i=1
10
Pm
ai
≥ Pi=1
, a i , bi > 0 ,
m
bi
i=1 bi
m
X
a2
i
with ai = |Ci | + 1, bi = #{j ∈ Ci : Wj ≤ −t} + 1, we obtain
#{j ∈ S0 : Wj ≥ t}
|Nt | + m
F = E Pm
E
F −1
#{j ∈ S0 : Wj ≤ −t} + m
i=1 (#{j ∈ Ci : Wj ≤ −t} + 1)
"
#
m
X
(|Ci | + 1)2
1
≤E
F −1
|Nt | + m
#{j ∈ Ci : Wj ≤ −t} + 1
i=1
m
X
|Ci | + 1
|Ci | + 1
E
F −1
=
|Nt | + m
#{j ∈ Ci : Wj ≤ −t} + 1
i=1
m
X
#{j ∈ Ci : Wj ≥ t}
|Ci | + 1
=
1+E
F
− 1.
|Nt | + m
#{j ∈ Ci : Wj ≤ −t} + 1
(36)
i=1
Here, we use #{j ∈ Ci : Wj ≤ −t} + 1 + #{j ∈ Ci : Wj ≥ t} = |Ci | + 1 to obtain the first and the
last equality, and use the fact that |Nt |, |Ci | are measurable w.r.t F to yield the second equality.
Note that ξ and η are independent, sign(ξj ), j ∈ Ci are mutually independent, and ξS0 have
mean zero multivariate normal distributions. Conditional on F, sign(Wj ) = sign(f (η)j )sign(ξj ), j ∈
Ci are independent and thus sign(Wj ), j ∈ Ci obeys a binomial distribution. We yield
#{j ∈ Ci : Wj ≥ t}
#{j ∈ Ci : Wj > 0}
E
F =E
F ≤ 1.
(37)
#{j ∈ Ci : Wj ≤ −t} + 1
#{j ∈ Ci : Wj < 0} + 1
Therefore, the last line in (36) is bounded by
m
−1+
X
2
1
2(|Ci | + 1) = −1 +
· (|Nt | + m) = 1 .
|Nt | + m
|Nt | + m
(38)
i=1
Combining (36) with (38) yields (35). Finally, we integrate both sides of (35) to conclude the proof.
Remark 2. In order to control the expectation in (34) by 1, we consider the conditional expectation
(35) and add the maximal group size m (or the band width in the banded construction) in the
denominator. We show that this is optimal.
Let us consider the block diagonal construction with β = 0 (no signal). Recall that the covariance matrix of ξ is block diagonal, i.e. B = diag(D11 , D22 , ..., Dkk ) . Assume that the size of each
block is m and the off-diagonal elements of Dii are close to the diagonal elements. Accordingly,
ξi , ξj are strongly correlated if i, j are in the same block or are independent if i, j come from
different blocks. Let t = 0 in (35). Conditional on η, we have sign(Wj ) = sign((f (η))j )sign(ξj ).
If sign((f (η))j )’s are the same in each block, i.e. sign((f (η))i,j ) = sign((f (η))i,l ), we yield
Z1 ,
p
X
i=1
1Wi >0 =
m
k X
X
i=1 j=1
1Wi,j >0 =
m
k X
X
i=1 j=1
1ξi,j ·(f (η))i,j >0 ≈ m
k
X
i=1
1ξi,1 ·(f (η))i,1 >0 = m
k
X
1Wi,1 >0 .
i=1
Here, ξi,j and Wi,j are the j-th element in block i. In the above derivation, we have used that
ξi , ξj P
are strongly correlated if i, j are in the same block. Similar analysis can be applied to
Z2 , pi=1 1Wi <0 . Plugging these estimates into the expectation, we obtain
"
#
" Pk
#
P
m ki=1 1Wi,1 >0
1
Z1
W
>0
i,1
i=1
E(
F) ≈ E
F =E
F .
P
P
Z2 + m
m + m ki=1 1Wi,1 <0
1 + ki=1 1Wi,1 <0
Since ξi,1 are mutually independent with mean 0, 1Wi,1 >0 are i.i.d and obeys a binomial distribution.
Thus, the expectation on the right hand side is 1 − 2−k . Consequently, for any m̃ < m, it is likely
that the expectation E(Z1 /(Z2 + m̃) F) exceeds 1.
11
In the above analysis, we rely on the fact that the amplitude of WS0 and sign(WS0 ) are controlled
by independent random variables η, ξ. In our numerical experiments, we find that the signed max
statistic, i.e. W = |β̂| ∨ |β̃| · sign(|β̂| − |β̃|), is more powerful than the statistic in (34). To
establish (34) for the signed max statistic, we cannot apply the same argument since the amplitude
of WS0 depends on η and |ξS0 |. If the pseudo knockoff matrix is generated by the block diagonal
construction, we have the following result.
Theorem 2.3. Assume that the pseudo knockoff matrix is generated by the block diagonal construction. For any t > 0, (34) is true for the signed max statistic.
Proof. From the amplitude property of the signed max statistic, |WS0 | depends on η, |ξS0 |. Let F be
the σ-field generated by η and |ξS0 |. In the following, we will use the same notations Nt , Ci (Ci ⊂ S0 )
as in the proof of Theorem 2.2. Note that (36) does not rely on the independence of ξCi . Based on
the previous proof, we only need to verify (37).
For the block diagonal construction, the elements of Ci come from different blocks. Note that
(2)
(2)
V ar(ξ) = B = diag(S11 , S22 , ..., Skk ) and ξS0 = ǫS0 (see (18)). We can change the sign of ǫS0 in any
(2)
block Si1 i1 , Si2 i2 , ..., Sij ij without changing |ξS0 | and the joint distribution of ǫS0 . Consequently,
conditional on F, the elements of sign(ξCi ) are mutually independent. Using the independence of
sign(ξCi ) together with the sign property of W , we conclude that the elements of sign(WCi ) are
i.i.d random variables satisfying
P (sign(WCi ,j ) = 1) = P (sign(WCi ,j ) = −1) = 1/2.
Finally, conditional on F, #{j ∈ Ci : Wj ≥ t} ∼ Binomial(|Ci |, 1/2) and #{j ∈ Ci : Wj ≤
−t} = |Ci | − #{j ∈ Ci : Wj ≥ t} imply (37).
Variance of the numerator We show that the variance of the numerator in (34) for t = 0+
is O(mp). Denote N0 , {j ∈ S0 : (f (η))j 6= 0}. By definition, N0 is determined by η. Based on
the sign property of W and the fact that ξ is a multi-normal random variable, conditional on η,
N0 = {j ∈ S0 : Wj 6= 0} and {j ∈ S0 : Wj > 0} = {j ∈ N0 : Wj > 0} almost surely. For the block
diagonal or banded pseudo knockoff, we can divide N0 into m groups C1 , C2 , ..., Cm (Ci ⊂ S0 ) such
that the elements of ξCi are mutually independent. Applying the Cauchy-Schwarz inequality yields
m
m
X
X
V ar(#{j ∈ Ci : Wj > 0} η)
#{j ∈ Ci : Wj > 0} η) ≤ m
V ar(#{j ∈ S0 : Wj > 0} η) = V ar(
i=1
i=1
m X
m X
X
X
m|N0 ||
1
)=
= O(mp). (39)
(
V ar(1Wj >0 |η)) = m
(
=m
4
4
i=1 j∈Ci
i=1 j∈Ci
To obtain the second and the third equality, we apply a property of Wj in the proof of Theorem
2.2 and Theorem 2.3 that sign(Wj ), j ∈ Ci are i.i.d conditional on η with P (sign(Wj ) = 1|η) =
P (sign(Wj ) = −1|η) = 1/2.
It is interesting to perform a similar variance estimate for the knockoff filter. Since the knockoff
statistic W has the property that sign(Wj ), j ∈ S0 are i.i.d conditional on |W | with P (sign(Wj ) =
1|F) = P (sign(Wj ) = −1|F) = 1/2, we obtain V ar(#{j ∈ S0 : Wj > 0} F) = |S0 |/4 = O(p),
where F is the σ- field generated by |W |.
3
Numerical results for the pseudo knockoff filter
In this section, we perform a number of numerical experiments to test the robustness of the pseudo
knockoff filter and study the performance of various methods.
Default Setting
12
i.i.d
Notations. βi ∼ {±A} means that βi takes value A or −A independently with equal probability 1/2. We denote the orthogonal pseudo knockoff, the pseudo knockoff with the block diagonal
construction, and the pseudo knockoff with the banded construction as orthogonal, block diagonal
and banded.
Data Throughout all simulations, the noise ǫ ∈ Rn is a standard Gaussian, i.e. ǫ ∼ N (0, In ).
Given some covariance matrix Σ, we first draw the rows of the design matrix X ∈ Rn×p from
a multivariate normal distribution N (0, Σ), and then normalize the columns of X. The pseudo
knockoff matrix (orthogonal, block diagonal or banded) is generated according to Section 2.3; the
knockoff matrix is generated by the standard SDP construction in [3]. To generate the signal
i.i.d
strength β ∈ Rp , we choose k coefficients βi1 , βi2 , ..., βik randomly and set βij ∼ {±A}.Finally,
the response variable y ∈ Rn is generated from y = Xβ + ǫ. Without specification, the sample size
is p = 500, n = 1500, the sparsity is k = 30, the signal amplitude is A = 3.5 and the covariance
matrix is Σ = Ip .
Pseudo Knockoff+ Threshold We introduce the pseudo knockoff+ adaptive threshold Tm
in (13) and discuss how (13) and (12) contribute to the FDR control. In Remark 2, we argue
that the inequality (12) is tight in some extreme case. In general, adding m to the denominator
is an over-estimate. Our numerical study indicates that we can replace m by 1 and the associated
expectation is still close to 1 in the numerical examples that we consider in this paper. In the
following numerical experiments, we choose T1 as the adaptive threshold to select features. We
note that this criterion is exactly the same as the original knockoff+ adaptive threshold. The
nominal FDR level in the definition of T1 is q = 20%.
Metrics We use the following metrics to study the robustness of various methods: the FDR
(the mean false discovery proportion), the power (the proportion of true discoveries) and the
expectation, which is defined as the mean of
#{j : Wj ≥ T1 & βj = 0}
.
#{j : Wj ≤ −T1 & βj = 0} + 1
(40)
Methods The methods that we focus on include the orthogonal, the block diagonal, the
banded pseudo knockoff filters with the half Lasso statistic (λ = 0.75), and the knockoff filter with
the orthogonal matching pursuit (OMP) statistic [3]. The reason we choose the OMP statistic for
the knockoff filter is because the OMP statistic seems to give the most robust performance among
other knockoff statistics from our numerical study in this paper and in [6]. For the banded pseudo
knockoff filter, we construct Wj = (β̂j + β̃j ) · sign(β̂j − β̃j ); for the other methods, we use the signed
max statistic, i.e. Wj = |β̂j | ∨ |β̃j | · sign(|β̂j | − |β̃j |).
3.1
Numerical evidence of FDR control for the pseudo knockoff filter
In this subsection, we perform extensive numerical experiments to test whether the pseudo knockoff
filter has FDR control. For this purpose, we apply it to select features in the linear model y = Xβ+ǫ
with different design matrices under various extreme conditions.
The default simulated data is discussed at the beginning of Section 3 and we vary one of
the default settings in each experiment as follows (one setting is varied while keeping the others
unchanged).
(a) Sparsity: k varies from 10, 20, 30, ..., 90, 100.
(b) Signal amplitude: A varies from 2.8, 2.9, ..., 4.2.
(c) Correlation Structure: We use the covariance matrix Σ ∈ R500×500 , Σij = ρ|i−j| and vary
the correlation level ρ = 0, 0.1, ..., 0.9.
(d) The sample size: We vary the sample size n = 150l, p = 50l and sparsity k = 10l with
l ∈ {2, 3, ..., 12}.
13
FDR
Expectation
Power(%)
Group Structure: We consider a design matrix X ∈ R1500×500 with a group structure. Specifically, we assume that the features Xj can be clustered into 100 groups with 5 features in each
group. To generate a different group structure, we choose the covariance matrix Σii = 1, Σij = ρ
for i 6= j in the same group and Σij = γ · ρ for i 6= j in different groups and generate the design
matrix X as in the previous discussion.
(e) The within-group correlation: γ = 0 is fixed and ρ varies from 0, 0.1, 0.2, ..., 0.9.
(f ) The between-group correlation: ρ = 0.5 is fixed and γ varies from 0, 0.1, 0.2, ..., 0.9.
We test the orthogonal, banded, and block diagonal pseudo knockoff filters and pay particular
attention to the FDR control, the power and the expectation (mean of the ratio defined in (40)).
Each experiment is repeated 200 times to calculate the mean FDR, the mean power, and the
expectation. The design matrix X and the pseudo knockoff matrices X̃ are fixed over these trials.
80
80
80
60
60
60
40
40
40
20
20
20
0
0
0
1
1
1
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.2
0.15
0.15
0.15
0.1
0.1
0.1
0.05
0.05
0.05
10
40
70
100 2.8
Sparsity k
3.1
3.4
3.7
4
Orthogonal
Banded
Block Diagonal
0
0.2
Signal amplitude A
0.4
0.6
Correlation ρ
0.8
FDR
Expectation
Power(%)
Figure 1: Testing the orthogonal, the banded, and the block diagonal pseudo knockoff+ at a nominal
FDR q = 20% by varying sparsity, the signal amplitude, or the feature correlation.
80
80
80
60
60
60
40
40
40
20
20
20
0
0
0
1
1
1
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.2
0.15
0.15
0.15
0.1
0.1
0.1
0.05
0.05
0.05
100
200
300
400
500
Number of feature p
600
0
0.2
0.4
0.6
0.8
Within−group correlation ρ
Orthogonal
Banded
Block Diagonal
0
0.2
0.4
0.6
0.8
Between−group correlation γ
Figure 2: Testing the orthogonal, the banded, and the block diagonal pseudo knockoff+ at a
nominal FDR q = 20% by varying one of the following: the number of features p, the within-group
correlation, and the between-group correlation.
The dotted line in Figure 1 and Figure 2 represents the prescribed FDR q or constant 1 as a
14
reference. In all figures, we observe that the FDR is controlled by the nominal level q = 20%. From
the results of the expectation, we observe that all of them are close to or less than 1, which can
be explained by Theorem 2.2, though the definition of this expectation (averaged ratio defined in
(40)) is different from that in (34). Meanwhile, the orthogonal pseudo knockoff+ offers more power
than the banded and the block diagonal pseudo knockoff filters.
We have also applied the least squares statistic in these experiments. For the banded and the
block diagonal pseudo knockoff+, we observe that the least squares statistic controls FDR. For the
orthogonal pseudo knockoff, if ξi are strongly correlated (covariance matrix is 2(X T X)−1 ), the least
squares statistic is not robust and cannot control FDR. By adding a half penalized term in (19),
the performance of the half Lasso statistic is more robust than that of the least squares statistic.
We will provide some partial analysis of the orthogonal pseudo knockoff later on.
3.2
The pseudo knockoff filter in some correlated scenarios
In the correlated case, the diagonal matrix diag(s) in the original knockoff filter constrained by
(2) is small and leads to poor performance of the knockoff filter. A main advantage of the pseudo
knockoff filter is that it relaxes the constraint of X̃ in (15). In some correlated scenarios with
a special structure, we can construct the pseudo knockoff matrix adapting to such structure and
improve the power. To illustrate the effectiveness of the pseudo knockoff filter+, we perform a
comparison between the knockoff filter and the pseudo knockoff filter.
Methods We use the representative statistics that give the most robust performance in both the
knockoff filter and the pseudo knockoff filter based on our numerical experiments for the examples
that we consider. Specifically, we compare the knockoff filter with the OMP statistic with the
pseudo knockoff with the half Lasso statistic (λ = 0.75).
FDR
Power(%)
Group Structure We consider a design matrix X ∈ R1500×500 with a group structure. In
particular, we consider experiment (e) in Section 3.1. The within-group correlation factor ρ varies
from 0.5, 0.55, 0.6, ..., 0.95 and the between-group correlation factor is γ = 0. We use a slightly
larger signal amplitude A = 5 and consider two sparsity cases: k = 30 and k = 100. In all
other settings, we use the default values. By taking advantage of the a priori knowledge of the
correlation structure of X, we use the block diagonal pseudo knockoff filter. We choose m = 5 in
our construction of the block diagonal pseudo knockoff matrix. We also compare the orthogonal
pseudo knockoff using the half Lasso statistic with the knockoff filter using the OMP statistic.
100
100
80
80
60
60
40
40
20
20
0
0.2
0
0.2
0.1
0.1
0
0.5
0.6
0.7
0.8
0
0.5
0.9
Sparsity k=30, Within−group Corr ρ
Orthogonal
Block Diagonal
Knockoff OMP
0.6
0.7
0.8
0.9
Sparsity k=100, Within−group Corr ρ
Figure 3: Comparing the orthogonal, the block diagonal pseudo knockoff filter and the knockoff
filter with the OMP statistic at nominal FDR q = 20% by varying the within-group correlation.
In both figures, the pseudo knockoff filter controls FDR successfully. In the very sparse case
15
(left sub figure), the block diagonal pseudo knockoff filter is not as robust as the knockoff filter.
Based on our analysis of the least squares statistic, we know that the half Lasso statistic favors
the non-sparse case. We observe that it offers more power than the knockoff filter with the OMP
statistic in the right figure where the sparsity is k = 100.
FDR
Power(%)
Decaying Structure We consider a design matrix X ∈ R1500×500 with some decaying structure.
Specifically, the design matrix X is generated from N (0, Σ) with Σij = ρ|i−j| , where ρ varies from
0.5, 0.55, ..., 0.95. We use a slightly larger signal amplitude A = 5 and consider two sparsity cases:
k = 30 and k = 100. Other settings use the default values. We know a priori that the off-diagonal
elements of Σ−1 decay rapidly, i.e. |(Σ−1 )i,j | → 0 as |i − j| increases. Thus, it makes sense to apply
the banded pseudo knockoff filter. In addition, we apply the orthogonal pseudo knockoff and the
knockoff with the OMP statistic for comparison.
100
100
80
80
60
60
40
40
20
20
0
0.2
0
0.2
0.1
0.1
0
0.5
0.6
0.7
0.8
Sparsity k=30, Correlation ρ
0
0.5
0.9
Orthogonal
Banded
Knockoff OMP
0.6
0.7
0.8
Sparsity k=100, Correlation ρ
0.9
Figure 4: Comparing the orthogonal, the banded pseudo knockoff filter and the knockoff filter with
the OMP statistic at a nominal FDR q = 20% by varying the within-group correlation.
From Figure 4, the pseudo knockoff filters control FDR in both cases. In the non-sparse case,
the two pseudo knockoff filters offer more power than the knockoff filter with the OMP statistic.
In these correlated examples with a special correlation structure, we see that designing a special
pseudo knockoff that adapts to the problem could increase the power of the pseudo knockoff filter.
Moreover, we observe that the orthogonal pseudo knockoff filter offers more power than the other
pseudo knockoff filters in each experiment.
4
Partial analysis of the orthogonal pseudo knockoff
From the previous numerical results, we observe that the orthogonal pseudo knockoff offers more
power than other pseudo knockoff filters and still retains robust FDR control. In this section, we
perform further numerical experiments to gain some insight and provide some partial analysis.
4.1
Approximate monotonicity of the expectation of (34)
The FDR control of the banded or the block diagonal pseudo knockoff relies on Theorem 2.2 and
the expectation inequality (34). This result is based on the fact that ξS0 can be classified into
several groups such that the random variables in the same group are mutually independent. For
the orthogonal pseudo knockoff, we have not made use of any special correlation structure of the
design matrix. Thus, we cannot expect that (34) or a similar result holds.
We denote the expectation of the key ratio quantity in (34) with m = 1 as a function of t:
#{j : Wj ≥ t, βj = 0}
.
F (t) , E
#{j : Wj ≤ −t, βj = 0} + 1
16
Inspired by the supermartingale argument in [3], we investigate whether F (t) decreases monotonically with respect to t. We use settings (a), (c), (e) and (f) in Section 3.1, respectively, to
study numerically the behavior of F (t) for a range of t (averaged over 200 trials). Note that these
settings correspond to different correlation structures of X. We apply the half Lasso signed max
statistic (λ = 0.75) and plot the heat map of the expectation. The x-axis corresponds to the range
of t varying from 0, 0.2, 0.4, ..., 5 and the y-axis corresponds to the values of the sparsity or the
correlation.
Correlation ρ
Sparsity k
100
80
60
40
20
0.4
0.8
1.2
1.6
2
2.4
2.8
3.2
3.6
4
4.4
0.9
0.7
0.5
0.3
0.1
0
0.4
0.8
1.2
1.6
2
2.4
2.8
3.2
3.6
4
4.4
1
0.7
0.8
0.5
0.6
0.3
0.4
0.1
0.2
4.8
0
Bet−group Corr γ
Wit−group Corr ρ
0
0.9
0.4
0.8
1.2
1.6
2
2.4
2.8
3.2
3.6
4
4.4
4.8
0.9
0
1
0.7
0.8
0.5
0.6
0.3
0.4
0.1
0.2
4.8
0
0.4
0.8
1.2
Threshold t
1.6
2
2.4
2.8
3.2
3.6
4
4.4
4.8
Threshold t
Figure 5: Computing the variance of F (t) in different scenarios using the orthogonal pseudo knockoff
filter. We vary one of the following default parameters in each sub figure: sparsity k, the correlation
factor ρ (Σij = ρ|j−j|), the within-group correlation factor ρ and the between-group correlation γ.
In each experiment, F (t) varies slightly for small t and decreases almost monotonically for large
t . Based on these numerical observations, we focus on F (0+) in order to gain some insight on the
behavior of F (t).
4.2
Property of the Orthogonal Pseudo Knockoff
Symmetry Property (23) Since X T X̃ = 0 is symmetric, the symmetry property Proposition
2.1 holds for the orthogonal pseudo knockoff.
The identical distribution property From (17) and (18), the covariance matrix of η and ξ
are A = 4[(X + X̃)T (X + X̃)]−1 and B = 4[(X − X̃)T (X − X̃)]−1 , respectively. The orthogonality
condition implies
A = 2(X T X)−1 = 2Σ−1 = B.
Note that E(η) = β = E(ξ) and that η and ξ are independent multivariate normal distributions. η
and ξ are independent and identically distributed.
Control of the ratio We introduce the following notations, which will be used frequently later.
Σ = X T X,
D = diag(Σ−1 ) = diag(d1 , d2 , .., dp ),
Z1 , #{j ∈ S0 : Wj > 0},
g
−1 = D −1/2 Σ−1 D −1/2 ,
Σ
Z2 , #{j ∈ S0 : Wj < 0}.
g
−1 ) = 1. We have the following results.
By definition, we have (Σ
ii
17
(41)
Theorem 4.1. For any δ ∈ (0, 1), the orthogonal pseudo knockoff with the statistic defined in (22)
satisfies
g
−1 )
1+δ
(1 + 3π)λmax (Σ
Z1
≥
E|Nη |−1 ,
(42)
≤
P
2
Z2
1−δ
πδ
1
Z1
E
1 Z1 ≤ 1+δ ≤
,
(43)
Z2 Z2 1−δ
1−δ
g
g
−1 ) denotes the largest eigenvalue of Σ
−1 . In
where Nη , #{j ∈ S0 : Wj 6= 0} and λmax (Σ
particular, if Wi 6= 0 ∀i ∈ S0 almost surely, we have
Z1
1+δ
1
1
g
g
−1
−1
−2
−1
+ 3 min λmax (Σ ), 4(si di )
(44)
≥
min λmax (Σ ), 2(si di )
P
≤ 2
Z2
1−δ
δ |S0 | π
2
−1
g
g
−1 )/(δ 2 |S |)) = O(λ
−1
which is of order O(λmax (Σ
0
max (Σ )/(δ p)). Here, di = (Σ )ii , si is the
P
−α
−1
, α ∈ {1, 2}.
knockoff factor in (2) and (si di )−α , |S0 |
j∈S0 (si di )
Remark 3. (42) controls the ratio Z1 /Z2 in a probabilistic sense. By the definition of F (t), F (0+) =
g
−1 are all 1.
E(Z1 /(Z2 + 1)) and is bounded by E(Z1 /Z2 ). Note that the diagonal elements of Σ
P
p
g
g
g
g
−1 ) =
−1
−1
−1
From p = T r(Σ
i=1 λi (Σ ) and λi (Σ ) > 0, we have λmax (Σ ) < p. We will demonstrate
g
−1 is well-conditioned and the number of null features |S | is large enough,
later that if the matrix Σ
0
the ratio Z1 /Z2 is bounded by a constant close to 1 with high probability.
We would like to draw a connection between the probabilistic upper bound (44) and the knockoff
filter. Since the amplitude of si is associated with the power of the knockoff filter, the right hand
side relates to the powerof the knockoff.
Based on Σii = 1 and 0 ≺ Σ, one can prove di ≥ 1. To
1 vT
, v ∈ R(p−1) , 0 ≺ E ∈ R(p−1)×(p−1) . The positive definiteness of
see this, we denote Σ =
v E
Σ guarantees 0 < 1 − v T E−1 v ≤ 1. It follows d1 = (1 − v T E−1 v)−1 ≥ 1. Similarly, we can prove
di ≥ 1. If si is not small and is bounded from below (e.g. using the modified SDP construction
in [6]), then the quantity s−1
cannot be too large. We can obtain a good control of Z1 /Z2 in the
i
probabilistic sense if |S0 | is large enough.
4.3
The Mean-Variance Argument
From the sign property of W , we know sign(WS0 ) = sign((f (η))S0 ) · sign(ξS0 ). Denote Yi = 1Wi >0 .
In order to estimate the variance of Z1 , Z2 , we first analyze the covariance of each pair (Yi , Yj ), i, j ∈
S0 .
Lemma 4.2. Conditional on η, for any null variable i, j, we have
3 g
1 g
−1 )2 .
(Σ−1 )ij (1(f (η))i >0 − 1(f (η))i <0 )(1(f (η))j >0 − 1(f (η))j <0 ) + (Σ
Cov(Yi , Yj |η) ≤
ij
2π
2
e −1 )ij is defined in (41).
where (Σ
(45)
We will defer the proof to the Appendix.
The Mean and Variance of Zi Denote Nη , {j ∈ S0 : (f (η))j 6= 0}. Obviously, Nη is
measurable with respect to the σ-field generated by η. Since ξ = β̂ ls − β̃ ls 6= 0 almost surely, we
have Wj = 0 ⇐⇒ f (η)j = 0 almost surely and thus Nη = {j ∈ S0 : Wj 6= 0} almost surely.
The means of Yi and Zj are straight forward. For i ∈ S0 , we have
E(Yi |η) = 1fi (η)>0 E(1ξi >0 |η) + 1fi (η)<0 E(1ξi <0 |η) = 1fi (η)>0 E(1ξi >0 ) + 1fi (η)<0 E(1ξi <0 ),
1
1
= (1fi (η)>0 + 1fi (η)<0 ) =
(46)
2
2
|Nη |
, Z1 + Z2 = |Nη |,
E(Z1 |η) = E(Z2 |η) =
2
18
where we use the fact that η and ξ are independent. Next, we show that
3
1
g
g
−1
−1
−1
−2
min λmax (Σ ), 2(si di )
V ar(Z1 |η) ≤ |Nη |
+ min λmax (Σ ), 4(si di )
.
(47)
2π
2
P
Here (si di )−α = |Nη |−1 i∈Nη (si di )−α . Denote wi , 1fi (η)>0 − 1fi (η)<0 . Using (45), we have
V ar(Z1 |η) ≤
X
i,j∈Nη
T (Σ
g
−1 )
g
−1 )
X (Σ
wN
Nη Nη wNη
3 g
3
ij
η
2
g
−1
−1 )2
Cov(Yi , Yj |η) ≤
wi wj + (Σ )ij =
+ T r((Σ
Nη Nη )
2π
2
2π
2
i,j∈Nη
(48)
−1/2
−1
−1/2
g
g
−1
−1
From diag(s)
(2) and the definition of Σ (41), we have Σ 2D
(diag(s)) D
.
g
g
−1
−1
Meanwhile, Σ λmax (Σ )I. It follows
T
T
T
−1/2
−1 −1/2
g
g
−1 )
−1 )I
wN
(
Σ
w
≤
min
w
λ
(
Σ
w
,
2w
(D
(diag(s))
D
)
w
Nη Nη Nη
Nη N η Nη
|Nη | Nη
Nη max
Nη
η
g
−1 ), 2(s d )−1
(49)
= |Nη | min λmax (Σ
i i
2X T X
g
−1 )I, D−1/2 (diag(s))−1 D−1/2 are diagonal and |w | = 1. (s d )−1
Here, we use the fact that λmax (Σ
i i
i
is the average of (si di )−1 over i ∈ Nη . For any positive definite matrix 0 ≺ P Q, we have
T r(Q2 − P2 ) = T r((Q − P)(Q + P)) = T r((Q − P)1/2 (Q − P)1/2 (Q + P)1/2 (Q + P)1/2 )
=T r((Q − P)1/2 (Q + P)1/2 (Q + P)1/2 (Q − P)1/2 ) = T r(RRT ) ≥ 0, R = (Q − P)1/2 (Q + P)1/2
−1/2 (diag(s))−1 D−1/2 )
g
−1 )
Taking P = (Σ
Nη Nη , Q = 2(D
Nη Nη , we yield
−1/2
g
−1 )2
T r((Σ
(diag(s))−1 D−1/2 )2Nη Nη ) = 4|Nη |(si di )−2
Nη Nη ) ≤ 4T r((D
P 2
g
−1 ) = 1. We have
Note that for any 0 A, T r(A2 ) =
λi (A) ≤ λmax (A)T r(A) and (Σ
ii
g
g
g
g
−1
−1
−1 )2
−1
T r((Σ
Nη Nη ) ≤ λmax (Σ )T r((Σ )Nη Nη ) = |Nη |λmax (Σ ).
(50)
(51)
Combining (48), (49),(50) and (51) yields (47).
Proof of Theorem 4.1
Proof. Conditional on η, we apply E(Z1 |η) = E(Z2 |η) = |Nη |/2 = (Z1 + Z2 )/2 (46), (47), and the
Chebyshev inequality to yield
1
1 V ar(Z2 η)
P (Z2 ≤ (1 − δ)EZ2 η) = P (Z2 ≥ (1 + δ)EZ2 η) = P (|Z2 − EZ2 | ≥ δEZ2 η) ≤ 2
2
2δ (E(Z2 η))2
1
1
g
g
−1 ), 2(s d )−1 + 3 min λ
−1 ), 4(s d )−2
min λmax (Σ
≤ 2
(
Σ
max
i i
i i
δ |Nη | π
g
−1 )
(1 + 3π)λmax (Σ
(52)
≤
πδ2 |Nη |
P
Here (si di )−α = |Nη |−1 i∈Nη (si di )−α , α = 1, 2. The first identity holds since the symmetry
d
property (23) implies that Z2 = Z1 = |Nη | − Z2 . The estimate (42) follows from integrating the
last inequality in (52). If Wi 6= 0 ∀i ∈ S0 almost surely, we have Nη = S0 almost surely. It follows
P
|Nη | = |S0 | and (si di )−α = |S0 |−1 i∈S0 (si di )−α almost surely. Consequently, we can integrate the
second inequality in (52) to yield (44).
Next, we prove (43). Recall that conditional on η, Z1 + Z2 = |Nη | , nη . We know Z1 ≤
nη
1+δ
obtain
2 ⇐⇒ Z2 ≥ (1− δ) 2. Consequently, we
1−δ Z
2E(Z1 1Z2 ≥(1−δ)nη /2 |η)
Z1
Z1
2E(Z1 |η)
1
1 Z1 ≤ 1+δ η = E
1Z2 ≥(1−δ)nη /2 η ≤
≤
=
E
Z2 Z2 1−δ
Z2
(1 − δ)nη
(1 − δ)nη
1−δ
Integrating both sides yields the desired result (43).
19
4.4
Some Special Design Matrices
For some special design matrices, we can improve the estimate of V ar(Z1 ) in (47) and get better
control of Z1 /Z2 . In our simulations, we observe that the orthogonal pseudo knockoff filter offers
robust FDR control and works better than other pseudo knockoff filters and the knockoff filter
with the OMP statistic in these scenarios. We would like to offer a partial explanation of this
phenomenon.
A diagonally dominated case Let X ∈ Rn×p and Σ = X T X. We consider several classes of
design matrices described below.
(a) For any i 6= j, hXi , Xj i , XiT Xj = ρ, ρ ∈ [0, 1).
(b) Assume that X can be clustered into k groups, X = (XC1 , XC2 , ..., XCk ). The within-group
correlation of group i is ρi for some ρi ∈ [0, 1) and the between-group correlation is zero.
(c) The sizes of different groups are equal. The within-group correlation is ρ and the betweengroup correlation is γ · ρ.
Case (b) and (c) correspond to setting (e) and (f) in Section 3.1. Denote E , Σ−1 for convenience.
g
−1 ) = E /(E1/2 E1/2 ). For the design matrices described above, we can show that
From (41), (Σ
ij
ij
ii
jj
P
−1
−1
Σ−1 is diagonally dominated, i.e.
j6=i |(Σ )ij | < Σii .The proof is a bit technical and tedious.
We will omit the proof here. From Lemma 4.2, we have
Cov(Yi , Yj |η) ≤
3 g
1 g
g
−1 )2 ≤ c |(Σ
−1 ) |,
(Σ−1 )ij wi wj + (Σ
0
ij
ij
2π
2
c0 =
1
3
+ <2
2π 2
(53)
Using the fact that Σ−1 is diagonally dominated, we can improve the estimate of V ar(Z1 |η) in (47)
1/2
1/2
X
X
X
X
|E
|
|E
|
|E
|
ij
ij
ij
g
−1 ) | = c
c0 |(Σ
V ar(Z1 η) ≤
≤ c0
ij
0
1/2 1/2
E
E
ii
jj
i,j∈Nη Eii Ejj
i,j∈Nη
i,j∈Nη
i,j∈Nη
1/2
1/2
1/2
1/2
X 1 X
X 2Eii
X 2Ejj
X 1 X
= 2c0 |Nη |.
|Eij |
|Eij | ≤ c0
=c0
Eii
Ejj
Eii
Ejj
i∈Nη
j∈Nη
j∈Nη
i∈Nη
i∈Nη
j∈Nη
P
Here, we have used Eij = Eji and the diagonal dominated assumption to yield
j∈Nη |Eij | ≤
Pp
|E
|
≤
2E
.
Accordingly,
the
estimate
of
Z
/Z
in
Theorem
4.1
can
be
improved.
ij
ii
1
2
j=1
Proposition 4.3. Assume that Σ−1 = (X T X)−1 is diagonally dominated, the orthogonal pseudo
knockoff satisfies
i
Z1
1+δ
2 + 6π h
−1
P
,
≥
E
(#{j
∈
S
:
W
=
6
0})
≤
0
j
Z2
1−δ
πδ2
If Wi 6= 0 ∀i ∈ S0 almost surely, the upper bound becomes (2 + 6π)/(πδ2 |S0 |) = O((δ2 p)−1 ).
Exponentially Decaying Class Assume that |(Σ−1 )ij | ≤ Cρ|i−j| for ρ ∈ [0, 1) and some constant C. The design matrix in setting (c) in Section 3.1 has a similar structure. One can prove
that (Σ−1 )ii ≥ 1 using the fact that Σii = 1 and Σ is positive definite. By our assumption, we have
g
g
g
−1 ) | ≤ |(Σ−1 ) | ≤ Cρ|i−j| . Hence, we have λ
−1
−1
|(Σ
ij
ij
max (Σ ) ≤ ||Σ ||l1 ≤ 2C/(1 − ρ). Using (47)
and Theorem 4.1, we yield
2c0 C|Nη |
Z1
4c0 C
1+δ
4c0 C
g
−1 )|N | ≤
,
P
E|Nη |−1 ≈ 2
.
≥
≤ 2
V ar(Z1 η) ≤ c0 λmax (Σ
η
1−ρ
Z2
1−δ
δ (1 − ρ)
δ (1 − ρ)|S0 |
where c0 = (1+3π)/(2π). If 1−ρ ≥ c1 for some positive constant c1 not too small, this probabilistic
upper bound is of the order O((δ2 |S0 |)−1 ) = O((δ2 p)−1 ).
20
Summary Let Fkf be the σ- field generated by all |Wi | in the knockoff filter, B = 4[(X −
X̃)T (X − X̃)], N0 , #{i ∈ S0 : |Wi | =
6 0}, m be the maximal group size or the band width, and
Z1 (t) , #{i ∈ S0 : Wi ≥ t}, Z2 (t) , #{i ∈ S0 : Wi ≤ −t} ∀t > 0. In particular, Z1 (0) , #{i ∈
S0 : Wi > 0}, Z2 (0) , #{i ∈ S0 : Wi < 0}. We summarize the variance estimates for several classes
of pseudo knockoff filters and the knockoff filter as follows. We note that X̃ in all filters satisfies
X T X = X̃ T X̃, and B Σ−1 being diagonal is equivalent to the second constraint in (1) and (2).
Filter
Knockoff
Block
Diagonal
Banded
Orthogonal
Constraint on X̃
B Σ−1
being diagonal
B Σ−1 being
block diagonal
B Σ−1
being banded
B = 2Σ−1 or
X T X̃ = 0
FDR Control
Exact. For adaptive T ,
E(Z1 (T ))/(Z2 (T ) + 1) ≤ 1
Partial. For fixed t > 0,
E(Z1 (t)/(Z2 (t)) + m) ≤ 1
Partial. For fixed t > 0,
E(Z1 (t)/(Z2 (t)) + m) ≤ 1
P (Z1 (0)/Z2 (0) ≥ (1 + δ)/(1 − δ))
g
−1 )/|N |)
≤ 2c0 (δ)−2 E(λmax (Σ
0
Conditional Variance
V ar(Z1 (0)|Fkf ) = |N0 |/4
V ar(Z1 (0)|η) ≤ |N0 |m/4
V ar(Z1 (0)|η) ≤ |N0 |m/4
V ar(Z1 (0)|η)
g
−1 )
≤ c0 |N0 |λmax (Σ
We focus on the conditional variance of the pseudo knockoff. The estimate of V ar(Z1 (0)|η)
of the block diagonal or the banded pseudo knockoff is given by (39) and V ar(Z1 (0)|F) of the
knockoff is discussed in the paragraph after (39). For small m, V ar(Z1 |η) of the block diagonal
or the banded pseudo knockoff is of the same order as that of the knockoff. For the orthogonal
g
−1 ) and c = (1 + 3π)/(2π).
pseudo knockoff, V ar(Z1 (0)|η) in the above table is c0 |N0 |λmax (Σ
0
g
g
−1 is well-conditioned, e.g. Σ
−1 is diagonally dominated or X is one of the design matrices
If Σ
g
−1 ) is small and V ar(Z (0)|η) is O(|N |). Note that
we discussed in this section, then λmax (Σ
1
0
E(Z1 |η) = E(Z2 |η) = |N0 |/2. If V ar(Z1 (0)|η) is bounded by C|N0 |α for α < 2 and some universal
constant C, the mean-variance argument provides a relatively tight upper bound (that is close to 1)
for Z1 (0)/Z2 (0) as |N0 | tends to infinity. This preliminary analysis offers some insight as to why the
pseudo knockoff filter may offer FDR control similar to that of the knockoff filter in the examples
that we consider here. Based on the mean-variance argument, it is reasonable to conjecture that
if V ar(Z1 (0)|η) ≤ C|N0 | for some universal constant C, then the pseudo knockoff may have FDR
control similar to that of the knockoff.
4.5
An extreme example
The banded or the block diagonal pseudo knockoff filters are constructed for those design matrices
with some special correlation structure, i.e. either the off-diagonal elements of (X T X)−1 decay
rapidly or (X T X)−1 can be approximated by a block diagonal matrix or a banded matrix. In this
subsection, we consider a class of design matrices that violate these assumptions. Specifically, the
covariance matrix M to generate X ∈ R1500×500 (X ∼ N (0, M)) satisfies (M−1 )ij = ρ ∀i 6= j and
(M−1 )ii = 1 for some ρ ≥ 0. We need to normalize the columns of X to be a unit vector. We
focus on two sparsity levels with k = 30 and k = 100, and choose the signal amplitude to be A = 5
i.i.d
(βi ∼ {±A}). We let ρ vary from 0, 0.1, 0.2, ..., 0.9. We construct the pseudo knockoff matrix
according to Section 2.3 with m = 2 (m is the largest block size or the band width).
In Figure 6, we can see that the orthogonal pseudo knockoff with the half Lasso statistic has
FDR control and considerable power for varying ρ, while the knockoff filter with the OMP statistic
has almost no power. The banded and the block diagonal pseudo knockoff filters with the half
Lasso statistic are also considered. They control FDR but lose a lot of power for ρ > 0. Since the
off-diagonal elements of Σ−1 do not decay and Σ−1 is not block diagonal or banded, there is no
advantage of using these pseudo knockoff filters in this extreme example.
To understand why the knockoff filter with the OMP or the Lasso Path statistics performs
21
Power(%)
FDR
100
100
80
80
60
60
40
40
20
20
0
1
0
1
0.5
0.5
0
0
0.2
0.4
0.6
Sparsity k=30, Correlation ρ
0
0.8
Orthogonal half Lasso
Orthogonal LS
Knockoff OMP
Knockoff Lasso Path
0
0.2
0.4
0.6
Sparsity k=100, Correlation ρ
0.8
Figure 6: Comparing the orthogonal pseudo knockoff filter with the half Lasso and least squares
statistics with the knockoff filter with the OMP and the Lasso path statistics at a nominal FDR
q = 20% by varying the correlation ρ. The figures show the mean power and the mean FDR
averaged over 200 trials.
poorly in this example, we analyze the correlation between Xj and X̃j . Based on the second
constraint of the knockoff filter (2), we use an argument similar to (49). More specifically, we
Pp
1
g
−1 v ≤ 2vT D−1/2 (diag(s))−1 D−1/2 v with v = (1, 1, .., 1)T ∈ RP . This gives
have vT Σ
i=1 si di ≥
P (Σg
−1 )
ij
g
g
−1 is defined in (41). By the construction of X, we have (Σ
−1 ) = 1 and
, where Σ
ii
i,j
2
−1
g
−1
(Σ )ij ≈ ρ, ∀i 6= j. Note that di ≥ 1. The lower bound of mean(s ) is approximately equal
i
to 12 (1 + (p − 1)ρ) ≈ pρ/2. If si generated by the SDP construction distributes evenly, we obtain
ej are strongly
si = O((pρ)−1 ). For ρ 6= 0 and p large enough, si is very small and thus Xj and X
e
correlated. We have computed the mean correlation of Xj , Xj numerically (the knockoff matrix
is generated via the SDP construction). For the mildest case of ρ = 0.1, the mean correlations
ej for the banded, the block diagonal pseudo knockoff filters, the knockoff filter, and the
of Xj , X
orthogonal pseudo knockoff filter are 0.976, 0.981, 0.971, and 0, respectively. The strong correlation
ej explains the weak power of the block diagonal and the banded pseudo knockoff
between Xj , X
filters and the knockoff filter with the OMP statistic. The orthogonal pseudo knockoff filter still
ej to be orthogonal by construction.
has a reasonable amount of power since we force Xj , X
From the FDR result, we see that the Lasso path statistic is not robust. Note that the Lasso
path is not computed exactly, but approximated by a fine grid of λ. Typically, the number of grid
points to approximate λ is chosen to be 5p. Since Xj and X̃j are strongly correlated, this level of
grid resolution for λ may not be fine enough to separate the entrance time of Xj and X̃j . From
our simulations, we observe that Xj and X̃j enter into the model at the same λ, which leads to
Wj = 0. In order to determine which one enters the model earlier, we need to refine the grid.
However, refining the grid resolution leads to a considerable increase in the computational cost of
the Lasso path statistic. To show that the FDR of the Lasso Path statistic is indeed under control
as predicted by the knockoff theory in [3], we reduce the sample size to p = 100, n = 300 (sparsity
k = 20 and k = 6), and increase the number of realizations to 1000 and implement the Lasso path
statistic with the number of grid points for λ equal to 50p. Using this level of resolution, we observe
that the mean FDR is controlled at q = 20%. However, the statistical power of the Lasso Path
statistic is also greatly reduced and is less than 10% for large values of ρ (ρ ≥ 0.5). Based on the
performance of these two knockoff statistics, we find that the OMP statistic is more robust.
Variance of the orthogonal pseudo knockoff filter with different statistics Although
the orthogonal pseudo knockoff filter with the least squares statistic maintains a lot of power, its
22
mean FDR is more than the prescribed q = 20% (i.e. we lose the FDR control for the least squares
g
−1 ) is of order 1 + (p − 1)ρ ≈ pρ. Therefore, the
statistic). Based on our construction of X, λmax (Σ
2
estimate of V ar(Z1 |η) in (47) is O(p ) if Nη = S0 and |S0 | ≈ p. Consequently, the upper bound
in Theorem 4.1 is O(δ−2 ) and we may lose control of Z1 /Z2 for some statistics. To understand
the difference between the mean FDR of the half Lasso and the least squares statistic, we compute
V ar(Z1 ) numerically over 200 trials. The results are plotted in the figure below.
Variance of the half Lasso
Variance of least squares
0.16
ρ = 0.1
ρ = 0.3
ρ = 0.5
ρ = 0.7
ρ = 0.9
0.12
0.08
2
0.07
Var(Z1) / p
Var(Z1) / p
0.14
0.1
0.06
0.05
0.04
0.03
0.02
0.08
0.01
0.06
100
200
300
400
500
600
700
800
Number of features p
0
100
200
300
400
500
600
700
800
Number of features p
Figure 7: Numerical computation of the variance of the orthogonal pseudo knockoff filter with
different statistics by varying the sample size (n, p) = (3p, p). Here the sparsity k = 0.2p. Different
curves show the variance of Z1 in different correlation structures.
We focus on the order of V ar(Z1 ). The left figure shows that V ar(Z1 ) of the half Lasso statistic
is O(p) and the constant is about 0.1, while the right figure shows that V ar(Z1 ) of the least squares
statistic is O(p2 ) and the constant varies for different ρ. The results of ρ = 0, 0.2, 0.4, 0.6, 0.8 are
qualitatively similar. The variance of the least squares obtained from our simulation matches well
with our theoretical estimate and thus its FDR may lose control in the extreme example (Figure 6).
The numerical variance of the half Lasso statistic, however, is order O(p), which is much better than
O(p2 ) in the estimate (47). Therefore, the mean-variance argument shows that the upper bound in
(42) is O((δ2 p)−1 ), which explains why we still have FDR control of the orthogonal pseudo knockoff
filter with the half Lasso signed max statistic in this extreme example.
From the numerical results in this subsection and Section 3, we see that a half penalty in
(19), e.g. the l1 penalty λ|| · ||1 , reduces the variance of Z1 significantly, which leads to robust
performance of the orthogonal pseudo knockoff filter. We do not fully understand why a half Lasso
penalty reduces the variance from O(p2 ) to O(p) in the extreme example. From our numerical
experiments, we observe that the combination of the orthogonal pseudo knockoff and the half
Lasso signed max statistic seems to offer the most power among other pseudo knockoff filters. It
would be worthwhile to further study the effect of a half penalty to take full advantage of the
orthogonal pseudo knockoff filter.
5
Concluding remarks
In this paper, we introduce three classes of pseudo knockoff filters. These pseudo knockoff filters
preserve some essential features of the original knockoff filter but offer more flexibility in constructing the knockoff matrix, especially when the features are highly correlated or when the inverse of
the covariance matrix, Σ−1 , decays very slowly away from its main diagonal. We provide a partial analysis of the pseudo knockoff filters by investigating the expectation inequality (9), which
is a variant of the corresponding expectation inequality (10) for the knockoff filter. We establish
this expectation inequality for the block diagonal and the banded pseudo knockoff filters. For the
orthogonal pseudo knockoff filter, we provide a probabilistic analysis based on the mean-variance
argument. This analysis gives an upper bound on the the expectation of interest. In the case when
23
Σ−1 is diagonally dominated or when Σ−1 has some special structure, this probabilistic analysis
provides a relatively tight upper bound for the expectation of interest. This analysis offers some
partial explanation why the pseudo knockoff filters provide FDR control in the examples that we
consider in this paper. We have also performed a number of numerical experiments to test the
performance of the pseudo knockoff filters and compare them with the knockoff filter. In the case
when the features are highly correlated and the non-sparse case (e.g. about 20% non-null features),
the orthogonal pseudo knockoff filter offers more power than the knockoff filter with the OMP or
Lasso Path statistic. In the extreme case when Σ−1 decays very slowly away from its main diagonal, the knockoff filter with the OMP or the Lasso Path statistic tends to lose considerable power.
The orthogonal pseudo knockoff filter still offers reasonably high power in this case. The study
presented in this paper is still at a very preliminary stage and further study is required to fully
understand the advantages and the limitations of the pseudo knockoff filters.
Appendix
Verification of the construction (31) of X̃.
tions show that
Let X̃ be constructed from (31). Direct calcula-
(X − X̃)T (X − X̃) = [(2XΣ−1 − 2UC)B−1 ]T [(2XΣ−1 − 2UC)B−1 ]
=4B−1 (Σ−1 X T XΣ−1 + CT UT UC)B−1 = 4B−1 (Σ−1 + CT C)B−1 = 4B−1 .
(X + X̃)T (X − X̃) = [X(2I − 2Σ−1 B−1 ) + 2UCB−1 ]T [2XΣ−1 B−1 − 2UCB−1 ]
=4(I − Σ−1 B−1 )T X T XΣ−1 B−1 − 4B−1 CT UT UCB−1
=4(I − B−1 Σ−1 )B−1 − 4B−1 C T CB−1 = 4(I − B−1 Σ−1 )B−1 − 4B−1 (B − Σ−1 )B−1 = 0 .
Here, we use UT X = X T U = 0. The first identity implies (29) and the second is exactly the
orthogonal condition (16).
Proof of Lemma 4.2
Conditional on η, we can determine Nη = {j ∈ S0 : Wj 6= 0}. Recall that
d
ξ and η are independent and ξS0 = −ξS0 . We have E(1ξi >0 |η) = E(1ξi >0 ) = 1/2, i ∈ So and for
any i, j ∈ Nη ,
1
1
1
E(1ξi >0 1ξj <0 ) − + E(1ξi >0 1ξj >0 ) − = E(1ξi >0 ) − = 0.
4
4
2
1
1
Similarly, E(1ξi <0 1ξj >0 ) − 4 = −(E(1ξi >0 1ξj >0 ) − 4 ), ∀i, j ∈ S0 . Meanwhile, the symmetry of ξS0
implies E(1ξi <0 1ξj <0 ) = E(1ξi >0 1ξj >0 ). Therefore, we obtain
Cov(Yi , Yj |η) = E(Yi Yj |η) − E(Yi |η)E(Yj |η) = E(Yi Yj |η) −
1
4
1
=E[(1fi (η)>0 1ξi >0 + 1fi (η)<0 1ξi <0 ) · (1fj (η)>0 1ξj >0 + 1fj (η)<0 1ξj <0 )|η] −
4
1
1
+ 1fi (η)>0 1fj (η)<0 E(1ξi >0 1ξj <0 ) −
=1fi (η)>0 1fj (η)>0 E(1ξi >0 1ξj >0 ) −
4
4
(54)
1
1
+ 1fi (η)<0 1fj (η)>0 E(1ξi <0 1ξj >0 ) −
+ 1fi (η)<0 1fj (η)<0 E(1ξi <0 1ξj <0 ) −
.
4
4
1
=(E(1ξi >0 1ξj >0 ) − )(1fi (η)>0,fj (η)>0 − 1fi (η)>0,fj (η)<0 − 1fi (η)<0,fj (η)>0 + 1fi (η)<0,fj (η)<0 )
4
1
1
=(E(1ξi >0 1ξj >0 ) − )(1fi (η)>0 − 1fi (η)<0 )(1fj (η)>0 − 1fj (η)<0 ) = (E(1ξi >0 1ξj >0 ) − )wi wj ,
4
4
where wi , 1fi (η)>0 − 1fi (η)<0 . By definition, wi = 1 or −1. From Cov(ξ) = B = 2Σ−1 , we
ξi
Bii Bij
∼ N 0,
. Since normalizing
know that the joint distribution of (ξi , ξj ) is
ξj
Bji Bjj
24
1 µij
∼ N 0,
, where µij =
µij 1
g
−1 ) ,
µ = µij = (Σ
ij
(55)
ξi , ξj does not change their sign, we assume that
1/2 1/2
g
−1 ) (see (41)). Define
Bij /(Bii Bjj ) = (Σ
ij
ξi
ξj
and let P (ξi , ξj ) and Ps (·) be the probability distribution function of (ξi , ξj ) and the standard
normal distribution, respectively.
Using 0 ≤ ex − 1 − x ≤
Ps (ξi )Ps (ξj ) up to µ2
x2 x
2 (e 1x>0
+ 1) by taking x , −
µ2 ξi2 +µ2 ξj2 −2µξi ξj
,
2(1−µ2 )
we expand P (ξi , ξj ) −
p
Ps (ξi )Ps (ξj )
p
(1 + x − 1 − µ2 + ex − 1 − x)wi wj
1 − µ2
p
Ps (ξi )Ps (ξj )
≤ p
((1 + x − 1 − µ2 )wi wj + ex − 1 − x)
1 − µ2
p
Ps (ξi )Ps (ξj )
x2 (ex 1x>0 + 1)
)
≤ p
((1 + x − 1 − µ2 )wi wj +
2
1 − µ2
p
Ps (ξi )Ps (ξj )
x2
x2 1x>0
= p
.
(1 + x − 1 − µ2 )wi wj +
+ P (ξi , ξj )
2
2
1 − µ2
[P (ξi , ξj ) − Ps (ξi )Ps (ξj )]wi wj =
Integrating both sides with respect to ξi , ξj in the region ξi , ξj > 0 gives
Z
p
Ps (ξi )Ps (ξj )
x2
1
2
p
wi wj ≤
dξi dξj
(1 + x − 1 − µ )wi wj +
E(1ξi >0 1ξj >0 ) −
4
2
1 − µ2
ξi ,ξj >0
Z
x2 1x>0
P (ξi , ξj )
+
dξi dξj , I + II + III .
(56)
2
ξi >0,ξj >0
Since Ps (·) is standard Gaussian distribution and x = −
moments in I, II explicitly. For I, we have
I=
≤
=
1
µ2
1
µ
1
−
(p
− 1) +
4
2π(1 − µ2 )3/2 4 (1 − µ2 )3/2
1 − µ2
µ2 ξi2 +µ2 ξj2 −2µξi ξj
,
2(1−µ2 )
!
we can calculate all the
wi wj
µ2
µwi wj
µ
1
1
1
+
((1 − µ2 )−3/2 − 1) + ( p
− 1) −
2π
2π
4
4 (1 − µ2 )3/2
1 − µ2
(57)
µ2 (1 − µ2 )−3/2 − 1
1
µ2
1
µwi wj
µwi wj
p
+
−
+ c1 (µ)µ2 ,
+ p
,
2
2
2
2π
2π
µ
2π
4 1−µ 1−µ
1+ 1−µ
where
the coefficients of µ2 and is bounded near µ = 0. We use E(ξ1ξ>0 ) =
√ c1 (µ)2 ≥ 0 collects
1
1/ 2π, E(ξ 1ξ>0 ) = 2 for the standard Gaussian ξ to obtain the first equality, and |wi | = |wj | = 1
to obtain the inequality. For the second term, we get
Z
µ2 (1 − π8 µ + 2µ2 )
µ2
2
2 2
P
(ξ
)P
(ξ
)(2ξ
ξ
−
µ(ξ
+
ξ
))
dξ
dξ
=
II =
, c2 (µ)µ2 ,
s i s j
i j
i j
i
j
8(1 − µ2 )5/2 ξi ,ξj >0
8(1 − µ2 )5/2
(58)
(1− 8 µ+2µ2 )
π
where c2 (µ) = 8(1−µ
2 )5/2 ≥ 0 is bounded near µ = 0. Since ξi , ξj > 0 and x = −
µ ≤ 0 implies x ≤ 0, or equivalently 1x>0 ≤ 1µ>0 . Note that
x=−
µ2 ξi2 +µ2 ξj2 −2µξi ξj
,
2(1−µ2 )
µ2 ξi2 + µ2 ξj2 − 2µξi ξj
2µ2 ξi ξj − 2|µ|ξi ξj
|µ|ξi ξj
≤
−
=
, ∀ ξi , ξj > 0.
2
2
2(1 − µ )
2(1 − µ )
(1 + |µ|)
25
For ξi , ξj > 0, we have x2 1x>0 ≤
|µ|ξi ξj
1+|µ|
2
1µ>0 . Therefore, we obtain
1 |µ2 1µ>0 |
1
E(ξi2 ξj2 1ξi ,ξj >0 )
III = E(x2 1x>0 1ξi ,ξj >0 ) ≤
2
2 (1 + |µ|)2
1 |µ2 1µ>0 |
1 |µ2 1µ>0 | 3
4
4
1/2
≤
(E(ξ
1
)E(ξ
1
))
=
, 1µ>0 c3 (µ)µ2 ,
i ξi >0
j ξj >0
2 (1 + |µ|)2
2 (1 + |µ|)2 2
where c3 (µ) =
3
4(1+|µ|)2
(59)
is bounded near µ = 0. Combining (56), (57), (58) and (59), we yield
1
µ
µ
Cov(Yi , Yj |η) = E(1ξi >0 1ξj >0 ) −
wi wj +(c1 (µ)+c2 (µ)+c3 (µ)1µ>0 )µ2 ,
wi wj +c(µ)µ2 .
wi wj ≤
4
2π
2π
Here, c(µ) = c1 (µ) + c2 (µ) + c3 (µ)1µ>0 . Since ci (µ) is a non-negative and an explicit function
of µ, it is not difficult to show that c(µ) < 23 for |µ| < 12 . For |µ| > 1/2, we use the estimate
Cov(Yi , Yj |η) ≤ 1/4 ≤ −1/2π + 32 µ2 . Finally, we conclude
Cov(Yi , Yj |η) ≤
1
µ
µ
3
wi wj + c(µ)µ2 ∧ ≤
wi wj + µ2 ,
2π
4
2π
2
g
−1 ) is defined in (55). This proves Lemma 4.2.
where µ = (Σ
ij
Acknowledgements. The research was in part supported by NSF Grants DMS 1318377 and DMS
1613861. The research of the first author was performed during his visit to ACM at Caltech. We
would like to thank Professor Emmanuel Candes for his many valuable comments and suggestions
to our work. We would also like to thank Professor Lucas Janson for his interest and comments
on the earlier version of this manuscript and Dr. Pengfei Liu for the insightful discussions on the
pseudo knockoff.
References
[1] Y. Benjamini and Y. Hochberg. Controlling the false discovery rate: a practical and powerful
approach to multiple testing. J. Roy. Statist. Soc. Ser. B, 57(1):289300, 1995.
[2] Y. Benjamini and D. Yekutieli. The control of the false discovery rate in multiple testing under
dependency. Ann. Statist., 29(4):11651188, 2001.
[3] R. Foygel Barber, E. J. Candès. Controlling the false discovery rate via knockoffs. Ann. Statist.,
43(5):20552085, 2015.
[4] R. Foygel Barber, E. J. Candès. A knockoff filter for high-dimensional selective inference.
arXiv:1602.03574v1, 2016.
[5] E. J. Candès, Y. Fan, L. Janson, and J. Lv, Panning for gold: Model-free knockoffs for highdimensional controlled variable selection.arXiv:1610.02351, 2016.
[6] J. Chen, A. Hou, and T. Y. Hou, Some analysis of the knockoff filter and its variants for highly
correlated features, preprint, 2016.
[7] R. Dai and R. Foygel Barber. The knockoff filter for FDR control in group-sparse and multitask
regression. arXiv preprint arXiv:1602.03589,2016.
[8] M. G. G’Sell, S. Wager, A. Chouldechova, and R. Tibshirani. False discovery rate control for
sequential selection procedures, with application to the Lasso. 2013. arXiv:1309.5352.
26
[9] L. Janson and W. Su. Familywise error rate control via knockoffs.arXiv:1505.06549v3, 2015.
[10] H. Liu, K. Roeder, and L. Wasserman. Stability approach to regularization selection (StARS)
for high dimensional graphical models. Adv. Neural Inf. Process. Syst., 23:14321440, 2010.
[11] N. Meinshausen and P. Bühlmann. Stability selection. Journal of the Royal Statistical Society:
Series B (Statistical Methodology), 72(4):417473, 2010.
[12] A. J. Miller. Selection of subsets of regression variables. J. Roy. Statist. Soc. Ser. A,
147(3):389425, 1984.
[13] A. Miller. Subset selection in regression, volume 95 of Monographs on Statistics and Applied
Probability. Chapman & Hall/CRC, Boca Raton, FL, second edition, 2002.
[14] S. Reid and R. Tibshirani. Sparse regression and marginal testing using cluster prototypes.
Biostat., 17(2):364-376, 2016.
[15] W. Su, J. Qian, and L. Liu. Communication-efficient false discovery rate control via knockoff
aggregation. arXiv:1506.05446v2,2015.
[16] Q. Xu, J. Xiong, X. Cao, and Y. Yao. False discovery rate control and statistical quality
assessment of annotators in crowdsourced ranking. arXiv:1605.05860,2016
27
| 10math.ST
|
Continuous-Time Influence Maximization of Multiple Items
Scalable Influence Maximization for Multiple Products in
Continuous-Time Diffusion Networks
Nan Du
[email protected]
Google Research, 1600 Amphitheatre Pkwy, Mountain View, CA 94043
Yingyu Liang
[email protected]
Department of Computer Science, Princeton University, Princeton, NJ 08540
arXiv:1612.02712v2 [cs.SI] 29 Jan 2017
Maria-Florina Balcan
[email protected]
School of Computer Science, Carnegie Mellon University, Pittsburgh, PA 15213
Manuel Gomez-Rodriguez
[email protected]
MPI for Software Systems, Kaiserslautern, Germany 67663
Hongyuan Zha
Le Song
[email protected]
[email protected]
College of Computing, Georgia Institute of Technology, Atlanta, GA 30332
Abstract
A typical viral marketing model identifies influential users in a social network to maximize a single product adoption assuming unlimited user attention, campaign budgets, and
time. In reality, multiple products need campaigns, users have limited attention, convincing users incurs costs, and advertisers have limited budgets and expect the adoptions to
be maximized soon. Facing these user, monetary, and timing constraints, we formulate the
problem as a submodular maximization task in a continuous-time diffusion model under the
intersection of one matroid and multiple knapsack constraints. We propose a randomized
algorithm estimating the user influence1 in a network (|V| nodes, |E| edges) to an accuracy
of with n = O(1/2 ) randomizations and Õ(n|E| + n|V|) computations. By exploiting the
influence estimation algorithm as a subroutine, we develop an adaptive threshold greedy
algorithm achieving an approximation factor ka /(2 + 2k) of the optimal when ka out of
the k knapsack constraints are active. Extensive experiments on networks of millions of
nodes demonstrate that the proposed algorithms achieve the state-of-the-art in terms of
effectiveness and scalability.
Keywords: Influence Maximization, Influence Estimation, Continuous-time Diffusion
Model, Matroid, Knapsack
1. Introduction
Online social networks play an important role in the promotion of new products, the spread
of news, the success of political campaigns, and the diffusion of technological innovations. In
these contexts, the influence maximization problem (or viral marketing problem) typically
has the following flavor: identify a set of influential users in a social network, who, when
1. Partial results in the paper on influence estimation have been published in a conference paper: Nan Du,
Le Song, Manuel Gomez-Rodriguez, and Hongyuan Zha. Scalable influence estimation in continuous
time diffusion networks. In Advances in Neural Information Processing Systems 26, 2013.
1
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
convinced to adopt a product, shall influence other users in the network and trigger a large
cascade of adoptions. This problem has been studied extensively in the literature from both
the modeling and the algorithmic aspects (Richardson and Domingos, 2002; Kempe et al.,
2003; Leskovec et al., 2007; Chen et al., 2009, 2010a,b, 2011, 2012; Ienco et al., 2010; Goyal
et al., 2011a,b; Gomez-Rodriguez and Schölkopf, 2012), where it has been typically assumed
that the host (e.g., the owner of an online social platform) faces a single product, endless
user attention, unlimited budgets and unbounded time. However, in reality, the host often
encounters a much more constrained scenario:
• Multiple-Item Constraints: multiple products can spread simultaneously among
the same set of social entities. These products may have different characteristics, such
as their revenues and speed of spread.
• Timing Constraints: the advertisers expect the influence to occur within a certain
time window, and different products may have different timing requirements.
• User Constraints: users of the social network, each of which can be a potential
source, would like to be exposed to only a small number of ads. Furthermore, users
may be grouped by their geographical locations, and advertisers may have a target
population they want to reach.
• Product Constraints: seeking initial adopters entails a cost to the advertiser, who
needs to pay to the host and often has a limited amount of money.
For example, Facebook (i.e., the host) needs to allocate ads for various products with
different characteristics, e.g., clothes, books, or cosmetics. While some products, such as
clothes, aim at influencing within a short time window, some others, such as books, may
allow for longer periods. Moreover, Facebook limits the number of ads in each user’s sidebar (typically it shows less than five) and, as a consequence, it cannot assign all ads to a
few highly influential users. Finally, each advertiser has a limited budget to pay for ads on
Facebook and thus each ad can only be displayed to some subset of users. In our work,
we incorporate these myriads of practical and important requirements into consideration in
the influence maximization problem.
We account for the multi-product and timing constraints by applying product-specific
continuous-time diffusion models. Here, we opt for continuous-time diffusion models instead of discrete-time models, which have been mostly used in previous work (Kempe et al.,
2003; Chen et al., 2009, 2010a,b, 2011, 2012; Borgs et al., 2012). This is because artificially discretizing the time axis into bins introduces additional errors. One can adjust
the additional tuning parameters, like the bin size, to balance the tradeoff between the
error and the computational cost, but the parameters are not easy to choose optimally.
Extensive experimental comparisons on both synthetic and real-world data have shown
that discrete-time models provide less accurate influence estimation than their continuoustime counterparts (Gomez-Rodriguez et al., 2011; Gomez-Rodriguez and Schölkopf, 2012;
Gomez-Rodriguez et al., 2013; Du et al., 2013a,b).
However, maximizing influence based on continuous-time diffusion models also entails
additional challenges. First, evaluating the objective function of the influence maximization
problem (i.e., the influence estimation problem) in this setting is a difficult graphical model
2
Continuous-Time Influence Maximization of Multiple Items
inference problem, i.e., computing the marginal density of continuous variables in loopy
graphical models. The exact answer can be computed only for very special cases. For example, Gomez-Rodriguez and Schölkopf (2012) have shown that the problem can be solved
exactly when the transmission functions are exponential densities, by using continuous time
Markov processes theory. However, the computational complexity of such approach, in general, scales exponentially with the size and density of the network. Moreover, extending the
approach to deal with arbitrary transmission functions would require additional nontrivial
approximations which would increase even more the computational complexity. Second,
it is unclear how to scale up influence estimation and maximization algorithms based on
continuous-time diffusion models to millions of nodes. Especially in the maximization case,
the influence estimation procedure needs to be called many times for different subsets of
selected nodes. Thus, our first goal is to design a scalable algorithm which can perform
influence estimation in the regime of networks with millions of nodes.
We account for the user and product constraints by restricting the feasible domain over
which the maximization is performed. We first show that the overall influence function of
multiple products is a submodular function and then realize that the user and product constraints correspond to constraints over the ground set of this submodular function. To the
best of our knowledge, previous work has not considered both user and product constraints
simultaneously over general unknown different diffusion networks with non-uniform costs.
In particular, (Datta et al., 2010) first tried to model both the product and user constraints
only with uniform costs and infinite time window, which essentially reduces to a special
case of our formulations. Similarly, (Lu et al., 2013) considered the allocation problem
of multiple products which may have competitions within the infinite time window. Besides, they all assume that multiple products spread within the same network. In contrast,
our formulations generally allow products to have different diffusion networks, which can
be unknown in practice. Soma et al. (2014) studied the influence maximization problem
for one product subject to one knapsack constraint over a known bipartite graph between
marketing channels and potential customers; Ienco et al. (2010) and Sun et al. (2011) considered user constraints but disregarded product constraints during the initial assignment;
and, Narayanam and Nanavati (2012) studied the cross-sell phenomenon (the selling of the
first product raises the chance of selling the second) and included monetary constraints for
all the products. However, no user constraints were considered, and the cost of each user
was still uniform for each product. Thus, our second goal is to design an efficient submodular maximization algorithm which can take into account both user and product constraints
simultaneously.
Overall, this article includes the following major contributions:
• Unlike prior work that considers an a priori described simplistic discrete-time diffusion model, we first learn the diffusion networks from data by using continuous-time
diffusion models. This allows us to address the timing constraints in a principled way.
• We provide a novel formulation of the influence estimation problem in the continuoustime diffusion model from the perspective of probabilistic graphical models, which
allows heterogeneous diffusion dynamics over the edges.
3
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
• We propose an efficient randomized algorithm for continuous-time influence estimation, which can scale up to millions of nodes and estimate the influence of each node
to an accuracy of using n = O(1/2 ) randomizations.
• We formulate the influence maximization problem with the aforementioned constraints
as a submodular maximization under the intersection of matroid constraints and knapsack constraints. The submodular function we use is based on the actual diffusion
model learned from the data for the time window constraint. This novel formulation provides us a firm theoretical foundation for designing greedy algorithms with
theoretical guarantees.
• We develop an efficient adaptive-threshold greedy algorithm which is linear in the
e
number of products and proportional to O(|V|
+ |E ∗ |), where |V| is the number of
∗
nodes (users) and |E | is the number of edges in the largest diffusion network. We then
prove that this algorithm is guaranteed to find a solution with an overall influence
ka
of at least 2+2k
of the optimal value, when ka out of the k knapsack constraints
are active. This improves over the best known approximation factor achieved by
polynomial time algorithms in the combinatorial optimization literature. Moreover,
whenever advertising each product to each user entails the same cost, the constraints
reduce to an intersection of matroids, and we obtain an approximation factor of 1/3,
which is optimal for such optimization.
• We evaluate our algorithms over large synthetic and real-world datasets and show that
our proposed methods significantly improve over previous state-of-the-arts in terms
of both the accuracy of the estimated influence and the quality of the selected nodes
in maximizing the influence over independently hold-out real testing data.
In the remainder of the paper, we will first tackle the influence estimation problem in
section 2. We then formulate different realistic constraints for the influence maximization
in section 3 and present the adaptive-thresholding greedy algorithm with its theoretical
analysis in section 4; we investigate the performance of the proposed algorithms in both
synthetic and real-world datasets in section 5; and finally we conclude in section 6.
2. Influence Estimation
We start by revisiting the continuous-time diffusion model by Gomez-Rodriguez et al. (2011)
and then explicitly formulate the influence estimation problem from the perspective of
probabilistic graphical models. Because the efficient inference of the influence value for
each node is highly non-trivial, we further develop a scalable influence estimation algorithm
which is able to handle networks of millions of nodes. The influence estimation procedure
will be a key building block for our later influence maximization algorithm.
2.1 Continuous-Time Diffusion Networks
The continuous-time diffusion model associates each edge with a transmission function, that
is, a density over the transmission time along the edge, in contrast to previous discrete-time
models which associate each edge with a fixed infection probability (Kempe et al., 2003).
4
Continuous-Time Influence Maximization of Multiple Items
Moreover, it also differs from discrete-time models in the sense that events in a cascade
are not generated iteratively in rounds, but event timings are sampled directly from the
transmission function in the continuous-time model.
Continuous-Time Independent Cascade Model. Given a directed contact network,
G = (V, E), we use the independent cascade model for modeling a diffusion process (Kempe
et al., 2003; Gomez-Rodriguez et al., 2011). The process begins with a set of infected
source nodes, A, initially adopting certain contagion (idea, meme or product) at time zero.
The contagion is transmitted from the sources along their out-going edges to their direct
neighbors. Each transmission through an edge entails random waiting times, τ , drawn from
different independent pairwise waiting time distributions(one per edge). Then, the infected
neighbors transmit the contagion to their respective neighbors, and the process continues.
We assume that an infected node remains infected for the entire diffusion process. Thus, if
a node i is infected by multiple neighbors, only the neighbor that first infects node i will
be the true parent. As a result, although the contact network can be an arbitrary directed
network, each diffusion process induces a Directed Acyclic Graph (DAG).
Heterogeneous Transmission Functions. Formally, the pairwise transmission function fji (ti |tj ) for a directed edge j → i is the conditional density of node i getting infected
at time ti given that node j was infected at time tj . We assume it is shift invariant:
fji (ti |tj ) = fji (τji ), where τji := ti − tj , and causal: fji (τji ) = 0 if τji < 0. Both parametric
transmission functions, such as the exponential and Rayleigh function (Gomez-Rodriguez
et al., 2011), and nonparametric functions (Du et al., 2012) can be used and estimated from
cascade data.
Shortest-Path Property. The independent cascade model has a useful property we
will use later: given a sample of transmission times of all edges, the time ti taken to infect
a node i is the length of the shortest path in G from the sources to node i, where the edge
weights correspond to the associated transmission times.
2.2 Probabilistic Graphical Model for Continuous-Time Diffusion Networks
The continuous-time independent cascade model is essentially a directed graphical model for
a set of dependent random variables, that is, the infection times ti of the nodes, where the
conditional independence structure is supported on the contact network G. Although the
original contact graph G can contain directed loops, each diffusion process (or a cascade)
induces a directed acyclic graph (DAG). For those cascades consistent with a particular
DAG, we can model the joint density of ti using a directed graphical model:
p ({ti }i∈V ) =
Y
i∈V
p (ti |{tj }j∈πi ) ,
(1)
where each πi denotes the collection of parents of node i in the induced DAG, and each
term p(ti |{tj }j∈πi ) corresponds to a conditional density of ti given the infection times of
the parents of node i. This is true because given the infection times of node i’s parents, ti
is independent of other infection times, satisfying the local Markov property of a directed
graphical model. We note that the independent cascade model only specifies explicitly
the pairwise transmission function of each directed edge, but does not directly define the
conditional density p(ti |{tj }j∈πi ).
5
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
However, these conditional densities can be derived from the pairwise transmission functions based on the Independent-Infection property:
p (ti |{tj }j∈πi ) =
X
j∈πi
fji (ti |tj )
Y
l∈πi ,l6=j
S(ti |tl ),
(2)
which is the sum of the likelihoods that node i is infected by each parent node j. More
precisely, each term in the summation can be interpreted as the likelihood fji (ti |tj ) of node
i being infected at ti by node j multiplied by the probability S(ti |tl ) that it has survived
from the infection of each other parent node l 6= j until time ti .
Perhaps surprisingly, the factorization in Equation (1) is the same factorization that
can be used for an arbitrary induced DAG consistent with the contact network G. In this
case, we only need to replace the definition of πi (the parent of node i in the DAG) to the
set of neighbors of node i with an edge pointing to node i in G. This is not immediately
obvious from Equation (1), since the contact network G can contain directed loops which
seems to be in conflict with the conditional independence semantics of directed graphical
models. The reason why it is possible to do so is as follows: any fixed set of infection
times, t1 , . . . , td , induces an ordering of the infection times. If ti ≤ tj for an edge j → i
in G, hji (ti |tj ) = 0, and the corresponding term in Equation (2) is zeroed out, making the
conditional density consistent with the semantics of directed graphical models.
Instead of directly modeling the infection times ti , we can focus on the set of mutually
independent random transmission times τji = ti − tj . Interestingly, by switching from a
node-centric view to an edge-centric view, we obtain a fully factorized joint density of the
set of transmission times
Y
p {τji }(j,i)∈E =
(j,i)∈E
fji (τji ),
(3)
Based on the Shortest-Path property of the independent cascade model, each variable ti can
be viewed as a transformation from the collection of variables {τji }(j,i)∈E . More specifically,
let Qi be the collection of directed paths in G from the source nodes to node i, where each
path q ∈ Qi contains a sequence of directed edges (j, l). Assuming all source nodes are
infected at time zero, then we obtain variable ti via
X
ti = gi {τji }(j,i)∈E |A = min
q∈Qi
(j,l)∈q
τjl ,
(4)
where the transformation gi (·|A) is the value of the shortest-path minimization. As a
special case, we can now compute the probability of node i infected before T using a set of
independent variables:
Pr {ti ≤ T |A} = Pr gi {τji }(j,i)∈E |A ≤ T .
(5)
The significance of the relation is that it allows us to transform a problem involving
a sequence of dependent variables {ti }i∈V to one with independent variables {τji }(j,i)∈E .
Furthermore, the two perspectives are connected via the shortest path algorithm in weighted
directed graph, a standard well-studied operation in graph analysis.
6
Continuous-Time Influence Maximization of Multiple Items
2.3 Influence Estimation Problem in Continuous-Time Diffusion Networks
Intuitively, given a time window, the wider the spread of infection, the more influential the
set of sources. We adopt the definition of influence as the expected number of infected nodes
given a set of source nodes and a time window, as in previous work (Gomez-Rodriguez and
Schölkopf, 2012). More formally, consider a set of source nodes A ⊆ V, |A| ≤ C which get
infected at time zero. Then, given a time window T , a node i is infected within the time
window if ti ≤ T . The expected number of infected nodes (or the influence) given the set
of transmission functions {fji }(j,i)∈E can be computed as
σ(A, T ) = E
hX
i∈V
i X
I {ti ≤ T |A} =
i∈V
Pr {ti ≤ T |A} ,
(6)
where I {·} is the indicator function and the expectation is taken over the the set of dependent
variables {ti }i∈V . By construction, σ(A, T ) is a non-negative, monotonic nondecreasing
submodular function in the set of source nodes shown by Gomez-Rodriguez and Schölkopf
(2012).
Essentially, the influence estimation problem in Equation (6) is an inference problem for
graphical models, where the probability of event ti ≤ T given sources in A can be obtained
by summing out the possible configuration of other variables {tj }j6=i . That is
Z
∞
Z
T
···
Pr{ti ≤ T |A} =
0
Z
···
ti =0
∞ Y
j∈V
0
p tj |{tl }l∈πj
Y
j∈V
dtj ,
(7)
which is, in general, a very challenging problem. First, the corresponding directed graphical
models can contain nodes with high in-degree and high out-degree. For example, in Twitter,
a user can follow dozens of other users, and another user can have hundreds of “followers”.
The tree-width corresponding to this directed graphical model can be very high, and we
need to perform integration for functions involving many continuous variables. Second,
the integral in general can not be evaluated analytically for heterogeneous transmission
functions, which means that we need to resort to numerical integration by discretizing the
domain [0, ∞). If we use N levels of discretization for each variable, we would need to
enumerate O(N |πi | ) entries, exponential in the number of parents.
Only in very special cases, can one derive the closed-form equation for computing
Pr{ti ≤ T |A}. For instance, Gomez-Rodriguez and Schölkopf (2012) proposed an approach
for exponential transmission functions, where the special properties of exponential density
are used to map the problem into a continuous time Markov process problem, and the computation can be carried out via a matrix exponential. However, without further heuristic
approximation, the computational complexity of the algorithm is exponential in the size
and density of the network. The intrinsic complexity of the problem entails the utilization
of approximation algorithms, such as mean field algorithms or message passing algorithms.
We will design an efficient randomized (or sampling) algorithm in the next section.
2.4 Efficient Influence Estimation in Continuous-Time Diffusion Networks
Our first key observation is that we can transform the influence estimation problem in
Equation (6) into a problem with independent variables. With the relation in Equation (5),
7
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
we can derive the influence function as
X
Pr gi {τji }(j,i)∈E |A ≤ T
σ(A, T ) =
i∈V
i
hX
I gi {τji }(j,i)∈E |A ≤ T ,
=E
i∈V
(8)
where the expectation is with respect to the set of independent variables {τji }(j,i)∈E . This
equivalent formulation suggests a naive sampling (NS) algorithm for approximating σ(A, T ):
draw n samples of {τji }(j,i)∈E , run a shortest path algorithm for each sample, and finally
average the results (see Appendix A for more details). However, this naive sampling approach has a computational complexity of O(nC|V||E| + nC|V|2 log |V|) due to the repeated
calling of the shortest path algorithm. This is quadratic to the network size, and hence not
scalable to millions of nodes.
Our second key observation is that for each sample {τji }(j,i)∈E
P, we are only interested in
the neighborhood size of the source nodes, i.e., the summation i∈V I {·} in Equation (8),
rather than in the individual shortest paths. Fortunately, the neighborhood size estimation
problem has been studied in the theoretical computer science literature. Here, we adapt a
very efficient randomized algorithm by Cohen (1997) to our influence estimation problem.
This randomized algorithm has a computational complexity of O(|E| log |V| + |V| log2 |V|)
and it estimates the neighborhood sizes for all possible single source node locations. Since
it needs to run once for each sample of {τji }(j,i)∈E , we obtain an overall influence estimation
algorithm with O(n|E| log |V| + n|V| log2 |V|) computation, nearly linear in network size.
Next we will revisit Cohen’s algorithm for neighborhood estimation.
2.4.1 Randomized Algorithm for Single-Source Neighborhood-Size
Estimation
Given a fixed set of edge transmission times {τji }(j,i)∈E and a source node s, infected at
time zero, the neighborhood N (s, T ) of a source node s given a time window T is the set
of nodes within distance T from s, i.e.,
N (s, T ) = i gi {τji }(j,i)∈E ≤ T, i ∈ V .
(9)
Instead of estimating N (s, T ) directly, the algorithm will assign an exponentially distributed
random label ri to each network node i. Then, it makes use of the fact that the minimum of a
set of exponential random variables {ri }i∈N (s,T ) is still an exponential random variable, but
with its parameter being equal to the total number of variables, that is, if each ri ∼ exp(−ri ),
then the smallest label within distance T from source s, r∗ := mini∈N (s,T ) ri , will distribute
as r∗ ∼ exp {−|N (s, T )|r∗ }. Suppose we randomize over the labeling m times and obtain
m such least labels, {r∗u }m
u=1 . Then the neighborhood size can be estimated as
m−1
|N (s, T )| ≈ Pm u .
u=1 r∗
(10)
which is shown by Cohen (1997) to be an unbiased estimator of |N (s, T )|. This is an elegant
relation since it allows us to transform the counting problem in (9) to a problem of finding
the minimum random label r∗ . The key question is whether we can compute the least label
r∗ efficiently, given random labels {ri }i∈V and any source node s.
8
Continuous-Time Influence Maximization of Multiple Items
Cohen (1997) designed a modified Dijkstra’s algorithm (Algorithm 3) to construct a data
structure r∗ (s), called least label list, for each node s to support such query. Essentially, the
algorithm starts with the node i with the smallest label ri , and then it traverses in breadthfirst search fashion along the reverse direction of the graph edges to find all reachable nodes.
For each reachable node s, the distance d∗ between i and s, and ri are added to the end
of r∗ (s). Then the algorithm moves to the node i0 with the second smallest label ri0 , and
similarly find all reachable nodes. For each reachable node s, the algorithm will compare
the current distance d∗ between i0 and s with the last recorded distance in r∗ (s). If the
current distance is smaller, then the current d∗ and ri0 are added to the end of r∗ (s). Then
the algorithm move to the node with the third smallest label and so on. The algorithm is
summarized in Algorithm 3 in Appendix B.
Algorithm 3 returns a list r∗ (s) per node s ∈ V, which contains information about
distance to the smallest reachable labels from s. In particular, each list contains pairs of
distance and random labels, (d, r), and these pairs are ordered as
∞ > d(1) > d(2) > . . . > d(|r∗ (s)|) = 0
(11)
r(1) < r(2) < . . . < r(|r∗ (s)|) ,
(12)
where {·}(l) denotes the l-th element in the list. (see Appendix B for an example).
If we want to query the smallest reachable random label r∗ for a given source s and a
time T , we only need to perform a binary search on the list for node s:
r∗ = r(l) , where d(l−1) > T ≥ d(l) .
(13)
Finally, to estimate |N (s, T )|, we generate m i.i.d. collections of random labels, run Algorithm 3 on each collection, and obtain m values {r∗u }m
u=1 , which we use in Equation (10) to
estimate |N (i, T )|.
The computational complexity of Algorithm 3 is O(|E| log |V| + |V| log2 |V|), with expected size of each r∗ (s) being O(log |V|). Then the expected time for querying r∗ is
O(log log |V|) using binary search. Since we need to generate m set of random labels and
run Algorithm 3 m times, the overall computational complexity for estimating the singlesource neighborhood size for all s ∈ V is O(m|E| log |V| + m|V| log2 |V| + m|V| log log |V|).
For large-scale network, and when m min{|V|, |E|}, this randomized algorithm can be
much more efficient than approaches based on directly calculating the shortest paths.
2.4.2 Constructing Estimation for Multiple-Source Neighborhood Size
When we have a set of sources, A, its neighborhood is the union of the neighborhoods of
its constituent sources
[
N (A, T ) =
N (i, T ).
(14)
i∈A
This is true because each source independently infects its downstream nodes. Furthermore,
to calculate the least label list r∗ corresponding to N (A, T ), we can simply reuse the least
label list r∗ (i) of each individual source i ∈ A. More formally,
r∗ = mini∈A minj∈N (i,T ) rj ,
9
(15)
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
where the inner minimization can be carried out by querying r∗ (i). Similarly, after we
obtain m samples of r∗ , we can estimate |N (A, T )| using Equation (10). Importantly, very
little additional work is needed when we want to calculate r∗ for a set of sources A, and
we can reuse work done for a single source. This is very different from a naive sampling
approach where the sampling process needs to be done completely anew if we increase the
source set. In contrast, using the randomized algorithm, only an additional constant-time
minimization over |A| numbers is needed.
2.4.3 Overall Algorithm
So far, we have achieved efficient neighborhood size estimation of |N (A, T )| with respect to a
given set of transmission times {τji }(j,i)∈E . Next, we will estimate the influence by averaging
over multiple sets of samples for {τji }(j,i)∈E . More specifically, the relation from (8)
m−1
σ(A, T ) = E{τji }(j,i)∈E [|N (A, T )|] = E{τji } E{r1 ,...,rm }|{τji } Pm u ,
u=1 r∗
(16)
suggests the following overall algorithm :
Continuous-Time Influence Estimation (ConTinEst):
1. Sample n sets of random transmission times {τijl }(j,i)∈E ∼
Q
(j,i)∈E
fji (τji ).
Q
2. Given a set of {τijl }(j,i)∈E , sample m sets of random labels {riu }i∈V ∼
i∈V exp(−ri ).
P
P
ul
3. Estimate σ(A, T ) by sample averages σ(A, T ) ≈ n1 nl=1 (m − 1)/ m
r
.
ul =1 ∗
What is even more important is that the number of random labels, m, does not need
to be very large. Since the estimator for |N (A, T )| is unbiased (Cohen, 1997), essentially
the outer-loop of averaging over n samples of random transmission times further reduces
the variance of the estimator in a rate of O(1/n). In practice, we can use a very small m
(e.g., 5 or 10) and still achieve good results, which is also confirmed by our later experiments. Compared to (Chen et al., 2009), the novel application of Cohen’s algorithm arises
for estimating influence for multiple sources, which drastically reduces the computation
by cleverly using the least-label list from single source. Moreover, we have the following
theoretical guarantee (see Appendix C for the proof).
Theorem 1 Draw the following number of samples for the set of random transmission
times
CΛ
2|V|
n ≥ 2 log
(17)
α
where Λ := maxA:|A|≤C 2σ(A, T )2 /(m − 2) + 2V ar(|N (A, T )|)(m − 1)/(m − 2) + 2a/3 and
|N (A, T )| ≤ |V|, and for each set of random transmission times, draw m sets of random
labels. Then |b
σ (A, T ) − σ(A, T )| ≤ uniformly for all A with |A| ≤ C, with probability at
least 1 − α.
10
Continuous-Time Influence Maximization of Multiple Items
The theorem indicates that the minimum number of samples, n, needed to achieve
certain accuracy is related to the actual size of the influence σ(A, T ), and the variance
of the neighborhood size |N (A, T )| over the random draw of samples. The number of
random labels, m, drawn in the inner loop of the algorithm will monotonically decrease the
dependency of n on σ(A, T ). It suffices to draw a small number of random labels, as long
as the value of σ(A, T )2 /(m − 2) matches that of V ar(|N (A, T )|). Another implication is
that influence at larger time window T is harder to estimate, since σ(A, T ) will generally
be larger and hence require more random samples.
3. Constraints of Practical Importance
By treating our proposed influence estimation algorithm ConTinEst as a building block, we
can now tackle the influence maximization problem under various constraints of practical
importance. Here, since ConTinEst can estimate the influence value of any source set
with respect to any given time window T , the Timing Constraints can thus be naturally
satisfied. Therefore, in the following sections, we mainly focus on modeling the MultipleItem Constraints, the User Constraints and the Product Constraints.
3.1 Multiple-Item Constraints
Multiple products can spread simultaneously across the same set of social entities over
different diffusion channels. Since these products may have different characteristics, such
as the revenue and the speed of spread, and thus may follow different diffusion dynamics,
we will use multiple diffusion networks for different types of products.
Suppose we have a set of products L that propagate on the same set of nodes V. The
diffusion network for product i is denoted as Gi = (V, Ei ). For each product i ∈ L, we search
for a set of source nodes Ri ⊆ V to which we can assign the product i to start its campaign.
We can represent the selection of Ri ’s using an assignment matrix A ∈ {0, 1}|L|×|V| as
follows: Aij = 1 if j ∈ Ri and Aij = 0 otherwise. Based on this representation, we define
a new ground set Z = L × V of size N = |L| × |V|. Each element of Z corresponds to the
index (i, j) of an entry in the assignment matrix A, and selecting element z = (i, j) means
assigning product i to user j (see Figure 1 for an illustration). We also denote Z∗j = L×{j}
and Zi∗ = {i} × V as the j-th column and i-th row of matrix A, respectively. Then, under
the above mentioned additional requirements, we would like to find a set of assignments
S ⊆ Z so as to maximize the following overall influence
f (S) =
X
ai σi (Ri , Ti ),
(18)
i∈L
where σi (Ri , Ti ) denote the influence of product i for a given time Ti , {ai > 0} is a set of
weights reflecting the different benefits of the products and Ri = {j ∈ V : (i, j) ∈ S}. We
now show that the overall influence function f (S) in Equation (18) is submodular over the
ground set Z.
Lemma 2 Under the continuous-time independent cascade model, the overall influence
f (S) is a normalized monotone submodular function of S.
11
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
j
i
Figure 1: Illustration of the assignment matrix A associated with partition matroid M1
and group knapsack constraints. If product i is assigned to user j, then Aij = 1 (colored
in red). The ground set Z is the set of indices of the entries in A, and selecting an element
(i, j) ∈ Z means assigning product i to user j. The user constraint means that there are at
most uj elements selected in the j-th column; the product constraint means that the total
cost of the elements selected in the i-th row is at most Bi .
Proof By definition, f (∅) = 0 and f (S) is monotone. By Theorem 4 in Gomez-Rodriguez
and Schölkopf (2012), the component influence function σi (Ri , Ti ) for product i is submodular in Ri ⊆ V. Since non-negative linear combinations of submodular functions are
P still submodular, fi (S) := ai σi (Ri , Ti ) is also submodular in S ⊆ Z = L×V, and f (S) = i∈L fi (S)
is submodular.
3.2 User Constraints
Each social network user can be a potential source and would like to be exposed only to a
small number of ads. Furthermore, users may be grouped according to their geographical
locations, and advertisers may have a target population they want to reach. Here, we will
incorporate these constraints using the matroids which are combinatorial structures that
generalize the notion of linear independence in matrices (Schrijver, 2003; Fujishige, 2005).
Formulating our constrained influence maximization task with matroids allows us to design
a greedy algorithm with provable guarantees.
Formally, suppose that each user j can be assigned to at most uj products. A matroid
can be defined as follows:
Definition 3 A matroid is a pair, M = (Z, I), defined over a finite set (the ground set)
Z and a family of sets (the independent sets) I, that satisfies three axioms:
1. Non-emptiness: The empty set ∅ ∈ I.
2. Heredity: If Y ∈ I and X ⊆ Y , then X ∈ I.
3. Exchange: If X ∈ I, Y ∈ I and |Y | > |X|, then there exists z ∈ Y \ X such that
X ∪ {z} ∈ I.
12
Continuous-Time Influence Maximization of Multiple Items
An important type of matroid is the partition matroid where the ground set Z is partitioned into disjoint subsets Z1 , Z2 , . . . , Zt for some t and
I = {S | S ⊆ Z and |S ∩ Zi | ≤ ui , ∀i = 1, . . . , t}
for some given parameters u1 , . . . , ut . The user constraints can then be formulated as
Partition matroid M1 : partition the ground set Z into Z∗j = L × {j} each of
which corresponds to a column of A. Then M1 = {Z, I1 } is
I1 = {S|S ⊆ Z and |S ∩ Z∗j | ≤ uj , ∀j} .
3.3 Product Constraints
Seeking initial adopters entails a cost to the advertiser, which needs to be paid to the
host, while the advertisers of each product have a limited amount of money. Here, we will
incorporate these requirements using knapsack constraints which we describe below.
Formally, suppose that each product i has a budget Bi , and assigning item i to user j
costs cij > 0. Next, we introduce the following notation to describe product constraints
over the ground set Z. For an element z = (i, j) ∈ Z, define its cost as P
c(z) := cij . Abusing
the notation slightly, we denote the cost of a subset S ⊆ Z as c(S) := z∈S c(z). Then, in
a feasible solution S ⊆ Z, the cost of assigning product i, c(S ∩ Zi∗ ), should not be larger
than its budget Bi .
Now, without loss of generality, we can assume Bi = 1 (by normalizing cij with Bi ),
and also cij ∈ (0, 1] (by throwing away any element (i, j) with cij > 1), and define
Group-knapsack: partition the ground set into Zi∗ = {i} × V each of which
corresponds to one row of A. Then a feasible solution S ⊆ Z satisfies
c(S ∩ Zi∗ ) ≤ 1, ∀i.
Importantly, these knapsack constraints have very specific structure: they are on different groups of a partition {Zi∗ } of the ground set and the submodular function f (S) =
P
i ai σi (Ri , Ti ) is defined over the partition. In consequence, such structures allow us to
design an efficient algorithm with improved guarantees over the known results.
3.4 Overall Problem Formulation
Based on the above discussion of various constraints in viral marketing and our design
choices for tackling them, we can think of the influence maximization problem as a special
case of the following constrained submodular maximization problem with P = 1 matroid
constraints and k = |L| knapsack constraints,
maxS⊆Z
f (S)
(19)
subject to c(S ∩ Zi∗ ) ≤ 1,
S∈
P
\
1 ≤ i ≤ k,
Ip ,
p=1
where, for simplicity, we will denote all the feasible solutions S ⊆ Z as F. This formulation
in general includes the following cases of practical importance :
13
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
Uniform User-Cost. An important case of influence maximization, which we denote as
the Uniform Cost, is that for each product i, all users have the same cost ci∗ , i.e., cij = ci∗ .
Equivalently, each product i can be assigned to at most bi := bBi /ci∗ c users. Then the
product constraints are simplified to
Partition matroid M2 : for the product constraints with uniform cost, define a
matroid M2 = {Z, I2 } where
I2 = {S|S ⊆ Z and |S ∩ Zi∗ | ≤ bi , ∀i} .
In this case, the influence maximization problem defined by Equation (19) becomes the
problem with P = 2 matroid constraints and no knapsack constraints (k = 0). In addition,
if we assume only one product needs campaign, the formulation of Equation (19) further
reduces to the classic influence maximization problem with the simple cardinality constraint.
User Group Constraint. Our formulation in Equation (19) essentially allows for general matroids which can model more sophisticated real-world constraints, and the proposed
formulation, algorithms, and analysis can still hold. For instance, suppose there is a hierarchical community structure on the users, i.e., a tree T where leaves are the users and
the internal nodes are communities consisting of all users underneath, such as customers
in different countries around the world. In consequence of marketing strategies, on each
community C ∈ T , there are at most uC slots for assigning the products. Such constraints
are readily modeled by the Laminar Matroid, which generalizes the partition matroid by
allowing the set {Zi } to be a laminar family (i.e., for any Zi 6= Zj , either Zi ⊆ Zj , or
Zj ⊆ Zi , or Zi ∩ Zj = ∅). It can be shown that the community constraints can be captured
by the matroid M = (Z, I) where I = {S ⊆ Z : |S ∩ C| ≤ uC , ∀C ∈ T }. In the next section, we first present our algorithm, then provide the analysis for the uniform cost case and
finally leverage such analysis for the general case.
4. Influence Maximization
In this section, we first develop a simple, practical and intuitive adaptive-thresholding greedy
algorithm to solve the continuous-time influence maximization problem with the aforementioned constraints. Then, we provide a detailed theoretical analysis of its performance.
4.1 Overall Algorithm
There exist algorithms for submodular maximization under multiple knapsack constraints
achieving a 1 − 1e approximation factor by (Sviridenko, 2004). Thus, one may be tempted
to convert the matroid constraint in the problem defined by Equation (19) to |V| knapsack
constraints, so that the problem becomes a submodular maximization problem under |L| +
|V| knapsack constraints. However, this naive approach is not practical for large-scale
scenarios because the running time of such algorithms is exponential in the number of
knapsack constraints. Instead, if we opt for algorithms for submodular maximization under
k knapsack constraints and P matroids constraints, the best approximation factor achieved
1
(Badanidiyuru and Vondrák, 2014). However,
by polynomial time algorithms is P +2k+1
14
Continuous-Time Influence Maximization of Multiple Items
Algorithm 1: Density Threshold Enumeration
1
2
3
Input: parameter δ; objective f or its approximation fb; assignment cost c(z), z ∈ Z
Set d = max {f ({z}) : z ∈ Z};
n
o
2d
2d
, (1 + δ) P +2k+1
, . . . , P 2|Z|d
do
for ρ ∈ P +2k+1
+2k+1
Call Algorithm 2 to get Sρ ;
Output: argmaxSρ f (Sρ )
this is not good enough yet, since in our problem k = |L| can be large, though P = 1 is
small.
Here, we will design an algorithm that achieves a better approximation factor by exploiting the following key observation about the structure of the problem defined by Equation (19): the knapsack constraints are over different groups Zi∗ of the whole ground set,
and the objective function is a sum of submodular functions over these different groups.
The details of the algorithm, called BudgetMax, are described in Algorithm 1. BudgetMax enumerates different values of a so-called density threshold ρ, runs a subroutine
to find a solution for each ρ, which quantifies the cost-effectiveness of assigning a particular product to a specific user, and finally outputs the solution with the maximum objective
value. Intuitively, the algorithm restricts the search space to be the set of most cost-effective
allocations. The details of the subroutine to find a solution for a fixed density threshold
ρ are described in Algorithm 2. Inspired by the lazy evaluation heuristic (Leskovec et al.,
2007), the algorithm maintains a working set G and a marginal gain threshold wt , which
geometrically decreases by a factor of 1 + δ until it is sufficiently small to be set to zero. At
each wt , the subroutine selects each new element z that satisfies the following properties:
1. It is feasible and the density ratio (the ratio between the marginal gain and the cost)
is above the current density threshold;
2. Its marginal gain
f (z|G) := f (G ∪ {z}) − f (G)
is above the current marginal gain threshold.
The term “density” comes from the knapsack problem, where the marginal gain is the mass
and the cost is the volume. A large density means gaining a lot without paying much. In
short, the algorithm considers only high-quality assignments and repeatedly selects feasible
ones with marginal gain ranging from large to small.
Remark 1. The traditional lazy evaluation heuristic also keeps a threshold, however, it
only uses the threshold to speed up selecting the element with maximum marginal gain.
Instead, Algorithm 2 can add multiple elements z from the ground set at each threshold, and
thus reduces the number of rounds from the size of the solution to the number of thresholds
O( 1δ log Nδ ). This allows us to trade off between the runtime and the approximation ratio
(refer to our theoretical guarantees in section 4.2).
15
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
Algorithm 2: Adaptive Threshold Greedy for Fixed Density
Input: parameters ρ, δ; objective f or its approximation fb; assignment cost
c(z), z ∈ Z;set of feasible solutions F; and d from Algorithm 1.
1 Set dρ = max {f ({z}) : z ∈ Z, f ({z}) ≥ c(z)ρ};
dρ
δd
2 Set wt = (1+δ)t for t = 0, . . . , L = argmini wi ≤ N and wL+1 = 0;
3
4
5
6
7
Set G = ∅;
for t = 0, 1, . . . , L, L + 1 do
for z 6∈ G with G ∪ {z} ∈ F and f (z|G) ≥ c(z)ρ do
if f (z|G) ≥ wt then
Set G ← G ∪ {z};
Output: Sρ = G
Remark 2. Evaluating the influence of the assigned products f is expensive. Therefore,
we will use the randomized algorithm in Section 2.4.3 to compute an estimation fb(·) of the
quantity f (·).
4.2 Theoretical Guarantees
Although our algorithm is quite intuitive, it is highly non-trivial to obtain the theoretical
guarantees. For clarity, we first analyze the simpler case with uniform cost, which then
provides the base for analyzing the general case.
4.2.1 Uniform Cost
As shown at the end of Section 3.4, the influence maximization, in this case, corresponds
to the problem defined by Equation (19) with P = 2 and no knapsack constraints. Thus,
we can simply run Algorithm 2 with ρ = 0 to obtain a solution G, which is roughly a
1
P +1 -approximation.
Intuition. The algorithm greedily selects feasible elements with sufficiently large marginal
gain. However, it is unclear whether our algorithm will find good solutions and whether it
will be robust to noise. Regarding the former, one might wonder whether the algorithm will
select just a few elements while many elements in the optimal solution O will become infeasible and will not be selected, in which case the greedy solution G is a poor approximation.
Regarding the latter, we only use the estimation fb of the influence f (i.e., |fb(S) − f (S)| ≤
for any S ⊆ Z), which introduces additional error to the function value. A crucial question,
which has not been addressed before (Badanidiyuru and Vondrák, 2014), is whether the
adaptive threshold greedy algorithm is robust to such perturbations.
Fortunately, it turns out that the algorithm will provably select sufficiently many elements of high quality. First, the elements selected in the optimal solution O but not selected
in G can be partitioned into |G| groups, each of which is associated with an element in G.
Thus, the number of elements in the groups associated with the first t elements in G, by
the property of the intersection of matroids, are bounded by P t. See Figure 2 for an illustration. Second, the marginal gain of each element in G is at least as large as that of any
16
Continuous-Time Influence Maximization of Multiple Items
Gt
g1
gt−1
g2
g|G|
gt
G
······
······
O\G
C1
Ct−1
C2
Ct
C|G|
St
i=1 Ci
Figure 2: Notation for analyzing Algorithm 2. The elements in the greedy solution G are
arranged according to the order in which Algorithm 2 selects them in Step 3. The elements
in the optimal solution O but not in the greedy solution G are partitioned into groups
{Ct }1≤t≤|G| , where Ct are those elements in O \ G that are still feasible before selecting gt
but are infeasible after selecting gt .
element in the group associated with it (up to some small error). This means that even
if the submodular function evaluation is inexact, the quality of the elements in the greedy
solution is still good. The two claims together show that the marginal gain of O \ G is not
much larger than the gain of G, and thus G is a good approximation for the problem.
Formally, suppose we use an inexact evaluation of the influence f such that |fb(S) −
f (S)| ≤ for any S ⊆ Z, let product i ∈ L spread according to a diffusion network
Gi = (V, Ei ), and i∗ = argmaxi∈L |Ei |. Then, we have:
Theorem 4 Suppose fb is evaluated up to error = δ/16 with ConTinEst. For influence
maximization with uniform cost,
Algorithm 2 (with
ρ = 0) outputs a solution G with f (G) ≥
|Ei∗ |+|V|
|L||V|
1−2δ
e
+ δ3
.
3 f (O) in expected time O
δ2
The parameter δ introduces a tradeoff between the approximation guarantee and the
runtime: larger δ decreases the approximation ratio but results in fewer influence evaluations. Moreover, the runtime has a linear dependence on the network size and the number
of products to propagate (ignoring some small logarithmic terms) and, as a consequence,
the algorithm is scalable to large networks.
Analysis. Suppose G = {g1 , . . . , g|G| } in the order of selection, and let Gt = {g1 , . . . , gt }.
Let Ct denote all those elements in O \ G that satisfy the following: they are still feasible
before selecting the t-th element gt but are infeasible after selecting gt . Equivalently, Ct are
all those elements j ∈ O \ G such that (1) j ∪ Gt−1 does not violate the matroid constraints
but (2) j ∪ Gt violates the matroid constraints. In other words, we can think of Ct as the
optimal elements “blocked” by gt . Then, we proceed as follows.
S
By the property of the intersection of matroids, the size of the prefix ti=1 Ct is bounded
by P t. As a consequence of this property, for any Q ⊆ Z, the sizes of any two maximal
independent subsets T1 and T2 of Q can only differ by a multiplicative factor at most P .
This can be realized with the following argument. First, note that any element z ∈ T1 \ T2 ,
17
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
{z} ∪ T2 violates at least one of the matroid constraints since T2 is maximal. Then, let
{Vi }1≤i≤P denote all elements in T1 \ T2 that violate the i-th matroid, and partition T1 ∩ T2
arbitrarily among these Vi ’s so that they cover T1 . In this construction, the size of each Vi
must be at most |T2 |, since otherwise by the Exchange axiom, there would exist z ∈ Vi \ T2
that can be added to T2 , without violating the i-th matroid, leading to a contradiction.
Therefore, |T1 | is at most P times |T2 |.
St
Next, we apply the above property as follows. Let Q be the union of Gt and
St i=1 Ct .
t
On one hand, G is a maximal independent subset of Q, since no element in i=1
S Ct can
be added to Gt without violating the matroid constraints. On the other hand,S ti=1 Ct is
an independent subset of Q, since it is part of the optimal solution. Therefore, ti=1 Ct has
size at most P times |Gt |, which is P t. Note that the properties of matroids are crucial for
this analysis, which justifies our formulation using matroids. In summary, we have
Pt
Claim 1
i=1 |Ci | ≤ P t, for t = 1, . . . , |G|.
Now, we consider the marginal gain of each element in Ct associated with gt . First,
suppose gt is selected at the threshold τt > 0. Then, any j ∈ Ct has marginal gain bounded
by (1 + δ)τt + 2, since otherwise j would have been selected at a larger threshold before τt
by the greedy criterion. Second, suppose gt is selected at the threshold wL+1 = 0. Then,
any j ∈ Ct has marginal gain approximately bounded by Nδ d. Since the greedy algorithm
must pick g1 with fb(g1 ) = d and d ≤ f (g1 ) + , any j ∈ Ct has marginal gain bounded by
δ
N f (G) + O(). Putting everything together we have:
Claim 2 Suppose gt is selected at the threshold τt . Then f (j|Gt−1 ) ≤ (1+δ)τt +4+ Nδ f (G)
for any j ∈ Ct .
Since the evaluation of the marginal gain of gt should be at least τt , this claims essentially
indicates that the marginal gain of j is approximately bounded by that of gt .
Since there are not many elements in Ct (Claim 1) and the marginal gain of each of its
elements is not much larger than that of gt (Claim 2), we can conclude that the marginal
S|G|
gain of O \ G = i=1 Ct is not much larger than that of G, which is just f (G).
Claim 3 The marginal gain of O \ G satisfies
X
f (j|G) ≤ [(1 + δ)P + δ]f (G) + (6 + 2δ)P |G|.
j∈O\G
P
Finally, since by submodularity, f (O) ≤ f (O ∪ G) ≤ f (G) + j∈O\G f (j|G), Claim 3
shows that f (G) is close to f (O) up to a multiplicative factor roughly (1 + P ) and additive
factor O(P |G|). Given that f (G) > |G|, it leads to roughly a 1/3-approximation for our
influence maximization problem by setting = δ/16 when evaluating fb with ConTinEst.
Combining the above analysis and the runtime of the influence estimation algorithm, we
have our final guarantee in Theorem 4. Appendix D.1 presents the complete proofs.
4.2.2 General Case
In this section, we consider the general case, in which users may have different associated
costs. Recall that this case corresponds to the problem defined by Equation (19) with P = 1
matroid constraints and k = |L| group-knapsack constraints. Here, we will show that there
is a step in Algorithm 1 which outputs a solution Sρ that is a good approximation.
18
Continuous-Time Influence Maximization of Multiple Items
Intuition. The key idea behind Algorithm 1 and Algorithm 2 is simple: spend the budgets
efficiently and spend them as much as possible. By spending them efficiently, we mean to
only select those elements whose density ratio between the marginal gain and the cost is
above the threshold ρ. That is, we assign product i to user j only if the assignment leads
to large marginal gain without paying too much. By spending the budgets as much as
possible, we mean to stop assigning product i only if its budget is almost exhausted or no
more assignments are possible without violating the matroid constraints. Here we make use
of the special structure of the knapsack constraints on the budgets: each constraint is only
related to the assignment of the corresponding product and its budget, so that when the
budget of one product is exhausted, it does not affect the assignment of the other products.
In the language of submodular optimization, the knapsack constraints are on a partition
Zi∗ of the ground set and the objective function is a sum of submodular functions over the
partition.
However, there seems to be a hidden contradiction between spending the budgets efficiently and spending them as much as possible. On one hand, efficiency means the density
ratio should be large, so the threshold ρ should be large; on the other hand, if ρ is large,
there are just a few elements that can be considered, and thus the budget might not be
exhausted. After all, if we set ρ to be even larger than the maximum possible value, then
no element is considered and no gain is achieved. In the other extreme, if we set ρ = 0 and
consider all the elements, then a few elements with large costs may be selected, exhausting
all the budgets and leading to a poor solution.
Fortunately, there exists a suitable threshold ρ that achieves a good tradeoff between
the two and leads to a good approximation. On one hand, the threshold is sufficiently small,
so that the optimal elements we abandon (i.e., those with low-density ratio) have a total
gain at most a fraction of the optimum; on the other hand, it is also sufficiently large, so
that the elements selected are of high quality (i.e., of high-density ratio), and we achieve
sufficient gain even if the budgets of some items are exhausted.
Theorem 5 Suppose fb is evaluated up to error = δ/16 with ConTinEst. In Algorithm 1,
there exists a ρ such that
f (Sρ ) ≥
max {ka , 1}
f (O)
(2|L| + 2)(1 + 3δ)
where ka is the number of active knapsack constraints:
ka = |{i : Sρ ∪ {z} 6∈ F, ∀z ∈ Zi∗ }| .
e
The expected running time is O
|Ei∗ |+|V|
δ2
+
|L||V|
δ4
.
1
Importantly, the approximation factor improves over the best known guarantee P +2k+1
=
1
2|L|+2 for efficiently maximizing submodular functions over P matroids and k general knapsack constraints. Moreover, since the runtime has a linear dependence on the network size,
the algorithm easily scales to large networks. As in the uniform cost case, the parameter δ
introduces a tradeoff between the approximation and the runtime.
19
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
Analysis. The analysis follows the intuition. Pick ρ =
solution, and define
2f (O)
P +2k+1 ,
where O is the optimal
O− := {z ∈ O \ Sρ : f (z|Sρ ) < c(z)ρ + 2} ,
O+ := {z ∈ O \ Sρ : z 6∈ O− } .
Note that, by submodularity, O− is a superset of the elements in the optimal solution that
we abandon due to the density threshold and, by construction, its marginal gain is small:
f (O− |Sρ ) ≤ ρc(O− ) + O(|Sρ |) ≤ kρ + O(|Sρ |),
where the small additive term O(|Sρ |) is due to inexact function evaluations. Next, we
proceed as follows.
First, if no knapsack constraints are active, then the algorithm runs as if there were no
knapsack constraints (but only on elements with density ratio above ρ). Therefore, we can
apply the same argument as in the case of uniform cost (refer to the analysis up to Claim 3
in Section 4.2.1); the only caveat is that we apply the argument to O+ instead of O \ Sρ .
Formally, similar to Claim 3, the marginal gain of O+ satisfies
f (O+ |Sρ ) ≤ [(1 + δ)P + δ]f (Sρ ) + O(P |Sρ |),
where the small additive term O(P |Sρ |) is due to inexact function evaluations. Using
1
that f (O) ≤ f (Sρ ) + f (O− |Sρ ) + f (O+ |Sρ ), we can conclude that Sρ is roughly a P +2k+1
approximation.
Second, suppose ka > 0 knapsack constraints are active and the algorithm discovers that
the budget of product i is exhausted when trying to add element z to the set Gi = G∩Zi∗ of
selected elements at that time. Since c(Gi ∪ {z}) > 1 and each of these elements has density
above ρ, the gain of Gi ∪ {z} is above ρ. However, only Gi is included in our final solution,
so we need to show that the marginal gain of z is not large compared to that of Gi . To do
so, we first realize that the algorithm greedily selects elements with marginal gain above
a decreasing threshold wt . Then, since z is the last element selected and Gi is nonempty
(otherwise adding z will not exhaust the budget), the marginal gain of z must be bounded
by roughly that of Gi , which is at least roughly 21 ρ. Since this holds for all active knapsack
ka
constraints, then the solution has value at least k2a ρ, which is an P +2k+1
-approximation.
Finally, combining both cases, and setting k = |L| and P = 1 as in our problem, we
have our final guarantee in Theorem 5. Appendix D.2 presents the complete proofs.
5. Experiments on Synthetic and Real Data
In this section, we first evaluate the accuracy of the estimated influence given by ConTinEst and then investigate the performance of influence maximization on synthetic and
real networks by incorporating ConTinEst into the framework of BudgetMax. We show
that our approach significantly outperforms the state-of-the-art methods in terms of both
speed and solution quality.
20
Continuous-Time Influence Maximization of Multiple Items
−3
x 10
0.08
200
relative error
100
relative error
0.06
150
Influence
Core-perphery
8
0.04
0.02
50
NS
ConTinEst
0
2
4
6
8
0 2
10
10
T
6
4
2
3
10
#samples
0
5 10
4
10
20
30
#labels
40
50
20
30
#labels
40
50
20
30
#labels
40
50
−3
x 10
20
0.08
8
10
0.04
0.02
5
NS
ConTinEst
0
relative error
0.06
relative error
Influence
Random
15
2
4
6
8
0 2
10
10
T
6
4
2
3
10
#samples
0
4
10
10
−3
x 10
9
0.08
8
8
6
5
4
0.04
0.02
3
2
1
relative error
0.06
relative error
Influence
Hierarchy
7
NS
ConTinEst
2
4
6
8
T
(a) Influence vs. time
10
0 2
10
6
4
2
3
10
#samples
4
10
(b) Influence vs. #samples
0
10
(c) Error vs. #labels
Figure 3: Influence estimation for core-periphery, random, and hierarchical networks with
1,024 nodes and 2,048 edges. Column (a) shows estimated influence by NS (near ground
truth), and ConTinEst for increasing time window T ; Column (b) shows ConTinEst’s
relative error against number of samples with 5 random labels and T = 10; Column (c) reports ConTinEst’s relative error against the number of random labels with 10,000 random
samples and T = 10.
21
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
5.1 Experiments on Synthetic Data
We generate three types of Kronecker networks (Leskovec et al., 2010) which are synthetic
networks generated by a recursive Kronecker product of a base 2-by-2 parameter matrix
with itself to generate self-similar graphs. By tuning the base parameter matrix, we are
able to generate the Kronecker networks which can mimic different structural properties
of many real networks. In the following, we consider networks of three different types of
structures: (i ) core-periphery networks (parameter matrix: [0.9 0.5; 0.5 0.3]), which mimic
the information diffusion traces in real-world networks (Gomez-Rodriguez et al., 2011), (ii )
random networks ([0.5 0.5; 0.5 0.5]), typically used in physics and graph theory (Easley
and Kleinberg, 2010) and (iii ) hierarchical networks ([0.9 0.1; 0.1 0.9]) (Clauset et al.,
2008). Next, we assign a pairwise transmission function for every directed edge in each
type of network and set its parameters at random. In our experiments, we use the Weibull
distribution from (Aalen et al., 2008),
β t β−1 −(t/α)β
f (t; α, β) =
e
, t ≥ 0,
(20)
α α
where α > 0 is a scale parameter and β > 0 is a shape parameter. The Weibull distribution
(Wbl) has often been used to model lifetime events in survival analysis, providing more
flexibility than an exponential distribution. We choose α and β from 0 to 10 uniformly at
random for each edge in order to have heterogeneous temporal dynamics. Finally, for each
type of Kronecker network, we generate 10 sample networks, each of which has different α
and β chosen for every edge.
5.1.1 Influence Estimation
To the best of our knowledge, there is no analytical solution to the influence estimation
given Weibull transmission function. Therefore, we compare ConTinEst with the Naive
Sampling (NS) approach by considering the highest degree node in a network as the source,
and draw 1,000,000 samples for NS to obtain near ground truth. In Figure 3, Column (a)
compares ConTinEst with the ground truth provided by NS at different time window
T , from 0.1 to 10 in networks of different structures. For ConTinEst, we generate up
to 10,000 random samples (or sets of random waiting times), and 5 random labels in the
inner loop. In all three networks, estimation provided by ConTinEst fits the ground truth
accurately, and the relative error decreases quickly as we increase the number of samples
and labels (Column (b) and Column (c)). For 10,000 random samples with 5 random labels,
the relative error is smaller than 0.01.
5.1.2 Influence Maximization with Uniform Cost
In this section, we first evaluate the effectiveness of ConTinEst to the classic influence
maximization problem where we have only one product to assign with the simple cardinality
constraint on the users. We compare to other influence maximization methods developed
based on discrete-time diffusion models: traditional greedy by (Kempe et al., 2003), with
discrete-time Linear Threshold Model (LT) and Independent Cascade Model (IC) diffusion
models, and the heuristic methods SP1M, PMIA and MIA-M by (Chen et al., 2009, 2010a,
2012). For Influmax, since it only supports exponential pairwise transmission functions,
22
400
ConTinEst(Wbl)
ConTinEst(Exp)
Greedy(IC)
SP1M
PMIA
MIA-M
400
350
300
influence
300
350
250
200
150
200
250
200
150
150
100
50
50
50
10
20
30
40
0
0
50
10
20
#sources
400
ConTinEst(Wbl)
ConTinEst(Exp)
Greedy(IC)
SP1M
PMIA
MIA-M
400
350
350
300
influence
300
30
40
0
0
50
10
20
#sources
250
200
150
30
40
50
4
5
#sources
300
ConTinEst(Wbl)
ConTinEst(Exp)
Greedy(IC)
SP1M
PMIA
MIA-M
ConTinEst(Wbl)
ConTinEst(Exp)
Greedy(IC)
SP1M
PMIA
MIA-M
250
200
250
influence
450
influence
ConTinEst(Wbl)
ConTinEst(Exp)
Greedy(IC)
SP1M
PMIA
MIA-M
250
100
100
0
0
Influence by Time
300
ConTinEst(Wbl)
ConTinEst(Exp)
Greedy(IC)
SP1M
PMIA
MIA-M
influence
450
influence
Influence by Size
Continuous-Time Influence Maximization of Multiple Items
200
150
150
100
100
100
50
50
50
0
0
1
2
3
4
0
0
5
1
2
T
3
4
0
0
5
1
2
T
(a) Core-periphery
3
T
(b) Random
(c) Hierarchy
Figure 4: Influence σ(A, T ) achieved by varying number of sources |A| and observation
window T on the networks of different structures with 1,024 nodes, 2,048 edges and heterogeneous Weibull transmission functions. Top row: influence against #sources by T = 5;
Bottom row: influence against the time window T using 50 sources.
Uniform Cost
x 10
BudgetMax
GreedyDegree
Random
x 10
2.5
influence
influence
1.5
4
3
1
Uniform Cost
1.5
1
16
32
# products
(a) By products
64
0
4
1.5
1
BudgetMax
GreedyDegree
Random
0.5
0.5
8
x 10
2
2
0.5
0
4
Uniform Cost
4
2.5
BudgetMax
GreedyDegree
Random
influence
4
2
8
12
16
product constraints
20
(b) By product constraints
0
2
4
6
8
user constraints
10
(c) By user constraints
Figure 5: Over the 64 product-specific diffusion networks, each of which has 1,048,576
nodes, the estimated influence (a) for increasing the number of products by fixing the
product-constraint at 8 and user-constraint at 2; (b) for increasing product-constraint by
user-constraint at 2; and (c) for increasing user-constraint by fixing product-constraint at
8. For all experiments, we have T = 5 time window.
we fit an exponential distribution per edge by NetRate (Gomez-Rodriguez et al., 2011).
Furthermore, Influmax is not scalable. When the average network density (defined as the
average degree per node) of the synthetic networks is ∼ 2.0, the run time for Influmax
23
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
is more than 24 hours. In consequence, we present the results of ConTinEst using fitted
exponential distributions (Exp). For the discrete-time IC model, we learn the infection
probability within time window T using Netrapalli’s method (Netrapalli and Sanghavi,
2012). The learned pairwise infection probabilities are also served for SP1M and PMIA,
which approximately calculate the influence based on the IC model. For the discretetime LT model, we set the weight of each incoming edge to a node u to the inverse of
its in-degree, as in previous work (Kempe et al., 2003), and choose each node’s threshold
uniformly at random. The top row of Figure 4 compares the expected number of infected
nodes against the source set size for different methods. ConTinEst outperforms the rest,
and the competitive advantage becomes more dramatic the larger the source set grows.
The bottom row of Figure 4 shows the expected number of infected nodes against the time
window for 50 selected sources. Again, ConTinEst performs the best for all three types
of networks.
Next, using ConTinEst as a subroutine for influence estimation, we evaluate the performance of BudgetMax with the uniform-cost constraints on the users. In our experiments
we consider up to 64 products, each of which diffuses over one of the above three different
types of Kronecker networks with ∼ one million nodes. Further, we randomly select a subset
of 512 nodes VS ⊆ V as our candidate target users, who will receive the given products,
and evaluate the potential influence of an allocation over the underlying one-million-node
networks. For BudgetMax, we set the adaptive threshold δ to 0.01 and the cost per user
and product to 1. For ConTinEst, we use 2,048 samples with 5 random labels on each
of the product-specific diffusion networks. We repeat our experiments 10 times and report
the average performance.
We compare BudgetMax with a nodes’ degree-based heuristic, which we refer to as
GreedyDegree, where the degree is treated as a natural measure of influence, and a baseline
method, which assigns the products to the target nodes randomly. We opt for the nodes’
degree-based heuristic since, in practice, large-degree nodes, such as users with millions of
followers in Twitter, are often the targeted users who will receive a considerable payment if
he (she) agrees to post the adoption of some products (or ads) from merchants. GreedyDegree proceeds as follows. It first sorts the list of all pairs of products i and nodes j ∈ VS in
descending order of node-j’s degree in the diffusion network associated to product i. Then,
starting from the beginning of the list, it considers each pair one by one: if the addition of
the current pair to the existing solution does not violate the predefined matroid constraints,
it is added to the solution, and otherwise, it is skipped. This process continues until the end
of the list is reached. In other words, we greedily assign products to nodes with the largest
degree. Due to the large size of the underlying diffusion networks, we do not apply other
more expensive node centrality measures such as the clustering coefficient and betweenness.
Figure 5 summarizes the results. Panel (a) shows the achieved influence against number
of products, fixing the budget per product to 8 and the budget per user to 2. As the number
of products increases, on the one hand, more and more nodes become assigned, so the total
influence will increase. Yet, on the other hand, the competition among products for a
few existing influential nodes also increases. GreedyDegree achieves a modest performance,
since high degree nodes may have many overlapping children. In contrast, BudgetMax, by
taking the submodularity of the problem, the network structure and the diffusion dynamics
of the edges into consideration, achieves a superior performance, especially as the number
24
Continuous-Time Influence Maximization of Multiple Items
of product (i.e., the competition) increases. Panel (b) shows the achieved influence against
the budget per product, considering 64 products and fixing the budget per user to 2. We
find that, as the budget per product increases, the performance of GreedyDegree tends to
flatten and the competitive advantage of BudgetMax becomes more dramatic. Finally,
Panel (c) shows the achieved influence against the budget per user, considering 64 products
and fixing the budget per product to 8. We find that, as the budget per user increases, the
influence only increases slowly. This is due to the fixed budget per product, which prevents
additional new nodes to be assigned. This meets our intuition: by making a fixed number
of people watching more ads per day, we can hardly boost the popularity of the product.
Additionally, even though the same node can be assigned to more products, there is hardly
ever a node that is the perfect source from which all products can efficiently spread.
5.1.3 Influence Maximization with Non-Uniform Cost
In this section, we evaluate the performance of BudgetMax under non-uniform cost constraints, using again ConTinEst as a subroutine for influence estimation. Our designing of
user-cost aim to mimic a real scenario, where advertisers pay much more money to celebrities with millions of social network followers than to normal citizens. To do so, we let
1/n
ci ∝ di where ci is the cost, di is the degree, and n ≥ 1 controls the increasing speed of
cost with respect to the degree. In our experiments, we use n = 3 and normalize ci to be
within (0, 1]. Moreover, we set the product-budget to a base value from 1 to 10 and add a
random adjustment drawn from a uniform distribution U (0, 1).
We compare our method to two modified versions of the above mentioned nodes’ degreebased heuristic GreedyDegree and to the same baseline method. In the first modified
version of the heuristic, which we still refer to as GreedyDegree, takes both the degree
and the corresponding cost into consideration. In particular, it sorts the list of all pairs of
products i and nodes j ∈ VS in descending order of degree-cost ratio dj /cj in the diffusion
network associated to product i, instead of simply the node-j’s degree, and then proceeds
similarly as before. In the second modified version of the heuristic, which we refer as
GreedyLocalDegree, we use the same degree-cost ratio but allow the target users to be
partitioned into distinct groups (or communities) and pick the most cost-effective pairs
within each group locally instead. Figure 6 compares the performance of our method with
the competing methods against four factors: (a) the number of products, (b) the budget
per product, (c) the budget per user and (d) the time window T , while fixing the other
factors. In all cases, BudgetMax significantly outperforms the other methods, and the
achieved influence increases monotonically with respect to the factor value, as one may have
expected. In addition, in Figure 6(e), we study the effect of the Laminar matroid combined
with group knapsack constraints, which is the most general type of constraint we handle in
this paper (refer to Section 3.4). The selected target users are further partitioned into K
groups randomly, each of which has, Qi , i = 1 . . . K, limit which constrains the maximum
allocations allowed in each group. In practical scenarios, each group might correspond to
a geographical community or organization. In our experiment, we divide the users into 8
equal-size groups and set Qi = 16, i = 1 . . . K to indicate that we want a balanced allocation
in each group. Figure 6(e) shows the achieved influence against the budget per user for
K = 8 equally-sized groups and Qi = 16, i = 1 . . . K. In contrast to Figure 6(b), the total
25
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
x 10
1.5
8000
influence
influence
10000
4
2
BudgetMax
GreedyDegree
Random
6000
4000
Non−uniform cost
4
2
BudgetMax
GreedyDegree
Random
1
0.5
8
16
32
# products
64
(a) By products
8
10
0
1
1.5
2
2.5
product budget
3
(b) By product budgets
Non−uniform cost
1
0
2
4
6
8
user constraints
10
(c) By user constraints
Non−uniform cost
12000
BudgetMax
GreedyDegree
Random
10000
6
10
influence
influence
Non−uniform cost
BudgetMax
GreedyDegree
Random
0.5
2000
0
4
x 10
1.5
influence
Non−uniform cost
12000
4
10
8000
6000
4000
BudgetMax
GreedyDegree
GreedyLocalDegree
Random
2000
2
10
2
5
time
10
15
(d) By time
0
2
4
6
8
user constraints
10
(e) By group limits
Figure 6: Over the 64 product-specific diffusion networks, each of which has a total
1,048,576 nodes, the estimated influence (a) for increasing the number of products by fixing
the product-budget at 1.0 and user-constraint at 2; (b) for increasing product-budget by
fixing user-constraint at 2; (c) for increasing user-constraint by fixing product-budget at
1.0; (d) for different time window T; and (e) for increasing user-constraint with group-limit
16 by fixing product-budget at 1.0.
26
Continuous-Time Influence Maximization of Multiple Items
Uniform cost
Uniform cost
70
1
60
0.8
0.6
time(s)
accuracy
50
0.4
40
30
20
0.2
10
BudgetMax(Adaptive)
BudgetMax(Lazy)
0
1 5 10 15 20 25 30 35 40 45 50
δ
(a) δ vs. accuracy
BudgetMax(Adaptive)
BudgetMax(Lazy)
0
1 5 10 15 20 25 30 35 40 45 50
δ
(b) δ vs. time
Figure 7: The relative accuracy and the run-time for different threshold parameter δ.
estimated influence does not increase significantly with respect to the budget (i.e., number
of slots) per user. This is due to the fixed budget per group, which prevents additional new
nodes to be assigned, even though the number of available slots per user increases.
5.1.4 Effects of Adaptive Thresholding
In Figure 7, we investigate the impact that the threshold value δ has on the accuracy and
runtime of our adaptive thresholding algorithm and compare it with the lazy evaluation
method. Note that the performance and runtime of lazy evaluation do not change with
respect to δ because it does not depend on it. Panel (a) shows the achieved influence against
the threshold δ. As expected, the larger the δ value, the lower the accuracy. However, our
method is relatively robust to the particular choice of δ since its performance is always over
a 90-percent relative accuracy even for large δ. Panel (b) shows the runtime against the
threshold δ. In this case, the larger the δ value, the lower the runtime. In other words,
Figure 7 verifies the intuition that δ is able to trade off the solution quality of the allocation
with the runtime time.
5.1.5 Scalability
In this section, we start with evaluating the scalability of the proposed algorithms on the
classic influence maximization problem where we only have one product with the cardinality
constraint on the users.
We compare it to the state-of-the-art method Influmax (Gomez-Rodriguez and Schölkopf,
2012) and the Naive Sampling (NS) method in terms of runtime for the continuous-time
influence estimation and maximization. For ConTinEst, we draw 10,000 samples in the
outer loop, each having 5 random labels in the inner loop. We plug ConTinEst as a subroutine into the classic greedy algorithm by (Nemhauser et al., 1978). For NS, we also draw
10,000 samples. The first two experiments are carried out in a single 2.4GHz processor.
27
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
> 24 hours
5
10
4
2
10
time(s)
time(s)
time(s)
3
3
10
2
10
1
10
1
1
10
ConTinEst
0
2
3
4
NS
Influmax
5 6 7
#sources
8
10
ConTinEst
0
10
9 10
1
3
10
2
10
10
1
4
10
10
> 24 hours
5
10
4
10
10
> 24 hours
5
10
2
(a) Core-Periphery
3
4
NS
Influmax
5 6 7
#sources
8
ConTinEst
0
10
9 10
1
2
(b) Random
3
4
NS
Influmax
5 6 7
#sources
8
9 10
(c) Hierarchy
Figure 8: Runtime of selecting increasing number of sources on Kronecker networks of 128
nodes and 320 edges with T = 10.
> 24 hours
5
10
4
3
2
10
time(s)
time(s)
time(s)
4
10
10
3
10
2
10
1
10
1
1
10
ConTinEst
0
2
2.5
NS
3 3.5
density
Influmax
4
(a) Core-Periphery
4.5
5
10
ConTinEst
0
10
1.5
3
10
2
10
10
> 24 hours
5
10
4
10
10
1.5
> 24 hours
5
10
2
2.5
NS
3 3.5
density
(b) Random
Influmax
4
4.5
5
ConTinEst
0
10
1.5
2
2.5
NS
3 3.5
density
Influmax
4
4.5
5
(c) Hierarchy
Figure 9: Runtime of selecting 10 sources in networks of 128 nodes with increasing density
by T = 10.
Figure 8 compares the performance of increasingly selecting sources (from 1 to 10) on
small Kronecker networks. When the number of selected sources is 1, different algorithms
essentially spend time estimating the influence for each node. ConTinEst outperforms
other methods by order of magnitude and for the number of sources larger than 1, it can
efficiently reuse computations for estimating influence for individual nodes. Dashed lines
mean that a method did not finish in 24 hours, and the estimated run time is plotted.
Next, we compare the run time for selecting 10 sources with increasing densities (or the
number of edges) in Figure 9. Again, Influmax and NS are order of magnitude slower due
to their respective exponential and quadratic computational complexity in network density.
In contrast, the run time of ConTinEst only increases slightly with the increasing density
since its computational complexity is linear in the number of edges. We evaluate the speed
on large core-periphery networks, ranging from 100 to 1,000,000 nodes with density 1.5
in Figure 10. We report the parallel run time only for ConTinEst and NS (both are
implemented by MPI running on 192 cores of 2.4Ghz) since Influmax is not scalable. In
contrast to NS, the performance of ConTinEst increases linearly with the network size
and can easily scale up to one million nodes.
28
Continuous-Time Influence Maximization of Multiple Items
6
10
> 48 hours
5
time(s)
10
4
10
3
10
2
10
ConTinEst
NS
2
10
3
10
4
10
#nodes
5
10
6
10
Figure 10: For core-periphery networks by T = 10, runtime of selecting 10 sources with
increasing network size from 100 to 1,000,000 by fixing 1.5 network density.
Uniform Cost
2
Uniform Cost
10
50
time(s)
time(s)
45
1
10
40
0
10
4
8
16
32
# products
64
(a) Speed by products
35 4
10
5
10
#nodes
(b) Speed by nodes
6
10
Figure 11: Over the 64 product-specific diffusion networks, each of which has 1,048,576
nodes, the runtime (a) of allocating increasing number of products and (b) of allocating 64
products to 512 users on networks of varying size. or all experiments, we have T = 5 time
window and fix product-constraint at 8 and user-constraint at 2.
29
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
3.5
ConTinEst
IC
LT
SP1M
PMIA
3
MAE
2.5
2
1.5
1
0.5
10
20
30
T
40
50
Figure 12: In MemeTracker dataset, comparison of the accuracy of the estimated influence
in terms of mean absolute error.
Finally, we investigate the performance of BudgetMax in terms of runtime when using ConTinEst as subroutine to estimate the influence. We can precompute the data
structures and store the samples needed to estimate the influence function in advance.
Therefore, we focus only on the runtime for the constrained influence maximization algorithm. BudgetMax runs on 64 cores of 2.4Ghz by using OpenMP to accelerate the first
round of the optimization. We report the allocation time for increasing number of products
in Figure 11(a), which clearly shows a linear time complexity with respect to the size of
the ground set. Figure 11(b) evaluates the runtime of allocation by varying the size of the
network from 16,384 to 1,048,576 nodes.
5.2 Experiments on Real Data
In this section, we first quantify how well our proposed algorithm can estimate the true
influence in the real-world dataset. Then, we evaluate the solution quality of the selected
sources for influence maximization under different constraints. We have used the public
MemeTracker datasets (Leskovec et al., 2009), which contains more than 172 million news
articles and blog posts from 1 million mainstream media sites and blogs.
5.2.1 Influence Estimation
We first trace the flow of information from one site to another by using the hyperlinks among
articles and posts as in the work of Gomez-Rodriguez et al. (2011); Du et al. (2012). In
detail, we extracted 10,967 hyperlink cascades among top 600 media sites. We then evaluate
the accuracy of ConTinEst as follows. First, we repeatedly split all cascades into a 80%
training set and a 20% test set at random for five times. On each training set, we learn one
continuous-time model, which we use for ConTinEst, and a discrete-time model, which we
use for the competitive methods: IC, SP1M, PMIA and MIAM-M. For the continuous-time
30
Continuous-Time Influence Maximization of Multiple Items
80
60
70
60
40
influence
influence
50
30
50
40
30
20
ConTinEst
Greedy(IC)
SP1M
PMIA
MIAM-M
10
0
0
10
20
30
40
ConTinEst(Wbl)
Greedy(IC)
SP1M
PMIA
MIA-M
20
10
0
0
50
5
#sources
10
15
20
T
(a) Influence vs. #sources
(b) Influence vs. time
Figure 13: In MemeTracker dataset, (a) comparison of the influence of the selected nodes by
fixing the observation window T = 5 and varying the number sources, and (b) comparison
of the influence of the selected nodes by fixing the number of sources to 50 and varying the
time window.
model, we opt for NetRate (Gomez-Rodriguez et al., 2011) with exponential transmission
functions (fixing the shape parameter of the Weibull family to be one) to learn the diffusion
networks by maximizing the likelihood of the observed cascades. For the discrete-time
model, we learn the infection probabilities using the method by Netrapalli and Sanghavi
(2012).
Second, let C(u) be the set of all cascades where u was the source node. By counting the total number of distinct nodes infected before T in C(u), we can quantify the real
influence of node u up to time T . Thus, we can evaluate the quality of the influence estimation by computing the average (across nodes) Mean Absolute Error (MAE) between
the real and the estimated influence on the test set, which we show in Figure 12. Clearly,
ConTinEst performs the best statistically. Since the length of real cascades empirically
conforms to a power-law distribution, where most cascades are very short (2-4 nodes), the
gap of the estimation error is not too large. However, we emphasize that such accuracy
improvement is critical for maximizing long-term influence since the estimation error for individuals will accumulate along the spreading paths. Hence, any consistent improvement in
influence estimation can lead to significant improvement to the overall influence estimation
and maximization task, which is further confirmed in the following sections.
5.2.2 Influence Maximization with Uniform Cost
We first apply ConTinEst to the continuous-time influence maximization task with the
simple cardinality constraint on the users. We evaluate the influence of the selected nodes in
the same spirit as influence estimation: the true influence is calculated as the total number
of distinct nodes infected before T based on C(u) of the selected nodes. Figure 13 shows
that the selected sources given by ConTinEst achieve the best performance as we vary the
number of selected sources and the observation time window.
31
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
Next, we evaluate the performance of BudgetMax on cascades from Memetracker
traced from quotes which are short textual phrases spreading through the websites. Because
all published documents containing a particular quote are time-stamped, a cascade induced
by the same quote is a collection of times when the media site first mentioned it. In detail, we
use the public dataset released by Gomez-Rodriguez et al. (2013), which splits the original
Memetracker dataset into groups, each associated to a topic or real-world event. Each group
consists of cascades built from quotes which were mentioned in posts containing particular
keywords. We considered 64 groups, with at least 100,000 cascades, which play the role
of products. Therein, we distinguish well-known topics, such as “Apple” and “Occupy
Wall-Street”, or real-world events, such as the Fukushima nuclear disaster in 2013 and the
marriage between Kate Middleton and Prince William in 2011.
We then evaluate the accuracy of BudgetMax in the following way. First, we evenly
split each group into a training and a test set and then learn one continuous-time model and
a discrete-time model per group using the training sets. As previously, for the continuoustime model, we opt for NetRate (Gomez-Rodriguez et al., 2011) with exponential transmission functions, and for the discrete-time model, we learn the infection probabilities using
the method by Netrapalli and Sanghavi (2012), where the step-length is set to one. Second,
we run BudgetMax using both the continuous-time model and the discrete-time model.
We refer to BudgetMax with the discrete-time model as the Greedy(discrete) method.
Since we do have no ground-truth information about cost of each node, we focus our experiments using a uniform cost. Third, once we have found an allocation over the learned
networks, we evaluate the performance of the two methods using the cascades in the test
set as follows: given a group-node pair (i, j), let C(j) denote the set of cascades induced by
group i that contains node j. Then, we take the average number of nodes coming after j
for all the cascades in C(j) as a proxy of the average influence induced by assigning group
i to node j. Finally, the influence of an allocation is just the sum of the average influence
of each group-node pair in the solution. In our experiments, we randomly select 128 nodes
as our target users.
Figure 14 summarizes the achieved influence against four factors: (a) the number of
products, (b) the budget per product, (c) the budget per user and (d) the time window T,
while fixing the other factors. In comparison with the Greedy(IC) and a random allocation,
BudgetMax finds an allocation that indeed induces the largest diffusion in the test data,
with an average 20-percent improvement overall.
In the end, Figure 15 investigates qualitatively the actual allocations of groups (topics or real-world events; in red) and sites (in black). Here, we find examples that intuitively one could expect: “japantoday.com” is assigned to Fukushima Nuclear disaster or
“finance.yahoo.com” is assigned to “Occupy Wall-street”. Moreover, because we consider
several topics and real-world events with different underlying diffusion networks, the selected nodes are not only very popular media sites such as nytimes.com or cnn.com but also
several modest sites (Bakshy et al., 2011), often specialized or local, such as freep.com or
localnews8.com.
32
Continuous-Time Influence Maximization of Multiple Items
Uniform cost
influence
3000
Uniform cost
4000
BudgetMax
Greedy(discrete)
Random
3000
influence
4000
2000
1000
2000
1000
0
4
8
16
32
# products
0
4
64
(a) By products
3000
1200
2500
1000
2000
1500
BudgetMax
Greedy(discrete)
Random
8
16
32
# user constraints
(c) By user constraints
64
Uniform cost
1400
influence
influence
Uniform cost
1000
8
16
32
# product constraints
(b) By product constraints
3500
500
4
BudgetMax
Greedy(discrete)
Random
800
BudgetMax
Greedy(discrete)
Random
600
400
64
200
2
5
time
(d) By time
10
15
Figure 14: Over the inferred 64 product-specific diffusion networks, the true influence
estimated from separated testing data (a) for increasing the number of products by fixing
the product-constraint at 8 and user-constraint at 2; (b) for increasing product-constraint
by fixing user-constraint at 2; (c) for increasing user-constraint by fixing product-constraint
at 8; (d) for different time window T.
6. Conclusion
We have studied the influence estimation and maximization problems in the continuous-time
diffusion model. We first propose a randomized influence estimation algorithm ConTinEst,
which can scale up to networks of millions of nodes while significantly improves over previous state of the art methods in terms of the accuracy of the estimated influence. Once
we have a subroutine for efficient influence estimation in large networks, we then tackle
the problem of maximizing the influence of multiple types of products (or information) in
33
e
Up
ris
als
et
Al
-Q
ae
da
an
y Se
Sy
ri
Nav
om
om
Fukush
mlive.c
s8.c
m
es
m
co
.co
m
.co
dailybre
io
.
es
nda
il
0.
gsto
13
di
ar
tim
livin
al1
co
sta
n
me
wg
du
c
lo
hin
ele
lnew
loca
ima Nu
clear
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
c
Oc
y
up
e
nc
Pri
y.co
m
internationalentertainment.blogspot.com
obs
ve J
rotests
Europe Debt
om
dalje.c
m
h.co
kwc
g
.or
ub
h
s
m
co
icle
s.
art
w
ne
bf
m
w
fina
ntr
m
om
m
uk
.
l.co
te.com
daytondailynews.com
co
.co
ay
nd
om
.c
ne
m
ily.
oo.c
es
ily
.yah
ed
a
co
tim
oca
uts
a
n.
ny
tod
an
jap
rnel
advoca
ws
.c
om
ieg
o.c
om
free
p.co
m
hurriye
tdailyn
ews.co
m
cn
newark
da
nce
ce
bou
or
m'
Mass P
m
ng
ing
dd
e
sW
llia
Wi
Ste
eze.co
ba
a
W
tre
lls
Figure 15: The allocation of memes to media sites.
realistic continuous-time diffusion networks subject to various practical constraints: different products can have different diffusion structures; only influence within a given time
window is considered; each user can only be recommended a small number of products;
and each product has a limited campaign budget, and assigning it to users incurs costs.
We provide a novel formulation as a submodular maximization under an intersection of
matroid constraints and group-knapsack constraints, and then design an efficient adaptive
threshold greedy algorithm with provable approximation guarantees, which we call BudgetMax. Experimental results show that the proposed algorithm performs remarkably
better than other scalable alternatives in both synthetic and real-world datasets. There are
also a few interesting open problems. For example, when the influence is estimated using
ConTinEst, its error is a random variable. How does this affect the submodularity of the
influence function? Is there an influence maximization algorithm that has better tolerance
to the random error? These questions are left for future work.
34
Continuous-Time Influence Maximization of Multiple Items
Acknowledgments
Nan Du is supported by the Facebook Graduate Fellowship 2014-2015. Maria-Florina Balcan and Yingyu Liang are supported in part by NSF grant CCF-1101283 and CCF-0953192,
AFOSR grant FA9550-09-1-0538, ONR grant N00014-09-1-075, a Microsoft Faculty Fellowship, and a Raytheon Faculty Fellowship. Le Song is supported in part by NSF/NIH
BIGDATA 1R01GM108341, ONR N00014-15-1-2340, NSF IIS-1218749, NSF CAREER IIS1350983, Nvidia and Intel. Hongyuan Zha is supported in part by NSF/NIH BIGDATA
1R01GM108341, NSF DMS-1317424 and NSF CNS-1409635.
References
Odd Aalen, Oernulf Borgan, and Håkon K Gjessing. Survival and event history analysis: a
process point of view. Springer, 2008.
Ashwinkumar Badanidiyuru and Jan Vondrák. Fast algorithms for maximizing submodular
functions. In SODA. SIAM, 2014.
Eytan Bakshy, Jake M. Hofman, Winter A. Mason, and Duncan J. Watts. Everyone’s an
influencer: Quantifying influence on twitter. In WSDM, pages 65–74, 2011.
Christian Borgs, Michael Brautbar, Jennifer Chayes, and Brendan Lucier. Influence maximization in social networks: Towards an optimal algorithmic solution. arXiv preprint
arXiv:1212.0884, 2012.
Wei Chen, Yajun Wang, and Siyu Yang. Efficient influence maximization in social networks. In Proceedings of the 15th ACM SIGKDD international conference on Knowledge
discovery and data mining, pages 199–208. ACM, 2009.
Wei Chen, Chi Wang, and Yajun Wang. Scalable influence maximization for prevalent viral
marketing in large-scale social networks. In Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 1029–1038. ACM,
2010a.
Wei Chen, Yifei Yuan, and Li Zhang. Scalable influence maximization in social networks
under the linear threshold model. In Data Mining (ICDM), 2010 IEEE 10th International
Conference on, pages 88–97. IEEE, 2010b.
Wei Chen, Alex Collins, Rachel Cummings, Te Ke, Zhenming Liu, David Rincon, Xiaorui
Sun, Yajun Wang, Wei Wei, and Yifei Yuan. Influence maximization in social networks
when negative opinions may emerge and propagate. In SDM, pages 379–390. SIAM, 2011.
Wei Chen, Wei Lu, and Ning Zhang. Time-critical influence maximization in social networks
with time-delayed diffusion process. In AAAI, 2012.
Aaron Clauset, Cristopher Moore, and M.E.J. Newman. Hierarchical structure and the
prediction of missing links in networks. Nature, 453(7191):98–101, 2008.
35
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
Edith Cohen. Size-estimation framework with applications to transitive closure and reachability. Journal of Computer and System Sciences, 55(3):441–453, 1997.
Samik Datta, Anirban Majumder, and Nisheeth Shrivastava. Viral marketing for multiple
products. In Proceedings of the 2010 IEEE International Conference on Data Mining,
pages 118–127, 2010.
Nan Du, Le Song, Alex Smola, and Ming Yuan. Learning networks of heterogeneous influence. In Advances in Neural Information Processing Systems 25, pages 2789–2797,
2012.
Nan Du, Le Song, Manuel Gomez-Rodriguez, and Hongyuhan Zha. Scalable influence
estimation in continuous time diffusion networks. In Advances in Neural Information
Processing Systems 26, 2013a.
Nan Du, Le Song, Hyenkyun Woo, and Hongyuan Zha. Uncover topic-sensitive information
diffusion networks. In Artificial Intelligence and Statistics (AISTATS), 2013b.
David Easley and Jon Kleinberg. Networks, Crowds, and Markets: Reasoning About a
Highly Connected World. Cambridge University Press, 2010.
Satoru Fujishige. Submodular functions and optimization, volume 58. Elsevier Science
Limited, 2005.
Manuel Gomez-Rodriguez and Bernhard Schölkopf. Influence maximization in continuous
time diffusion networks. In Proceedings of the 29th International Conference on Machine
Learning, pages 313–320, 2012.
Manuel Gomez-Rodriguez, David Balduzzi, and Bernhard Schölkopf. Uncovering the temporal dynamics of diffusion networks. In Proceedings of the 28th International Conference
on Machine Learning, 2011.
Manuel Gomez-Rodriguez, Jure Leskovec, and Bernhard Schölkopf. Structure and Dynamics of Information Pathways in On-line Media. In Proceedings of the 6th International
Conference on Web Search and Web Data Mining, 2013.
Amit Goyal, Wei Lu, and Laks V. S. Lakshmanan. Celf++: optimizing the greedy algorithm
for influence maximization in social networks. In WWW (Companion Volume), pages 47–
48, 2011a.
Amit Goyal, Wei Lu, and Laks V. S. Lakshmanan. Simpath: An efficient algorithm for
influence maximization under the linear threshold model. In ICDM, pages 211–220,
2011b.
Dino Ienco, Francesco Bonchi, and Carlos Castillo. The meme ranking problem: Maximizing
microblogging virality. In ICDM Workshops, 2010.
David Kempe, Jon Kleinberg, and Éva Tardos. Maximizing the spread of influence through
a social network. In Proceedings of the ninth ACM SIGKDD international conference on
Knowledge discovery and data mining, pages 137–146. ACM, 2003.
36
Continuous-Time Influence Maximization of Multiple Items
Jure Leskovec, Andreas Krause, Carlos Guestrin, Christos Faloutsos, Jeanne VanBriesen,
and Natalie Glance. Cost-effective outbreak detection in networks. In P. Berkhin, R. Caruana, and X. Wu, editors, Conference on Knowledge Discovery and Data Mining, pages
420–429. ACM, 2007. URL http://doi.acm.org/10.1145/1281192.1281239.
Jure Leskovec, Lars Backstrom, and Jon Kleinberg. Meme-tracking and the dynamics of
the news cycle. In Proceedings of the 15th ACM SIGKDD international conference on
Knowledge discovery and data mining, pages 497–506. ACM, 2009.
Jure Leskovec, Deepayan Chakrabarti, Jon Kleinberg, Christos Faloutsos, and Zoubin
Ghahramani. Kronecker graphs: An approach to modeling networks. Journal of Machine Learning Research, 11(Feb):985–1042, 2010.
Wei Lu, Francesco Bonchi, Goyal Amit, and Laks V. S. Lakshmanan. The bang for the
buck: fair competitive viral marketing from the host perspective. In KDD, pages 928–936,
2013.
Ramasuri Narayanam and Amit A Nanavati. Viral marketing for product cross-sell through
social networks. In Machine Learning and Knowledge Discovery in Databases. 2012.
George Nemhauser, Laurence Wolsey, and Marshall Fisher. An analysis of the approximations for maximizing submodular set functions. Mathematical Programming, 14:265–294,
1978.
Praneeth Netrapalli and Sujay Sanghavi. Learning the graph of epidemic cascades. In
SIGMETRICS/PERFORMANCE, pages 211–222. ACM, 2012. ISBN 978-1-4503-1097-0.
Matthew Richardson and Pedro Domingos. Mining knowledge-sharing sites for viral marketing. In Proceedings of the eighth ACM SIGKDD international conference on Knowledge
discovery and data mining, pages 61–70. ACM, 2002.
Alexander Schrijver. Combinatorial Optimization, volume 24 of Algorithms and Combinatorics. Springer, 2003.
Tasuku Soma, Naonori Kakimura, Kazuhiro Inaba, and Ken-ichi Kawarabayashi. Optimal
budget allocation: Theoretical guarantee and efficient algorithm. In Proceedings of The
31st International Conference on Machine Learning, pages 351–359, 2014.
Tao Sun, Wei Chen, Zhenming Liu, Yajun Wang, Xiaorui Sun, Ming Zhang, and Chin-Yew
Lin. Participation maximization based on social influence in online discussion forums. In
Proceedings of the International AAAI Conference on Weblogs and Social Media, 2011.
Maxim Sviridenko. A note on maximizing a submodular set function subject to knapsack
constraint. Operations Research Letters, 32:41–43, 2004.
37
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
Appendix A. Naive Sampling Algorithm
The graphical model perspective described in Section 2.2 suggests a naive sampling (NS)
algorithm for approximating σ(A, T ):
n o
n
l
1. Draw n samples,
τji
, i.i.d. from the waiting time product distribution
(j,i)∈E l=1
Q
(j,i)∈E fji (τji );
2. For each sample
n o
l
τji
and for each node i, find the shortest path from source
n o
l
nodes to node i; count the number of nodes with gi
τji
≤ T;
(j,i)∈E
(j,i)∈E
3. Average the counts across n samples.
Although the naive sampling algorithm can handle arbitrary transmission function,
it is not scalable to networks with millions of nodes. We need to compute the shortest path for each node and each sample, which results in a computational complexity of
O(n|E| + n|V| log |V|) for a single source node. The problem is even more pressing in the
influence maximization problem, where we need to estimate the influence of source nodes
at different location and with increasing number of source nodes. To do this, the algorithm
needs to be repeated, adding a multiplicative factor of C|V| to the computational complexity (C is the number of nodes to select). Then, the algorithm becomes quadratic in the
network size. When the network size is in the order of thousands and millions, typical in
modern social network analysis, the naive sampling algorithm become prohibitively expensive. Additionally, we may need to draw thousands of samples (n is large), further making
the algorithm impractical for large-scale problems.
Appendix B. Least Label List
The notation “argsort((r1 , . . . , r|V| ), ascend)” in line 2 of Algorithm 3 means that we sort
the collection of random labels in ascending order and return the argument of the sort as
an ordered list.
• Node labeling :
e(0.2) < b(0.3) < d(0.4) < a(1.5) < c(1.8) < g(2.2) < f (3.7)
2
))252)
))553)1
))554)
3
))15))
• Neighborhoods:
N (c, 2) = {a, b, c, e}; N (c, 3) = {a, b, c, d, e, f };
2
2
4
555
1
))155)
• Least-label list:
r∗ (c) : (2, 0.2), (1, 0.3), (0.5, 1.5), (0, 1.8)
1
))35))
• Query: r∗ (c, 0.8) = r(a) = 1.5
))552)
Figure 16: Graph G = (V, E), edge weights {τji }(j,i)∈E , and node labeling {ri }i∈V with the
associated output from Algorithm 3.
38
Continuous-Time Influence Maximization of Multiple Items
Algorithm 3: Least Label List
Input: a reversed directed graph G = (V, E) with edge weights {τji }(j,i)∈E , a node
labeling {ri }i∈V
Output: A list r∗ (s) for each s ∈ V
1 for each s ∈ V do ds ← ∞, r∗ (s) ← ∅
2 for i in argsort((r1 , . . . , r|V| ), ascend) do
3
empty heap H ← ∅;
4
set all nodes except i as unvisited;
5
push (0, i) into heap H;
6
while H 6= ∅ do
7
pop (d∗ , s) with the minimum d∗ from H;
8
add (d∗ , ri ) to the end of list r∗ (s);
9
ds ← d∗ ;
10
for each unvisited out-neighbor j of s do
11
set j as visited;
12
if (d, j) in heap H then
13
Pop (d, j) from heap H;
14
Push (min {d, d∗ + τjs } , j) into heap H;
15
else if d∗ + τjs < dj then
16
Push (d∗ + τjs , j) into heap H;
Figure 16 shows an example of the Least-Label-List. The nodes from a to g are assigned
to exponentially distributed labels with mean one shown in each parentheses. Given a query
distance 0.8 for node c, we can binary-search its Least-label-list r∗ (c) to find that node a
belongs to this range with the smallest label r(a) = 1.5.
Appendix C. Theorem 1
Theorem 1 Sample the following number of sets of random transmission times
CΛ
2|V|
n > 2 log
α
where Λ := maxA:|A|≤C 2σ(A, T )2 /(m − 2) + 2V ar(|N (A, T )|)(m − 1)/(m − 2) + 2a/3,
|N (A, T )| 6 a, and for each set of random transmission times, sample m set of random
labels. Then we can guarantee that
|b
σ (A, T ) − σ(A, T )| 6
simultaneously for all A with |A| 6 C, with probability at least 1 − α.
Proof Let Sτ := |N (A, T )| for a fixed set of {τji } and then σ(A, T ) = Eτ [Sτ ]. The
randomized
estimator Sbτ = (m −
Pm ualgorithm with mbrandomizations produces an bunbiased
2
1)/( u=1 r∗ ) for Sτ , i.e., Er|τ [Sτ ] = Sτ , with variance Er|τ [(Sτ −Sτ ) ] = Sτ2 /(m−2) (Cohen,
1997).
39
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
Then Sbτ is also an unbiased estimator for σ(A, T ), since Eτ,r [Sbτ ] = Eτ Er|τ [Sbτ ] =
Eτ [Sτ ] = σ(A, T ). Its variance is
V ar(Sbτ ) := Eτ,r [(Sbτ − σ(A, T ))2 ] = Eτ,r [(Sbτ − Sτ + Sτ − σ(A, T ))2 ]
= Eτ,r [(Sbτ − Sτ )2 ] + 2 Eτ,r [(Sbτ − Sτ )(Sτ − σ(A, T ))] + Eτ,r [(Sτ − σ(A, T ))2 ]
= Eτ [Sτ2 /(m − 2)] + 0 + V ar(Sτ )
= σ(A, T )2 /(m − 2) + V ar(Sτ )(m − 1)/(m − 2)
Then using Bernstein’s inequality, we have, for our final estimator σ
b(A, T ) =
that
!
n2
Pr {|b
σ (A, T ) − σ(A, T )| > } 6 2 exp −
2V ar(Sbτ ) + 2a/3
1
n
Pn
l=1 Sτ l ,
b
(21)
where Sbτ < a 6 |V|.
Setting the right hand side of relation (21) to α, we have that, with probability 1 − α,
sampling the following number sets of random transmission times
2
2V ar(Sbτ ) + 2a/3
log
n>
2
α
2
2σ(A, T ) /(m − 2) + 2V ar(Sτ )(m − 1)/(m − 2) + 2a/3
2
=
log
2
α
we can guarantee that our estimator to have error |b
σ (A, T ) − σ(A, T )| 6 .
If we want to insure that |b
σ (A, T ) − σ(A, T )| 6 simultaneously hold for all A such
that |A| 6 C |V|, we can first use union bound with relation (21). In this case, we have
that, with probability 1 − α, sampling the following number sets of random transmission
times
2|V|
CΛ
n > 2 log
α
we can guarantee that our estimator to have error |b
σ (A, T ) − σ(A, T )| 6 for all A with
|A| 6 C. Note that we have define the constant Λ := maxA:|A|≤C 2σ(A, T )2 /(m − 2) +
2V ar(Sτ )(m − 1)/(m − 2) + 2a/3.
Appendix D. Complete Proofs for Section 4
D.1 Uniform Cost
In this section, we first prove a theorem for the general problem defined by Equation (19),
considering a normalized monotonic submodular function f (S) and general P (Theorem 7)
and k = 0, and then obtain the guarantee for our influence maximization problem (Theorem 4).
40
Continuous-Time Influence Maximization of Multiple Items
Suppose G = g1 , . . . , g|G| in the order of selection, and let Gt = {g1 , . . . , gt }. Let Ct
denote all those elements in O \ G that satisfy the following: they are still feasible before
selecting the t-th element gt but are infeasible after selecting gt . Formally,
Ct = z ∈ O \ G : {z} ∪ Gt−1 ∈ F, {z} ∪ Gt 6∈ F .
In the following, we will prove three claims and then use them to prove the Theorems 7
and 4. Recall that for any i ∈ Z and S ⊆ Z, the marginal gain of z with respect to S is
denoted as
f (z|S) := f (S ∪ {z}) − f (S)
and its approximation is denoted by
fb(z|S) = fb(S ∪ {z}) − fb(S).
Also, when |f (S) − fb(S)| ≤ for any S ⊆ Z, we have
|fb(z|S) − f (z|S)| ≤ 2
for any z ∈ Z and S ⊆ Z.
Pt
Claim 1.
i=1 |Ci | ≤ P t, for t = 1, . . . , |G|.
Proof We first show the following property about matroids: for any Q ⊆ Z, the sizes of
any two maximal independent subsets T1 and T2 of Q can only differ by a multiplicative
factor at most P . Here, T is a maximal independent subset of Q if and only if:
• T ⊆ Q;
• T ∈F =
TP
i=1 Ip ;
• T ∪ {z} 6∈ F for any z ∈ Q \ T .
To prove the property, note that for any element z ∈ T1 \ T2 , {z} ∪ T2 violates at least
one of the matroid constraints since T2 is maximal. Let {Vi }1≤i≤P denote all elements in
T1 \ T2 that violate the i-th matroid, and then partition T1 ∩ T2 using these Vi ’s so that they
cover T1 . Note that the size of each Vi must be at most that of T2 , since otherwise by the
Exchange axiom, there would exist z ∈ Vi \ T2 that can be added to T2 without violating
the i-th matroid, leading to a contradiction. Therefore, |T1 | is at most P times |T2 |.
St
Now we apply the property to prove the claim. Let Q be the union of Gt and
St i=1 Ct .
t
On one hand, G is a maximal independent subset of Q, since no element in i=1
S Ct can
be added to Gt without violating the matroid constraints. On the other hand,S ti=1 Ct is
an independent subset of Q, since it is part of the optimal solution. Therefore, ti=1 Ct has
size at most P times |Gt |, which is P t.
Claim 2. Suppose gt is selected at the threshold τt . Then, f (j|Gt−1 ) ≤ (1 + δ)τt + 4 +
δ
N f (G), ∀j ∈ Ct .
Proof First, consider τt > wL+1 = 0. Since gt is selected at the threshold τt , we have that
fb(gt |Gt−1 ) ≥ τt and thus f (gt |Gt−1 ) ≥ τt − 2. Any j ∈ Ct could have been selected at an
41
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
earlier stage, since adding j to Gt−1 would not have violated the constraint. However, since
j 6∈ Gt−1 , that means that fb(j|Gt−1 ) ≤ (1 + δ)τt . Then,
f (j|Gt−1 ) ≤ (1 + δ)τt + 2.
Second, consider τt = wL+1 = 0. For each j ∈ Ct , we have fb(j|G) <
greedy algorithm must pick g1 with fb(g1 ) = d and d ≤ f (g1 ) + , then
δ
N d.
Since the
δ
f (G) + 4.
N
The claim follows by combining the two cases.
f (j|G) <
Claim 3. The marginal gain of O \ G satisfies
X
f (j|G) ≤ [(1 + δ)P + δ]f (G) + (6 + 2δ)P |G|.
j∈O\G
Proof Combining Claim 1 and Claim 2, we have:
X
j∈O\G
f (j|G) =
|G|
X
X
f (j|G) ≤ (1 + δ)
|G|
X
|Ct |τt + δf (G) + 4
t=1
t=1 j∈Ct
≤ (1 + δ)
|G|
X
|G|
X
|Ct |
t=1
|Ct |τt + δf (G) + 4P |G|.
t=1
P|G|
P|G|
Further, t=1 |Ct |τt ≤ P t=1 τt by Claim 1Pand a technicalP
lemma (Lemma 6). Finally,
the claim follows from the fact that f (G) = t f (gt |Gt−1 ) ≥ t (τt − 2).
Pt
Lemma 6 If
i=1 σi−1 ≤ tPfor t = 1, . . . , K and ρi−1 ≥ ρi for i = 1, . . . , K − 1 with
P
K
ρi , σi ≥ 0, then K
i=1 ρi−1 .
i=1 ρi σi ≤
Proof Consider the linear program
V
=
s.t.
max
σ
K
X
ρi σi
i=1
t
X
σi−1 ≤ t, t = 1, . . . , K,
i=1
σi ≥ 0, i = 1, . . . , K − 1
with dual
W
=
s.t.
min
u
K−1
X
K
X
tut−1
i=1
ut ≥ ρi , i = 0, . . . , K − 1,
t=i
ut ≥ 0, t = 0, . . . , K − 1.
42
Continuous-Time Influence Maximization of Multiple Items
As ρi ≥ ρi+1 , thePsolution ui = ρi − ρP
i+1 , i = 0, . . . , K − 1 (where ρK = 0) is dual feaK
K
t(ρ
−
ρ
)
=
sible with value
t
i=1 ρi−1 . By weak linear programming duality,
t=1 P t−1
PK
K
i=1 ρi−1 .
i=1 ρi σi ≤ V ≤ W ≤
Theorem 7 Suppose we use Algorithm 2 to solve the problem defined by Equation (19)
with k = 0, using ρ = 0 and fb to estimate the function f , where |fb(S) − f (S)| ≤ for all
S ⊆ Z. It holds that the algorithm returns a greedy solution G with
f (G) ≥
1
4P |G|
f (O) −
(1 + 2δ)(P + 1)
P + cf
where O is the optimal solution, using O( Nδ log Nδ ) evaluations of fb.
Proof By submodularity and Claim 3, we have:
X
f (O) ≤ f (O ∪ G) ≤ f (G) +
f (j|G) ≤ (1 + δ)(P + 1)f (G) + (6 + 2δ)P |G|,
j∈O\G
which leads to the bound in the theorem.
Since there are O( 1δ log Nδ ) thresholds, and there are O(N ) evaluations at each threshold, the number of evaluations is bounded by O( Nδ log Nδ ).
Theorem 7 essentially shows f (G) is close to f (O) up to a factor roughly (1 + P ), which
then leads to the following guarantee for our influence maximization problem. Suppose product i ∈ L spreads according to diffusion network Gi = (V, Ei ), and let i∗ = argmaxi∈L |Ei |.
Theorem 4. In the influence maximization problem with uniform cost, Algorithm 2 (with
ρ
= 0) is able to
output a solution G that satisfies f (G) ≥ 1−2δ
3 f (O) in expected time
|Ei∗ |+|V|
|L||V|
e
O
.
+ 3
2
δ
δ
Proof In the influence maximization problem, the number of matroids is P = 2. Also note
that |G| ≤ f (G) ≤ f (O), which leads to 4|G| ≤ 4f (O). The approximation guarantee
then follows from setting ≤ δ/16 when using ConTinEst (Du et al., 2013a) to estimate
the influence.
The runtime is bounded as follows. In Algorithm 2, we need to estimate the marginal
gain of adding one more product to the current solution. In ConTinEst (Du et al., 2013a),
building the initial data structure takes time
1
|V|
2
O (|Ei∗ | log |V| + |V| log |V|) 2 log
δ
δ
and afterwards each function evaluation takes time
1
|V|
O
log log |V| .
log
δ2
δ
As there are O Nδ log Nδ evaluations where N = |L||V|, the runtime of our algorithm follows.
43
Du, Liang, Balcan, Gomez-Rodriguez, Zha and Song
D.2 General case
As in the previous section, we first prove that a theorem for the problem defined by Equation (19) with general normalized monotonic submodular function f (S) and general P
(Theorem 8), and then obtain the guarantee for our influence maximization problem (Theorem 5).
Theorem 8 Suppose Algorithm 1 uses fb to estimate the function f which satisfies |fb(S) −
f (S)| ≤ for all S ⊆ Z. Then, there exists a ρ such that
f (Sρ ) ≥
max {1, |Aρ |}
f (O) − 8|Sρ |
(P + 2k + 1)(1 + 2δ)
where Aρ is the set of active knapsack constraints:
Aρ = {i : Sρ ∪ {z} 6∈ F, ∀z ∈ Zi∗ }.
2
f (O). By submodularity, we
Proof Consider the optimal solution O and set ρ∗ = P +2k+1
h
i
2|Z|d
2d
have d ≤ f (O) ≤ |Z|d, so ρ ∈ P +2k+1 , P +2k+1 , and there is a run of Algorithm 2 with ρ
such that ρ∗ ∈ [ρ, (1 + δ)ρ]. In the following we consider this run.
Case 1 Suppose |Aρ | = 0. The key observation in this case is that since no knapsack
constraints are active, the algorithm runs as if there were only matroid constraints. Then,
the argument for matroid constraints can be applied. More precisely, let
O+ := {z ∈ O \ Sρ : f (z|Sρ ) ≥ c(z)ρ + 2}
O− := {z ∈ O \ Sρ : z 6∈ O+ } .
Note that all elements in O+ are feasible. Following the argument of Claim 3 in Theorem 7,
we have:
f (O+ |Sρ ) ≤ ((1 + δ)P + δ)f (Sρ ) + (4 + 2δ)P |Sρ |.
(22)
Also, by definition, the marginal gain of O− is:
f (O− |Sρ ) ≤ kρ + 2|O− | ≤ kρ + 2P |Sρ |,
(23)
where the last inequality follows from the fact that Sρ is a maximal independent subset, O−
is an independent subset of O ∪ Sρ , and the sizes of any two maximal independent subsets
in the intersection of P matroids can differ by a factor of at most P . Plugging (22)(23) into
f (O) ≤ f (O+ |Sρ ) + f (O− |Sρ ) + f (Sρ ), we obtain the bound
f (Sρ ) ≥
(6 + 2δ)P |Sρ |
f (O)
−
.
(P + 2k + 1)(1 + δ) (P + 1)(1 + δ)
Case 2 Suppose |Aρ | > 0. For any i ∈ Aρ (i.e., the i-th knapsack constraint is active),
consider the step when i is added to Aρ . Let Gi = G ∩ Zi∗ , and we have c(Gi ) + c(z) > 1.
44
Continuous-Time Influence Maximization of Multiple Items
Since every element g we include in Gi satisfies fb(g|G) ≥ c(g)ρ with respect to the solution
Gi when g is added. Then f (g|G) = fi (g|Gi ) ≥ c(g)ρ − 2, and we have:
fi (Gi ∪ {z}) ≥ ρ[c(Gi ) + c(z)] − 2(|Gi | + 1) > ρ − 2(|Gi | + 1).
(24)
Note that Gi is non-empty since otherwise the knapsack constraint will not be active. Any
element in Gi is selected before or at wt , so fi (Gi ) ≥ wt − 2. Also, note that z is not
selected in previous thresholds before wt , so fi ({z} |Gi ) ≤ (1 + δ)wt + 2 and thus,
fi ({z} |Gi ) ≤ (1 + δ)fi (Gi ) + 2(2 + δ).
(25)
Combining Eqs. 24 and 25 into fi (Gi ∪ {z}) = fi (Gi ) + fi ({z} |Gi ) leads to
fi (Gi ) ≥
2(|Gi | + 3 + δ)
1
2(|Gi | + 3 + δ)
ρ
−
≥
ρ∗ −
(2 + δ)
(2 + δ)
2(1 + 2δ)
(2 + δ)
f (O)
≥
− 5|Gi |.
(P + 2k + 1)(1 + 2δ)
Summing up over all i ∈ Aρ leads to the desired bound.
Suppose item i ∈ L spreads according to the diffusion network Gi = (V, Ei ). Let i∗ =
argmaxi∈L |Ei |. By setting = δ/16 in Theorem 8, we have:
Theorem 5. In Algorithm 1, there exists a ρ such that
f (Sρ ) ≥
max {ka , 1}
f (O)
(2|L| + 2)(1 + 3δ)
where ka is the
number of active
knapsack constraints. The expected runtime to obtain the
|L||V|
|Ei∗ |+|V|
e
solution is O
+ δ4
.
δ2
45
| 8cs.DS
|
arXiv:1211.1477v1 [math.AC] 7 Nov 2012
ON THE FINITENESS AND STABILITY OF CERTAIN SETS OF
ASSOCIATED PRIME IDEALS OF LOCAL COHOMOLOGY MODULES
Nguyen Tu Cuonga and Nguyen Van Hoangb,c
a
Institute of Mathematics, Hanoi, Vietnam
Thai Nguyen University of Education, Thai Nguyen, Vietnam
c
Meiji Institute for Advanced Study of Mathematical Sciences, Meiji University,
Kawasaki, Japan
b
Abstract1 . Let (R, m) be a Noetherian local ring, I an ideal of
R and N a finitely generated R-module. Let k≥ − 1 be an integer
and r = depthk (I, N ) the length of a maximal N -sequence in dimension > k in I defined by M. Brodmann and L. T. Nhan (Comm.
Algebra, 36 (2008), 1527-1536). For a subset S ⊆ Spec R we set
S≥k = {p ∈ S | dim(R/p)≥k}. We first prove in this paper that
AssR (HIj (N ))≥k is a finite set for all j≤r . Let N = ⊕n≥0 Nn be a
finitely generated graded R-module, where R is a finitely generated
standard graded algebra over R0 = R. Let r be the eventual value
of depth
S k (I, Nn ). Then our second result says that for all l≤r the
sets j≤l AssR (HIj (Nn ))≥k are stable for large n.
1
Introduction
Let (R, m) be a Noetherian local ring, I an ideal of R and N a finitely generated
R-module. In 1990, C. Huneke [11, Problem 4] asked whether the set of associated primes of HIj (N ) is finite for all finitely generated modules N and all I.
Affirmative answers were given by Huneke-R.Y. Sharp [12] and G. Lyubeznik
[16] for equicharacteristic regular local rings. Although, A. Singh [20] and M.
Katzman [13] provided examples of finitely generated modules having some local
cohomology modules with infinite associated prime ideals, the problem is still
true in many situations, such as [4], [14], [16], [17], [19]. However, little is known
about the finiteness of AssR (HIj (N )). Brodmann-L.T. Nhan introduced the notion of N -sequence in dimension > k in [5]: Let k be an integer with k ≥ −1.
A sequence x1 , . . . , xr of elements of m is called an N -sequence in dimension
1 Key words and phrases: associated prime, local cohomology, generalized local cohomology, N -sequence in dimension > k.
2000 Subject Classification: 13D45, 13D07, 13C15
a E-mail: [email protected]
b E-mail: [email protected]
1
> k if xi ∈
/ p for all p ∈ AssR (N/(x1 , . . . , xi−1 )N ) with dim(R/p) > k and all
i = 1, . . . , r. They also showed that every maximal N -sequence in dimension
> k in I has the same length. This common length is denoted by depthk (I, N ).
Note that depth−1 (I, N ) is just depth(I, N ), depth0 (I, N ) is the filter depth
f-depth(I, N ) defined by R. Lü-Z. Tang [15], and depth1 (I, N ) is the generalized depth gdepth(I, N ) defined by Nhan [19]. Let S be a subset of Spec(R)
and k ≥ −1 an integer. We set S≥k = {p ∈ S | dim(R/p) ≥ k}. The first main
result of this paper is to prove the following theorem.
Theorem 1.1. Let (R, m) be a Noetherian local ring, I an ideal of R and
N a finitely generated R-module. Let k be an integer with k ≥ −1 and r =
depthk (I, N ). If r < ∞ and x1 , . . . , xr is an N -sequence in dimension > k in
I, then for any integer j ≤ r the set AssR (HIj (N ))≥k is finite. Moreover, we
have
[
[
AssR (HIj (N ))≥k =
AssR (N/(x1 , . . . , xj )N )≥k ∩ V (I)
j≤l
j≤l
for all l ≤ r.
Let R = ⊕n≥0 Rn be a finitely generated standard graded algebra over R0 =
R and ⊕n≥0 Nn a finitely generated graded R-module. In 1979, M. Brodmann
[2] had proved that the set AssR (Nn ) is stable for large n. Based on this result
he showed in [3, Theorem 2 and Proposition 12] that the integer depth(I, Nn )
takes a constant value for large n. We generalized this last one in [8] and showed
that depthk (I, Nn ) takes
S a constant value rk for large n. Moreover, in [8], we
also prove that the set j≤r1 AssR (HIj (Nn )) ∪ {m} is stable for large n, where
r1 is the stable value of depth1 (I, Nn ) for large n. Then the second main result
of this paper is the following theorem.
Theorem 1.2. Let (R, m) be a Noetherian local ring and I an ideal of R. Let
R = ⊕n≥0 Rn be a finitely generated standard graded algebra over R0 = R and
N = ⊕n≥0 Nn a finitely generated graded R-module. For each integer k ≥ −1,
let
value of depthk (I, Nn ). Then for each integer l ≤ r the set
S r be the eventual
j
Ass
(H
(N
))
is stable for large n.
R
n
≥k
j≤l
I
For two finitely generated R-modules M and N , J. Herzog [10] defined the
generalized local cohomology module HIj (M, N ) of M and N with respect to I
by
HIj (M, N ) = limn ExtjR (M/I n M, N ).
−→
It is clear that if M = R, HIj (R, N ) is just the ordinary local cohomology module
HIj (N ) of N with respect to I. Therefore all questions about local cohomology
modules discussed above can be asked for generalized local cohomology modules. Thus, to prove the two theorems above, we will show in this paper the
generalizations of them for generalized local cohomology. Let us explain the
organization of the paper. In the next section, we present auxiliary results on
generalized local cohomology modules, the maximal length of an N -sequence in
dimension > k in certain ideals of R and relationships between them. In Section 3, we show a generalization of Theorem 1.1 for generalized local cohomology
2
modules (Theorem 3.1). As consequences, we get again one of main results [19,
Theorem 3.1] of Nhan (see Corollary 3.3) and a generalization of the main results of Brodmann-Nhan [5, Theorems 1.1 and 1.2] (see Corollary 3.5). The last
section is devoted to prove Theorem 4.4, which is a generalization of Theorem
1.2 for generalized local cohomology modules. Then we can derive from Theorem 4.4 a generalization of the main result of [8, Theorem 1.2] for generalized
local cohomology modules (see Corollary 4.5) and some new results about the
asymptotic stability of the sets of associated prime ideals of Ext-modules (see
Corollaries 4.6 and 4.7).
2
Generalized local cohomology module
Throughout this paper we use following notations: we denote by (R, m) a
Noetherian local ring with the maximal ideal m, and M , N finitely generated R-modules; let S be a subset of Spec R and k ≥ −1 an integer, we set
S≥k = {p ∈ S | dim(R/p) ≥ k} and S>k = {p ∈ S | dim(R/p) > k}; for an ideal
I of R, IM is the annihilator of the R-module M/IM .
We first recall the notion of generalized local cohomology module defined by
Herzog in [10].
Definition 2.1. For an integer j ≥ 0, the j th generalized local cohomology
module HIj (M, N ) of M and N with respect to I is defined by
HIj (M, N ) = limn ExtjR (M/I n M, N ).
−→
It is clear that HIj (R, N ) is just the ordinary local cohomology module
of N with respect to I. On the other hand, Ext-modules can be also
determined in many cases by generalized local cohomology modules.
HIj (N )
Lemma 2.2. (cf. [7, Lemma 2.1]) If I ⊆ ann(M ) or ΓI (N ) = N then
j
HIj (M, N ) ∼
= ExtR (M, N ) for all j ≥ 0.
The following definition of a generalization of regular sequences is introduced
by Brodmann-Nhan in [5, Definition 2.1].
Definition 2.3. Let k ≥ −1 be an integer. A sequence x1 , . . . , xr of elements of m is called an N -sequence in dimension > k if xi ∈
/ p for all p ∈
AssR (N/(x1 , . . . , xi−1 )N )>k and all i = 1, . . . , r.
It is clear that x1 , . . . , xr is an N -sequence in dimension > −1 if and only if
it is a regular sequence of N ; and x1 , . . . , xr is an N -sequence in dimension > 0
if and only if it is a filter regular sequence of N introduced by P. Schenzel, N.
V. Trung and the first author in [9]. Moreover, x1 , . . . , xr is an N -sequence in
dimension > 1 if and only if it is a generalized regular sequence of N defined by
Nhan in [19].
Remark 2.4. (i) It is easy to see that if x1 , . . . , xr is an N -sequence in dimension > k, then so is xn1 1 , . . . , xnr r for all n1 , . . . , nr ∈ N.
3
(ii) Note that every maximal N -sequence in dimension > k in an ideal I has
the same length, this common length is denoted by depthk (I, N ) (see [8]). Furthermore, depth−1 (I, N ) is the usual depth depth(I, N ) of N in I, depth0 (I, N )
is the filter depth f-depth(I, N ) of N in I denoted by Lü-Tang in [15], and
depth1 (I, N ) is just the generalized depth gdepth(I, N ) of N in I defined by
Nhan [19].
Lemma 2.5. ([8, Lemma 2.3]) Let k ≥ −1 be an integer. Then
depthk (I, N ) = inf{depthk−i (Ip , Np ) | p ∈ SuppR (N/IN )≥i }
for all 0 ≤ i ≤ k + 1, where we use the convention that inf(∅) = ∞.
The following two results are useful in the sequel.
Lemma 2.6. (cf. [1, Proposition 5.5]) The following equality is true
depth(IM , N ) = inf{i | HIi (M, N ) 6= 0}.
Lemma 2.7. ([6, Theorem 2.4]) Let r = depth(IM , N ). Assume that r < ∞
and let x1 , . . . , xr be a regular sequence of N in IM . Then
AssR HIr (M, N ) = AssR N/(x1 , . . . , xr )N ∩ V (IM ).
3
Proof of Theorem 1.1 and its consequences
The following theorem for generalized local cohomology modules implies Theorem 1.1 when M = R.
Theorem 3.1. Let (R, m) be a Noetherian local ring, I an ideal of R and
M, N finitely generated R-modules. Let k be an integer with k ≥ −1 and r =
depthk (IM , N ). If r < ∞ and x1 , . . . , xr is an N -sequence in dimension > k in
IM , then for any integer l ≤ r we have
[
[
AssR (HIj (M, N ))≥k =
AssR (N/(x1 , . . . , xj )N )≥k ∩ V (IM ).
j≤l
j≤l
Therefore AssR (HIj (M, N ))≥k is a finite set for all j ≤ r.
S
Proof. Let l be an integer with 0 ≤ l ≤ r. For each p ∈ j≤l AssR (HIj (M, N ))≥k
/ AssR (HIj (M, N ))
there exists an integer j0 ≤ l such that p ∈ AssR (HIj0 (M, N )) and p ∈
for all j < j0 . Assume that p ∈
/ AssR (N/(x1 , . . . , xj )N ) for all j < j0 , then
pRp ∈
/ AssRp ((N/(x1 , . . . , xj )N )p ). Hence
AssRp ((N/(x1 , . . . , xj )N )p ) = AssRp ((N/(x1 , . . . , xj )N )p )≥1
for all j < j0 . Since x1 /1, . . . , xj0 /1 is a filter regular sequence of Np by Lemma
2.5 and the hypothesis of x1 , . . . , xj0 , it follows that x1 /1, . . . , xj0 /1 is a regular
4
sequence of Np , and so depth((IM )p , Np ) ≥ j0 . Since p ∈ AssR (HIj0 (M, N )),
HIj0 (M, N )p 6= 0. Thus depth((IM )p , Np ) ≤ j0 by Lemma 2.6, and hence we
obtain that depth((IM )p , Np ) = j0 . This yields by Lemma 2.7 that
AssRp (HIj0 (M, N )p ) = AssRp ((N/(x1 , . . . , xj0 )N )p ) ∩ V ((IM )p ).
Hence p ∈ AssR (N/(x1 , . . . , xj0 )N ) ∩ V (IM ). Thus
[
[
AssR (HIj (M, N ))≥k ⊆
AssR (N/(x1 , . . . , xj )N )≥k ∩ V (IM ).
j≤l
j≤l
S
Conversely, for any p ∈ j≤l AssR (N/(x1 , . . . , xj )N )≥k ∩V (IM ), there is an integer e ≤ l such that p ∈ AssR (N/(x1 , . . . , xe )N ) and p ∈
/ AssR (N/(x1 , . . . , xj )N )
for all j < e. So pRp ∈
/ AssRp ((N/(x1 , . . . , xj )N )p ) for all j < e. Hence
AssRp ((N/(x1 , . . . , xj )N )p ) = AssRp ((N/(x1 , . . . , xj )N )p )≥1
for all j < e. Therefore x1 /1, . . . , xe /1 is a regular sequence of Np , and so
that depth((IM )p , Np ) ≥ e. Keep in mind that p ∈ AssR (N/(x1 , . . . , xe )N ), so
depth((IM )p , (N/(x1 , . . . , xe )N )p ) = 0. Thus we get depth((IM )p , Np ) = e. It
follows by Lemma 2.7 that p ∈ AssR (HIe (M, N )). This implies that
[
[
AssR (HIj (M, N ))≥k ⊇
AssR (N/(x1 , . . . , xj )N )≥k ∩ V (IM ),
j≤l
j≤l
and the theorem follows.
Theorem 1.1 and Theorem 3.1 lead to many consequences. It was showed in
[5, Proposition
S 2.6] that if x1 , . . . , xr is a permutable N -sequence in dimension
> k then n1 ,...,nr ∈N AssR (N/(xn1 1 , . . . , xnr r )N )≥k is a finite set. By virtue of
Theorem 3.1 we can prove this result without the hypothesis that x1 , . . . , xr is
a permutable N -sequence in dimension > k.
Corollary 3.2. Let k be an integer with k ≥ −1 and x1 , . . . , xr an N -sequence
in dimension > k (not necessary permutable). Then
[
[
n
AssR (N/(xn1 1 , . . . , xj j )N )≥k =
AssR (N/(x1 , . . . , xj )N )≥k
j≤r
for all n1 , . . . , nr ∈ N. In particular,
is a finite set.
j≤r
S
n1 ,...,nr ∈N
S
n
j≤r
AssR (N/(xn1 1 , . . . , xj j )N )≥k
Proof. Let n1 , . . . , nr be positive integers, then xn1 1 , . . . , xnr r is an N -sequence
in dimension > k by Remark 2.4. For each integer i with 0 ≤ i ≤ r, we set
Ii = (x1 , . . . , xi ). Since
AssR (N/(xn1 1 , . . . , xni i )N )≥k = AssR (N/(xn1 1 , . . . , xni i )N )≥k ∩ V (Ii )
5
for any r-tuple of positive integers n1 , . . . , nr , it follows from Theorem 1.1 that
[[
[[
n
AssR (HIji (N ))≥k =
AssR (N/(xn1 1 , . . . , xj j )N )≥k ∩ V (Ii )
i≤r
i≤r
j≤i
=
[
j≤i
n
AssR (N/(xn1 1 , . . . , xj j )N )≥k .
j≤r
Then the corollary follows from the fact that the set on the left hand of the
equality is independent of the choice of n1 , . . . , nr .
Setting k = 1 in Corollary 3.2 we get the following corollary which covers
Theorem 3.1 of [19].
Corollary 3.3. Let x1 , . . . , xr be a generalized regular sequence of N . Then
[
[
n
AssR (N/(xn1 1 , . . . , xj j )N ) ∪ {m} =
AssR (N/(x1 , . . . , xj )N ) ∪ {m}
j≤r
j≤r
for all n1 , . . . , nr ∈ N. In particular,
is a finite set.
S
n1 ,...,nr ∈N
S
n
j≤r
AssR (N/(xn1 1 , . . . , xj j )N )
Corollary 3.4. Let k be an integer with k ≥ −1. Set r = depth
pk (ann(M ), N ).
If r < ∞ and x1 , . . . , xr is an N -sequence in dimension > k in ann(M ), then
for any integer l ≤ r we have
[
[
AssR (ExtjR (M, N ))≥k =
AssR (N/(x1 , . . . , xj )N )≥k ∩ V (ann(M )).
j≤l
j≤l
p
√
Proof. Set I =
ann(M ). It is easy to see that IM = IM = I. Hence
r = depthk (IM , N ) and x1 , . . . , xr is an N -sequence in dimension > k in IM .
j
Moreover, we have in this case that ExtjR (M, N ) ∼
= HI (M, N ) by Lemma 2.2.
Therefore the conclusion follows from Theorem 3.1.
Let j ≥ 0, t > 0 be integers. Let a = (a1 , . . . , as ) be elements of R and
t = (t1 , . . . , ts ) an s-tuple of positive integers. For an ideal I we set
T j (I t , N ) = AssR (ExtjR (R/I t , N )) and
T j (at , N ) = AssR (ExtjR (R/(at11 , . . . , atss ), N )).
Then we get the following relation between these sets.
Corollary 3.5. Let k ≥ −1 be an integer, I an ideal of R and r =√depthk (I, N ).
If r < ∞ and x1 , . . . , xr is an N -sequence in dimension > k in I. Then, for
any system of generators a1 , . . . , as of I and all integer l ≤ r, we have
[
[
[
T j (I t , N )≥k =
AssR (N/(x1 , . . . , xj )N )≥k ∩ V (I) =
T j (at , N )≥k
j≤l
j≤l
j≤l
6
for all t ∈ N and all tS∈ Ns . In particular, for any integer j ≤ r the sets
S
j t
j t
t∈N T (I , N )≥k and
t∈Ns T (a , N )≥k are contained in the finite set
[
i≤j
Proof. Since
√
It =
AssR (N/(x1 , . . . , xi )N )≥k ∩ V (I).
q
√
I = (at11 , . . . , atss ) for all positive integer t, t1 , . . . , ts ,
thanks to Corollary 3.4 we obtain by setting M = R/I t and M = R/(at11 , . . . , atss )
that
[
[
AssR (ExtjR (R/I t , N ))≥k =
AssR (N/(x1 , . . . , xj )N )≥k ∩ V (I) and
j≤l
[
j≤l
AssR (ExtjR (R/(at11 , . . . , atss ), N ))≥k =
j≤l
[
j≤l
AssR (N/(x1 , . . . , xj )N )≥k ∩ V (I)
for all l ≤ r, and the corollary follows.
Remark 3.6. The main result of Brodmann-Nhan [5, Theorems 1.1 and 1.2]
i
says that S
for a non negative integer
S r, if dim Supp HI (N ) ≤ k for all i < r, then
the sets t∈N T j (I t , N )≥k and t∈Ns T j (at , N )≥k are contained in the finite
S
set i≤j AssR (ExtiR (R/I, N )) for all j ≤ r. Moreover, if x1 , . . . , xr is at the
same time a permutable N -sequence in dimension > k and a permutable I-filter
regular sequence in I, then these sets are contained in the finite set
[
AssR (N/(x1 , . . . , xj )N )≥k+1 ∪
AssR (N/(x1 , . . . , xi )N )k ,
(*)
i≤j
where AssR (N/(x1 , . . . , xi )N )k = {p ∈ AssR (N/(x1 , . . . , xi )N ) | dim R/p = k}.
In fact, since the S
equality in Corollary 3.5
S do not depend on t and t, thus for
any j ≤ r the sets t∈N T j (I t , N )≥k and t∈Ns T j (at , N )≥k are contained in the
S
S
set i≤j T i (I, N )≥k = i≤j AssR (ExtiR (R/I, N ))≥k . This shows that Corollary
3.5 covers Theorem 1.1 of [5]. Moreover, let p ∈ T j (I t , N )≥k ∪ T j (at , N )≥k for
some t, t. If dim(R/p) = k then p belongs to the set (∗) by Corollary 3.5. If
dim(R/p) ≥ k + 1 then depth(Ip , Np ) = r. Therefore j = r in this case, and so
p ∈ T r (I, N ). Since ExtrR (R/I, N ) ∼
= HIr (R/I, N ) by Lemma 2.2, it follows by
Lemma 2.7 that p ∈ Ass(N/(x1 , . . . , xr )N )≥k+1 . Thus Corollary 3.5 also covers
Theorem 1.2 of [5] for local rings without the assumption that x1 , . . . , xr is at
the same time a permutable N -sequence in dimension > k and a permutable
I-filter regular sequence in I.
4
Proof of Theorem 1.2 and its consequences
In this section we always denote by R = ⊕n≥0 Rn a finitely generated standard
graded algebra over R0 = R and N = ⊕n≥0 Nn a finitely generated graded
7
R-module. Firstly, we recall two results on the asymptotic stability of associated prime ideals [2] (see also [18, Theorem 3.1]) and of generalized depths
depthk (I, Nn ) [8, Theorem 1.1] as follows.
Lemma 4.1. The set AssR (Nn ) is stable for large n.
Lemma 4.2. ([8, Theorem 1.1]) Let k be an integer with k ≥ −1. Then
depthk (I, Nn ) is independent of n for large n.
In view of Lemma 4.2, if r = depthk (I, Nn ) for all large n then we call that r is
the eventual value of depthk (I, Nn ).
Lemma 4.3. Let k be an integer with k ≥ −1 and r the eventual value of
depthk (I, Nn ). Assume that 1 ≤ r ≤ ∞. Then there exists a sequence x1 , . . . , xr
in I which is an Nn -sequence in dimension > k for all large n.
Proof. Let r ≥ 1 be the eventual value of depthk (I, Nn ). By Lemma 4.1 we can
choose x1 ∈ I such that x1 ∈
/ p for all p ∈ AssR (Nn )>k and all large n. Then
depthk (I, Nn /x1 Nn ) = r − 1 for all large n. By the induction assumption, there
exists a sequence x2 , . . . , xn ∈ I to be an Nn /x1 Nn -sequence in dimension > k
for all large n. Thus x1 , . . . , xr is the sequence as required.
Proof of Theorem 1.2. Theorem 1.2 follows immediately from the following
theorem when M = R.
Theorem 4.4. Let (R, m) be a Noetherian local ring, I an ideal of R and M
finitely generated R-module. Let R = ⊕n≥0 Rn be a finitely generated standard
graded algebra over R0 = R and N = ⊕n≥0 Nn a finitely generated graded Rmodule. For each integer k ≥ −1, let r be the eventual value of depthk (IM , Nn ).
Then for each integer l ≤ r, the set
[
AssR (HIj (M, Nn ))≥k
j≤l
is finite and stable for large n.
Proof. Let r be the eventual S
value of depthk (IM , Nn ). We need to show that
for any integer l ≤ r the set j≤l AssR (HIj (M, Nn ))≥k is finite and stable for
large n.
By Lemmas 4.1 and 4.2, we can find a large integer t such that AssR (Nn /IM Nn )
is stable and r = depthk (IM , Nn ) for all n ≥ t. Set d = dim(Nt /IM Nt ). We
consider three cases as follows.
Case 1. If d < k, then r = ∞ and the conclusion follows as
[
AssR (HIj (M, Nn ))≥k = ∅
j≤l
for all n ≥ t.
8
Case 2. If d = k, then r = ∞. It follows for all positive integers n ≥ t that
[
AssR (HIj (M, Nn ))≥d ⊆ SuppR (Nn /IM Nn )≥d = AssR (Nn /IM Nn )≥d ,
j≤l
since the set AssR (Nn /IM Nn )≥d only consists of all minimal elements with
dim(R/p) = d. Hence
[[
[
X=
AssR (HIj (M, Nn ))≥d ⊆
AssR (Nn /IM Nn ),
n≥t j≤l
n≥t
therefore X is finite by Lemma 4.1. Thus we can take t sufficiently large such
that for each p ∈ X then
[
p∈
AssR (HIj (M, Nn ))≥d
j≤l
for infinitely many n ≥ t. Now for each p ∈ X we set s to be the eventual
value of depth((IM )p , (Nn )p ), then we can write s = depth((IM )p , (Nn )p ) for
all n ≥ n(p) for some integer n(p) ≥ t. Then HIs (M, Nn )p 6= 0 for all n ≥ n(p).
Moreover, by the definition of X, it implies that s ≤ l. Hence p is a minimal
element of SuppR (HIs (M, Nn )), and so that
[
p ∈ AssR (HIs (M, Nn ))≥d ⊆
AssR (HIj (M, Nn ))≥d
j≤l
for all n ≥ n(p). Therefore the set
[
AssR (HIj (M, Nn ))≥d
j≤l
is finite and stable for all n ≥ max{n(p) | p ∈ X}.
Case 3. Assume that d > k. Then r < ∞. If l = 0, then
AssR (HI0 (M, Nn ))≥k = AssR (Nn )≥k ∩ V (IM )
is finite and stable for large n by Lemma 4.1. Assume that 1 ≤ l ≤ r. In
this case, by Lemma 4.3, there exists a sequence x1 , . . . , xr in IM which is an
Nn -sequence in dimension > k for all n ≥ u for some integer u ≥ t. By Theorem
3.1, we get the following equality
[
[
AssR (HIj (M, Nn ))≥k =
AssR (Nn /(x1 , . . . , xj )Nn )≥k ∩ V (IM )
j≤l
j≤l
for all n ≥ u. Therefore we conclude by Lemma 4.1 that the following set
[
AssR (HIj (M, Nn ))≥k
j≤l
is finite and stable for large n as required.
9
The rest of this section devotes to consider many consequences of Theorem
4.4. By replacing k = −1, 0, 1 in Theorem 4.4, we get the following result
which is an extension of the main result in [8, Theorem 1.2] for generalized local
cohomology modules.
Corollary 4.5. Let r, r0 and r1 be the eventual values of depth(IM , Nn ),
f-depth(IM , Nn ) and gdepth(IM , Nn ), respectively. Then the following statements are true.
(i) The set AssR HIr (M, Nn ) is finite and stable for large n.
S
(ii) For each integer l0 ≤ r0 the set j≤l0 AssR (HIj (M, Nn )) is finite and stable
for large n.
S
(iii) For each integer l1 ≤ r1 the set j≤l1 AssR (HIj (M, Nn )) ∪ {m} is finite
and stable for large n.
Proof. (i) Since HIj (M, Nn ) = 0 for all j < r and all large n, so that
[
AssR (HIj (M, Nn ))≥−1 = AssR HIr (M, Nn )
j≤r
is finite and stable for large n by Theorem 4.4.
(ii) The conclusion follows from Theorem 4.4 and the equality
[
[
AssR (HIj (M, Nn ))
AssR (HIj (M, Nn ))≥0 =
j≤l0
j≤l0
for all l0 ≤ r0 and all large n.
(iii) For each l1 ≤ r1 we have
[
[
AssR (HIj (M, Nn ))≥1 ∪ {m} =
AssR (HIj (M, Nn )) ∪ {m}
j≤l1
j≤l1
for all large n. Therefore the result follows by Theorem 4.4.
p
Note that if I = ann(M ) then HIj (M, Nn ) =√ExtjR (M, Nn ) for all j ≥ 0 by
Lemma 2.2. Moreover, in this case we have IM = IM = I. Therefore Theorem
4.4 leads to the following immediate consequence.
Corollary 4.6. Let k be an integer with k ≥ −1 S
and r the eventual value of
depthk (ann(M ), Nn ). Then for any l ≤ r the set j≤l AssR (ExtjR (M, Nn ))≥k
is stable for large n.
For an integer j ≥ 0, an ideal I of R, a system a = (a1 , . . . , as ) of elements
of R, and a s-tuple t = (t1 , . . . , ts ) ∈ Ns , as in the previous section we set
T j (I t , Nn ) = AssR (ExtjR (R/I t , Nn )),
T j (at , Nn ) = AssR (ExtjR (R/(at11 , . . . , atss ), Nn )).
Then we have
10
Corollary 4.7. Let k ≥ −1 be an integer, I an ideal of R and r the eventual
value of depthk (I, Nn ). Let a1 , . . . , as be a system of generators of I and an
integer l ≤ r. Then for all positive integers t, t1 , . . . , ts the following sets
[
[
T j (I t , Nn )≥k and
T j (at , Nn )≥k
j≤l
j≤l
are stable for large n. Moreover, if r < ∞ then
[
[
T j (I t , Nn )≥k =
T j (at , Nn )≥k
j≤l
j≤l
are stable for all large n.
Proof. By Lemma 4.2, we find a large integer u such that r =√depthk (I,
√ Nn )
t =
I
I =
for
all
n
≥
u.
Let
t,
t
,
.
.
.
,
t
be
positive
integers.
Since
1
s
q
t1
ts
(a1 , . . . , as ) , therefore
√
r = depthk ( I, Nn ) = depthk (I t , Nn ) = depthk ((at11 , . . . , atss ), Nn )
for all n ≥ u. This implies
Corollary 4.6 Sby setting M = R/I t and
S from
t1
j t
ts
M = R/(a1 , . . . , as ) that j≤l T (I , Nn )≥k and j≤l T j (at , Nn )≥k are stable
for large n. We now
∞. Then, for any n ≥ u, we have by
S assume that r < S
Corollary 3.5 that j≤l T j (I t , Nn )≥k = j≤l T j (at , Nn )≥k for all for all positive
integers t, t1 , . . . , ts . Thus they are stable for all large n.
Acknowledgements: The authors would like to thank the referee for pointing out a mistake in Remark 3.6. This work is funded by Vietnam National
Foundation for Science and Technology Development (NAFOSTED), and the
second author is also partially supported by Vietnam Ministry of Training and
Education under grant number B2012-TN03-02.
References
[1] M. H. Bijan-Zadeh, A common generalization of local cohomology theories, Glasgow Math. J., 21 (1980), 173-181.
[2] M. Brodmann, Asymptotic stability of AssR (M/I n M ), Proc. Amer. Math. Soc.,
(1) 74 (1979), 16-18.
[3] M. Brodmann, The Asymptotic nature of the analytic spread, Math. Proc. Camb.
Phil. Soc., 86 (1979), 35-39.
[4] M. Brodmann and A. L. Faghani, A Finiteness result for associated primes of
local cohomology modules, Proc. Amer. Math. Soc., (10) 128 (2000), 2851 - 2853.
[5] M. Brodmann and L. T. Nhan, A finiteness result for associated primes of certain
Ext-modules, Comm. Algebra, 36 (2008), 1527-1536.
11
[6] N. T. Cuong and N. V. Hoang, Some finite properties of generalized local cohomology modules, East-West J. Math., (2) 7 (2005), 107-115.
[7] N. T. Cuong and N. V. Hoang, On the vanishing and the finiteness of supports of
generalized local cohomology modules, Manuscripta Math., (1) 126 (2008), 59-72
.
[8] N. T. Cuong, N. V. Hoang and P. H. Khanh, Asymptotic stability of certain
sets of associated prime ideals of local cohomology modules, Comm. Algebra, 38
(2010), 4416-4429.
[9] N. T. Cuong, P. Schenzel and N. V. Trung, Verallgemeinerte Cohen-Macaulay
Moduln, Math. Nachr., 85 (1978), 57-73.
[10] J. Herzog, Komplexe, Auflösungen und Dualität in der Lokalen Algebra, Habilitationsschrift, Universität Regensburg, 1970.
[11] C. Huneke, Problems on local cohomology, Free resolutions in commutative algebra and algebraic geometry (Sundance, Utah, 1990), Res. Notes Math., 2 (1992),
93 - 108.
[12] C. Huneke and R. Y. Sharp, Bass numbers of local cohomology modules, Trans.
Amer. Math. Soc., 339 (1993), 765 - 779.
[13] M. Katzman, An example of an infinite set of associated primes of a local cohomology module, J. Algebra, 252 (2002), 161-166.
[14] K. Khashyarmanesh and Sh. Salarian, On the associated primes of local cohomology modules, Comm. Algebra, 27 (1999), 6191 - 6198.
[15] R. Lü and Z. Tang, The f-depth of an ideal on a module, Math. Proc. Camb. Phil.
Soc., (7) 130 (2001), 1905-1912.
[16] G. Luybeznik, Finiteness properties of local cohomoly modules (an application of
D-modules to commutative algebra), Invent. Math., 113 (1993), 41-55.
[17] T. Marley, Associated primes of local cohomology module over rings of small
dimension, Manuscripta Math., (4) 104 (2001), 519 - 525.
[18] L. Melkersson, On asymptotic stability for sets of prime ideals connected with the
powers of an ideal, Math. Proc. Camb. Phil. Soc., 107 (1990), 267-271.
[19] L. T. Nhan, On generalized regular sequences and the finiteness for associated
primes of local cohomology modules, Comm. Algebra, 33 (2005), 793-806.
[20] A. Singh, p-torsion elements in local cohomology modules, Math. Res. Lett., 7
(2000), 165-176.
12
| 0math.AC
|
Stability and Stabilization of Fractional-order Systems with
Different Derivative Orders: An LMI Approach
Pouya Badri1, Mahdi Sojoodi1*
1
Advanced Control Systems Laboratory, School of Electrical and Computer Engineering, Tarbiat Modares
University, Tehran, Iran.
Abstract— Stability and stabilization analysis of fractional-order linear time-invariant (FO-LTI)
systems with different derivative orders is studied in this paper. First, by using an appropriate
linear matrix function, a single-order equivalent system for the given different-order system is
introduced by which a new stability condition is obtained that is easier to check in practice than
the conditions known up to now. Then the stabilization problem of fractional-order linear systems
with different fractional orders via a dynamic output feedback controller with a predetermined
order is investigated, utilizing the proposed stability criterion. The linear matrix inequality based
procedure of developing stabilizing output feedback control is preserved in spite of the complexity
of assuming the most complete linear controller model, with direct feedthrough parameter. The
proposed stability and stabilization theorems are applicable to FO-LTI systems with different
fractional orders in one or both of
and
intervals. Eventually, some numerical
examples are presented to confirm the obtained analytical results.
Keywords: Fractional-order system, different fractional orders, stability, stabilization, linear matrix
inequality, dynamic output feedback.
I. INTRODUCTION
Recently, study of applications of fractional calculus in the modeling and control of various real-world
systems have attracted increasing interest [1, 2]. This is mainly due to the fact that fractional-order
models can more precisely describe systems having responses with anomalous dynamics and long
memory transients. Furthermore, many natural, biological, and practical engineering systems have
inherent properties which can be better explained by fractional-order models. Accordingly, controllerdesigning methods for systems modeled by fractional-order dynamics are of great interest. Utilizing
fractional-order controllers is a proper method for controlling such systems, because of their high
performance and robustness [3, 4]. Moreover, stability and stabilization have become fundamental issues
for all control systems including fractional-order systems, because in the most of control systems,
stability is the first objective to be accomplished. In this paper, we are particularly concerned with
fractional-order linear time-invariant (LTI) systems with different orders. This kind of system
representation is used to analyze linear electrical circuits composed of resistors, supercondensators
(ultracapacitors), coils, and voltage (current) sources [5], and formulate the problem of model reference
adaptive control of fractional-order systems presented in [6]. Relaxation processes, viscoelastic materials
models, and diffusion phenomena can be easily modeled by multi-order fractional differential equations
[7].
The stability of fractional-order feedback systems was investigated in [8], by adapting classical root
locus plot analysis to some viscoelastic structures. For fractional differential systems in state-space form,
both internal and external stabilities are investigated in [9]. Stability of FO-LTI systems with different
fractional orders is addressed in [10], using its characteristic equation. Multi-term fractional differential
equations such as Bagley–Torvik and Basset equations were discussed in [11], where the solution is
The authors are with the School of Electrical and Computer Engineering, Tarbiat Modares University, Tehran, Iran, P.O. Box 14115.
[email protected], [email protected] (corresponding author to provide phone: +98 21 8288-3902)
proposed under the assumption that fractional derivatives are rational numbers. Stabilization of uncertain
multi-order fractional systems with fractional orders
is investigated in [12] using extended
state observer.
The utilization of linear matrix inequality (LMI)-based methods has been increased by the evolution of
efficient numerical procedures to solve convex optimization problems [13–17] defined by LMI
conditions. Appropriate LMI conditions for asymptotically stability analysis of FO-LTI systems with
fractional orders in
and
intervals are presented in [18] and [19], respectively. The
problem of pseudo-state feedback stabilization of fractional-order systems is presented in [19]. Necessary
and sufficient conditions for the stability and stabilization of fractional-order interval systems is
addressed in [20]. Furthermore, in [21], the robust stability and stabilization of fractional-order linear
systems with positive real uncertainty are investigated.
State feedback controller is utilized in the most of mentioned studies, in which all individual states are
needed. However, in some cases measuring all states is impossible due to economic issues or physical
limitations, where using output feedback control seems to be effective [14]. Moreover, it should be noted
that dynamic controllers are superior to the static ones because of their more performance efficiency and
degrees of freedom in achieving control objectives [22].
A large number of controller design procedures lead to high order controllers with expensive
implementation, undesirable reliability, maintenance difficulties, and potential numerical errors. On the
other hand, plant or controller order reduction methods do not always guarantee the desired closed-loop
performance. Therefore, designing a controller with a low and fixed order, chosen by the designer
according to the requirements of the problem, is an efficient solution [14].
Motivated by the above discussion, in this paper, using an appropriate linear matrix function, a singleorder equivalent system for the given different-order system is introduced by which a new stability
condition is derived that is easier to check in practice comparing to the conditions known so far. As far as
we know, there is no result on the analytical design of a stabilizing dynamic output feedback controller
for fractional-order systems with different orders in the literature. In this paper, the stabilization problem
of fractional-order linear systems with different fractional orders through a dynamic output feedback
controller with a predetermined order is investigated based on the proposed stability theorem. The LMIbased approach of developing stabilizing controller is preserved, despite the intricacy of assuming the
most complete model of linear controller with direct feedthrough parameter. Moreover, the proposed
stability and stabilization theorems are applicable to FO-LTI systems with different fractional orders in
one or both of
and
intervals.
The rest of this paper is organized as follows. In section II, some preliminary concepts, lemmas, and
definitions are included. Problem formulation is presented in section III. Section IV is devoted to the
main results of the paper where a single-order equivalent system for the given different-order system is
presented, a new stability condition is investigated for such systems, and the stabilization analysis of
fractional-order systems with different orders is addressed. Some numerical examples are provided in
Section V to illustrate the efficiency and validity of the proposed method. Eventually, section VI draws
the conclusion.
(
) is the greatest
Notation: Throughout the paper is the symbol of pseudo inverse,
).
common divisor of n-tuple (
and
represent the -th row and -th column of matrix
, respectively. Moreover,
is a matrix of suitable dimensions with zero entries and denotes the
imaginary unit. Function space
,
- is the space of functions , having continuous -th derivative.
(
)
Furthermore, symbols , ⌈ ⌉, ⌊ ⌋, and
stand for natural number set, ceiling function, floor function,
and -th derivative of , respectively.
II. PRELIMINARIES
In this section, some basic concepts of fractional-order calculus and useful definitions and lemmas are
presented.
The following Caputo fractional derivative definition is adopted throughout this paper, due to its
integer–order initial value that makes it applicable to physical systems.
(
)
(
)
. /
( )
,
where ( ) is Gamma function defined by ( )
and
to or greater than . In the sequel, wherever the notation
operator
is meant.
is used, the Caputo fractional derivative
is the smallest integer that is equal
For the proofs of the theorems in this paper, we will use the following lemmas.
,
with
Lemma 1 [11]. Let
there exists some
- for some
and
and some
. Moreover, let
,
-. Then
be such that
(1)
Lemma 2 [11]. Let
,
- and
( )
,
.
⌈ ⌉. Moreover, assume that
and
( )
Lemma 3 [23]. Commensurate system
satisfied
|
(
( ))|
,
-. Then,
( ) is stable if the following conditions are
(2)
,
for all eigenvalues
( ) of the matrix .
( )
Lemma 4 [19]: Let
,
and
. The fractional-order system
is asymptotically stable if and only if there exists a positive definite Hermitian matrix
such that
̅̅ /
.
(
̅ ̅)
(
(
)
and the set
)
,
({
in which
(3)
,
where
Definition 1. For any square matrix
function is defined as follows:
)
with
( )
,
( ).
III. PROBLEM FORMULATION
Consider the following fractional-order linear time-invariant FO-LTI system
the linear matrix
(4)
( )
( )
[
( )
]
( )
(5)
( )
( )
( )
subject to the initial conditions
( )(
)
.
( )
(6)
⌊ ⌋,
/
(
)
in which
different fractional orders of the system,
( )
(
Furthermore,
,
matrices with appropriate dimensions.
( ),
Defining
denotes the pseudo-state vector,
is the control input, and
)
( )
, and
, and
are the
is the output vector.
are known constant
, one can obtain the following
equivalent pseudo-state space system
()
( )
( )
( ),
in which
( ),
(7)
(
(8)
) ,
and
(
)
(9)
,
with initial conditions
( )
( )(
and
(
)
( )
{
(
( )
{
)
( )
⌊ ⌋
(10)
is easily obtained using the linear matrix function F in Definition 1, as follows
(
)
(
)
also we have
( )
where
and .
)
,
,
(
[
]
,
, and
(
)
,
(11)
)
-
,
(12)
are submatrices in the -th row and -th column of partitioned matrices
,
,
IV. MAIN RESULTS
In this section, first, an equivalent single-order system for the given different-order system (5) is
derived using the matrix function (4). Then, utilizing the equivalent system, a new stability condition is
derived for system (5) and an LMI approach is proposed for designing a dynamic output feedback control
law to stabilize the fractional-order system (5), using the stabilizing controller parameters of system (7).
A. Equivalency
First, equivalency of systems (5) and (7) is proved in the following theorem.
Theorem 1. Consider the fractional-order linear system (5). Defining
to be the greatest common
divisor of s the FO-LTI systems (5) and (7) subject to the initial conditions ( ) and ( ) are
equivalent in the following sense:
⌈ ⌉
1. Whenever , defined in (7) and (8), with
,
- for some
is the pseudo-state of
the FO-LTI system (7), is the pseudo-state of (5).
2. Whenever is the pseudo-state of (5), , defined in (7) and (8), is the pseudo-state of the FO-LTI
system (7).
Proof. In the case that all fractional derivatives of FO-LTI system (5) are equal, we have
which brings about the equality of systems (5) and (7). Therefore, from now on we assume that
at least one of the s is different from the others. As a result we have
.
In
order
to
prove
the
first
claim,
we
have
to
assume
that
(
) is the pseudo-state vector of system (7),
and we define
,
. Since
, by a repeated application of Lemma 1 combined with
(9), we obtain
,
,
(
)
(
(
)
(13)
(
)
)
(
)
,
( )
According to the rows of
and , which are made from
that
, the last equation can be rewritten as follows
(
( )
By definition,
( )
( )
[
]
)
( )
and
rows respectively, and the fact
(14)
( )
(
)
( )
and considering (14) for
we have
( )
( )
Therefore, ( ) is the pseudo-state vector of system (5). In addition, it is obvious from equation system
(13) that for
we have
( )
( )
( )
( )
( )
( )
( )
Therefore, ( ) also satisfies the initial conditions (6).
(
) is the pseudo-state of system
For the second claim we have to assume that
(5) with the initial conditions (6). The equation system (13) is valid in this case too, and therefore it
( )
follows that is the pseudo-state vector of system (7) with the initial conditions ( ) ( ) .
/
.
( )
/
⌊ ⌋ whenever
of Lemma 2 reveals that
( )(
)
.
and ( )
( )
/
(
)
( )
for
. Finally, an application
in the other cases, and it completes the proof. ∎
B. STABILITY
In this subsection a necessary and sufficient condition is established for the asymptotical stability of
system (5) with ( )
.
Theorem 2. The fractional-order linear system (5) with ( )
orders , if and only if there exists a Hermitian matrix
̅ ̅)
(
̅ ̅)
(
is asymptotically stable for different
such that
(15)
,
in which
and is defined in (11).
Proof. It can be easily concluded that since the fractional orders
are different
from each other, then
. Therefore, It follows from Lemma 4 that the fractional-order system
(7) is asymptotically stable if and only if there exists a positive definite Hermitian matrix
such that (15) holds, and since according to Theorem 1 the pseudo states of original system (5) are
in the subset of system (7) pseudo states, asymptotically stability of (7) results that of the system (5). For
the proof of necessity we suppose that system (5) is asymptotically stable. As derived in [5], pseudostates of system (5), i.e.
( )
, are equal to some finite linear combination of
⌈ ⌉
(
(
)
pseudo-states of system (7), i.e.
(
)
functions, with this in mind and according to (13),
( )
, are some finite linear combination of
⌈ ⌉
)
(
(
(
)
)
)
functions. Due to the asymptotically
stability of ( ) together with the fact that ( ) is a vanishing term and (
(
) )
holds, it can be concluded that pseudo-states of system (7) are some finite linear
combinations of asymptotically stable functions, which brings about the asymptotically stability of
( )
, and it completes the proof. ∎
C. Stabilization
The main purpose of this subsection is to design a fixed-order dynamic output feedback controller that
asymptotically stabilizes the FO-LTI system (5) using LMIs. Hence, the following dynamic output
feedback controller is presented
( )
( )
( )
( )
(16)
( )]
[
( )
( )
( )
with
, in which
is the arbitrary order of the controller, and
,
,
controller unknown parameters with appropriate dimensions.
The resulting closed-loop augmented FO-LTI system using (5) and (16) is as follows
, and
are the
( )
( )
( )
( )
( )
( )
( )
( )
(17)
( )]
[
where
( )
[
( )
]
( )
[
Next, a stabilization result is established.
]
(18)
Theorem 3. Considering closed-loop system in (17), if a positive definite Hermitian matrix
(
) (
)
in the form of
(
)
alongside with matrices
[
(19)
exist in a way that
,
]
(
where
(
̅ )
)
̅ ̅)
(
then, the dynamic output feedback controller parameters of
and
̅ ̅̅̅)
̅ ̅̅̅)
(
(
(20)
̅̅)
̅̅)
(
(
(21)
make the FO-LTI system in (5) asymptotically stable.
Proof. It follows from Theorem 1 that the closed-loop FO-LTI system (17) is asymptotically stable if
(
) (
) such that
there exists a positive definite Hermitian matrix
(22)
(
(
̅ ̅)
̅ ̅)
where
(23)
(
*
+)
in (4). By some manipulations, it can be easily
in which
is the linear matrix function defined
concluded that
(
*
where matrices
matrix
[
]
,
, and
(
) (
(24)
]
are defined in (11) and (12). According to (22) and (24), and assuming
)in the form of (19), the inequality (22) can be written as follows
(
+)
[
̅ ̅)
(
̅̅)
(25)
The matrix inequality (25) is not linear due to several multiplications of variables. Therefore, by
changing variables as
the linear matrix inequality of (20) is obtained. ∎
V.
(26)
SIMULATION
In this section, some numerical examples are given to validate the effectiveness of the derived results.
Various efficient LMI Parsers and Solvers can be used to determine the feasibility of the proposed
constraints and obtain the parameters of the controller design problem. In this paper, we use YALMIP
parser [24] and SeDuMi [25] solver in Matlab tool [26].
A. Example 1
Consider system (5) with the following parameters
(27)
According to the stability criterion proposed in [10], characteristic equation of above system is as
follows
[
]
(28)
The roots of characteristic equation (28) along with the stability boundary
are depicted in Fig.
1. It can be concluded from [10] that the system (27) is asymptotically stable since roots of equation (28)
satisfy the condition | ( )|
, in which
.
The same conclusion can be drawn from proposed Theorem 2 with a relatively simpler method. The
eigenvalues of
and stability boundaries
, with
are demonstrated in Fig. 2.
( )
( ) is stable for
According to Lemma 3 and Fig. 2, the system
(
(
)),
since the eigenvalues of are located on the left side of
boundaries. Subsequently, according to
Theorem 2, system (5) with the parameters in (27) is asymptotically stable.
Since the characteristic equation (28) is of order 372 and
, checking the
eigenvalues of is far easier than checking the arguments of roots of characteristic equation (28), which
is obvious from Figures 1 and 2.
Fig. 1. Roots of characteristic equation (28) and stability boundaries
with
for the system of Example 1.
Fig. 2. The eigenvalues of
and stability boundaries
, with
for the system of Example 1.
B. Example 2
Consider the dynamic output feedback stabilization problem of the fractional-order system (5) with the
following parameters
(29)
The eigenvalues of
and stability boundaries
, with
are depicted in Fig. 3.
( )
( ) is unstable for
According to Lemma 3 and Fig. 3, the system
(
(
)),
since one of the eigenvalues of
are located on the right side of
boundaries. Therefore,
according to Theorem 2, system (5) with the parameters in (29) is unstable. However, according to
Theorem 3, it can be concluded that this fractional-order system is asymptotically stabilizable utilizing
the obtained dynamic output feedback controllers of arbitrary orders in the form of (16), tabulated in
Table 1. The eigenvalues of
are located in stability region which is depicted in Fig. 3. According to
Theorem 3, stability of
implies the stability of
. Fig. 4 shows the time response of open-loop
system and the resulted closed-loop system of form (17) via obtained controller with
. It is
obvious from Fig. 4 that all states of the closed-loop system asymptotically converge to zero.
0
1
,
0 1
-
TABLE 1. CONTROLLER PARAMETERS OF EXAMPLE 2 OBTAINED BY THEOREM 3
0
0
0
0
-0.21
1
-1.10
1.77
1.73
-0.16
2
0
1
0
1
,
- -0.16
Fig. 3. The eigenvalues of ,
and stability boundaries
, with
for the system of Example 2 with
.
Fig. 4. The time response of the open-loop (red) and closed-loop system in Example 2 via obtained output feedback
controllers with
(black),
(blue)
(red).
C. Example 3
Dynamic output feedback stabilization problem of the fractional-order system (5) is considered with
the following parameters
(30)
The eigenvalues of and stability boundaries
, with
are demonstrated in Fig. 5.
( )
( ) is unstable for
According to Lemma 3 and Fig. 5, the system
(
(
)),
since not all of the eigenvalues of are located on the left side of
boundaries. Thus, according
to Theorem 2, system (5) with the parameters in (30) is unstable. However, according to Theorem 3, it
can be concluded that this fractional-order system is asymptotically stabilizable utilizing the obtained
dynamic output feedback controllers of arbitrary orders in the form of (16), tabulated in Table 2. The
eigenvalues of
are located in stability region which is depicted in Fig. 3. According to Theorem 3,
stability of
implies the stability of
. Fig. 6 shows the time response of open-loop system and the
0
1
0 1
,
-
resulted closed-loop system of form (17) via obtained controllers with
and . It is obvious from
Fig. 6 that all the closed-loop system states asymptotically converge to zero. Furthermore, control effort
signals of the obtained controllers are depicted in Fig. 7 in order to have an accurate comparison. It can
be concluded from Fig. 6 and Fig. 7 that, with almost the same control effort amount, the obtained
dynamic output feedback controllers, even with a low order of
, have more appropriate control
actions compared to that of the static controller of
. Moreover, it is obvious that for output
feedback stabilizing of fractional-order systems with orders in both
and
intervals,
a unified method is utilized.
TABLE 2. CONTROLLER PARAMETERS OF EXAMPLE 3 OBTAINED BY THEOREM 3
0
0
1
-21.78
2
Fig. 5. The eigenvalues of ,
0
0
1 ,
1 0
, and stability boundaries
, with
0
1.28
2.46
1.20
-
1.14
for the system of Example 3 with
.
Fig. 6. The time response of the open-loop (red) and closed-loop system in Example 3 via obtained output feedback
controllers with
(black),
(blue).
Fig.7. The control effort of the dynamic output feedback controllers of Example 3 obtained by Theorem 3.
VI. CONCLUSION
In this paper using an appropriate linear matrix function, a single-order equivalent system for the given
different-order system is introduced through which a new stability condition has been obtained that is
easier to check in practice compared with the conditions known up to now (Theorem 1 and 2). Then the
stabilization problem of fractional-order linear systems with different fractional orders via a dynamic
output feedback controller with a predetermined order has been addressed based on the proposed stability
criterion (Theorem 3). The LMI-based approach of developing stabilizing controller is preserved, in spite
of the intricacy of assuming the most complete model of linear controller. The proposed stability and
stabilization theorems are applicable to FO-LTI systems with different fractional orders in one or both of
and
intervals. Eventually, some numerical examples are presented to show the
validity of the proposed theorems.
REFERENCES
1. Badri V, Tavazoei MS, “Achievable Performance Region for a Fractional-Order Proportional and
Derivative Motion Controller,” IEEE Trans. Ind. Electron., vol. 62, no. 11, pp. 7171–7180 (2015).
2. Badri V, Tavazoei MS, “Some Analytical Results on Tuning Fractional-Order [ProportionalIntegral] Controllers for Fractional-Order Systems,” IEEE Trans. Control Syst. Technol., vol. 24,
no. 3, pp. 1059–1066 (2016).
3. Podlubny I, “Fractional-order systems and
-controllers,” IEEE Trans. Autom. Control, vol.
44, no. 1, pp. 208–214 (1999).
4. Jin Y, Chen YQ, Xue D, “Time-constant robust analysis of a fractional order [proportional
derivative] controller,” IET Control Theory Appl., vol. 5, no. 1, pp. 164–172 (2011).
5. Kaczorek T, “Positive Linear Systems Consisting of n Subsystems With Different Fractional
Orders,” IEEE Trans. Circuits Syst. Regul. Pap., vol. 58, no. 6, pp. 1203–1210 (2011).
6. Wei Y, Hu Y, Song L, Wang Y, “Tracking Differentiator Based Fractional Order Model Reference
Adaptive Control: The
Case,” in 53rd IEEE Conference on Decision and Control, pp.
6902–6907 (2014).
7. Taghavian H, Tavazoei MS, “Robust stability analysis of uncertain multiorder fractional systems:
Young and Jensen inequalities approach,” Int. J. Robust Nonlinear Control (2017).
8. Skaar SB, Michel AN, Miller RK, “Stability of viscoelastic control systems,” IEEE Trans. Autom.
Control, vol. 33, no. 4, pp. 348–357 (1988).
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
Matignon D, “Stability Results For Fractional Differential Equations With Applications To Control
Processing,” in In Computational Engineering in Systems Applications, pp. 963–968 (1996).
Deng W, Li C, Lü J, “Stability analysis of linear fractional differential system with multiple time
delays,” Nonlinear Dyn., vol. 48, no. 4, pp. 409–416 (2007).
Diethelm K, The Analysis of Fractional Differential Equations: An Application-Oriented
Exposition Using Differential Operators of Caputo Type. Springer (2010).
Chen L, Chen G, Wu R, Tenreiro Machado J a., Lopes AM, Ge S, “Stabilization of Uncertain
Multi-Order Fractional Systems Based on the Extended State Observer,” Asian J. Control, DOI:
10.1002/asjc.1618 (2018).
Sabatier J, Moze M, Farges C, “LMI Stability Conditions for Fractional Order Systems,” Comput
Math Appl, vol. 59, no. 5, pp. 1594–1609 (2010).
Badri P, Amini A, Sojoodi M, “Robust fixed-order dynamic output feedback controller design for
nonlinear uncertain suspension system,” Mech. Syst. Signal Process., vol. 80, pp. 137–151 (2016).
Zhang H, Wang X, Lin X, “Stability and Control of Fractional Chaotic Complex Networks with
Mixed Interval Uncertainties,” Asian J. Control, vol. 19, no. 1, pp. 106–115 (2017).
Amini A, Mohammadi A, Asif A, “Event-based consensus for a class of heterogeneous multi-agent
systems: An LMI approach,” in 2017 IEEE International Conference on Acoustics, Speech and
Signal Processing (ICASSP), pp. 3306–3310 (2017).
Thuan MV, Huong DC, “New Results on Stabilization of Fractional-Order Nonlinear Systems via
an LMI Approach,” Asian J. Control, DOI: 10.1002/asjc.1644 (2018).
Moze M, Sabatier J, Oustaloup A “LMI Tools for Stability Analysis of Fractional Systems,”
IDETC/CIE2005 Proceedings, pp. 1611–1619 (2005).
Farges C, Moze M, Sabatier J, “Pseudo-state feedback stabilization of commensurate fractional
order systems,” Automatica, vol. 46, no. 10, pp. 1730–1734 (2010).
Lu J-G, Chen G, “Robust Stability and Stabilization of Fractional-Order Interval Systems: An LMI
Approach,” IEEE Trans. Autom. Control, vol. 54, no. 6, pp. 1294–1299 (2009).
Ma Y, Lu J, Chen W, “Robust stability and stabilization of fractional order linear systems with
positive real uncertainty,” ISA Trans., vol. 53, no. 2, pp. 199–209 (2014).
Sontag ED, Mathematical Control Theory: Deterministic Finite Dimensional Systems. Springer
Science & Business Media, (2013).
Matignon D, “Stability properties for generalized fractional differential systems,” ESAIM Proc.,
vol. 5, pp. 145–158 (1998).
Löfberg J, “YALMIP : a toolbox for modeling and optimization in MATLAB,” in 2004 IEEE
International Symposium on Computer Aided Control Systems Design, pp. 284–289 (2004).
Labit Y, Peaucelle D, Henrion D, “SEDUMI INTERFACE 1.02: a tool for solving LMI problems
with SEDUMI,” in Proceedings. IEEE International Symposium on Computer Aided Control
System Design, pp. 272–277 (2002).
Higham DJ, Higham NJ, MATLAB Guide: Second Edition. SIAM, (2005).
| 3cs.SY
|
arXiv:1610.00525v1 [math.AC] 3 Oct 2016
LINEARITY DEFECT OF THE RESIDUE FIELD OF SHORT LOCAL RINGS
RASOUL AHANGARI MALEKI
Abstract. Let (R, m, k) be a Noetherian local ring with maximal ideal m and residue field k.
The linearity defect of a finitely generated R-module M , which is denoted ldR (M ), is a numerical
measure of how far M is from having linear resolution. We study the linearity defect of the residue
field. We give a positive answer to the question raised by Herzog and Iyengar of whether ldR (k) < ∞
implies ldR (k) = 0, in the case when m4 = 0.
1. Introduction and notation
This paper is concerned with the notion of the linearity defect of the residue field of a commutative
Noetherian local ring. This invariant was introduced by Herzog and Iyengar [3] and has been
further studied by Iyengar and Römer [4], Şega [6] and Nguyen [5]. Let us recall the definition of
the linearity defect. Throughout this paper (R, m, k) will denote a commutative Noetherian local
ring with maximal ideal m and residue field k. Let
∂
∂n−1
∂
n
1
F : · · · → Fn −→
Fn−1 −−−→ · · · → F1 −→
F0 → 0
be a minimal complex (i.e. ∂i (Fi ) ⊆ mFi−1 for all i ≥ 0) of finitely generated free R-modules. Then
the complex has a filtration {Fp F}p≥0 with (Fp F)i = mp−i Fi for all p and i where, by convention,
mj = R for all j ≤ 0. The associated graded complex with respect to this filtration is called the
linear part of F and denoted by linR (F). Let N be an R-module. The notation Rg will stand for the
associated graded ring ⊕i≥0 mi /mi+1 and N g for the associated graded Rg -module ⊕i≥0 mi N/mi+1 N .
By construction, linR (F) is a graded complex of graded free Rg -modules and has the property that
g
linR
n (F) = Fn (−n), for all n. For more information about this complex, we again refer to [3] and
[4]. Let M be a finitely generated R-module. The linearity defect of M is defined to be the number
ldR (M ) := sup{i ∈ Z| Hi (linR (F)) 6= 0},
where F is a minimal free resolution of M . By definition, ldR (M ) can be infinite and ldR (M ) ≤ d is
finite if and only if (Syzd (M ))g has a linear resolution over the standard graded algebra Rg , where
Syzd (M ) is the dth syzygy module of M . In particular, ldR (M ) = 0 if and only if M g has a linear
resolution over Rg and then linR (F) is a minimal graded free resolution of M g . The notion of the
linearity defect can be defined, in the same manner, for graded modules over a standard graded
2010 Mathematics Subject Classification. 13D07 (primary), 13D02 (secondary) .
Key words and phrases. Minimal free resolutions, Linearity defect.
This research was in part supported by a grant from IPM (No. 94130028). This work was also supported by the
Iran National Science Foundation (INSF) grant No. 95001343.
1
2
RASOUL AHANGARI MALEKI
algebra A over a field k. In [3], the authors proved that if ldA (k) < ∞, then ldA (k) = 0. Motivated
by this known result in the graded case, the following natural question raised in [3].
Question 1. If ldR (k) < ∞, does it follow that ldR (k) = 0?
If Rg is Cohen-Macaulay, Şega [6] showed that the question has positive answer in the case that
R is a complete intersection. Also, she gave an affirmative answer when m3 = 0. Another positive
answer to the question is given by the author and Rossi [1] when R is of homogeneous type, that
Rg
is dimk TorR
i (k, k) = dimk Tori (k, k) for all i.
In this paper we show that this problem has an affirmative answer when m4 = 0. The proof
relies on the existence of a DG algebra structure of a minimal free resolution of residue field k.
2. Preliminaries and the main result
Şega provided an interpretation of linearity defect in term of vanishing of special maps. For each
n ≥ 0 and i ≥ 0 we consider the map
n+1
n
υin (M ) : TorR
) → TorR
i (M, R/m
i (M, R/m )
induced by the natural surjection R/mn+1 → R/mn . For simplicity, we set υin := υin (k) when
M = k.
Theorem 2.1. [6, Theorem 2.2] Let M be a finitely generated R-module and d be an integer. Then
the following conditions are equivalent.
(1) ldR (M ) ≤ d;
(2) υin (M ) = 0 for all i ≥ d + 1 and all n ≥ 0.
Remark 2.2. Let i ≥ 0. Assume that F is a minimal free resolution of a finitely generated R-module
′
M . Then by [6, 2.3 (2 )], the following statements are equivalent.
(1) υi1 (M ) = 0;
(2) if x ∈ Fi satisfies ∂i (x) ∈ m2 Fi−1 , then x ∈ mFi .
Let S be a unitary commutative ring. Given an S-complex C, we write |c| = i (the homological
degree of c) when c ∈ Ci . When we write c ∈ C we mean c ∈ Ci for some i. A (graded commutative)
DG algebra over S is a non-negative S-complex (D, ∂) with a morphism of complexes called the
product
µD : D ⊗ S D → D
a ⊗ b 7→ ab
satisfying the following properties:
(i) unital: there is an element 1 ∈ D0 such that 1a = a1 = a for a ∈ D;
(ii) associative: a(ba) = (ab)c for all a, b, c ∈ D;
(iii) graded commutative: ab = (−1)|a||b| ba ∈ D|a|+|b| for all a, b ∈ D and a2 = 0 when |a| is odd.
The fact that µ is a morphism of complexes is expressed by the Leibniz rule:
∂(ab) = ∂(a)b + (−1)|a| a∂(b)
For more information on DG algebras we refer to [2].
3
Remark 2.3. If (D, ∂) is a DG algebra over S. Using Leibniz rule, one can see that the subcomplex
of cycles Z(D) is a DG subalgebra of D and the boundaries B(D) is a DG ideal of Z(D). Thus the
product on D induces a product on the homology H(D) = Z(D)/B(D). In particular, ⊕i≥0 Hn (D)
is a graded module over commutative ring H0 (D).
Tate constructed a DG algebra (free) resolution of k. Furthermore, such a resolution can be
chosen to be minimal, see [2, Theorem 6.3.5], which we refer a minimal Tate resolution of k over
R.
The following lemma shows that the linear part of a minimal Tate resolution of k inherits a DG
algebra structure from that of the resolution.
Lemma 2.4. Let (F, ∂) be a minimal Tate resolution of k. Then linR (F) has a DG algebra structure
induced by that of F.
Proof. Let µF : F ⊗R F → F be a morphism of complexes which defines the product on F. Set
S = Rg . Since F is minimal we see that F ⊗R F is a minimal complex as well. Hence the morphism
induces a morphism of graded S-complexes linR (µF ) : linR (F ⊗R F) → linR (F) such that if i, n ≥ 0
and x∗ is the image of an element x ∈ mi (F ⊗R F)n in mi (F ⊗R F)n /mi+1 (F ⊗R F)n , then linR (µF )
maps x∗ into the image of µ(x) in mi Fn /mi+1 Fn . There is also a natural isomorphism of graded
S-complexes λ : linR (F) ⊗S linR (F) → linR (F ⊗R F) such that if i, j, n, m ≥ 0 and x ∈ mi Fn and
y ∈ mj Fm with images x∗ in mi Fn /mi+1 Fn and y ∗ in mj Fm /mj+1 Fm respectively, then λ maps
x∗ ⊗ y ∗ into the image of x ⊗ y in mi+j (F ⊗R F)n+m /mi+j+1 (F ⊗R F)n+m , see [4, Lemma 2.7].
Now, define (the product)
R
µlin
(F)
: linR (F) ⊗S linR (F) → linR (F)
as the composition linR (µF ) ◦ λ. Since µ satisfies conditions (i), (ii), (iii) of the definition of DG
R
algebras, one can see that µlin (F) satisfies the same properties as well. Therefore the linear part
of F is a DG algebra over S augmented to k.
Let m∗ denote the homogeneous maximal ideal of Rg . The following is a direct consequence of
the above lemma.
Corollary 2.5. If F is a minimal free resolution of k, then m∗ Hn (linR (F)) = 0 for all n.
Proof. The assertion follows from Remark 2.3 with considering the fact that H0 (linR (F)) = Rg /m∗ .
In what will follow, let (F, ∂) be a minimal free resolution of residue field k with differential
map ∂. The differential map of linR (F) which is induced by ∂ will be denoted by ∂ ∗ . We recall
that linR (F)n = Fng (−n). For any i, n ≥ 0 and x ∈ mi Fn , ∂ ∗ maps x + mi+1 Fn , the image of x in
mi Fn /mi+1 Fn , into the image of ∂(x) in mi+1 Fn−1 /mi+2 Fn−1 that is ∂(x) + mi+2 Fn−1 .
Proposition 2.6. Let d be an integer. If ldR (k) ≤ d, then the following hold.
(1) υd1 = 0.
∗ .
(2) m∗ Ker ∂d∗ = m∗ Im ∂d+1
4
RASOUL AHANGARI MALEKI
∗ .
Proof. For the simplicity, we set Z = Ker ∂d∗ and B = Im ∂d+1
(1) If υd1 6= 0, then there exists an element e ∈ Fd \ mFd such that ∂d (e) ∈ m2 Fd−1 , by 2.2 .
Let e∗ be the image of e in the quotient module Fd /mFd . Then ∂d∗ (e∗ ) = 0 and so e∗ is a cycle in
linR (F). Applying 2.5, we have m∗ Z ⊆ B and therefore m∗ e∗ ⊆ B. As e∗ is an element of a basis
of the free module Fdg (−d) and B ⊆ m∗ Fdg (−d), it is straightforward to see that m∗ e∗ is a direct
summand of B. By the hypothesis, B has a linear resolution. This implies that the same property
holds for m∗ e∗ . Therefore k has a linear resolution over Rg and consequently linR (F) is acyclic.
Hence υd1 = 0 and this is a contradiction.
For (2), it is enough to show that m∗ Z ⊆ m∗ B. First we claim that m∗ Z is generated in degree
at least d + 2 and m∗ Z ⊆ B. Indeed since υd1 = 0, applying Remark 2.2, one has Z ⊆ m∗ Fdg (−d)
and consequently m∗ Z is generated in degree at least d + 2. The second part of the claim follows
from Corollary 2.5.
On the other hand, B has a linear resolution, by the hypothesis. Hence B is generated by elements
of degree d + 1 and then all its elements of degree at least d + 2 contained in m∗ B. Now, putting
these two considerations together, we get m∗ Z ⊆ m∗ B.
Now, we are ready to prove our main result.
Theorem 2.7. Assume that R is Artinian with m4 = 0. If ldR (k) < ∞, then ldR (k) = 0.
Proof. Let d be a non-negative integer and ldR (k) ≤ d. We prove by descending induction on d.
The case where d = 0 is clear. Let d > 0. Applying Proposition 2.6, we have υd1 = 0. Since m4 = 0,
it follows from [6, Theorem 7.1] that υd2 = 0. Again since m4 = 0 obviously υdi = 0 for all i ≥ 3,
by the definition of the map υdi . Therefore, from 2.1 we get ldR (k) ≤ d − 1. This completes the
induction and finishes the proof.
Acknowledgments. The author would like to express great thanks to the referee for valuable
comments and suggestions which have improved the exposition of this paper. This research was in
part supported by a grant from IPM (No. 94130028). This work was also jointly supported by the
Iran National Science Foundation (INSF) and Alzahra University grant No. 95001343.
References
[1] R. Ahangari Maleki, M.E. Rossi, Regularity and linearity defect of modules over local rings, J. Commut. Algebra
6 (2014) 485-504.
[2] L.L. Avramov, Infinite free resolutions, in: Six Lectures on Commutative Algebra, Bellaterra, 1996, in: Progr.
Math., vol. 166, Birkhäuser, Basel, 1998, pp. 1-118.
[3] J. Herzog, S. Iyengar, Koszul modules, J. Pure Appl. Algebra 201 (2005) 154–188.
[4] S. Iyengar, T. Römer, Linearity defects of modules over commutative rings, J. Algebra 322 (2009) 3212-3237.
[5] H.D. Nguyen, Notes on the linearity defect and applications, arXiv:1411.0261.
[6] L.M. Şega, On the linearity defect of the residue field, J. Algebra 384 (2013) 276–290.
Rasoul Ahangari Maleki, School of Mathematics, Institute for Research in Fundamental Sciences
(IPM), P.O. Box: 19395-5746, Tehran, Iran
E-mail address: [email protected], [email protected]
| 0math.AC
|
End-to-End Prediction of Buffer Overruns from Raw Source Code via Neural
Memory Networks
arXiv:1703.02458v1 [cs.SE] 7 Mar 2017
Min-je Choi and Sehun Jeong and Hakjoo Oh and Jaegul Choo
Department of Computer Science and Engineering
Korea University, Seoul
{devnote5676, gifaranga, hakjoo_oh, jchoo}@korea.ac.kr
Abstract
Detecting buffer overruns from a source code is
one of the most common and yet challenging tasks
in program analysis. Current approaches based
on rigid rules and handcrafted features are limited in terms of flexible applicability and robustness due to diverse bug patterns and characteristics
existing in sophisticated real-world software programs. In this paper, we propose a novel, datadriven approach that is completely end-to-end without requiring any hand-crafted features, thus free
from any program language-specific structural limitations. In particular, our approach leverages a recently proposed neural network model called memory networks that have shown the state-of-the-art
performances mainly in question-answering tasks.
Our experimental results using source code samples demonstrate that our proposed model is capable of accurately detecting different types of buffer
overruns. We also present in-depth analyses on how
a memory network can learn to understand the semantics in programming languages solely from raw
source codes, such as tracing variables of interest,
identifying numerical values, and performing their
quantitative comparisons.
1
Introduction
Detecting potential bugs in software programs has long been
a challenge ever since computers were first introduced. To
tackle this problem, researchers in the domain of programming languages developed various techniques called static
analysis, which tries to find potential bugs in source codes
without having to execute them based on a solid mathematical framework [Cousot and Cousot, 1977]. However, designing a static analyzer is tightly coupled with a particular programming language, and it is mainly based on a rigid set of
rules designed by a few experts, considering numerous types
of possible program states and bug cases. Thus, even with
its slight syntax changes frequently found in real-world settings, e.g., several variants of ANSI C languages, a significant
amount of engineering effort is required to make a previously
designed analyzer applicable to the other similar languages.
To overcome these limitations, one can suggest datadriven, machine learning-based approaches as the rapid
growth of deep neural networks in natural language processing has proved its effectiveness in solving similar problems
such as defect predictions. Studies show that deep convolutional neural networks (CNNs) and recurrent neural networks
(RNNs) are capable of learning patterns or structures within
text corpora such as source codes, so they can be applied to
programming language tasks such as bug localization [Lam
et al., 2016], syntax error correction [Bhatia and Singh, 2016;
Pu et al., 2016], and code suggestion [White et al., 2015].
Despite their impressive performances at detecting syntaxlevel bugs and code patterns, deep neural networks have
shown less success at understanding how data values are
transferred and used within source codes. This semantic level
of understanding requires not only knowledge on the overall
structure but also the capability to track the data values stored
in different variables and methods. Although the aforementioned deep learning models may learn patterns and structures, they cannot keep track of how values are changed. This
restriction greatly limits their usefulness in program analysis
since run-time bugs and errors are usually much more difficult
to detect and thus are often treated with greater importance.
In response, we introduce a new deep learning model with
the potential of overcoming such difficulties: memory networks [Weston et al., 2015b; Sukhbaatar et al., 2015]. Memory networks are best described as neural networks with external memory ‘slots’ to store previously introduced information for future uses. Given a question, it accesses relevant memory slots via an attention mechanism and combines
the values of the accessed slots to reach an answer. While
long short-term memories (LSTMs) and earlier models also
have external memories, theirs tend to evolve as longer sequences of information are fed in to the network, thus failing
to fully preserve and represent information introduced at earlier stages. Memory networks on the other hand can preserve
the given information even during long sequences.
This unique aspect of memory networks makes it and its
variant models [Kumar et al., 2016; Henaff et al., 2016] perform exceptionally well at question answering tasks, e.g., the
Facebook bAbI task [Weston et al., 2015a], a widely-used
QA benchmark set. The structure of these tasks comprises
a story, a query, and an answer, from which a model has to
predict the correct answer to the task mentioned in the query
by accessing relevant parts of the given story. These tasks are
logical questions such as locating an object, counting numbers, or basic induction/deduction. All questions can be correctly answered by referring to appropriate lines of the given
story.
We point out that this task setting is in fact similar to that
of a buffer overrun analysis that requires the understanding of
previous lines in a source code to evaluate whether a buffer
access is valid. Both tasks require knowledge not only on how
each line works but also on how to select the best relevant
information from previous lines. It is this very situation at
which our work sets a starting point.
In this study we set the objective as demonstrating a datadriven model free of hand-crafted features and rules, and yet
capable of solving tasks with the complexity of buffer overrun analyses. We present how memory networks can be effectively applied to tasks that require the understanding of not
only syntactic aspects of a source code but also more complex
tasks such as how values are transferred along code lines. We
present how our models can learn the concept of numbers
and numerical comparison simply by training on such buffer
overrun tasks without any additional information. We also introduce a generated source code dataset that was used to compensate for difficulties we faced in our data-driven approach.
As far as our knowledge goes, our proposed approach is the
first to use deep learning to directly tackle a run-time error
prediction task such as buffer overruns.
In Section 2, we cover previous work related to our task.
In Section 3, we redefine our tasks, introduce our generated
dataset and its purposes, and propose characteristics of the
memory network model and how it is applied to this domain.
In Section 4, we report experimental results and further discuss the performance of memory networks and notable characteristics it learned during the process. In Section 5 we conclude our work and discuss future work as well as the potential of memory networks for future tasks.
2
Related Work
To improve traditional static analysis techniques in the programming language domain, data-driven approaches based
on machine learning have been recently studied. Obtaining
general properties of a target program, namely, invariants, is
one of the prime examples. When concrete data of target programs such as test cases or logs are available, data-driven approaches can be used to identify general properties [Sharma
et al., 2012; Sharma et al., 2013b; Sharma et al., 2013a;
Sankaranarayanan et al., 2008b; Sankaranarayanan et al.,
2008a; Nori and Sharma, 2013], similar to static analysis
techniques. This use case is particularly useful when a target
program has inherent complexity that makes contemporary
static analyzers to compromise either of precision and cost,
but is bundled with test cases that can cover most of cases.
Meanwhile, following the upsurge in the developing field
of neural computing and deep learning, many models have
been applied to natural language texts, especially in identifying language structure and patterns. Socher et al. [Socher et
al., 2013] introduced recursive neural networks which parse
a sentence into subsections. [Sutskever et al., 2014] proposed
RNNs that learn structures of long text sequences. Source
codes of a program can also be seen as a text corpus with its
own grammar structure, thus being applicable for such neural network models. [Karpathy et al., 2015] showed that a
character-level LSTM trained with Linux kernel codes is capable of detecting features such as brackets or sentence length
as well as generating simulated codes that greatly resemble
actual ones in syntactic structure. Motivated by such results
in the pattern discovery of source codes, several approaches
have been taken to solve practical issues in source code analysis. [Pu et al., 2016] and [Bhatia and Singh, 2016] gathered
data from programming assignments submitted for a MOOC
class to train a correction model which corrects syntax errors in assignments. [Huo et al., 2016] and [Lam et al.,
2016] applied attention-based CNN models to detect buggy
source codes. [Allamanis et al., 2014] learned coding styles
by searching for patterns with neural networks. While these
approaches proved that neural networks are capable of detecting patterns within codes, they are limited to detecting only
syntax errors or bugs and not the transition of values stored
inside variables or functions of a source code program.
Neural networks with external memories have shown better performances in inference or logical tasks compared to
contemporary models. Following the introduction of neural Turing machines [Graves et al., 2014] and memory networks [Weston et al., 2015b; Sukhbaatar et al., 2015], many
variants of these models were applied to various tasks other
than QA tasks such as sentiment analysis, part-of-speech tagging [Kumar et al., 2016], and information extraction from
documents [Miller et al., 2016]. Yet, so far there has been
no work that applies a memory network-based model to tasks
with the complexity of semantic analysis in source codes.
3
Model Description
In this section, we first provide the rationale for solving buffer
overruns as a QA task. We also introduce a source codebased training dataset that we designed. Lastly, we describe
the structure of our model which is based on the memory
network [Sukhbaatar et al., 2015] and how it predicts buffer
overruns from a source code.
3.1
Benchmark Source Code Generation
(a) bAbI task example
(b) buffer overrun code sample
Figure 1: Comparison of a bAbI and a buffer overrun tasks
Comparison of bAbI tasks and buffer overruns. We return to our statement that analyzing buffer overruns is similar
to solving bAbI tasks. Consider Fig. 1. The bAbI task shown
in Fig. 1(a) is given a story (lines 1-7) and a query (line 8).
A solver model understands this task by looking at ‘John’
void fun ()
{
int ent2=4;
char ent9[10]=“”;
char ent8[24]=“”;
memset(ent9,‘K’,10);
ent9[9]=‘\0’;
memset(ent8,‘R’,23);
ent8[23]=‘0’;
ent9[27]=‘H’;
return;
}
// end
(a) Level 1:
buffer access
direct
void fun ()
{
int ent4=30;
char ent2=‘x’;
char* ent9;
char ent5[10]=“”;
memset(ent5,‘K’,10-1);
ent5[10-1]=‘0’;
ent9=(char*)malloc
(42*sizeof(char))
ent9[0]=‘0’;
strcpy(ent9,ent5);
return;
} // end
void fun ()
{
int ent4=30;
char ent5[10]=“”;
char* ent9;
memset(ent5,‘M’,10-1);
ent5[9]=‘\0’;
ent9=(char*)malloc
(ent4*sizeof(char));
ent9[0]=‘\0’;
strcpy(ent9,ent5);
return;
}
// end
(b) Level 2: strcpy access
(c) Level 3: int allocation
void fun ()
{
int ent1=2;
ent1=25;
char* ent2;
char ent4[ent1]=“”;
ent2=(char*)malloc
(14*sizeof(char))
memset(ent4,‘K’,ent1);
ent2[0]=‘K’;
memcpy(ent2,ent4,
ent1*sizeof(char));
return;
} //end
(d) Level 4: memcpy
access, int reallocation
Figure 2: Different levels of buffer overrun tasks
and ‘where’ from the query and then attends the story to find
lines related to ‘John.’ Lines 3 and 5 are chosen as candidates. The model understands from the sequential structure
that line 5 contains more recent, thus relevant information. In
the end, the model returns the answer ‘garden’ by combining
the query and the information from line 5.
Meanwhile, the task of Fig. 1(b) is to discriminate whether
the buffer access made at line 6 is valid. Our analyzer first understands that its objective is to compare the size of the character array entity_2 and the integer variable entity_1. Next, it
searches for the length of entity_2 at line 3, where 53 is allocated to the variable. It also gains knowledge from line 2 that
entity_1 is equivalent to 70. The remaining task is to compare
the integer variables 53 and 70 and return an alarm (unsafe) if
the index exceeds the length of the character array. One can
think of lines 1-5 as a story and line 6 as a query, perfectly
transforming this problem into a bAbI task.
Limitations of test suites. Although test suites such as
Juliet Test Suite for C programming language [Boland and
Black, 2012] are designed for benchmarking buffer overrun and other program analysis tasks, the data is not diverse
enough. Code samples differ by only a small fraction such as
a different variable nested in a conditional statement or loop,
while a large portion of code appears repeatedly over several
samples. A data-driven model will inevitably learn from only
the small variations and ignore a large portion of the code
where much of the valuable information is stored.
Program structure. We tackle this problem of data inadequacy by generating our own training source code dataset.1
Our dataset adopts buffer access functions and initialization
methods from Juliet to maintain at least an equal level of task
complexity, while also preserving an underlying structure that
makes it applicable for deep learning approaches. Each sample is a void function of 10 to 30 lines of C code and consists
of three stages: initialization, allocation, and query. During
the initialization stage, variables are initialized as either characters, character arrays, or integers. At the allocation stage,
these variables are assigned values using randomly generated integers between 1 to 100. Buffer sizes are allocated
to character arrays with malloc and memset functions. At the
1
The generated dataset and generator codes are available at https://github.com/mjc92/buffer_overrun_
memory_networks
query stage, a buffer access is attempted on one of the allocated character arrays via a direct access on an array index
(Fig. 2(a)). We formulate this task into a binary classification problem where an ‘unsafe’ sign is returned if a character
array is accessed with a string or index that exceeds its size.
Naming conventions. We assume that a limited number of individual variables appear in each program sample. Each variable is given the name entity_n where n ∈
{i|0 5 i 5 Nupper , i ∈ Z} and Nupper is an integer set by
default to 10. Each n is assigned randomly to variables and
invariant of their introduced order or data type. One can
imagine a situation where an agent (variable) is given a fake
ID (entity name) for a particular task (sample). The agent
learns to complete the task with that fake ID, then discards
it upon task completion, and selects a new one for the subsequent task. In this manner, we can prevent entities from learning task-specific knowledge and instead train them as representations of universal variables which can replace any kind
of variable that appears in a program. We can easily apply our
model to real-life source codes using this naming convention
by simply changing the names of newly introduced variables
and methods to different entities.
Adding complexity. Our model has to adapt to more realistic source codes with complex structures. Possible settings
that complicate our task include
• selecting only the appropriate variables out of several
dummy variables,
• introducing different buffer access methods requiring
the comparison of two character arrays such as strcpy
or memcpy functions,
• allocating the sizes of character arrays not with integers
but indirectly with previously assigned integer variables,
• reallocating integer variables prior to or after their use in
allocating a character array.
We first assign a number of dummy variables to each sample
program. Each dummy variable is initialized and allocated
in the same manner as the ones actually used in the buffer
access. We include the use of strcpy (Fig. 2(b)) / memcpy
(Fig. 2(d)) functions for buffer accesses. We also add cases
where character arrays are allocated not directly by integers,
but indirectly with additionally introduced integer variables
(Fig. 2(c)). Given this setting, the model has to learn to store
the integer value allocated to the integer variable first, then
Mval
memory blocks
attention vector
Eval
search lines
query line
1 void fun()
2 {
3 int entity_19 = 70;
4 char entity_14[53] = “”;
5 char entity_7[73] = “”;
6 memset(entity_14,‘K’,52);
7 entity_14[52]=‘\0’;
8 entity_14[entity_9] = ‘r’;
9 return;
10 }
Eaddr
(a) Input encoding
response vector
o3
o2
o1
Maddr
answer
u3
u2
u1
p1 p2 p3
W
u0
(b)
Memory storage
(c) Memory search
(d)
Multiple hops and output
Figure 3: Our proposed memory network-based model for buffer overrun tasks
use that value to obtain the length of the character array. Finally, we add further cases where the additional integer variable itself is reallocated (Fig. 2(d)), either before or after it is
used to define the character array length. Now the model has
to learn to choose whether the previously assigned or reallocated value was used for allocating a character array.
Our generated source codes are equivalent to an expanded
version of Juliet test suite in terms of data flow, that is, the
flow of data values defined and used within a source code.
Compared to codes in Juliet which only define the source and
destination variables that will be used for buffer access, ours
include dummy variables which are defined and used similarly. The inclusion of various settings such as memory allocation using assigned integer variables instead of raw integers and reassignment of variables also increase data flow. In
overall, our source codes provide a tougher environment for
models to solve buffer overrun tasks than the existing Juliet
dataset as there are more variations to consider.
3.2
Model Structure
The overall structure of our model is displayed in Fig. 3.
Input encoding (Fig. 3(a)). The memory network takes
in as input a program code X consisting of n search
lines X1 , X2 , · · · , Xn and a single buffer access line or
query Xq . A single program line Xm is a list of words
l
2
1
. With V as the vocabulary size, we define
, · · · , wm
wm
, wm
l
xm as the V -dimensional one-hot vector representation of a
l
word wm
. We set an upper limit N for the max number of
lines a memory can store, and we pad zeros for the remaining
lines if a program is shorter than N lines.
Note that every word in the source code is treated as a word
token. This includes not only variable names (entity), type
definitions (int) and special characters, (‘[0 , ‘∗0 ), but also integers as well. This setting matches our concept of an endto-end model that does not require explicit parsing. While
it is possible to apply parsers to extract numbers and represent them differently from other word tokens, this would contradict our goal of applying a purely data-driven approach.
Treating integers as individual word tokens means that our
model will not be given any prior information regarding the
size differences between numbers, and thus has to learn such
numerical concepts by itself. We further discuss this in Section 4.
Next, we compute vector representations for each sentence
using its words. Each word is represented in a d-dimensional
vector using an embedding matrix Eval ∈ Rd×V . We also
multiply a column vector lj to the j-th word vector for each
word to allocate different weights according to word positions. This concept known as position encoding [Sukhbaatar
et al., 2015] enables our model to discriminate the different
roles of variables when two or more identical words appear
in a single sentence. Without such settings, our model may
fail to discriminate between source and destination variables
such as in a strcpy function. The memory representation mi
of line i consisting of J words and the k-th element ljk of the
position encoding vector lj ∈ Rd for word j in the line i are
obtained as
mi = Σtj=1 lj · Axji ,
(1)
ljk = (1 − j/J) − (k/d) (1 − 2j/J) ,
(2)
where ‘·’ is element-wise multiplication.
Memory storage (Fig. 3(b)).
Next, we allocate
our encoded sentences mi into matrices called memory
blocks. Fig. 3(b) shows two memory
blocks, the mem
ory value block Mval∈ RN ×d and the memory address
block Maddr ∈ RN ×d . Each sentence is allocated into one
row of memory block, namely a memory slot. Mval stores semantical information about the contents of a code line while
Maddr stores information for locating how much to address
each line. For this reason, sentences are encoded using two
different word embedding matrices, Eval and Eaddr for Mval
and Maddr , respectively.
Memory search (Fig. 3(c)). The query is encoded into a
representation using Eaddr . We denote the initial query embedding as u0 . By computing the inner products between the
query embedding and each slot of the memory address block,
then applying a softmax function to the resulting vector, we
obtain the attention vector p which indicates how related each
line is to the query. The i-th element of p is obtained as
T
pi = softmax u0 Maddr ,
(3)
with the softmax function as
softmax (zi ) = ezi /Σj ezi .
The response vector o is computed as in
o = Σi pi (Mval )i .
(4)
(5)
Table 1: Comparison on generated source codes. Inside brackets are the standard deviations
CNN
LSTM
Memory
network
acc
level 1
F1
auc
acc
level 2
F1
auc
acc
level 3
F1
auc
acc
level 4
F1
0.67
0.69
0.75
0.73
0.71
0.81
0.61
0.61
0.66
0.62
0.62
0.67
(0.01)
(0.02)
(0.01)
(0.02)
(0.02)
(0.01)
(0.01)
(0.03)
(0.01)
(0.01)
(0.03)
(0.01)
0.8
0.84
0.92
0.82
0.80
0.90
0.69
0.66
0.76
0.67
0.64
0.75
(0.01)
(0.01)
(0.00)
(0.01)
(0.02)
(0.01)
(0.00)
(0.01)
(0.01)
(0.01)
(0.02)
(0.01)
0.84
0.84
0.92
0.86
0.85
0.93
0.83
0.83
0.90
0.82
0.82
0.90
(0.01)
(0.01)
(0.01)
(0.01)
(0.02)
(0.01)
(0.02)
(0.02)
(0.02)
(0.02)
(0.02)
(0.02)
This vector contains information collected over all lines of
the memory value block according to their attention weights
obtained from the memory address block. This is equivalent
to searching the memory for different parts of information
with respect to a given query.
Multiple hops and output (Fig. 3(d)). The response vector o can be either directly applied to a weight matrix W to
produce an output, or added to strengthen the query u. In the
latter case, the query is updated as in Eq. (6) by simply adding
the response vector to the previous query embedding.
uk+1 = uk + ok
learning models for such tasks, we used existing deep learning models often used for text classification tasks as baselines. That is, we included a CNN for text classification [Kim,
2014] and a two-layer LSTM binary classifier. All models
were trained with Adam [Kingma and Ba, 2014] at a learning
rate of 1e-2. We used the classification accuracy, F1 score,
and the area under the ROC curve (AUC) as performance
metrics. We averaged the scores of the ten best cases with
the smallest training error.
Table 2: Different levels of test sets
(6)
Level 1
We repeat from Eq. (3) to obtain a new response vector. Our model iterates through multiple hops where at each
hop the desired information to be obtained from the memory
slightly changes. This accounts for situations where a model
has to first look for lines where an array is allocated, and then
gather information from lines stating the size of the variables
used for allocating the array size. The final output is a floating value ranging from 0 (unsafe) to 1 (safe), which we round
to the nearest integer to obtain a binary prediction result.
4
Experiments
In this section, we present both quantitative and qualitative
results of our experiments on model performance and learned
characteristics.
4.1
auc
Quantitative Evaluation
Experiment settings. Our main dataset consists of C-style
source codes discussed in Section 3.2. We used a single training set consisting of 10,000 sample programs. We generated
four test sets with 1,000 samples each and assigned them levels one to four, with a higher level indicating a more complex
condition (see Table 2). Samples ranged from 8 to 33 lines of
code, with an average of 16.01. A total of 196 unique words
appeared in the training set. A maximum of four dummy variables were added to each sample. We used random integers
between 0 and 100 for buffer allocation and access. We conducted every experiment on a Intel (R) Xeon (R) CPU E52687W v3 @ 3.10GHz machine equipped with two GeForce
GTX TITAN X GPUs. All models were implemented with
Tensorflow 0.12.1 using Python 2.7.1 on an Ubuntu 14.04 environment.
Model Comparison. We set our memory network to three
hops with a memory of 30 lines and the embedding size of
d = 32. As there has been no previous work on using deep
Direct buffer access
Access by strcpy / memcpy
Allocation by int variable
Reallocation of int variable
√
Level 2
√
√
Level 3
√
√
√
Level 4
√
√
√
√
Performance results shown in Table 2 demonstrate that all
models decrease in performance as task levels increase, due
to our level assignment. Of the deep learning models, only
memory networks performed consistently at a high level on
all four level settings with accuracy rates higher than 80%.
This is expected since their hops allow them to solve even
complex situations such as variable reallocation and different buffer access types. Meanwhile, CNNs failed to complete
even the simplest tasks since they cannot capture the sequential information in input sentences and instead apply convolutional filters to words of all regions on an equal basis. Any
positional information is discarded.
Memory networks require substantially shorter computing
time compared to other models, requiring an average of 0.63s
per epoch, while LSTMs and CNNs each require 7.13s and
13.21s. As for the latter models, the number of computations
is proportional to the number of words that appear in a code.
However, memory networks sum up all words of a sentence
to form a single representation, and thus computation time
relies on the number of lines instead of individual words. This
significantly reduces computation time.
Interestingly, LSTM models also performed well when set
to easier tasks. Results show that LSTMs performed comparably to memory networks, even equaling them on Level 1
tasks. However, its performance sharply dropped when used
on higher level tasks. This partial success of LSTMs relates
to the simple structure of Level 1 tasks. The size of the character array always appears before the index to access, so the
model can cheat by comparing the only two numbers that appear within the entire code. This cheating becomes obsolete
(a) Cosine similarities
(b) L2-norm distances
(c) Representations of learned
word embeddings
(d) Visualization with t-SNE
Figure 4: Visualizations of word embedding vectors of numbers 1-100. Red and blue indicate high and low values, respectively.
as higher-level tasks require knowledge only obtainable by
attending previous lines in a stepwise manner.
4.2
search
range
query
results
Qualitative Analysis
code lines
1st hop 2nd hop 3rd hop
void fun ()
0.0E+00 0.0E+00
0.0E+00
{
0.0E+00 0.0E+00
0.0E+00
int entity_1 = 31;
0.0E+00 0.0E+00
1.1E-22
entity_1 = 48;
0.0E+00 0.0E+00
1.0E+00
char entity_7[6] = "";
0.0E+00 1.0E+00
0.0E+00
char* entity_9;
4.8E-31 0.0E+00
0.0E+00
entity_9 = (char*)malloc(entity_1*sizeof(char));
1.0E+00 0.0E+00
0.0E+00
entity_9[0] = '0';
0.0E+00 0.0E+00
0.0E+00
memset(entity_7,'z',6-1);
0.0E+00 0.0E+00
0.0E+00
entity_7[6-1] = '0';
0.0E+00 0.0E+00
0.0E+00
memcpy(entity_9,entity_7,6*sizeof(char));
48>6?
Answer: Safe
Prediction: Safe
Prediction confidence: 99.99%
Figure 5: Prediction result with attention per hop
We further examine the performance of our memory network model and the steps it takes to obtain a correct answer.
We also present visualization results on how our model learns
the concepts of numbers and numerical comparison without
being explicitly supervised about such tasks.
Tracking hop-wise results. In order to prove that our
model solves the tasks in our desired manner, that is, by
attending and collecting relevant information from different
parts of the memory at different hops, we analyze individual
prediction cases by inspecting which parts of information our
model has obtained from taking each hop.
Fig. 5 displays an example of buffer overrun analysis using
our model. We can observe that when given a strcpy buffer
access as a query, the model’s initial attention shifts to the
sentence where the destination buffer (entity_3) is allocated.
The model decides here to next look for entity_9, which contains the size used for allocating to entity_3. During the next
hop it attends the line where the source buffer (entity_2) is
allocated and obtains data of 99, the size of entity_2. At the
last hop the memory network visits entity_9 and obtains 69.
After the three hops, the destination size 69 is compared with
source size 99, and being a smaller number, returns ‘unsafe’
as a result. The prediction confidence in Fig. 5 indicates how
close the predicted value is to the ground answer.
Numerical concepts automatically learned. Recall from
Section 3 that our model was not given any prior information
regarding the notion of quantitative values. Interestingly, our
model learned to compare between different numbers. Fig. 4
displays visualization results using only the word embedding
vectors corresponding to the 100 numbers.
Figs. 4(a) and (b) display the cosine similarities and the
L2-norm distances of all numbers from 1 to 100, with 1 at
the topmost left-hand side. The colors observed at the first
and third quadrants from both figures show that numbers
with large differences are trained to minimize cosine similarities while maximizing L2-norm distances, thus spacing
themselves apart. In contrast, similar numbers in the second
and fourth quadrants have opposite characteristics, meaning
they are similarly placed.
The word embedding vectors of numbers across all d dimensions as seen in Fig. 4(c) further demonstrate a clear sequential order between numbers. The highlighted column
forms a strong color spectrum starting from a low value which
gradually increases as the corresponding number increases
from 1 to 100. As all word embeddings were initialized with
random values at the beginning, this spectrum indicates that
our model learns by itself to assign such values for comparison purposes.
Last of all, Fig. 4(d) is a t-SNE representation of all word
embedding vectors. The black gradation indicates the word
embeddings of numbers, with denser colors indicating larger
numbers. We notice that they are embedded in a consistent direction in an increasing order. While this again shows
how our model learns numerical characteristics, we also discover that dots in red, which correspond to entities from Section 3, stand out. As mentioned earlier, entities correspond
to the variables that appear in source codes as integer variables or character buffers. This implies that our model learns
to train word embeddings differently according to their purposes within a code.
5
Conclusions and Future Work
In this work, we proposed a memory network-based model
for predicting buffer overruns in programming language analysis. Our work is the first to apply a deep learning-based approach to a problem in the field of program analysis that requires both syntactic and semantic knowledge. Performance
results show that memory networks are superior to other models in solving buffer overrun tasks across all difficulty levels. We also presented that our model successfully learns the
notion of numbers and their quantitative comparisons from
merely textual data in an end-to-end setting.
Our work has room to improve in many interesting aspects.
We can expand our model to cover different program analysis tasks such as pointer analysis, interval analysis, and flow-
sensitivity analysis, which share similar semantic natures. We
can apply advanced variants of memory networks to handle
various conditions in source codes such as if and for statements. Our knowledge of models learning numerical representations can further aid deep learning models compatible
with arithmetic and logical reasoning. All of these combined,
our work marks a stepping stone to a fully data-driven program analyzer.
References
[Allamanis et al., 2014] Miltiadis Allamanis, Earl T. Barr, Christian Bird, and Charles Sutton. Learning natural coding conventions. In Proceedings of the 22Nd ACM SIGSOFT International
Symposium on Foundations of Software Engineering, FSE 2014,
pages 281–293, New York, NY, USA, 2014. ACM.
[Bhatia and Singh, 2016] Sahil Bhatia and Rishabh Singh. Automated correction for syntax errors in programming assignments
using recurrent neural networks. CoRR, abs/1603.06129, 2016.
[Boland and Black, 2012] T. Boland and P. E. Black. Juliet 1.1
c/c++ and java test suite. Computer, 45(10):88–90, Oct 2012.
[Cousot and Cousot, 1977] Patrick Cousot and Radhia Cousot. Abstract interpretation: a unified lattice model for static analysis of
programs by construction or approximation of fixpoints. In Proceedings of the 4th ACM SIGACT-SIGPLAN symposium on Principles of programming languages, pages 238–252. ACM, 1977.
[Graves et al., 2014] Alex Graves, Greg Wayne, and Ivo Danihelka.
Neural turing machines. CoRR, abs/1410.5401, 2014.
[Henaff et al., 2016] Mikael Henaff, Jason Weston, Arthur Szlam,
Antoine Bordes, and Yann LeCun. Tracking the world state with
recurrent entity networks. CoRR, abs/1612.03969, 2016.
[Huo et al., 2016] Xuan Huo, Ming Li, and Zhi-hua Zhou. Learning unified features from natural and programming languages for
locating buggy source code. In Proceedings of the 25th International Joint Conference on Artificial Intelligence (IJCAI 2016),
pages 1606–1612, 2016.
[Karpathy et al., 2015] Andrej Karpathy, Justin Johnson, and
Li Fei-Fei. Visualizing and understanding recurrent networks.
CoRR, abs/1506.02078, 2015.
[Kim, 2014] Yoon Kim. Convolutional neural networks for sentence classification. In Proceedings of the 2014 Conference on
Empirical Methods in Natural Language Processing (EMNLP
2014), pages 1746–1751, 2014.
[Kingma and Ba, 2014] Diederik P. Kingma and Jimmy Ba. Adam:
A method for stochastic optimization. CoRR, abs/1412.6980,
2014.
[Kumar et al., 2016] Ankit Kumar, Ozan Irsoy, Peter Ondruska,
Mohit Iyyer, James Bradbury, Ishaan Gulrajani, Victor Zhong,
Romain Paulus, and Richard Socher. Ask me anything: Dynamic
memory networks for natural language processing. In ICML,
2016.
[Lam et al., 2016] An Ngoc Lam, Anh Tuan Nguyen, Hoan Anh
Nguyen, and Tien N. Nguyen. Combining deep learning with
information retrieval to localize buggy files for bug reports. In
Proceedings - 2015 30th IEEE/ACM International Conference
on Automated Software Engineering, ASE 2015, pages 476–481.
IEEE, 2016.
[Miller et al., 2016] Alexander H. Miller, Adam Fisch, Jesse
Dodge, Amir-Hossein Karimi, Antoine Bordes, and Jason Weston. Key-value memory networks for directly reading documents. CoRR, abs/1606.03126, 2016.
[Nori and Sharma, 2013] Aditya V Nori and Rahul Sharma. Termination proofs from tests. In Proceedings of the 2013 9th Joint
Meeting on Foundations of Software Engineering, pages 246–
256. ACM, 2013.
[Pu et al., 2016] Yewen Pu, Karthik Narasimhan, Armando SolarLezama, and Regina Barzilay. sk-p: a neural program corrector
for moocs. In Companion Proceedings of the 2016 ACM SIGPLAN International Conference on Systems, Programming, Languages and Applications: Software for Humanity, pages 39–40,
New York, New York, USA, 2016. ACM Press.
[Sankaranarayanan et al., 2008a] Sriram
Sankaranarayanan,
Swarat Chaudhuri, Franjo Ivančić, and Aarti Gupta. Dynamic
inference of likely data preconditions over predicates by tree
learning. In Proceedings of the 2008 international symposium
on Software testing and analysis, pages 295–306. ACM, 2008.
[Sankaranarayanan et al., 2008b] Sriram
Sankaranarayanan,
Franjo Ivančić, and Aarti Gupta. Mining library specifications
using inductive logic programming. In Proceedings of the
30th international conference on Software engineering, pages
131–140. ACM, 2008.
[Sharma et al., 2012] Rahul Sharma, Aditya V Nori, and Alex
Aiken. Interpolants as classifiers. In International Conference
on Computer Aided Verification, pages 71–87. Springer, 2012.
[Sharma et al., 2013a] Rahul Sharma, Saurabh Gupta, Bharath Hariharan, Alex Aiken, Percy Liang, and Aditya V Nori. A data
driven approach for algebraic loop invariants. In European Symposium on Programming, pages 574–592. Springer, 2013.
[Sharma et al., 2013b] Rahul Sharma, Saurabh Gupta, Bharath
Hariharan, Alex Aiken, and Aditya V Nori. Verification as learning geometric concepts. In International Static Analysis Symposium, pages 388–411. Springer, 2013.
[Socher et al., 2013] Richard Socher, Alex Perelygin, and Jy Wu.
Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on
Empirical Methods in Natural Language Processing (EMNLP
2013), pages 1631–1642, 2013.
[Sukhbaatar et al., 2015] Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory networks. In
Advances in Neural Information Processing Systems, pages 1–11,
2015.
[Sutskever et al., 2014] Ilya Sutskever, Oriol Vinyals, and Quoc V
Le. Sequence to sequence learning with neural networks. In Advances in Neural Information Processing Systems, pages 3104–
3112, 2014.
[Weston et al., 2015a] Jason Weston, Antoine Bordes, Sumit
Chopra, and Tomas Mikolov. Towards ai-complete question answering: A set of prerequisite toy tasks. CoRR, abs/1502.05698,
2015.
[Weston et al., 2015b] Jason Weston, Sumit Chopra, and Antoine
Bordes. Memory networks. In Int’l Conf. on Learning Representations, pages 1–15, 2015.
[White et al., 2015] Martin White, Christopher Vendome, Mario
Linares-Vásquez, and Denys Poshyvanyk. Toward deep learning
software repositories. In Mining Software Repositories (MSR),
2015 IEEE/ACM 12th Working Conference on, pages 334–345.
IEEE, 2015.
| 9cs.NE
|
arXiv:1712.04794v3 [math.GR] 15 Jan 2018
Generators and closed classes of groups
Ramón Flores and José L. Rodrı́guez
January 15th, 2017
Abstract
We show that in the category of groups, every singly-generated class which is
closed under isomorphisms, direct limits and extensions is also singly-generated
under isomorphisms and direct limits, and in particular is co-reflective. We also
establish several new relations between singly-generated closed classes.
1
Introduction
In a category of R-modules over a certain ring R, the study of the closure of different
classes under different operations has been a hot research topic for decades (see [15] for
a thorough approach). The operations include extensions, subgroups, quotients, etc. In
the case of R = Z, we are dealing, of course, with operations in the category of abelian
groups.
There is another different but related problem. Given a collection D of objects in a
class C, the smallest class that contains D and is closed under certain operations is called
the class generated by D under these operations. When this class is equal to C, it is
said that C is generated by D under the operations, and the elements of D are called
generators of C. In this situation, if D consists of only one object, the class C is said to be
singly-generated under the operations. Now the question is clear: given a class C defined
0
Mathematics Subject Classification (2010): Primary: 20E34; Secondary: 20E22, 55P60.
Keywords: co-reflections, cellular covers, direct limits, extensions, radicals.
The first author was partially supported by FEDER-MEC grants MTM2010-20692 and MTM-201676453-C2-1-P. Second-named author was supported by grant MTM2016-76453-C2-2-P, and Grant FQM213 of the Junta de Andalucı́a.
1
by a certain property, is it possible to construct a generator for it? A standard reference
for the topic is [1].
In this paper, we will deal with different versions of these two problems in the category
of groups. Rather than from the vast literature in Module and Category Theory related to
these topics, our motivation and strategy come from the field of Homotopy Theory. From
the nineties, A.K. Bousfield, E. Farjoun and W. Chachólski ([4], [9], [13]) among others,
developed an ambitious program whose main goal was to classify spaces and spectra in
terms of classes defined by their generators and certain operations, particularly homotopy direct limits (cellular classes) and fibrations (periodicity classes). Taking account of
this framework, the fundamental group functor gives a natural (but non-trivial) way to
understand closed classes of groups from the same viewpoint.
We are mainly interested in three classes of operations in the class of groups, namely
taking direct limits, constructing extensions from a given kernel and cokernel, and taking
quotients. The relation between the closed classes under these operations has been already
studied in [10] from a different point of view. We focus on singly-generated classes, and
in fact our first main result is a generation result. Given a group A, we denote by C(A)
the class generated by A under direct limits, and by C(A) the class generated by A under
direct limits and extensions. Then:
Theorem 3.7 For any group A, there exists a group B such that C(A) = C(B).
Moreover, we are able to describe B in terms of groups of the class C(A). This
statement allows to generalize results of [18], where the case in which the Schur multiplier
of A is trivial was studied, and also of [2] for the case of acyclic groups. It can also be
seen as the group-theoretic version of the main theorem of [11], and it is in the spirit (see
last section) of the generators of radicals described in [8].
Section 4 is devoted to establish relations between the closed classes under the operations mentioned above. The hierarchy is the following:
C(A)
/ C(A)
_
_
Cq (A)
/ Ct (A)
where Cq (A) is the closure of C(A) under quotients and Ct (A) is the closure of C(A) under
quotients and extensions. All these classes turn out to be co-reflective. Furthermore, we
2
find an example where the four classes are distinct (see Example 4.7).
The co-reflections associated to the classes C(A), Cq (A) and Ct (A) are given respectively
by the A-cellular cover CA , the A-socle SA and the A-radical TA , which are well-known
notions in Group/Module Theory (see definitions in Section 2). However, for the class
C(A), the existence of the corresponding co-reflection DA had been previously established
[18] only for groups for which there exists a Moore space, and this goal was achieved
only by topological methods. This gap in the theory is filled in this paper making use of
Proposition 3.7, that guarantees the existence of DA for any group A. In general, given a
group G, the philosophy is the following: a singly-generated class C and the co-reflection
F associated to a certain operation, the kernel of the homomorphism F (G) → G measures
how far G is from belonging to the class generated from C under the operation. We prove
how the co-reflections define the corresponding closed classes, describe when possible the
kernels, and discuss, for different instances of the generator, the (in)equality between the
corresponding classes. In particular, we prove the following:
Proposition 4.10 Suppose that for a group A the class C(A) is closed under extensions,
i.e. C(A) = C(A). Then Cq (A) = Ct (A).
The converse (stated as Question 4.12) is not true in general. Section 5 is devoted to
give a counterexample that involves the Burnside idempotent radical and the Thompson
group. It is an interesting question to find groups for which this converse is true. In
fact, we are not even aware of any example of abelian group A such that SA = TA , see
Proposition 4.9 and the discussion that precedes it.
All the classes that appear in this paper will be supposed to be closed under isomorphisms.
2
(Co)reflections
In this section we introduce the main notions that will be used throughout the rest of our
paper. We will start with the classical categorical definition of co-reflection, that will be
crucial for us.
Definition 2.1 Given a inclusion of categories C ⊆ C ′ , where C is full in C ′ , C is coreflective in C ′ if the inclusion functor possesses a right adjoint, called co-reflection.
3
Given a co-reflective category, the co-reflection is unique up to natural isomorphism.
A good survey about the topic from the categorical point of view can be found in [1].
The following definitions will be of great interest, and will be used in particular in the
last section to define some interesting co-reflective subcategories of groups.
Definition 2.2 Let A and H be groups.
1. The A-socle SA H is the subgroup of H generated by the images of all the homomorphisms A → H.
2. The A-radical TA H of H is the largest normal subgroup of H such that
Hom(A, H/TA H) = 0.
3. The A-cellular cover of H is the unique group CA H ∈ C(A), up to isomorphism,
that can be constructed from copies of A by iterated direct limits, and such that
there exists a homomorphism CA H → H which induces a bijection Hom(A, CA H) ≃
Hom(A, H). Observe that this is equivalent to Hom(L, CA H) ≃ Hom(L, H) being
a bijection for every L ∈ C(A).
Remark 2.3 Note that similar definitions are possible if we change A by a collection
A of groups indexed over a certain set I. In this case, the generator would be the free
product of isomorphism classes of elements of A. Moreover, it is possible to give versions
of the socle and the radical relative to a class of epimorphisms of groups. See [1] for a
general treatment of these topics.
Let us recall more constructive versions of some of these groups. The A-radical TA H
S
is the union TA (H) = i T i of a chain starting at T 0 = SA (H) and defined inductively
S
by T i+1 /T i = SA (H/T i ), and T λ = α T α if λ is a limit ordinal. In turn, the A-cellular
cover CA G is constructed by taking coequalizers of free products of copies of A, being the
homomorphisms induced by coequalizers. See details in ([18], Proposition 2.1).
In the concrete case of the A-cellular cover, there is the following description, which
is very useful:
Theorem 2.4 ([18], Theorem 2.7) Given groups A and H, there is a central extension
K ֒→ CA (H) ։ SA (H),
(2.1)
such that Hom(Aab , K) = 0 and the natural map Hom(A, H) → H 2 (A; K) is trivial.
Moreover, this extension is universal with respect to this property.
4
The group constructions described in this section will give rise to different co-reflections,
as we will see in Section 4.
We will also need to recall some basics on (co)-reflections of spaces. Analogously to
groups, given a pointed space M, the cellular class C(M) generated by M is the smallest
class that contains M and is closed under pointed homotopy colimits. The corresponding
co-reflection CWM X → X for a pointed space X, called the M-cellular cover of X, was
explicitly described by Chachólski in his thesis [9] in the following way: consider the
W
homotopy cofiber Cof M X of the evaluation map [M,X]∗ M → X. Denote by PM the Mnullification functor, i.e. the homotopical localization with respect to a constant map (see
[13], chapter 1). Then the homotopy fiber of Y → PΣM Cof M X is homotopy equivalent to
CWM X.
In [18] the relation between cellular covers of groups and spaces was described: if M
is a 2-dimensional complex with π1 M = A, then π1 CWM K(N, 1) = CA N for every group
N.
3
Generating classes of groups
As said in the introduction, one of the goals of this paper is to describe a generator for
the class C(A), the smallest class that contains A and is closed under direct limits and
extensions. The elements of this class are usually called A-acyclic groups. In this section
we will explain how to construct a generator for this class:
Definition 3.1 Given a class C of groups closed under direct limits, a cellular generator
for the class C is a group B such that C = C(B).
In our context, C will usually be the class C(A) for some group A. These generators are
not unique, and in fact several models for particular instances of A have been proposed
in the literature ([2], [18]). We now present a general construction, which is inspired
in the methodology of Chachólski-Parent-Stanley [11], that solved successfully the same
problem in the category of topological spaces. Concretely, we translate to the grouptheoretic context their idea of small space; aside this, the proofs are very different, as
their methods are homotopical in nature.
We start with a set-theoretic definition:
Definition 3.2 For a given infinite regular cardinal λ a group H is called λ-small (or
small with respect to λ) if #H is smaller than the cofinality of λ.
5
Given a class C of groups closed under direct limits, let C[λ] be the subclass of groups
H in C such that for each λ-small group K, every group homomorphism K → H factors
as K → C → H for some λ-small group C in C. The following holds:
Proposition 3.3 Let λ be an infinite regular cardinal. A group H is in C[λ] if and only
if it is isomorphic to a direct limit of λ-small groups in C.
Proof. Suppose H ∈ C[λ]. Consider the direct system of all λ-small subgroups of H that
belong to C, ordered by inclusion, and let G be its direct limit. We will check that G = H.
By contradiction, suppose there exists an element a ∈ G\H, and call A the subgroup that
φ0
it generates. The group A is λ-small, and hence there is a factorization A → C0 → H of
the inclusion A → H such that C0 is λ-small and belongs to C. Call K0 the image of φ0 ,
which is clearly λ-small. If K0 = A, then A is a retract of C0 , hence belongs to C (as a
retract is a direct limit) and therefore a ∈ G, which is a contradiction, so we are done.
If not, then A
K0 , and we may proceed in the same way with the inclusion K0 < H,
φ1
obtaining a factorization K0 → C1 → H. Iterating the process if necessary, we obtain a
string of homomorphisms such that for every i there is a strict inclusion Ki
Ki+1 :
A ֒→ C0 ։ K0 ֒→ C1 ։ K1 ֒→ C2 ։ K2 ֒→ . . . ,
where Ci is λ-small in C for every i. As this string of arrows is a factorization of the
inclusion A → H, it must eventually stabilize, and this implies that for a certain i
(possibly transfinite ordinal) we have Ki = Ci = Ki+1 , and hence Ki is a retract of Ci
and belongs to C. Again, this yields a ∈ G, contradiction, and then G = H.
Conversely, assume that H is a direct limit of λ-small groups in C, and let us prove
that H belongs to C[λ]. As every direct limit of groups is a composition of limits over
filtered categories and push-outs, we only need to check these two cases.
If H is the direct limit of the λ-small groups {Gα } over a filtered category, a standard
compactness argument states that every homomorphism K → H factors through Gα for a
certain α. Hence H belongs to C[λ]. Now if H is a pushout of groups, the short argument
of the proof of Lemma 4.8.4 in [11] applies to the group-theoretic case, and implies that
H ∈ C[λ]. So we are done.
✷
Now we can establish closure properties of C[λ]. The first is a straightforward consequence of the previous result:
6
Corollary 3.4 If C is closed under direct limits, C[λ] is so for every infinite regular
cardinal λ.
Proof. Let H be a direct limit of groups in C[λ]. By the previous proposition, H is a
direct limit of λ-small groups in C, and, every direct limit of λ-small groups in C belongs
to C[λ], so we are done.
✷
Observe that the proof establishes that the λ-small groups in C generate C[λ] as a
cellular class.
The following result is the group-theoretical analogue of [11, Theorem 6.1], although
the proof is very different and purely group-theoretical.
Lemma 3.5 If C is any class of groups closed under extensions, then C[λ] is also closed
under extensions.
π
Proof. Let 1 → X → Y → Z → 1 be a short exact sequence with X and Z in C[λ]. We
have to prove that Y ∈ C[λ]. Let j : K → Y be any group homomorphism, with K λsmall. Since Z ∈ C[λ], πj factors through some λ-small group Z0 in C. Taking the pullback
π0
along Z0 → Z we obtain a short exact sequence 1 → X ∼
= Ker π0 → Y0 → Z0 → 1. Let
j0 : K → Y0 be the induced homomorphism and consider K0 the subgroup of Y0 generated
by j0 (K) and s(Z0 ) where s is a chosen section (non necessarily homomorphism) of π0 .
The intersection K ′ = K0 ∩ X is λ-small because K0 is λ-small, and it is normal in K0
with K0 /K ′ ∼
= Z0 . Since X ∈ C[λ], the inclusion K ′ ֒→ X factors as K ′ ֒→ X0 → X for
some small group X0 ∈ C.
Let now P0 be the push-out of K0 ← K ′ → X0 , and denote by (X0 )N the normal
closure of X0 in P0 . The quotient P0 /(X0 )N ∼
= Z0 , and P0 is a λ-small group in C, because
both K ′ and X0 are λ-small. Observe that (X0 )N need not be in C. In any case, it is λsmall and therefore the inclusion (X0 )N → X factors through some λ-small group X1 ∈ C.
Iterating this process we get an increasing chain of short sequences over X → Y → Z
which are, alternatively, in C, but not exact, and exact but not in C. Here the subscript
“∞” denotes the colimit of each linear sequence of homomorphisms.
X0 → (X0 )N → X1 → ... → Xi → (Xi )N → ... → X∞ = (X∞ )N
↓
↓
P0 →
P0
↓
↓
Z0 →
Z0
↓
↓
→ P1 → ... → Pi →
↓
↓
→ Z0 → ... → Z0 →
7
↓
Pi
↓
→ ... →
↓
Z0
P∞
↓
→ ... →
Z0
In the limit we get a short exact sequence 1 → X∞ = (X∞ )N → P∞ → Z0 → 1, where
X∞ is λ-small and is in C, hence P∞ is λ-small and is in C too, because C is closed under
extensions. Hence, K → Y factors K → P∞ → Y , and this shows that C[λ] is closed
under extensions.
✷
For a given cardinal λ, let Bλ the free product of a set of representatives of isomorphism
classes of λ-small groups in C(A). We will need the following result:
Lemma 3.6 For every infinite regular cardinal, C(A)[λ] = C(Bλ ).
Proof. If K is an element of C(A)[λ], it is by Proposition 3.3 a direct limit of λ-small
groups in C(A). But every such λ-small group is a retract of Bλ , so it belongs to C(Bλ ),
and hence so does K since C(Bλ ) is closed under direct limits. Conversely, every group
H in C(Bλ ) is an (iterated) direct limit of copies of Bλ , and in particular a direct limit of
λ-small groups, so H ∈ C(A)[λ] and we are done.
✷
Now we are ready to prove the main theorem in this section:
Theorem 3.7 For any group A, there exists a group B such that C(A) = C(B).
Proof. Let A be any group, λ an infinite regular limit cardinal with cf (λ) ≥ #A.
Then, by Lemma 3.6, we have C(A)[λ] = C(B), where B = Bλ . In particular, the class
C(A)[λ] is closed under direct limits. The conditions over λ imply that A is λ-small,
and hence C(A) ⊆ C(A)[λ]. But by Lemma 3.5 C(A)[λ] is closed under extensions, so
C(A) ⊆ C(A)[λ]. On the other hand, C(A)[λ] is a subclass of C(A) by definition, and the
result follows.
✷
It is likely that the assumptions over the cardinals may be relaxed, but here we are
not interested in some models of the cellular generators, but rather in their existence. In
the next section we will use different cellular generators for concrete computations and in
particular, and perhaps more important, to define in general the co-reflection associated
to the class C(A) for every group A.
4
4.1
Closed classes
Classes of groups and associated co-reflections
Let A be a group. The goal of this section is to analyze the relations between four different
classes that contain the group A, namely:
8
• The cellular class C(A), the closure of {A} under direct limits. Their elements are
called A-cellular groups.
• The socular class Cq (A), the closure of {A} under direct limits and quotients. Their
elements are called A-generated or A-socular groups.
• The radical class Ct (A), the closure of {A} under direct limits, extensions and quotients. Their elements are called A-constructible or A-radical groups.
• The acyclic class C(A), the closure of {A} under direct limits and extensions. Their
elements are called A-acyclic groups.
The hierarchy between these classes can be depicted in the following way:
C(A)
/ C(A)
_
_
Cq (A)
(4.2)
/ Ct (A)
We use in particular the terminology of W. Chachólski, who studied the problem in the
non singly-generated case in [10].
It is known that a group G is A-cellular if and only if CA G = G ([18], section 2). We
will check now that similar statements hold for A-generated and A-constructible groups.
Probably this result is known to experts, but we do not know any explicit proof of it. We
offer purely algebraic arguments here.
Proposition 4.1 Let A and G be groups. Then:
• G is A-generated if and only if SA G = G.
• G is A-constructible if and only if TA G = G.
Proof. For the first part, it is clear that A is A-generated. Now, if G is a direct limit
of A-generated groups, it is in fact a quotient of a free product of copies of A-generated
groups (which is A-generated itself), and then it is A-generated. Finally, every quotient
of an A-generated group is A-generated. Conversely, if SA G = G, G is a quotient of a
free product of copies de A (by definition of socle), and hence belongs to the closure of A
under direct limits and quotients.
9
Let us check the second statement. Consider the class C of groups such that G = TA G.
Clearly A belongs to C, so let see that C is closed under quotients and extensions. For
the first, recall that for every homomorphism of groups f : H → G, f (TA H) ⊆ TA G.
Then, if H = TA H and G is a quotient of H, G = TA G, and hence G belongs to C. On
the other hand, assume that G is an extension G1 → G → G2 , with G1 and G2 in C. As
G1 = TA G1 , G1 ⊆ TA G. Hence, there is a projection G2 → G/TA G. But G2 = TA G2 and
TA (G/TA G) = 0, thus G/TA G = 0 and G = TA G. So C is also closed under extensions,
and then Ct (A) ⊆ C.
Now assume G = TA G. Then G is constructed out of A by means of A-socles, extensions and a direct union, and thus, by the previous case and the fact that a direct union
is a direct limit, G belongs to the closure of A by direct limits, quotients and extensions.
So C ⊆ Ct (A), and we are done.
✷
The previous result provides co-reflections for the respective classes.
Corollary 4.2 In the previous notation, the categories C(A), Cq (A) and Cr (A) are coreflective. The co-reflections are respectively given by the A-cellular cover, the A-socle
and the A-radical.
✷
In this context, the next question is immediate: giving a group A, is the class C(A)
co-reflective? We will see next that Theorem 3.7 gives a general answer to this question.
A partial solution was obtained in [18], where the authors assume the existence of a 2dimensional Moore space for A in order to construct the co-reflection. Recall that given
a group A, a Moore space of type M(A, 1) is a CW-complex X such that π1 X = A and
H2 X = 0. Then the following is proved:
Theorem 4.3 ([18], Theorem 3.1) Let A be the fundamental group of a two-dimensional
Moore space. Then the inclusion C(A) ֒→ Groups admits a right adjoint DA . Furthermore, for each group H, there is a central extension
L ֒→ DA (H) ։ TA (H)
such that Hom(Aab , L) = Ext(Aab , L) = 0, and this extension is universal (initial) with
respect to this property.
This result presents interesting features. First, it gives a manageable way to compute
the value of the co-reflection DA H over any group H, as L is identified as the fundamental
10
group of the homotopy fiber of the M(A, 1)-nullification of K(TA H, 1), i.e. localization
with respect to the constant map over M(A, 1) (see [13], chapter 1). Moreover, the
similarity with Theorem 2.4 should be remarked, and in fact for every group H there exists
a natural homomorphism CA N → DA N which induces a map between the corresponding
extensions, and in particular the inclusion SA N ֒→ TA N.
Remark that the co-reflection is constructed here in a strongly non-algebraic but homotopical way. This homotopical nature of the construction made it impossible to define
DA for a general A, as it depended on the existence of a two-dimensional Moore space
for the group. We will avoid the difficulty by defining DA as a cellular cover in a pure
group-theoretical way.
We present now our general construction of the co-reflection DA for every group A.
Proposition 4.4 Let A be a group, and B a cellular generator of the class C(A). Then
a group G is A-acyclic if and only if CB G = G.
Proof. It is a straightforward consequence of Theorem 3.7.
✷
Corollary 4.5 For every group A, the class C(A) is co-reflective, and the co-reflection is
given by CB .
As co-reflections are unique, the value of CB does not depend on the concrete model
for the generator B; hence, in the following this co-reflection will be denoted by DA .
Note that there is no ambiguity in the notation, as again by uniqueness, the value of the
co-reflection coincides with the one defined in Theorem 4.3, for the instances of A for
which that theorem is valid. Finally, remark that our model for B in the previous section
is built in a pure algebraic way, and hence it is the construction of DA .
The following proposition provides examples of concrete computations of DA , in particular when H2 A 6= 0.
Proposition 4.6 Let H and G be finite non-trivial groups, with H simple, |H|
|G|
and such that there is exactly a subgroup of G that is isomorphic to H. Then DH G = H,
and G/H does not belong to C(H).
Proof. As H is simple and normal in G, SH G = TH G = H. But for every group A we
always have DA = DA TA , and hence
DH G = DH TH G = DH H = H,
11
as H is clearly H-acyclic. Moreover, if G/H belonged to C(H), then G would also belong,
and that is impossible because DH G 6= G.
✷
As said above, this result gives information for classes C(A) with H2 A 6= 0, a very
difficult issue without a general construction of DA . For example, if H = An , G = Σn
and n ≥ 5, the proposition implies that DAn Σn = An , and hence neither Σn nor Z/2 can
be constructed out of An by direct limits and extensions.
4.2
Comparing the classes
From now on we will concentrate in measuring the difference between the classes above.
Let F be one of the four co-reflections. Given groups A and G, F G measures in general
how close the group G is from belonging to the corresponding class. Moreover, the
homomorphisms SA G → G, TA G → G and SA G → TA G are always injective, while the
homomorphisms DA G → TA G and CA G → SA G are always surjective, so we have a
commutative diagram:
CA G
DA G
//
SA G
_
//
TA G
(4.3)
Note that in general, CA G → DA G is neither surjective nor injective.
Using an example, we start by showing that the four inclusions of classes in the diagram
4.2 above are strict in general. Of course, the previous results imply that two such classes
are identical if and only if their associated co-reflections are the same.
Example 4.7 Consider the group
G = ha, b, c, d | a4 = b4 = c4 = d4 = 1, abab = cdcdi.
This group is generated by order 4 elements, and hence TZ/2 G = G. On the other hand,
it is easy to check that the Z/2-socle of G is generated by the squares of the generators,
and then
SZ/2 G = he, f, g, h | e2 = f 2 = g 2 = h2 = 1, ef ef = ghghi.
Observe that SZ/2 G is the group N described in Example 2.11 of [18]. Hence, as CZ/2 G =
CZ/2 SZ/2 G, the cellular cover of G is an extension of SZ/2 G by Z. Finally, the same
12
reasoning of the mentioned example proves that DZ/2 G is an extension of G by Z[1/2].
As these four groups are different, we have that the four inclusions of classes are strict in
the case A = Z/2.
Next we will discuss when the socular class Cq (A) is closed under extensions, and hence
equal to the radical class Ct (A). For example, if A = Z/p for some prime p, it is known
that there are many groups G such that G/Sp G has again p-torsion, and same happens if
we change Z/p by ∗j≥1Z/pj ; this contrasts with the case of abelian groups, in which for
L
A = j≥1 Z/pj , TA N = SA N for every abelian group.
When A is the additive group of the rational numbers, the situation is similar. If N
is abelian, SQ N = TQ N, because the image of every homomorphism Q → N is divisible
and then splits out of N. However, if N is not abelian the equality is not true in general.
We have found no example of this situation in the literature, so we describe one in the
sequel:
Example 4.8 Let N be the push-out of inclusions Q ←֓ Z ֒→ Z[1/p]. As N is torsionfree, every homomorphism Q → N is injective. Moreover, given a subgroup A < N
isomorphic to Q, SQ N is isomorphic to the subgroup generated by the conjugates of A by
elements of Z[1/p]. This implies that N/SQ N = Z[1/p]/Z, which is isomorphic to Z/p∞ .
As there exists an epimorphism Q → Z/p∞ , we conclude that N = TQ N, which strictly
contains SQ N.
It can be stated in an analogous way that SZ[1/p] N 6= TZ[1/p] N, as Q/Z has non-trivial
homomorphisms from Z[1/p]; moreover, similar arguments can be used to prove that for a
non-trivial set of primes J, the co-reflection SZ[J −1 ] is never equal to TZ[J −1 ] . In particular
we obtain:
Proposition 4.9 If A is an additive subgroup of Q, SA = TA if and only if A is trivial
or A = Z.
Proof. It is immediate from the previous discussion and Example 5.2 of [18] (see also
[14]).
✷
The moral here is that it is not easy to find examples for which SA = TA . However,
there are at least two ways of constructing such examples. The first family arises in the
context of varieties of groups. Consider a set of equations, and the verbal subgroup W
13
defined by these equations in the free group F∞ . It is proved in [8] that for every W
there exists a locally free group A such that the homomorphisms of W into any group G
identify the W -perfect radical, and in particular, this fact guarantees that SA = TA . The
inspiring (and first) “generator” of this kind was the acyclic group of Berrick-Casacuberta
[2], and a closely related group will be the crucial ingredient of Proposition 5.2 below.
The second family of examples can be constructed as a byproduct of Theorem 3.7,
taking account of the following two results:
Proposition 4.10 Suppose that for a group A the class C(A) is closed under extensions,
i.e. C(A) = C(A). Then Cq (A) = Ct (A).
Proof. In this case, for every group G, CA G = DA G, and then there is a diagram
DA G
//
SA G
j
Id
DA G
//
TA G
As j should be an epimorphism, and it is always a monomorphism, it is an isomorphism.
✷
Corollary 4.11 Let A be a group. Then for any cellular generator B of the class C(A)
we have SB = TB .
Recall that by Theorem 3.7, it is always possible to construct such B out of any given
G, and this provides many examples of groups for which SA = TA . It is natural and
interesting to ask if there is a converse of the previous proposition, so we will devote the
rest of the paper to the following problem:
Question 4.12 If TA = SA , for a certain group A, is it true that C(A) = C(A)?
We will see in Proposition 5.2 that the general answer is negative. Before, we will
discuss in which cases the conjecture can be true. The following result is easy to prove
and gives a sufficient condition:
Proposition 4.13 Let A be a group, and assume that C(A) = Cq (A) and TA = SA . Then
C(A) = C(A).
14
Proof. As the members of the classes are defined precisely as the groups for which the
corresponding co-reflection is the identity, TA = SA implies Cq (A) = Ct (A), and then by
hypothesis C(A) = Ct (A). But the inclusions C(A) ⊆ C(A) ⊆ Ct (A) always hold, so the
result follows.
✷
It can be seen that the condition of the proposition holds for the acyclic group of
Berrick-Casacuberta. However, the following example shows that the condition is not
necessary in general for cellular generators of closures under extensions:
Example 4.14 Given a prime p, consider a cellular generator B of the class C(Z/p). If
C(B) = Cq (B), this would imply that CB = SB , and equivalently DZ/p = TZ/p . But this is
not true in general, as can be deduced for instance from Example 4.7. Observe that we
obtain in particular that C(B) = C(B) 6= Cq (B) = Ct (B) in this case.
We finish this section with a technical description of the difference between DA and
CA when SA = TA . Remember from the previous section that for every group N we have
constructed DA N as the cellular cover of N with respect to a cellular generator B of C(A).
As CA N is B-cellular, there is a canonical homomorphism CA N → CB N, which induces
in turn a diagram of extensions:
KO A
KB
CBO N
/
/
CA N
SBO N
/
/
SA N
Here KA and KB are the corresponding kernels. Observe that SA N = TA N implies
that SB N = TA N, the right vertical homomorphism is the identity, and the difference
between CA N and CB N(= DA N) is exactly the difference between KA and KB .
We will need some homotopical tools in our description of the kernel of CA N → CB N.
In the sequel we will denote respectively by MA and MB 2-dimensional complexes with
π1 MA = A and π1 MB = B, respectively, and by Cof A and Cof B the corresponding
homotopy cofibres defined at the end of Section 2 . Without loss of generality, we can
assume that N = TA N = SA N, and denote X = K(N, 1). The long exact sequence of
the fibration X → PΣMA Cof A gives rise to an extension of groups:
π2 PΣMA Cof A → π1 (CWA X) → N.
15
On the other hand, Mayer-Vietoris sequence of the cofibration gives rise to another extension:
HA → π2 Cof A →
M
LA ,
where the group on the left is the quotient of H2 N under the image of the induced
L
homomorphism
H2 (A) → H2 (N) and the group on the right is a subgroup of the
L
direct sum
Aab . Similar sequences exist replacing A by B.
According to Section 7 of [5], the effect of PΣMA over the second homotopy group of
a simply-connected space is to divide by the Aab -radical. As TA = TB , we have KA =
π2 Cof A /TA (π2 Cof A ) and KB = π2 Cof B /TA (π2 Cof B ). Moreover, as the homomorphism
W
KA → KB is induced by the map Cof A → Cof B induced in turn by the inclusion MA →
W
MB , we obtain the following:
Proposition 4.15 In the previous notation, the kernel of the homomorphism CA N →
DA N is isomorphic to the kernel of the homomorphism
π2 Cof A /TA (π2 Cof A ) → π2 Cof B /TA (π2 Cof B ).
Let us remark a relationship between Cof A and Cof B . Consider
W
MA and
W
MB , with
′
the wedges respectively indexed by [MA , X]∗ and [MB , X]∗ . If B is the homotopy cofiber
W
W
of the map MA → MB induced by inclusion, there is also a cofibration sequence
ΣB ′ → Cof A → Cof B . As PΣMA B ′ is contractible, ([13], 1.D.3) implies that there is
a homotopy equivalence PΣMA ΣCof A ≃ PΣMA ΣCof B . As π3 ΣCof A = π2 Cof A and same
happens to Cof B , this seems an interesting way to investigate the relation between KA
and KB . However, when H2 A 6= 0, MA cannot be a Moore space, and these nullifications
are hard to compute. In the other case, Proposition 5.1 of [18] (based in turn in Section
5 of [5]) implies that certain localizations or completions of π2 Cof A and π2 Cof B are the
same. See also section 5.2 of [16].
In general, the relation between the group co-reflections CA and CB (= DA ) has been
investigated when H2 A = 0, and it is well understood when A is a cyclic group or a
subring of Q (see for example [18]). As seen above, the latter is not useful in our context,
as SA 6= TA for these groups. However, in the next section, we will take profit of some
features of the case H2 A = 0 to describe a counterexample for Question 4.12.
16
5
The Burnside radical and the Thompson group
In the following A will denote the generator constructed in ([8], Theorem 3.3) of the
Burnside idempotent radical for a prime p. This is a locally-free group, and then it is
torsion-free. In our computations we will need its lowest ordinary homology:
Lemma 5.1 The first homology group H1 (A) is a free module over Z[1/p], and H2 A = 0.
Proof. As the group is locally free, H2 A = 0. Moreover, the construction of A in the
mentioned paper (as a telescope of homomorphisms that send generators to products of
p-powers) implies that the first homology group is the direct limit of a system
M
M
Z → ...,
Z→
Z→
I2
I1
where:
• Every direct sum is taken over a finite set of indexes.
• If j < k, then Ij < Ik .
• For every k, the homomorphism
L
Ik
Z→
L
Ik+1
Z takes the |Ik | components of the
left group to the first |Ik | of the right group in their order, and the homomorphism
is multiplication by p.
Hence, the direct system that defines H1 A is a countable direct sum of copies of the
p
p
p
system Z → Z → Z → . . . and the result follows.
✷
Now consider the Thompson group F, defined by the presentation
F = hx0 , x1 , x2 , . . . | x−1
i xn xi = xn+1 for i < ni.
It is known that this is an infinite simple group such that H2 F is a free abelian group in
two generators, see [7].
Proposition 5.2 For the group A, one has SA = TA , but DA 6= CA .
Proof. As A generates the idempotent radical associated to a variety of groups, TA = SA .
We will check in this proof that DA F 6= CA F .
First we will compute explicitly DA F. As the radical is always a normal subgroup
and F is simple and not equal to its exponent p reduction, F = TA F . According to
17
Theorem 4.3, we need to compute the kernel of the homomorphism DA F → F , which is
in turn the fundamental group of the homotopy fibre of the map K(F, 1) → PM K(F, 1),
being M = M(A, 1) the corresponding 2-dimensional Moore space (which exists by the
local freeness of A). We can describe nullifications of nilpotent spaces, so we consider
W
the evaluation [M,K(F,1)] M → K(F, 1) extended over all the homotopy classes of maps
M → K(F, 1). Let C be the homotopy cofibre of this map. As the wedge is M-acyclic,
PM K(F, 1) = PM C. As F = TA F , C is simply-connected.
Now we need to compute π2 C, which by Mayer-Vietoris is defined by an extension:
M
H2 F ֒→ π2 C ։
Z[1/p].
Here we have used that H2 A = 0 and the perfectness of F . Now according to ([5],
Theorem 7.5), π2 PM C is the Ext-p-completion of π2 C. In the sequel we will call P the
group in the right hand-side of the extension.
Taking account of the fact that Hom(Z/p∞ , G) = 0 if G is free abelian and ([6],VI.2.5),
we obtain an extension:
M
Ext(Z/p∞ ,
Z[1/p]) ֒→ Ext(Z/p∞ , π2 C) ։ Ext(Z/p∞ , Z ⊕ Z).
As the Ext-p-completion of every p-divisible group is trivial, the term in the left is so,
and the homomorphism in the right is an isomorphism. Hence by ([6],VI.2.6) we have
π2 PM C = Z∧p ⊕ Z∧p , and DA F is defined by an extension:
Z∧p ⊕ Z∧p ֒→ DA F ։ F.
Now by Theorem 2.4, we must prove that the kernel of CA F → F cannot be isomorphic to
two copies of the p-adic integers. We know that such kernel is isomorphic to π2 C/TA π2 C.
As π2 C is abelian, TA π2 C = TZ[1/p] π2 C, and as the homomorphism f : π2 C → P is
an epimorphism, same is true for π2 C/TZ[1/p] π2 C → P/f (TZ[1/p] π2 C). But as Z[1/p] is
p-divisible, f (TZ[1/p] π2 C) is so, and then it is a free Z[1/p]-submodule of P , as Z[1/p] is
a principal ideal domain. Then it is easy to check that no non-trivial quotient of a free
Z[1/p]-module by a Z[1/p]-free submodule can be divisible.
Assume now by reductio ab absurdum that π2 C/TZ[1/p] π2 C is isomorphic to Z∧p ⊕ Z∧p .
Then P/f (TZ[1/p] π2 C) should be non-trivial, otherwise Z∧p ⊕ Z∧p would be isomorphic to
a free abelian group. Moreover, as Z∧p ⊕ Z∧p is q-divisible for every q 6= p, its quotient
P/f (TZ[1/p] π2 C) should be too. But P/f (TZ[1/p] π2 C) is p-divisible but not divisible, so we
have a contradiction, and then DA F 6= CA F .
18
✷
Remark 5.3 We do not know if π2 C/TZ[1/p] π2 C is a free abelian group in two generators,
because we are unable to show if the extension that defines π2 C splits or not. Recall that
Ext(Z[1/p], Z) = Z∧p /Z is uncountable, but only the trivial extension of Z by Z[1/p] splits.
Acknowledgments. We warmly thank Jérôme Scherer for several discussions about
the paper and for his thorough revision of the manuscript, which has led to a substantial
improvement of the presentation. We also thank Fernando Muro for useful discussions.
The first author expresses his gratitude to the Department of Mathematics of the University of Almerı́a for their kind hospitality.
References
[1] J. Adámek, J. Rosický, Locally presentable and accessible categories, Vol. 189, Cambridge Univ. Press, 1994.
[2] A. Berrick and C. Casacuberta, A universal space for plus-constructions, Topology
38 (1999), 467–477.
[3] A. K. Bousfield, Homotopy factorization systems, J. Pure Appl. Algebra 9 (1977),
207–220.
[4] A. K. Bousfield, Localization and periodicity in unstable homotopy theory, J. Amer.
Math. Soc. 7 (1994), 831–873.
[5] A. K. Bousfield, Homotopical localizations of spaces, Amer. J. Math. 119 (1997),
1321–1354.
[6] A. K. Bousfield, D. M. Kan Localization and completion in homotopy theory, Bull.
Amer. Math. Soc. 77 (1971), no. 6, 1006–1010
[7] K. S. Brown, The homology of Richard Thompson’s group F, Topological and asymptotic aspects of group theory, 47—59, Contemp. Math., 394, Amer. Math. Soc.,
Providence, RI, 2006.
[8] C. Casacuberta, J. L. Rodrı́guez, and D. Scevenels, Singly generated radicals associated with varieties of groups, Proceedings of Groups St. Andrews 1997 in Bath I,
19
London Math. Soc., Lecture Note Series 260, Cambridge University Press (1999),
202–210.
[9] W. Chachólski, On the functors CWA and PA , Duke Math. J. 84, (1996), 599–631.
[10] W. Chachólski, Covers of groups, An alpine expedition through algebraic topology,
109–131, Contemp. Math., 617, Amer. Math. Soc., Providence, RI, 2014.
[11] W. Chachólski, P. E. Parent, and D. Stanley, Cellular generators, Proc. Amer. Math.
Soc. 132, 11 (2004), 3397–3410.
[12] W. Chachólski, E. Damian, E. Farjoun, and Y. Segev, The A-core and A-cover of a
group, J. Algebra, 321, Issue 2, (2009), 631–666.
[13] E. Farjoun, Cellular spaces, null spaces and homotopy localization, Lecture Notes in
Math. 1622 Springer-Verlag, Berlin Heidelberg New York, 1996.
[14] T. Fay, E. Oxford, and G. Walls, Singly generated socles and radicals, in Abelian
group theory (Honolulu, Hawaii, 1983), Lecture Notes in Math. 1006, Springer,
Berlin-New York (1983), 671–684.
[15] R. Göbel and J. Trlifaj, Approximations and endomorphism algebras of modules,
GEM 41, W. de Gruyter, Berlin, 2012.
[16] J.L. Rodrı́guez, On homotopy colimits of spaces with a single homology or homotopy
group, Ph. D. thesis, Universitat Autònoma de Barcelona, 1997.
[17] J. L. Rodrı́guez and D. Scevenels, Universal epimorphic equivalences for group localizations, J. Pure Appl. Algebra, 148 (3) (2000), 309–316.
[18] J. L. Rodrı́guez and J. Scherer, Cellular approximations using Moore spaces. Cohomological methods in homotopy theory (Bellaterra, 1998), 357–374, Progr. Math.,
196, Birkhäuser, Basel, 2001.
[19] J. L. Rodrı́guez and J. Scherer, A connection between cellularization for groups and
spaces via two-complexes, J. Pure Appl. Algebra, 212 (2008), 1664–1673.
[20] K. Varadarajan. Groups for which Moore spaces M(π, 1) exist, Ann. of Math. 84
(1966), 368–371.
20
Ramón Flores
Departamento de Geometrı́a y Topologı́a, Universidad de Sevilla-IMUS
E-41012 Sevilla, Spain, e-mail: [email protected]
José L. Rodrı́guez
Departamento de Matemáticas, Universidad de Almerı́a
E-04120 Almerı́a, Spain, e-mail: [email protected]
21
| 4math.GR
|
arXiv:1703.04741v3 [cs.AI] 31 Oct 2017
Towards Moral Autonomous Systems
Vicky Charisi 1 , Louise Dennis 2 , Michael Fisher2 , Robert Lieck3 ,
Andreas Matthias4 , Marija Slavkovik5 , Janina Loh (Sombetzki)6 ,
Alan F. T. Winfield7 , and Roman Yampolskiy8
1
University of Twente
University of Liverpool
3
Universität Stuttgart
4
Lingnan University
5
[email protected], University of Bergen
6
University of Vienna
7
University of the West of England, Bristol
8
University of Louisville
2
November 1, 2017
Abstract
Both the ethics of autonomous systems and the problems of their
technical implementation have by now been studied in some detail.
Less attention has been given to the areas in which these two separate
concerns meet. This paper, written by both philosophers and engineers of autonomous systems, addresses a number of issues in machine
ethics that are located at precisely the intersection between ethics and
engineering. We first discuss the main challenges which, in our view,
machine ethics posses to moral philosophy. We them consider different approaches towards the conceptual design of autonomous systems
and their implications on the ethics implementation in such systems.
Then we examine problematic areas regarding the specification and
verification of ethical behavior in autonomous systems, particularly
with a view towards the requirements of future legislation. We discuss
transparency and accountability issues that will be crucial for any future wide deployment of autonomous systems in society. Finally we
consider the, often overlooked, possibility of intentional misuse of AI
systems and the possible dangers arising out of deliberately unethical
design, implementation, and use of autonomous robots.
1
Keywords: Robot ethics, Machine ethics, Artificial morality, Autonomous
systems, Verification, Transparency, Unethical AI
This article is the result of a series of discussions taken within the scope of
the Dagstuhl Seminar 6222 Fisher et al. [2016]. Dennis, Fisher and Winfield
wish to thank EPSRC for their support, particularly via the “Verifiable
Autonomy” research project (EP/L024845 and EP/L024861) Corresponding
Author: Marija Slavkovik, University of Bergen, P.O.Box 7802, 5020 Bergen,
Norway, [email protected]
1
Introduction
The so called “trolley problem” is a thought experiment introduced in Foot
[1967], whose ethical conundrum continues to fascinate today Wallach and
Allen [2008]. Roughly, it can be summarized as follows: there is a runaway
trolley on a railroad track and ahead on the track there are five people tied
up, unable to escape being killed by the trolley. The track splits in two by
a lever you control. The lever can divert the trolley on to a second track
where there is one person tied up and unable to move. Is it more ethical to
divert the train, or let it run its course?
The emergence of driver-less cars in regular traffic has brought the trolley
problem to public attention. Articles such as “Should Your Car Kill You
To Save Others?”1 are flooding popular science media. It is easy, given the
same problem context of traffic, to get sidetracked into confusing solving
the trolley problem with controlling the impact driver-less cars will have
on traffic and our society as a whole. This, of course, is not the case.
Enabling machines to exhibit ethical behavior is a very complex and very
real time-sensitive issue. The driver-less cars are only the forefront of a
whole generation of intelligent systems that can operate autonomously and
will operate as part of our society. Which ethical theory to employ for
deciding whose death to avoid in a difficult traffic situation is a difficult
problem. This, however is not necessarily the most important problem we
would need to solve. The goal of this position paper is to discuss what
does engineering machine ethics entail. Once we as a society have discerned
what is the right thing for an artificial autonomous system to do, how can
we make sure the machine does it?
The problem of identifying, discerning, and recommending concepts of
right and wrong is the domain of moral philosophy. Moral philosophy, together with the law, act as a system of recommendations regarding which
1
http://www.popularmechanics.com/cars/a21492/the-self-driving-dilemma/
2
possible actions are to be considered right or wrong (ignoring, for the moment, ethics systems that do not specifically address the morality of individual actions, e.g. character-based ethics, and which are less useful for the
problem at hand).
Driver-less vehicles are the most visible of a whole range of technologies. This range includes also the assisted living technologies, as well as
the various embedded decision aid software and solutions. We have been
using intelligent systems with varying degree of autonomy for the past fifty
years: industrial robots, intelligent programming of household appliances,
automated trains, etc. However, what all of these machines have in common is that they either operate in a segregated space, a so called working
envelope, or they have no capabilities to cause damage to their environment.
Driver-less vehicles are obviously going to “break” both these restrictions.
How can we build intelligent autonomous systems that uphold the ethical
values of the society in which they are embedded? This is the main concern
of machine ethics, a new interdisciplinary area of research within Artificial Intelligence (AI) Moor [2006]; Allen et al. [2006]; Wallach et al. [2008];
Wallach and Allen [2008]; Anderson et al. [2016].
It is clear that choosing the best moral theory to implement in a particular intelligent autonomous system is not a simple question. It is primarily
a question for moral philosophy and opens new challenges in this field. In
Section 2 we give a brief overview of these challenges.
For a long time, people and societies have been the only intelligent
decision-makers. Moral philosophy has been developed with the often implicit assumption that the moral agent is a human. It is not at all clear to
what extent existing moral theories extend to non-human decision-makers.
Even if it is shown to be easy to replace a human agent with an artificial agent in a moral theory, and some societal decision is made concerning
which ethical behavior in machines is desirable or sufficient, we are still faced
with a set of problems regarding the implementation of ethical reasoning.
These are the problems we analyze here: implementation, verification, trust,
confidence and transparency and the prevention of intentionally unethical
systems.
Moral theories are inherently ambiguous in recommendations of moral
behavior, thus reflecting the context dependency of what constitutes a moral
choice. We already know that artificial systems, when compared to people,
are not good at handling ambiguity. Enabling machines to deal with contextambiguity in decision-making is a core Artificial Intelligence problem Russell
and Norvig [2015]. In Section 3 we give an overview of the most intuitive
approaches to implementing ethical behavior in autonomous systems and
3
discuss the advantages and shortcomings of these approaches.
Human societies have a multitude of means for ensuring its members
behave within the socially accepted boundaries of morality. We can say
that a person behaves ethically because they have an individual and personal
motivation to do so, without going into how this motivation is formed. The
motivation for an artificial agent to behave ethically originates not personally
from the agent, but from other actors. These actors can broadly be organized
intro three groups: the designers of the artificial agent, its users, and the
various societal regulators whose job it is to make sure that order in society
is maintained. It is all of these actors that need to be reassured that their
own particular motivations for the AI system to behave ethically are met.
Hence, a big concern when engineering machine ethics is not only that ethical
behavior is accomplished, but also that the ethical behavior can be verified.
This is the issue we discuss in Section 4.
In Section 5 we focus on issues of transparency and accountability for
machine ethics implementations. Since there are several actors outside of
the ethical agent who are supplying the motivation for ethical behavior, the
implementation of this behavior must be transparent to those actors to the
extent and in a manner sufficient for their needs. Transparency is a key
element in enabling society to have the right amount of trust and confidence
in the operations of an AI system.
Lastly in Section 6 we discuss the possible dangers for society that lie
in engineering machine ethics. Like all technology, AI systems can also be
abused to further criminal activities. AI systems can be deliberately built to
behave unethically and illegally, but they also can be induced, deliberately
or by accident, into exhibiting socially undesirable behavior.
The main contribution of this position paper is an integral overview of
the immediate challenges and open questions faced when pursuing the problem of engineering of machine ethics. This paper is a result of a week long
discussion among experts of different fields within the scope of the Dagstuhl
Seminar 162222 , and incorporates various ideas that arose as a result of
discussions among interdisciplinary experts. Position papers that focus on
machine ethics as a whole have been produced and they offer interesting insights in the problem as a whole, see for example Moor [2006]; Anderson and
Leigh Anderson [2007]; Bonnefon et al. [2016]; Anderson et al. [2016], but
to the best of our knowledge, this is the only document devoted specifically
to the problem of engineering machine ethics.
2
The report on this seminar is available Fisher et al. [2016].
4
2
Challenges for Machine Moral Philosophy
Normative ethics is the branch of moral philosophy concerned with developing methods for deciding how one aught to act. The purpose of devising a
moral theory, within normative ethics, is to develop a consistent and coherent system that can be followed to unambiguously identify an action as good
or bad. Numerous specific theories have been developed, the most notable
of which are perhaps utilitarianism Harsanyi [1977], Kantianism Ellington
[1993], and Ross’s ethical theory Ross [1930].
All of the moral theories so far developed in philosophy have been built
around several underlying assumptions. First and foremost is the assumption that the reasoning and decision making when following the moral theory
is done by an agent that is a human. The immediate specific question we
can pose to moral philosophy is: given a specific normative moral theory,
how is the theory affected if an artificial agent replaces a human agent in it?
A more subtle assumption in normative ethics is the assumption that the
agent has de facto the ability of being a moral agent. This means that the
theory is developed for a being that is capable of understanding concepts
of right and wrong. The debate of whether an artificially created entity
can be a moral agent is still far from settled Etzioni and Etzioni [2017].
Wallach and Allen [Wallach and Allen, 2008, Chapter 2] hint at the idea
that how ethically sensitive an artificial system can be depends on how able
of autonomous action it is. At this point in the development of the field of
machine ethics it is fair to summarise that the capacity for moral agency of
an artificial system is more of a sliding scale rather than a Boolean value.
There are at least two questions to moral philosophy we can pose. The first
is: how to define a scale of moral agency to describe the moral abilities
of current and future artificial agents. The second question again is about
going back to specific normative moral theories and checking how the theory
is affected by replacing a full moral agent with an agent that is “lower” on the
newly devised moral agency scale. Perhaps the concept of moral agency is
altogether inadequate when discussing autonomous intelligent systems and
a new concept needs to be devised.
There is a further weaker assumption in normative ethics, which is the
assumption that the agents of the theory are able to accept and act upon
considered judgements Elgin [1996]. Consider judgments are “common sense
agreements” about what is good or bad in particular instances or cases. For
example, the idea that murder is bad except in exceptional circumstances,
is a considered judgment. This, let us call it considered judgment ability
assumption, is reflected in the numerous moral dilemmas encountered in
5
the normative ethics literature, the trolly problem being the most popular
example.
A dilemma in normative ethics, understood in a very broad sense, is
a problem of choosing between two options each of which violating one or
more considered judgment. An artificial agent does not possess a common
sense understanding unless one is programmed into him. The engineering
of common sense, or rather background knowledge, has been a notoriously
elusive problem in artificial intelligence. The dilemmas we encounter in
machine ethics reflect this. Consider for example the so called Cake or
Death problem introduced in Armstrong [2015] which describes a situation
in which an agent is unsure if if killing people is ethical or baking them a
nice cake is ethical.
Solving Cake or Death is not an ethical problem in moral philosophy,
it is trivial. However, the question of which are the essential considered
judgments that necessarily have to be implemented in machines is not a
trivial problem. This would depend on the nature and abilities of a specific
class of artificial agents. In contrast to focussing on a general theory of
discerning right from wrong, there is a need for normative ethics to identify
and develop a minimal such theory.
Building moral machines by implementing human morality is a natural
approach. After all, it is human society that those machine are entering
and it is human sensibilities and values that they have to uphold. An alternative, or perhaps parallel approach would be to build a normative ethics
theory exclusively for artificial agent. An example of such a theory comes
to us from science fiction - the three laws of robotics of Prof. Isaak Asimov
Asimov [1950]. The shortcomings of Asimov’s laws of robotics have been
extensively argued by the author himself, but they have also been given
a serious philosophical consideration Leigh Anderson [2008] and attempts
have been made for their implementation Dennis et al. [2015]; Vanderelst
and Winfield [2017]; Caycedo Alvarez et al. [2017].
The development of machine moral theory is an interesting open area
for study in normative ethics. The question that has to be addressed first
perhaps is what constitutes a viable, desirable, or good theory for artificial
agents? We are perhaps primed by the cultural influence of Asimov’s laws
of robotics to ask: what is it that a robot should never do? But thinking
in absolutes is not likely to be viable for machine moral theories any more
than it is viable for human normative ethics. Regardless of what comes our
of normative ethics in the future, any moral theory developed for artificial
agents must be developed to the point of being implementable. Prescriptions of good behaviour suffice for people, for machines we need algorithms.
6
An algorithm necessarily includes a specification of all possible scenarios
and context in which an ethical decision can be made. Therefore it is necessary that machine normative ethicists, computer scientists and engineers
collaborate closely towards developing machine moral philosophy, with this
collaboration perhaps being one of the challenges as well.
3
Different Approaches, their Advantages and Challenges
An intelligent system is one that is capable of communicating with, and
reasoning about, its environment and other systems. An autonomous system
is one that is capable of, to a certain extent, unsupervised operation and
decision-making. Wallach, Allen, and Smit Wallach et al. [2008]; Wallach
and Allen [2008] argue that very intuitively, ethical behavior in machines
can be accomplished in at least two different ways. The first approach is to
identify a set of ethical rules, perhaps by choosing a normative ethic theory,
around which a decision-making algorithm can be implemented. The second
approach is to have a machine evolve or “learn” to discern right from wrong
without it having be explicitly guided by any one ethic theory. They refer to
these two approaches as the top-down and bottom-up approach respectively.
A hybrid approach in their sense is one in which an agent starts with a set
of rules or values and modifies them into a system for discerning right from
wrong.
Artificial Intelligence (AI) has grown into a large field that incorporates
many approaches, which can be, very tentatively, classified into soft computing approaches, which include statistical methods, machine learning and
probabilistic reasoning, and traditional symbolic AI methods, which includes
logic-based reasoning Russell and Norvig [2015]. The question of how to
implement machine ethics in an intelligent autonomous system necessarily
hinges on the AI methods that system uses. Different AI approaches would
be subject to different machine ethics implementations and we need to consider their malleability to machine ethic approaches, as well as their risks
and advantages in this respect.
We here roughly classify the current and future machine ethics implementations based on the main AI approach used into soft machine ethics
and symbolic machine ethics mirroring the two largest traditional branches
of AI methodology. We discuss both of these approaches and their advantages and challenges. We should note at this point that a hybrid approach
here would be one that combines symbolic methods, such as for example rule
7
based reasoning, with soft methods such as for example supervised learning.
The bottom-up approach of Wallach et al. [2008]; Wallach and Allen
[2008] naturally lends itself to be approached by using soft computing AI
methods, whereas their top-down approach is perhaps best “served” by symbolic AI methods.
3.1
Soft and Symbolic AI Methods for machine ethics
Within engineering, a top-down approach towards solving a task consists in
breaking down the task iteratively into smaller sub-tasks until one obtains
tasks that can be directly implemented. Problems best solvable by a topdown approach are ones in which the problem, and its context, are fully
understood and can be formally specified. This is normally the case when
the problem occurs in a controlled environment. These problem properties
are also ones required for a successful solution by implementation using
symbolic AI methods such as rule based reasoning.
There are numerous ways in which a symbolic AI approach can be taken
to develop ethical behaviour in a system. The most frequent in the literature
is to constrain the choices of the system using rules derived from an ethical
theory. This is the approach taken in Arkin et al. [2012], for developing
the concept of ethical governor, and also in Dennis et al. [2016b] where the
ethical theory used is a version of Ross’s ethical theory Ross [1930]. The
Dennis et al. [2016b] work considers a hybrid autonomous system three-layer
architecture: a continuous system controlled by a rational software agent,
which makes discrete decisions, via a continuous control layer that allows
for a dynamic environment interaction and feedback. The rational software
agent is provided with an ethical policy, a total order over abstract ethical
principles such as “do no harm”, “do not damage property” etc. The agent
relies on external entities to identify if, and which, of her possible actions
impinges on some of these abstract ethical principles. Having her actions
annotated, the agent chooses between possible actions by selecting the one
that is minimally unethical with respect to the given ethical policy. In
contrast, Bendel [2016] proposes a method for building annotated decision
trees for making simple ethical choices.
A bottom-up approach to problem solving in engineering starts with describing instances of desired solutions by using adequate parameters and
the proceeding to build up a procedure for identifying solutions based on
these parameters. Machine learning methods in AI take a bottom-up approach to problem solving. There are several examples of using machine
learning to implement machine ethics, such as for example Anderson and
8
Anderson [2014] andAbel et al. [2016]. In Anderson and Anderson [2014]
inductive logic programming is used over a corpus of particular cases of
ethical dilemma to discover ethical preference principles. Each case relates
two actions, one more ethical than the other. The preference between the
actions depends on ethically relevant features that actions involve such as
harm, benefit, respect for autonomy, etc. Each feature is represented as an
integer that specifies the degree of its presence (positive value) or absence
(negative value) in a given action. The system is able to extract an ethical
rule from the cases it is presented with and thus to a certain extent is able to
learn to discern right from wrong. In Abel et al. [2016] reinforcement learning is used to learn what the most moral of two actions is, by rewarding the
“correct” decisions an agent makes and “punishing” the bad “wrong” ones.
Whether a soft or symbolic AI method is used depends on the nature
of the problem that needs to be solved. The two families of methods tackle
problem from different sides and are not mutually exclusive. Each of the
methods comes with its own advantages and challenges with respect to building ethical behaviour in an intelligent autonomous system.
3.2
Advantages and challenges of using Symbolic AI Methods
Symbolic AI methods are best suited for ethical reasoning in limited domains
when the context of the decision-making problems can be predicted. To use
symbolic based reasoning, an ethical theory needs to chosen before constructing and deploying the system and this theory does not change throughout
the system’s usage. This allows for a thorough and well-informed process of
decision-making and the verification of the system prior its practical application. Different ethical principles and theories may explicitly be encoded
into the system giving clear options to decide upon. Any parameters that
are left open for definition by the customer or to be learned from interaction
with the environment have a clear function and it is possible to verify that
they do not violate more general ethical principles.
Symbolic AI methods also come with their set of challenges and limitations. General ethical guidelines are typically formulated on a very abstract
level. Much philosophical discourse on ethics is concerned with problems
occurring when applying such general guidelines to concrete situations. It
is thus not clear if, and how, general ethical guidelines can be leveraged to
solve concrete problems of decision making in practice. For a real-world system the connection to the non-discrete sensory-motor level must be made.
There are many ways to transform continuous sensor values into discrete
9
propositions that can be used in reasoning. General guidelines or even single terms and concepts are only (if at all) implementable in a reduced way,
i.e. restricted to one preferably clear interpretation. Due to their context
sensitive definition it is not possible to consider every possible interpretation
of an abstract guideline or term in implementing them in an artificial system
Matthias [2011].
Furthermore, symbolic AI approaches risk conflicts between the implemented ethical theories and principles. If only one theory is implemented,
e.g. Kant’s Categorical Imperative Ellington [1993] or Isaac Asimov’s first
Law of Robotics3 , then this theory would determine the specific maxims that
are to be defined situationally by the artificial system. Winfield Winfield et
al. [2014] describes experimental trials of a minimally ethical robot which
implements Asimov’s three laws of robotics. The chosen theory must be
such as to allow implementable rules to be derived from it. Such a monistic
approach assumes that there exist no moral dilemmas, i.e. that the implemented theory is able to give a conflict-free rule to make a decision in every
context (Winfield et al. [2014] experimentally shows how a single ethical rule
performs when faced with a balanced ethical dilemma).
Deciding on a specific set of ethical principles involves settling longstanding philosophical disputes in an ad-hoc way. It is possible that governmental bodies might take the lead in outlining high-level ethical guidance
to designers. For example, Germany’s ministry of transport recently announced the intention to set out a basic ethical policy to be followed by car
designers stipulating that property damage takes always precedence over
personal injury, that there must be no classification of people, for example,
on size, age and the like, and that – ultimately – it is the manufacturer who
is liable4 .
Symbolic AI approaches require that any kind of “common sense” is
explicitly coded using the formal language of the method. This severely
impacts the scalability of the machine ethics solution. As it is well understood in the AI sub-discipline of knowledge representation and reasoning,
the more expressive the formal language for encoding knowledge is, the more
computationally expensive reasoning becomes. It is not a problem of having
or symbolically encoding a large amount of information, but a problem of
computing logical entailment or consistency, which are in the core of deep
reasoning and are known to be of non-deterministic computational com3
The laws can be found quoted in Wikipedia, at http://en.wikipedia.org/wiki/
Three_Laws_of_Robotics
4
http://www.wiwo.de/politik/europa/selbstfahrende-autos-dobrindtgruendet-ethikkommission-fuer-automatisiertes-fahren/14513384.html
10
plexity. Therefore, it is unsurprising that the most recent major AI breakthroughs have been achieved using statistical processing of information and
shallow reasoning. That being said, some symbolic-methods, such as model
checking, are scalable within reasonable parameters, and have been vastly
deployed in the information processing industry.
3.3
Advantages and challenges of using Soft AI Methods
Soft AI methods are best applicable when we know the kind of data the AI
system receives from interacting with the environment and, while the overall
objective might not be well known or specified, we still have an idea how
to process these data in a useful manner. For instance, the vision pipeline
of a household robot is designed to extract obstacles (walls, tables, etc.),
objects of interest (books on a shelf etc.), and its own position from the
sensory data because this information is useful for a wide range of tasks
it will be required to perform. Most real-world AI systems will be partly
designed using a soft AI method at least on the lower sensory-motor level.
Because soft AI methods are based on synthesis of actions and choices, with
respect to the task of building an ethical AI the major question here is: how
do components designed in a bottom-up fashion affect the overall ethical
properties of the system?
Soft AI methods do not require predetermines moral principles , ethical
theories or sets of rules, but instead formulate basal parameters and intend
to implement competences whereby an artificial system acts autonomously.
This can be done, for example, via trial and error or other modes of learning
such as imitation, induction and deduction, exploration, learning through
reward, association and conditioning Cangelosi and Schlesinger [2014]. Soft
AI methods can be separated into models of evolution Froese and Di Paolo
[2010] and models of human socialization Fong et al. [2003]; Breazeal and
Scassellati [2002]. The former simulate evolutionary moral learning, by assessing slightly different programs in an artificial system to evaluate an ethical case. Those programs that can solve the ethical task sufficiently go
through to a “next round” where they are (re)combined to solve further
ethical tasks. Evolutionary approaches can be used in earlier stages of moral
development before considering models of human socialization.
Models of human socialization consider the role of empathy and emotion
for moral learning. They assume that a robot learns morality via empathy Slote [2007]. What is controversial in the philosophical discourse is
that there exist two types of empathy Stüeber [2006]: perceptual empathy,
when an emotion triggers an equivalent or congruent reaction in the ob11
server Misselhorn [2009], and imaginative empathy that requires a change
in perspective in the form of empathising with the other, putting oneself
in the observed other’s shoes. Perceptual empathy is explicable with the
help of specific theories of mind or neuronal resonance and mirror neurons
and has been implemented in a rudimentary fashion in artificial systems
Balconi and Bortolotti [2012]; Rizzolatti and Fabbri-Destro [2008]; Mataric
[2000]. Ekman Ekman [1992] implements perceptual empathy in the form
of a basal affect program as an autonomous reaction scheme as a route to
the implementation of morality in robots. Young children and chimpanzees
are equipped with this fundamental form of empathy which forms the basis for pre-social behavior Warneken and Tomasello [2009]; Hoffman [2001].
Imaginative empathy is much more complex and develops on the basis of
perceptual empathy only. It is exhibited only in human socialisation, not
in non-human primates. This form of empathy is cognitively more ambitious and is involved in more complex moral reasoning and acting Gallagher
[2012]. We are not aware of any attempt to implement imaginative empathy
in artificial systems.
Since, by means of a soft AI solution the AI system becomes a moral
agent (if only in the narrowest sense of the word) one might ask whether it is
likely to be more adaptable to making ethical choices in situations that are
not pre-determined (which is a strong limitation to using the symbolic AI
methods). Since the system learns its own ethical rules, it circumvents, to an
extent (one could argue), the need to choose one particular ethical theory
to implement. But this seems at least questionable. Every self-learning
system must still be configured to pay attention to particular features of the
data set, and to ignore others. Looking at the consequences of an action,
instead of the agent’s motivation (for example) is such a choice of features
that essentially determines the choice of moral theory. It seems difficult to
judge at this point whether we can hope to create ethical-theory-agnostic
AI systems, since every choice of relevant data features is already, to some
extent, a choice of moral theory.
A major challenge with using soft AI methods is that it is hard to certify
whether the system fulfils any requirements one might want to impose. Indeed this is a challenge for all machine learning systems. A machine learning
solution virtually behaves as a black box - the approach solves a problem
successfully most of the time, but it is unclear whether a solution can be
expected for sure, or why a particular solution was learned or developed for
a particular problem. Soft AI methods, and machine learning in particular,
have had a dramatic success recently, with machine learning methods being used in a variety of problems and contexts. This success has prompted
12
for calls to ensure that some level of explainability for the choices of the
system is required, which in turn have given rise for the Explainable AI
(XAI) DARPA programme 5 . In Anderson and Leigh Anderson [2015] we
find one of the earliest specific implementations of XAI in machine ethics.
Their system extracts a moral rule from a collection of cases and is able to
explain why a particular decision is identified as more ethical referring to
the learning data.
Nonetheless, the black box nature of soft AI methods is likely to mean
that these solutions are unsuitable for implementation in critical systems.
This fundamental problem occurs irrespective of whether the ethical system
itself or only low-level sub-systems are built using a soft AI solution.
3.4
Modular and Hybrid approaches
Both the soft and symbolic AI methods come with advantages and challenges, but they also can complement each-other. A system is an entity
comprised of several entities, thus in principle an AI system can be built using components that exploit both solution approaches. We are unaware of
any implemented hybrid ethical reasoning system6 , but we can very briefly
discuss some recommendations for how such a system can be created.
One approach would be to separate decision-making by the ethical principles it involves. For example, decisions involving the possibility of human
death should be made using a pre-programmed ethical policy, while decisions involving violation of autonomy can be based on ethical preferences
learned through interaction with the system’s owner. Another approach
would be to separate decision-making in different contexts, with soft AI
methods being allowed as the default ethical decision-making method, while
symbolic AI approaches being required to be implemented for certain specific pre-determined contexts. Alternatively a system can be designed so
it first learns to recognize the ethical implications of its actions and then
those implications can be used to follow an implemented ethical theory when
choosing some specific course of action.
Implementing ethical reasoning within a system is not sufficient, we must
execute such implementation in a way that allows for verification of the
quality of the resulting ethical behavior. The designers and manufacturers
5
https://www.darpa.mil/program/explainable-artificial-intelligence
Although the title of van Rysewyk and Pontier [2015] claims a hybrid implemented
system upon closer inspection it is not clear in which sense is the solution not a pure
bottom-up approach in the sense of Wallach and Allen [2008], while the authors themselves
do not offer an analysis of this type in their paper.
6
13
of AI systems necessarily have to offer reasons for their users to trust the
artificial ethical system, and they also need to foresee possible malfunctions
and provide means to deal with them.
4
Specification and Verification of Ethical Behavior
Within our society, entities that are in a position to do us harm, be it
a complex machine production tool, the surgeon operating on our unconscious body, the other drivers on the highway, or a chainsaw, are subject
to licensing and certification. Certification informs consumers and experts
of the properties of a product, a system, or a person in a position of responsibility. Knowing that a standard has been met allows individuals to
have confidence in using machinery and to trust the decisions and actions of
professionals. Tools and systems are certified to operate within designated
parameters, while under (well defined) proper care. Certification confirms
that the manufacturer has taken all steps necessary to avoid or minimize
foreseeable risks that arise in relation to the usage of the tool. Certification
for persons in position of responsibility is more complex because it involves
a (possibly continuous) examination to demonstrate that the certified person has the understanding and skills necessary to perform his/her duties.
Typically, this involves regulations prescribing expected behavior — often,
humans must pass an examination concerning these regulations. Once we
move to an autonomous system, with no human directly in control, what are
our means to ensure that a systems actually matches the relevant criteria?
In order to be confident in a system’s behavior we need to specify what
we can expect the system to do in a particular circumstance, verify that
the system does actually achieve this, and validate that our requirements
are actually what the end-users want. There exist a vast range of different techniques, for example developed over many years within the field of
Software Engineering Sommerville [2001]. These techniques range from the
formal, such as proof, through structured, such as testing, to informal, such
as user validation. All these approaches can, in principle, be applied across
the range of autonomous systems, including robotics Fisher et al. [2013].
4.1
Who is the confirmation of ethical behavior for?
What constitutes an appropriate specification and verification methodology for ethical behavior depends on who is to use the results. In the case
14
of intelligent autonomous systems at least three interested parties can be
discerned: the designers including developers and engineers working on developing and maintaining the systems, the end-users, owners or customers,
and lastly various government and trade regulatory bodies and insurance
agents. Although these three categories are the evident interested parties,
this issue of interest discernment is an open problem in its own right, and as
some preliminary investigations show7 finer discernment may be required.
Although those actually constructing the AI system may have an intimate knowledge of its internal workings, it is still important that developers
and engineers not only have confidence in their prototypes but have techniques for highlighting where issues still remain. The technology itself should
not be a black box, but should be open to maintenance and analysis, and
must be flexible enough to be improved dynamically.
For end-users, customers and owners, the primary concern is that the
AI system they interact with is safe and behaves ethically with respect to
the ethical norms they themselves follow, as long as these are within the
scope of what is considered ethical and legal within their society. Trust is
a key issue and, in order to have trust extended to AI systems, the user
needs to be informed of its range of capabilities. The future of AI systems
and their proper integration within our society is subject, paradoxically, to
undue levels of both optimism and pessimism in terms of the extent to which
people can trust such systems. Close attention must be paid to nurturing
the appropriate level of trust.
AI systems are an exciting technological development that have long
been anticipated as part of the future in various works of fiction and there is
the temptation to play-up their apparent capabilities, particularly by early
marketing when the producers are still seeking financiers for their products.
This could lead to the customers placing an unwarranted level of trust in
some technology, even when adequate disclaimers and use guidelines are
outlined by the manufacturer, which can in turn lead to disastrous consequences8 . Such misplacement of trust is dangerous for users in the present,
and may cause society to over-react in order to limit integration of technologies which given proper time to adequately develop would have been
advantageous to the same society.
The appearance of trustworthiness is similarly an issue when people interact with an AI system. For example, a robot might appear “experienced,”
7
http://robohub.org/should-a-carebot-bring-an-alcoholic-a-drink-pollsays-it-depends-on-who-owns-the-robot/
8
https://www.theguardian.com/technology/2016/jun/30/tesla-autopilotdeath-self-driving-car-elon-musk
15
“benevolent,” or “sympathetic”. Such appearances are of particular concern
for AI systems that are integrated in assisted living technologies. Concerns
have been raised with respect to the impact assisted living technologies can
have on the elderly Sharkey and Sharkey [2012]. Similarly, it has been
shown that children who interact with robots derive expectations of them
and ascribe abilities to them. We need to develop an understanding of the
potential long-term effects of robots on child development Matthias [2015].
Trust should play a considerable role in choosing an ethical theory to
implement in AI systems. The ethical theory that is easiest to implement
may not necessarily be the one that is most trusted by society. This was
demonstrated in the case of utilitarianism and driver-less cars Bonnefon et
al. [2016].
It is important to note that trust Arkin et al. [2012] is not equal to ethics.
Trust is a social construct intimately concerned with how each individual
views the behavior of a robot or system. There may well be some varieties of
objective trustworthiness, but there will remain many varieties of subjective
trustworthiness. Many items affect users’ level of trust Salem et al. [2015],
for example, the relationship between trust and harm. If you could show
that robot causes no harm, would you trust it more?
Those who must regulate AI systems and their integration within society
also need confidence in the system. In addition, the Insurance industry
needs to be clear where responsibility Sombetzki [2015] lies and so where
liability lies. The concept of liability is likely to be complex and may be
split over several actors, such as the manufacturers/designers, the operators
and the environment. Regulation is crucial and first steps have been taken
to go beyond safety and reliability regulations Bryson and Winfield [2017];
International Organization for Standardization (ISO) [2014] into considering
the ethical aspects that should be taken into account British Standards
Institution (BSI) [2016].
Any system operating in the real world would eventually find itself in a
situation in which it will malfunction and AI systems are no exception. The
question is thus how certain can one be in the verified ethical behavior of
an AI system and what measures can be taken to mitigate the consequences
of, and learn from, a system’s potential failure. This is the issue of having
confidence in the system.
In terms of safety standards, the “gold standard” is currently that of aircraft autopilot software where safety is measured as the number of accidents
per miles flown. We might think that, for AI systems, at least as much confidence is needed. But, is the standard too high or even achievable? There
are of course, noticeable differences between aircraft and other autonomous
16
systems. While the operational environment of an aircraft is very controlled
and limited, as the aircraft must adhere to a strictly defined flying corridor,
the severity of accidents is very high. E.g., any malfunction in the air is
certainly fatal for all of the aircraft passengers measuring in the hundreds,
whereas a miscalculation on the road does not need to be fatal since cars
carry fewer passengers than airplanes. It is possible that the hours of operation per accident alone is not always the best measure to assess safety
of an autonomous system, but that the severity of the damage caused and
the number of individuals involved in an accident should also be taken into
account Patchett et al. [2015]; Kelly and McDermid [2001]; Denney and Pai
[2014]; Webster et al. [2014].
4.2
What do we want the system to do?
A key problem is specifying what our expectations of an AI system are. Although this is beginning to be codified where safety is considered, for example through robot safety standards9 , it is less clear where the ethical/moral
requirements should come from and in what form should they be represented? The BS8611 standard British Standards Institution (BSI) [2016],
for example, does not prescribe what the ethical requirements should be,
but maps out the issues over which ethical decisions should be considered.
An obvious route for ethical and legal requirements is through regulatory
or standards bodies. These entities have the ability to set overall standards,
potentially with the help of domain experts. In addition, designers may well
have built in specific ethical codes that go beyond (though do not contradict) those prescribed by regulations. Finally, the user herself may wish to
input her ethical preferences, ensuring that the AI acts in a way that is
personally acceptable. Since there are multiple actors that need to define
and refine the ethical requirements of the system, each with varying levels
of technical expertise, the issue arises of how the ethical requirements are
represented for the machine and for concerned actors. No one clear methodology emerges. One possibility is to have them represented in the form of
a set of legal or formal rules, as argued in Saptawijaya and Moniz Pereira
[2016]. Another possibility is to use a set of example scenarios developed to
test specific ethical choices, as in Anderson and Anderson [2014]. A third,
but by no means final, possibility is as a statistical envelope around a large
(and possibly random) set of test cases, against which the AI system must
be exhaustively assessed.
9
See International Organization for Standardization (ISO) [2016] for a range or robotic
safety standards.
17
4.3
How do we show that the AI systems meets the expectations?
There is a well-established body of work tackling the Verification and Validation (V&V) of systems, both hardware-centred and software-rich. The
aim of Verification is to ensure that a system meets its requirements; Formal
Verification takes this further, not only having precise formal requirements,
but carrying out a comprehensive mathematical analysis of the system to
‘prove’ whether it corresponds to these formal requirements. There are
many varieties of formal verification, the most popular being model checking Clarke et al. [1999]; Armstrong et al. [2012], whereby formal requirements are checked (usually automatically) against all possible executions of
the system. Verification, via model checking, is widely used especially for
the analysis of the safety and reliability of robotic systems both in terms
of physical navigation Mitsch et al. [2013] and in terms of internal decisionmaking Dennis et al. [2016a]. What is being verified is that the behaviour
of a particular system conforms to defined expectations. In terms of ethical/moral verification, it seems clear that if an AI system acts by following
mathematically specified rules, we can potentially formally verify its highlevel behavior. Only recently, however, has the use of formal verification for
ethical or moral issues begun to be addressed Dennis et al. [2016b, 2015].
A practical alternative to fully formal verification is to use sophisticated
coverage-driven analysis methods, appealing to Monte-Carlo techniques and
dynamic test refinement in order to systematically “cover” a wide range
of practical situations. Especially where real-world interactions and devices are involved, testing is likely to be crucial. Indeed, testing for safety
and reliability of robotic systems is well-established Mossige et al. [2015].
Such model-based testing is a well-developed technology but, as we move to
more complex (ethical) issues sophisticated extensions may well be required.
Though such approaches are typically used before deployment, related techniques provide a basis for run-time verification and compliance testing Rosu
and Havelund [2005]. Testing is not as exhaustive as formal proof, but can
cover many more scenarios.
Validation is the process of confirming that the final system has the
intended behavior once it is active in its target environment, and is often
concerned with satisfying external stakeholders. For example, does our system match ethical standards or legal rules set by regulators? Does our
system perform acceptably from a customer point of view, and how well do
users feel that it works Lehmann et al. [2013]? There are many approaches
to carrying out validation, typically involving the assessment of accuracy,
18
repeatability, trust, usability, resilience, etc. All must be extended to cope
with ethical and moral concerns.
It is clear that the strength and breadth of V&V research should allow
us to extend and develop this towards ethical and moral concerns. However,
a number of issues remain, as follows.
• If the core software is not purely rule-based, for example involving
some sub-symbolic learning procedures, then we will need a symbolic
representation of the learned content if we are to carry out formal
verification of the above form. One of the limitations of both formal
verification and testing is likely to be in verifying learning procedures,
especially where new ethical principles and preferences of behavior are
learned.
• Fully formal verification is likely to be unrealistic for complete, complex systems both because of non-symbolic components (as mentioned
above) and because of practical complexity limits.
However, we can formally verify parts of the system under particular circumstances. There are things that can be proved about core parts of the
system and about the system’s outputs. Consequently, formal verification
techniques can provide some evidence. In assessing how much confidence we
need in the V&V of AI system ethics, it may be possible to leave the burden
of this decision to the regulator, manufacturer or end-user as appropriate.
So long as a clear indication of the extent of the V&V of a system exists
a user or other interested may take the decision about the risk involved
in using the system. Note that we can potentially separate regulation from
verification and so allow a variety of different V&V techniques to be applied.
Lastly, we would like to include here the existing efforts of validating a
system that uses soft AI methods, which is the discussion of Ethical Turing
Tests. Ethical Turing Tests were introduced in Allen et al. [2000]. In Anderson and Anderson [2014] this idea is further fleshed out and implemented.
Under an ethical Turing test, both the AI system and an ethicist resolve
the same dilemmas. The system passes the test if its choices are sufficiently
similar to the ones of the ethicist. Whether a variant of a Turing test is
a sufficient indicator of a certain type of “human-like” behaviour from a
machine is a topic that has been argued as long as the artificial intelligence
field exists. All the issues that have been raised, and exhaustively discussed
in artificial intelligence, against the Turing test original can be argued to
hold for an Ethical Turing test.
19
5
Transparency and accountability
The opacity and transparency of deep neural network algorithms has become a major research subject area in recent years. Without knowing how
the algorithm functions concerns about inherent biases in data and algorithm itself create difficulties in either discerning forensically or through
explainability how decisions were made. Furthermore, there are questions
as to whether the actual logic used by learning systems can be explained to
people, or whether any explanation would be created after the fact.
The problem with interpretability/explainability is that in some cases
it may be impossible to provide a complete and accurate explanation for
how a black box system has arrived at its decision. As complexity of systems increases, it is not unusual for the algorithm to extract a milliondimensional feature vector and assign unique weights to each feature. Any
human-readable explanation for the decision will include some top N most
important features and completely ignore the rest. A human-comprehensible
explanation cant be too long or too complex. A good metaphor for this is
how we explain things to children if they are not old enough to fully appreciate nuances of the problem. Where do kids come from? You buy them
at the store! Any human-friendly explanation from a sufficiently complex
system has to be a partially inaccurate simplification or a complete lie.
The choice of the relevant criteria for an AI system to be deemed ethical
will eventually need to be taken by society as a whole. Therefore transparency is of utmost importance and thus ensuring transparency is a major
challenge. To this end it is necessary to identify what has to be transparent
to whom, and how this can be realized.
Transparency is a key requirement for ethical machines. Important attributes flow from transparency including trust, because it is hard to trust
a machine unless you have some understanding of what it is doing and why,
and accountability, because without transparency it becomes very difficult
to understand who is responsible when a machine does not behave as we
expect it to10 . An ethical machine will need to be transparent to different
stakeholders in different ways – each suited to that particular stakeholder.
In this section we consider the transparency needs of a range of stakeholders before considering aspects of transparency common to all. This section
outlines how and why transparency is important to four different groups
of stakeholders: users, regulators (including accident investigators), ethicists/lawyers and society at large. Each group has different transparency
10
Although it is important to note that transparency is not the same as accountability.
20
needs, some of which will have to be met by allowing an AI system’s ethics,
and ethical logic, to be human readable, or through public engagement.
Other needs will require new human-robot interfaces.
Some literature exists on the topic of transparency in AI and autonomous
systems. Owotoki and Mayer-Lindenberg Owotoki and F. [2007] proposes a
theoretical framework for providing transparency in computational intelligence (CI) in order to expose the underlying reasoning process of an agent
embodying CI models. In a recent book Taylor and Kelsey Taylor and Kelsey
[2016] make the case for the importance of transparency in AI systems to
an open society. For autonomous robots Wortham et al. [2016] describes
early results showing that building transparency into robot action-selection
can help users build a more accurate understanding of the robot. There is
also no doubt that transparency is high on policy agenda: the 2016 UK Parliamentary Select Committee on Science and Technology’s final report on
Robotics and AI expresses concerns over both decision making transparency
and accountability and liability11 . Indeed the EU’s new General Data Protection Regulation, due to take effect as law in 2018, creates a “right to
explanation” such that a user will be able to ask for an explanation of an
algorithmic decision that was made about them Goodman and Flaxman
[2016].
5.1
Transparency to the user
Although the critical importance of the human-machine interface is well
understood, what is not yet clear is the extent to which an ethical machine’s
ethics should be transparent to its user. It would seem to be unwise to
rely on a user to discover a machine’s ethics by trial and error, but at the
same time a machine that requires its user to undergo a laborious process
of familiarisation may well be unworkable.
For care robots for instance it may be appropriate for the user to configure the “ethics” settings (perhaps expressing the user’s preference for more
or less privacy) or, at the very least, allowing the user to choose between
a small number of “preset” ethics options. There is of course always some
danger that many users will rely on the default setting. What is clear is
that how these options are presented to the user is very important Matthias
[2015]; they should for instance help and guide the user in thinking about
their ‘value hierarchy’. The robot might for instance explain to the user
11
http://www.publications.parliament.uk/pa/cm201617/cmselect/cmsctech/145/
145.pdf
21
“what would happen” in different situations and hence guide their preferences Theodorou et al. [2016].
For other robot types, driverless cars for instance, the ethics settings may
be fixed (perhaps by law) and therefore not user configurable. However,
the need for the user to understand how the car would behave in certain
situations remains critically important – especially if the car’s design (or
the law) requires her to act as a safety driver and assume manual control
when the autopilot cannot cope. Even for fully autonomous cars in which
the user is only ever a passenger, the person with legal responsibility for the
car should be aware of the car’s ethics settings. For fully autonomous cars
there should still be some user interface so that the passenger can discover,
or perhaps ask for help, if the vehicle become unexpectedly immobile or
starts behaving erratically.
5.2
Transparency to regulatory bodies
It is clear that the ethics of ethical robots needs to be transparent to those
responsible for (i) certifying the safety of ethical machines, and (ii) accident
investigators. Both regulators and accident investigators will be working
within a governance framework which includes standards and protocols. The
role of the protocols is to set out how robots are certified against those
standards, and – following an accident – how the accident is investigated.
Regulators will need the ability to determine that a machine’s ethics
comply with the appropriate standards12 , and making such a determination
will require those ethics to be coded and embedded into the robot in a
readable way. We might imagine something like a standard Ethics Markup
Language (EML – perhaps based on XML) which codes the ethics. The
EML script would be embedded in the robot in a way that is accessible
to the regulator, noting that the script will need to be secured to prevent
attack from hackers.
Accident Investigators. When serious accidents happen, as they inevitably will (see Section 4.1), they will need to be investigated. To allow such investigation, data must to be recorded, suggesting the need for a
robot equivalent of the flight data recorder. Therefore an Ethical Black
Box (EBB) is proposed Winfield and Jirotka [2017]– a device that records all
relevant data including, crucially, internal state data on the robot’s ethical
governor. Although the data stored by the EBB would be vital for investigating all aspects of an accident, including causes unrelated to the robot’s
12
Standards Which do not yet exist.
22
ethics, here we are interested in accidents which might have been caused
by a fault or deficiency in the robot’s ethics programming. By recording
the sequence of internal states of the ethical reasoning in the moments before the accident, the EBB would allow an investigator to discover exactly
why the robot made an incorrect decision. Information that would be important both in determining accountability, and to make recommendations
for upgrading the robot’s ethics and prevent the same accident happening
again.
Specifying the EBB is beyond the scope of this paper, this work has
been carried further in Winfield and Jirotka [2017]. It is however clear
that research is needed to determine what data the EBB must record, the
frequency and time window of that data, and how the privacy of that data
is maintained. One thing we can be sure of however is the need for an
industry standard EBB (as in the aviation industry). Different EBBs and
EBB standards will of course be needed for different applications, but for
driverless cars for instance, a single standard EBB should be mandated.
Such an EBB would itself require an industry standard, and protocols for
certification, fitting and maintenance of EBBs.
5.3
Transparency to ethicists / lawyers
A third group of stakeholders includes lawyers, who might be required to
advocate for AI systems’ owners, or on behalf of anyone who makes a claim
against an AI system’s owner, or ethicists who might, for instance, be required to act as expert witnesses, in a court of law. If we consider an accident
in which a robot’s ethics are implicated (see Section 5.2 above), it is clear
that both lawyers and ethicists will need to understand (i) a robot’s ethics,
(ii) the process the robot uses to make an ethical decision (in other words
how its ethical reasoning works), and (iii) the data captured by the ethical
black box. Providing this kind of transparency to lawyers and ethicists will
not only be necessary, but is also likely to be challenging, as robot manufacturers and designers may regard such details, especially (ii), as proprietary
IP.
Another category of expert stakeholder includes psychologists, who might
be required to either evaluate robots for their potential to cause psychological harm to the user, or as expert witnesses, in providing an investigation
with an expert evaluation of the psychological harm caused to robot user(s)
in a particular case.
23
5.4
Transparency to the whole of society
AI systems – and especially ethical AI systems – are a disruptive technology,
with potentially significant societal and economic impact, thus an easily
overlooked but important stakeholder is society as a whole. We only need
to consider driverless cars and trucks to appreciate the level of potential
disruption, to jobs and transport policy for instance, as already reflected in
the level of public and press interest in this technology.
It is therefore very important that the ethics of ethical AI systems should
be transparent to society at large, for two reasons. First, because citizens
should be able to make informed judgments about the kinds of AI system
they wish to have in their lives, and even more importantly those they do not
want in their lives, so that they can lobby their elected representatives and
ensure that government policy properly reflects those views. And second,
if society is to have confidence in the ethics of a class of ethical AI system
(driverless cars, for example) then it should accept a degree of collective
responsibility for those ethics.
5.5
Technical means to bring about transparency
It is clear that the different stakeholders outlined above have very different
transparency needs. Some of those needs are met through making the ethical
rules and logic readable (for instance for regulators, ethicists or lawyers), but
for others transparency can only be met through technical means. Here we
briefly outline several approaches to meeting those needs.
• Assisted living AI systems would benefit from a “Why did you do
that?” button which, when pressed, causes the robot to explain –
perhaps using speech synthesized text – why it carried out the previous
action. We could call the system behind this an “explanation module”.
For an AI system with a fixed set of responses the explanation module
should be relatively easy to implement, but for an AI system which
learns its ethics such an implementation could be challenging; in either
case the explanation module and its user interface would need very
careful design in order to meet the needs of a non-technical user.
• An ethical AI system which makes use of simulation based internal
models as part of some ethical governor (for example Winfield et al.
[2014]) might allow us to go further than the “Why did you do that?”
button, by making the robot’s internal simulation accessible to the
user. This would enable the user to ask the robot “What would you
do?” in a given situation. Clearly such a facility would need a much
24
more sophisticated user interface than a button press, but through visualisation tools we can imagine the user watching the robot’s internal
simulation running through various scenarios on a connected laptop
or tablet device. Note that a similar visualisation interface would be
of great value to accident investigators (Section 5.2), and expert witnesses or lawyers (section 5.3) to play back a robot’s internal simulation
in the moments leading up to an accident, and what the alternatives
open to the robot at the time might have been.
• The technical requirements for an ethical back box (EBB) were already
outlined in Section 5.2 above.
6
Dangerous and Deliberately Unethical AI
Finally, it is important to be aware of the ways people may abuse or manipulate AI systems. As with all technology AI systems can also be deliberately
abused for malice or to further one’s illegal goals Yampolskiy [2016]. While
our primary concern is to contribute towards designing AI systems that behave ethically within a human society Sotala and Yampolskiy [2015]; Yampolskiy [2015a] and promote human and animal welfare, some concern also
needs to be raised about how that AI system can protect itself against abuse
Yampolskiy and Spellchecker [2016]. By abuse we, of course, do not mean
mistreating the AI system in the sense in which a person or an animal can
be mistreated, but taking advantage of the capabilities and opportunities
offered by the AI system to commit criminal acts.
The abuse of an AI system can be achieved by hacking an existing system
or by deliberately creating an unethical AI system Pistono and Yampolskiy
[2016]; Vanderelst and Winfield [2016]. Hacking itself can be accomplished
in several ways. The code of the AI system might be directly hacked. But a
system can also be manipulated by interaction and such manipulation does
not necessarily require technical knowledge. This is illustrated by the shortlived Tay experiment. Tay was an artificial intelligence chatter-bot released
by Microsoft Corporation on March 23, 2016 and taken offline 16 hours after
launch13 . Tay was programmed to learn from conversation, however it took
the netizens a very short time to “train” it into making morally questionable
statements.
Manipulation by interaction can be accomplished both deliberately and
by accident. A learning based system can be led Yampolskiy [2014] intro
eliciting bad conclusions through crafted case descriptions, etc. By this
13
http://www.bbc.com/news/technology-35890188
25
means one can slowly train systems away from moral behavior. As an example of accidental manipulation consider the example of children learning
that driverless cars slow down in their presence, they might choose to make a
game out of it. Children playing with car’s reactions might annoy passengers
by causing delay; and might ultimately lead to the disabling of safeguards.
Purposeful creation of Malevolent AI can be attempted by a number of
diverse agents with varying degrees of competence and success. Each such
agent would bring its own goals/resources into the equation, but what is important to understand here is just how prevalent such attempts will be and
how numerous such agents can be. For example, we should be concerned
about: the military developing cyber-weapons and robot soldiers to achieve
dominance; governments attempting to use AI to establish hegemony, control people, or take down other governments; corporations trying to achieve
monopoly, destroying the competition through illegal means; villains trying
to take over the world and using AI as a dominance tool; black hats attempting to steal information, resources or destroy cyber infrastructure targets;
doomsday cults attempting to bring the end of the world by any means; the
depressed looking to commit suicide by AI; psychopaths trying to add their
name to history books in any way possible; criminals attempting to develop
proxy systems to avoid risk and responsibility; AI risk deniers attempting
to demonstrate that AI is not a risk factor and so ignoring caution; and even
AI safety researchers, if unethical, attempting to justify funding and secure
jobs by purposefully developing problematic AI.
The ethical and unethical behaviors of an AI system are not necessarily symmetrical. Existing systems define only a small part of the problem
space Yampolskiy [2015b]. Apart from ethical and unethical behavior, an AI
system can also exhibit a behavior that has neither been programmed nor
predicted as a particular combination of otherwise ethical rules and choices.
Lastly we must mention the potential for “cultural imperialism” when
designing the ethical behavior of an AI system. With globalisation, a product’s production and consumers are diverse. What constitutes ethical behavior in one region may even be considered unethical in another. All the
involved actors, the designers, users and society, both on the supplier and
on the demand end of the AI system need to be aware of the reality that
the supplier society ethics influences the ethical behavior of the AI system,
which in turn influences the ethics of the society in which the AI system
operates.
26
7
Summary
Moral philosophy has a very rich history of studying how to discern right
from wrong in a systematic, consistent and coherent way. Today we have
a real need for a functional system of ethical reasoning as AI systems that
function as part of our society are ready to be deployed. Building an AI
system that behaves ethically is a multifaceted challenge. The questions of
which ethical theory should be used to govern the AI system’s behavior has
received most of the attention. Here, we focus on the problem that comes
next, after what is right or wrong for a machine to do is decided – how to
implement the ethical behavior.
The problem of engineering ethical behavior is made complex because
of the prime motivators for such behavior. For humans, the motivation
for behaving ethically is primarily internal. Without falling into difficult
philosophical arguments on the existence of free will, we accept that people
are capable of behaving ethically because they choose to do so, although,
of course, they too can be motivated towards ethical behaviour by incentives, punishments and assignment of liability. For AI systems, however,
the motivation towards ethical behaviour is exclusively external because it
can always be traced back to their design and it cannot be reinforced in
the same way as it can be done with people. This motivation furhtermore
comes from several stakeholders. We cannot claim that the full list of these
stakeholders can even be known before the AI systems are fully deployed,
but we can discern between the three most evident groups of stakeholders:
the designers, the users and the various regulatory organs of society. Each
of these stakeholders needs to play their role in deciding what is the best
ethical behavior for a given AI system, but they also need to be convinced
in an adequate way that the implemented behavior actually yields the desired results. A moral AI system needs to be adequately transparent and
accountable to each group of stakeholders.
Unlike people, who more or less share the same “hardware” and reasoning
capabilities, machines and AI systems can be built using many different
approaches. The implementation of ethical reasoning will depend not only
on what the stakeholders need and desire, but also on what is possible
given the chosen problem-solving implementation. We discussed the two
basic implementation approaches reflecting two large families of AI methods:
the soft AI and the symbolic AI families, and identify the challenges and
advantages of each.
An AI system capable of ethical behavior is necessarily a complex system.
With complex systems two things are evident: that they will malfunction
27
and that they can be used to attain criminal goals. We discuss methods of
verifying that an AI system behaves as designed within specified parameters,
but we also discuss how the engineering of the ethical behavior impacts
available options once a system malfunctions. Lastly we discuss in broad
strokes what the stakeholders need to be aware of in terms of abuse of
an AI system with ethical behavior capabilities, both when that abuse is
intentional and accidental.
References
D. Abel, J. MacGlashan, and M. L. Littman. Reinforcement Learning As a
Framework for Ethical Decision Making. In Workshops at the Thirtieth
AAAI Conference on Artificial Intelligence, 2016.
C. Allen, G. Varner, and J. Zinser. Prolegomena to any future artificial moral
agent. Journal of Experimental and Theoretical Artificial Intelligence,
12:251–261, 2000.
C. Allen, W. Wallach, and I. Smit. Why machine ethics? IEEE Intelligent
Systems, 21(4):12–17, 2006.
M. Anderson and S. Leigh Anderson. Geneth: A general ethical dilemma
analyzer. In Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intelligence, July 27 -31, 2014, Québec City, Québec, Canada., pages
253–261, 2014.
M. Anderson and S. Leigh Anderson. Machine ethics: Creating an ethical
intelligent agent. AI Magazine, 28(4):15–26, 2007.
M. Anderson and S. Leigh Anderson. Toward ensuring ethical behavior
from autonomous systems: a case-supported principle-based paradigm.
Industrial Robot, 42(4):324–331, 2015.
M. Anderson, S. Leigh Anderson, and Vincent Berenz. Ensuring ethical
behavior from autonomous systems. In Artificial Intelligence Applied to
Assistive Technologies and Smart Environments, Papers from the 2016
AAAI Workshop, Phoenix, Arizona, USA, February 12, 2016, 2016.
R.C. Arkin, P. Ulam, and A. R. Wagner. Moral Decision Making in Autonomous Systems: Enforcement, Moral Emotions, Dignity, Trust, and
Deception. Proceedings of the IEEE, 100(3):571–589, 2012.
28
P.J. Armstrong, M. Goldsmith, G. Lowe, J. Ouaknine, H. Palikareva, A. W.
Roscoe, and J. Worrell. Recent Developments in FDR. In P. Madhusudan
and S. A. Seshia, editors, Computer Aided Verification: 24th International
Conference, CAV 2012, Berkeley, CA, USA, July 7-13, 2012 Proceedings,
volume 7358 of LNCS, pages 699–704. Springer Berlin Heidelberg, 2012.
S. Armstrong. Motivated value selection for artificial agents. In Artificial
Intelligence and Ethics, Papers from the 2015 AAAI Workshop, Austin,
Texas, USA, January 25, 2015., 2015. http://aaai.org/ocs/index.
php/WS/AAAIW15/paper/view/10183.
I. Asimov. I, Robot. Gnome Press, 1950.
M. Balconi and A. Bortolotti. Detection of the facial expression of emotion and self-report measures in empathic situations are influenced by
sensorimotor circuit inhibition by low-frequency rtms. Brain Stimulation,
5(3):330 – 336, 2012.
O. Bendel. Annotated decision trees for simple moral machines. 2016.
J.F. Bonnefon, A. Shariff, and I. Rahwan. The social dilemma of autonomous
vehicles. Science, 352(6293):1573–1576, 2016.
C. Breazeal and B. Scassellati. Robots that imitate humans. Trends in
Cognitive Sciences, 6(11):481–487, 2002.
British Standards Institution (BSI). BS 8611 Robots and Robotic Devices —
Guide to the ethical design and application. http://www.bsigroup.com,
2016.
J. Bryson and A.F.T. Winfield. Standardizing ethical design for artificial
intelligence and autonomous systems. IEEE Computer, 50(5):116–119,
2017.
A. Cangelosi and M. Schlesinger. Developmental Robotics: From Babies to
Robots. The MIT Press, 2014.
M. Caycedo Alvarez, Ø. S. Berge, A. S. Berget, E. S. Bjørknes, D.V.K.
Johnsen, F. O. Madsen, and M. Slavkovik. Implementing Asimov’s first
law of robotics. In 30th Norsk Informatikkonferanse, NIK 2017, Westerdals Oslo ACT, November, 27-29, 2017, 2017. forthcoming.
E. M. Clarke, O. Grumberg, and D. Peled. Model Checking. MIT Press,
1999.
29
E. Denney and J.P. Pai. Automating the assembly of aviation safety cases.
IEEE Trans. Reliability, 63(4):830–849, 2014.
L. A. Dennis, M. Fisher, and A. F. T. Winfield. Towards Verifiably Ethical
Robot Behaviour. In Proc. AAAI Workshop on AI and Ethics, 2015.
http://aaai.org/ocs/index.php/WS/AAAIW15/paper/view/10119.
L. A. Dennis, M. Fisher, N. K. Lincoln, A. Lisitsa, and S. M. Veres. Practical Verification of Decision-Making in Agent-Based Autonomous Systems.
Automated Software Engineering, 23(3):305–359, 2016.
L. A. Dennis, M. Fisher, M. Slavkovik, and M. P. Webster. Formal Verification of Ethical Choices in Autonomous Systems. Robotics and Autonomous Systems, 77:1–14, 2016.
P. Ekman. Are there basic emotions? Psychological Review, 99(3):550–553,
1992.
C. Elgin. Considered Judgment. Princeton: New Jersey: Princeton University Press, 1996.
J. W. Ellington. Translation of: Grounding for the Metaphysics of Morals:
with On a Supposed Right to Lie because of Philanthropic Concerns by
Kant, I. [1785]. Hackett Publishing Company, 1993.
A. Etzioni and O. Etzioni. Incorporating ethics into artificial intelligence.
The Journal of Ethics, pages 1–16, 2017.
M. Fisher, L. A. Dennis, and M. Webster. Verifying Autonomous Systems.
ACM Communications, 56(9):84–93, 2013.
M. Fisher, C. List, M. Slavkovik, and A. F. T. Winfield. Engineering moral
agents - from human morality to artificial morality (dagstuhl seminar
16222). Dagstuhl Reports, 6(5):114–137, 2016.
T. Fong, I. Nourbakhsh, and K. Dautenhahn. A survey of socially interactive robots. Robotics and Autonomous Systems, 42(3–4):143 – 166, 2003.
Socially Interactive Robots.
P. Foot. The problem of abortion and the doctrine of double effect. Oxford
Review, 5:5–15, 1967.
T. Froese and E. A. Di Paolo. Modelling social interaction as perceptual
crossing: An investigation into the dynamics of the interaction process.
Connection Science, 22(1):43–68, March 2010.
30
S. Gallagher. Empathy, simulation, and narrative. Science in Context,
25(3):355–381, 009 2012.
B. Goodman and S. Flaxman. Eu regulations on algoritic decision-making
and a “right to explanation”. arXiv preprint arXiv:1606.08813, 2016.
J.C. Harsanyi. Rule utilitarianism and decision theory. Erkenntnis (1975-),
11(1):25–53, 1977.
M. Hoffman. Empathy and Moral Development: Implications for Caring and
Justice. Cambridge University Press, 2001.
International Organization for Standardization (ISO). ISO 13482: Robots
and robotic devices — Safety requirements for Personal Care Robots.
http://www.iso.org, 2014.
International Organization for Standardization (ISO). TC299 — Robotics,
2016.
T.P. Kelly and J.A. McDermid. A systematic approach to safety case maintenance. Rel. Eng. & Sys. Safety, 71(3):271–284, 2001.
H. Lehmann, D. S. Syrdal, K. Dautenhahn, G.J. Gelderblom, S. Bedaf, and
F. Amirabdollahian. What Should a Robot do for you? Evaluating the
Needs of the Elderly in the UK. In Proc. 6th Int. Conf. on Advances in
Computer-Human Interactions, pages 83–88, 2013.
S. Leigh Anderson.
Asimov’s ’three laws of robotics’ and machine
metaethics. AI & Society, 22(4):477–493, 2008.
M. J. Mataric. Getting humanoids to move and imitate. IEEE Intelligent
Systems, 15(4):18–24, July 2000.
A. Matthias. Algorithmic moral control of war robots: Philosophical questions. Law, Innovation and Technology, 3(2):279–301, 2011.
A. Matthias. Robot lies in health care: when is deception morally permissible? Kennedy Institute of Ethics Journal, 25(2):169–162, 2015.
C. Misselhorn. Empathy with inanimate objects and the uncanny valley.
Minds and Machines, 19(3):345, 2009.
S. Mitsch, K. Ghorbal, and A. Platzer. On Provably Safe Obstacle Avoidance for Autonomous Robotic Ground Vehicles. In Robotics: Science and
Systems IX, 2013.
31
J. H. Moor. The nature, importance, and difficulty of machine ethics. IEEE
Intelligent Systems, 21(4):18–21, July 2006.
M. Mossige, A. Gotlieb, and H. Meling. Testing Robot Controllers using Constraint Programming and Continuous Integration. Information
& Software Technology, 57:169–185, 2015.
P. Owotoki and Mayer-Lindenberg F. Transparency of computational intelligence models. In Research and Development in Intelligent Systems
XXIII, The 26th SGAI International Conference on Innovative Techniques and Applications of Artificial Intelligence, Proceedings, pages 387–
392. Springer, 2007.
C. Patchett, M. Jump, and M. Fisher. Safety and Certification of Unmanned
Air Systems. In Engineering and Technology Reference. 2015.
F. Pistono and R. V Yampolskiy. Unethical research: How to create a malevolent artificial intelligence. In 25th International Joint Conference on Artificial Intelligence (IJCAI-16). Ethics for Artificial Intelligence Workshop
(AI-Ethics-2016), 2016.
G. Rizzolatti and M. Fabbri-Destro. The mirror system and its role in
social cognition. Current Opinion in Neurobiology, 18(2):179 – 184, 2008.
Cognitive neuroscience.
W.D. Ross. The Right and the Good. Oxford University Press, 1930.
G. Rosu and K. Havelund. Rewriting-Based Techniques for Runtime Verification. Automated Software Engineering, 12(2):151–197, 2005.
S.J. Russell and P. Norvig. Artificial Intelligence: A Modern Approach.
Pearson Education, 3 edition, 2015.
M. Salem, G. Lakatos, F. Amirabdollahian, and K. Dautenhahn. Would You
Trust a (Faulty) Robot?: Effects of Error, Task Type and Personality on
Human-Robot Cooperation and Trust. In Proc. 10th Annual ACM/IEEE
International Conference on Human-Robot Interaction (HRI), pages 141–
148. ACM, 2015.
A. Saptawijaya and L. Moniz Pereira. Logic programming for modeling
morality. Logic Journal of the IGPL, 24(4):510–525, 2016.
A. Sharkey and N. Sharkey. Granny and the robots: ethical issues in robot
care for the elderly. Ethics and Information Technology, 14(1):27–40, 2012.
32
M. A. Slote. The Ethics of Care and Empathy. Routledge, 2007.
J. Sombetzki. Responsibility in Crisis — From the Traditional Concept of
Responsibility to Systems Responsibility, 2015.
I. Sommerville. Software Engineering. Pearson Studium, 2001.
K. Sotala and R. V Yampolskiy. Responses to catastrophic agi risk: A
survey. Physica Scripta, 90(1), 2015.
K. Stüeber. Rediscovering Empathy: Agency, Folk Psycholgy, and the Human Sciences. MIT Press, 2006.
R. Taylor and T. Kelsey. Transparency and the open society: Practical
lessons for effective policy. Policy Press, Bristol UK, 2016.
A. Theodorou, R. Wortham, and J.J. Bryson. Why is my robot behaving
like that? designing transparency for real time inspection of autonomous
robots. In AISB Workshop on Principles of Robotics, April 2016, Sheffield
UK, Proceedings, 2016.
S. P. van Rysewyk and M. Pontier. A Hybrid Bottom-Up and Top-Down
Approach to Machine Medical Ethics: Theory and Data, pages 93–110.
Springer International Publishing, Cham, 2015.
D. Vanderelst and A.F. Winfield. The dark side of ethical robots. arXiv
preprint arXiv:1606.02583, 2016.
D. Vanderelst and A. Winfield. An architecture for ethical robots inspired
by the simulation theory of cognition. Cognitive Systems Research, 2017.
W. Wallach and C. Allen. Moral Machines: Teaching Robots Right from
Wrong. Oxford University Press, 2008.
W. Wallach, C.n Allen, and I. Smit. Machine morality: Bottom-up and topdown approaches for modelling human moral faculties. AI Soc., 22(4):565–
582, March 2008.
F. Warneken and M. Tomasello. Varieties of altruism in children and chimpanzees. Trends in Cognitive Sciences, 13(9):397 – 402, 2009.
M. Webster, N. Cameron, M. Fisher, and M. Jump. Generating Certification
Evidence for Autonomous Unmanned Aircraft Using Model Checking and
Simulation. Journal of Aerospace Information Systems, 11(5):258–279,
2014.
33
A.F.T. Winfield and M. Jirotka. The case for an ethical black box. In
Towards Autonomous Robotic Systems - 18th Annual Conference, TAROS
2017, Guildford, UK, July 19-21, 2017, Proceedings, pages 262–273, 2017.
A. F. T. Winfield, C. Blum, and W. Liu. Towards an Ethical Robot: Internal
Models, Consequences and Ethical Action Selection, pages 85–96. Springer
International Publishing, 2014.
R.H. Wortham, A. Theodorou, and J.J. Bryson. What does the robot think?
transparency as a fundamental design requirement for intelligent systems.
In IJCAI-2016 Ethics for Artificial Intelligence Workshop, July 2016, New
York USA, Proceedings, 2016.
R. V Yampolskiy and MS Spellchecker. Artificial intelligence safety and
cybersecurity: a timeline of ai failures. arXiv preprint arXiv:1610.07997,
2016.
R.V. Yampolskiy. Utility function security in artificially intelligent agents.
Journal of Experimental & Theoretical Artificial Intelligence, 26(3):373–
389, 2014.
R. V. Yampolskiy. Artificial superintelligence: A futuristic approach, 2015.
R. V Yampolskiy. The space of possible mind designs. In Artificial General
Intelligence: 8th International Conference, AGI 2015, AGI 2015, Berlin,
Germany, July 22-25, 2015, Proceedings, volume 9205, page 218. Springer,
2015.
R. V Yampolskiy. Taxonomy of pathways to dangerous artificial intelligence.
In Workshops at the Thirtieth AAAI Conference on Artificial Intelligence,
2016.
34
| 2cs.AI
|
Verifying Chemical Reaction Network Implementations:
A Pathway Decomposition Approach ∗
Seung Woo Shin1 , Chris Thachuk2 , and Erik Winfree2
arXiv:1411.0782v3 [cs.CE] 10 Aug 2017
1
2
University of California, Berkeley
California Institute of Technology
Abstract
The emerging fields of genetic engineering, synthetic biology, DNA computing, DNA nanotechnology, and molecular programming herald the birth of a new information technology that acquires information by directly sensing
molecules within a chemical environment, stores information in molecules such as DNA, RNA, and proteins, processes that information by means of chemical and biochemical transformations, and uses that information to direct
the manipulation of matter at the nanometer scale. To scale up beyond current proof-of-principle demonstrations, new
methods for managing the complexity of designed molecular systems will need to be developed. Here we focus on
the challenge of verifying the correctness of molecular implementations of abstract chemical reaction networks, where
operation in a well-mixed “soup” of molecules is stochastic, asynchronous, concurrent, and often involves multiple
intermediate steps in the implementation, parallel pathways, and side reactions. This problem relates to the verification of Petri nets, but existing approaches are not sufficient for providing a single guarantee covering an infinite set
of possible initial states (molecule counts) and an infinite state space potentially explored by the system given any
initial state. We address these issues by formulating a new theory of pathway decomposition that provides an elegant
formal basis for comparing chemical reaction network implementations, and we present an algorithm that computes
this basis. Our theory naturally handles certain situations that commonly arise in molecular implementations, such as
what we call “delayed choice,” that are not easily accommodated by other approaches. We further show how pathway
decomposition can be combined with weak bisimulation to handle a wider class that includes most currently known
enzyme-free DNA implementation techniques. We anticipate that our notion of logical equivalence between chemical
reaction network implementations will be valuable for other molecular implementations such as biochemical enzyme
systems, and perhaps even more broadly in concurrency theory.
Keywords: chemical reaction networks; molecular computing; DNA computing; formal verification; molecular programming; automated design
1
Introduction
A central problem in molecular computing and bioengineering is that of implementing algorithmic behavior
using chemical molecules. The ability to design chemical systems that can sense and react to the environment finds applications in many different fields, such as nanotechnology [8], medicine [13], and robotics
[18]. Unfortunately, the complexity of such engineered chemical systems often makes it challenging to
∗
Preliminary versions of this manuscript appeared in the proceedings of VEMDP 2014 and are available on arXiv:1411.0782
[cs.CE].
1
ensure that a designed system really behaves according to specification. Furthermore, since experimentally
synthesizing chemical systems can require considerable resources, mistakes are generally expensive, and it
would be useful to have a procedure by which one can theoretically verify the correctness of a design using
computer algorithms prior to synthesis. In this paper we propose a theory that can serve as a foundation for
such automated verification procedures.
Specifically, we focus our attention on the problem of verifying chemical reaction network (CRN) implementations. Informally, a CRN is a set of chemical reactions that specify the behavior of a given chemical
system in a well mixed solution. For example, the reaction equation A + B → C means that a reactant
molecule of type A and another of type B can be consumed in order to produce a product molecule of type
C. A reaction is applicable if all of its reactants are present in the solution in sufficient quantities. In case
both A + B → C and C → A + B are in the CRN, we may also use the shorthand notation A + B
C.
In general, the evolution of the system from some initial set of molecules is a stochastic, asynchronous,
and concurrent process. While abstract CRNs provide the most widely used formal language for describing
chemical systems, and have done so for over a century, only recently have abstract CRNs been used explicitly as a programming language in molecular programming and bioengineering. This is because CRNs are
often used to specify the target behavior for an engineered chemical system (see Figure 1). How can one
realize these “target” CRNs experimentally? Unfortunately, synthesizing chemicals to efficiently interact —
and only as prescribed — presents a significant, if not infeasible, engineering challenge. Fortunately, any
target CRN can be emulated by a (generally more complex) “implementation” CRN. For example, in the
field of DNA computing, implementing a given CRN using synthesized DNA strands is a well studied topic
that has resulted in a number of translation schemes [36, 5, 31].
In order to evaluate CRN implementations prior to their experimental demonstration, a mathematical
model describing the expected molecular interactions is necessary. For this purpose, software simulators
that embody the relevant physics and chemistry can be used. Beyond performing simulations – which by
themselves can’t provide absolute statements about the correctness of an implementation – it is often possible to describe the model of the molecular implementation as a CRN. That is, software called “reaction
enumerators” can, given a set of initial molecules, evaluate all possible configuration changes and interactions, possibly generating new molecular species, and repeating until the full set of species and reactions
have been enumerated. In the case of DNA systems, there are multiple software packages available for
this task [24, 17]. More general biochemical implementations could be modeled using languages such as
BioNetGen [16] and Kappa [10].
Given a “target” CRN which specifies a desired algorithmic behavior and an “implementation” CRN
which purports to implement the target CRN, how can one check that the implementation CRN is indeed
correct? As we shall see, this question involves subtle issues that make it difficult to even define a notion
of correctness that can be universally agreed upon, despite the fact that in this paper we study a somewhat
simpler version of the problem in which chemical kinetics, i.e. rates of chemical reactions, is dropped from
consideration. However, we note that this restriction is not without its own advantages. For instance, when
basing a theory on chemical kinetics, it is of interest to accept approximate matches to the target behavioral
dynamics [38, 39], which may overlook certain logical flaws in the implementation that occur rarely. While
theories of kinetic equivalence are possible and can in principle provide guarantees about timing [7], they
can be difficult to apply to molecular engineering in practice. In contrast, a theory that ignores chemical
kinetics can be exact and therefore emphasize the logical aspect of the correctness question.
The main challenge in this verification problem lies in the fact that the implementation CRN is usually
much more complex than the target CRN. This is because each reaction in the target CRN, which is of
course a single step in principle, gets implemented as a sequence of steps which may involve “intermediate”
species that were not part of the original target CRN. For example, in DNA-based implementations, the implementation CRN can easily involve an order of magnitude more reactions and species than the target CRN
(the size will depend upon the level of detail in the model of the implementation [24, 17, 33, 14]). Given
2
that the intermediate species participating in implementations of different target reactions can potentially
interact with each other in spurious ways, it becomes very difficult to verify that such an implementation
CRN is indeed “correct.”
A+B →C +D
C +A→C +C
A
i
i+B
j
i+j →C +k
k
D
C +A
m+n
m+n→C +C
A
i
i+B
j
j
C +k
k→D
C +A
m+n
m+n→C +C
CRN1
CRN2
CRN3
A→i
i+B
j
j →C +k
k
D
C +A
m+n
m+n→C +C
A
i
i+B
j
j →C +k
k
D
C +A
m+n
m+n→C +C
A + g1
i + g2
i+B
j + g3
g4 + j → C + k + w1
g5 + k
D + w2
C +A
m+n
g6 + m + n → C + C + w3
CRN4
CRN5
CRN6
Figure 1: An example of CRN implementation. CRN1 represents the “target” CRN, i.e., the behavior we
desire to implement, whereas CRN2-5 are potential “implementations” of this target CRN. In these CRNs,
the lowercase species are “intermediate” species of the implementations, while the uppercase species are
“formal” species. CRN6 illustrates the way in which “fuel” and “waste” species may appear in a typical
DNA-based system, with fuel species denoted by gi and waste species denoted by wi . Removing inert waste
species and ever-present fuel species from CRN6 yields CRN5.
It is not immediately obvious how to precisely define what makes an implementation correct or incorrect,
so it is helpful to informally examine a few examples. Figure 1 illustrates various different ways that a
proposed implementation can be “incorrect.” For instance, one can easily see that CRN2 is clearly not
a good implementation of CRN1, because it implements the reaction A + A + B → C + D in place
of A + B → C + D. CRN3 is incorrect in a more subtle way. While a cursory look may not reveal
any immediate problem with this implementation, one can check that CRN3 can get from the initial state1
⦃A, A, B⦄ to a final state ⦃A, B, C⦄, whereas there is no way to achieve this using reactions from CRN1.2
CRN4 is incorrect in yet another way. Starting from the initial state ⦃A, C⦄, one can see that the system
will sometimes get “stuck” in the state ⦃i, C⦄, unable to produce ⦃C, C⦄, with i becoming an intermediate
species that is not really “intermediate.” Now, CRN5 seems to be free of any such issue, but with what
confidence can we declare that it is a correct implementation of CRN1, having seen the subtle ways that an
implementation can go wrong? A goal of this paper is to provide a mathematical definition of “correctness”
of CRN implementations which can be used to test them in practice.
In our further discussions, we will restrict our attention to implementation CRNs that satisfy the condition that we call “tidiness.” Informally stated, tidy CRNs are implementation CRNs which do not get
“stuck” in the way that CRN4 got stuck above, i.e., they always can “clean up” intermediate species. This
1
2
In this paper, we use the notation ⦃ · ⦄ to denote multisets.
The pathway is (A → i, i + B → j, j → C + k, C + A → m + n, m + n → C + C, C + k → j, j → i + B, i → A).
3
means that any intermediate species that are produced during the evolution of the system can eventually
turn back into species of the target CRN. Of course, the algorithm we present in this paper for testing our
definition of correctness will also be able to test whether the given implementation is tidy.
Finally, we briefly mention that many CRN implementations also involve what are called “fuel” and
“waste” species, in addition to the already mentioned intermediate species. Fuel species are helper species
that are assumed to be always present in the system at fixed concentration, whereas waste species are chemically inert species that sometimes get produced as a byproduct of implemented pathways (see CRN6 of
Figure 1 or for a more detailed explanation Example #1 of Section 6). While our core theory addresses the
version of the problem in which there is no fuel or waste species, as we demonstrate in Section 5, it can
easily be extended to handle the general case with fuel and waste species, using existing tools.
2
Motivations for a new theory
To one who is experienced in formal verification, the problem seems to be closely related to various wellstudied notions such as reachability, (weak) trace equivalence, (weak) bisimulation, serializability, etc. In
this section, we briefly demonstrate why none of these traditional notions seems to give rise to a definition
which is entirely satisfactory for the problem at hand.
The first notion we consider is reachability between formal states [28, 27, 15]. We call the species
that appear in both the target and the implementation CRNs “formal,” to distinguish them from species that
appear only in the implementation CRN, which we call “intermediate.” Formal states are defined to be states
which do not contain any intermediate species. Since we are assuming that our implementation CRN is tidy,
it then makes sense to ask whether the target CRN and the implementation CRN have the same reachability
when we restrict our attention to formal states only — this is an important distinction from the traditional
Petri net reachability-equivalence problem. That is, given some formal state, what is the set of formal states
that can be reached from that state using reactions from one CRN, as opposed to the other CRN? Do the
target CRN and the implementation CRN give rise to exactly the same reachability for every formal initial
state? While it is obvious that any “correct” implementation must satisfy this condition, it is also easy to see
that this notion is not very strong. For example, consider the target CRN {A → B, B → C, C → A} and
the implementation CRN {A → i, i → C, C → j, j → B, B → k, k → A}. The two CRNs are implementing opposite behaviors in the sense that starting from one A molecule, the target CRN will visit formal
states in the clockwise order ⦃A⦄, ⦃B⦄, ⦃C⦄, ⦃A⦄, ⦃B⦄, ⦃C⦄, . . ., whereas the implementation CRN will
visit formal states in the counter-clockwise order ⦃A⦄, ⦃C⦄, ⦃B⦄, ⦃A⦄, ⦃C⦄, ⦃B⦄, . . .. Nonetheless, they
still give rise to the same reachability between purely formal states.
Trace equivalence [15, 21] is another notion of equivalence that is often found in formal verification
literature. To our knowledge, it has not been applied in the context of CRN equivalence. We interpret its
application in this context as follows. Weak trace equivalence requires that it should be possible to “label”
the reactions of the implementation CRN to be either a reaction of the target CRN or a “null” reaction.
This labeling must be such that for any formal initial state, any sequence of reactions that can take place
in the target CRN should also be able to take place in the implementation CRN and vice versa, up to the
interpretation specified by the given labeling. However, it turns out to be an inappropriate notion in our
setting. For example, consider the target CRN {A
B, B
C, C
A} and the implementation CRN
{A
i, B
i, C
i}. The dynamics of the implementation appear correct since each reaction of
the target CRN can be simulated in the implementation CRN in the obvious way by exactly two reactions:
the first reaction consumes the reactant and produces an intermediate species i while the second reaction
consumes i and produces the intended formal species. However, these CRNs are not (weak-)trace equivalent.
Consider that every reaction of the implementation CRN must be labeled by one of the six formal reactions
(since the implementation CRN also consists of six reactions) and none can be labeled as a “null” reaction.
4
Since any initial reaction of the implementation CRN must begin in a formal state, and since there are only
three reactions that can occur from one of the three formal states, then any trace of the target CRN that begins
with one of the other three possible reactions cannot be simulated by the implementation CRN. Consider a
second example with target CRN {A
B, B → C, C → A} and implementation CRN {A
B, B →
j, j → C, C → A, C → ∅, ∅
i} where the implementation reactions {j → C, C → ∅, ∅ → i, i → ∅}
are labeled as “null” and the other reactions are labeled in the obvious way that is consistent with formal
species names. The implementation CRN exemplifies a common shortcoming of trace equivalence: inability
to distinguish the two systems with respect to deadlock. In our example the implementation CRN can in
principle simulate all finite and infinite traces of the target CRN, but once the first “null” reaction C → ∅
occurs then only “null” reactions can follow. In essence, the implementation CRN can become “stuck”
whereas the target CRN cannot. While (weak-)trace equivalence cannot distinguish based on deadlock
conditions as in our second example, other equivalence notions such as bisimulation can.
Bisimulation [29, 4] is perhaps the most influential notion of equivalence in state transition systems such
as CRNs, Petri nets, or concurrent systems [19, 34, 11]. A notion of CRN equivalence based on the idea
of weak bisimulation is explored in detail in [12, 22], and indeed it proves to be much more useful than
the above two notions. For bisimulation equivalence of CRNs, each intermediate species is “interpreted” as
some combination of formal species, such that in any state of the implementation CRN, the set of possible
next non-trivial reactions is exactly the same as it would be in the formal CRN. (Here, a “trivial” reaction
is one where the interpretation of the reactants is identical to the interpretation of the products.) However,
one potential problem of this approach is that it demands a way of interpreting every intermediate species
in terms of formal species. Therefore, if we implement the target CRN {A → B, A → C, A → D}
as {A → i, i → B, i → C, A → j, j → D}, we cannot apply this bisimulation approach because
the intermediate i cannot be interpreted to be any of A, B, or C. Namely, calling it A would be a bad
interpretation because i can never turn into D. Calling it B would be bad because i can turn into C whereas
B should not be able to turn into C. For the same reason calling it C is not valid either.
Perhaps this example deserves closer attention. We name this type of phenomenon the “delayed choice”
phenomenon, to emphasize the point that when A becomes i, although it has committed to becoming either
B or C instead of D, it has delayed the choice of whether to become B or C until the final reaction takes
place. This is the same phenomenon occurring in the first example given when discussing (weak-)trace
equivalence. Neither (weak-)trace equivalence nor bisimulation can be applied in systems that exhibit “delayed choice”. There are two reasons that the phenomenon is interesting; firstly, there may be a sense in
which it is related to the efficiency of the implementation, because the use of delayed choice may allow
for a smaller number of intermediate species in implementing the same CRN. Secondly, this phenomenon
actually does arise in actual systems, as presented in [17].
We note an important distinction between the various notions of equivalence discussed here and those
found in the Petri net literature. Whereas two Petri nets are compared for (reachability/trace/bisimulation)equivalence for a particular initial state [21], we are concerned about the various notions of equivalence of
two CRNs for all initial states. This distinction may limit the applicability of common verification methodologies and software tools [20, 3], since the set of initial states is by necessity always infinite (and the set
of reachable states from a particular initial state may also be infinite). Finally, we note that [25] proposes
yet another notion of equivalence based on serializability from database and concurrency theory. The serializability result works on a class of implementations that are “modular”. Formal reactions are encoded
by a set of implementation reactions and species. Roughly speaking, modular implementations ensure that
each formal reaction has a unique and correct encoding that does not “cross-talk” with the encodings of
other formal reactions. In general, this results in a one-to-one mapping between formal reactions and their
encodings. Implementation CRNs satisfying the formal modularity definitions of [25] will correctly emulate
their target CRN. However, this class of implementation CRNs precludes those that utilize “delayed choice”.
Interestingly, when restricted to “modular” implementations, the notion of serializability and our notion of
5
pathway decomposition have a close correspondence.
Our approach (originally developed in [35]) differs from any of the above in that we ignore the target
CRN and pay attention only to the implementation CRN. Namely, we simply try to infer what CRN the
given implementation would look like in a hypothetical world where we cannot observe the intermediate
species. We call this notion “formal basis.” We show that not only is the formal basis unique for any valid
implementation, but it also has the convenient property that a CRN that does not have any intermediate
species has itself as its formal basis. This leads us to a simple definition of CRN equivalence; we can
declare two CRNs to be equivalent if and only if they have the same formal basis. Therefore, unlike trace
equivalence or weak bisimulation [12, 22], our definition is actually an equivalence relation and therefore
even allows for the comparison of an implementation with another implementation.
3
3.1
Theory
Overview
In previous sections we saw that a reaction which is a single step in the target CRN gets implemented as
a pathway of reactions which involves intermediate species whose net effect only changes the number of
“formal” species molecules. For instance, the pathway A → i, i + B → j, j → C + k, k → D involves
intermediate molecules i, j, and k but the net effect of this pathway is to consume A and B and produce C
and D. In this sense this pathway may be viewed as an implementation of A + B → C + D.
In contrast, we will not want to consider the pathway A → i, i → B, B → j, j → C to be an
implementation of A → C, even though its net effect is to consume A and produce C. Intuitively, the
reason is that this pathway, rather than being an indivisible unit, looks like a composition of smaller unit
pathways each implementing A → B and B → C.
The core idea of our definition, which we call pathway decomposition, is to identify all the pathways
which act as indivisible units in the above sense. The set of these “indivisible units” is called the formal basis
of the given CRN. If we can show that all potential pathways in the CRN can be expressed as compositions
of these indivisible units, then that will give us ground to claim that this formal basis may be thought of as
the target CRN that the given CRN is implementing.
3.2
Basic definitions
The theory of pathway decomposition will be developed with respect to a chosen set F of species called
the formal species; all other species will be intermediate species. All the definitions and theorems below
should be implicitly taken to be with respect to the choice of F. As a convenient convention, we use upper
case and lower case letters to denote formal and intermediate chemical species, respectively.
Definition 1. A state is a multiset of species. If every species in a state S is a formal species, then S is
called a formal state. In this paper we will use + and − to denote multiset sum and multiset difference
respectively, e.g., S + T will denote the sum of two states S and T .
Definition 2. If S is a state, Formal(S) denotes the multiset we obtain by removing all the intermediate
species from S.
Definition 3. A reaction is a pair of multisets of species (R, P ) and it is trivial if R = P . Here, R is called
the set of reactants and P is called the set of products. We say that the reaction (R, P ) can occur in the
state S if R ⊆ S. If both R and P are formal states, then (R, P ) is called a formal reaction. If r = (R, P ),
we will sometimes use the notation r̄ to denote the reverse reaction (P, R).
6
Definition 4. If (R, P ) is a reaction that can occur in the state S, we write S ⊕(R, P ) to denote the resulting
state S − R + P . As an operator, ⊕ is left-associative.
Definition 5. A CRN is a (nonempty) set of nontrivial reactions. A CRN that contains only formal reactions
is called a formal CRN.
Definition 6. A pathway p of a CRN C is a (finite) sequence of reactions (r1 , . . . , rk ) with ri ∈ C for all i.
We say that a pathway can occur in the state S if all its reactions can occur in succession starting from S.
Note that given any pathway, we can find a unique minimal state from which the pathway can occur. We will
call such state the minimal initial state, or simply the initial state of the pathway. Correspondingly, the
final state of a pathway will denote the state S ⊕ r1 ⊕ r2 ⊕ · · · ⊕ rk where S is the (minimal) initial state of
the pathway. If both the initial and final states of a pathway are formal, but not necessarily the intermediate
states, it is called a formal pathway. A pathway is called trivial if its initial state equals its final state. In
this paper, we will write p + q to denote the concatenation of two pathways p and q.
To absorb these definitions, we can briefly study some examples. Consider the chemical reaction 2A +
B → C. According to our definitions, this will be written (⦃A, A, B⦄, ⦃C⦄). Here, ⦃A, A, B⦄ is called
the reactants and ⦃C⦄ is called the products, just as one would expect. Note that this reaction can occur
in the state ⦃A, A, A, B, B⦄ but cannot occur in the state ⦃A, B, C, C, C, C⦄ because the latter state does
not have all the required reactants. If the reaction takes place in the former state, then the resulting state
will be ⦃A, B, C⦄ and thus we can write ⦃A, A, A, B, B⦄ ⊕ (⦃A, A, B⦄, ⦃C⦄) = ⦃A, B, C⦄. In this
paper, although we formally define a reaction to be a pair of multisets, we will interchangeably use the
chemical notation whenever it is more convenient. For instance, we will often write 2A + B → C instead
of (⦃A, A, B⦄, ⦃C⦄).
Note that we say that a pathway p = (r1 , r2 , . . . , rk ) can occur in the state S if r1 can occur in S, r2 can
occur in S ⊕ r1 , r3 can occur in S ⊕ r1 ⊕ r2 , and so on. For example, consider the pathway that consists of
2A + B → C and B + C → A. This pathway cannot occur in the state ⦃A, A, B⦄ because even though the
first reaction can occur in that state, the resulting state after the first reaction, which is ⦃C⦄, will not have
all the reactants required for the second reaction to occur. In contrast, it is easy to see that this pathway can
occur in the state ⦃A, A, B, B⦄, which also happens to be its minimal initial state.
We also point out that we cannot directly express a reversible reaction in this formalism. Thus, a reversible reaction will be expressed using two independent reactions corresponding to each direction, e.g.,
A
B will be expressed as two reactions: A → B and B → A.
Before we proceed, we formally define the notion of tidiness which we informally introduced in Section 1.
Definition 7. Let p be a pathway with a formal initial state and T its final state. Then, a (possibly empty)
pathway p0 = (r1 , . . . , rk ) is said to be a closing pathway of p if p0 can occur in T and T ⊕ r1 ⊕ · · · ⊕ rk
is a formal state. A CRN is weakly tidy if every pathway with a formal initial state has a closing pathway.
As was informally explained before, this means that the given CRN is always capable of cleaning up all
the intermediate species. For example, the CRN {A → i, i + B → C} will not be weakly tidy because if
the system starts from the state ⦃A⦄, it can transition to the state ⦃i⦄ and become “stuck” in a non-formal
state: there does not exist a reaction to convert the intermediate species i back into some formal species.
For a more subtle example, let us consider the CRN {A → i + B, i + B → B}, which is weakly
tidy according to the definition as stated above. In fact, it is easy to see that this implementation CRN will
never get stuck when it is operating by itself, starting with any formal initial state. However, this becomes
problematic when we begin to think about composing different CRNs. Namely, when intermediate species
require other formal species in order to get removed, the implementation CRN may not work correctly if
some other formal reactions are also operating in the system. For instance, if the above implementation runs
7
in an environment that also contains the reaction B → C, then it is no longer true that the system is always
able to get back to a formal state.
This is not ideal because the ability to compose different CRNs, at least in the case where they do not
share any intermediate species, is essential for CRN implementations to be useful. To allow for this type
of composition, and more importantly to allow for the proofs of Theorems in Section 3.4.2 and to make the
algorithm defined in Section 4 tractable, we define a stronger notion of tidiness which is preserved under
such composition.
Definition 8. A closing pathway is strong if its reactions do not consume any formal species. A CRN is
strongly tidy if every pathway with a formal initial state has a strong closing pathway.
In the rest of the paper, unless indicated otherwise, we will simply say tidiness to mean strong tidiness.
Similarly, we will simply say closing pathway to mean strong closing pathway. For some examples of
different levels of tidiness, see Figure 2.
A→i+j
i+j →B
A→i
i+B →C
A→i+D
D→E
E+i→C
strongly tidy
not tidy
weakly tidy
Figure 2: Some examples of tidy and non-tidy CRNs
3.3
Pathway decomposition
Now we formally define the notion of pathway decomposition. Following our intuition from Section 3.1,
we first define what it means to implement a formal reaction.
Definition 9. Consider a pathway p = (r1 , . . . , rk ) and let Si = S ⊕ r1 ⊕ · · · ⊕ ri , so that S0 , S1 , . . . , Sk are
all the states that p goes through. Then, p is regular if there exists a turning point reaction rj = (R0 , P 0 )
such that Formal(Si ) ⊆ S for all i < j, Formal(Si ) ⊆ T for all i ≥ j, and Formal(Sj−1 − R0 ) = ∅.
Definition 10. We say that a pathway p = (r1 , . . . , rk ) implements a formal reaction (R, P ) if it is regular
and R and P are equal to the initial and final states of p, respectively.
While the first condition is self-evident, the second condition needs a careful explanation. It asserts that
there should be a point in the pathway prior to which we only see the formal species from the initial state
and after which we only see the formal species from the final state. The existence of such a “turning point”
allows us to interpret the pathway as an implementation of the formal reaction (R, P ) where in a sense the
real transition is occurring at that turning point. Importantly, this condition rules out such counterintuitive
implementations as (A → i, i → C+j, C+j → k, k → B) or (A → i+B, i+B → j+A, j+A → B) as
implementations of A → B. Note that a formal pathway that consumes but does not produce formal species
prior to its turning point, and thereafter produces but does not consume formal species, is by this definition
regular, and this is the “typical case.” However our definition also allows additional flexibility; for example,
the reactants can fleetingly bind, as B does in the second and third reactions of (A → i, i + B → j, j →
B + i, i + B → C), whose turning point is unambiguously the last reaction. One may also wonder why
we need the condition Formal(Sj−1 − R0 ) = ∅. This is to prevent ambiguity that may arise in the case of
catalytic reactions. Consider the pathway (A → i + A, i → B). Without the above condition, both reactions
8
in this pathway qualify as a turning point, but the second reaction being interpreted as the turning point is
counterintuitive because the product A gets produced before the turning point.
One problem of the above definition is that it interprets the pathway (A → i, A → i, i → B, i → B)
as implementing A + A → B + B. As explained in Section 3.1, we would like to be able to identify such a
pathway as a composition of smaller units.
Definition 11. We say that a pathway p can be partitioned into two pathways p1 and p2 if p1 and p2 are
subsequences of p (which need not be contiguous, but must preserve order) and every reaction in p belongs
to exactly one of p1 and p2 . Equivalently, we can say p is formed by interleaving p1 and p2 .
Definition 12. A formal pathway p is decomposable if p can be partitioned into p1 and p2 that are each
formal pathways. A nonempty formal pathway that is not decomposable is called prime.
For example, consider the formal pathway p = (A → i, B → i, i → C, i → D, D → j, j → E).
This pathway is not prime because it can be decomposed into two formal pathways p1 = (A → i, i → C)
and p2 = (B → i, i → D, D → j, j → E). Note that within each of the two subsequences, reactions must
appear in the same order as in the original pathway p. In this example, p1 is already a prime pathway after the
first decomposition, whereas p2 can be further decomposed into (B → i, i → D) and (D → j, j → E). In
this manner, any nonempty formal pathway can eventually be decomposed into one or more prime pathways.
Note that such a decomposition may not be unique, e.g., p can also be decomposed into (A → i, i → D),
(B → i, i → C), and (D → j, j → E).
Definition 13. The set of prime pathways in a given CRN is called the elementary basis of the CRN. The
formal basis is the set of (initial state, final state) pairs of the pathways in the elementary basis.
Note that the elementary basis and/or the formal basis can be either finite or infinite. The elementary
basis may contain trivial pathways, and the formal basis may contain trivial reactions.
Definition 14. A CRN is regular if every prime pathway implements some formal reaction (in particular, it
must have a well-defined turning point reaction as defined in Definition 9). Equivalently, a CRN is regular
if every prime pathway is regular.
Definition 15. Two tidy and regular CRNs are said to be pathway decomposition equivalent if their formal
bases are identical, up to addition or removal of trivial reactions.
For clarity, we remind the reader here that the definitions and theorems in this section are implicitly
taken to be with respect to the choice of F. In particular, this means that each choice of F gives rise to a
j
A
i
D
m
F
A
i
j
i
k
B
j
l
E
C
k
j
B
A
i
k
B
D
C
{A + B → D + E, D → F}
{A → B, C → D}
E
B
D
{A → E, B + C → B + D}
Figure 3: Three examples of decomposable formal pathways and the formal bases of their corresponding
CRNs. The partition of reactions is marked by lines of different types and colors. In the right most example,
the decomposed pathway denoted by blue lines with circles (which shows up as A → E in the formal basis)
is not regular, and therefore pathway decomposition equivalence does not apply.
9
different pathway decomposition equivalence relation. For instance, CRNs {A → i, i → C} and {A → C}
are clearly pathway decomposition equivalent with respect to the conventional choice of F, which contains
exactly those species named with upper case letters, but if e.g. we defined F0 = F ∪ {i}, these two CRNs
would not be pathway decomposition equivalent with respect to F0 .
3.4
3.4.1
Theorems
Properties
It is almost immediate that pathway decomposition equivalence satisfies many nice properties, some of
which are expressed in the following theorems.
Theorem 3.1. For any fixed choice of F, pathway decomposition equivalence with respect to F is an equivalence relation, i.e., it satisfies the reflexive, symmetric, and transitive properties.
Theorem 3.2. If C is a formal CRN, its formal basis is itself.
Corollary 3.3. If C1 and C2 are formal CRNs, they are pathway decomposition equivalent if and only if
C1 = C2 , up to removal or addition of trivial reactions.
Theorem 3.4. Any formal pathway of C can be generated by interleaving one or more prime pathways of C.
It is perhaps worth noting here that the decomposition of a formal pathway may not always be unique.
For example, the pathway (A → i, B → i, i → C, i → D) can be decomposed in two different ways:
(A → i, i → C) and (B → i, i → D), and (A → i, i → D) and (B → i, i → C). Pathway
decomposition differs from other notions such as (weak) bisimulation or (weak) trace equivalence in that
it allows such degeneracy of interpretations. We note that such degeneracy, which is closely related to the
previously mentioned delayed choice phenomenon, may permit a more efficient implementation of a target
CRN in terms of the number of species or reactions used in the implementation CRN. For example, if we
wish to implement the formal CRN consisting of the twelve reactions A
B, A
C, A
D, B
C,
B
D and C
D, it may be more efficient to implement it as the following eight reactions: A
i,
B
i, C
i and D
i.
The following theorems illuminate the relationship between a tidy and regular CRN C and its formal
basis F and how to better understand this degeneracy of interpretations.
Definition 16. Let C be a tidy and regular CRN and F its formal basis. Suppose p = (r1 , . . . , rk ) is a
formal pathway in C (i.e. ri ∈ C for all i) and q = (s1 , . . . , sl ) is a formal pathway in F (i.e. si ∈ F for all
i). Then, we say p can be interpreted as q if
1. q can occur in the initial state S of p,
2. S ⊕ r1 ⊕ · · · ⊕ rk = S ⊕ s1 ⊕ · · · ⊕ sl , and
3. there is a decomposition of p such that if we replace a selected turning point reaction of each prime
pathway with the corresponding reaction of F and remove all other reactions, the result is q.
It is clear that the interpretation may not be unique, because there can be many different decompositions
of p as well as many different choices of the turning point reactions. For example, consider the pathway
p = (A → i, B → j, i → C, j → D), which has a unique decomposition into pathways (A → i, i → C)
and (B → j, j → D). In each of these constituent pathways, there are two ways to select a turning point
reaction. If we picked A → i and j → D, the process in condition 3 would yield (A → C, B → D) as the
interpretation of p. On the other hand, if we selected i → C and B → j as our turning point reactions, p
would end up being interpreted as (B → D, A → C).
10
One might also wonder why we do not simply require that p and q must have the same initial states. This
is because of a subtlety in the concept of the minimal initial state, which arises due to a potential parallelism
in the implementation. For instance, consider the pathway (A → i, B → A, i → B). This pathway, which
can be interpreted as two formal reactions A → B and B → A occuring in parallel, has initial state ⦃A, B⦄.
However, no such parallelism is allowed in the formal CRN and thus this pathway is forced to correspond
to either (A → B, B → A) or (B → A, A → B), neither of which has initial state ⦃A, B⦄.
Theorem 3.5. Suppose C is a tidy and regular CRN and F is its formal basis.
1. For any formal pathway q in F, there exists a formal pathway p in C whose initial and final states are
equal to those of q, such that p can be interpreted as q.
2. Any formal pathway p in C can be interpreted as some pathway q in F.
Proof. 1. Replace each reaction in q with the corresponding prime pathway of C.
2. Fix a decomposition of p and pick a turning point for each prime pathway. Replace the turning points
with the corresponding formal basis reaction and remove all other reactions. We call the resulting
pathway q. Then it suffices to show that q can occur in the initial state S of p. We show this by a
hybrid argument.
Define pj to be the pathway obtained by replacing the first j turning points in p by the corresponding
formal basis elements and removing all other reactions that belong to those prime pathways. In particular, note that p0 = p and pl = q. We show that pj can occur in the initial state of pj−1 for all j > 0.
First, write pj−1 = (r1 , . . . , rm ) and pj = (ri1 , . . . , rik , sj , rik+1 , . . . , rin ). Then it follows from the
definition of a turning point that Formal(S ⊕ ri1 ⊕ · · · ⊕ rix−1 ) ⊇ Formal(S ⊕ r1 ⊕ · · · ⊕ rix −1 ) for
every 1 ≤ x ≤ k. Therefore (ri1 , . . . , rik ) can occur in S. (Note that we need not worry about the
intermediate species because (ri1 , . . . , rik ) has a formal initial state.) Moreover, since the definition of
a turning point asserts that all the reactants must be consumed at the turning point, it also implies that
Formal(S ⊕ ri1 ⊕ · · · ⊕ rik ) ⊇ Formal(S ⊕ r1 ⊕ · · · ⊕ rt−1 − X + R) where rt = (X, Y ) denotes the
turning point that is being replaced by sj in this round and R denotes the reactants of sj . Therefore,
sj can occur in S ⊕ ri1 ⊕ · · · ⊕ rik . Finally, it again follows from the definition of a turning point that
Formal(S ⊕ ri1 ⊕ · · · ⊕ rik ⊕ sj ⊕ rik+1 ⊕ · · · ⊕ rix−1 ) ⊇ Formal(S ⊕ r1 ⊕ · · · ⊕ rix −1 ) for every
k + 1 ≤ x ≤ n. We conclude that (ri1 , . . . , rik , sj , rik+1 , . . . , rin ) = p0j can occur in S.
It is interesting to observe that tidiness is not actually used in the proof of Theorem 3.5 above (nor in
that of Theorem 3.6 below), so that condition could be removed from the theorem statement. We retain the
tidiness condition to emphasize that this is when the theorem characterizes the behavior of the CRN; without
tidiness, a CRN could have many relevant behaviors that take place along pathways that never return to a
formal state, and these behaviors would not be represented in its formal basis.
In our final theorem we prove that pathway decomposition equivalence implies formal state reachability
equivalence. Note that the converse is not true because {A → B, B → C, C → A} is not pathway
decomposition equivalent to {A → C, C → B, B → A}.
Theorem 3.6. If two tidy and regular CRNs C1 and C2 are pathway decomposition equivalent, they give rise
to the same reachability between formal states.
Proof. Suppose formal state T is reachable from formal state S in C1 , i.e. there is a formal pathway p in
C1 whose initial state is S and final state is T . By Theorem 3.5, it can be interpreted as some pathway q
consisting of the reactions in the formal basis of C1 . Since C1 and C2 have the same formal basis, by another
application of Theorem 3.5, there exists some formal pathway p0 in C2 that can be interpreted as q. That is,
the initial and final states of p0 are S and T respectively, which implies that T is reachable from S in C2 also.
By symmetry between C1 and C2 , the theorem follows.
11
3.4.2
Modular composition of CRNs
As we briefly mentioned in Section 3.2, it is very important for the usefulness of a CRN implementation
that it be able to be safely composed with other CRNs. For instance, consider the simplest experimental
setup of putting the molecules of the implementation CRN in the test tube and measuring the concentration
of each species over time. In practice, the concentration measurement of species A is typically carried out
by implementing a catalytic reaction that uses A to produce fluorescent material. Therefore even this simple
scenario already involves a composition of two CRNs, namely the implementation CRN itself and the CRN
consisting of the measurement reactions. It is evident that the ability to compose CRNs would become even
more essential in more advanced applications.
In this section, we prove theorems that show that pathway decomposition equivalence is preserved under
composition of CRNs, as long as those CRNs do not share any intermediate species.
Theorem 3.7. Let C and C 0 be two CRNs that do not share any intermediate species. Then, C ∪ C 0 is tidy if
and only if both C and C 0 are tidy.
Proof. For the forward direction, by symmetry it suffices to show that C is tidy. We begin by proving the
following lemma.
Lemma 3.8. Let C and C 0 be two CRNs that do not share any intermediate species. Let p be any formal
pathway in C ∪ C 0 . If we partition p into two pathways p1 and p2 such that p1 is a pathway of C and p2 is a
pathway of C 0 , then each of p1 and p2 is formal.
Proof. Since C and C 0 do not share any intermediate species, it follows that all the intermediate species in
the initial state of p1 will also show up in the initial state of p and all the intermediate species in the final
state of p1 will also show up in the final state of p. Hence p1 must be formal. The case for p2 follows by
symmetry.
Now let p be any pathway in C with a formal initial state. Since p is also a pathway in C ∪ C 0 , it has a
closing pathway q in C ∪ C 0 . Since s = p + q is a formal pathway, we can partition it into s1 and s2 as in the
above lemma. In particular, since all the reactions in p belong to C, we have s1 = p + q1 and s2 = q2 where
q1 and q2 are a partition of q such that q1 is a pathway of C and q2 is a pathway of C 0 . Since s1 is formal by
the lemma, q1 is a closing pathway of p. Hence, C is tidy.
For the reverse direction, suppose p is a pathway of C ∪ C 0 that has a formal initial state. Since C and
0
C do not share intermediate species, we can partition the intermediate species found in the final state of p
into two multisets A and A0 , corresponding to the intermediate species used by C and C 0 respectively. Now,
if we remove from p all the reactions that belong to C 0 and call the resulting pathway q, then the multiset
of all the intermediate species found in the final state of q will be exactly A. This is because the removed
reactions, which belonged to C 0 , cannot consume or produce any intermediate species used by C. Since C
is tidy, q has a closing pathway r. This time, remove from p all the reactions that belong to C and call the
resulting pathway q 0 . By a symmetric argument, q 0 must have a closing pathway r0 . Now observe that r + r0
is a closing pathway for p.
Theorem 3.9. Let C and C 0 be two CRNs that do not share any intermediate species. Then, C ∪ C 0 is regular
if and only if both C and C 0 are regular.
Proof. For the forward direction, simply observe that any prime pathway p of C is also a prime pathway of
C ∪ C 0 and therefore must be regular. Hence, C is regular. By symmetry, C 0 is also regular.
For the reverse direction, let p be a prime pathway in C ∪ C 0 . Partition p into two subsequences q and q 0 ,
which contains all reactions of p which came from C and C 0 respectively. Since the two CRNs do not share
any intermediate species, it is clear that q and q 0 must both be formal. Since p was prime, it implies that one
12
of q and q 0 must be empty. Therefore, p is indeed a prime pathway in either C or C 0 , and since each was a
regular CRN, p must be regular.
Theorem 3.10. Let C and C 0 be two tidy and regular CRNs that do not share any intermediate species, and
F and F 0 their formal bases respectively. Then the formal basis of C ∪ C 0 is exactly F ∪ F 0 .
Proof. Let p be a prime pathway in C ∪ C 0 . By the same argument as in the proof of Theorem 3.9, p is a
prime pathway of either C or C 0 . Therefore, the formal basis of C ∪ C 0 is a subset of F ∪ F 0 . The other
direction is trivial.
We note that the ability to compose CRNs has another interesting consequence. Frequently, molecular
implementations of CRNs involve intermediate species that are specific to a pathway implementing a particular reaction, such that intermediates that belong to pathways that implement different reactions do not react
with each other. This is a strong constraint on the architecture of the implementations that can facilitate their
verification. (For instance, it has been observed and used by Lakin et al. in [25].) We observe that in such
cases Theorem 3.10 provides an easier way to find the formal basis of the implementation CRN. Namely,
we can partition the CRN into disjoint subsets that do not share intermediate species with one another, find
the formal basis of each subset, and then take the union of the found formal bases. For example, if the implementation CRN was {A → i, i → B, A → j, j → C, j + C → k, k → D}, then it can be partitioned
into CRNs {A → i, i → B} and {A → j, j → C, j + C → k, k → D} such that they do not share
intermediate species with each other. It is straightforward to see that the formal bases of these two subsets
are {A → B} and {A → C, A + C → D} respectively, so the formal basis of the whole implementation
CRN must be {A → B, A → C, A + C → D}. Similarly, Theorems 3.7 and 3.9 ensure that we can
test for tidiness and regularity of the implementation CRN by testing tidiness and regularity of each of these
subsets.
4
Algorithm
In this section, we present a simple algorithm for finding the formal basis of a given CRN. The algorithm
can also test tidiness and regularity.
Our algorithm works by enumerating pathways that have formal initial states. The running time of our
algorithm depends on a quantity called maximum width, which can be thought of as the size of the largest
state that a prime pathway can ever generate. Unfortunately it is easy to see that this quantity is generally
unbounded; e.g., {A → i, i → i + i, i → ∅} has a finite formal basis {A → ∅} but it can generate arbitrarily
large states.3 However, since such implementations are highly unlikely to arise in practice, in this paper we
focus on the bounded width case. We note that even in the bounded width case it is still nontrivial to come
up with an algorithm that finishes in finite time, because it is unclear at what width we can safely stop the
enumeration.
4.1
Exploiting bounded width
We begin by introducing a few more definitions and theorems.
Definition 17. A pathway that has a formal initial state is called semiformal.
Definition 18. A semiformal pathway p is decomposable if p can be partitioned into two nonempty subsequences (which need not be contiguous) that are each semiformal pathways.
It is obvious that this reduces to our previous definition of decomposability if p is a formal pathway.
3
Clearly, there may also be cases where the formal basis itself is infinite, e.g. {A → i, i → i + i, i → B}.
13
Definition 19. Let p = (r1 , . . . , rk ) be a pathway and let Si = S ⊕ r1 ⊕ · · · ⊕ ri where S is the initial state
of p. The width of p is defined to be maxi |Si |.
Definition 20. The branching factor of a CRN C is defined to be the following value.
max max{|R|, |P |}
(R,P )∈C
We note that many implementations that arise in practice have small branching factors (e.g. [36, 5, 31]).
Theorem 4.1. Suppose that pathway p is obtained by interleaving pathways p1 , . . . , pl . Let S be the initial
state of p and S1 , . . . , Sl the initial states of p1 , . . . , pl respectively. Then, S ⊆ S1 + S2 + · · · + Sl .
Theorem 4.2. If p is an undecomposable semiformal pathway of width w > 0, there exists an undecomposable semiformal pathway of width smaller than w but at least (w − b)/b, where b is the branching factor of
the CRN. (Note that if w is small, the lower bound (w − b)/b might be negative. In this case, it would simply mean that there exists an undecomposable semiformal pathway of width 0, which would be the empty
pathway.)
Proof. Since w > 0, p is nonempty. Let p−1 denote the pathway obtained by removing the last reaction
0
the states that p−1
(R, P ) from p. Also, let S0 , . . . , Sk be the states that p goes through, and S00 , . . . , Sk−1
0
goes through. Si is potentially unequal to Si because if the last reaction in p consumes some new formal
species, then the minimal initial state of p−1 might be smaller than that of p.
It is obvious that the minimal initial state of p−1 is smaller than the minimal initial state of p by at most
|R|, i.e., |S0 | − |S00 | ≤ |R|. This means that for all 0 ≤ i ≤ k − 1, we have that |Si | − |Si0 | ≤ |R|. Clearly, if
there exists some 0 ≤ i ≤ k − 1 such that |Si | = w, then |Si0 | ≥ |Si | − |R| = w − |R| ≥ w − b, so p−1 has
width at least w − b. If there exists no such i, then we have that |Sk | = w. Clearly, |Sk−1 | = |Sk | − |P | + |R|
and it follows that
0
0
| ≤ |R|.
| = |Sk−1 | − |Sk−1
|Sk | − |P | + |R| − |Sk−1
0
0
| ≥ w − |P | ≥ w − b. Thus,
| ≤ |P |. Since |Sk | = w, we have that |Sk−1
This is equivalent to |Sk | − |Sk−1
p−1 achieves width at least w − b.
Then, we decompose p−1 until it is no longer decomposable. As a result, we will end up with l ≥ 1
undecomposable pathways p1 , p2 , . . . , pl which by interleaving can generate p−1 . Also, they are all semiformal. First, we show that l is at most b. Assume towards a contradiction that l > b. Then, by the
pigeonhole principle, there exists i such that (R − Formal(R), P ) can occur in the sum of the final states of
p1 , . . . , pi−1 , pi+1 , . . . , pl (since |R − Formal(R)| ≤ b and (R − Formal(R), P ) can occur in the sum
of the final states of p1 , . . . , pl , the at most b reactants of (R − Formal(R), P ) are distributed among
l > b pathways and there exists at least one pi that does not provide a reactant and can be omitted).
Then, consider the decomposition (pi , p0i ) of p−1 where p0i denotes the pathway we obtain by interleaving p1 , . . . , pi−1 , pi+1 , . . . , pl in the same order that those reactions occur in p−1 . By Theorem 4.1, p0i is
semiformal. Since pj ’s are all semiformal, this means that the intermediate species in the final state of p0i
will be exactly the same as those in the sum of the final states of p1 , . . . , pi−1 , pi+1 , . . . , pl . That is, the
final state of p0i contains all the intermediate species that (R, P ) needs to occur, i.e., p0i with (R, P ) appended at the end should have a formal initial state. However, this means that p is decomposable which is a
contradiction. Hence, l ≤ b.
Now, note that if we have P
l pathways each with widths w1 , . . . , wl , any pathway obtained by interleaving
P
them can have width at most li=1 wi . Since p−1 had width at least w − b, we have that w − b ≤ li=1 wi .
P
Then, if wi < (w − b)/b for all i, then li=1 wi < w − b, which is contradiction. Thus, we conclude that
at least one of p1 , . . . , pl has width greater than or equal to (w − b)/b. It is also clear that its width cannot
exceed w. Thus, we have found a pathway p0 which
14
1. has a smaller length than p, and
2. has width at least (w − b)/b and at most w.
If p0 has width exactly w, then we have failed to meet the requirements of the claim. However, since we have
decreased the length of the pathway by at least one, and the width of a zero-length pathway is always 0, we
can eventually get a smaller width than w by repeating this argument. The first time that the width decreases,
we will have found a pathway p0 that satisfies the theorem statement, because in that case conditions (1) and
(2) must hold by the arguments above.
Corollary 4.3. Suppose w and wmax are integers such that (w + 1)b ≤ wmax . Then, if there is no undecomposable semiformal pathway of width greater than w and less than or equal to wmax , then there exists no
undecomposable semiformal pathway of width greater than w.
Proof. Assume towards a contradiction that there exists an undecomposable semiformal pathway p of width
w0 > w. If w0 ≤ wmax , then it is an immediate contradiction. Thus, assume that w0 > wmax . By Theorem
4.2, we can find a smaller undecomposable semiformal pathway q of width v where (w0 − b)/b ≤ v < w0 .
Since w0 > wmax , we have that v ≥ (w0 − b)/b > (wmax − b)/b ≥ ((w + 1)b − b)/b = w. If v ≤ wmax ,
we have a contradiction. If v > wmax , then take q as our new p and repeat the above argument. Since v is
smaller than w0 by at least one, we will eventually reach a contradiction.
Thus, there exists no undecomposable semiformal pathway of width greater than w.
4.2
Overview
While Corollary 4.3 gives us a way to exploit the bounded width assumption, it is still unclear whether the
enumeration can be made finite, because the number of undecomposable semiformal pathways of bounded
width may still be infinite. For an easy example, if the CRN consists of {A → i, i → j, j → i, j → B},
we have infinitely many undecomposable semiformal pathways of width 1, because after the initial reaction
A → i, the segment i → j, j → i can be repeated arbitrarily many times without ever making the pathway
decomposable. In this section, we sketch at high level how this difficulty is resolved in our finite-time
algorithm.
The principal technique that lets us avoid infinite enumeration of pathways is memoization. To use
memoization, we first define what is called the signature of a pathway, which is a collection of information
about many important properties of the pathway, such as its initial and final states, decomposability, etc.
It turns out that the number of possible signatures of bounded width pathways is always finite, even if the
number of pathways themselves may be infinite. This means that the enumeration algorithm does not need
to duplicate pathways with the same signatures, provided the signatures alone give us sufficient information
for determining the formal basis and for testing tidiness and regularity of the CRN.
Therefore, the algorithm consists in enumerating all semiformal pathways of width up to (w + 1)b,
where w is the maximum width of the undecomposable semiformal pathways discovered so far, while excluding pathways that have the same signatures as previously discovered pathways. It is important to emphasize that no a priori knowledge of the width bound is assumed, and the algorithm is guaranteed to halt as
long as there exists some finite bound. While the existence of this algorithm shows that the problem of finding the formal basis is decidable with the bounded width assumption, the worst-case time complexity seems
to be adverse as is usual for algorithms based on exhaustive search. It is an open question to understand
the computational complexity of this problem as well as to find an algorithm that has better practical performance. Another important open question is whether the problem without the bounded width assumption is
decidable.
15
4.3
Signature of a pathway
While Corollary 4.3 gives us a way to make use of the bounded width assumption, it is still unclear
whether the enumeration can be made finite, because the number of undecomposable semiformal pathways
of bounded width may still be infinite. To resolve this problem, we need to define a few more concepts.
Definition 21. Let p be a semiformal pathway. The decomposed final states (DFS) of p is defined as the
set of all unordered pairs (T1 , T2 ) that can be obtained by decomposing p into two semiformal pathways and
taking their final states. Note that for an undecomposable pathway, the DFS is the empty set.
Definition 22. Let p = (r1 , . . . , rk ) be a semiformal pathway. Also, let Si = S ⊕ r1 ⊕ · · · ⊕ ri where
S is the initial state of p. The formal closure of p is defined as the unique minimal state S 0 such that
Formal(Si ) ⊆ S 0 for all i.
Definition 23. The regular final states (RFS) of p is defined as the set of all minimal states T such that
there exists a potential turning point reaction rj = (R, P ) which satisfies Formal(Si ) ⊆ S for all i < j,
Formal(Si ) ⊆ T for all i ≥ j, and Formal(Sj−1 − R) = ∅.
Some explanation is in order. Although the RFS definition applies equally to semiformal pathways that
are or could be regular, and to semiformal pathways that are not and cannot be regular, the RFS provides
a notion of “what the final state would/could be if the pathway were regular”. As examples, first consider
the semiformal pathway (A → i, B + i → j, j → X + k). The second and third reactions are potential
turning points, and the RFS is {⦃X⦄}. One can easily check that if the pathway were to be completed in a
way that its final state does not contain X, the resulting pathway cannot be regular (e.g. were it to be closed
by X + k → Y , the pathway becomes irregular). Now consider (A → i, i → B + j, B + j → k). Only
the first two reactions are potential turning points, and the RFS is {⦃B⦄}. One can also check in this case
that the only way that this semiformal pathway can be completed as a regular pathway is for it to have B in
its final state. Finally consider (A → i, i → A + j, A + j → B), which is in fact a regular formal pathway
implementing A → B. Because every reaction is a potential turning point by our definition, the RFS is
{⦃A, B⦄, ⦃B⦄}. One of these states is the actual final state, corresponding to the actual turning point, and
therefore we can see that this formal pathway is regular.
Definition 24. The signature of the pathway is defined to be the 6-tuple of the initial state, final state, width,
formal closure, DFS, and RFS.
Theorem 4.4. If m is any finite number, the set of signatures of all semiformal pathways of width up to m
is finite.
Proof. Clearly, there is only a finite number of possible initial states, final states, widths, formal closures,
and RFS. Also, since there is only a finite number of possible final states, there is only a finite number of
possibilities for DFS.
Theorem 4.5. Suppose p1 and p2 are two pathways with the same signature. Then, for any reaction r,
p1 + (r) and p2 + (r) also have the same signature.
Proof. Let p01 = p1 + (r) and p02 = p2 + (r). It is trivial that p01 and p02 have the same initial and final states,
formal closure, and width.
First, we show that p01 and p02 have the same DFS. Suppose (T1 , T2 ) is in the DFS of p01 . That is, there
exists a decomposition (q10 , q20 ) of p01 where q10 and q20 have final states T1 and T2 . The last reaction r is
either contained in q10 or q20 . Without loss of generality, suppose the latter is the case. Then, if q1 = q10 and
q2 + (r) = q2 , then (q1 , q2 ) should decompose p1 , which is a prefix of p01 . Since p1 and p2 have the same
DFS, there should be a decomposition (s1 , s2 ) of p2 that has the same final states as q1 and q2 . Clearly,
16
(s1 , s2 + (r)) should be a decomposition of p02 and thus (T1 , T2 ) is also in the DFS of p02 . By symmetry, it
follows that p01 and p02 have the same DFS.
Now we argue that p01 and p02 should have the same RFS. Suppose T is contained in the RFS of p01 .
1. If the potential turning point for T in p01 is the last reaction r, with r = (R, P ), then it must be the
case that T = Formal(P ). Because p2 has the same formal closure and final state as p1 , which was
sufficient to ensure that r was a valid potential turning point in p01 , r will also be a valid potential
turning point in p02 . Consequently, T is also in the RFS of p02 .
2. Otherwise, the potential turning point reaction for T in p01 , call it t = (R, P ), also appears in p1 . Since
the initial state of p1 must be a subset of the initial state of p01 , t is also a potential turning point for
p1 . Since “midway through” the potential turning point reaction, all formal species must be gone, we
conclude that in fact p1 and p01 have the same initial state. That is, R contains no formal species that
aren’t already in the final state of p1 . Thus, all shared states after t are the same, and some subset T 0
of T must be contained in the RFS of p1 . By assumption, T 0 is also in the RFS of p2 , and p2 has the
same final state as p1 . Since R contains no formal species that aren’t already in the final state of p2 , the
initial states of p2 and p02 are the same. Consequently, the potential turning point of p2 corresponding
to T 0 is also a potential turning point for p02 . This ensures that T is in the RFS for p02 .
Theorem 4.6. A nonempty pathway p is a prime pathway if and only if its signature satisfies the following
conditions:
1. The initial and final states are formal.
2. The DFS is the empty set.
4.4
Algorithm for enumerating signatures
It is now clear that we can find the formal basis by enumerating the signatures of all undecomposable
semiformal pathways. In this section we present a simple algorithm for achieving this.
function enumerate(p, w, ret)
if p is not semiformal or has width greater than w then return ret
sig = signature of p
if sig is in ret then return ret
add sig to ret
for every reaction rxn
ret = enumerate(p + [rxn], w, ret)
end for
return ret
end function
function main()
w_max = 0
b = branching factor of the given CRN
while true
signatures = enumerate([], w_max, {})
w = maximum width of an undecomposable pathway in signatures
if (w+1)*b <= w_max then break
w_max = (w+1)*b
17
end while
return signatures
end function
The subroutine enumerate is a function that enumerates the signatures of all semiformal pathways of
width at most w. Note that it uses memoization to avoid duplicating pathways that have identical signatures,
as justified by Theorem 4.5. Because of this memoization, Theorem 4.4 ensures that this subroutine will
terminate in finite time.
The subroutine main repeatedly calls enumerate, increasing the width bound according to Corollary
4.3. It is obvious that main will terminate in finite time if and only if there exists a bound to the width of
an undecomposable semiformal pathway.
It is out of scope of this paper to attempt theoretical performance analysis of this algorithm or to study
the computational complexity of finding the formal basis. While there are obvious further optimizations
by which the performance of the above algorithm can be improved, we meet our goal of this paper in
demonstrating the existence of a finite time algorithm and leave further explorations as a future task.
4.5
Testing tidiness and regularity
Finally, we discuss how to use the enumerated signatures to test tidiness and regularity of the given CRN.
Theorem 4.7. A CRN is tidy if and only if every undecomposable semiformal pathway has a closing pathway.
Proof. The forward direction is trivial. For the reverse direction, we show that if a CRN is not tidy, there
exists an undecomposable semiformal pathway that does not have a closing pathway.
By definition, there exists a semiformal pathway p that does not have a closing pathway. Consider a
minimal-length example of such a pathway. If p is undecomposable, then we are done. So suppose that p
is decomposable into two semiformal pathways p1 and p2 . By the minimality of p, both pathways p1 and
p2 must have closing pathways. However, since the final state of p has the same intermediate species as the
sum of the final states of p1 and p2 (by Theorem 4.1 and the fact that p1 and p2 are semiformal), the two
closing pathways concatenated will be a closing pathway of p (because a closing pathway does not consume
any formal species). This contradicts that p does not have a closing pathway, and thus we conclude that the
case where p is decomposable is impossible.
Theorem 4.8. Let p be an undecomposable semiformal pathway that has a closing pathway. Then p also
has a closing pathway q such that p + q is undecomposable.
Proof. Let q be a minimal-length closing pathway for p. Note that p + q is a formal pathway. If p + q is
undecomposable, we are done. So suppose that p + q decomposes into two formal pathways p1 and p2 ,
which by definition must both be nonempty. Then it must be the case that one of p1 or p2 contains all the
reactions of p, because otherwise p must be decomposable as well. Without loss of generality, suppose p1
contains all the reactions of p. Then p2 consists only of reactions from q. This means that the reactions
of q that went into p1 constitute a shorter closing pathway q for p, contradicting the minimality of q. We
conclude that p + q must have been undecomposable.
To test tidiness, we attempt to find a closing pathway for each undecomposable semiformal pathway p
enumerated by the main algorithm. Theorem 4.7 ensures that it suffices to consider only these pathways.
We do this by enumerating the signatures of all semiformal pathways of the form p + q where q is a pathway
that does not consume a formal species, but only those of width up to wmax (wmax is the maximum width of
18
the undecomposable semiformal pathways discovered by the main algorithm). Theorem 4.8 ensures that it
is safe to enforce this width bound.
The testing of regularity is trivial, using the following theorem.
Theorem 4.9. A prime pathway is regular if and only if its RFS contains its final state.
Proof. The potential turning point corresponding to the final state proves regularity. If the final state is
lacking in the RFS, then none of the potential turning points qualify as a turning point and the pathway is
not regular.
We emphasize that these methods work only because of the bounded width assumption we made on
undecomposable semiformal pathways. Without this assumption, it is unclear whether these problems still
remain decidable.
4.6
Optimization techniques
In this section, we discuss some optimization techniques that can be used to improve the performance of the
main enumeration algorithm. While we provide no theoretical analysis of these techniques, we report that
there are test instances on which these techniques speed up the algorithm by many orders of magnitude.
Definition 25. If S is a state, Intermediate(S) denotes the multiset that consists of exactly all the intermediate species in S.
Theorem 4.10. If p is an undecomposable semiformal pathway of CRN C with an initial state of size m > 0,
there exists an undecomposable semiformal pathway of C with an initial state of size smaller than m but at
least min(R,P )∈C {(m − |Formal(R)|)/|Intermediate(R)|}.
Proof. Since m > 0, p is nonempty. Let p−1 denote the pathway obtained by removing the last reaction
(R, P ) from p. Let x be the number of formal species in R. Also, let S and S−1 denote the initial states of
p and p−1 respectively.
It is obvious that the initial state of p−1 is smaller than the initial state of p by at most x, i.e., |S−1 | ≥
|S| − x. Then, we decompose p−1 until it is no longer decomposable. As a result, we will end up with l
undecomposable pathways p1 , p2 , . . . , pl which by interleaving can generate p−1 . Also, they are all semiformal. First, we show that l is at most y, where y is the number of intermediate species in R (clearly,
x + y = |R|). Assume towards a contradiction that l > y. Then, by the pigeonhole principle, there exists
i such that (Intermediate(R), P ) can occur in the sum of the final states of p1 , . . . , pi−1 , pi+1 , . . . , pl (as in
the proof of Theorem 4.2). Then, consider the decomposition (pi , p0i ) of p−1 where p0i denotes the pathway
we obtain by interleaving p1 , . . . , pi−1 , pi+1 , . . . , pl in the same order that those reactions occur in p−1 . By
Theorem 4.1, p0i is semiformal. Since pj ’s are all semiformal, this means that the intermediate species in the
final state of p0i will be exactly the same as those in the sum of the final state of p1 , . . . , pi−1 , pi+1 , . . . , pl .
That is, the final state of p0i contains all the intermediate species that (Intermediate(R), P ) needs to occur,
i.e., p0i with (R, P ) appended at the end should have a formal initial state. However, this means that p is
decomposable which is a contradiction. Hence, l ≤ y.
Now, note that if we have l semiformal pathways whose initial states P
have size m1 , . . . , ml , any pathway
obtained by interleaving them can have an initial state of size at most li=1 mi . Since p−1 had an initial
state of size at least m − x and l ≤ y, we can conclude that at least one of p1 , . . . , pl has an initial state of
size at least (m − x)/y. It is also clear that the size of its initial state cannot exceed m. Thus, we have found
a pathway p0 which
1. has a smaller length than p, and
2. has an initial state of size at least min(R,P )∈C {(m − |Formal(R)|)/|Intermediate(R)|} and at most m.
19
If p0 has an initial state of size exactly m, then we have failed to meet the requirements of the claim.
However, since we have decreased the length of the pathway by at least one and the initial state of a zerolength pathway is of size 0, we can eventually get an initial state of size smaller than m by repeating this
process. The first time that the size of the initial state decreases, we will have found a pathway p0 that satisfies
the theorem statement because in that case conditions (1) and (2) must hold by the arguments above.
The above theorem allows us to maintain a bound i max on the size of initial states during enumeration,
in a similar manner to how w max is maintained. Since our enumeration algorithm is essentially brute-force,
imposing this additional bound may significantly reduce the number of pathways that need to be enumerated.
Moreover, the proof of the above theorem immediately lets us optimize the constants in Theorem 4.2.
Theorem 4.11. If p is an undecomposable semiformal pathway of width w > 0, there exists an undecomposable semiformal pathway of width smaller than w but at least (w − b)/br , where
br = max {|Intermediate(R)|}.
(R,P )∈C
Proof. Same as the proof of Theorem 4.2, except that we argue l ≤ br instead of l ≤ b, using the argument
from the proof of Theorem 4.10.
The following theorem helps us further eliminate a huge number of pathways from consideration.
Definition 26. Let p be a semiformal pathway. We say that p is strongly decomposable if p can be decomposed into two semiformal pathways p1 and p2 such that at least one of p1 and p2 is a formal pathway.
Theorem 4.12. Let p be a semiformal pathway. If it is strongly decomposable, any semiformal pathway
that contains p as a prefix is decomposable.
Proof. Suppose p is strongly decomposable into formal pathway p1 and semiformal pathway p2 . We show
that for any p0 , if p + p0 is semiformal, then it is decomposable into p1 and p2 + p0 . It suffices to show that
p2 +p0 is semiformal. Assume towards a contradiction that the initial state of p2 +p0 contains an intermediate
species. Since p2 is semiformal, it means that there is an intermediate species x contained in S − T , where
T is the final state of p2 and S is the initial state of p0 . Let T 0 be the final state of p. Since p1 is formal,
Intermediate(T 0 ) = Intermediate(T ). Hence, x is also contained in S − T 0 , which means that x appears in
the initial state of p + p0 . This is a contradiction to our initial assumption that p + p0 was semiformal. Hence,
p2 + p0 is semiformal and therefore p + p0 is decomposable.
Finally, we show that CRNs that possess a certain structure can be processed extremely quickly. Since
most published implementations do have this structure (e.g. [36, 5, 31]), this observation is very useful in
practice.
Definition 27. Let C be a CRN. We define the following two sets, which partition the set of intermediate
species of C according to whether they ever participate in a reaction as a reactant.
W (C) = {species x in C : x is not formal and x never appears as a reactant in reactions of C}
N W (C) = {species x in C : x is not formal and x ∈
/ W (C)}
Moreover, for any state S, we will denote by S N W (C) the multiset containing exactly those species of S that
belong to N W (C).
Definition 28. A CRN C is said to have monomolecular substructure if for every reaction (R, P ) ∈ C,
both |RN W (C) | and |P N W (C) | are at most one.
20
Theorem 4.13. Let C be a CRN that has monomolecular substructure and p = (r1 , . . . , rk ) any undecomposable semiformal pathway of C. Also, let S0 be the initial state of p and Si = S0 ⊕ r1 ⊕ · · · ⊕ ri all the
N W (C)
states that p goes through. Then, for any 0 ≤ i ≤ k, we have |Si
| ≤ 1.
Proof. We prove this by induction on k. If k = 1, the claim holds trivially. Now assume that the claim holds
for all pathways of length up to k −1. Let p−1 be the pathway obtained by removing the last reaction rk from
p. Note that rk consumes up to one intermediate species because the CRN has monomolecular substructure.
Moreover, rk must consume at least one intermediate species because otherwise p can be decomposed into
(r1 , . . . , rk−1 ) and (rk ). Therefore rk consumes exactly one intermediate species x, which by definition
N W (C)
must be in N W (C). By induction hypothesis, |Si
| ≤ 1 for all 0 ≤ i ≤ k − 1 (the initial state of p−1
and the initial state of p differ only by formal species) and in particular the final state of p−1 contains at
most one intermediate species that belongs to N W (C). This implies that this intermediate species must be
x, because otherwise p = p−1 + (rk ) would not be semiformal. The last reaction rk consumes this x and
N W (C)
produces at most one intermediate species that belongs to N W (C), which means that |Sk
| ≤ 1. The
theorem now follows by induction.
The above theorem implies that when we run our algorithm on a CRN with monomolecular substructure,
there is no need to enumerate semiformal pathways that ever go through a state that contains more than one
species from N W (C).
4.7
Testing pathway decomposition equivalence
In this section, we have presented an algorithm for enumerating the formal basis of a given CRN, which
is guaranteed to halt if there is a finite bound to the width of an undecomposable semiformal pathway.
Moreover, this algorithm can also be used to test whether the CRN is tidy and regular.
Hence, we are finally in a position to be able to verify the correctness of CRN implementations; namely,
using the above algorithm we can test whether the target CRN and the implementation CRN are pathway
decomposition equivalent. Since it immediately follows from definition that the target CRN is tidy and regular and that its formal basis is equal to itself, this verification amounts to checking that the implementation
CRN is tidy and regular and that its formal basis is equal to the target CRN up to addition or removal of
trivial reactions. All of these tasks can easily be achieved using our algorithm.
We note that because of Theorem 3.1 our theory applies also to the more general scenario of comparing
two arbitrary CRNs. In this case, one would need to enumerate the elementary and formal bases of both
CRNs, verify that both CRNs are tidy and regular, and finally check that their formal bases are identical up
to addition or removal of trivial reactions.
5
Handling the general case
In this section, we discuss some important issues that pertain to practical applications and hint at the possibility of further theoretical investigations.
As we briefly mentioned earlier, many CRN implementations that arise in practice involve not only
formal and intermediate species but also what are called fuel and waste species. Fuel species are chemical
species that are assumed to be always present in the system at fixed concentration, as in a buffer. For
instance, DNA implementations [36, 5, 31] often employ fuel species that are present in the system in large
concentrations and have the ability to transform formal species into various other intermediates. This type
of “implementation” is also prevalent in biological systems, where the concentrations of energy-carrying
species such as ATP, synthetic precursors such as NTPs, and general-purpose enzymes such as ribosomes
and polymerases, are all maintained in roughly constant levels by the cellular metabolism.
21
In CRN verification, the standard approach to fuel species is to preprocess implementation CRNs such
that all occurrences of fuel species are simply removed. For instance, if the CRN contained reaction A+g →
i + t where g and t are fuel species, the preprocessed CRN will only have A → i. The justification for this
type of preprocessing is that since fuel species are always present in the system in large concentrations by
definition, consuming or producing a finite number of fuel species molecules do not have any effect on the
system. In particular, it can be shown that holding fuel species at constant concentration, versus simply
removing them from the reactions while appropriately adjusting the reaction rate constants, leads to exactly
the same mass-action ODE’s and continuous-time Markov chains.
On the other hand, implementations sometimes produce “waste” species as byproducts. Waste species
are supposed to be chemically inert and thus cannot have interaction with other formal or intermediate
species. However, in practice it is often difficult to implement a chemical species which is completely inert
and therefore they may interact with other species in trivial or nontrivial ways. Therefore the main challenge
is to ensure that these unwanted interactions do not give rise to a logically erroneous behavior. One way
to deal with this problem is to first verify that such waste species are indeed “effectively inert” and then
preprocess them in a similar manner to fuel species. To achieve this we need to answer two important
questions: first, how to satisfactorily define “effectively inert” and second, how such waste species may be
algorithmically identified.
Another related problem which must be solved before we can use pathway decomposition is that some
implementations may have multiple chemical species that are interpreted as the same formal species. (For
example, see DNA implementations [36, 5] with “history domains.” An example is given in Section 6.)
Since our mathematical framework implicitly assumes one-to-one correspondence between formal species
of the target CRN and formal species of the implementation CRN, it is not immediately clear how we can
apply our theory in such cases.
Interestingly, the weak bisimulation-based approach to CRN equivalence proposed in [12, 22] does
not seem to suffer from any of these problems, because it in fact does not make a particular distinction
between these different types of species except fuel species. Rather, it requires that there must be a way to
interpret each species that appears in the implementation CRN as one or more formal species. For instance,
if {A
i, B + i
j, j → C} is proposed as an implementation of A + B → C, the weak bisimulation
approach will interpret A and i as ⦃A⦄, B as ⦃B⦄, j as ⦃A, B⦄, and C as ⦃C⦄. Therefore the state
of the system at any moment will have an instantaneous interpretation as some formal state, which is not
provided by pathway decomposition. On the other hand, the weak bisimulation approach cannot handle
interesting phenomena that are allowed in the pathway decomposition approach, most notably the delayed
choice phenomenon explained in Section 2.
Our proposed solution to the problem of wastes and multiple formal labeling is a compositional hybrid
approach between weak bisimulation and pathway decomposition. Namely, we take the implementation
CRN from which only the fuel species have been preprocessed, and tag as “formal” species all the species
that have been labeled by the user as either an implementation of a target CRN species or a waste. All
other species are tagged as “intermediates”. Then we can apply the theory of pathway decomposition to
find its formal basis (with respect to the tagging, as opposed to the smaller set of species in the target
CRN). Note that waste species must be tagged as “formal” rather than “intermediate” because they will
typically accumulate, and thus tagging them as “intermediate” would result in a non-tidy CRN to which
pathway decomposition theory does not apply. Finally, we verify that the resulting formal basis of tagged
species is weak bisimulation equivalent to the target CRN under the natural interpretation, which interprets
implementations of each target CRN species as the target CRN species itself and wastes as “null.” If the
implementation is incorrect, or if some species was incorrectly tagged as “waste”, the weak bisimulation
test will fail. See Figure 4 for example.
On the other hand, we note that the weak bisimulation approach can sometimes handle interesting cases
which pathway decomposition cannot. For instance, the design proposed in [31] for reversible reactions
22
A1 → i
i → B1 + W
A2 → j
j → B2
W + j → B1
A1 → B1 + W
A2 → B 2
A2 + W → B 1
A→B
Implementation CRN
Formal basis
Under weak bisimulation
Figure 4: The compositional hybrid approach for verifying an implementation of the formal CRN {A → B}.
We first apply pathway decomposition, treating the upper case species as formal species and lower case
species as intermediate species. Then, we apply weak bisimulation using the natural interpretation which
interprets A1 and A2 as {|A|}, B1 and B2 as {|B|}, and W as ∅. Thus, in two steps, the implementation
CRN has been shown to be a correct implementation of {A → B}.
implements A+B
C+D as {A
i, i+B
j, j
k+C, k
D}. Note that this implementation CRN
is not regular according to our theory because of the prime pathway A → i, i + B → j, j → k + C, k + C →
j, j → i + B, i → A. Interestingly, this type of design seems to directly oppose the foundational principles
of the pathway decomposition approach. One of the key ideas that inspired pathway decomposition is that
of “base touching,” namely the idea that even though the evolution of the system involves many intermediate
species, a pathway implementing a formal reaction must eventually produce all its formal products and thus
“touch the base.” This principle is conspicuously violated in the above pathway, because while the only
intuitive way to interpret it is as A + B → C + D and then C + D → A + B, the first part does not touch
the base by producing a D molecule. In contrast, the weak bisimulation approach naturally has no problem
handling this implementation: i is interpreted as ⦃A⦄, j is interpreted as ⦃A, B⦄, and k is interpreted as
⦃D⦄.
The fact that the two approaches are good for different types of instances motivates us to further generalize the compositional hybrid approach explained above. To define the generalized compositional hybrid approach, we begin by formally introducing the weak bisimulation approach of [12, 22]. As we have
seen above, the weak bisimulation approach requires an “interpretation map” m from species of the implementation CRN to states of the target CRN. For instance, in the above example m was defined as
m(A) = m(i) = ⦃A⦄, m(B) = ⦃B⦄, m(j) = ⦃A, B⦄, m(C) = ⦃C⦄, and m(D) = m(k) = ⦃D⦄.
Although the domain of m is technically species of the implementation CRN, there is an obvious sense in
which we can also apply
P it to states, reactions, or pathways. Thus when convenient we will abuse notation to mean m(S) = x∈S m(x) for a state S, m(r) = (m(R), m(P )) for a reaction r = (R, P ), and
m(p) = (m(r1 ), m(r2 ), . . . , m(rk )) for a pathway p = (r1 , . . . , rk ). Then, the following definition and
theorem are adapted from [12, 22] to fit our definitions of chemical reactions and pathways.
Definition 29. (Section 3.2 of [22]) A target CRN C1 and an implementation CRN C2 are weak bisimulation
equivalent under interpretation m if
1. for any state S in C1 , there exists a state S 0 in C2 such that m(S 0 ) = S,
2. for any state S 0 in C2 and S = m(S 0 ),
(a) if r ∈ C1 can occur in S, then there exists a pathway p = (s1 , . . . , sk ) in C2 such that S ⊕ r =
m(S 0 ⊕ s1 ⊕ · · · ⊕ sk ) and m(p) is equal to (r) up to addition or removal of trivial reactions, and
(b) if r0 ∈ C2 can occur in S 0 , then m(r0 ) is either a reaction in C1 or a trivial reaction, and thus
m(S 0 ⊕ r0 ) = S ⊕ m(r0 ).
23
Theorem 5.1. (An immediate corollary of Theorem 1 of [22]) If a target CRN C1 and an implementation
CRN C2 are weak bisimulation equivalent under interpretation m, then the following holds:
1. If S is a state in C1 , p is a pathway in C1 that can occur in S, and S 0 is a state in C2 such that m(S 0 ) = S,
then there exists a pathway p0 in C2 such that p0 can occur in S 0 and m(p0 ) is equal to p up to addition
or removal of trivial reactions.
2. If S 0 is a state in C2 and p0 is a pathway in C2 that can occur in S 0 , then there exists a pathway p in C1
such that p can occur in m(S 0 ) and p is equal to m(p0 ) up to addition or removal of trivial reactions.
Similarly to Theorem 3.5, the above theorem establishes a kind of pathway equivalence between the
target CRN and the implementation CRN. Now, we can formally define the generalized compositional hybrid
approach as follows.
Definition 30. Suppose we are given a target CRN C1 and an implementation CRN C2 . Let F and S denote
the species of C1 and C2 respectively. Let X ⊆ S be the set of species that have been labeled by the user as
implementations of target CRN species or wastes. In the compositional hybrid approach, we say C2 is a
correct implementation of C1 if there exists some X ⊆ V ⊆ S such that
1. C2 with respect to V as formal species is tidy and regular, and
2. the formal basis of C2 with respect to V as formal species is weak bisimulation equivalent to C1 under
some interpretation that respects the labels on X provided by the user.
The flexibility to vary V can be useful: for example, intermediates that are involved in “delayed choice”
pathways can be kept out of V so as to be handled by pathway decomposition, whereas intermediates involved in the aforementioned reversible reaction pathways can be retained within V so as to be handled by
weak bisimulation.
Finally, we prove a theorem analogous to Theorems 3.5 and 5.1, in order to provide an intuitive justification for the adequacy of the above definition. We begin by extending the notion of interpretation of
pathways that we introduced in Section 3.4 to include the concept of interpretation map.
Definition 31. Suppose V denotes the set of species of C2 that are being tagged as formal species in the
compositional hybrid approach. Let m be an interpretation map from V to states of C1 . We say a formal
pathway p = (r1 , . . . , rk ) in C2 can be interpreted as a pathway q = (s1 , . . . , sl ) in C1 under m if
1. q can occur in m(S), where S is the initial state of p,
2. m(S ⊕ r1 ⊕ · · · ⊕ rk ) = m(S) ⊕ s1 ⊕ · · · ⊕ sl , and
3. there is a decomposition of p such that if we replace the turning point reaction of each prime pathway
with the corresponding element of C1 (i.e. the corresponding formal basis reaction mapped through
m) and remove all other reactions, then the resulting pathway is equal to q up to addition or removal
of trivial reactions.
Then, the following theorem provides a sense in which two CRNs that are “equivalent” according to the
compositional hybrid approach indeed do have equivalent behaviors.
Theorem 5.2. Suppose an implementation CRN C2 is a correct implementation of the target CRN C1 according to the compositional hybrid approach. Then, there exists a mapping m from V to states of C1 such
that the following two conditions hold.
1. Let q and S be a pathway and a state in C1 such that q can occur in S. Then, for any state S 0 in C2 that
uses species from V such that m(S 0 ) = S, there exists a formal pathway p in C2 that can occur in S 0
and can be interpreted as q under m.
24
2. Any formal pathway p in C2 can be interpreted as some pathway q in C1 under m.
Proof. Let m be the interpretation map provided by the weak bisimulation equivalence [12, 22], and I the
formal basis of C2 with respect to V as formal species.
1. By Theorem 5.1, we have a pathway p0 in I that can occur in S 0 and m(p0 ) is equal to q up to addition
or removal of trivial reactions. Now replace each reaction in p0 by the prime pathway that implements
that reaction and call the resulting pathway p. Clearly, p can occur in S 0 . To show that p can be
interpreted as q under m, observe that the first condition of Definition 31 follows from the fact that q
can clearly occur in m(S 0 ) and S 0 is a superset of the initial state of p (because p can occur in S 0 ). Since
p and p0 have the same initial and final states and m(p0 ) = q, we also satisfy the second condition. The
final condition trivially follows from the way p was constructed and the fact that m(p0 ) was equal to q
up to addition or removal of trivial reactions.
2. By Theorem 3.5, p can be interpreted as some pathway p0 in I. Let q be the pathway we obtain by
removing all the trivial reactions from m(p0 ). Now we show that p can be interpreted as q under m.
For the first condition of Definition 31, we use Theorem 5.1 to see that q can occur in m(S 0 ) where S 0
is the initial state of p0 . Since p0 can occur in the initial state of p, this implies that q can also occur in
m(S) where S is the initial state of p. The second condition follows immediately from the way q was
constructed and the fact that p and p0 have the same net effect. The final condition follows from the
fact that p can be interpreted as p0 in I and that m(p0 ) = q up to removal of trivial reactions.
As we shall see in Section 6, the compositional hybrid approach allows for the verification of interesting
real-life systems that neither pathway decomposition nor bisimulation is able to handle individually. In fact,
the compositional hybrid approach seems to be the most general approach proposed thus far in terms of the
range of implementations that it can address, which, to our best knowledge, includes all currently known
enzyme-free DNA implementation techniques. At the same time, we remark that its definition as presented
in this paper does not seem to be completely satisfactory. To see why, consider the CRN {A → i, i + B1
j1 , i + B2
j2 , j1 → C, j2 → C} as an implementation of {A + B → C}. Intuitively, it seems that the
compositional hybrid approach should have no problem handling this example with V = {A, B1 , B2 , C}
and m(A) = ⦃A⦄, m(B1 ) = m(B2 ) = ⦃B⦄, m(C) = ⦃C⦄. Surprisingly, it turns out that the prime
pathway A → i, i + B1 → j1 , j1 → i + B1 , i + B2 → j2 , j2 → C is not regular under this choice of V,
because the product B1 is produced before the reactant B2 is consumed. Of course, the compositional hybrid
approach can still handle this implementation because we can always choose V = {A, B1 , B2 , C, i, j1 , j2 }
and delegate the whole verification to the bisimulation part. Nonetheless, it is troubling that the above
pathway is considered irregular because if indeed B1 and B2 both represent B, then there is a sense in
which this pathway should really be thought of as A → i, i + B → j1 , j1 → i + B, i + B → j2 , j2 → C
and hence be considered regular.
Towards the resolution of the above issue, we may want to imagine a modified version of hybrid approach where pathway decomposition and bisimulation are not merely composed as in the above definition,
but combined in a more integrated manner. For example, we have considered a kind of “integrated” hybrid
approach in which regularity and the delimiting condition of weak bisimulation [12, 22] are tested only after
we apply the interpretation m to the prime pathways in the elementary basis. While empirical results suggest
that such modifications may successfully fix the issue described above, their theoretical implications are yet
to be understood.
25
6
Case studies
In this section, we study five real-life examples from the field of DNA computing in order to demonstrate
how the theory developed in this paper can be applied in practice. The code that was used to test these examples is included as part of the Nuskell suite for compiling and verifying DNA implementations (previously
called BioCRN [35]). Nuskell interfaces the formal basis enumeration algorithm from Section 4 with other
software pieces to form the following pipeline for verifying CRN implementations. First, the given target
CRN is converted into a set of DNA molecules using the Nuskell compiler [35]. Second, all the reactions
that can occur between these DNA molecules are enumerated using Grun et al.’s domain-level DNA reaction enumerator [17], from which the fuel species are pruned out as described in Section 5. The resulting
reactions constitute the implementation CRN. Finally, either pathway decomposition, weak bisimulation
[12, 22], or the compositional hybrid approach can be applied to the target CRN and the implementation
CRN to verify that the two are indeed equivalent.
The current version of Nuskell implements a special case of the compositional hybrid approach in which
V = X (see Definition 30). In other words, the Nuskell compiler provides the verifier with not only the target
and implementation CRNs, but also information on formal and waste labeling. For formal labeling, it uses
a pattern matching algorithm described in [35] to decide which species in the implementation CRN should
correspond to which formal species. For waste labeling, it currently uses the following criterion proposed
in [35]:
Definition 32. A species is a non-waste if it is formal or it is a reactant of a reaction that involves at least
one non-waste either as a reactant or a product. An intermediate species that is not a non-waste is a waste
species.
Given this information, our verifier first finds the formal basis of the implementation CRN with respect
to exactly those species labeled as formal or waste species by the Nuskell compiler, and then verifies that this
formal basis is weak bisimulation equivalent to the target CRN under the natural interpretation that accords
with the compiler’s labeling.
Example #1
For the first example, we will implement the target CRN {A → X +Y +Z, A → X +Y, A → X, A → B}
using the translation scheme proposed in [36]. Figure 5 shows how a unimolecular reaction (i.e. reaction
with exactly one reactant) gets implemented in this translation scheme.
To understand how this translation works, first note that the figure makes use of domain-level annotation
as opposed to sequence-level annotation: that is, the DNA strands in Figure 5 are specified by numbered
segments, or “domains,” instead of the actual nucleotide base sequences of A, G, C, and T. Here, the star is
used to indicate sequence complementarity, e.g. segments 1 and 1* are complementary to each other. Under
the assumption that domains otherwise have very little complementarity, this abstraction is very useful in
modeling complex DNA systems.
In this implementation, g and t are fuel species that are assumed to be present in large concentration.
Hence, when the molecule A is present in the solution, A and g may collide and bind to each other by the
1 and 1* domains that are complementary to each other. When this happens, since the adjacent domains 2
and 3 on A and 2* and 3* on the bottom strand of g are also complementary to each other, the hybridization
can continue to the right, by a process called branch migration, thus displacing the top strand of g and
producing two species on the right-hand side of the first reaction. The resulting molecule i can then react
with another fuel species t to produce the desired products X and Y . Note that g and t can be easily modified
to implement reactions with different numbers of product molecules, e.g. A → X or A → X + Y + Z.
26
history species
domain identifier
?
1
2
3
3
10 4
i
10
3
7
11
? 1
1*
1* 2* 3*
A
2
2
4
10 4
3* 10* 4*
3
2
2* 3*
W1
g
11 7
2
5
6
11 7
8
9
2
3
4 11 7
10
3* 10* 4* 11* 7*
11*7*
W2
t
3
10 4
11 7
i
history species
domain identifier
10 4
5
X
6
history species
domain identifier
11 7
8
9
Y
Figure 5: The implementation of A → X + Y according to the translation scheme from [36]. Dotted boxes
indicate fuel species. Note that in this translation scheme each formal species molecule will retain a “history
domain” specific to the gate from which it was produced.
Note also that this scheme makes use of “history domains” in implementing formal species, represented
by the question mark in the domain specification. For instance, in this example, any single-stranded molecule
that has an arbitrary domain followed by domains 1, 2, and 3 is considered to be A. This is necessary because
if the implementation consists of multiple modules like the one depicted in this figure, A molecules produced
by different modules will have different history domains, each specific to the gate from which the molecule
was produced. However, all those different versions of A will then be able to participate in the same set
of downstream modules, because as can be seen in the figure, the history domains do not participate in the
reactions employed by those modules.
If we follow this translation scheme blindly, we would require exactly two fuel species for each unimolecular reaction in the target CRN. However, in the case of our target CRN {A → X + Y + Z, A →
X + Y, A → X, A → B}, there is an optimization technique we can use to reduce the number of fuel
species, exploiting the fact that the first three reactions in this target CRN are very similar to one another.
Namely, it turns out that in this case we can share one fuel species g among those three reactions, therefore
using only 6 fuel species to implement the four reactions in the target CRN rather than 2×4 = 8. In practice,
researchers who experiment with actual DNA systems generally want to employ such optimizations whenever possible, because they are often crucial to the cost and efficacy of the experiment. Figure 6 illustrates
how the optimized implementation works for this example.
In the next step, we preprocess the fuel and waste species to obtain a simpler CRN that involves only
formal and intermediate species. Although in principle waste species are handled using the compositional
hybrid approach, we will assume for the sake of presentation that in this example we can treat waste species
in the same way as fuel species. We note that this assumption is not far-fetched because in this example it
is rather obvious that the waste species do not participate in any reaction at all. After this preprocessing, the
resulting implementation CRN looks as follows:
A→i
i→X +Y +Z
i→X +Y
i→X
A→j
j→B
Moreover, the optimized implementation no longer uses multiple history domains for a single formal species,
so the theory of pathway decomposition can be directly applied to this implementation.
27
0
1
2
3
3
10 4
3
11
12
13
0 1
1*
1* 2* 3*
A
2
2
4
10
7
2
3
9
6
15
8
14
11 7
12 13
5
10 4
11*7*
3* 10* 4*
i
2
12*13*
3
10
3
10 4
11 7 12 13
9
6
8
11 7
5
10 4
2
11*7*
3
3
10 4
11 7 12 13
6
2
3* 10* 4*
i
3
10
1
2
3
A
2
3 16 17
j
2
3
16
17
0 1
1*
2
10 4
11
12
18
3* 16* 17*
19
2
3
3
2* 3*
16 17
3* 16* 17*
g6
6
11 7
8
9
Y
W5
5
6
X
13
W1
g5
11 7 12 13
12 13 14 15
Z
13
10 4
11 7
8
9
Y
5
6
X
W4
1* 2* 3*
16 17
4
7
5
X
3* 10* 4*
g4
0
10 4
3* 10* 4* 11* 7*
W3
5
10 4
4 11 7 12
10
g3
2
12 13
4 11 7
W2
3* 10* 4*
i
10 4
i
3* 10* 4* 11* 7* 12* 13*
g2
2
3
W1
g1
11 7 12 13
2
2* 3*
2
3 16 17
j
16 17 18 19
B
Figure 6: An optimized implementation of target CRN {A → X + Y + Z, A → X + Y, A → X, A → B}.
Before we proceed, we also remark that this example contains a notable instance of the delayed choice
phenomenon, and therefore cannot be verified by either weak bisimulation [12, 22] nor serializability [25].
Namely, the intermediate i has multiple fates ⦃X, Y, Z⦄, ⦃X, Y ⦄, and ⦃X⦄, and hence it is unclear what its
instantaneous interpretation should be. Indeed, we cannot interpret i to be any of ⦃X, Y, Z⦄, ⦃X, Y ⦄, and
⦃X⦄ because then the CRN would appear to contain reactions X + Y + Z → X + Y , X + Y → X + Y + Z,
and X → X + Y + Z, respectively. Neither can we interpret i to be A, because i cannot turn into B.
In contrast, pathway decomposition has no difficulty verifying this implementation CRN. Running the
algorithm for enumerating the basis, we find that its elementary basis is
{(A → i, i → X + Y + Z),
(A → i, i → X + Y ),
(A → i, i → X),
(A → j, j → B)},
from which it is clear that the CRN is tidy and regular, and moreover its formal basis equals the target CRN
that we desired to implement. Hence, this implementation is pathway decomposition equivalent to the target
CRN.
Example #2
Our second example is a verification of the network condensation procedure proposed in [17]. The domainlevel reaction enumerator from [17] can produce the output using several different semantics. One is “detailed” semantics, in which all internal configuration changes within DNA molecules are enumerated step
28
by step, one domain change at a time. Another is “condensed” semantics, in which internal configuration
changes that occur within a single molecule are considered to be one step. For instance, we note that Figure
5 is an example of condensed semantics. If the first reaction in Figure 5 was enumerated using detailed
semantics instead, it would be enumerated as three reactions instead of one, where the first reaction would
be A binding to g by domain 1, the second reaction would be domain 2 on A hybridizing with domain 2* on
the bottom strand of g (i.e. domain 2* on the top strand of g would now be displaced), and the third reaction
would be domain 3 on A hybridizing with domain 3* on the bottom strand of g (i.e. the top strand is now
completely released). For practical purposes, it is often convenient to use condensed semantics, which produces many fewer species and reactions while still capturing the essential behavioral features of the given
system.
While [17] provides its own theoretical justification for the correctness of this condensation procedure,
it would be interesting to verify using pathway decomposition that the CRNs generated by the two different
semantics are indeed equivalent. For example, let us consider the DNA system in Figure 7, enumerated using
detailed and condensed semantics of Grun et al.’s enumerator [17]. This example was taken from Figure
4 of [17] and was slightly modified to highlight the delayed choice phenomenon inherent in the system.
Condensed semantics identifies species that differ only by a reversible change of internal configuration with
one another, resulting in “resting sets” of species variants that are easily interconvertible. In this example,
species G and i4 are grouped together and will be treated as one species G in the condensed CRN (and
similarly D and i7 as one species D). In general, this grouping gives rise to subtle issues that necessitate
the use of the compositional hybrid approach (see Example #4), but we will show that the system at hand
is simple enough that we can verify it using only pathway decomposition. To achieve this, we treat the
CRN generated by condensed semantics as a target CRN and the CRN generated by detailed semantics as
an implementation CRN.
First of all, observe that the weak bisimulation approach of [12, 22] is not sufficient to verify the correctness of this condensation. For example, the intermediate species i19 does not admit an appropriate
instantaneous interpretation. If it is interpreted to be ⦃A, B⦄, the system would appear to contain the reaction A + B → C + D. If it is interpreted to be ⦃C, D⦄, the system would appear to contain the reaction
C + D → A + B. It cannot be interpreted as ⦃G, T ⦄, because it cannot react with an X molecule to produce
W.
In contrast, it is easily verified using our algorithm that the detailed CRN is tidy and regular and moreover
its formal basis is {D → D, G → G, G + T → A + B, G + T → C + D, T + X → W }, which is
equal to the condensed CRN up to addition or removal of trivial reactions. Therefore the condensation in
this example is correct according to pathway decomposition equivalence.
Example #3
As a third example, we consider the same translation scheme and optimization technique as in Example
#1, this time applied to a different target CRN {A + X → X + X + A, A + X → X + X, A + X →
X, A → A + X + X, A → A + X}. While this example is very similar in flavor to Example #1, we can
no longer directly verify it with pathway decomposition because now there can be multiple history domains
for a single formal species. As can be seen in Figure 8, there are four different implementation species that
correspond to the formal species X and two different implementation species that correspond to the formal
species A. Moreover, as can be seen in Figure 9, bisimulation still does not apply because of the delayed
choice of species j. Hence, in this case the compositional hybrid approach is necessary to establish the
equivalence between the target CRN and the implementation CRN.
To verify this implementation using the compositional hybrid approach, we first enumerate species and
reactions using [17], then preprocess the fuel species to obtain the implementation CRN of Figure 9, and
finally run our formal basis enumerator on it treating Ai ’s, Xi ’s, and Wi ’s as formal species. It verifies that
29
a)
i4
G
T
X
i14
i13
W
b)
T
G
X
W
i19
A
A
B
C
c)
unimolecular reaction
2
G
C
3
condensed reaction
A
T
t
2
D
D
i7
bimolecular reaction
B
i4
3
B
W
i19
X
Figure 7: An example DNA system enumerated using detailed and condensed semantics. a) Detailed semantics enumeration. b) Condensed semantics enumeration. c) Key species shown in detail (adapted from
Figure 4 of [17]).
30
12
1
2
3
13
1
A
2
3
A
2
34
5
10
6
4
11
4
12
1
5
6
11 4
X
14 4
5
6
6
X
15 4
5
X
6
2
5
6* 10* 4*
g1
g1h
6
11 4
5
6
12 1
2
3
10 4
3
13
1
14
4
15
5
6* 10* 4*
12*1*
11*4*
g2
X
5
34
10 4
1* 2* 3*4* 5* 6*
10 4
2
6
11 4
5
6
10 4
5
6
6* 10* 4*
11*4*
g3
g4
4
13 1
1* 2* 3*
3* 13* 1*
g5
g6
2
3
14 4
14*4*
5
6
15 4
15*4*
5
6
13 1
3* 13* 1*
2
3
14 4
5
6
14*4*
g7
Figure 8: An optimized implementation of target CRN {A+X → X +X +A, A+X → X +X, A+X →
X, A → A + X + X, A → A + X}. (Left) All species of A and X with distinct history domains. (Right)
All fuel species of the optimized implementation.
the implementation CRN is tidy and regular and returns the formal basis shown in Figure 9. While the formal
basis turns out to be very large because of the existence of multiple history domains, it is easily shown to be
weak bisimulation equivalent to the target CRN under the obvious interpretation m(Ai ) = ⦃A⦄, m(Xi ) =
⦃X⦄, m(Wi ) = ∅. Therefore the given target CRN and implementation CRN are equivalent according to
the compositional hybrid approach.
We remark that even though this implementation CRN contained 25 species and 19 reactions, the basis
enumeration algorithm took less than one second to finish on an off-the-shelf laptop computer. This performance is perhaps surprising considering that the algorithm is based on brute-force enumeration, and it
suggests that despite the adverse worst-case time complexity, the algorithm may still be practical for many
instances that arise in practice. In fact, when we made full use of the optimization techniques outlined in
Sections 3.4.2 and 4.6, the algorithm terminated on almost all of our test instances in less than ten seconds.
The performance was particularly strong for implementations that had monomolecular substructure. For
example, the translation scheme from [36] applied on a target CRN consisting of 20 reactions produces an
implementation CRN consisting of 394 reactions and 387 species. However, since this implementation CRN
is modular and has monomolecular substructure, our verifying algorithm was able to successfully verify it
in mere 4 seconds. On the other hand, we report that the algorithm failed to terminate in an hour on some
instances that did not have monomolecular substructure.
Example #4
In our fourth example, we illustrate why pathway decomposition may not suffice for the verification of the
condensation procedure in [17] and how the compositional hybrid approach can be used to remedy this
problem. This example is almost identical to the system in Example #2, except that we remove species X
and W from the system and add a different species Y , which is merely the DNA strand consisting of a single
domain ‘a’.
Since there are many species in this system that have an unhybridized a* domain (e.g. G, i4, A, i7, and
D), molecule Y can react with those species to form various other species, thus giving rise to a reaction
network that is much more complex than in Figure 7. Most importantly, we note that molecule i7 can
also bind with Y to form a species which we shall call V . Since i7 is identified with D in condensed
semantics, this reaction will appear as D + Y → V in the condensed CRN. However, if we try to run
pathway decomposition on the detailed CRN treating as formal species only those species that appear in the
condensed CRN, we will find that the prime pathway G+T → i13, i13 → i19, i19 → C +i7, i7+Y → V
is irregular and thus pathway decomposition does not apply to this system (see Figure 10). The problem
31
A1 → A1
A2 → A2
A1 → i1
i1 → A1
A2 → i2
i2 → A2
i1 + X1 → j + W3
i1 + X2 → j + W4
i1 + X3 → j + W5
i1 + X4 → j + W6
i2 + X1 → j + W7
i2 + X2 → j + W8
i2 + X3 → j + W9
i2 + X4 → j + W10
j → X3 + W11
j → X3 + X4 + W12
j → A2 + X3 + X4 + W13
A1 → k + W 1
A2 → k + W 2
k → A1 + X1 + X2 + W14
k → A1 + X1 + W15
Implementation CRN
A1 + X1 → X3 + W11 + W3
A1 + X2 → X3 + W11 + W4
A1 + X3 → X3 + W11 + W5
A1 + X4 → X3 + W11 + W6
A2 + X1 → X3 + W11 + W7
A2 + X2 → X3 + W11 + W8
A2 + X3 → X3 + W11 + W9
A2 + X4 → X3 + W11 + W10
A1 + X1 → X3 + X4 + W12 + W3
A1 + X2 → X3 + X4 + W12 + W4
A1 + X3 → X3 + X4 + W12 + W5
A1 + X4 → X3 + X4 + W12 + W6
A2 + X1 → X3 + X4 + W12 + W7
A2 + X2 → X3 + X4 + W12 + W8
A2 + X3 → X3 + X4 + W12 + W9
A2 + X4 → X3 + X4 + W12 + W10
A1 + X1 → A2 + X3 + X4 + W13 + W3
A1 + X2 → A2 + X3 + X4 + W13 + W4
A1 + X3 → A2 + X3 + X4 + W13 + W5
A1 + X4 → A2 + X3 + X4 + W13 + W6
A2 + X1 → A2 + X3 + X4 + W13 + W7
A2 + X2 → A2 + X3 + X4 + W13 + W8
A→A
A+X →X
A2 + X3 → A2 + X3 + X4 + W13 + W9
A2 + X4 → A2 + X3 + X4 + W13 + W10
A+X →X +X
A1 → A1 + X1 + X2 + W14 + W1
A+X →A+X +X
A1 → A1 + X1 + W15 + W1
A→A+X +X
A2 → A1 + X1 + X2 + W14 + W2
A→A+X
A2 → A1 + X1 + W15 + W2
After applying the interpretation
Formal basis
Figure 9: Applying the compositional hybrid approach to Example #3.
32
A
T
A
C
?→A
?→B
?→B
B+A→C
C +B →D
D→?
B
B
C
D
C
T
C
Y
Y
i14
i19
i13
i4
D
(a)
i7
V
G
D
G
D
V
(b)
Figure 10: Conceptual issues for pathways involving resting sets with multiple interconvertible species.
Species are illustrated as dots, while resting sets are indicated by circles encoding the interconvertible
species. Although unimolecular reactions between the species within a given resting set form a strongly
connected graph, for clarity they are not explicitly shown. (a) A graph of the resting sets and potential reaction pathways for the adjacent implementation CRN. Reaction pathways do not necessarily involve formal species, but can instead utilize equivalent species from the same resting set. (b) A subgraph of the resting sets and reaction pathways for the CRN of Example #4 that illustrates the pathway
G + T → i13, i13 → i19, i19 → C + i7, i7 + Y → V .
here is that species i7 should have been considered a formal species because it is identified with D, even
though its name does not appear in the condensed CRN. Hence, in order to apply pathway decomposition
properly, we would need a way to inform the theory that i7 is also a copy of the species D.
We note that this problem is very similar to the problem of multiple history domains that we discussed
in Example #3. Just as A1 and A2 had both to be considered an implementation of A, in this example
we need to ensure that both D and i7 are considered an implementation of D. Thus, we can apply the
compositional hybrid approach to such systems by first running pathway decomposition on the detailed
CRN, treating as formal species all species from the condensed CRN and any other species that are identified
with those species within resting sets, and then verifying that the resulting formal basis is weak bisimulation
equivalent to the condensed CRN. In our example, this would correspond to treating species like i4 and i7
as formal species, and then using the interpretation m(G) = m(i4) = ⦃G⦄, m(D) = m(i7) = ⦃D⦄ in the
bisimulation step of the compositional hybrid approach. All of this information, i.e. which species should be
considered formal and what interpretation should be used, is provided to our verifier software by the reaction
enumerator of [17]. This way, the irregular pathway from the previous paragraph would no longer be prime,
because it can now be decomposed into G + T → i13, i13 → i19, i19 → C + i7 and i7 + Y → V . The
full result of a compositional hybrid approach verification of this example is shown in Figure 11, where we
can easily check that the formal basis of the detailed CRN is weak bisimulation equivalent to the condensed
CRN under the interpretation m.
Like Example #3, we note that this is an example of a DNA system that neither pathway decomposition
nor bisimulation can verify, but the compositional hybrid approach can.
33
i13
i14
A+Y →Z
i13
i19
D → i7
i14 → A + B
D + Y → i41
i38
i40
i7 → D
i38
i46
i7 + Y → V
i40 → B + Z
G → i4
i46 → B + Z
G+T →A+B
i46 → C + V
G + T → C + i7
i19 → A + B
G+Y →U
i19 → C + i7
i4 → G
A+Y →Z
i4 + T → A + B
D
i4 + T → C + i7
i7
D + Y → i41
i4 + Y → i42
i7 + Y → V
T +U →B+Z
G
T +U →C +V
i4
G + T → i13
T + i42 → B + Z
G+Y →U
T + i42 → C + V
i4 + T → i14
U → i42
i4 + Y → i42
i42 → U
T + U → i38
i41 → V
T + i42 → i40
U
i42
V
i41
V → i41
Formal basis
Implementation CRN
A+Y →Z
D+Y →V
m(A) = ⦃A⦄,
m(B) = ⦃B⦄,
G+T →A+B
m(C) = ⦃C⦄,
m(D) = m(i7) = ⦃D⦄,
G+T →C +D
m(G) = m(i4) = ⦃G⦄,
m(T ) = ⦃T ⦄,
G+Y →U
m(i42) = m(U ) = ⦃U ⦄,
m(Y ) = ⦃Y ⦄,
T +U →B+Z
m(i41) = m(V ) = ⦃V ⦄,
m(Z) = ⦃Z⦄
T +U →C +V
Bisimulation interpretation m
Condensed CRN
Figure 11: Applying the compositional hybrid approach to Example #4.
34
Example #5
For the final example, we investigate the following CRN as an implementation of {A + B → C + D + E}:
A
i+j
i+B
k+l
k→C
l
m+n
m→D
j+n→E
Unlike the previous examples, this CRN was not constructed by a direct application of a published CRN
implementation scheme, although it was inspired by the reaction network given rise to by the “garbage
collection” module of [5]. Nonetheless, this system is very interesting because it distinguishes pathway
decomposition from bisimulation [12, 22] or serializability [25] without making use of the delayed choice
phenomenon. Therefore it suggests that delayed choice may in fact be just one example of many interesting
behaviors that are allowed by pathway decomposition but not by other approaches.
First, we observe that this implementation would be deemed “incorrect” by bisimulation or serializability. To see this for bisimulation, we simply note that the interpretation of k must be C because
of reaction k → C. However, this would result in k + l → i + B being interpreted as some reaction that consumes at least one C and produces at least one B, which means that this CRN cannot be
a correct implementation of {A + B → C + D + E} according to bisimulation. To see it for serializability, we note that, roughly translated into our language, serializability requires every prime pathway in a module to have a well-defined turning point reaction and moreover visit the same set of states
prior to the turning point reaction. In this example, the turning point reaction of a prime pathway would
be defined as the first irreversible reaction that occurs in the pathway, e.g. reaction k → C in pathway
(A → i + j, i + B → k + l, k → C, l → m + n, m → D, j + n → E) and reaction j + n → E in
pathway (A → i + j, i + B → k + l, l → m + n, j + n → E, k → C, m → D).4 Noting that these
two pathways do not visit the same set of states prior to the respective turning point reactions, we conclude
that this implementation cannot be handled by serializability. However, we do remark that there could be a
relaxed version of serializability which can handle this example.
In contrast to the above two approaches, pathway decomposition does not have any difficulty with this
example. Not only does it allow prime pathways implementing the same reaction to have different choices
of turning point reactions, but it also allows them to go through different sets of states as long as each of
those prime pathways is regular. In fact, our basis enumeration algorithm easily finds the formal basis of the
given implementation CRN to be {A → A, A + B → A + B, A + B → C + D + E}, proving that it is a
correct implementation of {A + B → C + D + E} according to pathway decomposition.
7
Conclusions
The development of pathway decomposition theory was motivated by a desire for a general notion of CRN
behavioral equivalence, up to ignoring rate constants and implementation intermediates. An overarching
challenge is that, despite the set of species and set of reactions both being finite, the set of possible initial
states may be infinite, the set of system states reachable from a given initial state may be infinite, and the set
of possible pathways from a given initial state may be infinite – yet we desire a guarantee that the available
4
Lakin et al.’s definition of a turning point (which they call a “commit reaction”) is slightly different from our definition. For
detail, refer to [25].
35
behaviors within two CRNs are essentially identical, and we desire that this guarantee may be found (or
refuted) algorithmically in all cases. These factors eliminate many standard approaches – such as those that
only handle finite state spaces – from consideration. How well does the pathway decomposition approach
meet these goals?
The central concepts of pathway decomposition are quite general, allowing application of the theory
to a wide range of CRNs, but some important limitations are imposed. First, species must be divided into
formal species and intermediate species, and we are only concerned with “formal” pathways of reactions
that start with purely formal states and end in purely formal states. (I.e. our theory does not concern itself
with what may or may not happen when you start the CRN with intermediate species; they occur only in the
middle of formal pathways.) The basic idea is that any such formal pathway can be decomposed (perhaps
not uniquely) into interleaved sub-pathways until non-decomposable (“prime”) pathways are reached. The
set of all such prime pathways defines the formal basis for the CRN – the corresponding set of initial and
final states for the set of prime pathways – and two CRNs with the same formal basis are deemed equivalent.
The consequence is that the sets of formal pathways in the two CRNs can be put in correspondence with
each other by decomposing into prime pathways and replacing each prime pathway by its formal basis
reaction. In this sense, anything that one CRN can do, can also be done by the other. However, pathway
decomposition theory applies only to CRNs that are tidy (any state reached from a formal state can clean
up all intermediate species and return to a formal state) and regular (every prime pathway consists of a
consumptive phase followed by a productive phase, separated by a turning point reaction).
The choices implicit in the formulation of pathway decomposition allow for a general and elegant theory. A primary feature is that other than regularity, there are no structural constraints on what goes on inside
the prime pathways. This provides the potential for intermediates within a pathway to perform a non-trivial
(deterministic or non-deterministic) computation. In particular, intermediates can be shared between prime
pathways, and intermediates may not “know” which pathway they are on – a phenomenon we call “delayed
choice”. This is substantially less restrictive than other related methods for CRN implementation verification [12, 22, 25]. On the other hand, in the case that a CRN can be divided into two parts with distinct
intermediates, a modularity property holds such that the formal bases of the two parts can be considered
independently (as in a previous method [25]). This greatly facilitates algorithmic verification of CRN implementations. Finally, because the formal basis of an implementation CRN is unique (up to choice of
which species are considered formal and which are intermediates), a given implementation CRN cannot be
considered a correct implementation of two distinct formal CRNs – a natural property that, again, does not
hold for some related methods [12, 22].
On the other hand, the elegance and generality of pathway decomposition theory come at a cost. Because
the core theory only addresses tidy CRNs, fuel and waste species must be removed by pre-processing outside
the core pathway decomposition theory. Because fuels are presumed to be held at constant concentrations,
they can be eliminated from the CRN representation with a change of rate constants and absolutely no effect
on the dynamics. However, implementation species considered “waste” are often not entirely inert, but
rather their interactions with the system are such that their presence or absence does not affect the possible
formal pathways – for example, they might interact only with other “waste” species. Further, because the
core theory requires a one-to-one correspondence between formal species and selected representative species
in the implementation, the core theory is insufficient for implementation schemes where a formal species
may be represented by molecules with variable regions, such as the “history domains” of Soloveichik et al
[36]. To accommodate these concerns, we developed a compositional hybrid theory, which allows pathway
decomposition to be first applied to the implementation CRN with all waste species and all representations
of formal species being designated as “formal”, after which bisimulation [12, 22] is applied to the resulting
formal basis to establish correctness with respect to the original formal CRN. For both the core theory and
the compositional hybrid theory, “correctness” of an implementation provides guarantees that every pathway
in the formal CRN can occur in the implementation and vice versa, under an appropriate interpretation of
36
implementation pathways.
The compositional hybrid theory is conceptually sufficient for verifying – or finding errors in – CRNs
implemented according to most published translation schemes5 . However, the algorithmic challenges of
finding the formal basis for pathway decomposition theory, and of finding the interpretation function for
bisimulation theory, are substantial: the verification of correctness pertains to all possible initial states of
the CRNs (which are infinite in number) and all possible pathways from these states (also infinite). To meet
this challenge, we developed the notion of “signatures” for partial pathways, proving that they are bounded
in number for implementations for which the prime pathways have bounded width, and thus proving that at
least in this case our algorithm is guaranteed to terminate, even if there are an infinite number of distinct
prime pathways. Although the worst-case complexity of our algorithm is unknown, in practice implemented
CRNs have a modularity property that is easily recognized and exploited, often allowing verification to
complete in a matter of seconds for systems of the scale that is currently experimentally feasible6 . Although
verifying CRN bisimulation equivalence in the general case is PSPACE-complete [22], the bisimulation test
implemented for the compositional hybrid theory in Nuskell is a restricted case, and in practice this has not
proven to be the limiting step for difficult verification cases.
During the course of our investigations, we encountered a number of CRN implementations that intuitively seem correct, but which are not accepted by our theory, pointing to the need for a yet more general notion of correctness. We give four examples here. The first was previously mentioned: our notions of regularity and turning points, which appear necessary for correctly implementing irreversible reactions, preclude the use of physically reversible implementations of logically reversible reactions, such as
{A+B
i, i
j, j
C +D}. Such implementations appear in [31] and can reduce energy consumption
and reduce implementation complexity exponentially [37, 9]. Interestingly, the serializability approach to
verification [25] shares this restriction, but the bisimulation approach [12, 22] easily accommodates physically reversible implementations. Technically speaking, the compositional hybrid theory, which generalizes
both pathway decomposition and bisimulation, can handle the above example, but it is only by “abusing” the
theory by setting all species to be “formal” for its pathway decomposition step. As a second example, one
might have a reaction implementation where a “waste” species is produced prior to the turning point, e.g.,
{A → i + W, i → A, B + i → C}. The compositional hybrid theory will not accept this implementation
because in the initial pathway decomposition step, where waste W is temporarily considered “formal”, the
prime pathway (A → i + W, B + i → C) is not a regular implementation of A + B → C + W because
W is produced prior to the arrival of B. Although this type of situation does not commonly arise in DNA
strand displacement systems, it could be ameliorated by simply eliminating waste species (as defined in
Definition 32) from the implementation CRN prior to verification. A more general solution would be to
propose an “integrated hybrid theory” where regularity is tested only after the bisimulation interpretation
has been applied; however, this complicates attempts to prove a theorem relating formal and implementation
pathways analogous to Theorem 5.2. The third example concerns situations where formal species interact in
non-meaningful ways, resulting in pathways that have no net effect. As an illustration, in the implementation {A
i, i + B → C, i + D
j}, the pathway (A → i, i + D → j, j → D + i, i + B → C) is a prime
pathway taking A + B + D to D + C, but it is not regular. Pathways like (i + D → j, j → i + D), which
end where they begin and which never produce a net formal species that wasn’t previously consumed, could
be called futile loops; they would arise, for example, when the implementation model explicitly accounts
for fleeting binding between molecules that results in no transformation. In some published CRN-to-DNA
5
We have applied pathway decomposition to translation schemes from [36, 5, 6, 31, 8, 25], verifying implementations using
many schemes, revealing errors and suggesting fixes for some schemes, identifying concerns such a potential leak pathways in
other schemes, and encountering limitations of the theory for still other cases. These investigations will be reported in more detail
elsewhere; some general observations are mentioned below.
6
For example, a system of ten up-to-bimolecular reactions compiles, according to the translation scheme in [36], into a implementation with 185 species and 150 reactions and verifies in about ten seconds on a 2013 MacBook Pro.
37
translation schemes, especially those that involve variable “history” domains and multiple reversible steps
prior to the turning point (e.g. [5], and [36] generalized to trimolecular reactions), the involvement of futile
loops can result in irregular prime pathways that interlink multiple history-distinct versions of the same signal, such as (A → i, i + B1 → j1 , j1 → B1 + i, i + B2 → j2 , j2 → C). Interestingly, the integrated hybrid
theory could also address this particular problem, if it could be rigorously justified. Alternatively, it would
be desirable to allow decomposition by removing such futile loops (in which case the above pathway would
be considered to implement A + B → C, a sensible result) but unfortunately modifying our definitions to
allow this leads to additional complexities for the notion of a “signature”, and we have not been able to generalize our algorithm while retaining a proof that it is guaranteed to terminate. The fourth example highlights
a situation where our algorithm’s combinatorial explosion makes verification infeasible. Specifically, most
cases where verification is fast and scalable involve translation schemes that result in implementations that
are modular and have monomolecular substructure, as per Sections 3.4.2 and 4.6. However, the “garbage
collection” stages in the translation schemes of [5] do not have monomolecular substructure, and our algorithms are capable of verifying only the simplest instances. Interestingly, if the garbage collection stages are
removed or key garbage collection species are held at constant concentrations as fuel, then monomolecular
substructure is restored and verification proceeds apace. In summary, there is room for a deeper understanding of the notion of logical correctness for CRN implementations, and of the relative capabilities of different
existing theories, such as composition and modularity properties.
Even in their present form, existing theories and algorithms for establishing the correctness of CRN
implementations can play an important role in the development of rigorous compilers for molecular programming. We envision that future compilers for molecular programming will conform to the standards
established in electrical and computer engineering: complexity is managed using an abstraction hierarchy
that allows a program specification in a high-level language (such as Verilog) to be translated through a
series of intermediate-level languages to a physically implementable low-level language (such as transistorlevel netlists). Furthermore, the language at each level of the hierarchy has well-defined semantics (the
mathematical model describing the behavior of the program) and most importantly, formal proofs can establish that the compiler’s transformation from a higher-level language to a lower-level language preserves the
essential behavioral invariants [26, 32]. For molecular programming with dynamic DNA nanotechnology,
formal CRNs could serve as a higher-level language, while domain-level DNA strand displacement models
could serve as a lower-level language – the semantics of which provide the implementation CRN. Note that
in contrast to traditional compilers that (ideally) may be proved to be correct for all source programs, our
pathway decomposition theory and algorithms are best suited for evaluating the correctness on a case-bycase basis. While proving the general correctness of a CRN-to-DNA translation scheme would obviate the
need for time-consuming verification algorithms to be run, the advantages of case-by-case verification are
(1) when new translation schemes are proposed, they may immediately be used with confidence prior to
establishing what may be a difficult general-case proof; (2) when a new low-level semantics is considered
(e.g. either more or less detail in the molecular model), again there is no need to attempt a new proof for
the general case; (3) in cases where a translation scheme is in fact not correct in the general case, it may
still be used with confidence for CRNs that it does implement correctly; and (4) since it is highly desirable to make experimental systems as simple as possible, formal verification can be used to establish or
refute the correctness of arbitrary attempts to optimize and simplify the DNA-level design. These ideas have
been implemented in the verifying compiler, Nuskell, which has already been used to catch bugs in several
translation schemes [35].
Although the task of building an abstraction hierarchy for molecular programming with dynamic DNA
nanotechnology seems particularly tractable, in principle the same formalism could be used to establish the
correctness of other types of molecular and biochemical systems. For example, systems of protein enzymes
and nucleic acid substrates have been used to construct cell-free biochemical circuits [23, 30]; given a specification for the desired behavior as a formal CRN together with a CRN describing the actual implementation
38
details, one could use pathway decomposition theory to examine its logical correctness. Similarly, pathway
decomposition could provide an alternative perspective on the validity of descriptions at multiple levels of
details for biological networks studied in systems biology, or as an evaluation of coarse-graining and model
reduction techniques [1]. However, there are several limitations to pathway decomposition theory for these
purposes. First, in many such cases rate constants are important, but pathway decomposition theory does
not consider them. Related, in many cases approximate implementations are sufficient – for example, if
non-conforming pathways occur very rarely in the discrete stochastic (Gillespie) dynamics for the CRN, or
if non-conforming pathways “average out” in the continuous deterministic (ordinary differential equation)
dynamics [38, 39, 7]. Finally, for some purposes the target behavior that the implementation aims to achieve
is not best described as a CRN, but rather by some other specification language such as temporal logic.
Many of these issues are explored in the literature on Petri nets [19]. However, a more fundamental concern
is that CRNs are not an efficient representation for describing combinatorial processes in biology, for which
more effective models have been developed [16, 10] and analyzed [2]. It is reasonable to presume that such
models could in the future provide a programming language for more sophisticated molecular machines.
Nonetheless, the notion of logical correctness of CRN implementations that is provided by pathway decomposition theory has already proved its effectiveness for catching logical errors in CRN-to-DNA translation
schemes and appears to be particularly suitable for incorporation into automated verifying compilers for
molecular programming.
Acknowledgments
We appreciate helpful discussions with John Baez, Luca Cardelli, Vincent Danos, Qing Dong, Robert Johnson, Stefan Badelt, and Matthew Lakin. S.W.S. was supported by California Institute of Technology’s
Summer Undergraduate Research Fellowship 2009, NSF grant CCF-0832824, ARO Grant W911NF-09-10440 and NSF Grant CCF-0905626. C.T. was supported by NSF grants CCF-1213127, SHF-1718938, and a
Banting Fellowship. E.W. was supported by NSF grants CCF-0832824, CCF-1213127, and CCF-1317694.
References
[1] Jörg Ackermann, Jens Einloft, Joachim Nöthen, and Ina Koch. Reduction techniques for network
validation in systems biology. Journal of Theoretical Biology, 315:71–80, 2012.
[2] Nicolas Behr, Vincent Danos, and Ilias Garnier. Stochastic mechanics of graph rewriting. In Proceedings of the 31st Annual ACM/IEEE Symposium on Logic in Computer Science, pages 46–55. ACM,
2016.
[3] Robert Brayton and Alan Mishchenko. ABC: An academic industrial-strength verification tool. In
Computer Aided Verification, pages 24–40. Springer, 2010.
[4] Peter Buchholz. Bisimulation relations for weighted automata. Theoretical Computer Science,
393:109–123, 2008.
[5] Luca Cardelli. Strand algebras for DNA computing. Natural Computing, 10:407–428, 2011.
[6] Luca Cardelli. Two-domain DNA strand displacement. Mathematical Structures in Computer Science,
23:247–271, 2013.
[7] Luca Cardelli. Morphisms of reaction networks that couple structure to function. BMC Systems Biology, 8:84, 2014.
[8] Yuan-Jyue Chen, Neil Dalchau, Niranjan Srinivas, Andrew Phillips, Luca Cardelli, David Soloveichik,
and Georg Seelig. Programmable chemical controllers made from DNA. Nature nanotechnology,
8:755–762, 2013.
39
[9] Anne Condon, Alan J Hu, Ján Maňuch, and Chris Thachuk. Less haste, less waste: on recycling and
its limits in strand displacement systems. Interface Focus, 2:512–521, 2012.
[10] Vincent Danos, Jérôme Feret, Walter Fontana, Russell Harmer, and Jean Krivine. Rule-based modelling of cellular signalling. In Luı́s Caires and Vasco T. Vasconcelos, editors, International Conference
on Concurrency Theory (CONCUR 2007), volume 4703 of Lecture Notes in Computer Science, pages
17–41. Springer, 2007.
[11] Vincent Danos and Jean Krivine. Transactions in RCCS. In In Proc. of CONCUR, LNCS 3653, pages
398–412. Springer, 2005.
[12] Qing Dong. A bisimulation approach to verification of molecular implementations of formal chemical
reaction networks. Master’s thesis, Stony Brook University, 2012.
[13] Shawn M. Douglas, Ido Bachelet, and George M. Church. A logic-gated nanorobot for targeted transport of molecular payloads. Science, 335:831–834, 2012.
[14] Jonathan P. K. Doye, Thomas E. Ouldridge, Ard A. Louis, Flavio Romano, Petr Šulc, Christian Matek,
Benedict E. K. Snodin, Lorenzo Rovigatti, John S. Schreck, Ryan M. Harrison, and William P. J.
Smith. Coarse-graining DNA for simulations of DNA nanotechnology. Physical Chemistry Chemical
Physics, 15:20395–20414, 2013.
[15] Javier Esparza and Mogens Nielsen. Decidability issues for Petri nets. Petri Nets Newsletter, 94:5–23,
1994.
[16] James R. Faeder, Michael L. Blinov, and William S. Hlavacek. Rule-based modeling of biochemical
systems with BioNetGen. Methods in Molecular Biology, 500:113–67, 2009.
[17] Casey Grun, Karthik Sarma, Brian Wolfe, Seung Woo Shin, and Erik Winfree. A domain-level DNA
strand displacement reaction enumerator allowing arbitrary non-pseudoknotted secondary structures.
arXiv preprint arXiv:1505.03738, 2015. Originally appeared in Verification of Engineered Molecular
Devices and Programs (VEMDP) 2014.
[18] Hongzhou Gu, Jie Chao, Shou-Jun Xiao, and Nadrian C Seeman. A proximity-based programmable
DNA nanoscale assembly line. Nature, 465:202–205, 2010.
[19] Monika Heiner, David Gilbert, and Robin Donaldson. Petri nets for systems and synthetic biology.
Lecture Notes in Computer Science, 5016:215–264, 2008.
[20] Gerard J Holzmann. The model checker SPIN. IEEE Transactions on Software Engineering, 23:279–
295, 1997.
[21] Petr Jančar, Javier Esparza, and Faron Moller. Petri nets and regular processes. Journal of Computer
and System Sciences, 59:476–503, 1999.
[22] Robert F. Johnson, Qing Dong, and Erik Winfree. Verifying chemical reaction network implementations: A bisimulation approach. In Yannick Rondelez and Damien Woods, editors, DNA Computing
and Molecular Programming, volume 9818 of Lecture Notes in Computer Science (LNCS), pages
114–134. Springer, 2016.
[23] Jongmin Kim, Kristin S. White, and Erik Winfree. Construction of an in vitro bistable circuit from
synthetic transcriptional switches. Molecular Systems Biology, 2:68, 2006.
[24] Matthew R. Lakin and Andrew Phillips. Visual DSD. Microsoft Research.
[25] Matthew R. Lakin, Darko Stefanovic, and Andrew Phillips. Modular verification of chemical reaction
network encodings via serializability analysis. Theoretical Computer Science, 632:21–42, 2016.
[26] Xavier Leroy. Formal verification of a realistic compiler. Communications of the ACM, 52:107–115,
2009.
40
[27] Richard Lipton. The reachability problem requires exponential space. Research Report 62, Department
of Computer Science, Yale University, New Haven, Connecticut, 1976.
[28] Ernst Mayr. Persistence of vector replacement systems is decidable. Acta Informatica, 15:309–318,
1981.
[29] Robin Milner. Communication and Concurrency. Prentice Hall, 1989.
[30] Kevin Montagne, Raphael Plasson, Yasuyuki Sakai, Teruo Fujii, and Yannick Rondelez. Programming
an in vitro DNA oscillator using a molecular networking strategy. Molecular Systems Biology, 7:466,
2011.
[31] Lulu Qian, David Soloveichik, and Erik Winfree. Efficient Turing-universal computation with DNA
polymers. In Yasubumi Sakakibara and Yongli Mi, editors, DNA Computing and Molecular Programming, volume 6518 of Lecture Notes in Computer Science (LNCS), pages 123–140, 2011.
[32] Xavier Rival. Symbolic transfer function-based approaches to certified compilation. In Proceedings of
the 31st ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages, POPL ’04,
pages 1–13, New York, NY, USA, 2004. ACM.
[33] Joseph Schaeffer. Stochastic Simulation of the Kinetics of Multiple Interacting Nucleic Acid Strands.
PhD thesis, California Institute of Technology, 2013.
[34] Philippe Schnoebelen and Natalia Sidorova. Bisimulation and the reduction of Petri nets. Lecture
Notes in Computer Science, 1825:409–423, 2000.
[35] Seung Woo Shin. Compiling and verifying DNA-based chemical reaction network implementations.
Master’s thesis, California Institute of Technology, 2011.
[36] David Soloveichik, Georg Seelig, and Erik Winfree. DNA as a universal substrate for chemical kinetics. Proceedings of the National Academy of Sciences, 107:5393–5398, 2010.
[37] Chris Thachuk and Anne Condon. Space and energy efficient computation with DNA strand displacement systems. In Darko Stefanovic and Andrew Turberfield, editors, DNA Computing and Molecular
Programming, volume 7433 of Lecture Notes in Computer Science (LNCS), pages 135–149. Springer,
2012.
[38] David Thorsley and Eric Klavins. Model reduction of stochastic processes using Wasserstein pseudometrics. In American Control Conference, pages 1374–1381, 2008.
[39] David Thorsley and Eric Klavins. Approximating stochastic biochemical processes with Wasserstein
pseudometrics. IET Systems Biology, 4:193–211, 2010.
41
| 5cs.CE
|
Submitted to the Annals of Statistics
DYNAMIC NETWORK MODELS AND GRAPHON
ESTIMATION
arXiv:1607.00673v2 [math.ST] 7 Apr 2018
By Marianna Pensky∗
University of Central Florida
In the present paper we consider a dynamic stochastic network
model. The objective is estimation of the tensor of connection probabilities Λ when it is generated by a Dynamic Stochastic Block Model
(DSBM) or a dynamic graphon. In particular, in the context of the
b of Λ and
DSBM, we derive a penalized least squares estimator Λ
b
show that Λ satisfies an oracle inequality and also attains minimax
lower bounds for the risk. We extend those results to estimation of
Λ when it is generated by a dynamic graphon function. The estimators constructed in the paper are adaptive to the unknown number
of blocks in the context of the DSBM or to the smoothness of the
graphon function. The technique relies on the vectorization of the
model and leads to much simpler mathematical arguments than the
ones used previously in the stationary set up. In addition, all results
in the paper are non-asymptotic and allow a variety of extensions.
1. Introduction. Networks arise in many areas of research such as sociology, biology, genetics, ecology, information technology to list a few. An
overview of statistical modeling of random graphs can be found in, e.g., Kolaczyk (2009) and Goldenberg et al. (2011). While static network models are
relatively well understood, the literature on the dynamic network models is
fairly recent.
In this paper, we consider a dynamic network defined as an undirected
graph with n nodes with connection probabilities changing in time. Assume
that we observe the values of a tensor Bi,j,l ∈ {0, 1} at times tl where
0 < t1 < · · · < tL = T . For simplicity, we assume that time instants are
equispaced and the time interval is scaled to one, i.e. tl = l/L. Here Bi,j,l = 1
if a connection between nodes i and j is observed at time tl and Bi,j,l =
0 otherwise. We set Bi,i,l = 0 and Bi,j,l = Bj,i,l for any i, j = 1, · · · n
and l = 1, · · · , L, and assume that Bi,j,l are independent Bernoulli random
variables with Λi,j,l = P(Bi,j,l = 1) and Λi,i,l = 0. Below, we study two
∗
Supported in part by National Science Foundation (NSF), grants DMS-1407475 and
DMS-1712977
MSC 2010 subject classifications: Primary 60G05; secondary 05C80, 62F35
Keywords and phrases: dynamic network, graphon, stochastic block model, nonparametric regression, minimax rate
1
2
M. PENSKY
types of objects: a Dynamic Stochastic Block Model (DSBM) and a dynamic
graphon.
The DSBM can be viewed as a natural extension of the Stochastic Block
Model (SBM) which, according to Olhede and Wolfe (2014), provides an
universal tool for description of time-independent stochastic network data.
In a DSBM, all n nodes are grouped into m classes Ω1 , · · · , Ωm , and probability of a connection Λi,j,l is entirely determined by the groups to which
the nodes i and j belong at the moment tl . In particular, if i ∈ Ωk and
j ∈ Ωk′ , then Λi,j,l = Gk,k′ ,l . Here, G is the connectivity tensor at time tl
(l)
with Gk,k′ ,l = Gk′ ,k,l . Denote by nk the number of nodes in class k at the
moment tl , k = 1, . . . , m, l = 1, . . . , L.
A dynamic graphon can be defined as follows. Let ζ = (ζ1 , · · · , ζn ) be a
random vector sampled from a distribution Pζ supported on [0, 1]n . Although
the most common choice for Pζ is the i.i.d. uniform distribution for each ζi ,
we do not make this assumption in the present paper. We further assume
that there exists a function f : [0, 1]3 → [0, 1] such that for any t one has
f (x, y, t) = f (y, x, t) and
(1)
Λi,j,l = f (ζi , ζj , tl ),
i, j = 1, · · · , n, l = 1, · · · , L.
Then, function f summarizes behavior of the network and can be called
dynamic graphon, similarly to the graphon in the situation of a stationary network. This formulation allows to study a different set of stochastic
network models than the DSBM.
It is known that graphons play an important role in the theory of graph
limits described in Lovász and Szegedy (2006) and Lovász (2012). The definition of the dynamic graphon above fully agrees with their theory. Indeed,
for every l = 1, · · · , L, the limit of Λ∗,∗,l as n → ∞ is f (·, ·, tl ). We shall
further elaborate on the notion of the dynamic graphon in Section 7.
In the last few years, dynamic network models attracted a great deal of
attention (see, e.g., Durante et al. (2015), Durante et al. (2016), Han et
al. (2015), Kolar et al. (2010), Leonardi et al. (2016), Matias and Miele
(2015), Minhas et al. (2015), Xing et al. (2010), Xu (2015), Xu and Hero
III (2014) and Yang et al. (2011) among others). Majority of those paper
describe changes in the connection probabilities and group memberships
via various kinds of Bayesian or Markov random field models and carry
out the inference using the EM or iterative optimization algorithms. While
procedures described in those papers show good computational properties,
they come without guarantees for the estimation precision. The only paper
known to us that is concerned with estimation precision in the dynamic
DYNAMIC NETWORK MODELS AND GRAPHON ESTIMATION
3
setting is by Han et al. (2015) where the authors study consistency of their
procedures when n → ∞ or L → ∞.
On the other hand, recently, several authors carried out minimax studies
in the context of stationary network models. In particular, Gao et al. (2015)
developed upper and minimax lower bounds for the risk of estimation of the
matrix of connection probabilities. In a subsequent paper, Gao et al. (2016)
generalized the results to a somewhat more general problem of estimation
of matrices with bi-clustering structures. In addition, Klopp et al. (2017)
extended these results to the case when the network is sparse in a sense
that probability of connection is uniformly small and tends to zero as n →
∞. Also, Zhang and Zhou (2016) investigated minimax rates of community
detection in the two-class stochastic block model.
The present paper has several objectives. First, we describe the nonparametric DSBM model that allows for smooth evolution of the tensor
G of connection probabilities as well as changes in group memberships in
time. Second, we introduce vectorization of the model that enables us to take
advantage of well studied methodologies in nonparametric regression estimation. Using these techniques, we derive penalized least squares estimators
b of Λ and show that they satisfy oracle inequalities. These inequalities do
Λ
not require any assumptions on the mechanism that drives evolution of the
group memberships of the nodes in time and can be applied under very mild
conditions. Furthermore, we consider a particular situation where only at
most n0 nodes can change their memberships between two consecutive time
points. Under the latter assumption, we derive minimax lower bounds for
the risk of an estimator of Λ and confirm that the estimators constructed
in the paper attain those lower bounds. Moreover, we extend those results
to estimation of the tensor Λ when it is generated by a graphon function.
We show that, for the graphon, the estimators are minimax optimal within
a logarithmic factor of L. Estimators, constructed in the paper, do not require knowledge of the number of classes m in the context of the DSBM,
or a degree of smoothness of the graphon function f if Λ is generated by a
dynamic graphon.
Note that unlike in Klopp et al. (2016) we do not consider a network that
is sparse in a sense that probabilities of connections between classes are
uniformly small. However, since our technique is based on model selection,
it allows to study a network where some groups do not communicate with
each other and obtain more accurate results. Moreover, as we show in Section
6, by adjusting the penalty, one can provide adaptation to uniform sparsity
assumption if the number of nodes in each class is large enough.
The present paper makes several key contributions. First, to the best of
4
M. PENSKY
our knowledge, the time-dependent networks are usually handled via generative models that assume some probabilistic mechanism which governs
the evolution of the network in time. The present paper offers the first fully
non-parametric model for the time-dependent networks which does not make
any of such assumptions. It treats connection probabilities for each group
as functional data, allows group membership switching and enables one to
exploit stability in the group memberships over time. Second, the paper
provides the first minimax study of estimation of the tensor of connection
probabilities in a dynamic setting. The estimators constructed in the paper
are adaptive to the number of blocks in the context of the DSBM and to
the smoothness of the graphon function in the case of a dynamic graphon.
Moreover, the approach of the paper is non-asymptotic, so it can be used
irrespective of how large the number of nodes n, the number of groups m
and the number of time instants L are and what the relationship between
these parameters is. Third, in order to handle the tensor-variate functional
data, we use vectorization of the model. This technique allows to reduce the
problem of estimation of an unknown tensor of connection probabilities to
a solution of a functional linear regression problem with sub-gaussian errors. The technique is very potent and is used in a novel way. In particular,
it leads to much more simple mathematics than in Gao et al. (2015) and
Klopp et al. (2017). In the case of a time-independent SBM, it immediately
reduces the SBM to a linear regression setting. In addition, by using the
properties of the Kronecker product, we are able to reduce the smoothness
assumption on the connection probabilities to sparsity assumption on their
coefficients in one of the common orthogonal transforms (e.g., Fourier or
wavelet). Fourth, we use the novel structure of the penalty a part of which
is proportional to the logarithm of the cardinality of the set of all possible
clustering matrices over L time instants. The latter allows to accommodate
various group membership switching scenarios and is based on the Packing
lemma (Lemma 4) which can be viewed as a version of the VarshamovGilbert lemma for clustering matrices. In particular, while all papers that
studied the SBM dealt with the situation where no restrictions are placed on
the set of clustering matrices, our approach allows to impose those restrictions. Finally, the methodologies of the paper admit various generalizations.
For example, they can be adapted to a situation where the number of nodes
in the network depends on time, or the connection probabilities have jump
discontinuities, or when some of the groups have no connection with each
other. Section 6 shows that the technique can be adapted to an additional
uniform sparsity considered in Klopp et al. (2017) if the number of nodes in
each class is large enough.
DYNAMIC NETWORK MODELS AND GRAPHON ESTIMATION
5
The rest of the paper is organized as follows. In Section 2, we introduce
the notations and describe the vectorization of the model. In Section 3, we
b of the tensor Λ. In Secconstruct the penalized least squares estimators Λ
tion 4, we derive the oracle inequalities for their risks. In Section 5, we obtain
b
the minimax lower bounds for the risk that confirm that the estimators Λ
are minimax optimal. Section 6 shows how our technique provides adaptation to uniform sparsity assumption studied in Klopp et al. (2017). Section 7
develops the nearly minimax optimal (within a logarithmic factor of L) estimators of Λ when the network is generated by a graphon. Finally, Section 8,
provides a discussion of various generalizations of the techniques proposed
in the paper. The proofs of all statements are placed into the Supplemental
Material.
2. Notation, discussion of the model and data structures.
2.1. Notation. For any two positive sequences {an } and {bn }, an ≍ bn
means that there exists a constant C > 0 independent of n such that
C −1 an ≤ bn ≤ Can for any n. For any set Ω, denote cardinality of Ω by |Ω|.
For any x, [x] is the largest integer no larger than x.
For any vector t ∈ Rp , denote its ℓ2 , ℓ1 , ℓ0 and ℓ∞ norms by, respectively,
ktk, ktk1 , ktk0 and ktk∞ . Denote by kt1 − t2 kH the Hamming distance
between vectors t1 and t2 . Denote by 1 and 0 the vectors that have, respectively, only unit or zero elements. Denote by ej the vector with 1 in the j-th
position and all other elements equal to zero.
For a matrix A, its i-th row and j-th columns are denoted, respectively,
by Ai,∗ and A∗,j . Similarly, for a tensor A ∈ Rn1 ×n2 ×n3 , we denote its lth (n1 × n2 )-dimensional sub-matrix by A∗,∗,l . Let vec(A) be the vector
obtained from matrix A by sequentially stacking its columns. Denote by
A ⊗ B the Kronecker product of matrices A and B. Also, Ik is the identity
matrix of size k. For any subset J of indices, any vector t and any matrix
A, denote the restriction of t to indices in J by tJ and the restriction of A
to columns A∗,j with j ∈ J by AJ . Also, denote by t(J) the modification of
vector t where all elements tj with j ∈
/ J are set to zero.
For any matrix A, denote its spectral and Frobenius norms by, respectively, kAkop and kAk. Denote kAkH ≡ kvec(A)kH , kAk∞ = kvec(A)k∞
and kAk0 ≡ kvec(A)k0 . For any tensor A ∈ Rn1 ×n2 ×n3 , denote kAk2 =
P
n3
2
l=1 kA∗,∗,l k .
Denote by M(m, n) a collection of membership (or clustering) matrices
Z ∈ {0, 1}n×m , i.e. matrices such that Z has exactly one 1 per row and
Zik = 1 iff a node i belongs to the class Ωk and is zero otherwise. Denote by
6
M. PENSKY
C(m, n, L) a set of clustering matrices such that
(2)
C(m, n, L) ⊆
L
Y
l=1
M(m, n).
2.2. Discussion of the model. Note that the values of Bi,j,l are independent given the values of Λi,j,l , that is, Bi,j,l are independent in the sense
that their deviations from Λi,j,l are independent from each other. Therefore,
the values of Bi,j,l are linked to each other in the same way as observations
of a continuous function with independent Gaussian errors are related to
each other. Moreover, in majority of papers treating dynamic block models
(see, e.g., Durante et al. (2015), Han et al. (2015), Matias and Miele (2017),
Yang et al. (2011) among others), similarly to the present paper, the authors
assume that observations Bi,j,l are independent given Λi,j,l . Note that this is
not an artificial construct: Durante et al. (2015), for example, use the model
for studying international relationships between countries over time.
The only difference between the present paper and the papers cited above
is that we assume that the underlying connection probabilities G∗,∗,l are
functionally linked (e.g., smooth) rather than being probabilistically related. Indeed, many papers that treat dynamic block models assume some
Bayesian generative mechanism on the values of connection probabilities as
well as on evolution of clustering matrices. In particular, they impose some
prior distributions that relate G∗,∗,l+1 to G∗,∗,l and Z̃(l+1) to Z̃(l) , the matrices of underlying probabilities and the clustering matrices for consecutive
time points. Since the proposed generative mechanism may be invalid, we
avoid making assumptions about the probabilistic structures that generate
connection probabilities and group memberships, and treat the network as
a given object. However, our model enforces, in a sense, a more close but
yet flexible relation between the values of Bi,j,l since G∗,∗,l are functionally
(and not stochastically) related. Moreover, our theory allows to place any
restrictions on the set of clustering matrices.
To illustrate this point, consider just one pair of nodes (i, j) and assume
that these nodes do not switch their memberships between times tl and tl+1
and also that Gi,j,l is continuous at tl . It is easy to see that if Gi,j,l is
close to zero (or one), then Gi,j,l+1 is also close to zero (or one) and, hence,
Bi,j,l and Bi,j,l+1 are likely to be equal to zero (or one) simultaneously. This
relationship takes place in general.
To simplify the narrative, just for this paragraph, denote bl = Bi,j,l ,
bl+1 = Bi,j,l+1 , gl = Gi,j,l and gl+1 = Gi,j,l+1 . In order we are able to assert
conditional probabilities P(Bi,j,l+1 = 1|Bi,j,l = 1) ≡ P(bl+1 = 1|bl = 1) and
DYNAMIC NETWORK MODELS AND GRAPHON ESTIMATION
7
P(Bi,j,l+1 = 0|Bi,j,l = 0) ≡ P(bl+1 = 0|bl = 0), consider the situation where
gl and gl+1 are random variables with the joint pdf p(gl , gl+1 ) such that,
given gl , on the average gl+1 is equal to gl : E(gl+1 |gl ) = gl . Assume, as it
is done in the present paper, that, given gl , values of bl are independent
Bernoulli variables, so that
b
l+1
(1 − gl+1 )1−bl+1 .
p(bl , bl+1 |gl , gl+1 ) = glbl (1 − gl )1−bl gl+1
It is straightforward to calculate marginal probabilities P(bl = 1) = E(gl ),
P(bl+1 = 1) = E(gl+1 ) = E[E(gl+1 |gl )] = E(gl ) and the joint probability
P(bl = 1, bl+1 = 1) = E(gl+1 gl ) = E[E(gl+1 gl |gl )] = E(gl2 ) which yields
P(bl+1 = 1|bl = 1) − P(bl+1 = 1) =
E(gl2 )
Var(gl )
− E(gl ) =
>0
E(gl )
E(gl )
unless Var(gl ) = 0. The latter means that, even in the presence of the
assumption of the conditional independence, the probability of interaction
at the moment tl+1 is larger if there were an interaction at the moment tl
than it would be in the absence of this assumption. Similarly, repeating the
calculation with gl and gl+1 replaced by 1 − gl and 1 − gl+1 , obtain
P(bl+1 = 0|bl = 0) − P(bl+1 = 0) =
Var(gl )
> 0.
E(1 − gl )
In the absence of the probabilistic assumptions on gl and gl+1 , we cannot
evaluate those conditional probabilities but the relationship persists in this
situation as well.
2.3. Vectorization of the model. Note that tensor Λ of connection probabilities has a lot of structure. On one hand, it is easy to check that
(3)
Λ∗,∗,l = Z̃(l) G∗,∗,l (Z̃(l) )T ,
Bi,j,l ∼ Bernoulli(Λi,j,l ),
where Z̃(l) ∈ M(m, n) is the clustering matrix at the moment tl . On the
other hand, for every k1 and k2 , vectors Gk1 ,k2 ,∗ ∈ RL are comprised of
values of some smooth functions and, therefore, have low complexity. Usually, efficient representations of such vectors are achieved by applying some
orthogonal transform H (e.g., Fourier or wavelet transform), however, we
cannot apply this transform to the original data tensor for two reasons.
First, the errors in the model are not Gaussian, so application of H will
convert the data tensor with independent Bernoulli components into a data
tensor with dependent entries that are not Bernoulli variables any more. In
addition, application of this transform to the original data will not achieve
8
M. PENSKY
our goals since, although vectors Gk1 ,k2 ,∗ represent smooth functions, vectors Λi,j,∗ do not, due to possible switches in the group memberships. In
addition, for every l, matrix Λ∗,∗,l in (3) forms the so called bi-clustering
structure (see, e.g., Gao et al. (2016)) which makes recovery of G∗,∗,l much
harder than in the case of a usual regression model.
In order to handle all these intrinsic difficulties, we apply operation of
vectorization to Λ∗,∗,l . Denote
(4)
λ(l) = vec(Λ∗,∗,l ),
b(l) = vec(B∗,∗,l ),
g(l) = vec(G∗,∗,l ).
Then, Theorem 1.2.22(i) of Gupta and Nagar (2000) yields
(5)
(l)
(l)
λ(l) = (Z̃(l) ⊗ Z̃(l) )g(l) , bi ∼ Bernoulli(λi ), i = 1, · · · , n2 , l = 1, · · · , L.
(l)
Note that bi in (5) are independent for different values of l but not i due
(l)
(l)
to the symmetry. In addition, the values of bi and λi that are corresponding to diagonal elements of matrices B∗,∗,l and Λ∗,∗,l , are equal to zero by
construction. Since all those values are not useful for estimation, we remove
redundant entries from vectors λ(l) and b(l) for every l = 1, · · · , L. Specifically, in (5), we remove the elements in λ(l) and the rows in (Z̃(l) ⊗ Z̃(l) )
(l)
(l)
corresponding, respectively, to Λi1 ,i2 ,l and (Z̃i1 ,∗ ⊗ Z̃i2 ,∗ ) with i1 ≥ i2 . We
denote the reductions of vectors λ(l) , b(l) and matrices (Z̃(l) ⊗ Z̃(l) ) by, respectively, θ (l) , a(l) and C̃(l) obtaining
(6) θ(l) = C̃(l) g(l) ,
(l)
(l)
ai ∼ Bernoulli(θ i ), i = 1, · · · , n(n − 1)/2, l = 1, · · · , L.
(l)
(l′ )
Note that unlike in the case of b(l) , elements ai and ai′ are independent
whenever i 6= i′ or l 6= l′ . The interesting thing here is that matrices C̃(l)
are still clustering matrices, i.e., C̃(l) ∈ M(n(n − 1)/2, m2 ). Indeed, C̃(l) are
binary matrices such that, for i corresponding to (i1 , i2 ) with i1 < i2 and k
(l)
(l)
(l)
corresponding to (k1 , k2 ) in (Z̃i1 ,k1 ⊗ Z̃i2 ,k2 ) one has C̃i,k = 1 if and only if
the nodes i1 ∈ Ωk1 and i2 ∈ Ωk2 .
Observe that although we removed the redundant elements from vectors
(l)
λ and b(l) , we have not done so for the vectors g(l) . Indeed, since matrices
G∗,∗,l are symmetric, the elements of vectors g(l) corresponding to Gk1 ,k2 ,l
and Gk2 ,k1 ,l with k1 6= k2 are equal to each other. For the sake of eliminating
such redundancy (and, hence, the need of tracing the equal elements in
the process of estimation), for indices k corresponding to pairs of classes
(l)
(k1 , k2 ) with k1 > k2 , we remove entries gk from vectors g(l) and denote
the resulting vectors by q(l) . In order an equivalent of the relation (6) still
holds with vectors q(l) instead of g(l) , we add together columns of matrices
DYNAMIC NETWORK MODELS AND GRAPHON ESTIMATION
9
Fig 1. Vectorization of the probability tensor Λ with n = 4, m = 2, N = n(n−1)/2 =
6, M = m(m + 1)/2 = 3 and L = 3. Left panel, top: transforming Λ∗,∗,l into
θ (l) , l = 1, 2, 3. Left panel, middle: Λ∗,∗,1 = Z̃(1) G∗,∗,1 (Z̃(1) )T . Left panel, bottom:
θ (1) = C̃(1) g(1) = C(1) q(1) . In the left panel, redundant elements of Λ are white,
redundant elements of G are yellow. Right panel: θ = Cq.
C̃(l) corresponding to (k1 , k2 ) and (k2 , k1 ) with k1 < k2 , obtaining new
matrices C(l) . It is easy to see that, for every l, since C(l) is obtained from
C̃(l) by adding columns together and since each row of C̃(l) has exactly one
unit element with the rest of them being zeros, C(l) is again a clustering
matrix of size [n(n − 1)/2] × [m(m + 1)/2]. In particular, for indices i and
k corresponding to nodes (i1 , i2 ) and classes (Ωk1 , Ωk2 ) with i1 < i2 and
(l)
k1 ≤ k2 , one has Ci,k = 1 if i1 ∈ Ωk1 and i2 ∈ Ωk2 or i1 ∈ Ωk2 and i2 ∈ Ωk1 ;
(l)
Ci,k = 0 otherwise. The process of vectorization of the model and removing
redundancy is presented in Figure 1.
Using C(l) and q(l) , one can rewrite equations (6) as
(7)
a(l) = θ (l) + ξ (l)
with
θ(l) = C(l) q(l) ,
l = 1, · · · , L,
where C(l) ∈ M(M, N ), θ (l) ∈ RN , q(l) ∈ RM , N = n(n − 1)/2 and M =
(l)
m(m + 1)/2. Here, for every i and l, components ai of vector a(l) are
(l)
(l)
independent Bernoulli variables with P(ai = 1) = θ i , so that components
of vectors ξ (l) are also independent for different values of i or l.
If we had the time-independent SBM (L = 1) and the clustering matrix
were known, equation (7) would reduce estimation of q(1) to the linear re-
10
M. PENSKY
gression problem with independent sub-gaussian (Bernoulli) errors. Since in
(l)
the case of the DSBM, for each i, the elements gi , l = 1, · · · , L, of vector gi
represent the values of a smooth function, we combine vectors in (7) into matrices. Specifically, we consider matrices A, Θ, Ξ ∈ RN ×L and Q ∈ RM ×L
with columns a(l) , θ (l) , ξ (l) and q(l) , respectively. Note that if the group
memberships of the nodes were constant in time, so that C(l) ∈ {0, 1}N ×M
were independent of l, formula (7) would imply
(8)
A = Θ + Ξ,
Θ = ZQ if
C(l) = Z, l = 1, · · · , L.
However, we consider the situations where nodes can switch group memberships in time and (8) is not true.
For this reason, we proceed with further vectorization. We denote a =
vec(A), θ = vec(Θ) and q = vec(Q) and observe that vectors a, θ ∈ RN L
and q ∈ RM L are obtained by stacking vectors a(l) , θ (l) and q(l) in (7)
vertically for l = 1, · · · , L. Define a block diagonal matrix C ∈ {0, 1}N L×M L
with blocks C(l) , l = 1, · · · , L, on the diagonal. Then, (7) implies that
(9)
a=θ+ξ
with
θ = Cq = C vec(Q),
where ai are independent Bernoulli(θ i ) variables, i = 1, · · · , N L.
Observe that if the matrix C were known, then equations in (9) would
represent a regression model with independent Bernoulli errors. Moreover,
(l) T (l)
(l) 2
matrix CT C is diagonal since
q matrices (C ) C = (S ) , l = 1, · · · , L,
(l)
(l)
(l)
are diagonal with Sk1 ,k2 = Nk1 ,k2 , where Nk1 ,k2 is the number of pairs
(i1 , i2 ) of nodes such that i1 < i2 and one node is in class Ωk1 while another
is in class Ωk2 at time instant tl :
(l) (l)
if k1 6= k2 ;
n k1 n k2 ,
(l)
(10)
Nk1 ,k2 =
n(l) (n(l) − 1), if k = k .
1
2
k1
k1
Remark 1. (Directed graph). Similar vectorization algorithm can
be used when the dynamic network is constructed from directed graphs or
graphs with self loops. In the former case, the only redundant entries of
matrices Λ∗,∗,l would be the diagonal ones while, in the latter case, Λ has
no redundant elements and no row removal is necessary.
Remark 2. (Biclustering structures). Vectorization presented above
can significantly simplify the inference in the so called biclustering models
considered, for example, by Lee et al. (2010) and Gao et al. (2016). In those
DYNAMIC NETWORK MODELS AND GRAPHON ESTIMATION
11
models, one needs to recover matrix X from observations of matrix Y given
by Y = U1 XU2 +Ξ where matrices U1 and U2 are known and matrix Ξ has
independent zero-mean Gaussian or sub-gaussian entries. As long as there
are no structural assumptions on matrix X (such as, e.g., low rank), one can
apply vectorization and reduce the problem to the familiar non-parametric
regression problem of the form y = Ux + ξ where matrix U = U1 ⊗ U2
is known, ξ = vec(Ξ) is the vector with independent components and one
needs to recover x = vec(X) from observations y = vec(Y).
3. Assumptions and estimation for the DSBM. It is reasonable to
assume that the values of the probabilities q(l) of connections do not change
dramatically from one time instant to another. Specifically, we assume that
(1)
(L)
for various k = 1, · · · , M , vectors qk = (qk , · · · , qk ) represent values of
(l)
some smooth functions, so that qk = fk (tl ), l = 1, · · · , L. In order to quantify this phenomenon, we assume that vectors qk have sparse representation
in some orthogonal basis H ∈ RL×L with HT H = HHT = IL , so that vector
HqTk is sparse: it has only few large coefficients, the rest of the coefficients
are small or equal to zero. This is a very common assumption in functional
data analysis. For example, if H is the matrix of the Fourier transform and
fk belongs to a Sobolev space or H is a matrix of a wavelet transform and fk
belongs to a Besov space, the coefficients HqTk of qTk decrease rapidly and,
hence, vector HqTk is sparse. In particular, one needs only few elements in
vector HqTk to represent qk with high degree of accuracy. The extreme case
occurs when the connection probabilities do not change in time, so that vector qk has constant components: then, for the Fourier or a periodic wavelet
transform, the vector HqTk has only one non-zero element.
Denote D = QHT where matrix Q is defined in the previous section and
d = vec(D). Observe that vector d is obtained by stacking together the
columns of matrix D = QHT while its transpose DT = HQT has vectors
HqTk as its columns. Then, sparsity of the matrix D can be controlled by
imposing a complexity penalty on kdk0 = kDk0 = kDT k0 on matrix D.
Note that complexity penalty does not require the actual matrix D to have
only few non-zero elements, it merely forces the procedure to keep only few
large elements in D while setting the rest of the elements to zero and, hence,
acts as a kind of hard thresholding. Note that by Theorem 1.2.22 of Gupta
and Nagar (2000), one has
(11)
d = vec(QHT ) = (H ⊗ IM )vec(Q) = (H ⊗ IM )q = Wq,
where W = (H ⊗ IM ) is an orthogonal matrix such that WT W = WWT =
12
M. PENSKY
IM L . Denote
(12)
J ≡ JM = {j : dj 6= 0} , dJ C = 0,
so that J is the set of indices corresponding to nonzero elements of the vector
d.
Consider a set of clustering matrices C(m, n, L) satisfying (2). At this
point we impose very mild assumption on C(m, n, L):
(13)
log(|C(m, n, L)|) ≥ 2 log m.
Assumption (13) is used just for simplifying expression for the penalty. Indeed, until now, we allowed any collection of clustering matrices, so potentially, we can work with the case where all cluster memberships are fixed
in advance (although this would be a totally trivial case). Condition (13)
merely means that at least two nodes at some point in time can be assigned
arbitrarily to any of m classes. Later, we shall consider some special cases
such as fixed membership (no membership switches over time) or limited
change (only at most n0 nodes can change their memberships between two
consecutive time points).
We find m, J, d and C as one of the solutions of the following penalized
least squares optimization problem
b C)
b d,
b = argmin ka − CWT dk2 + Pen(|J|, m)
s.t. dJ c = 0
(14) (m,
b J,
m,J,d,C
where C ∈ C(m, n, L), a is defined in (9), d ∈ RM L , W ∈ RM L×M L , M =
m(m + 1)/2 and
11
25 m2 L
(15)
Pen(|J|, m) = 11 log(|C(m, n, L)|) +
|J| log
.
2
|J|
Observe that the penalty in (15) consists of two parts. The first part accounts for the complexity of clustering and, therefore, allows one to obtain
an estimator adaptive to the number of unknown groups m as long as the
we can express the complexity of clustering in terms of m, n and L. The
second term represents the price of estimating |J| elements of vector d and
finding those |J| elements in this vector of length m(m + 1)L/2.
Note that since minimization is carried out also with respect to m, optimization problem (14) should be solved separately for every m = 1, · · · , n,
bM , C
b M and JbM . After that, one needs to select the value M
c=
yielding d
m(
b m
b + 1)/2 that delivers the minimum in (14), so that
(16)
b=d
b c,
d
M
b =C
b c,
C
M
Jb = JbM
c.
DYNAMIC NETWORK MODELS AND GRAPHON ESTIMATION
13
c = (H ⊗ I c) and calculate
Finally, due to (12), we set W
M
c T d,
b
b=W
q
(17)
b = Cb
b q.
θ
b into the tensor and taking the symmetries
b by packing vector θ
We obtain Λ
into account.
4. Oracle inequalities for the DSBM. Denote the true value of tensor Λ by Λ∗ . Also, denote by m∗ the true number of groups, by q∗ and θ ∗
the true values of q and θ in (9) and by C∗ the true value of C. Denote
by D∗ and d∗ the true values of matrix D and vector d, respectively. Let
M ∗ = m∗ (m∗ + 1)/2 and W∗ = (H ⊗ IM ∗ ) be true values of M and W.
Note that vector θ ∗ is obtained by vectorizing Λ∗ and then removing the
redundant entries. Then, it follows from (9) that
(18)
a = θ∗ + ξ
with
θ ∗ = C∗ q∗ = C∗ (W∗ )T d∗ .
Due to the relation between the ℓ2 and the Frobenius norms, one has
kθ − θ ∗ k2 ≤ kΛ − Λ∗ k2 ≤ 2kθ − θ ∗ k2 ,
(19)
and the following statement holds.
Theorem 1. Consider a DSBM with a true matrix of probabilities Λ∗
b obtained according to (14)–(17). Let C(m, n, L) be a
and the estimator Λ
set of clustering matrices satisfying conditions (2) and (13). Then, for any
t > 0, with probability at least 1 − 9e−t , one has
(20)
b − Λ∗ k 2
kΛ
≤
n2 L
min
m,J,d
C∈C(m,n,L)
"
#
6 kCWT d(J) − θ∗ k2
38 t
4 Pen(|J|, m)
+ 2
+
2
2
n L
n L
n L
"
#
6 kCWT d(J) − θ∗ k2
4 Pen(|J|, m)
342
,
+
+ 2
n2 L
n2 L
n L
and
(21)
E
b − Λ∗ k 2
kΛ
n2 L
!
≤
min
m,J,d
C∈C(m,n,L)
where d(J) is the modification of vector d where all elements dj with j ∈
/J
are set to zero.
The proof of Theorem 1 is given in the Supplementary Material. Here,
we just explain its idea. Note that if the values of m and C are fixed, the
problem (14) reduces to a regression problem with a complexity penalty
b of d∗ is just a
Pen(|J|, m). Moreover, if J is known, the optimal estimator d
14
M. PENSKY
projection estimator. Indeed, denote ΥC = CWT and let ΥC,J = (CWT )J
b
be the reduction of matrix CWT to columns j ∈ J. Given m,
b Jb and C,
T
c
c
one obtains M = m(
b m
b + 1)/2, W = (H ⊗ IM
c), ΥC,J = (CW )J and
T
b
c
b
ΥC,
b Jb = (CW )Jb. Let
b TC,
b TC,
b b b)−1 Υ
b b b(Υ
bb b=Υ
(22) ΠC,J = ΥC,J (ΥTC,J ΥC,J )−1 ΥTC,J , Π
b Jb
b JbΥ
C,J
C,J
C,J
b b b, respecbe the projection matrices on the column spaces of ΥC,J and Υ
C,J
b is of the
b=Π
b b b a and vector d
b b bd
tively. Then, it is easy to see that Υ
C,J
form
C,J
b = (Υ
b Tb bΥ
b b b)−1 Υ
b Tb b a.
d
C,J C,
C,J
J
(23)
b can be obtained as a solution of the following
Hence, the values of m,
b Jb and C
optimization problem
b m,
(C,
b Jb) = argmin ka − ΠC,J ak2 + Pen(|J|, m)
s.t. C ∈ C(m, n, L),
m,J,C
where ΠC,J and Pen(|J|, m) are defined in (22) and (15), respectively. After
that, we use the arguments that are relatively standard in the proofs of
oracle inequalities for the penalized least squares estimators.
Note that kCWT d(J) − θ ∗ k2 in the right-hand sides of expressions (20)
and (21), is the bias term that quantifies how well one can estimate the
true values of probabilities θ ∗ by blocking them together, averaging the
values in each block and simultaneously setting all but |J| elements of vector
d to zero. If |J| is too small, then d will not be well represented by its
truncated version d(J) and the bias will be large. The penalty represents the
stochastic error and constitutes the ”price” for choosing too many blocks
and coefficients. In particular, the second term (11/2) |J| log 25 m2 L/|J|
in (15) is due to the need of finding and estimating |J| elements of the
Lm(m + 1)/2-dimensional vector. The first term, log(|C(m, n, L)|), accounts
for the difficulty of clustering and is due to application of the union bound
in probability.
Theorem 1 holds for any collection C(m, n, L) of clustering matrices satisfying assumption (13). In order to obtain some specific results, denote by
Z(m, n, n0 , L) the collection of clustering matrices corresponding to the situation where at most n0 nodes can change their memberships between any
two consecutive time points, so that
L−1
n
n
n0
(24)
|Z(m, n, n0 , L)| = m
m
,
n0
DYNAMIC NETWORK MODELS AND GRAPHON ESTIMATION
15
yielding |Z(m, n, 0, L)| = mn and |Z(m, n, n, L)| = mnL . Note that the case
of n0 = 0 corresponds to the scenario where the group memberships of the
nodes are constant and do not depend on time while the case of n0 = n
means that memberships of all nodes can change arbitrarily from one time
instant to another. Since
n
mne
,
log
mn0 ≤ n0 log
n0
n0
formulae (15) and (24) immediately yield the following corollary.
Corollary 1. Consider a DSBM with a true matrix of probabilities
b obtained according to (14)–(17) where C(m, n, L) =
Λ and estimator Λ
Z(m, n, n0 , L). Then, inequalities (20) and (21) hold with
∗
(25)
|J|
25 m2 L
mne
+
log
Pen(|J|, m) = 11 n log m + n0 (L − 1) log
n0
2
|J|
It is easy to see that the first term in (25) accounts for the uncertainty
of the initial clustering, the second term is due to the changes in the group
memberships of the nodes over time (indeed, if n0 = 0, this term just vanishes) while the last term is identical to the second term in the expression
for the generic penalty (15). While we elaborate only on the special case
where the collection of clustering matrices is given by (24), one can easily
produce results similar to Corollary 1 for virtually any nodes’ memberships
scenario.
Remark 3. (The SBM). Theorem 1 provides an oracle inequality in
the case of a time-independent SBM (L = 1). Indeed, in this case, by taking
H = 1 and W = IM , obtain for any t > 0
"
b − Λ∗ k2
6kCq(J) − θ ∗ k2 44 log m
EkΛ
(26)
≤
min
+
m,J,q
n2
n2
n
C∈M(m,n)
25 m2
22|J|
342
+
log
+ 2
2
n
|J|
n
and a similar result holds for the probability. Note that if |J| = m(m + 1)/2,
our result coincides with the one of Gao et al. (2015). However, if many
groups have zero probability of connection, then |J| is small and the righthand of (26) can be asymptotically smaller than n−1 log m+n−2 m2 obtained
in Gao et al. (2015). In addition, our oracle inequality is non-asymptotic and
the estimator is naturally adaptive to the unknown number of classes. (Gao
et al. (2016) obtained adaptive estimators but not via an oracle inequality).
16
M. PENSKY
Corollary 1 quantifies the stochastic error term in Theorem 1. The size of
the bias depends on the level of sparsity of coefficients of functions qk in the
basis H and on the constitution of classes. While one can study a variety of
scenarios, in order to be specific, we consider the case of a balanced network
model where the sizes of all the classes are proportional to each other, in
particular, for some absolute constants 0 < ℵ1 ≤ 1 ≤ ℵ2 < ∞, one has
(27)
ℵ1
n
n
(l)
≤ nk ≤ ℵ2 ,
m
m
k = 1, . . . , m, l = 1, . . . , L,
(l)
where nk the number of nodes in class k at the moment tl .
Note that the condition (27) is very common in studying random network
models (see, e.g., Gao et al. (2017) or Amini and Levina (2018) among
others). In addition, if class memberships are generated from the multinomial
distribution with the vector of probabilities (π1 , · · · , πm ), and C1 /m ≤ πi ≤
C2 /m for some constants 0 < C1 < C2 < ∞, as it is done in, e.g., Bickel
and Chen (2009), condition (27) holds with high probability.
In particular, we consider networks that satisfy condition (27) but yet allow only n0 nodes switch their memberships between time instances. We denote the corresponding set of clustering matrices by Zbal (m, n, n0 , L, ℵ1 , ℵ2 ).
It would seem that condition (27) should make clustering much simpler.
However, as Lemma 1 below shows, this reduction does not makes estimation significantly easier since the complexity of the set of balanced clustering
matrices log |Zbal (m, n, n0 , L, ℵ1 , ℵ2 )| is smaller than the complexity of the
set of unrestricted clustering matrices log |Z(m, n, n0 , L)| only by, at most,
a constant factor.
p
Lemma 1. (Balanced network model complexity) If n ≥ e n30 ,
then
1
mne
(28) log |Zbal (m, n, n0 , L, ℵ1 , ℵ2 )| ≥
.
n log m + (L − 1) n0 log
4
n0
Then, one can use the same penalty that was considered in Corollary 1,
so that Theorem 1 yields the following result.
Theorem 2. Consider a balanced DSBM satisfying condition (27). Let
Λ∗ be the true matrix of probabilities, m∗ be the true number of classes,
M ∗ = m∗ (m∗ + 1)/2, Q∗ be the true matrix p
of probabilities of connections
b is
for pairs of classes and D∗ = Q∗ H. If n ≥ e n30 and the estimator Λ
obtained as a solution of optimization problem (14) with the penalty (25)
DYNAMIC NETWORK MODELS AND GRAPHON ESTIMATION
17
where
(29)
J=
M
[
Jk ,
k=1
then, for any t > 0, with probability at least 1 − 9e−t , one has
M∗ X
∗
2
∗
2
X
b
4Pen(|J|, m )
38t
6ℵ2
kΛ − Λ k
∗ 2
≤
min
(D
)
+
+ 2
(30)
k,l
2
∗
2
2
J
n L
(m ) L
n L
n L
k=1 l∈
/ Jk
and a similar result holds for the expectation.
In order to obtain specific upper bounds in (30), we need to impose some
assumptions on the smoothness of functions Q∗k,∗ , k = 1, · · · , M ∗ . For the
sake of brevity, we assume that all vectors D∗k,∗ , k = 1, · · · , M ∗ , behave
similarly with respect to the basis H (generalization to the case where this
is not true is rather pedestrian but very cumbersome as we point out in
Section 8, Discussion).
(A0). There exist absolute constants ν0 and K0 such that
(31)
L
X
(l − 1)2ν0 (D∗k,l )2 ≤ K0 ,
l=1
k = 1, · · · , M ∗ .
Let conditions of Theorem 2 hold and D∗k,∗ satisfy asb is obtained as a solution of optimization
sumption (31). If the estimator Λ
problem (14), then for any t > 0, with probability at least 1 − 9e−t , one has
Corollary 2.
"
# 2ν0
∗ 2
2
2ν0 +1
∗ 2
∗
b
kΛ − Λ k
m
m
n
1
≤ K̃0 min
log
,
L
n2 L
n
m∗
n
∗
log m∗ n0
m ne
t
+
+ 2 log
+ 2
(32)
nL
n
n0
n L
and a similar result holds for the expectation. Here, K̃0 is an absolute constant that depends on K0 , ν0 , ℵ1 and ℵ2 only.
5. The lower bounds for the risk for the DSBM. In order to prove
that the estimator obtained as a solution of optimization problem (14) is
minimax optimal, we need to show that the upper bounds in Corollaries 1
18
M. PENSKY
and 2 coincide with the minimax lower bounds obtained under similar constraints. For the sake of derivation of lower bounds for the error, we impose
mild conditions on the orthogonal matrix H as follows: for any binary vector
ω ∈ {0, 1}L one has
√
√
(33)
kHT ωk∞ ≤ kωk1 / L and H 1 = Le1 ,
where 1 = (1, 1, · · · , 1)T and e1 = (1, 0, · · · , 0)T . Assumptions (33) are not
restrictive. In fact, they are satisfied for a variety of common orthogonal
transforms such as the Fourier transform or a periodic wavelet transforms.
First, we derive the lower bounds for the risk under the assumption that
vector d is l0 -sparse and has only s nonzero components. Let Gm,L,s be a
collection of tensors such that G ∈ Gm,L,s implies that the vectorized versions
q of G can be written as q = WT d with kdk0 ≤ s. In order to be more
specific, we consider the collection of clustering matrices Z(m, n, n0 , L) with
cardinality given by (24) that corresponds to the situation where at most
n0 nodes can change their memberships between consecutive time instants.
In this case, Pen(|J|, m) is defined in (25).
Theorem 3. Let orthogonal matrix H satisfy condition (33). Consider
the DSBM where G ∈ Gm,L,s with s ≥ κm2 where κ > 0 is independent
of m, n and L. Denote γ = min(κ, 1/2) and assume that L ≥ 2, n ≥ 2m,
n0 ≤ min(γn, 4/3 γn m−1/9 ) and s is such that
s2 log(2LM/s) ≤ 68LM n2 .
(34)
Then
(35)
inf
c
Λ
sup
G∈Gm,L,s
C∈Z(m,n,n0 ,L)
PΛ
(
b − Λk2
kΛ
≥ C(γ)
n2 L
s log(Lm2 /s)
+
n2 L
≥
log m n0
+ 2 log
nL
n
mne
n0
1
,
4
b is any estimator of Λ, P is the probability under the true value
where Λ
Λ
of the tensor Λ and C(γ) is an absolute constant that depends on γ only.
Theorem 3 ensures that if vector d has only s nonzero components, then
the upper bounds in Corollary 1 are optimal up to a constant. In order to
provide a similar assertion in the case of Corollary 2, we assume that rows of
matrix D are l2 -sparse. For this purpose, we consider a collection of tensors
Gm,L,ν0 such that G ∈ Gm,L,ν0 implies that Q = DH and rows Dk,∗ of matrix
DYNAMIC NETWORK MODELS AND GRAPHON ESTIMATION
19
D satisfy condition (31). Let as before Zbal (m, n, n0 , L, ℵ1 , ℵ2 ) be a collection of clustering matrices satisfying condition (27) and such that at most
n0 nodes change their memberships between two consecutive time instances.
The following statement ensures that the upper bounds in Corollary 2 are
minimax optimal up to a constant factor.
Theorem 4. Let orthogonal matrix H satisfy condition (33). Consider
the DSBM where G ∈ Gm,L,ν0 with ν0 > 1/2, L ≥ 2 and n ≥ 2m. Then, for
any absolute constants 0 < ℵ1 ≤ 1 ≤ ℵ2 < ∞, one has
(
"
(
)
2ν0
b − Λk2
kΛ
1 m 2 2ν0 +1 m 2
;
(36) inf sup PΛ
≥ C min
n2 L
L
n
n
c G∈Gm,L,s
Λ
C∈Z
bal
+
log m n0
+ 2 log
nL
n
mne
n0
1
≥ ,
4
b is any estimator of Λ, P
where Zbal stands for Zbal (m, n, n0 , L, ℵ1 , ℵ2 ), Λ
Λ
is the probability under the true value of the tensor Λ and C is an absolute
constant independent of n, m and L.
Theorems 3 and 4 confirm that the estimator constructed above is minimax optimal up to a constant if G ∈ Gm,L,s and C ∈ Z(m, n, n0 , L), or
G ∈ Gm,L,ν0 and C ∈ Zbal (m, n, n0 , L, ℵ1 , ℵ2 ).
Note that the terms log m/(nL) and n0 n−2 log(mne/n0 ) in (35) and (36)
correspond to, respectively, the error of initial clustering and the clustering
error due to membership changes. The remaining terms are due to nonparametric estimation and model selection. Assumptions (33) and (34) are purely
technical and are necessary to ensure that the “worst case scenario” tensor G
of connection probabilities has nonnegative components. As we mentioned
earlier, conditions (33) are totally non-restrictive. Condition (34) in Theorem 3 holds whenever representation of the tensor of probabilities in the
basis H is at least somewhat sparse. Indeed, if there is absolutely no sparsity
(which is a very implausible scenario when smooth functions are represented
in a basis) and s ≈ M L, then condition (34) reduces to m(m + 1)L ≤ Cn2
and will still be true if L is relatively small. If L is large, the situation where
s ≈ M L is very unlikely. Assumption that s ≥ κm2 for some κ > 0 independent of m, n and L, restricts the sparsity level and ensures that one does
not have too many classes where nodes have no interactions with each other
or members of other classes.
Finally, it is also worth keeping in mind that all assumptions in Theorems 3 and 4 are used for the derivation of the minimax lower bounds for
20
M. PENSKY
the risk and are not necessary for either the construction of the estimator
b of Λ in (14) or for the assessment of its precision in Theorems 1 and 2.
Λ
6. The uniformly sparse DSBM. In the current literature, the notion of the sparse SBM refers to the case where the entries of the matrix of the connection probabilities are uniformly small: Λ = ρn Λ(0) with
kΛ(0) k∞ = 1 and ρn → 0 as n → ∞. The concept is based on the idea
that when the number of nodes in a network grow, the probabilities of connections between them decline. The minimax study of the sparse SBM has
been carried out by Klopp et al. (2017). The logical generalization of the
sparse SBM of this type would be the sparse DSBM where the elements of
the tensor Λ are bounded above by ρn where ρn → 0 as n grows. We refer
to this kind of network as uniformly sparse.
On the other hand, not all networks become uniformly sparse as n → ∞.
Indeed, in the real world, when a network grows, the number of communities
increase and, while the probabilities of connections for majority of pairs of
groups become very small, some of the of pairs groups will still maintain high
connection probabilities. We refer to this type of network as non-uniformly
sparse. The idea of such a network has been elaborated in the recent paper of
Borgs et al. (2016). The authors considered heavy-tailed sparse graphs such
that, in the context of the SBM, one still has Λ = ρn Λ(0) but the elements
of Λ(0) are no longer bounded by one but by a quantity that grows with n.
While distinguishing between very small probabilities might be essential
in a clustering problem, it is not so necessary in the problem of estimation
of the tensor of the connection probabilities studied in the present paper.
Indeed, it is a common knowledge that, in the nonparametric regression
model, in order to obtain the best error rates, one needs to replace small
elements of the vector of interest by zeros rather than estimating them.
Similarly, if the network is non-uniformly sparse, i.e., some pairs of groups
have probabilities of connections equal or very close to zero, one would
obtain an estimator with better overall precision by setting those very small
connection probabilities to zeros. Although nowhere in the present paper we
make an assumption that a network is sparse and, moreover, consideration
of the non-uniformly sparse SBM or DSBM is not one of its objectives, the
paper naturally provides the tools for minimax optimal statistical estimation
in such models that deliver results with very little additional work.
In addition, the techniques developed in this paper allow, with some additional work, to extend results obtained in Klopp et al. (2017) to the dynamic setting. However, majority of their results depend upon solution of
optimization problem (14) under the restriction that kWT dk∞ ≤ ρn which
DYNAMIC NETWORK MODELS AND GRAPHON ESTIMATION
21
requires representation of the estimator via a different projection operator
and will result in more cumbersome calculations. Therefore, we avoid studying this new optimization problem and only extend Corollary 2.2 of Klopp
et al. (2017) that handles the case of the balanced model without placing
the above-mentioned restriction. For this purpose, consider a small ρn and
denote
(37)
rn (m) = max(ρn , m2 /n2 ).
Similarly to (14), we find m, J, d and C as one of the solutions of the following penalized least squares optimization problem
b C)
b d,
b = argmin ka − CWT dk2 + λ0 rn (m)Pen(|J|, m) s.t. dJ c = 0
(38) (m,
b J,
m,J,d,C
where C ∈ Zbal (m, n, n0 , L, ℵ1 , ℵ2 ), a is defined in (9), d ∈ RM L , W ∈
RM L×M L , M = m(m + 1)/2, Pen(|J|, m) is defined in (25) and λ0 is a tuning parameter that is bounded above and below by a constant.
In order the estimator has the uniform sparsity property, we need to make
sure that transformation H is such that, whenever it is used for sparse representation of smooth functions, the maximum absolute value of the estimator
obtained by truncation of the vector of coefficients is bounded above by a
constant factor of the maximum absolute value of the original function. In
particular, we denote the projection matrix on the column space of matrix
(CWT )J by ΠC,J and impose the following condition on the transformation
matrix H:
(A1). There exists an absolute constant B0 such that for any C ∈
Zbal (m, n, n0 , L, ℵ1 , ℵ2 ) and any vector θ
(39)
kΠ⊥
C,J θk∞ = kθ − ΠC,J θk∞ ≤ B0 kθk∞ .
Let, as before, Λ∗ be the true matrix of probabilities, m∗ be the true
number of classes, M ∗ = m∗ (m∗ + 1)/2, C∗ be the true clustering matrix,
Q∗ be the true matrix of probabilities of connections for pairs of classes,
D∗ = Q∗ H, d∗ = vec(D∗ ), θ ∗ = C∗ (W∗ )T d∗ and W∗ = H ⊗ IM ∗ .
Theorem 5. Consider a balanced DSBM satisfying condition (27). Let
∗
∗
matrix H be such
p that condition (39) is satisfied and kΛ k∞ ≤ ρn . If
∗
3
b is obtained as a solution of opρn ≥ ρn , n ≥ e n0 and the estimator Λ
timization problem (38), then, for an absolute constant C̃0 and any t > 0,
22
M. PENSKY
with probability at least 1 − 9e−t , one has
)
(
∗ 2
b − Λ∗ k2
kΠ⊥
rn (m∗ ) [Pen(|J|, m∗ ) + t]
kΛ
C∗ ,J θ k
≤ C̃0 min
+
(40)
J
n2 L
n2 L
n2 L
where ΠC∗ ,J is the projection matrix on the column space of (C∗ W∗T )J and
C̃0 is an absolute constant that depends on B0 , ℵ1 and ℵ2 only.
In particular, if condition (31) holds with K0 replaced with ρ∗n K0 , then
"
# 2ν0
∗ 2
2
2ν0 +1
∗ 2
∗
b
m
kΛ − Λ k
m
n
1
,
≤ K̃0 rn (m∗ ) min
log
2
∗
L
n L
n
m
n
m∗ ne
t
log m∗ n0
+ 2 log
+ 2
(41)
+
nL
n
n0
n L
Here, K̃0 is an absolute constant that depends on B0 , K0 , ν0 , ℵ1 and ℵ2
only. Results similar to (40) and (41) hold for the expectations.
7. Dynamic graphon estimation. Consider the situation where tensor Λ is generated by a dynamic graphon f , so that Λ is given by expression
(1) where function f : [0, 1]3 → [0, 1] is such that f (x, y, t) = f (y, x, t) for
any t and ζ = (ζ1 , · · · , ζn ) is a random vector sampled from a distribution
Pζ supported on [0, 1]n .
Given an observed adjacency tensor B sampled according to model (1),
the graphon function f is not identifiable since the topology of a network is
invariant with respect to any change of labeling of its nodes. Therefore, for
any f and any measure-preserving bijection µ : [0, 1] → [0, 1] (with respect
to Lebesgue measure), the functions f (x, y, t) and f (µ(x), µ(y), t) define the
same probability distribution on random graphs. For this reason, we are
considering equivalence classes of graphons. Note that in order for it to be
possible to compare clustering of nodes across time instants, we introduce
an assumption that there are no label switching in time, that is, every node
carries the same label at any time tl , so that function µ is independent of t.
Under this condition, we further assume that probabilities Λi,j,l do not
change drastically from one time point to another, i.e. that, for every x and
y, functions f (x, y, t) are smooth in t. We shall also assume that f is piecewise smooth in x and y. In order to quantify those assumptions, for each
x, y ∈ [0, 1]2 , we consider a vector f (x, y) = (f (x, y, t1 ), · · · , f (x, y, tL ))T and
an orthogonal transform H used in the previous sections. We assume that
elements vl (x, y) of vector v(x, y) = Hf (x, y) satisfy the following assumption:
DYNAMIC NETWORK MODELS AND GRAPHON ESTIMATION
23
(A2). There exist constants 0 = β0 < β1 < · · · < βr = 1 and ν1 , ν2 , K1 , K2 >
0 such that for any x, x′ ∈ (βi−1 , βi ] and y, y ′ ∈ (βj−1 , βj ], 1 ≤ i, j ≤ r, one
has
(42)
(43)
[vl (x, y) − vl (x′ , y ′ )]2 ≤ K1 [|x − x′ | + |y − y ′ |]2ν1 ,
L
X
(l − 1)2ν2 vl2 (x, y) ≤ K2 .
l=1
Note that, for a graphon corresponding to the DSBM model, on each of
the rectangles (βi−1 , βi ] × (βj−1 , βj ], functions vl (x, y) are constant, so that
vl (x, y) = 0 for l = 2, · · · , L, and ν1 = ∞.
We denote the class of graphons satisfying assumptions (1), (42) and
(43) by Σ(ν1 , ν2 , K1 , K2 ). In order to estimate the dynamic graphon, we
approximate it by an appropriate DSBM and then estimate the probability
tensor of the DSBM. Note that, since ν1 , ν2 , K1 and K2 in Assumption A
are independent of x and y, one can simplify the optimization procedure in
(14).
Let Q be the matrix defined in (8) and (9). Note that since random
variables ζ1 , . . . , ζn are time-independent, we can approximate the graphon
by a DSBM where group memberships of the nodes do not change in time.
Hence, , matrices C(l) are independent of l, so that C(l) = Z, (8) holds
and Θ = ZQ. Denote X = AHT . Denote by V and Φ the matrices of
the coefficients of Q and Θ in the transform H: V = QHT and Φ = ΘHT .
Then, by (8), Θ = ZVH and Φ = ZV. Note that each row of the matrices V
and Φ corresponds to one spatial location. Since, due to (43), the coefficients
in the transform H decrease uniformly irrespective of the location, one can
employ L1 < L columns instead of L columns in the final representations of
Q and Θ. In order to simplify our presentation, we denote L1 = Lρ where
0 < ρ ≤ 1 and use the optimization procedure (14) to find m, ρ, V(ρ) and Z
where V(ρ) is the submatrix of V with columns V∗,j , 1 ≤ j ≤ Lρ . Due to
|J| = M Lρ and 0.5 m2 Lρ ≤ |J| ≤ m2 Lρ , in this case optimization problem
(14) can be reformulated as
h
b (bρ) , Z)
b = argmin
(44)
kX(ρ) − ZV(ρ) k2
(m,
b ρb, V
m,ρ,V(ρ)
Z∈Z(m,n,0,L)
11 2 ρ
1−ρ
+11n log m + m L log(25L )
2
where Z(m, n, 0, L) is defined in (24). Then the estimation algorithm appears
as follows:
24
M. PENSKY
1. Apply transform H to the data matrix A obtaining matrix X = AHT .
2. Consider a set ℜ = {ρ ∈ (0, 1] : Lρ is an integer}. For every ρ ∈ ℜ,
remove all columns X∗,l with l ≥ Lρ + 1 obtaining matrix X(ρ) with
EX(ρ) = ZV(ρ) ≡ Φ(ρ) where matrix V(ρ) has Lρ columns.
b (bρ) , Z)
b as a solution of the optimization problem (44).
3. Find (m,
b ρb, V
(b
b
b
b
b by packing Θ
b into a tensor.
4. Choose Θ = ZV ρ) H and obtain Λ
b does not require knowledge of
Note that construction of the estimator Λ
ν1 , ν2 , K1 and K2 , so the estimator is fully adaptive. The following statement
b
provides a minimax upper bound for the risk of Λ.
Theorem 6. Let Σ ≡ Σ(ν1 , ν2 , K1 , K2 ) be the class of graphons satb is obtained as a solution of
isfying Assumptions (1), (42) and (43). If Λ
optimization problem (44) as described above, then
(45)
b − Λ∗ k2
EkΛ
sup
≤ C min
1≤h≤n−r
n2 L
f ∈Σ
0≤ρ≤1
Lρ−1 I(ρ < 1)
+ 2ρν2 +1
h2ν1
L
(h + r)2 (1 + (1 − ρ) log L) log(h + r)
+
+
n2 L1−ρ
nL
,
where the constant C in (45) depends on ν1 , ν2 , K1 and K2 only.
Note that h in (45) stands for h = m − r where m is the number of
blocks in the DSBM which approximates the graphon, hence, h ≤ n − r.
On the other hand, h ≥ 0 since one needs at least r blocks to approximate
the graphon that satisfies condition (42). Since the expression in the right
hand side of (45) is rather complex and is hard to analyze, we shall consider
only two regimes: a) r = rn,L ≥ 2 may depend on n and L and ν1 = ∞; or
b) r = r0 ≥ 1 is a fixed quantity independent of n and L. The first regime
corresponds to a piecewise constant (in x and y) graphon that generates
the DSBM while the second regime deals with the situation where f is a
piecewise smooth function of all three arguments with a finite number of
jumps. In the first case, we set h = 2, in the second case, we choose h to be
a function of n and L. By minimizing the right-hand side of (45), we obtain
the following statement.
b be obtained as a solution of optimization problem
Corollary 3. Let Λ
(44) as described above. Then, for Σ ≡ Σ(ν1 , ν2 , K1 , K2 ) and C independent
25
DYNAMIC NETWORK MODELS AND GRAPHON ESTIMATION
of n and L, one has
(46)
(
h
1
C
min
L
r 2
n
log
n
r
2
i 2ν2ν2 +1
;
r 2
n
)
+
C log r
nL ,
b −Λ k
EkΛ
≤
sup
)
(
n2 L
2ν1 ν2
f ∈Σ
1
(ν +1)(2ν
ν ν+1
1
1
2 +1)
log
L
log
L
1
+
; n2
C min L n2
∗ 2
r = rn,L ;
C log n
nL ,
r = r0 .
In order to assess optimality of the penalized least squares estimator obtained above, we derive lower bounds for the minimax risk over the set
Σ(ν1 , ν2 , K1 , K2 ). These lower bounds are constructed separately for each of
the two regimes.
Theorem 7. Let matrix H satisfy assumptions (33) and ν2 ≥ 1/2 in
(43). Then, for C independent of n and L, one has
)
(
b − Λk2
1
kΛ
≥ ∆(n, L) ≥ ,
sup
PΛ
(47)
inf
2
n L
4
c f ∈Σ(ν1 ,ν2 ,K1 ,K2 )
Λ
where
(48)
∆(n, L) =
(
h
1
C
min
L
C min L1
r 2
n
1
n2
i
2ν2
2ν2 +1
;
2ν1 ν2
(ν1 +1)(2ν2 +1)
r 2
n
;
)
1
n2
+
C log r
nL ,
ν1
ν1 +1
+
r = rn,L ;
C log n
nL ,
r = r0 .
It is easy to see that the value of ∆(n, L) coincides with the upper bound in
(46) up to a at most a logarithmic factor of n/r or L. In both cases, the first
quantities in the minimums correspond to the situation where f is smooth
enough as a function of time, so that application of transform H improves
estimation precision by reducing the number of parameters that needs to
be estimated. The second quantities represent the case where one needs to
keep all elements of vector d and hence application of the transform yields
no benefits. The latter can be due to the fact that ν2 is too small or L is too
low.
The upper and the lower bounds in Theorems 6 and 7 look somewhat
similar to the ones appearing in anisotropic functions estimation (see, e.g.,
Lepski (2015)). Note also that although in the case of a time-independent
graphon (L = 1), the estimation precision does not improve if ν1 > 1, this is
26
M. PENSKY
not true any more in the case of a dynamic graphon. Indeed, the right-hand
sides in (48) become significantly smaller when ν1 , ν2 or L grow.
Remark 4. (The DSBM and the dynamic graphon). Observe that
the definition (1) of the dynamic graphon assumes that vector ζ is independent of t. This is due to the fact that, to the best of our knowledge, the
notion of the dynamic graphon with ζ being a function of time has not yet
been developed by the probability community. For this reason, we restrict
our attention to the case where we are certain that, at any time point, the
graphon describes the limiting behavior of the network as n → ∞. Nevertheless, we believe that when the concept of the dynamic graphon is established,
our techniques will be useful for its estimation.
In the case of a piecewise constant graphon, our setting corresponds to
the situation where the nodes of the network do not switch their group
memberships in time, so that n0 = 0 in (24). Therefore, a piecewise constant graphon (r = rn,L , ν1 = ∞) is just a particular case of the general
DSBM since the latter allows any temporal changes of nodes’ memberships.
However, the dynamic piecewise constant graphon formulation enables us to
derive specific minimax convergence rates for estimators of Λ in terms of n,
L and r. On the other hand, the piecewise smooth graphon (r = r0 , ν1 < ∞)
is an entirely different object that is not represented by the DSBM.
8. Discussion. In the present paper we considered estimation of connection probabilities in the context of dynamic network models. To the best
of our knowledge, this is the first paper to propose a fully non-parametric
model for the time-dependent networks which treats connection probabilities
for each group as the functional data and allows to exploit the consistency
in the group memberships over time. The paper derives adaptive penalized
least squares estimators of the tensor of the connection probabilities in a
non-asymptotic setting and shows that the estimators are indeed minimax
optimal by constructing the lower bounds for the risk. This is done via vectorization technique which is very useful for the task in the paper and can
be very beneficial for solution of other problems such as, e.g., inference in
bi-clustering models mentioned in Remark 2. In addition, we show that the
correct penalty consists of two parts: the portion which accounts for the
complexity of estimation and the portion which accounts for the complexity
of clustering and is proportional to the logarithm of the cardinality of the
set of clustering matrices. The latter is a novel result and it is obtained by
using the innovative Packing lemma (Lemma 4) which can be viewed as a
version of the Varshamov-Gilbert lemma for clustering matrices. Finally, the
methodologies of the paper allow a variety of extensions.
DYNAMIC NETWORK MODELS AND GRAPHON ESTIMATION
27
1. (Inhomogeneous or non-smooth connection probabilities). Assumption (43) essentially implies that probabilities of connections are
spatially homogeneous and are represented by smooth functions of
time that belong to the same Sobolev class. The model, however, can
be easily generalized. First, by letting H be a wavelet transform and
assuming that for any fixed x and y, function f (x, y, ·) belongs to a
Besov ball, one can accommodate the case where f (x, y, ·) has jump
discontinuities. Furthermore, by using a weaker version of condition
(43), similarly to how this was done in Klopp and Pensky (2015), we
can treat the case where functions f (x, y, t) are spatially inhomogeneous.
2. (Time-dependent number of nodes). One can apply the theory
above even when the number of nodes in the network changes from
one time instant to another. Indeed, in this case we can form a set
that includes all nodes that have ever been in the network and denote
their number by n. Consider a class Ω0 such that all nodes in this class
have zero probability of interaction with each other or any other node
in the network. At each time instant, place all nodes that are not in
the network into the class Ω0 . After that, one just needs to modify
the optimization procedures by placing additional restriction that the
out-of-the-network nodes indeed belong to class Ω0 and that G0,k,l = 0
for any k = 0, 1, 2, · · · , m and l = 1, · · · , L.
3. (Adaptivity to clustering complexity). Although, in the case
of the DSBM, our estimator is adaptive to the unknown number of
classes, it requires knowledge about the complexity of the set of clustering matrices. For example, if at most n0 nodes can change their
memberships between two consecutive time points and n0 is a fixed
quantity independent of n and m, we can replace n0 by log n that
dominates n0 if n is large enough. However, if n0 depends on n and
m, development of an adaptive estimator would require an additional
investigation.
SUPPLEMENTARY MATERIAL. Supplement contains proofs of all
statements in the paper
References.
[1] Amini, A.A., Levina, E. (2018) On semidefinite relaxations for the block model. Ann.
Statist., 46, 149–179.
[2] Anagnostopoulos, A., Lacki, J., Lattanzi, S., Leonardi, S., Mahdian, M. (2016) Community Detection on Evolving Graphs. NIPS, 35223530.
28
M. PENSKY
[3] Bickel, P.J., Chen, A. (2009) A nonparametric view of network models and newmangirvan and other modularities. Proc. Nation. Acad. Sciences, 106(50), 21068–21073.
[4] Bickel, P.J., Chen, A., Levina, E. (2011) The method of moments and degree distributions for network models. Ann. Statist., 39, 2280–2301.
[5] Borgs, C., Chayes, J.T., Cohn, H., Ganguly, S. (2016) Consistent nonparametric
estimation of heavy-tailed sparse graphs. ArXiv:1508.06675v2
[6] Durante, D., Dunson, D.B., Vogelstein, J.T. (2015) Nonparametric Bayes Modeling
of Populations of Networks. Journ. Amer. Statist. Assoc., accepted.
[7] Durante, D., Yang, X. Dunson, D.B. (2016) Locally adaptive dynamic networks. Ann.
Appl. Stat., 10, 2203–2232.
[8] Gao, C., Lu, Yu, Zhou, H.H. (2015) Rate-optimal graphon estimation. Ann. Statist.,
43, 2624–2652.
[9] Gao, C., Ma, Z., Zhang, S.Y., Zhou, H.H. (2017) Achieving Optimal Misclassification
Proportion in Stochastic Block Models. Journ. Machine Learn. Research, 18, 1–45.
[10] Gao, C., Lu, Yu, Ma, Z., Zhou, H.H. (2016) Optimal Estimation and Completion of
Matrices with Biclustering Structures. Journ. Machine Learn. Research, 17, 1–29.
[11] Goldenberg, A., Zheng, A.X., Fienberg, S.E., Airoldi, E. M. (2010). A survey of
statistical network models. Found. Trends Mach. Learn., 2, 129–233.
[12] Gupta, A.K., Nagar, D.K. (2000) Matrix Variate Distributions. Chapman&
Hall/CRC Press.
[13] Han, Q., Xu, K.S., Airoldi, E.M. (2015) Consistent estimation of dynamic and multilayer block models. Proceedings of the 32nd International Conference on Machine
Learning, Lille, France, 2015, 1511–1520.
[14] Hsu, D., Kakade, S.M., Zhang, T. (2012) A tail inequality for quadratic forms of
subgaussian random vectors. Electron. Commun. Probab., 17(52), 1–6.
[15] Klopp, O., Pensky, M. (2015) Sparse high-dimensional varying coefficient model: nonasymptotic minimax study. Ann. Stat., 43, 1273–1299.
[16] Klopp, O., Tsybakov, A.B., Verzelen, N. (2017) Oracle inequalities for network models
and sparse graphon estimation. Ann. Statist., 45, 316–354.
[17] Kolaczyk, E. D. (2009). Statistical Analysis of Network Data: Methods and Models.
Springer.
[18] Kolar, M., Song, L., Ahmed, A., Xing, E.P. (2010) Estimating time-varying networks.
Ann. Appl. Statist., 4, 94–123.
[19] Lee, M., Shen, H., Huang, J. Z., Marron, J.S.(2010) Biclustering via sparse singular
value decomposition. Biometrics, 66, 1087–1095.
[20] Lepski, O. (2015) Adaptive estimation over anisotropic functional classes via oracle
approach. Ann. Statist., 43, 1178–1242.
[21] Lovász, L., Szegedy, B. (2006) Limits of dense graph sequences. Journ. Combinatorial
Theory, Ser. B, 96, 933–957.
[22] Lovász, L. (2012) Large Networks and Graph Limits. Colloquium Publications of the
American Mathematical Society, 60, American Mathematical Society, Providence,
RI.
[23] Massart, P. (2007) Concentration Inequalities and Model Selection. Springer-Verlag,
Berlin, Heidelberg.
[24] Matias, C., Miele, V. (2017) Statistical clustering of temporal networks through a
dynamic stochastic block model. Journ. Royal Stat. Soc., Ser. B, 79, 1119–1141.
[25] Minhas, S., Hoff, P.D., Warda, M.D. (2015) Relax, Tensors Are Here. Dependencies
in International Processes. ArXiv:1504.08218
[26] Olhede, S.C., Wolfe, P. J. (2014) Network histograms and universality of blockmodel
approximation. Proceed. Nat. Acad. Sci., 111, 14722–14727.
DYNAMIC NETWORK MODELS AND GRAPHON ESTIMATION
29
[27] Pollard, D. (1990) Empitrical Processes: Theory and Applications. Institute of Mathematical Statistics.
[28] Tsybakov, A. B. (2009). Introduction to Nonparametric Estimation. Springer, New
York.
[29] Vershynin, R. (2012). Introduction to the non-asymptotic analysis of random matrices. In Compressed Sensing, Theory and Applications, ed. Y. Eldar and G. Kutyniok,
Chapter 5. Cambridge University Press.
[30] Wolfe,
P.J.,
Olhede,
S.C.(2013)
Nonparametric
graphon
estimation.
ArXiv:1309.5936.
[31] Xing, E.P., Fu, W., Song, L. (2010) A state-space mixed membership blockmodel for
dynamic network tomography. Ann. Appl. Stat., 4, 535–566.
[32] Xu, K.S. (2015) Stochastic Block Transition Models for Dynamic Networks. Proceedings of the 18th International Conference on Artificial Intelligence and Statistics
(AISTATS) 2015, San Diego, CA, USA. JMLR: W&CP, 38.
[33] Xu, K.S., Hero III, A.O. (2014) Dynamic stochastic blockmodels for time-evolving
social networks. IEEE Journal of Selected Topics in Signal Processing, 8(4), 552–562.
[34] Yang, T., Chi, Y., Zhu, S., Gong, Y., Jin, R. (2011) Detecting communities and
their evolutions in dynamic social networksa Bayesian approach. Mach. Learn., 82,
157–189.
[35] Zhang, A.Y., Zhou, H.H. (2016) Minimax rates of community detection in stochastic
block models. Ann. Statist., 44, 2252–2280.
9. Supplemental Material: Proofs.
b C,
b m,
9.1. Proof of Theorem 1. Since (d,
b Jb) are solutions of optimization
problem (14), for any m, J, C and d one has
b b k2 + Pen(|Jb|, m)
bW
cT d
(S.1) ka − C
b ≤ ka − CWT d(J) k2 + Pen(|J|, m),
(J )
For any m, J, C and d, it follows from (18) that
ka − CWT d(J) k2 = k(a − C∗ W∗ T d∗ ) + (C∗ W∗ T d∗ − CWT d(J) )k2
= kξk2 + kCWT d(J) − C∗ W∗ T d∗ k2 + 2ξ T (C∗ W∗ T d∗ − CWT d(J) ).
Hence, plugging the last identity into the inequality (S.1), derive that for
any m, J, C and d
bW
cT d
b b − C∗ W∗T d∗ k2 ≤ kCWT d(J) − C∗ W∗T d∗ k2
(S.2)
kC
(J )
+ ∆ + Pen(|J|, m) − Pen(|Jb|, m).
b
bW
cT d
b b − CWT d(J) )| ≤ ∆1 + ∆2 + ∆3 with
Here ∆ = 2|ξ T (C
(J )
(S.3)
∆1 = 2|ξ T (C∗ W∗T d∗ − CWT d(J) )|,
b b b)C∗ W∗ T d∗ |, ∆3 = 2ξ T Π
b b b ξ,
∆2 = 2|ξ T (IN L − Π
C,J
C,J
b b =Π
bW
cT d
b b b a where a is given
since, due to (22) and (23), one has C
C,J
(J)
by (18). Now, we need to find upper bounds for each of the terms in (S.3).
By Lemma 2 with α = 1/2 and any t > 0, one has
n
o
(S.4)
P ∆1 − 0.5 kC∗ W∗ T d∗ − CWT d(J) k2 ≤ 4t ≥ 1 − 2e−t .
Note that
bW
cT d
b
kC
b
(J)
b b b)C∗ W∗ T d∗ k2
b b b ξk2 + k(IN L − Π
− C∗ W∗ T d∗ k2 = kΠ
C,J
C,J
b b b)C∗ W∗T d∗ k2 .
≥ k(IN L − Π
C,J
Therefore, applying an union bound over m = 1, · · · , n, C ∈ C(m, n, L) and
J with |J| = 1, · · · , M L, we derive that for any x > 0
n
o
b b − C∗ W∗ T d∗ k2 < 4x
bW
cT d
P ∆2 − 0.5 kC
(J)
o
n
b b b)C∗ W∗T d∗ k2 < 4x
≥P ∆2 − 0.5 k(IN L − Π
C,J
≥1 −
n
X
X
ML X
X
m=1 C∈C(m,n,L) j=1 |J|=j
n
b b b)CWT d∗ |
P 2|ξ T (IN L − Π
C,J
o
b b b)CWT d∗ k2 < 4x .
− 0.5 k(IN L − Π
C,J
1
Denote
(S.5) R(m, J, L) = log(|C(m, n, L)|) + |J| log m2 Le |J|−1 + 2 log(m|J|).
Then, taking into account that the number of sets J with |J| = j is
ML
j
≤
M Le
j
j
≤
m2 Le
j
j
and applying Lemma 2 with α = 1/2 and x = t + R(m, J, L), derive
n X
ML
o
X
−t
Tb
∗
∗T ∗ 2
b
b
c
m−2 j −2 .
b J, L) ≤ 4t ≥ 1−2e
P ∆2 − 0.5 kCW d(J)
b − C W d k − 4R(m,
n
m=1 j=1
√
P
−2 = π 2 /6 <
3, the last inequality yields
Since ∞
j=1 j
(S.6)
n
o
bW
cT d
b b − C∗ W∗ T d∗ k2 − 4R(m,
b L) ≤ 4t ≥ 1 − 6e−t .
P ∆2 − 0.5 kC
b
J,
(J)
Finally, in order to obtain an upper bound for ∆3 , apply Lemma 4 with
b b b and again use the union upper bound over m = 1, · · · , n, C ∈
A=Π
C,J
C(m, n, L) and J with |J| = 1, · · · , M L similarly to the way it was done
for ∆2 . Since for any projection matrix ΠC,J , one has kΠC,J kop = 1 and
kΠC,J k2 = |J|, obtain that for any t > 0
3
3t
b
b
(S.7)
P ∆3 − |J | − R(m,
b J, L) ≤
≥ 1 − e−t ,
2
2
where R(m, J, L) is defined in (S.5). Combining (S.2)–(S.7) and recalling
b=C
bW
cT d
b b and θ ∗ = C∗ W∗ T d∗ , obtain that, with probability at
that θ
(J )
least 1 − 9e−t , one has
(S.8)
b − θ ∗ k2 ≤
3kCWT d(J) − θ ∗ k2 + 11R(m, J, L) + 2|J| + 19t.
kθ
min
m,J,q
C∈C(m,n,L)
In order to complete the proof of (20), observe that 2 log(|J|) ≤ 2|J| and
log(|C(m, n, L)| ≥ 2 log m by (13). Therefore, one has
(S.9)
11R(m, J, L) + 2|J| ≤ 2Pen(|J|, m),
and (20) follows from (S.8), (S.9) and the second inequality in (19).
Finally, inequality (21)
R ∞can be proved by noting that for any random
b − Λ∗ k2 .
variable ζ one has Eζ ≤ 0 P(ζ > z)dz and using it with ζ = kΛ
2
9.2. Proofs of Lemma 1, Theorem 2 and Corollary 2. This section contains proofs of Lemma 1, Theorem 2 and Corollary 2.
Proof of Lemma 1. Since |Zbal (m, n, n0 , L, 1, 1)| ≤ |Zbal (m, n, n0 , L, ℵ1 , ℵ2 )|,
it is sufficient to prove (28) for ℵ1 = ℵ2 = 1. Note that
n
log |Zbal (m, n, n0 , L, 1, 1)| ≥ log(n!) − m log[(n/m)!] + (L − 1) log
n0
since there are nn0 ways to select n0 nodes out of n but there are more
than one way to put them back. Applying Lemma 5 with γ = 1 obtain that
log(n!) − m log[(n/m)!] ≥ n log(m)/4. In addition,
4
n
mne
n0
n0
n
n
log
log
=
≥
log
> n0 log
n0
4
4
n0
n0
n40
provided n4 /n40 ≥ (mne)/n0 which holds under conditions of the lemma due
to m ≤ n.
Proof of Theorem 2. Let m∗ be the true number of classes, M ∗ =
m∗ (m∗ + 1)/2. Let C = C∗ be the true clustering matrix and (C∗ )T C∗ =
(S∗ )2 where (S∗ )2 is the diagonal matrix with the number of nodes in respective pairs of classes on the diagonal. Let Q∗ be the true matrix of probabilities of connections for pairs of classes, D∗ = Q∗ H, d∗ = vec(D∗ ),
θ ∗ = C∗ (W∗ )T d∗ and W∗ = H ⊗ IM ∗ . We need to find an upper bound for
kC∗ (W∗ )T d∗ (J) − θ ∗ k2 in (20). Let Q∗ (J) be such that
vec(Q∗ (J) ) = q∗ (J) = (W∗ )T d∗ (J) = (W∗ )T (W∗ q∗ )(J) .
Then, by direct calculations, one obtains
i
h
iT
h
kC∗ (W∗ )T d∗ (J) − θ ∗ k2 = q∗ (J) − (S∗ )−2 C∗ θ ∗ (S∗ )2 q∗ (J) − (S∗ )−2 C∗ θ ∗
+ kθ ∗ k2 − k(S∗ )−1 (C∗ )T θ ∗ k2 .
Since k(S∗ )−1 (C∗ )T θ ∗ k = kθ ∗ k and (S∗ )−2 C∗ θ ∗ = q∗ = vec(Q∗ ) = vec(D∗ H),
obtain
kC∗ (W∗ )T d∗ (J) − θ ∗ k2 = kS∗ (Q∗ (J) − Q∗ )k2
∗
≤
M
X
k=1
(l)
max Nk
l
L h
X
l=1
(Q∗ (J) )k,l − Q∗ k,l
M∗
n 2 X
2
≤ ℵ2
kQ∗ (J) )k,∗ − Q∗ k,∗ k2
m∗
k=1
3
i2
(l)
Here, k is the index corresponding to a pair of classes (k1 , k2 ), Nk1 ,k2 is
defined in formula (10) and the second inequality follows from assumption
(27). In order to complete the proof, note that
k=1
∗
∗
∗
M
X
kQ
∗
(J) )k,∗
−Q
∗
2
k,∗ k
=
M
X
k=1
kD∗(Jk ) )k,∗
−
D∗k,∗ k2
=
M X
X
(D∗k,l )2 .
k=1 l∈J
/ k
Therefore, (20) implies (30).
Proof of Corollary 2. Observe that it follows from (31) that one can choose
Jk = {l : 1 ≤ l ≤ L0 } where L0 ≤ L. Then, (29) yields that |J| = L0 M ∗ .
Moreover, due to assumption (31), obtain
∗
M X
X
k=1 l∈J
/ k
0
(D∗k,l )2 ≤ K0 M ∗ L−2ν
.
0
Note also that if L0 = L, then there is no bias and the sum in (30) is identical
zero. Then, (30) becomes
b − Λ∗ k2 ≤ min {∆(n, L, m∗ ), 22M ∗ L}
kΛ
∗
m ne
+ 38 t,
+ 44 n log(m∗ ) + n0 L log
n0
where
25 L
2 2 −2ν0
∗ 2
∆(n, L, m ) = min
6K0 ℵ2 n L0
+ 22(m ) L0 log
0≤L0 <L
L0
∗
In order to obtain (32), minimize the right-hand side of the last expression
with respect to L0 < L and note that, if L0 < L, then log(L0 ) ≍ log(L) ≍
log(n/m∗ ).
9.3. Proofs of Theorems 3 and 4. This section contains the proofs of
the lower bounds for the error. The lower bounds in both, Theorem 3 and
Theorem 4, consist of two parts, the clustering error and the nonparametric
estimation error. We shall consider those terms separately.
Proof of Theorem 3. Although the upper bounds for the risk in Corollary 1 are derived for the case of general clustering matrices, due to the fact
that the balanced model clustering complexity is the same as complexity
of general clustering, we derive the lower bounds for the clustering error
4
for the case when C ∈ Zbal (m, n, n0 , L, ℵ1 , ℵ2 ). Moreover, since the case of
ℵ1 = ℵ2 = 1 is the most restrictive, we prove the clustering error for this case.
The clustering error. Without loss of generality, assume that γm and
γn are integers. Assume that connectivity tensor G does not change with l,
so G∗,∗,l = V is an m × m symmetric matrix. Let V be block diagonal and
such that the diagonal blocks are equal to zero and the non-diagonal blocks
are equal to F and FT , respectively, so that Vk1 ,k2 = 0 if 1 ≤ k1 , k2 ≤
(1 − γ)m or (1 − γ)m + 1 ≤ k1 , k2 ≤ m and Vk1 ,(1−γ)m+k2 = Fk1 ,k2 if
k1 = 1, · · · (1 − γ)m, k2 = (1 − γ)m + 1, · · · , m. Since components of vectors
Gk1 ,k2 ,∗ are constant for any k1 , k2 , then, due to condition (33), each of the
vectors HFk1 ,k2 has only one non-zero component, so that the set J has at
most γ(1 − γ)m2 < s nonzero elements.
Consider a collection of binary vectors ω ∈ {0, 1}(1−γ)m . By VarshamovGilbert Lemma (see Tsybakov (2009)), there exists a subset Ξ of those
vectors such that for any ω, ω ′ ∈ Ξ one has kω − ω ′ kH = kω − ω ′ k2 ≥
(1 − γ)m/8 ≥ m/16 and |Ξ| ≥ exp((1 − γ)m/8). Assume, without loss of
generality, that m is large enough, so that exp((1−γ)m/8) ≥ γm, otherwise,
choose a smaller value of γ (inequality exp((1−γ)m/8) ≥ γm is always valid
for γ ≤ 1/9). Choose γm vectors ω in Ξ, enumerate them as ω (1) , · · · , ω (γm)
and use them to form columns of matrix F as follows:
(S.10)
F∗,j = 0.5 1 + αω (j) ,
j = 1, · · · , γm.
Then, for any j, j ′ = 1, · · · , γm, obtain
(S.11)
kF∗,j − F∗,j ′ k2 ≥ α2 m/16,
where α is a positive constant that will be defined later. Note that for every
l and k one has
(S.12)
k
k
ke
1
l log
≤ log
, log(k!) = k log k−k+ log(2πk)+o(1),
≤ l log
l
l
2
l
where the o(1) term is smaller than 1. Therefore, it follows from the first
formula in (24) that
(S.13)
mne
mn
≤ log |Z(m, n, n0 , L)| ≤ n log m+n0 (L−1) log
n log m+n0 (L−1) log
n0
n0
The term n log m in (S.13) is due to the initial clustering while the term
n0 (L − 1) log (mn/n0 ) is due to temporal changes in the clusters’ memberships.
5
In what follows, we shall utilize clustering functions z (l) : [n] → [m] corresponding to clustering matrices C(l) such that z (l) (j) = k iff at the moment
tl node j belongs to class Ωk , k = 1, · · · , m.
Clustering error due to initial clustering. First consider the case
when initial clustering error dominates. If m = 2 or m takes a small value,
the proof is almost identical to the proof in Section 3.3 of Gao et al. (2015).
Hence, we shall skip this part and consider the case when m is large enough,
so that γm ≥ 2.
Following Gao et al. (2015), we consider clustering matrices and clustering
functions independent of l, so that z (l) ≡ z. Consider a sub-collection of
clustering matrices F(m, n, γ) ⊂ M(m, n) such that they cluster the first
n(1 − γ) nodes into the first m(1 − γ) classes uniformly and sequentially,
n/m nodes in each class, i.e., the first n/m nodes are placed into class Ω1 ,
the second n/m nodes into class Ω2 , and so on. The remaining γn nodes are
clustered into the remaining γm classes, n/m nodes into each class. Then,
by Lemma 5,
.
log |F(m, n, γ)| = log (γn)! [(n/m)!]γm ≥ γn log(γm)/4.
Now, apply Lemma 4 with γn and γm, respectively, instead of n and m and
r = γn/32. Derive that there exists a subset S(m, n, γ) of the set F(m, n, γ)
such that, for any C, C′ ∈ S(m, n, γ), one has 2 {#j : z(j) 6= z ′ (j)} = kC −
C′ kH ≥ γn/32. Also, by (S.66),
(S.14)
γn log(32mγe)
γn
γn
log(γm) −
≥
log(γm).
4
32
16
log |S(m, n, γ)| ≥
Let Λ and Λ′ be the tensors of probabilities corresponding to, respectively,
clustering matrices C, C′ ∈ S(m, n, γ) with related clustering functions z
and z ′ . Then, by (S.11), due to the fact that the first n(1 − γ) nodes are
clustered uniformly and sequentially, obtain
n(1−γ)
kΛ − Λ′ k2 = 2L
=
=
2Ln
m
2Ln
m
X
i=1
m(1−γ)
X
k=1
n
X
n
X
n
X
j=n(1−γ)+1
(Fz(i),z(j) − Fz ′ (i),z ′ (j) )2
(Fk,z(j) − Fk,z ′ (j) )2
j=n(1−γ)+1
j=n(1−γ)+1
kF∗,z(j) − F∗,z ′ (j) k2 ≥
6
2Ln α2 m
#j : z(j) 6= z ′ (j) ,
m 16
so that
(S.15)
kΛ − Λ′ k2 ≥ 2−9 Ln2 α2 γ.
On the other hand, if α ≤ 1/4, then, by Lemma 7, obtain that the Kullback
divergence is bounded above
(S.16)
K(PΛ , PΛ′ ) ≤ 8 kΛ − Λ′ k2 ≤ 16α2 n2 Lγ.
Set α2 = Cα log(γm)/nL and apply Theorem 2.5 of Tsybakov (2009). Due
to (S.14) and (S.16), if Cα is a small enough absolute constant,
16α2 n2 Lγ = 16Cα log(γm)n < (1/8) (γn/16) log(γm) ≤ (1/8) log |S(m, n, γ)|
and conditions of Theorem 2.5 are satisfied. Since Ln2 α2 γ = Cα n γ log(γm)
and log(γm) ≥ C(γ) log m for some constant C(γ) dependent on γ only,
derive
(
)
b − Λ∗ k2
kΛ
C(γ) log m
PΛ
sup
(S.17)
inf
≥ 1/4.
≥
n2 L
nL
c G∈Gm,L,s
Λ
C∈S(m,n,γ)
Clustering error due to changes in the memberships. Now, we
consider the case when the clustering error which is due to the temporal
changes in memberships dominates the error of initial clustering. Use the
same construction for G and F as
Q before. Consider the following collection of clustering matrices F = L
l=1 Fl where Fl is defined as follows.
When l is odd, Fl contains only one matrix that clusters nodes uniformly
and sequentially, i.e., the first n/m nodes go to class Ω1 , the second n/m
nodes go to class Ω2 and the last n/m nodes go to class Ωm . If l is even,
Fl = P(m, n, n0 , γ) where P(m, n, n0 , γ) is the set of clustering matrices
that corresponds to a perturbation of the uniform sequential clustering with
at most n0 nodes moved to different classes in the manner described below.
Let k0 be an integer such that
(S.18)
k0 ≤ n0 /(γm) < k0 + 1.
If k0 = 0, then n0 < γm and we choose n0 clusters out of the last γm
clusters, remove one element from each of those clusters and then put those
n0 elements back in such a manner that every element goes to a different
cluster and no elements goes back to its own cluster. If k0 ≥ 1, we remove
k0 elements from each of the last γm clusters and then put each of those
k0 -tuples back, one tuple per cluster, so that none of the tuple goes back
7
to its own cluster. Then, log |F| = [L/2] log |P(m, n, n0 , γ)| where [L/2] ≥
(L − 1)/2 is the largest integer not exceeding L/2 and
(
log γm
n0 + n0 log(n/m) + log[(n0 − 1)!], if k0 = 0;
log |P(m, n, n0 , γ)| =
γm log n/m
+ log[(γm − 1)!],
if k0 ≥ 1.
k0
If n0 < γm, so that k0 = 0, then, by (S.12), obtain that log |P(m, n, n0 , γ)| ≥
n0 log(γm/n0 ) + n0 log(n/m) = n0 log(γn/n0 ). If n0 ≥ γm and k0 ≥ 1, then,
by (S.12), obtain log |P(m, n, n0 , γ)| ≥ γmk0 log(n/(mk0 )). Since k0 + 1 ≤
2k0 , obtain that k0 ≥ n0 /(2mγ). Hence, for any k0 ≥ 0
(S.19)
log |P(m, n, n0 , γ)| ≥
n0
log (γn/n0 ) .
2
For every even value of l, apply Lemma 4 with γn and γm, respectively,
instead of n and m and r = n0 /40 obtaining that there exists a subset Sl (m, n, n0 , γ) of the set P(m, n, n0 , γ) such that, for any C(l) , C′(l) ∈
Sl (m, n, n0 , γ), one has
(S.20)
kC(l) − C′(l) kH ≥ n0 /40,
l = 2k, k = 1, · · · , [L/2].
By (S.19) and Lemma 4, for every even l, one has
n0
γn
80nemγ 2
nem
n0
n0
log |Sl (m, n, n0 , γ)| ≥
log
log
log
−
≥
2
n0
40
n0
40
n0
since, due to n0 ≤ 4/3 γn m−1/9 and (80e2 )1/18 ≤ 0.75, one has (γn/n0 )20 ≥
(80e2 ) [m (γn/n0 )]2 , so that
20 log(γn/n0 ) ≥ log(80n2 m2 e2 γ 2 /n20 ) = log(nem/n0 ) + log(80nemγ 2 /n0 ).
For odd values of l, let Sl (m, n, n0 , γ) contain just one clustering matrix
corresponding to the
Q uniform sequential clustering. Now, consider the set
S(m, n, n0 , γ, L) = L
l=1 Sl (m, n, n0 , γ) with
nem
(L − 1)n0
log
.
(S.21)
log |S(m, n, n0 , γ, L)| ≥
80
n0
Let C = (C(1) , C(2) , · · · , C(L) ) and C′ = (C′(1) , C′(2) , · · · , C′(L) ) be two
sets of clustering matrices with C(l) , C′(l) ∈ Sl (m, n, n0 , γ) and let z =
(z1 , · · · , zL ) and z ′ = (z1′ , · · · , zL′ ) be the corresponding clustering functions.
8
Let Λ and Λ′ be the tensors of probabilities corresponding to sets of clustering matrices C, C′ ∈ S(m, n, n0 , γ, L). Then, similarly to the previous case,
using (S.11), derive
[L/2] n(1−γ)
′ 2
kΛ − Λ k = 2
=
X
l=1
2n X
=
m
i=1
l=1
[L/2] m(1−γ)
2n X X
m
[L/2]
X
k=1
n
X
j=n(1−γ)+1
j=n(1−γ)+1
n
X
n
X
2
′ (i),z ′ (j) )
(Fz2l (i),z2l (j) − Fz2l
2l
2
′ (j) )
(Fk,z2l (j) − Fk,z2l
[L/2]
nα2 X
′ (j) k ≥
kF∗,z2l (j) − F∗,z2l
kC(2l) − C′(2l) kH ,
8
2
l=1 j=n(1−γ)+1
l=1
so that by (S.20),
(S.22)
kΛ − Λ′ k2 ≥ (L − 1)nn0 α2 /1280.
Again, similarly to the previous case, if α ≤ 1/4, then by Lemma 7, obtain
that the Kullback divergence is bounded above
(S.23)
K(PΛ , PΛ′ ) ≤ 8 kΛ − Λ′ k2 ≤ 8Lnn0 α2 .
Set α2 = Cα n−1 log(nem/n0 ) where Cα is an absolute constant and apply
Theorem 2.5 of Tsybakov (2009). Observe that if Cα is small enough, then,
due to (S.21) and (S.23), conditions of this theorem are satisfied, hence,
(S.24)
(
)
b − Λ∗ k2
nem
kΛ
n0
PΛ
sup
inf
≥ C(γ) 2 log
≥ 1/4.
n2 L
n
n0
c
G∈Gm,L,s
Λ
C∈S(m,n,n0 ,γ,L)
The nonparametric estimation error. Consider uniform sequential
clustering with n/m nodes in each group and group memberships remaining
the same for all l = 1, · · · , L. Let Q ∈ RM ×L be the matrix with columns
q(l) , l = 1, · · · , L, defined in Section 2.3. Denote V = QHT ∈ RM ×L and
recall that for G ∈ Gm,L,s , by (11) and (12), matrix V should have at most
s nonzero entries.
Let k0 = min(s/2, M ). Choose k0 rows among M rows of matrix V and
denote this set by X . If k0 = M , set X = {1, · · · , M }. For k ∈ X , set
Vk,1 6= 0. We have already distributed k0 non-zero entries and have s − k0
entries left. We distribute those entries into the k0 rows Vk,∗ where k ∈ X .
Let
(S.25)
s0 = [(s − k0 )/k0 ] = [s/k0 ] − 1
9
with
s/2 ≤ k0 s0 < s,
where [s/k0 ] is the largest integer no larger than s/k0 . Consider a set of
binary vectors ω ∈ {0, 1}L with exactly s0 ones in each vector. By Lemma
4.10 of Massart (2007), there exists a subset T of those vectors such that
for any ω, ω ′ ∈ T , one has
kω − ω ′ kH ≥ s0 /2 and log |T | ≥ 0.233 s0 log(L/s0 ).
N
Denote T̃ = k∈X Tk , where Tk is a copy of the set T corresponding to row
k of matrix V. For ω (k) ∈ Tk , set
√
(S.26) Vk,∗ = ( L/2, · · · , 0) + αm/n ω (k) , if k ∈ X , Vk,∗ = 0 if k ∈
/ X.
It is easy to see that matrix V has at most s nonzero entries as required.
Let V and V′ be matrices corresponding to sequences ω (k) and ω ′(k) in
Tk , k ∈ X . Let Λ and Λ′ be the tensors corresponding to V and V′ . Then,
due to (S.25) and the uniform sequential clustering, (S.26) implies that
kΛ − Λ′ k2 ≥
n 2
kΛ − Λ′ k2 ≤ 4
kV − V′ k2 ≥
m
m 2
n
α2
n 2
m
n 2 X
m
k∈X
′
kVk,∗ − Vk,∗
k2 ≥
α2 s
α2 k0 s0
≥
;
2
4
k0 s0 ≤ 4α2 s.
Set α2 = Cα log(L/s0 ). It is easy to check that, due to assumptions (33)
and (34), one has Qij ∈ [1/4, 3/4] for any i and j. Hence, by Lemma 7,
obtain K(PΛ , PΛ′ ) ≤ 8 kΛ − Λ′ k2 ≤ 32α2 s. If Cα ≤ 2−8 · 0.233, then
32α2 s < (1/8) log(T̃ ) and conditions of Theorem 2.5 of Tsybakov (2009)
hold.
Finally, in order to obtain the last term in (35), examine L/s0 . If s < 2M ,
then k0 = s/2, s0 = 1 and L/s0 = L = Lm2 /m2 ≥ γLm2 /s. If s ≥ 2M , then
k0 = M , s0 ≤ s/M and L/s0 ≥ LM/s ≥ Lm2 /(2s) ≥ γLm2 /s. Since for
some constant C(γ) > 0 independent of L and m, one has log(γLm2 /s) ≥
C(γ) log(Lm2 /s), obtain
(
)
b − Λ∗ k2
Lm2
s
kΛ
≥ 1/4.
PΛ
≥ Cγ 2 log
sup
(S.27) inf
n2 L
n L
s
c
G∈Gm,L,s
Λ
C∈S(m,n,n0 ,γ,L)
Finally, in order to obtain the lower bound in (35) observe that, for any
a, b, c ≥ 0, one has max(a, b, c) ≤ a + b + c ≤ 3 max(a, b, c) and then combine
(S.17), (S.24) and (S.27).
10
Proof of Theorem 4. Note that for Theorem 3 we proved the lower
bounds for the clustering error in the most restrictive case when C ∈
Zbal (m, n, n0 , L, 1, 1). Moreover, in this proof, the connection probabilities
are set to be constant over time, hence due to condition (33), D∗k,l = 0 for
l ≥ 2 in assumption (31), so (31) holds for any K0 and ν0 . Therefore, the
lower bounds for the risk due to clustering errors hold in this case and coincide with the lower bounds in Theorem 3. For this reason, we only need to
prove the lower bounds that are due to the nonparametric estimation error.
The nonparametric estimation error. Consider a set up where the
nodes are grouped into m classes, n/m nodes in each class, so the model is
fully balanced. Let Gk,k,l = 0 for any k = 1, · · · , m and l = 1, · · · , L.
Consider an even number L0 such that 1 ≤ L0 ≤ L/2 and a set of vectors
ω ∈ {0, 1}L0 with exactly L1 = L0 /2 nonzero entries. By Lemma 4.10 of
Massart (2007), there exists a subset T of those vectors such that for any
ω, ω ′ ∈ T one has
(S.28)
kω − ω ′ kH ≥ L0 /4,
log |T | ≥ 0.233(L0 /2) log 2 ≥ 0.08L0 .
Denote K = {(k1 , k2 ) : 1 ≤ kN
1 < k2 ≤ m} and let Tk1 ,k2 be the copies of T
for (k1 , k2 ) ∈ K. Denote T̃ = (k1 ,k2 )∈K Tk1 ,k2 and observe that
log |T̃ | = 0.5 m(m − 1) log |T | ≥ 0.02m2 L0 .
(k ,k )
Then, ω̃ ∈ T̃ are binary tensors with elements ω l 1 2 , l = 1, · · · , L0 , and
(k1 , k2 ) ∈ K. Consider a set of matrices D(ω̃ ) indexed by ω̃ such that for
the index k = 1, · · · , M, corresponding to (k1 , k2 ) ∈ K, one has
√
(ω̃ )
Dk,1 = L/2;
(ω̃ )
(k1 ,k2 )
Dk,l = α ω l−L
, l = L0 + 1, · · · , 2L0 , k = (k1 , k2 ) ∈ K.
0
In order condition (31) is satisfied, we set
(S.29)
−(2ν0 +1)
α2 ≤ C1 L0
with C1 ≤ min(K2 21−2ν0 , 1/8).
Denote by Λ and Λ′ the probability tensors corresponding, respectively, to
ω̃ and ω̃ ′ in T̃ . Then, due to (S.28) and the symmetry,
kΛ − Λ′ k2 ≥ α2
m
X
m
X
k1 =1 k2 =k1 +1
kω (k1 ,k2 ) − ω ′(k1 ,k2 ) kH
kΛ − Λ′ k2 ≤ α2 m(m − 1)(n/m)2 L0 ≤ α2 n2 L0 .
11
n 2
m
≥
α2 n2 L0
;
8
Note that one has 1/4 ≤ Qk,l ≤ 3/4 provided kHT ω (k1 ,k2 ) k∞ ≤ 1/4 for
(k1 , k2 ) ∈ K. By Assumption (33), the latter is guaranteed by α2 L20 /L ≤ 1/4,
so that, due to L ≥ 2L0 and ν0 ≥ 1/2, it is ensured by (S.29). Then, by
Lemma 7, one has
K(PΛ , PΛ′ ) ≤ 8 kΛ − Λ′ k2 ≤ 8α2 n2 L0 ≤ log |T̃ |/8
provided
α2 ≤ C2 (m/n)2 ,
where C2 is an absolute constant. Therefore, application of Theorem 2.5 of
Tsybakov (2009) yields that
(
)
b − Λk2
kΛ
1
inf sup PΛ
≥ ∆(n, L) ≥
2
n L
4
c G∈Gm,L,s
Λ
C∈Z
bal
with ∆(n, m, L) = C α2 L0 /L where C is an absolute constant.
Now, we denote C32 = (C2 /C1 )2−(2ν0 +1) and consider two cases. If n ≤
C3 mLν0 +1/2 , choose L0 = [(C1 /C2 )(n/m)2 ]1/(2ν0 +1) which leads to α2 =
−(2ν +1)
C1 L 0 0
= C2 (m/n)2 . It is easy to check that L0 ≥ 2 and that L0 ≤ L/2,
so that
2ν0
C m 2 2ν0 +1
∆(n, m, L) =
.
L
n
If n > C3 mLν0+1/2 , choose α2 = C2 (m/n)2 and set L0 = L/2. Then, (S.29)
holds and ∆(n, m, L) = C(m/n)2 which completes the proof of the lower
bound in this case.
9.4. Proof of Theorem 5 . Denote WT = V and recall that S2 = CT C
(l)
where S2 is the diagonal matrix with entries Nk , the number of nodes in
the pair of classes k = (k1 , k2 ) at time tl . Note that (CWT )J = CVJ , so
that ΠC,J = CVJ [VJT CT CVj ]−1 VJT CT can be written as
ΠC,J = CS−1 ΠS,J (CS−1 )T
with
ΠS,J = SVJ (VJT S2 VJ )−1 VJT S.
Here, ΠS,J is the projection matrix on the column space of matrix SVJ .
b C,
b m,
Since (d,
b Jb) are solutions of optimization problem (38), for any J
∗
and d = vec(D∗ ), one has
(S.30)
b b k2 + 2ρn Pen(|J|,
bW
cT d
b m)
ka − C
b
(J)
≤ ka − C∗ W∗T d∗ (J) k2 + 2ρn Pen(|J|, m∗ ),
12
It follows from (18) that for any m, J, d and C
ka − CWT d(J) k2 = k(a − C∗ W∗T d∗ ) + (C∗ W∗T d∗ − CWT d(J) )k2
= kξk2 + kCWT d(J) − C∗ W∗T d∗ k2 + 2ξ T (C∗ W∗T d∗ − CWT d(J) ).
Hence, plugging the last identity into the inequality (S.30), derive that
b b − C∗ W∗T d∗ k2 ≤ kC∗ W∗T d∗ (J) − C∗ W∗T d∗ k2
bW
cT d
(S.31) kC
(J)
+ ∆ + ρn Pen(|J|, m∗ ) − ρn Pen(|Jb|, m),
b
b b − C∗ W∗T d∗ (J) )|. Note that, due to (22) and
bW
cT d
where, ∆ = 2|ξ T (C
(J)
T
b
b
c
b b b a with a is given by (18), C∗ W∗T d∗ (J) =
(23), one has CW d b = Π
(J)
C,J
ΠC∗ ,J θ ∗ and C∗ W∗ T d∗ = θ ∗ . Therefore,
(S.32)
∆ ≤ ∆1 + ∆2 + ∆3 ,
⊥
∗
T b
∗
∆1 = 2|ξ T Π⊥
b Jbθ |,
C∗ ,J θ |, ∆2 = 2|ξ ΠC,
b b b ξ.
∆3 = 2ξ T Π
C,J
In order to obtain an upper bound for ∆1 and ∆2 , note that by Bernstein
inequality, for any C, m, θ, J and for any x > 0 with probability at least
1 − 2e−x , one has
q
⊥
2
2xρn kΠ⊥
θ|
≤
2
2|ξ T Π⊥
C,J
C,J θk + 4/3 kΠ C,J θk∞ x.
Due to 2ab ≤ a2 + b2 , obtain that with probability at least 1 − 2e−x
⊥
2
⊥
(S.33)
2|ξ T Π⊥
θ|
≤
0.5
kΠ
θk
+
4
ρ
+
kΠ
θk
/3
x
n
∞
C,J
C,J
C,J
Applying (S.33) to ∆1 with θ = θ ∗ and x = t > 0, using (39) and keeping
in mind that kΛ∗ k∞ = kθ ∗ k∞ ≤ ρ∗n ≤ ρn , obtain
(S.34)
o
n
P ∆1 − 0.5 kC∗ W∗ T d∗ − C∗ W∗ T d∗ (J) k2 − 4 ρn (1 + B0 /3)t ≤ 0 ≥ 1−2e−t .
In order to obtain an upper bound for ∆2 , note that, similarly to the proof
of Theorem 1, one has
b ⊥b bC∗ W∗ T d∗ k2 .
bW
cT d
b b −C∗ W∗ T d∗ k2 ≥ k(IN L −Π
b b b)C∗ W∗ T d∗ k2 = kΠ
kC
C,J
C,J
(J)
Hence, due to condition (39), since kθ ∗ k∞ ≤ ρ∗n ≤ ρn , obtain
o
n
∗
b b − C∗ W∗T d∗ k2 ≤ 4 ρn + 1/3 kΠ
b⊥
bW
cT d
θ
k
P ∆2 − 0.5 kC
b
b
∞ x ≥
C,J
(J )
o
n
∗
∗T ∗ 2
b⊥
P ∆2 − 0.5 kΠ
b JbC W d k − 4 ρn (1 + B0 /3)x ≤ 0
C,
13
b L) where R(m, J, L) is defined in (S.5). Applying inequalSet x = t+R(m,
b J,
ity (S.33) with θ = θ ∗ together with the union bound over m = 1, · · · , n,
C ∈ Zbal (m, n, n0 , L, ℵ1 , ℵ2 ) and J with |J| = 1, · · · , M L, we derive
b b − C∗ W∗ T d∗ k2
bW
cT d
P {∆2 − 0.5 kC
(J)
(S.35)
o
−4 ρn (1 + B0 /3)(t + R(m,
b Jb, L)) ≤ 0 ≥ 1 − 6e−t .
For an upper bound for ∆3 , write
bS
b −1 Π
bS
b −1 )T ξ = 2η T Π
b b b(C
b b bη.
∆3 = 2ξ T C
S,J
S,J
b = C, due to Corollary 5, vector η = (CS−1 )T ξ
Here, for any fixed C
has independent sub-Gaussian components such that, for any index i corresponding to a pair of nodes k = (k1 , k2 ) at time tl , one has
(l)
kη i k2ψ2 ≤ 1.5 e2 max(kθ ∗ k∞ , [Nk ]−1 ).
Observe that, due to (10) and (27),
(l)
(l) (l)
Nk1 ,k2 ≥ 0.5 nk1 nk2 ≥ 0.5 ℵ21 (n/m)2 .
Therefore,
kη i k2ψ2
≤ ρm,n
where ρm,n
2
3e2
∗ 2 m
max ρn , 2 2 .
=
2
ℵ1 n
Therefore, by Lemma 5.5 of Vershynin (2012), obtain that there exists an
absolute constant C0 such that for any vector t one has
E[exp(η T t)] ≤ exp(C0 ρm,n ktk2 ).
Applying Lemma 3 with σ 2 = 2C0 ρm,n , derive that for any fixed C and J,
one has
P {2ηΠS,J η − 4C0 ρm,n (2|J| + 3t) ≤ 0} ≥ 1 − e−t .
Again, taking a union bound over m = 1, · · · , n, C ∈ Zbal (m, n, n0 , L, ℵ1 , ℵ2 )
and J with |J| = 1, · · · , M L, we derive that, due to ρ∗n ≤ ρn , for some absolute constant C̃0 one has
m
b2
b
(S.36)
b J , L) + t] ≤ 0 ≥ 1 − e−t .
P ∆3 − C̃0 max ρn , 2 [R(m,
n
where R(m, J, L) is defined in (S.5). The rest of the proof of (40) is very
similar to the proof of Theorem 1. In order to establish (41), follow the
arguments of the proof of Corollary 2.
14
9.5. Proof of Theorem 6, Corollary 3 and Theorem 7 . This section contains proofs for the upper and the lower bounds for the risks in the case of
graphon estimation.
Proof of Theorem 6. To prove (45), we approximate the graphon by the
DSBM and use inequality (21) in Theorem 1. In order to find an upper bound
for the bias term, we need to cluster the nodes and create an approximate
connectivity tensor Q. For this purpose, let h ≥ 1 be a positive integer and
denote κj = 1 + [(βj − βj−1 )h] where [x] is the largest integer no larger than
x. For k = 1, · · · , κj and j = 1, · · · , r, consider a set of intervals
(1)
(2)
Uj,k = (Uj,k , Uj,k ] with
(1)
Uj,k = βj−1 +(k−1)/h,
(2)
Uj,k = min{βj−1 +k/h, βj }.
Intervals Uj,k subdivide every interval (βj−1 , βj ] into κj sub-intervals of
length at most 1/h and the total number of intervals is equal to
m=
r
X
κj .
j=1
Re-number the intervals consecutively as U1 , · · · , Um and observe that since
(βj − βj−1 )h ≤ κj ≤ 1 + (βj − βj−1 )h, one has
(S.37)
h ≤ m ≤ h + r.
The value m in (S.37) acts as a number of classes. Indeed, if ζi ∈ Uk , we
place node i into class Ωk and set Zi,j = I(j = k).
Let Θ∗ be the true tensor of connection probabilities. Set Φ∗ = Θ∗ HT ,
Q = (ZT Z)−1 ZT Θ∗ and V = QHT . Since H is an orthogonal matrix, the
bias term kCWT d(J) − θ ∗ k2 in the oracle inequality (21) is equal to
kCWT d(J) − θ ∗ k2 = kZV(ρ) H − Θ∗ k2 = kZV(ρ) − ΘHT k2
(S.38)
= kZV(ρ) − Φ∗ k2 = kZV(ρ) − Φ∗ (ρ) k2 + kΦ∗ − (Φ∗ )(ρ) k2 .
The first term in the right-hand side of (S.38) describes how well the first
Lρ columns of matrix V(ρ) represent the first Lρ columns (Φ∗ )(ρ) of matrix
Φ∗ . The second term kΦ∗ − (Φ∗ )(ρ) k2 = kΘ∗ − (Φ∗ )(ρ) Hk2 evaluates how
well Θ∗ is represented by Lρ columns of its coefficients in the transform H.
The upper bound for kZV(ρ) − Φ∗ (ρ) k2 can be found by repeating the
calculations in Lemma 2.1 of Gao et al. (2015) with the only difference that
f is replaced by vl and there is an additional sum over l = 1, · · · Lρ . Then,
we obtain
(S.39)
kZV(ρ) − Φ∗(ρ) k2 ≤ K1 22ν1 n2 Lρ h−2ν1 .
15
On the other hand, if ρ < 1, then by Assumption A,
(S.40)
kΦ∗ − (Φ∗ )(ρ) k2 ≤
n X
n
X
L
X
i1 =1 i2 =1 l=Lρ +1
vl2 (ζi1 , ζi2 ) ≤ K2 n2 L−2ν2 ρ
and kΦ∗ − (Φ∗ )(ρ) k2 = 0 if ρ = 1. Now, note that for given m and ρ, one
has |J| ≤ m2 Lρ , so that
(S.41)
Pen(|J|, m) ≤ C n log m + m2 Lρ log(25 L1−ρ ) .
Therefore, (S.37)–(S.41) yield (45).
Proof of Corollary 3. Note that if ν1 = ∞, one can set h = 2, so that
the first term in (45) is equal to zero, h + r ≤ 3r and
b − Λ∗ k2
EkΛ
≤C
n2 L
I(ρ < 1) r 2 1 + (1 − ρ) log L log r
+
+
.
L2ρν2 +1
n
L1−ρ
nL
Minimizing this expression with respect to ρ ∈ [0, 1] obtain the result in (46)
for r = rn,L . If r = r0 , then
(S.42)
b − Λ∗ k2
EkΛ
≤C
n2 L
Lρ−1 I(ρ < 1) h2 log L log h
+ 2ρν2 +1 + 2 1−ρ +
h2ν1
L
n L
nL
.
Minimizing (S.42) with respect to h and ρ, obtain that the values h∗ and
∗
L∗ = Lρ delivering the minimum in (S.42) are such that h∗ ≍ (n2 / log L)1/(2(ν1 +1) .
Hence, for some absolute constants C, one has log(h∗ ) ≤ C log n and
n
o
ν1
∗
L∗ = Lρ = min L, C n2 / log L (ν1 +1)(2ν2 +1) .
Therefore, (46) holds for r = r0 .
Proof of Theorem 7. We consider the cases when r = rn,L and r = r0 ,
corresponding to piecewise constant and piecewise smooth graphon, separately.
Piecewise constant graphon. Assume, without loss of generality, that
j = n/r is an integer. Consider a set up where the nodes are grouped into r
classes and values of ζj ’s are fixed:
ζkj+i = βk + (i − 1/2)(βk+1 − βk )/j,
16
k = 0, · · · , r − 1, i = 1, · · · , j.
Then, there are j nodes in each class. Let Gk,k,l = 0 for any k = 1, · · · , r
and l = 1, · · · , L.
Consider an even number L0 such that 1 ≤ L0 ≤ L/2 and a set of vectors
ω ∈ {0, 1}L0 with exactly L1 = L0 /2 nonzero entries. By Lemma 4.10 of
Massart (2007), there exists a subset T of those vectors such that for any
ω, ω ′ ∈ T one has
(S.43)
kω − ω ′ kH ≥ L0 /4,
log |T | ≥ 0.233(L0 /2) log 2 ≥ 0.08L0 .
Denote K = {(k1 , k2 ) : 1 ≤ kN
1 < k2 ≤ r} and let Tk1 ,k2 be the copies of T
for (k1 , k2 ) ∈ K. Denote T̃ = (k1 ,k2 )∈K Tk1 ,k2 and observe that
log |T̃ | = r(r − 1)/2 log |T | ≥ 0.02r 2 L0 .
(k ,k )
Then, ω̃ ∈ T̃ are binary tensors with elements ω l 1 2 ,l = 1, · · · , L0 and
(k1 , k2 ) ∈ K. Consider a set of functions f (ω̃ ) indexed by ω̃ such that, for
βk1 −1 < x ≤ βk1 and βk2 −1 < y ≤ βk2 , their coefficients in the transform H
are given by
√
(ω̃ )
v1 (x, y) = L/2;
(ω̃ )
vl
(k ,k )
1 2
(x, y) = αω l−L
, l = L0 +1, · · · , 2L0 , (k1 , k2 ) ∈ K.
0
Then, Assumption (42) holds. In order condition (43) is satisfied, we set
(S.44)
−(2ν2 +1)
α2 ≤ C1 L0
with C1 ≤ min(K2 21−2ν2 , 1/8).
Denote by Λ and Λ′ the probability tensors corresponding, respectively, to
ω̃ and ω̃ ′ in T̃ . Then, due to (S.43) and the symmetry,
kΛ − Λ′ k2 ≥ 2α2
r
X
r
X
k1 =1 k2 =k1 +1
kω (k1 ,k2 ) − ω ′(k1 ,k2 ) kH
kΛ − Λ′ k2 ≤ α2 r(r − 1)(n/r)2 L0 ≤ α2 n2 L0 .
n 2
r
≥
α2 n2 L0
;
8
Note that one has 1/4 ≤ f (x, y, t) ≤ 3/4 provided kHT ω (k1 ,k2 ) k∞ ≤ 1/4 for
for (k1 , k2 ) ∈ K. By Assumption (33), the latter is guaranteed by α2 L20 /L ≤
1/4, so that, due to L ≥ 2L0 and ν2 ≥ 1/2, it is ensured by (S.44). Then,
by Lemma 7, one has
K(PΛ , PΛ′ ) ≤ 8 kΛ − Λ′ k2 ≤ 8α2 n2 L0 ≤ log |T̃ |/8
provided
(S.45)
α2 ≤ C2 (r/n)2
17
where C2 is an absolute constant. Therefore, application of Theorem 2.5
of Tsybakov (2009) yields (47) with ∆(n, L) = C α2 L0 /L where C is an
absolute constant.
Now, we denote C32 = (C2 /C1 )2−(2ν2 +1) and consider two cases. If n ≤
C3 rLν2+1/2 , choose L0 = [(C1 /C2 )(n/r)2 ]1/(2ν2 +1) which leads to α2 =
−(2ν +1)
C1 L 0 2
= C2 (r/n)2 . It is easy to check that L0 ≥ 2 and that L0 ≤ L/2,
so that
2ν2
C r 2 2ν2 +1
∆(n, L) =
.
L
n
If n > C3 rLν2 +1/2 , choose α2 = C2 (r/n)2 and set L0 = L/2. Then, (S.44)
holds and ∆(n, L) = C(r/n)2 which completes the proof of the lower bound
when r = rn,L .
Piecewise smooth graphon. Since r = r0 is a fixed quantity, without
loss of generality, we set r = 1. Let ζj = j/n, j = 1, · · · , n, be fixed. Let h
be a positive integer, 1 ≤ h ≤ n, and denote δ = 1/h. Consider a kernel
function F (x) such that F (x) is n1 > ν1 times continuously differentiable
and for any x, x′ ∈ R and some CF > 0
(S.46)
supp(F ) = (−1/2; 1/2),
|F (x) − F (x′ )| ≤ CF |x − x′ |ν1 .
It follows from (S.46) that |F (x)| ≤ CF for any x. Denote
Ψk1 ,k2 (x, y) = h−ν1 F (h(x − uk1 ))F (h(y − uk2 ))
(S.47)
where uk = (k − 1/2)δ, k = 1, · · · , h. It is easy to see that Ψk1 ,k2 (x, y) =
Ψk2 ,k1 (y, x) and, for different pairs (k1 , k2 ), functions Ψk1 ,k2 (x, y) have disjoint supports. Similar to the case of the piecewise constant graphon, consider an even L0 ≤ L/2 and a set of vectors ω ∈ {0, 1}L0 with exactly L0 /2
nonzero entries. By Lemma 4.10 of Massart (2007), there exists a subset T
of those vectors such that (S.43) holds for any ω, ω ′Q∈ T . Let again Tk1 ,k2
be the copies of T for (k1 , k2 ) ∈ K and denote T̃ = (k1 ,k2)∈K Tk1 ,k2 where
K = {(k1 , k2 ) : 1 ≤ k1 < k2 ≤ h}. Then,
log |T˜ | = h(h + 1)/2 log |T | ≥ 0.04 L0 h2 .
(S.48)
For any ω̃ ∈ T̃ and l = L0 + 1, · · · , 2L0 , define
√
(ω̃ )
(S.49)
v1 (x, y) = L/2;
(ω̃ )
vl
(x, y) = α
h
X
h
X
(k ,k )
1 2
ω l−L
[Ψk1 ,k2 (x, y) + Ψk2 ,k1 (x, y)].
0
k1 =1 k2 =k1 +1
18
It is easy to see that v(ω̃ ) (x, y) = v(ω̃ ) (y, x) for any x, y ∈ [0, 1]. Now we
need to check that conditions (42) and (43) hold.
Note that for any x, y, x′ , y ′ , due to (S.46), obtain
|Ψk1 ,k2 (x, y) − Ψk1 ,k2 (x′ , y ′ )| ≤ h−ν1 |F (h(x − uk1 ) − F (h(x′ − uk1 )||F (h(y − uk2 )|
(S.50)
ν
+|F (h(y − uk2 ) − F (h(y ′ − uk2 )||F (h(x′ − uk1 )| ≤ Cψ |x − x′ | + |y − y ′ | 1 ,
where constant Cψ depends only on CF and ν1 . Since functions Ψk1 ,k2 (x, y)
have disjoint supports for different pairs of indices (k1 , k2 ), the sums below
have at most four nonzero terms. Then, (S.50) implies that
(ω̃ )
vl
(ω̃ )
(x, y) − vl
≤ 4α Cψ
(x′ , y ′ ) ≤ α
k1 ,k2 =1
ν
|x − x | + |y − y | 1 ,
′
′
h
X
|Ψk1 ,k2 (x, y) − Ψk1 ,k2 (x′ , y ′ )|
so that (42) holds if α ≤ K1 /(4Cψ ). Also, it is easy to check that, by (S.46),
(ω̃ )
one has [vl (x, y)]2 ≤ Cv2 α2 h−2ν1 where Cv depends only on CF and ν1 .
Therefore,
2L0
X
(ω̃ )
(l − 1)2ν2 [vl
l=L0 +1
2 +1
(x, y)]2 ≤ Cv2 α2 h−2ν1 L2ν
.
0
Hence, both assumptions, (42) and (43) are valid provided
p
−(ν +1/2)
K2 /Cv hν1 L0 2
.
(S.51)
α ≤ min K1 /(4Cψ ),
Denote by Λ and Λ′ the probability tensors corresponding to ω̃ and ω̃ ′ in T̃ ,
respectively. Let ṽ(x, y) be a vector with ṽ1 (x, y) = 0 and ṽl (x, y) = vl (x, y)
for l ≥ 2. By (33) and (S.49), since Ψk1 ,k2 (x, y) have disjoint supports, we
derive that for any ω̃ ∈ T̃ one has
√
√
|Λi,j,l − 1/2| ≤ kHṽ(x, y)k∞ ≤ kṽ(x, y)k1 / L ≤ Cv αL0 h−ν1 / L.
Hence, Λi,j,l , Λ′i,j,l ∈ [1/4; 3/4] provided
(S.52)
α≤
√
L hν1 /(4Cv L0 ).
19
Then, by Lemma 7, since each ω ∈ T has exactly L0 /2 nonzero entries
K(PΛ , PΛ′ ) ≤ 8kΛ − Λ′ k2 ≤ 8α2 L0 /2
= 16α2 h−2ν1 L0
n
X
h
X
i,j k1 ,k2
= 16α2 h−2ν1 L0
" n h
XX
i=1 k=1
i j 2
,
4 Ψk1 ,k2
n n
n X
h
X
i,j k1 ,k2
F 2 (h(i/n − uk1 ))F 2 (h(j/n − uk2 ))
#2
F 2 (h(i/n − uk ))
.
Here,
(S.53)
h
n X
X
i=1 k=1
≈n
F 2 (h(i/n − uk )) =
h Z uk +δ/2
X
k=1
uk −δ/2
h
X
k=1
X
uk −δ/2<i/n≤uk +δ/2
F 2 (h(i/n − uk ))dx = n
Z
1/2
F 2 (h(i/n − uk ))
F 2 (z)dz,
−1/2
so that
(S.54)
K(PΛ , PΛ′ ) ≤ 16 kF k42 α2 h−2ν1 n2 L0
where kF k2 is the L2 -norm of F . On the other hand, due to the first inequality in (S.43) and (S.53), obtain
2L
n
n
r
r
kΛ − Λ′ k2
2α2 X0 X X X X (k1 ,k2 )
(k ,k )
− ω ′l 1 2 ]2 [Ψk1 ,k2 (i/n, j/n)]2
[ω l
≥ 2
n2 L
n L
l=L0 +1 k1 =1 k2 =k1 +1 i=1 j=1
" n h
#2
α2 h−2ν1 L0 X X 2
α2 h−2ν1 L0
≥
F
(h(i/n
−
u
))
kF k42 .
≥
k
4 n2 L
8L
i=1 k=1
Application of Theorem 2.5 of Tsybakov (2009) yields that (47) holds with
∆(n, L) = C α2 h−2ν1 L0 /L provided K(PΛ , PΛ′ ) ≤ 1/8 log |T̃ |, which, due
to (S.48) and (S.54), is guaranteed by
4kF k42 α2 h−2ν1 n2 L0 ≤ (1/8) (0.04L0 h2 )/2
and leads to the following restriction on α:
√
(S.55)
α ≤ hν1 +1 n−1 /(40 2kF k22 ).
20
Set
(S.56)
h=n
1
ν1 +1
, L0 = min n
2ν1
(ν1 +1)(2ν2 +1)
L
,
2
(ν1 +1)(2ν2 +1)
2ν1
L
, nL =
2
and consider two cases.
If n ≤ nL , then L0 is given by the first expression in (S.56) and inequalities (S.51), (S.52) and (S.55) hold if α2 is a small enough absolute
constant that depends on ν1 , ν2 , K1 and K2 only. In this case, ∆(n, L) =
−
4ν1 ν2
C L−1 n (ν1 +1)(2ν2 +1) and (48) is valid.
If n > nL , then L0 = L/2 and again all inequalities (S.51), (S.52) and
(S.55) hold if α2 is a small enough absolute constant that depends on
ν1 , ν2 , K1 and K2 only. In this case, ∆(n, L) = C n
the proof.
2ν1
1 +1
−ν
which completes
9.6. Proofs of supplementary statements. In this section, we formulate
and prove supplementary lemmas used in the proofs of other statements.
Lemma 2. Let ai be independent Bernoulli(θi ) variables and consider
vectors a and θ with components ai and θi , respectively. Then, for any vector
z and any positive t and α one has
(S.57)
P 2|zT (a − θ)| > αkzk2 + 2t/α ≤ 2e−t , E exp(zT (a − θ) ≤ exp(kzk2 /8).
Proof. Validity of the Lemma follows from Hoeffding inequality (see, e.g.,
Massart (2007)).
Lemma 3. Let ξ be a vector with independent components, Eξ = 0 and
such that for any vector z and some σ > 0 one has
(S.58)
E exp(zT ξ) ≤ exp(kzk2 σ 2 /2).
Then, for any matrix A and any positive t one has
(S.59)
P kAξk2 > σ 2 2kAk2 + 3 kAk2op t ≤ e−t .
Proof. Denote Σ = AT A and use Theorem 2.1 of Hsu et al. (2012). Obtain,
for any t > 0,
q
2
2
2
≤ e−t .
P kAξk > σ Tr(Σ) + 2 t Tr(Σ ) + 2t kΣkop
21
√
Note that Tr(Σ2 ) ≤ kΣkop Tr(Σ) and 2 xy ≤ x + y for any positive x
and y. In order to complete the proof, recall that kΣkop = kAk2op and
Tr(Σ) = kAk2 .
Corollary 4. Let ai be independent Bernoulli(θi ) variables and consider vectors a and θ with components ai and θi , respectively. Then, for any
matrix A and any positive t one has
!
2
2
3t
kAk
kAk
op
(S.60)
P kA(a − θ)k2 >
≤ e−t .
+
2
4
Proof. Denote ξ = a − θ and note that, due to (S.57), condition (S.58)
holds with σ = 1/2.
Lemma 4. (The packing lemma). Let Z(m, n) ⊆ M(m, n) be a collection of clustering matrices. Then, there exists a subset Sn,m (r) ⊂ Z(m, n)
such that for C1 , C2 ∈ Z(m, n) one has kC1 − C2 kH = kC1 − C2 k2 ≥ r and
log |Sn,m (r)| ≥ log |Z(m, n)| − r log(nem/r).
Proof. For any clustering matrix C define the r-neighborhood of C as
n
o
B(C, r) = C̃ ∈ Z(m, n) : kC̃ − CkH ≤ r .
Let S̃n,m (r) be the covering set of Z(m, n) and |S̃n,m (r)| be the covering
number, the smallest number of closed balls of radius r whose union covers
Z(m, n). Let |Sn,m (r)| be the packing number of Z(m, n), the largest number
of points with the distance at least r between them. It is known (see Pollard
(1990), page 10) that
(S.61)
|S̃n,m (r)| ≤ |Sn,m (r)| ≤ |S̃n,m (r/2)|
P
Note that |Z(m, n)| ≤ C∈S̃n,m (r) |B(C, r)| ≤ |S̃n,m (r)| maxC∈S̃n,m (r) |B(C, r)|,
so that
,
(S.62)
|S̃n,m (r)| ≥ |Z(m, n)|
max
C∈S̃n,m (r)
|B(C, r)|
and, also,
(S.63)
ne r
nem r
n
|B(C, r)| ≤
mr ≤
mr =
.
r
r
r
22
Now, combining (S.61) – (S.63), obtain log |Sn,m (r)| ≥ log |S̃n,m (r)| ≥ log |Z(m, n)|−
r log(nem/r) which completes the proof.
Lemma 5. Let γm and n/m be positive integers. Then, for n0 ≥ 1 and
γm ≥ 2, one has
(S.64)
(S.65)
(S.66)
γn
log(γm);
log {(γn)!} − γm log {(n/m)!} ≥
4
n
γn
n0
mγ
n n0
0
(n0 − 1)! ≥ n0 log
;
log
+
log
m
n0
2
2
n0
6γn log(γm) − γn log(32 mγ e) ≥ 0.
Proof. Note that due to (S.12), one has
γm
log
A1 = log {(γn)!}−γm log {(n/m)!} ≈ γn log(γm)−
2
Consider a function
3γn
x
F (x) =
log x − log
4
2
2πγn
x
+
2πn
1
+ log(2πγn).
m
2
1
log(2πγn).
2
It is easy to check that F (1) = 0 and that F ′ (x) = 0.75 γn/x−0.5 log(2πγn/(xe)).
Replacing 2πγn/(xe) in F ′ (x) by z and noting that the inequality 3e/(4π)z >
log z is true for any z > 0, we confirm that F ′ (x) > 0. so that F (x) > 0 for
any x ≥ 1. Finally, in order to prove (S.64), observe that A1 = F (γm) +
γn log(γm)/4.
For the sake of proving (S.65), note that for every n0 ≥ 1 one has log[(n0 −
1)!] ≥ 0.5 n0 log(n0 /2) and apply the first inequality in (S.12).
The validity of inequality (S.66) follows from γm ≥ 2 and the fact that
log(64e) < 6.
Lemma 6. Let ai be independent Bernoulli(θi ) variables with 0 ≤ θi ≤
ρ < 1/2, i = 1, · · · , n. Denote ξi = ai − θi and
(S.67)
n
1 X
ξi .
ζ=√
n
i=1
Then, ζ is a sub-Gaussian random variable with sub-Gaussian norm
1/p 3e2
1
−1/2
p
E|ζ|
≤
(S.68)
kζkψ2 = sup p
max ρ,
.
2
n
p≥1
23
Proof. First, observe that we need to check inequality of the type (S.68)
p ≤
only for an even integer p = 2m. Recall that kζkψ2 ≤ K0 provided
√ E|ζ|
√ p
p/2
2m
2
2m
(K0 p) = (K0 p) for every p ≥ 1. Suppose that E|ζ| ≤ (K1 2m) =
(2K12 m)m for every integer m ≥ 1 and some some K1 > 0. For any p ≥ 1
choose m such that 2m − 2 < p ≤ 2m which implies 2m < p + 2. Then,
p
p
p
E|ζ|p ≤ Eζ 2m 2m ≤ (2K12 m)m 2m ≤ K12 (p + 2) 2
√
Observing that p + 2 ≤ 3p for p ≥ 1, obtain that E|ζ|p ≤ [( 3K1 )2 p]p/2 .
Hence, (S.68) is valid provided that for any m ≥ 1
!2m
n
X
1
(S.69) E(ζ)2m = E √
ξi
≤ (e2 mρn )m , ρn = max(ρ, 1/n).
n
i=1
Note that
E(ζ)2m =
X E(ξim1 . . . ξiml )
1
l
nm
where the sum is taken over all positive integers l, m1 , . . . , ml , i1 , . . . , il such
that 2m = m1 + m2 + ml and i1 , . . . , il are distinct. Note that Eξi = 0
for every i, hence, in the sum the terms with ml = 1 are equal to zero.
Therefore,
(S.70)
E(ζ)2m ≤
X
S(m, l)P (n, l)
l
E(ξim1 1 . . . ξiml l )
nm
where S(m, l) is the number of partitions of 2m unlabeled objects into l
distinct subparts of size at least two, and P (m, l) is the number of ways of
choosing l variables out of n (order matters since powers can be different):
2m − l − 1
n!
S(m, l) =
, P (n, l) =
≤ nl .
l−1
(n − l)!
Here, 2m − l − 1 ≥ l − 1, so that l ≤ m. Observe that due to
Eξik = θi (1 − θi )[(−θi )k−1 + (1 − θi )k−1 ] ≤ ρ,
(S.70) yields
E(ζ)2m ≤
min(m,n)
X
l=1
l
2m − l − 1
ρ
.
nm−l
l−1
By considering the cases ρ ≥ 1/n and ρ < 1/n separately, it is easy to show
that
ρl
≤ [max(ρ, n−1 )]m = ρm
n,
nm−l
24
so that
2m
E(ζ)
≤
ρm
n
min(m,n)
X
l=0
2m − l
l
≤ ρm
n
∞
X
(2m)l
l=0
l!
2m
≤ ρm
= (e2 ρn )m
n e
which implies (S.69) and, hence, (S.68).
Corollary 5. Let a ∈ [0, 1]n be a random vector with independent
Bernoulli(θi ) components. Let ξ = a − θ and Z ∈ M(M, N ) be a clustering
matrix with S2 = ZT Z = diag(N1 , · · · , NM ). Define η = S−1 ZT ξ. Then,
vector η has independent sub-Gaussian components with
kη k k2ψ2 ≤
3e2
max(kθk∞ , Nk−1 ).
2
Proof. Validity of the Lemma follows from the fact that each η k is of the
form (S.67) and is evaluated using a distinct set of components of vector ξ.
Lemma 7. (Proposition 4.2 of Gao, Lu and Zhou (2015)) Let
θ = {θi,j } ∈ [0, 1]n×n . Let Pθi,j denote the probability of Bernoulli (θi,j ) and
the probability Pθ stands for the product measure Pθ = ⊗i,j Pθi,j . Then, for
any θ, θ ′ ∈ [1/4, 3/4]n×n one has
X
′ 2
(θi,j − θi,j
) .
K(Pθ , Pθ ′ ) ≤ 8
i,j
Marianna Pensky
Department of Mathematics
University of Central Florida
Orlando FL 32816-1354, USA
E-mail: [email protected]
25
| 10math.ST
|
Low Rank Approximation with Entrywise `1-Norm Error
arXiv:1611.00898v1 [cs.DS] 3 Nov 2016
Zhao Song∗
[email protected]
UT-Austin
David P. Woodruff
[email protected]
IBM Almaden
Peilin Zhong
[email protected]
Columbia University
Abstract
We study the `1 -low rank approximation problem, where for a given n × d matrix A and
b for which
approximation factor α ≥ 1, the goal is to output a rank-k matrix A
b 1 ≤α·
kA − Ak
min
rank-k matrices A0
kA − A0 k1 ,
Pn Pd
where for an n × d matrix C, we let kCk1 = i=1 j=1 |Ci,j |. This error measure is known to
be more robust than the Frobenius norm in the presence of outliers and is indicated in models
where Gaussian assumptions on the noise may not apply. The problem was shown to be NP-hard
by Gillis and Vavasis and a number of heuristics have been proposed. It was asked in multiple
places if there are any approximation algorithms.
We give the first provable approximation algorithms for `1 -low rank approximation, showing
that it is possible to achieve approximation factor α = (log d)·poly(k) in nnz(A)+(n+d) poly(k)
time, where nnz(A) denotes the number of non-zero entries of A. If k is constant, we further
improve the approximation ratio to O(1) with a poly(nd)-time algorithm. Under the Exponential
1
Time Hypothesis, we show there is no poly(nd)-time algorithm achieving a (1 + log1+γ
)(nd)
approximation, for γ > 0 an arbitrarily small constant, even when k = 1.
We give a number of additional results for `1 -low rank approximation: nearly tight upper and
lower bounds for column subset selection, CUR decompositions, extensions to low rank approximation with respect to `p -norms for 1 ≤ p < 2 and earthmover distance, low-communication
distributed protocols and low-memory streaming algorithms, algorithms with limited randomness, and bicriteria algorithms. We also give a preliminary empirical evaluation.
∗
Work done while visiting IBM Almaden.
1
Introduction
Two well-studied problems in numerical linear algebra are regression and low rank approximation.
In regression, one is given an n × d matrix A, and an n × 1 vector b, and one seeks an x ∈ Rd which
minimizes kAx − bk under some norm. For example, for least squares regression one minimizes
kAx − bk2 . In low rank approximation, one is given an n × d matrix A, and one seeks a rank-k
b which minimizes kA − Ak
b under some norm. For example, in Frobenius norm low rank
matrix A
1/2
b F = P (Ai,j − A
bi,j )2
approximation, one minimizes kA − Ak
. Algorithms for regression are
i,j
often used as subroutines for low rank approximation. Indeed, one of the main insights of [DMM06b,
DMM06a, Sar06, DMM08, CW09] was to use results for generalized least squares regression for
Frobenius norm
P low rank approximation. Algorithms for `1 -regression, in which one minimizes
kAx − bk1 = i |(Ax)i − bi |, were also used [BD13, SW11] to fit a set of points to a hyperplane,
which is a special case of entrywise P
`1 -low rank approximation, the more general problem being to
b
bi,j |.
find a rank-k matrix A minimizing i,j |Ai,j − A
Randomization and approximation were introduced to significantly speed up algorithms for these
problems, resulting in algorithms achieving relative error approximation with high probability. Such
algorithms are based on sketching and sampling techniques; we refer to [Woo14b] for a survey. For
least squares regression, a sequence of work [Sar06, CW13, MM13, NN13, LMP13, BDN15, Coh16]
shows how to achieve algorithms running in nnz(A) + poly(d) time. For Frobenius norm low rank
approximation, using the advances for regression this resulted in nnz(A) + (n + d) poly(k) time
algorithms. For `1 -regression, sketching and sampling-based methods [Cla05, SW11, CDMI+ 13,
CW13, MM13, LMP13, WZ13, CW15b, CP15] led to an nnz(A) + poly(d) time algorithm.
Just like Frobenius norm low rank approximation is the analogue of least squares regression,
entrywise `1 -low rank approximation is the analogue of `1 -regression. Despite this analogy, no
non-trivial upper bounds with provable guarantees are known for `1 -low rank approximation. Unlike
Frobenius norm low rank approximation, which can be solved exactly using the singular value
decomposition, no such algorithm or closed-form solution is known for `1 -low rank approximation.
Moreover, the problem was recently shown to be NP-hard [GV15]. A major open question is whether
there exist approximation algorithms, sketching-based or otherwise, for `1 -low rank approximation.
Indeed, the question of obtaining betters algorithms was posed in section 6 of [GV15], in [Exc13],
and as the second part of open question 2 in [Woo14b], among other places. The earlier question of
NP-hardness was posed in Section 1.4 of [KV09], for which the question of obtaining approximation
algorithms is a natural followup. The goal of our work is to answer this question.
We now formally define the `1 -low rank approximation problem: we are given an n × d matrix
A and approximation factor α ≥ 1, and we would like, with large constant probability, to output a
b for which
rank-k matrix A
b 1 ≤α·
kA − Ak
min
rank-k matrices A0
kA − A0 k1 ,
(1)
P P
where for an n × d matrix C, we let kCk1 = ni=1 dj=1 |Ci,j |. This notion of low rank approximation has been proposed as a more robust alternative to Frobenius norm low rank approximation [KK03, KK05, KLC+ 15, Kwa08, ZLS+ 12, BJ12, BD13, BDB13, MXZZ13, MKP13, MKP14,
MKCP16, PK16], and is sometimes referred to as `1 -matrix factorization or robust PCA. `1 -low rank
approximation gives improved results over Frobenius norm low rank approximation since outliers
are less exaggerated, as one does not square their contribution in the objective. The outlier values
are often erroneous values that are far away from the nominal data, appear only a few times in the
data matrix, and would not appear again under normal system operation. These works also argue
`1 -low rank approximation can better handle missing data, is appropriate in noise models for which
1
the noise is not Gaussian, e.g., it produces the maximum likelihood estimator for Laplacian noise
[Gao08, KAC+ 08, VT01], and can be used in image processing to prevent image occlusion [YZD12].
To see that `1 -low rank approximation and Frobenius
norm
low rank approximation can give
very different results, consider the n × n matrix A = n0 B0 , where B is any (n − 1) × (n − 1)
matrix with kBkF < n. The best rank-1 approximation with Frobenius norm error is given by
b = n · e1 e> , where e1 is the first standard unit vector. Here A
b ignores all but the first row and
A
1
column of A, which may be undesirable in the case that this row and column represent an outlier.
b 1 = kBk1 . If, for example, B is the all 1s matrix, then A
b = [0, 0; 0, B] is a rank-1
Note kA − Ak
b 1 = n, and therefore this solution is a much better solution to the
approximation for which kA − Ak
>
2
`1 -low rank approximation problem than n · e1 e>
1 , for which kA − n · e1 e1 k1 = (n − 1) .
Despite the advantages of `1 -low rank approximation, its main disadvantage is its computationally intractability. It is not rotationally invariant and most tools for Frobenius low rank approximation do not apply. To the best of our knowledge, all previous works only provide heuristics.
We provide hard instances for previous work in Section L, showing these algorithms at best give a
poly(nd)-approximation (though even this is not shown in these works). We also mention why a
related objective function, robust PCA [WGR+ 09, CLMW11, NNS+ 14, NYH14, CHD16, ZZL15],
does not give a provable approximation
factor for `1 -low rank approximation. Using that for√ an
√
n × d matrix C, kCkF ≤ kCk1 ≤ ndkCkF , a Frobenius norm low rank approximation gives a nd
approximation for `1 -low rank approximation. A bit better is to use algorithms for low rank approxb
b
imation with respect to the sum of distances,
Pn i.e., to find a rank-k matrix A minimizing kA − Ak1,2 ,
where for an n × d matrix C, kCk1,2 = i=1 kCi k2 , where Ci is the i-th row of C. A sequence of
work [DV07, FMSW10, FL11, SV12, CW15a] shows how to obtain an O(1)-approximation
√ to this
problem in nnz(A)
√ + (n + d) poly(k) + exp(k) time, and using that kCk1,2 ≤ kCk1 ≤ dkCk1,2
results in an O( d)-approximation.
There are also many variants of Frobenius norm low rank approximation for which nothing is
known for `1 -low rank approximation, such as column subset selection and CUR decompositions, distributed and streaming algorithms, algorithms with limited randomness, and bicriteria algorithms.
Other interesting questions include low rank approximation for related norms, such as `p -low rank
bi,j )p . Note for
b minimizing Pn Pd (Ai,j − A
approximation in which one seeks a rank-k matrix A
i=1
j=1
1 ≤ p < 2 these are also more robust than the SVD.
1.1
Our Results
We give the first efficient algorithms for `1 -low rank approximation with provable approximation
guarantees. By symmetry of the problem, we can assume d ≤ n. We first give an algorithm
which runs in O(nnz(A)) + n · poly(k) time and solves the `1 -low rank approximation problem
with approximation factor (log
√ d) · poly(k). This is an exponential improvement over the previous
approximation factor of O( d), provided k is not too large, and is polynomial time for every k.
Moreover, provided nnz(A) ≥ n · poly(k), our time is optimal up to a constant factor as any relative
error algorithm must spend nnz(A) time. We also give a hard instance for our algorithm ruling out
log d
1/2−γ approximation for arbitrarily small constant γ > 0, and hard instances for a general
k log k + k
class of algorithms based on linear sketches, ruling out k 1/2−γ approximation.
e
Via a different algorithm, we show how to achieve an O(k)-approximation
factor in
e
e 2)
O(k)
O(k
poly(n)d
2
time. This is useful for constant k, for which it gives an O(1)-approximation
in poly(n) time, improving the O(log d)-approximation for constant k of our earlier algorithm.
The approximation ratio of this algorithm, although O(1) for constant k, depends on k. We also
b in poly(n) time for constant k for which kA − Ak
b 1 ≤
show one can find a rank-2k matrix A
2
C minrank-k matrices A0 kA − A0 k1 , where C > 1 is an absolute constant independent of k. We refer
b instead of a rankto this as a bicriteria algorithm. Finally, one can output a rank-k matrix A,
b
2k matrix A, in poly(n) time with the same absolute constant C approximation factor, under an
b are integers in the range {−b, −b+1, . . . , b} for an integer
additional assumption that the entries of A
b ≤ poly(n). Unlike our previous algorithms, this last algorithm has a bit complexity assumption.
Under the Exponential Time Hypothesis (ETH), we show there is no poly(n)-time algorithm
1
achieving a (1 + log1+γ
)-approximation, for γ > 0 an arbitrarily small constant, even when k = 1.
(n)
The latter strengthens the NP-hardness result of [GV15].
We also give a number of results for variants of `1 -low rank approximation which are studied for
Frobenius norm low rank approxiation; prior to our work nothing was known about these problems.
Column Subset Selection and CUR Decomposition: In the column subset selection
problem, one seeks a small subset C of columns of A for which there is a matrix X for which
kCX − Ak is small, under some norm. The matrix CX provides a low rank approximation to
A which is often more interpretable, since it stores actual columns of A, preserves sparsity, etc.
These have been extensively studied when the norm is the Frobenius or operator norm (see, e.g.,
[BMD09, DR10, BDM11] and the references therein). We initiate the study of this problem with respect to the `1 -norm. We first prove an existence result, namely, that there exist matrices A for which
any subset C of poly(k) columns satisfies minX kCX − Ak1 ≥ k 1/2−γ · minrank-k matrices A0 kA − A0 k1 ,
where γ > 0 is an arbitrarily small constant. This result is in stark contrast to the Frobenius
norm for which for every matrix there exist O( k ) columns for which the approximation factor is
1 + . We also show that our bound is nearly optimal in this √
regime, by showing for every matrix there exists a subset of O(k log k) columns providing an O( k log k)-approximation. One can
find such columns in poly(n)dO(k log k) time by enumerating and evaluating the cost of each subset.
Although
this is exponential in k, we show it is possible to find O(k log k) columns providing an
√
O( k log k log d)-approximation in polynomial time for every k.
We extend these results to the CUR decomposition problem (see, e.g., [DMM08, BW14]), in
which one seeks a factorization CU R for which C is a subset of columns of A, R is a subset of rows
of A, and kCU R − Ak is as small as possible. In the case of Frobenius norm, one can choose O(k/)
columns and rows, have rank(U ) = k, have kCU R − AkF be at most (1 + ) times the optimal cost,
and find the factorization in nnz(A) log n + n · poly(k/) time [BW14]. Using our column subset
selection results, we give an nnz(A) + n · poly(k) time algorithm choosing O(k log k) columns and
rows, for which rank(U ) = k, and for which kCU R − Ak1 is poly(k) log d times the cost of any
rank-k approximation to A.
`p -Low Rank Approximation and EMD-Low Rank Approximation: We also give the
first algorithms with provable approximation guarantees for the `p -low rank approximation problem,
1 ≤ p < 2, in which we are given an n × d matrix A and approximation factor α ≥ 1, and would
b for which
like, with large constant probability, to output a rank-k matrix A
b pp ≤ α ·
kA − Ak
min
rank-k matrices A0
kA − A0 kpp ,
(2)
P P
where for an n × d matrix C, kCkpp = ni=1 dj=1 |Ci,j |p . We obtain similar algorithms for this
problem as for `1 -low rank approximation. For instance, we obtain an nnz(A) + n · poly(k) time
algorithm with approximation ratio (log d) · poly(k). We also provide the first low rank approximab with a (log2 d) poly(k)
tion with respect to sum of earthmover distances (of the n rows of A and A)
approximation factor. This low rank error measure was used, e.g., in [SL09]. Sometimes such
applications also require a non-negative factorization, which we do not provide.
Distributed/Streaming Algorithms, and Algorithms with Limited Randomness: There
is a growing body of work on low rank approximation in the distributed (see, e.g., [TD99, QOSG02,
3
BCL05, BRB08, MBZ10, FEGK13, PMvdG+ 13, KVW14, BKLW14, BLS+ 16, BWZ16, WZ16]) and
streaming models (see, e.g., [McG06, CW09, KL11, GP13, Lib13, KLM+ 14, Woo14a]), though almost exclusively for the Frobenius norm. One distributed model is the arbitrary partition model
[KVW14] in which there are s servers, each holding an n×d matrix Ai , and they would like to output
a k × d matrix V > for which minU kU V > − Ak is as small as possible (or, a centralized coordinator
e
may want to output this). We give O(snk)-communication
algorithms achieving a poly(k, log(n))approximation for `1 -low rank approximation in the arbitrary partition model, which is optimal for
this approximation factor (see [BW14] where lower bounds for Frobenius norm approximation with
poly(n) multiplicative approximation were shown - such lower bounds apply to `1 low rank approximation). We also consider the turnstile streaming model [Mut05] in which we receive positive or
negative updates to its entries and wish to output a rank-k factorization at the end of the stream.
e
We give an algorithm using O(nk)+poly(k)
space to achieve a poly(k, log(n))-approximation, which
is space-optimal for this approximation factor, up to the degree of the poly(k) factor. To obtain
e
these results, we show our algorithms can be implemented using O(dk)
random bits.
We stress for all of our results, we do not make assumptions on A such as low coherence or
condition number; our results hold for any n × d input matrix A.
We report a promising preliminary empirical evaluation of our algorithms in Section L.
Remark 1.1. We were just informed of the concurrent and independent work [CGK+ 16], which also
obtains approximation algorithms for `1 -low rank approximation. That paper obtains a 2O(k) log de
approximation in (log d)k poly(nd) time. Their algorithm is not polynomial time once k = Ω(log
d),
whereas we obtain a polynomial time algorithm for every k (in fact nnz(A) + (n + d) poly(k) time).
Our approximation factor is also poly(k) log d, which is an exponential improvement over theirs in
terms of k. In [CGK+ 16] they also obtain a 2k -approximation in poly(nd)dO(k) time. In contrast,
e
e 2
e
we obtain an O(k)-approximation
in poly(nd)dO(k) 2O(k ) time. The dependence in [CGK+ 16] on k
in the approximation ratio is exponential, whereas ours is polynomial.
1.2
Technical Overview
Initial Algorithm and Optimizations: Let A∗ be a rank-k matrix for which kA − A∗ k1 =
minrank-k matrices A0 kA − A0 k1 . Let A∗ = U ∗ V ∗ be a factorization for which U ∗ is n × k and V ∗
is k × d. Suppose we somehow
knew U ∗ and consider the multi-response `1 -regression problem
P
minV kU ∗ V − Ak1 = minV di=1 kU ∗ Vi − Ai k1 , where Vi , Ai denote the i-th columns of V and
A, respectively. We could solve this with linear programming though this is not helpful for our
argument here.
Instead, inspired by recent advances in sketching for linear algebra (see, e.g., [Woo14b]
for a surP
vey), we could choose a random matrix S and solve minV kSU ∗ V − SAk1 = minV di=1 k(SU ∗ )Vi −
SAi k1 . If V is an approximate minimizer of the latter problem, we could hope V is an approximate
minimizer of
former problem. If also S has a small number t of rows, then we could instead
Pthe
d
solve minV i=1 k(SU ∗ )Vi − SAi k2 , that is, minimize the sum of Euclidean norms rather than the
sum of `1√
-norms. Since t−1/2 k(SU ∗ )Vi − SAi k1 ≤ k(SU ∗ )Vi − SAi k2 ≤ k(SU ∗ )Vi − SAi k1 , we would
obtain a t-approximation
to the problem minV kSU ∗ V − SAk1 . A crucial observation is that the
Pd
solution to minV i=1 k(SU ∗ )Vi − SAi k2 is given by V = (SU ∗ )† SA, which implies that V is in the
row span of SA. If also S were oblivious to U ∗ , then we could compute SA without ever knowing
U ∗ . Having a low-dimensional space containing a good solution in its span is our starting point.
For this to work, we need a distribution on oblivious matrices S with a small number of rows,
for which an approximate minimizer V to minV kSU ∗ V − SAk1 is also an approximate minimizer
to minV kU ∗ V − Ak1 . It is unknown if there exists a distribution on S with this property. What
4
is known is that if S has O(d log d) rows, then the Lewis weights (see, e.g., [CP15] and references
therein) of the concatenated matrix [U ∗ , A] give a distribution for which the optimal V for the latter
problem is a (1 + )-approximation to the former problem; see also earlier work on `1 -leverage scores
[Cla05, DDH+ 09] which have poly(d) rows and the same (1 + )-approximation guarantee. Such
distributions are √
not helpful here as (1) they are not oblivious, and (2) the number O(d log d) of
rows gives an O( d log d) approximation factor, which is much larger than what we want.
There are a few oblivious distributions S which are useful for single-response `1 -regression
min kU ∗ v − ak1 for column vectors v, a ∈ Rk [SW11, CDMI+ 13, WZ13]. In particular, if S is an
O(k log k) × n matrix of i.i.d. Cauchy random variables, then the solution v to min kSU ∗ v − Sak1 is
an O(k log k)-approximation to min kU ∗ v − ak1 [SW11]. The important property of Cauchy random
variables is that if X and Y are independent Cauchy random variables, then αX + βY is distributed
as a Cauchy random variable times |α| + |β|, for any scalars α, β ∈ R. The O(k log k) approximation
arises because all possible regression solutions are in the column span of [U ∗ , a] which is (k + 1)dimensional, and the sketch S gives an approximation factor of O(k log k) to preserve every vector
norm in this subspace. If we instead had a multi-response regression problem min kSU ∗ V ∗ − SAk1
the dimension of the column span of [U ∗ , A] would be d + k, and this approach would give an
O(d log d)-approximation. Unlike Frobenius norm multi-response regression min kSU ∗ V ∗ − SAkF ,
which can be bounded if S is a subspace embedding for U ∗ and satisfies an approximate matrix
product theorem [Sar06], there is no convenient linear-algebraic analogue for the `1 -norm.
We first note that since regression is a minimization problem, to obtain an O(α)-approximation
by solving the sketched version of the problem, it suffices that (1) for the optimal V ∗ , we have
kSU ∗ V ∗ −SAk1 ≤ O(α)kU ∗ V ∗ −Ak1 , and (2) for all V , we have kSU ∗ V −SAk1 ≥ Ω(1)·kU ∗ V −Ak1 .
We show (1) holds for α = O(log d) and any number of rows of S. Our analysis follows by
truncating the Cauchy random variables (SU ∗ Vj∗ − SAj )i for i ∈ [O(k log k)] and j ∈ [d], so
that their expectation exists, and applying linearity of expectation across the d columns. This is
inspired from an argument of Indyk [Ind06] for embedding a vector into a lower-dimensional vector
while preserving its `1 -norm; for single-response regression this is the statement that kSU ∗ v ∗ −
Sak1 = Θ(1)kU ∗ v − ak1 , implied by [Ind06]. However, for multi-response regression we have to
work entirely with expectations, rather than the tail bounds in [Ind06], since the Cauchy random
variables (SU ∗ Vj − SAj )i , while independent across i, are dependent across j. Moreover, our
O(log d)-approximation factor is not an artifact of our analysis - we show in Section G that there
is an n × d input matrix A for which
with probability
1 − 1/ poly(k), there is no k-dimensional
log d
1/2−γ
space in the span of SA achieving a t log t + k
-approximation, for S a Cauchy matrix with
t rows, where γ > 0 is an arbitrarily small constant. This shows (k log d)Ω(1) -inapproximability.
Thus, the fact that we achieve O(log d)-approximation instead of O(1) is fundamental for a matrix
S of Cauchy random variables or any scaling of it.
While we cannot show (2), we instead show for all V , kSU ∗ V − SAk1 ≥ kU ∗ V − Ak1 /2 −
O(log d)kU ∗ V ∗ − Ak1 if S has O(k log k) rows. This suffices for regression, since the only matrices V
for which the cost is much smaller in the sketch space are those providing an O(log d) approximation
in the original space. The guarantee follows from the triangle inequality: kSU ∗ V −SAk1 ≥ kSU ∗ V −
SU ∗ V ∗ k1 − kSU ∗ V ∗ − SAk1 and the fact that S is known to not contract any vector in the column
span of U ∗ if S has O(k log k) rows [SW11]. Because of this, we have kSU ∗ V − SU ∗ V ∗ k1 =
Ω(1)kU ∗ V −U ∗ V ∗ k1 = Ω(1)(kU ∗ V −Ak1 −kU ∗ V ∗ −Ak1 ), where we again use the triangle inequality.
We also bound the additive term kSU ∗ V ∗ − SAk1 by O(log d)kU ∗ V ∗ − Ak1 using (1) above.
Given that SA contains a good rank-k approximation in its row span, our algorithm with a
slightly worse poly(n) time and poly(k log(n))-approximation can be completely described here.
Let S and T1 be independent O(k log k) × n matrices of i.i.d. Cauchy random variables, and let R
5
and T2 be independent d × O(k log k) matrices of i.i.d. Cauchy random variables. Let
X = (T1 AR)† ((T1 AR)(T1 AR)† (T1 AT2 )(SAT2 )(SAT2 )† )k (SAT2 )† ,
which is the rank-k matrix minimizing kT1 ARXSAT2 − T1 AT2 kF , where for a matrix C, Ck is its
b = ARXSA as the solution to `1 -low rank
best rank-k approximation in Frobenius norm. Output A
b is a poly(k log(n))-approximation.
approximation of A. We show with constant probability that A
To improve the approximation factor, after computing SA, we `1 -project each of the rows of A
onto SA using linear programming or fast algorithms for `1 -regression [CW13, MM13], obtaining
an n × d matrix B of rank O(k log k). We then apply the algorithm in the previous paragraph with
A replaced by B. This ultimately leads to a log d · poly(k)-approximation.
To improve the running time from poly(n) to nnz(A) + n · poly(k), we show a similar analysis
holds for the sparse Cauchy matrices of [MM13]; see also the matrices in [WZ13].
CUR Decompositions: To obtain a CUR decomposition, we first find a log d·poly(k)-approximate
b as above. Let B1 be an n × k matrix whose columns span those of A,
b and
rank-k approximation A
∗
consider the regression minV kB1 V − Ak1 . Unlike the problem minV kU V − Ak1 where U ∗ was
unknown, we know B1 so can compute its Lewis weights efficiently, sample by them, and obtain a
regression problem minV kD1 (B1 V − A)k1 where D1 is a sampling and rescaling matrix. Since
kD1 (B1 V − A)k1 ≤ kD1 (B1 V − B1 V ∗ )k1 + kD1 (B1 V ∗ − A)k1 ,
where V ∗ = argminV kB1 V − Ak1 , we can bound the first term by O(kB1 V − B1 V ∗ k1 ) using that
D1 is a subspace embedding if it has O(k log k) rows, while the second term is O(1)kB1 V ∗ − Ak1
by a Markov bound. Note that kB1 V ∗ − Ak1 ≤ (log d) · poly(k) minrank-k matrices A0 kA − A0 k1 . By
switching to `2 as before, we see that Vb = (D1 B1 )† D1 A contains a (log d) poly(k)-approximation
in its span. Here D1 A is an actual subset of rows of A, as required in a CUR decomposition. Moreover the subset size is O(k log k). We can sample by the Lewis weights of Vb to obtain a subset C of O(k log k) rescaled columns of A, together with a rank-k matrix U for which
kCU R − Ak1 ≤ (log d) poly(k) minrank-k matrices A0 kA − A0 k1 .
Algorithm for Small k: Our CUR decomposition shows how we might obtain an O(1)-approximation
for constant k in poly(n) time. If we knew the Lewis weights of U ∗ , an α-approximate solution
to the problem minV kD1 (U ∗ V − A)k1 would be an O(α)-approximate solution to the problem
minV kU ∗ V √− Ak1 , where D1 is a sampling and rescaling matrix of O(k log k) rows of A. Moreover, an O( k log k)-approximate solution to minV kD1 (U√∗ V − A)k1 is given by V = (D1 U ∗ )† D1 A,
which implies the O(k log k) rows of D1 A contain an O( k log k)-approximation. For small k, we
can guess every subset of O(k log k) rows of A in nO(k log k) time (if d n, by taking transposes
at the beginning one can replace this with dO(k log k) time). For each guess, we set up the problem
minrank-k U kU (D1 A)−Ak1 . If D2 is a sampling and rescaling matrix according to the Lewis weights
of D1 A, then by a similar triangle√inequality argument as for our CUR decomposition, minimizing
kU (D1 A)D2 − AD2 k1 gives an O( k log k) approximation. By switching to `2 , this implies there is
an O(k log k)-approximation of the form AD2 W D1 A, where W is an O(k log2 k) × O(k log k) matrix
of rank k. By setting up the problem minrank-k W kAD2 W D1 A − Ak1 , one can sample from Lewis
weights on the left and right to reduce this to a problem independent of n and d, after which one
can use polynomial optimization to solve it in exp(poly(k)) time. One of our guesses D1 A will be
e
correct, and for this guess we obtain an O(k)-approximation.
For each guess we can compute its
cost and take the best one found. This gives an O(1)-approximation for constant k, removing the
O(log d)-factor from the approximation of our earlier algorithm.
6
Existential Results for Subset Selection: In our algorithm for small k, the first√step was
to show there exist O(k log k) rows of A which contain a rank-k space which is an O( k log k)approximation.
While for Frobenius norm one can find O(k) rows with an O(1)-approximation in their span,
one of our main negative results for `1 -low rank approximation is that this is impossible, showing
that the best approximation one can obtain with poly(k) rows is k 1/2−γ for an arbitrarily small
constant γ > 0. Our hard instance is an r × (r + k) matrix A in which the first k columns are i.i.d.
Gaussian, and the remaining r columns are an identity matrix. Here, r can be twice the number
of rows one is choosing. The optimal `1 -low rank approximation has cost at most r, obtained by
choosing the first k columns.
Let R ∈ Rr/2×k denote the first k entries of the r/2 chosen rows, and let y denote the first k
entries of an unchosen row. For r/2 > k, there exist many solutions x ∈ Rr/2 for which x> R = y.
However, we can show the following tradeoff:
whenever kx> R − yk1 <
√
k
poly(log k) ,
√
then kxk1 >
k
poly(log k) .
Then no matter which√linear combination x> of the rows of R one chooses to approximate y by,
k
either one incurs a poly(log
k) cost on the first k coordinates, or since A contains an identity matrix,
√
one incurs cost kxk1 >
k
poly(log k)
on the last r coordinates of x> R.
P
To show the tradeoff, consider an x ∈ Rr/2 . We decompose x = x0 + j≥1 xj , where xj agrees
1
with x on coordinates which have absolute value in the range √k log
· [2−j , 2−j+1 ], and is zero
c
k
otherwise. Here, c > 0 is a constant, and x0 denotes
the restriction of x to all coordinates of
√
1
k
absolute value at least √k logc k . Then kxk1 < logc k , as otherwise we are done. Hence, x0 has
small support. Thus, one can build a small net for all x0 vectors by choosing the support, then
placing a net on it. For xj for j > 0, the support sizes are increasing so the net size needed for
all xj vectors is larger. However, since xj has all coordinates of roughly the same magnitude on
its support, its `2 -norm is decreasing in j. Since (xj )> R ∼ N (0, kxj k22 Ik ), this makes it much less
likely that individual coordinates of (xj )> R can be large. Since this probability goes down rapidly,
we can
over the larger net size. What we show is that for any sum of the
Paffordj to union bound
k
form j≥1 x , at most 10 of its coordinates are at least log1 k in magnitude.
√
For kx> R − yk1 to be at most logckk , for at least k2 coordinates i, we must have |(x> R −
1
2
y)i | < √k log
With probability 1 − 2−Ω(k) , |yi | ≥ 100
on at least 2k
c .
3 coordinates. From the
k
k
previous paragraph, it follows there are at least k2 − 10
− k3 = Ω(k) coordinates i of x for which (1)
P
j
2
1
1
>
|(x R − y)i | < √k logc k , (2) | j≥1 xi | < log k , and (3) |yi | ≥ 100
. On these i, (x0 )> Ri must be in
1
an interval of width log1 k at distance at least 100
from the origin. Since (x0 )> R ∼ N (0, kx0 k22 Ik ), for
any value of kx0 k22 the probability this happens on Ω(k) coordinates is at most 2−Θ(k) . Since the
net size for x0 is small, we can union bound over every sequence x0 , x1 , . . . , coming from our nets.
Some care is needed to union bound over all possible subsets R of rows which can be chosen.
We handle this by conditioning on a few events of A itself, which imply corresponding events for
every subset of rows. These events are such that if R is the chosen set of half the rows, and S the
remaining set of rows of A, then the event that a constant fraction of rows in S are close to the row
span of R is 2−Θ(kr) , which is small enough to union bound over all choices of R.
Curiously, we also show there are some matrices A ∈ Rn×d for which any `1 rank-k approximation in the entire row span of A cannot achieve better than a (2 − Θ(1/d))-approximation.
7
Bicriteria Algorithm: Our algorithm for small k gives an O(1)-approximation in poly(n) time
for constant k, but the approximation factor depends on k. We show how one can find a rankb for which kA − Ak
b 1 ≤ C · OPT, where C is an absolute constant, and OPT =
2k matrix A
0
minrank-k matrices A0 kA − A k1 . We first find a rank-k matrix B 1 for which kA − B 1 k1 ≤ p · OPT for
a factor 1 ≤ p ≤ poly(n). We can use any of our algorithms above for this.
Next consider the problem minV ∈Rk×d kU ∗ V − (A − B 1 )k1 , and let U ∗ V ∗ be a best `1 -low rank
approximation to A − B 1 ; we later explain why we look at this problem. We can assume V ∗ is an
`1 well-conditioned basis [Cla05, DDH+ 09], since we can replace U ∗ with U ∗ R−1 and V ∗ with RV ∗
1
> ∗
for any invertible linear transformation R. For any vector x we then have kxk
f ≤ kx V k1 ≤ ekxk1 ,
where 1 ≤ e, f ≤ poly(k). This implies all entries of U ∗ are at most 2f kA − Bk1 , as otherwise
1
one could replace U ∗ with 0n×k and reduce the cost. Also, any entry of U ∗ smaller than kA−Bk
100enkp
∗
can be replaced with 0 as this incurs additive error OPT
100 . If we round the entries of U to integer
1
∗
multiples of kA−Bk
100enkp , then we only have O(enkpf ) possibilities for each entry of U , and still obtain
an O(1)-approximation. We refer to the rounded U ∗ as U ∗ , abusing notation.
Let D be a sampling and rescaling matrix with O(k log k) non-zero diagonal entries, corresponding to sampling by the Lewis weights of U ∗ . We do not know D, but handle this below. By the
triangle inequality, for any V ,
kD(U ∗ V − (A − B 1 ))k1 = kD(U ∗ V − U ∗ V ∗ )k1 ± kD(U ∗ V ∗ − (A − B 1 ))k1
= Θ(1)kU ∗ V − U ∗ V ∗ k1 ± O(1)kU ∗ V ∗ − (A − B 1 )k1 ,
where the Lewis weights give kD(U ∗ V −U ∗ V ∗ )k1 = Θ(1)kU ∗ V −U ∗ V ∗ k1 and a Markov bound gives
kD(U ∗ V ∗ − (A − B 1 ))k1 = O(1)kU ∗ V ∗ − (A − B 1 )k1 . Thus, minimizing kDU ∗ V − D(A − B 1 )k1
gives a fixed constant factor approximation to the problem minV ∈Rk×d kU ∗ V − (A − B 1 )k1 . The
non-zero diagonal entries of D can be assumed to be integers between 1 and n2 .
We guess the entries of DU ∗ and note for each entry there are only O(enkpf log(n2 )) possibilities.
One of our guesses corresponds to Lewis weight sampling by U ∗ . We solve for V and by the
guarantees of Lewis weights, the row span of this V provides an O(1)-approximation. We can find
the corresponding U via linear programming. As mentioned above, we do not know D, but can
enumerate over all D and all possible DU ∗ . The total time is npoly(k) .
After finding U , which has k columns, we output the rank-2k space formed by the column span
of [U, B 1 ]. By including the column span of B 1 , we ensure our original transformation of the problem minV ∈Rk×d kU ∗ · V − Ak1 to the problem minV ∈Rk×d kU ∗ · V − (A − B 1 )k1 is valid, since we can
first use the column span of B 1 to replace A with A − B 1 . Replacing A with A − B 1 ultimately
results in a rank-2k output. Had we used A instead of A − B 1 our output would have been rank
kAk1
k but would have additive error poly(k/)
. If we assume the entries of A are in {−b, −b + 1, . . . , b},
then we can lower bound the cost kU ∗ V − Ak1 , given that it is non-zero, by (ndb)−O(k) (if it is zero
then we output A) using Lemma 4.1 in [CW09] and relating entrywise `1 -norm to Frobenius norm.
We can go through the same arguments above with A − B replaced by A and our running time will
now be (ndb)poly(k) .
Hard Instances for Cauchy Matrices and More General Sketches: We consider a d × d
matrix A = Id + (log d)e>
1 e, where e1 = (1, 0, . . . , 0) and e = (1, 1, . . . , 1) and Id is the d × d identity.
For an O(k log k) × d matrix S of i.i.d. Cauchy random variables, SA = S + (log d)S1> e, where S1 is
the first column of S. For a typical column of SA, all entries are at most poly(k) log d in magnitude.
Thus, in order to approximate the first row of A, which is (log d)e, by x> SA for an x ∈ Rk log k , we
8
1
need kxk1 ≥ poly(k)
. Also kx> Sk1 = Ω(kxk1 d log d) with 1 − exp(−k log k) probability, for d large
enough, so by a net argument kxk1 ≤ poly(k) for all x.
However, there are entries of SA that are very large, i.e., about one which is r = Θ(dk log k) in
magnitude, and in general about 2i entries about r2−i in magnitude. These entries typically occur
in columns Cj of SA for which all other entries in the column are bounded by poly(k) in magnitude.
Thus, |x> Cj | ≈ r2−i for about 2i columns j. For each such column, if r2−i log d, then we incur
r log r
d log d
r2−i
cost poly(k)
in approximating the first row of A. In total the cost is poly(k)
= poly(k)
, but the optimal
cost is at most d, giving a
log d
poly(k)
log d
lower bound.
k log2 k
1/2−γ
k
lower bound for arbitrarily
lower bound. We optimize this to a
When k is large this bound deteriorates, but we also show a
small constant γ > 0. This bound applies to any oblivious sketching matrix. The idea is similar to
our row subset selection lower bound. Let A be as in our row subset selection lower bound, consider
SA, and write S = U ΣV > in its full SVD. Then SA is in the row span of the top O(k log k) rows
of V > A, since Σ only has O(k log k) non-zero singular values. Since the first k columns of A are
rotationally invariant, V > A has first k columns i.i.d. Gaussian and remaining columns equal to
V > . Call the first O(k log k) rows of V > A the matrix B. We now try to approximate a row of A
by a vector in the row span of B. There are two issues that make this setting different from row
subset selection: (1) B no longer contains an identity submatrix, and (2) the rows of B depend on
the rows of A. We handle the first issue by building nets for subsets of coordinates of x> V > rather
than x as before; since kx> V > k2 = kxk2 similar arguments can be applied. We handle the second
issue by observing that if the number of rows of B is considerably smaller than that of A, then the
distribution of B had we replaced a random row of A with zeros would be statistically close to i.i.d.
Gaussian. Hence, typical rows of A can be regarded as being independent of B.
Limited Independence, Distributed, and Streaming Algorithms: We show for an n × d
matrix A, if we left-multiply by an O(k log k) × n matrix S in which each row is an independent vece
tor of O(d)-wise
independent Cauchy random variables, SA contains a poly(k) log d-approximation
e
in its span. This allows players in a distributed model to share a common S by exchanging O(kd)
bits, independent of n. We use Lemma 2.2 of [KNW10] which shows for a reasonably
smoothP
approxP
imation f to an indicator function, E[f (X)] = E[f (Y )] + O(), where X = i ai Xi , Y = i ai Yi ,
e
a ∈ Rn is fixed, X is a vector of i.i.d. Cauchy random variables, and Y is a vector of O(1/)-wise
independent random variables.
To show the row span of SA contains a good rank-k approximation, we argue kSyk1 = Ω(kyk1 )
for a fixed y ∈ Rn with 1−exp(−k log k) probability. We apply the above lemma with = Θ(1). We
e
also need for an n × d matrix A with unit-`1 columns, that kSAk1 = O(kd).
We fool the expectation
of a truncated Cauchy by taking a weighted sum of O(log(dk)) indicator functions and applying the
e
above lemma with = Θ(1/d). An issue is there are Θ(kd)
Cauchy random variables corresponding
e
to the entries of SA, some of which can be as large as Θ(kd),
so to fool their expectation (after
2
e
e
e
truncation) we need = Θ(1/(dk)), resulting in O(dk ) seed length and ruining our optimal O(dk)
communication. We show we can instead pay a factor of k in our approximation and maintain
e
O(dk)-wise
independence. The distributed and streaming algorithms, given this, follow algorithms
for Frobenius norm low rank approximation in [KVW14, BWZ16].
Hardness Assuming Exponential Time Hypothesis: By inspecting the proof of NP-hardness
of [GV15], it at best gives a (1 + n1γ )-inapproximability for an arbitrarily small constant γ > 0. We
1
considerably strengthen this to (1 + log1+γ
)-inapproximability by taking a modified version of the
n
n × n hard instance of [GV15] and planting it in a 2o(n) × 2o(n) matrix padded with tiny values.
9
Under the ETH, the maximum cut problem that [GV15] and that we rely on cannot be solved in
2o(n) time, so our transformation is efficient. Although we use the maximum cut problem as in
[GV15] for our n × n hard instance, in order to achieve our inapproximability we need to use that
under the ETH this problem is hard to approximate even if the input graph is sparse and even up
to a constant factor; such additional conditions were not needed in [GV15].
`p -Low Rank Approximation and EMD-Low Rank Approximation: Our algorithms for
entrywise `p -Norm Error are similar to our algorithms for `1 . We use p-stable random variables in
place of Cauchy random variables, and note that the p-th power of a p-stable random variable has
similar tails to that of a Cauchy, so many of the same arguments apply. Our algorithm for EMD
low rank approximation immediately follows by embedding EMD into `1 .
Counterexamples to Heuristics: Let A = diag(n2+γ , n1.5+ , B, B) ∈ R(2n+2)×(2n+2) where
∈ (0, .5), γ > 0, and B is the n × n all 1s matrix. For this A we show the four heuristic algorithms [KK05, DZHZ06, Kwa08, BDB13] cannot achieve an nmin(γ,0.5−) approximation ratio
when the rank parameter k = 3.
1.3
Several Theorem Statements, an Algorithm, and a Roadmap
Algorithm 1 Main Meta-Algorithm
1:
2:
3:
4:
5:
6:
7:
procedure L1LowRankApprox(A, n, d, k)
. Theorem 1.2
Choose sketching matrix S (a Cauchy matrix or a sparse Cauchy matrix.)
Compute SA, form C by C i ← arg minx kxSA − Ai k1 . Form B = C · SA.
Choose sketching matrices T1 , R, D, T2 (Cauchy matrices or sparse Cauchy matrices.)
Solve minX,Y kT1 BRXY DBT2 − T1 BT2 kF .
return BRX,Y DB.
end procedure
Theorem 1.2 (Informal Version of Theorem C.6). Given A ∈ Rn×d , there is an algorithm which
in nnz(A) + (n + d) · poly(k) time, outputs a (factorization of a) rank-k matrix A0 such that with
kU V − Ak1 .
probability 9/10, kA0 − Ak1 ≤ (log d) poly(k)
min
U ∈Rn×k ,V ∈Rk×d
Theorem 1.3 (Informal Version of Theorem C.7). Given A ∈ Rn×d , there is an algorithm that
e
e 2
takes poly(n)dO(k) 2O(k ) time and outputs a rank-k matrix A0 such that, with probability 9/10,
e
kA0 − Ak1 ≤ O(k)
min
kU V − Ak1 . In addition, A0 is a CUR decomposition.
U ∈Rn×k ,V ∈Rk×d
Theorem 1.4 (Informal Version of Theorem G.28). For any k ≥ 1, and any constant c ≥ 1,
let n = k c . There exists a matrix A such that for any matrix A0 in the span of n/2 rows of A,
kA0 − Ak1 = Ω(k 0.5−α )
min
kU V − Ak1 , where α > 0 is an arbitrarily small constant.
U ∈Rn×k ,V ∈Rk×d
Road map Section A introduces some notation and definitions. Section B includes several useful
tools. We provide several `1 -low rank approximation algorithms in Section C. Section D contains
the no contraction and no dilation analysis for our main algorithm. The results for `p and earth
mover distance are presented in Section E and F. We provide our existential hardness results for
Cauchy matrices, row subset selection and oblivious subspace embeddings in Section G. We provide
our computational hardness results in Section H. We analyze limited independent random Cauchy
variables in Section I. Section K presents the results for the distributed setting. Section J presents
the results for the streaming setting. Section L contains the experimental results of our algorithm
and several heuristic algorithms, as well as counterexamples to heuristic algorithms.
10
References
[AB09]
Sanjeev Arora and Boaz Barak. Computational complexity: a modern approach. Cambridge University Press, 2009.
[ABIW09]
Alexandr Andoni, Khanh Do Ba, Piotr Indyk, and David P Woodruff. Efficient
sketches for earth-mover distance, with applications. In 50th Annual IEEE Symposium
on Foundations of Computer Science, pages 324–330, 2009.
[AGKM12]
Sanjeev Arora, Rong Ge, Ravindran Kannan, and Ankur Moitra. Computing a nonnegative matrix factorization - provably. In Proceedings of the 44th Symposium on
Theory of Computing Conference, STOC 2012, New York, NY, USA, May 19 - 22,
2012, pages 145–162, 2012.
[AIK08]
Alexandr Andoni, Piotr Indyk, and Robert Krauthgamer. Earth mover distance
over high-dimensional spaces. In Proceedings of the nineteenth annual ACM-SIAM
symposium on Discrete algorithms, pages 343–352. Society for Industrial and Applied
Mathematics, 2008.
[Bas14]
Saugata Basu. Algorithms in real algebraic geometry: a survey. arXiv preprint
arXiv:1409.1534, 2014.
[BCL05]
Zheng-Jian Bai, Raymond H Chan, and Franklin T Luk. Principal component analysis
for distributed data sets with updating. In Advanced Parallel Processing Technologies,
pages 471–483. Springer, 2005.
[BD13]
J. Paul Brooks and José H. Dulá. The `1 -norm best-fit hyperplane problem. Appl.
Math. Lett., 26(1):51–55, 2013.
[BDB13]
J. Paul Brooks, José H. Dulá, and Edward L Boone. A pure `1 -norm principal
component analysis. Computational statistics & data analysis, 61:83–98, 2013.
[BDL16]
Amitabh Basu, Michael Dinitz, and Xin Li. Computing approximate PSD factorizations. CoRR, abs/1602.07351, 2016.
[BDM11]
Christos Boutsidis, Petros Drineas, and Malik Magdon-Ismail. Near optimal columnbased matrix reconstruction. In IEEE 52nd Annual Symposium on Foundations of
Computer Science, FOCS 2011, Palm Springs, CA, USA, October 22-25, 2011, pages
305–314, 2011.
[BDN15]
Jean Bourgain, Sjoerd Dirksen, and Jelani Nelson. Toward a unified theory of sparse
dimensionality reduction in euclidean space. In Proceedings of the Forty-Seventh Annual ACM on Symposium on Theory of Computing, STOC 2015, Portland, OR, USA,
June 14-17, 2015, pages 499–508, 2015.
[BGS98]
Mihir Bellare, Oded Goldreich, and Madhu Sudan.
Free bits, pcps, and
nonapproximability—towards tight results. SIAM Journal on Computing, 27(3):804–
915, 1998.
[BIRW16]
Arturs Backurs, Piotr Indyk, Ilya Razenshteyn, and David P Woodruff. Nearlyoptimal bounds for sparse recovery in generic norms, with applications to k-median
sketching. In Proceedings of the Twenty-Seventh Annual ACM-SIAM Symposium on
Discrete Algorithms, pages 318–337. SIAM, 2016.
11
[BJ12]
J. Paul Brooks and Sapan Jot. Pcal1: An implementation in r of three methods for
`1 -norm principal component analysis. Optimization Online preprint, 2012.
[BKLW14]
Maria-Florina Balcan, Vandana Kanchanapally, Yingyu Liang, and David Woodruff.
Improved distributed principal component analysis. NIPS, 2014.
[BLS+ 16]
Maria-Florina Balcan, Yingyu Liang, Le Song, David Woodruff, and Bo Xie. Distributed kernel principal component analysis. KDD, 2016.
[BMD09]
Christos Boutsidis, Michael W Mahoney, and Petros Drineas. An improved approximation algorithm for the column subset selection problem. In Proceedings of the
twentieth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 968–977.
Society for Industrial and Applied Mathematics, 2009.
[BPR96]
Saugata Basu, Richard Pollack, and Marie-Françoise Roy. On the combinatorial and
algebraic complexity of quantifier elimination. J. ACM, 43(6):1002–1045, 1996.
[BPR05]
Saugata Basu, Richard Pollack, and Marie-Francoise Roy. Algorithms in real algebraic
geometry, volume 20033. Springer, 2005.
[BRB08]
Yann-Ael Le Borgne, Sylvain Raybaud, and Gianluca Bontempi. Distributed principal
component analysis for wireless sensor networks. Sensors, 2008.
[BW14]
Christos Boutsidis and David P Woodruff. Optimal cur matrix decompositions. In
Proceedings of the 46th Annual ACM Symposium on Theory of Computing, pages
353–362. ACM, 2014.
[BWZ16]
Christos Boutsidis, David P Woodruff, and Peilin Zhong. Optimal principal component analysis in distributed and streaming models. In Proceedings of the 48th Annual
ACM SIGACT Symposium on Theory of Computing, pages 236–249. ACM, 2016.
[CDMI+ 13]
Kenneth L Clarkson, Petros Drineas, Malik Magdon-Ismail, Michael W Mahoney,
Xiangrui Meng, and David P Woodruff. The fast cauchy transform and faster robust
linear regression. In Proceedings of the Twenty-Fourth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 466–477. Society for Industrial and Applied
Mathematics, 2013.
[CGK+ 16]
Flavio Chierichetti, Sreenivas Gollapudi, Ravi Kumar, Silvio Lattanzi, and Rina Panigrahy. On robust low-rank approximation. ArXiv e-prints, 2016.
[CHD16]
Kai-Yang Chiang, Cho-Jui Hsieh, and Inderjit S Dhillon. Robust principal component
analysis with side information. In Proceedings of The 33rd International Conference
on Machine Learning, pages 2291–2299, 2016.
[CK11]
Imre Csiszar and János Körner. Information theory: coding theorems for discrete
memoryless systems. Cambridge University Press, 2011.
[Cla05]
Kenneth L Clarkson. Subgradient and sampling algorithms for `1 regression. In
Proceedings of the sixteenth annual ACM-SIAM symposium on Discrete algorithms,
pages 257–266, 2005.
12
[CLM+ 15]
Michael B Cohen, Yin Tat Lee, Cameron Musco, Christopher Musco, Richard Peng,
and Aaron Sidford. Uniform sampling for matrix approximation. In Proceedings of
the 2015 Conference on Innovations in Theoretical Computer Science, pages 181–190.
ACM, 2015.
[CLMW11]
Emmanuel J Candès, Xiaodong Li, Yi Ma, and John Wright. Robust principal component analysis? Journal of the ACM (JACM), 58(3):11, 2011.
[Coh16]
Michael B. Cohen. Nearly tight oblivious subspace embeddings by trace inequalities.
In Proceedings of the Twenty-Seventh Annual ACM-SIAM Symposium on Discrete
Algorithms, SODA 2016, Arlington, VA, USA, January 10-12, 2016, pages 278–287,
2016.
[CP15]
Michael B. Cohen and Richard Peng. `p row sampling by lewis weights. In Proceedings
of the Forty-Seventh Annual ACM on Symposium on Theory of Computing, STOC
’15, pages 183–192, New York, NY, USA, 2015. ACM.
[CW87]
Don Coppersmith and Shmuel Winograd. Matrix multiplication via arithmetic progressions. In Proceedings of the nineteenth annual ACM symposium on Theory of
computing, pages 1–6. ACM, 1987.
[CW09]
Kenneth L. Clarkson and David P. Woodruff. Numerical linear algebra in the streaming model. In Proceedings of the 41st Annual ACM Symposium on Theory of Computing, STOC 2009, Bethesda, MD, USA, May 31 - June 2, 2009, pages 205–214,
2009.
[CW13]
Kenneth L. Clarkson and David P. Woodruff. Low rank approximation and regression
in input sparsity time. In Symposium on Theory of Computing Conference, STOC’13,
Palo Alto, CA, USA, June 1-4, 2013, pages 81–90, 2013.
[CW15a]
Kenneth L Clarkson and David P Woodruff. Input sparsity and hardness for robust
subspace approximation. In Foundations of Computer Science (FOCS), 2015 IEEE
56th Annual Symposium on, pages 310–329. IEEE, 2015.
[CW15b]
Kenneth L Clarkson and David P Woodruff. Sketching for m-estimators: A unified
approach to robust regression. In Proceedings of the Twenty-Sixth Annual ACM-SIAM
Symposium on Discrete Algorithms, pages 921–939. SIAM, 2015.
[DDH+ 09]
Anirban Dasgupta, Petros Drineas, Boulos Harb, Ravi Kumar, and Michael W Mahoney. Sampling algorithms and coresets for `p regression. SIAM Journal on Computing, 38(5):2060–2078, 2009.
[DMM06a]
Petros Drineas, Michael W. Mahoney, and S. Muthukrishnan. Subspace sampling
and relative-error matrix approximation: Column-based methods. In Approximation,
Randomization, and Combinatorial Optimization. Algorithms and Techniques, 9th International Workshop on Approximation Algorithms for Combinatorial Optimization
Problems, APPROX 2006 and 10th International Workshop on Randomization and
Computation, RANDOM 2006, Barcelona, Spain, August 28-30 2006, Proceedings,
pages 316–326, 2006.
[DMM06b]
Petros Drineas, Michael W. Mahoney, and S. Muthukrishnan. Subspace sampling
and relative-error matrix approximation: Column-row-based methods. In Algorithms
13
- ESA 2006, 14th Annual European Symposium, Zurich, Switzerland, September 1113, 2006, Proceedings, pages 304–314, 2006.
[DMM08]
Petros Drineas, Michael W. Mahoney, and S. Muthukrishnan. Relative-error CUR
matrix decompositions. SIAM J. Matrix Analysis Applications, 30(2):844–881, 2008.
[DR10]
Amit Deshpande and Luis Rademacher. Efficient volume sampling for row/column
subset selection. In Foundations of Computer Science (FOCS), 2010 51st Annual
IEEE Symposium on, pages 329–338. IEEE, 2010.
[DV07]
Amit Deshpande and Kasturi R. Varadarajan. Sampling-based dimension reduction
for subspace approximation. In Proceedings of the 39th Annual ACM Symposium on
Theory of Computing, San Diego, California, USA, June 11-13, 2007, pages 641–650,
2007.
[DZHZ06]
Chris Ding, Ding Zhou, Xiaofeng He, and Hongyuan Zha. R1-pca: rotational invariant `1 -norm principal component analysis for robust subspace factorization. In
Proceedings of the 23rd international conference on Machine learning, pages 281–288.
ACM, 2006.
[Exc13]
Stack Exchange. Low-rank matrix approximation in terms of entry-wise `1 norm,
2013.
[FEGK13]
Ahmed K Farahat, Ahmed Elgohary, Ali Ghodsi, and Mohamed S Kamel. Distributed
column subset selection on mapreduce. In Data Mining (ICDM), 2013 IEEE 13th
International Conference on, pages 171–180. IEEE, 2013.
[FL11]
Dan Feldman and Michael Langberg. A unified framework for approximating and
clustering data. In Proceedings of the 43rd ACM Symposium on Theory of Computing,
STOC 2011, San Jose, CA, USA, 6-8 June 2011, pages 569–578, 2011.
[FLP15]
Dimitris Fotakis, Michael Lampis, and Vangelis Th Paschos. Sub-exponential
approximation schemes for csps: From dense to almost sparse. arXiv preprint
arXiv:1507.04391, 2015.
[FMSW10]
Dan Feldman, Morteza Monemizadeh, Christian Sohler, and David P. Woodruff. Coresets and sketches for high dimensional subspace approximation problems. In Proceedings of the Twenty-First Annual ACM-SIAM Symposium on Discrete Algorithms,
SODA 2010, Austin, Texas, USA, January 17-19, 2010, pages 630–649, 2010.
[FT07]
Shmuel Friedland and Anatoli Torokhti. Generalized rank-constrained matrix approximations. SIAM Journal on Matrix Analysis and Applications, 29(2):656–659,
2007.
[Gao08]
Junbin Gao. Robust `1 principal component analysis and its bayesian variational
inference. Neural computation, 20(2):555–572, 2008.
[GP13]
M. Ghashami and J.M. Phillips. Relative errors for deterministic low-rank matrix
approximations. In SODA, 2013.
[GV15]
Nicolas Gillis and Stephen A Vavasis. On the complexity of robust pca and ell_1norm low-rank matrix approximation. arXiv preprint arXiv:1509.09236, 2015.
14
[GW95]
Michel X Goemans and David P Williamson. Improved approximation algorithms for
maximum cut and satisfiability problems using semidefinite programming. Journal of
the ACM (JACM), 42(6):1115–1145, 1995.
[Hås01]
Johan Håstad. Some optimal inapproximability results. Journal of the ACM (JACM),
48(4):798–859, 2001.
[Ind06]
Piotr Indyk. Stable distributions, pseudorandom generators, embeddings, and data
stream computation. Journal of the ACM (JACM), 53(3):307–323, 2006.
[IP11]
Piotr Indyk and Eric Price. K-median clustering, model-based compressive sensing,
and sparse recovery for earth mover distance. In Proceedings of the forty-third annual
ACM symposium on Theory of computing, pages 627–636. ACM, 2011.
[IPZ98]
Russell Impagliazzo, Ramamohan Paturi, and Francis Zane. Which problems have
strongly exponential complexity? In Foundations of Computer Science, 1998. Proceedings. 39th Annual Symposium on, pages 653–662. IEEE, 1998.
[IT03]
Piotr Indyk and Nitin Thaper. Fast image retrieval via embeddings. In Workshop on
Statistical and Computational Theories of Vision (at ICCV), 2003.
[KAC+ 08]
Fredrik Kahl, Sameer Agarwal, Manmohan Krishna Chandraker, David Kriegman,
and Serge Belongie. Practical global optimization for multiview geometry. International Journal of Computer Vision, 79(3):271–284, 2008.
[KK03]
Qifa Ke and Takeo Kanade. Robust subspace computation using `1 norm. Technical
Report CMU-CS-03-172, Carnegie Mellon University, Pittsburgh, PA., 2003.
[KK05]
Qifa Ke and Takeo Kanade. Robust `1 norm factorization in the presence of outliers
and missing data by alternative convex programming. In 2005 IEEE Computer Society
Conference on Computer Vision and Pattern Recognition (CVPR’05), volume 1, pages
739–746. IEEE, 2005.
[KKMO07]
Subhash Khot, Guy Kindler, Elchanan Mossel, and Ryan O’Donnell. Optimal inapproximability results for max-cut and other 2-variable csps? SIAM Journal on
Computing, 37(1):319–357, 2007.
[KL11]
J. Kelner and A. Levin. Spectral sparsification in the semi-streaming setting. In
Symposium on Theoretical Aspects of Computer Science (STACS), 2011.
[KLC+ 15]
Eunwoo Kim, Minsik Lee, Chong-Ho Choi, Nojun Kwak, and Songhwai Oh. Efficientnorm-based low-rank matrix approximations for large-scale problems using alternating rectified gradient method. IEEE transactions on neural networks and learning
systems, 26(2):237–251, 2015.
[KLM+ 14]
Michael Kapralov, Yin Tat Lee, Cameron Musco, Christopher Musco, and Aaron
Sidford. Single pass spectral sparsification in dynamic streams. In Foundations of
Computer Science (FOCS), 2014 IEEE 55th Annual Symposium on, pages 561–570.
IEEE, 2014.
[KNW10]
Daniel M Kane, Jelani Nelson, and David P Woodruff. On the exact space complexity of sketching and streaming small norms. In Proceedings of the twenty-first
15
annual ACM-SIAM symposium on Discrete Algorithms, pages 1161–1178. Society for
Industrial and Applied Mathematics, 2010.
[KV09]
Ravi Kannan and Santosh Vempala. Spectral algorithms. Foundations and Trends in
Theoretical Computer Science, 4(3-4):157–288, 2009.
[KVW14]
Ravindran Kannan, Santosh S Vempala, and David P Woodruff. Principal component
analysis and higher correlations for distributed data. In Proceedings of The 27th
Conference on Learning Theory, pages 1040–1057, 2014.
[Kwa08]
Nojun Kwak. Principal component analysis based on `1 -norm maximization. IEEE
transactions on pattern analysis and machine intelligence, 30(9):1672–1680, 2008.
[Lib13]
Edo Liberty. Simple and deterministic matrix sketching. In KDD, pages 581–588.
ACM, 2013.
[LMP13]
Mu Li, Gary L. Miller, and Richard Peng. Iterative row sampling. In 54th Annual
IEEE Symposium on Foundations of Computer Science, FOCS 2013, 26-29 October,
2013, Berkeley, CA, USA, pages 127–136, 2013.
[LOG16]
Wuchen Li, Stanley Osher, and Wilfrid Gangbo. Fast algorithm for earth mover’s
distance based on optimal transport and `1 type regularization i. arXiv preprint
arXiv:1609.07092, 2016.
[LPW09]
David Asher Levin, Yuval Peres, and Elizabeth Lee Wilmer. Markov chains and
mixing times. American Mathematical Soc., 2009.
[MBZ10]
Sergio V Macua, Pavle Belanovic, and Santiago Zazo. Consensus-based distributed
principal component analysis in wireless sensor networks. In Signal Processing Advances in Wireless Communications (SPAWC), 2010 IEEE Eleventh International
Workshop on, pages 1–5. IEEE, 2010.
[McG06]
Andrew McGregor. Open problems in data streams and related topics. In IITK
Workshop on Algorithms For Data Streams, 2006.
[MKCP16]
P. P. Markopoulos, S. Kundu, S. Chamadia, and D. A. Pados. Efficient `1 -Norm
Principal-Component Analysis via Bit Flipping. ArXiv e-prints, 2016.
[MKP13]
Panos P. Markopoulos, George N. Karystinos, and Dimitrios A. Pados. Some options
for `1 -subspace signal processing. In ISWCS 2013, The Tenth International Symposium on Wireless Communication Systems, Ilmenau, TU Ilmenau, Germany, August
27-30, 2013, pages 1–5, 2013.
[MKP14]
Panos P. Markopoulos, George N. Karystinos, and Dimitrios A. Pados. Optimal algorithms for `1 -subspace signal processing. IEEE Trans. Signal Processing, 62(19):5046–
5058, 2014.
[MM13]
Xiangrui Meng and Michael W Mahoney. Low-distortion subspace embeddings in
input-sparsity time and applications to robust linear regression. In Proceedings of
the forty-fifth annual ACM symposium on Theory of computing, pages 91–100. ACM,
2013.
16
[Moi13]
Ankur Moitra. An almost optimal algorithm for computing nonnegative rank. In
Proceedings of the Twenty-Fourth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2013, New Orleans, Louisiana, USA, January 6-8, 2013, pages 1454–
1464, 2013.
[Mut05]
S. Muthukrishnan. Data Streams: Algorithms and Applications. Foundations and
Trends in Theoretical Computer Science, 1(2):117–236, 2005.
[MXZZ13]
Deyu Meng, Zongben Xu, Lei Zhang, and Ji Zhao. A cyclic weighted median method
for `1 low-rank matrix factorization with missing entries. In AAAI, volume 4, page 6,
2013.
[NN13]
Jelani Nelson and Huy L Nguyên. Osnap: Faster numerical linear algebra algorithms
via sparser subspace embeddings. In Foundations of Computer Science (FOCS), 2013
IEEE 54th Annual Symposium on, pages 117–126. IEEE, 2013.
[NN14]
Jelani Nelson and Huy L Nguyên. Lower bounds for oblivious subspace embeddings.
In International Colloquium on Automata, Languages, and Programming, pages 883–
894. Springer, 2014.
[NNS+ 14]
Praneeth Netrapalli, UN Niranjan, Sujay Sanghavi, Animashree Anandkumar, and
Prateek Jain. Non-convex robust pca. In Advances in Neural Information Processing
Systems, pages 1107–1115, 2014.
[NYH14]
Feiping Nie, Jianjun Yuan, and Heng Huang. Optimal mean robust principal component analysis. In Proceedings of the 31st International Conference on Machine
Learning (ICML-14), pages 1062–1070, 2014.
[Pin60]
Mark S Pinsker. Information and information stability of random variables and processes. San Francisco: Holden-Day, 1964, originally published in Russian in 1960,
1960.
[PK16]
Young Woong Park and Diego Klabjan. Iteratively reweighted least squares algorithms for `1 -norm principal component analysis. arXiv preprint arXiv:1609.02997,
2016.
[PMvdG+ 13] Jack Poulson, Bryan Marker, Robert A van de Geijn, Jeff R Hammond, and Nichols A
Romero. Elemental: A new framework for distributed memory dense matrix computations. ACM Transactions on Mathematical Software (TOMS), 39(2):13, 2013.
[PP+ 08]
Kaare Brandt Petersen, Michael Syskind Pedersen, et al. The matrix cookbook.
Technical University of Denmark, 7:15, 2008.
[PSW16]
Eric Price, Zhao Song, and David P. Woodruff. Fast regression with an `∞ guarantee.
2016. manuscript.
[QOSG02]
Yongming Qu, George Ostrouchov, Nagiza Samatova, and Al Geist. Principal component analysis for dimension reduction in massive distributed data sets. In Proceedings
of IEEE International Conference on Data Mining (ICDM), 2002.
[Ren92a]
James Renegar. On the computational complexity and geometry of the first-order
theory of the reals, part I: introduction. preliminaries. the geometry of semi-algebraic
17
sets. the decision problem for the existential theory of the reals. J. Symb. Comput.,
13(3):255–300, 1992.
[Ren92b]
James Renegar. On the computational complexity and geometry of the first-order
theory of the reals, part II: the general decision problem. preliminaries for quantifier
elimination. J. Symb. Comput., 13(3):301–328, 1992.
[RSW16]
Ilya Razenshteyn, Zhao Song, and David P Woodruff. Weighted low rank approximations with provable guarantees. In Proceedings of the 48th Annual Symposium on
the Theory of Computing, 2016.
[RV10]
Mark Rudelson and Roman Vershynin. Non-asymptotic theory of random matrices:
extreme singular values. arXiv preprint arXiv:1003.2990, 2010.
[Sar06]
Tamás Sarlós. Improved approximation algorithms for large matrices via random
projections. In 47th Annual IEEE Symposium on Foundations of Computer Science
(FOCS 2006), 21-24 October 2006, Berkeley, California, USA, Proceedings, pages
143–152, 2006.
[SL09]
Roman Sandler and Michael Lindenbaum. Nonnegative matrix factorization with
earth mover’s distance metric. In Computer Vision and Pattern Recognition, 2009.
CVPR 2009. IEEE Conference on, pages 1873–1880. IEEE, 2009.
[Str69]
Volker Strassen. Gaussian elimination is not optimal.
13(4):354–356, 1969.
[SV12]
Nariankadu D. Shyamalkumar and Kasturi R. Varadarajan. Efficient subspace approximation algorithms. Discrete & Computational Geometry, 47(1):44–63, 2012.
[SW11]
Christian Sohler and David P Woodruff. Subspace embeddings for the `1 -norm with
applications. In Proceedings of the forty-third annual ACM symposium on Theory of
computing, pages 755–764. ACM, 2011.
[TD99]
Françoise Tisseur and Jack Dongarra. A parallel divide and conquer algorithm for the
symmetric eigenvalue problem on distributed memory architectures. SIAM Journal
on Scientific Computing, 20(6):2223–2236, 1999.
[TSSW00]
Luca Trevisan, Gregory B Sorkin, Madhu Sudan, and David P Williamson. Gadgets,
approximation, and linear programming. SIAM Journal on Computing, 29(6):2074–
2097, 2000.
[Tsy09]
Alexandre B Tsybakov. Introduction to nonparametric estimation. revised and extended from the 2004 french original. translated by vladimir zaiats, 2009.
[Ver14]
Sergio Verdú. Total variation distance and the distribution of relative information.
In ITA, pages 1–3. Citeseer, 2014.
[VT01]
H. L. Van Trees. Detection, estimation, and modulation theory. Wiley, 1 edition,
September 2001.
[WGR+ 09]
John Wright, Arvind Ganesh, Shankar Rao, Yigang Peng, and Yi Ma. Robust principal component analysis: Exact recovery of corrupted low-rank matrices via convex
optimization. In Advances in neural information processing systems, pages 2080–2088,
2009.
18
Numerische Mathematik,
[Wil12]
Virginia Vassilevska Williams. Multiplying matrices faster than coppersmithwinograd. In Proceedings of the forty-fourth annual ACM symposium on Theory of
computing, pages 887–898. ACM, 2012.
[Woo14a]
David P Woodruff. Low rank approximation lower bounds in row-update streams. In
Advances in Neural Information Processing Systems, pages 1781–1789, 2014.
[Woo14b]
David P. Woodruff. Sketching as a tool for numerical linear algebra. Foundations and
Trends in Theoretical Computer Science, 10(1-2):1–157, 2014.
[WZ13]
David P. Woodruff and Qin Zhang. Subspace embeddings and `p -regression using
exponential random variables. In COLT 2013 - The 26th Annual Conference on
Learning Theory, June 12-14, 2013, Princeton University, NJ, USA, pages 546–567,
2013.
[WZ16]
David P Woodruff and Peilin Zhong. Distributed low rank approximation of implicit
functions of a matrix. In ICDE, 2016.
[Yao77]
Andrew Chi-Chin Yao. Probabilistic computations: Toward a unified measure of
complexity. In Foundations of Computer Science, 1977., 18th Annual Symposium on,
pages 222–227. IEEE, 1977.
[YZD12]
Linbin Yu, Miao Zhang, and Chris Ding. An efficient algorithm for `1 -norm principal
component analysis. In 2012 IEEE International Conference on Acoustics, Speech
and Signal Processing (ICASSP), pages 1377–1380. IEEE, 2012.
[ZLS+ 12]
Yinqiang Zheng, Guangcan Liu, Shigeki Sugimoto, Shuicheng Yan, and Masatoshi
Okutomi. Practical low-rank matrix approximation under robust `1 -norm. In 2012
IEEE Conference on Computer Vision and Pattern Recognition, Providence, RI,
USA, June 16-21, 2012, pages 1410–1417, 2012.
[ZZL15]
Huishuai Zhang, Yi Zhou, and Yingbin Liang. Analysis of robust pca via local incoherence. In Advances in Neural Information Processing Systems, pages 1819–1827,
2015.
19
Contents
1 Introduction
1.1 Our Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Technical Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Several Theorem Statements, an Algorithm, and a Roadmap . . . . . . . . . . . . . .
1
2
4
10
A Notation
22
B Preliminaries
B.1 Polynomial system verifier . . . . . . . . . . . . . . . . . . . . . . . .
B.2 Cauchy and p-stable transform . . . . . . . . . . . . . . . . . . . . .
B.3 Lewis weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
B.4 Frobenious norm and `2 relaxation . . . . . . . . . . . . . . . . . . .
B.5 Converting entry-wise `1 and `p objective functions into polynomials
B.6 Converting entry-wise `1 objective function into a linear program . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
22
22
23
23
24
25
25
C `1 -Low Rank Approximation
C.1 Existence results via dense Cauchy transforms, sparse Cauchy transforms, Lewis weights
C.2 Input sparsity time, poly(k, log n, log d)-approximation for an arbitrary matrix A . .
C.3 poly(k, log d)-approximation for an arbitrary matrix A . . . . . . . . . . . . . . . . .
e
C.4 O(k)-approximation
for an arbitrary matrix A . . . . . . . . . . . . . . . . . . . . . .
C.5 Rank-2k and O(1)-approximation algorithm for an arbitrary matrix A . . . . . . . .
C.6 CUR decomposition for an arbitrary matrix A . . . . . . . . . . . . . . . . . . . . . .
C.7 Rank-r matrix B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
C.7.1 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
C.7.2 poly(k, r)-approximation for rank-r matrix B . . . . . . . . . . . . . . . . . .
26
26
28
31
32
35
37
39
39
41
D Contraction and Dilation Bound for `1
D.1 Definitions . . . . . . . . . . . . . . . . . . .
D.2 Properties . . . . . . . . . . . . . . . . . . .
D.3 Cauchy embeddings, no dilation . . . . . . .
D.4 Cauchy embeddings, no contraction . . . . .
D.5 Cauchy embeddings, k-dimensional subspace
D.6 Sparse Cauchy transform . . . . . . . . . . .
D.7 `1 -Lewis weights . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
44
44
45
48
50
51
53
55
E `p -Low Rank Approximation
E.1 Definitions . . . . . . . . . .
E.2 Properties . . . . . . . . . .
E.3 Tools and inequalities . . .
E.4 Dense p-stable transform . .
E.5 Sparse p-stable transform .
E.6 `p -Lewis weights . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
57
57
58
62
63
64
66
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
F EMD-Low Rank Approximation
66
F.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
F.2 Analysis of no contraction and no dilation bound . . . . . . . . . . . . . . . . . . . . 67
20
G Hardness results for Cauchy matrices, row subset selection, OSE
G.1 Hard instance for Cauchy matrices . . . . . . . . . . . . . . . . . . .
G.2 Hard instance for row subset selection . . . . . . . . . . . . . . . . .
G.3 Hard instance for oblivious subspace embedding and more row subset
G.3.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G.3.2 Main results . . . . . . . . . . . . . . . . . . . . . . . . . . . .
H Hardness
H.1 Previous results . . . . . . . . . . . . . . . . . . . . . . . . .
H.2 Extension to multiplicative error `1 -low rank approximation
H.3 Usin the ETH assumption . . . . . . . . . . . . . . . . . . .
H.4 Extension to the rank-k case . . . . . . . . . . . . . . . . . .
I
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. . . . . .
. . . . . .
selection
. . . . . .
. . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
68
68
72
77
77
79
.
.
.
.
91
92
92
94
97
Limited independent Cauchy random variables
101
I.1 Notations and Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
I.2 Analysis of limited independent random Cauchy variables . . . . . . . . . . . . . . . 101
J Streaming Setting
104
J.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
J.2 Turnstile model, poly(k, log(d), log(n)) approximation . . . . . . . . . . . . . . . . . 105
J.3 Row-update model, poly(k) log d approximation . . . . . . . . . . . . . . . . . . . . . 106
K Distributed Setting
K.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
K.2 Arbitrary-partition model, subspace, poly(k, log(d), log(n)) approximation . . .
K.3 Arbitrary-partition model, decomposition, poly(k, log(d), log(n)) approximation
K.4 Row-partition model, subspace, poly(k) log d approximation . . . . . . . . . . .
K.5 Row-partition model, decomposition, poly(k) log d approximation . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
108
110
110
112
112
114
L Experiments and Discussions
L.1 Setup . . . . . . . . . . . . . . . . . .
L.2 Counterexample for [DZHZ06] . . . . .
L.3 Counterexample for [BDB13] . . . . .
L.4 Counterexample for [Kwa08] . . . . . .
L.5 Counterexample for [KK05] . . . . . .
L.6 Counterexample for all . . . . . . . . .
L.7 Discussion for Robust PCA [CLMW11]
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
115
115
115
116
118
119
121
121
.
.
.
.
.
.
.
M Acknowledgments
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
122
21
A
Notation
Let N+ denote the set of positive integers. For any n ∈ N+ , let [n] denote the set {1, 2, · · · , n}. For
any p ∈ [1, 2], the `p -norm of a vector x ∈ Rd is defined as
kxkp =
d
X
|xi |p
1/p
.
i=1
For any p ∈ [1, 2), the `p -norm of a matrix A ∈ Rn×d is defined as
kAkp =
n X
d
X
|Aij |p
1/p
.
i=1 j=1
Let kAkF denote the Frobenius norm of matrix A. Let nnz(A) denote the number of nonzero entries
of A. Let det(A) denote the determinant of a square matrix A. Let A> denote the transpose of
A. Let A† denote the Moore-Penrose pseudoinverse of A. Let A−1 denote the inverse of a full rank
square matrix. We use Aj to denote the j th column of A, and Ai to denote the ith row of A. For an
n × d matrix A, for S a subset of [n] and T a subset of [d], we let AS denote the |S| × d submatrix
of A with rows indexed by S, while AT denotes the n × |T | submatrix of A with columns indexed
by T , and AST denote the |S| × |T | submatrix A with rows in S and columns in T .
e ) to be f · logO(1) (f ). In addition to O(·) notation, for two
For any function f , we define O(f
functions f, g, we use the shorthand f . g (resp. &) to indicate that f ≤ Cg (resp. ≥) for an
absolute constant C. We use f h g to mean cf ≤ g ≤ Cf for constants c, C. We use OPT to
denote minrank −k Ak kAk − Ak1 , unless otherwise specified.
B
B.1
Preliminaries
Polynomial system verifier
Renegar [Ren92a, Ren92b] and Basu et al. [BPR96] independently provided an algorithm for the
decision problem for the existential theory of the reals, which is to decide the truth or falsity of a
sentence (x1 , · · · , xv )F (f1 , · · · , fm ) where F is a quantifier-free Boolean formula with atoms of the
form sign(fi ) = σ with σ ∈ {0, 1, −1}. Note that this problem is equivalent to deciding if a given
semi-algebraic set is empty or not. Here we formally state that theorem. For a full discussion of
algorithms in real algebraic geometry, we refer the reader to [BPR05] and [Bas14].
Theorem B.1 (Decision Problem [Ren92a, Ren92b, BPR96]). Given a real polynomial system
P (x1 , x2 , · · · , xv ) having v variables and m polynomial constraints fi (x1 , x2 , · · · , xv )∆i 0, ∀i ∈ [m],
where ∆i is any of the “standard relations”: {>, ≥, =, 6=, ≤, <}, let d denote the maximum degree
of all the polynomial constraints and let H denote the maximum bitsize of the coefficients of all the
polynomial constraints. Then in
(md)O(v) poly(H),
time one can determine if there exists a solution to the polynomial system P .
Recently, this technique has been used to solve a number of low-rank approximation and matrix
factorization problems [AGKM12, Moi13, CW15a, BDL16, RSW16].
22
B.2
Cauchy and p-stable transform
Definition B.2 (Dense Cauchy transform). Let S = σ · C ∈ Rm×n where σ is a scalar, and each
entry of C ∈ Rm×n is chosen independently from the standard Cauchy distribution. For any matrix
A ∈ Rn×d , SA can be computed in O(m · nnz(A)) time.
Definition B.3 (Sparse Cauchy transform). Let Π = σ·SC ∈ Rm×n , where σ is a scalar, S ∈ Rm×n
has each column chosen independently and uniformly from the m standard basis vectors of Rm , and
C ∈ Rn×n is a diagonal matrix with diagonals chosen independently from the standard Cauchy
distribution. For any matrix A ∈ Rn×d , ΠA can be computed in O(nnz(A)) time.
Definition B.4 (Dense p-stable transform). Let p ∈ (1, 2). Let S = σ · C ∈ Rm×n where σ is a
scalar, and each entry of C ∈ Rm×n is chosen independently from the standard p-stable distribution.
For any matrix A ∈ Rn×d , SA can be computed in O(m nnz(A)) time.
Definition B.5 (Sparse p-stable transform). Let p ∈ (1, 2). Let Π = σ · SC ∈ Rm×n , where σ is a
scalar, S ∈ Rm×n has each column chosen independently and uniformly from the m standard basis
vectors of Rm , and C ∈ Rn×n is a diagonal matrix with diagonals chosen independently from the
standard p-stable distribution. For any matrix A ∈ Rn×d , ΠA can be computed in O(nnz(A)) time.
B.3
Lewis weights
We follow the exposition of Lewis weights from [CP15].
Definition B.6. For a matrix A, let ai denote ith row of A, and ai (= (Ai )> ) is a column vector.
The statistical leverage score of a row ai is
def
>
−1
>
−1/2
τi (A) = a>
ai k22 .
i (A A) ai = k(A A)
For a matrix A and norm p, the `p Lewis weights w are the unique weights such that for each row i
we have
wi = τi (W 1/2−1/p A).
or equivalently
2/p
>
1−2/p
a>
A)−1 ai = wi .
i (A W
Lemma B.7 (Lemma 2.4 of [CP15] and Lemma 7 of [CLM+ 15]). Given a matrix A ∈ Rn×d , n ≥ d,
for any constant C > 0, 4 > p ≥ 1, there is an algorithm which can compute C-approximate `p Lewis
weights for every row i of A in O((nnz(A) + dω log d) log n) time, where ω < 2.373 is the matrix
multiplication exponent[Str69, CW87, Wil12].
n×d (n ≥ d) with ` (4 > p ≥ 1) Lewis
Lemma B.8 (Theorem 7.1 of [CP15]). Given matrix
p
P A∈R
weights w, for any set of sampling probabilities pi , i pi = N ,
pi ≥ f (d, p)wi ,
1/p
if S ∈ RN ×n has each row chosen independently as the ith standard basis vector, times 1/pi , with
probability pi /N , then with probability at least 0.999,
1
∀x ∈ Rd , kAxk1 ≤ kSAxk1 ≤ 2kAxk1
2
Furthermore, if p = 1, N = O(d log d). If 1 < p < 2, N = O(d log d log log d). If 2 ≤ p < 4,
N = O(dp/2 log d).
23
B.4
Frobenious norm and `2 relaxation
Theorem B.9 (Generalized rank-constrained matrix approximations, Theorem 2 in [FT07]). Given
matrices A ∈ Rn×d , B ∈ Rn×p , and C ∈ Rq×d , let the SVD of B be B = UB ΣB VB> and the SVD of
C be C = UC ΣC VC> . Then,
B † (UB UB> AVC CC> )k C † =
kA − BXCkF ,
arg min
rank −k X∈Rp×q
where (UB UB> AVC VC> )k ∈ Rp×q is of rank at most k and denotes the best rank-k approximation to
UB UB> AVC VC> ∈ Rp×d in Frobenious norm.
Claim B.10 (`2 relaxation of `p -regression). Let p ∈ [1, 2). For any A ∈ Rn×d and b ∈ Rn , define
x∗ = arg minkAx − bkp and x0 = arg minkAx − bk2 . Then,
x∈Rd
x∈Rd
kAx∗ − bkp ≤ kAx0 − bkp ≤ n1/p−1/2 · kAx∗ − bkp .
Proof. The lower bound trivially holds by definition; we will focus on proving the upper bound.
Because Ax − b is an n-dimensional vector, ∀x,
1
n1/p−1/2
kAx − bkp ≤ kAx − bk2 ≤ kAx − bkp .
(3)
Then,
kAx0 − bkp
√
≤ nkAx0 − bk2
√
≤ nkAx∗ − bk2
√
≤ nkAx∗ − bkp
by LHS of Equation (3)
by x0 = arg minkAx − bk2
x
by RHS of Equation (3)
This completes the proof.
Claim B.11 (Frobenius norm relaxation of `p -low rank approximation). Let p ∈ [1, 2) and for any
matrix A ∈ Rn×d , define A∗ =
arg min kB − Akp and A0 =
arg min kB − AkF . Then
rank −k B∈Rn×d
rank −k B∈Rn×d
kA∗ − Akp ≤ kA0 − Akp ≤ (nd)1/p−1/2 kA∗ − Akp .
(4)
Proof. The lower bound of kA0 − Akp trivially holds by definition. We show an upper bound of
kA0 − Akp in the rest of the proof. For any A0 − A ∈ Rn×d , we have
1
(nd)1/p−1/2
kA0 − Akp ≤ kA0 − AkF ≤ kA0 − Akp .
Then,
kA0 − Akp
≤ (nd)1/p−1/2 kA0 − AkF
1/p−1/2
≤ (nd)
by LHS of Equation (5)
∗
by A0 = arg min kB − Akp
kA − AkF
rank −k B
1/p−1/2
≤ (nd)
∗
kA − Akp .
by RHS of Equation (5)
24
(5)
B.5
Converting entry-wise `1 and `p objective functions into polynomials
Claim B.12 (Converting absolute value constraints into variables). Given m polynomials f1 (x), f2 (x),
· · · , fm (x) where x ∈ Rv , solving the problem
minv
m
X
x∈R
|fi (x)|,
(6)
i=1
is equivalent to solving another minimization problem with O(m) extra constraints and m extra
variables,
m
X
min
v
x∈R ,σ∈Rm
σi fi (x)
i=1
s.t. σi2 = 1, ∀i ∈ [m]
fi (x)σi ≥ 0, ∀i ∈ [m].
Claim B.13. (Handling `p ) Given m polynomials f1 (x), f2 (x), · · · , fm (x) where x ∈ Rv and
p = a/b for positive integers a and b, solving the problem
minv
x∈R
m
X
|fi (x)|p ,
(7)
i=1
is equivalent to solving another minimization problem with O(m) extra constraints and O(m) extra
variables,
min
v
m
X
x∈R ,σ∈Rm
yi
i=1
s.t. σi2 = 1, ∀i ∈ [m]
fi (x)σi ≥ 0, ∀i ∈ [m]
(σi fi (x))a = yib , ∀i ∈ [m]
yi ≥ 0, ∀i ∈ [m].
B.6
Converting entry-wise `1 objective function into a linear program
Claim B.14. Given any matrix A ∈ Rn×d and matrix B ∈ Rk×d , the problem minU ∈Rn×k kU B−Ak1
can be solved by solving the following linear program,
min
U ∈Rn×k ,x∈Rn×d
n X
m
X
xi,j
i=1 j=1
Ui B j − Ai,j ≤ xi,j , ∀i ∈ [n], j ∈ [d]
Ui B j − Ai,j ≥ −xi,j , ∀i ∈ [n], j ∈ [d]
xi,j ≥ 0, ∀i ∈ [n], j ∈ [d],
where the number of constraints is O(nd) and the number of variables is O(nd).
25
C
`1 -Low Rank Approximation
This section presents our main `1 -low rank approximation algorithms. Section C.1 provides our
three existence results. Section C.2 shows an input sparsity algorithm with poly(k) log2 d log napproximation ratio. Section C.3 improves the approximation ratio to poly(k) log d. Section C.4
e
explains how to obtain O(k)
approximation ratio. Section C.5 improves the approximation ratio to
O(1) by outputting a rank-2k solution. Section C.6 presents our algorithm for CUR decomposition.
Section C.7 includes some useful properties. Our `1 -low rank approximation algorithm for a rank-r
(where k ≤ r ≤ (n, d) ) matrix is used as a black box (by setting r = poly(k)) in several other
algorithms.
C.1
Existence results via dense Cauchy transforms, sparse Cauchy transforms,
Lewis weights
The goal of this section is to present the existence results in Corollary C.2. We first provide some
bicriteria algorithms in Theorem C.1 which can be viewed as a “warmup”. Then the proof of our
bicriteria algorithm actually implies the existence results.
Theorem C.1. Given matrix A ∈ Rn×d , for any k ≥ 1, there exist bicriteria algorithms with
running time T (specified below), which output two matrices U ∈ Rn×m , V ∈ Rm×d such that, with
probability 9/10,
kU V − Ak1 ≤ α min kAk − Ak1 .
rank −k Ak
(I). Using a dense Cauchy transform, √
T = poly(n, d, k), m = O(k log k), α = O( k log k log d).
(II). Using a sparse Cauchy transform,
T = poly(n, d, k),m = O(k 5 log5 k), α = O(k 4.5 log4.5 k log d).
(III). Sampling by Lewis weights,
√
e
T = (nd)O(k) , m = O(k log k), α = O( k log k).
The matrices in (I), (II), (III) here, are the same as those in (I), (II), (III), (IV) of Lemma D.11.
Thus, they have the properties shown in Section D.2.
Proof. We define
OPT :=
min
rank −k Ak
kAk − Ak1 .
We define U ∗ ∈ Rn×k , V ∗ ∈ Rk×d to be the optimal solution such that kU ∗ V ∗ − Ak1 = OPT .
Part (I). Apply the dense Cauchy transform S ∈ Rm×n with m = O(k log k) rows, and β =
O(log d).
Part (II). Apply the sparse Cauchy transform S(= Π ∈ Rm×n ) with m = O(k 5 log5 k) rows, and
β = O(σ log d) = O(k 2 log2 k log d).
Part (III). Use S (= D ∈ Rn×k ) to denote an n × n matrix which is a sampling and rescaling
diagonal matrix according to the Lewis weights of matrix U ∗ . It has m = O(k log k) rows, and
β = O(1). Sometimes we abuse notation, and should regard D as a matrix which has size m × n,
where m = O(k log k).
We can just replace M in Lemma D.11 with U ∗ V ∗ − A, replace U in Lemma D.11 with U ∗ , and
replace c1 c2 with O(β). So, we can apply Lemma D.11 for S. Then we can plug it in Lemma D.8,
26
we have: with constant probability, for any c ≥ 1, for any V 0 ∈ Rk×d which satisfies
kSU ∗ V 0 − SAk1 ≤ c · min kSU ∗ V − SAk1 ,
(8)
kU ∗ V 0 − Ak1 ≤ c · O(β)kU ∗ V ∗ − Ak1 .
(9)
V ∈Rk×d
it has
Define Vbi = arg minkSU ∗ Vi − SAi k2 for each i ∈ [d]. By using Claim B.10 with n = m and
Vi ∈Rk
d = k, it shows
kSU ∗ Vb − SAk1 =
d
X
kSU ∗ Vbi − SAi k1 ≤
i=1
d
X
√
mkSU ∗ Vei − SAi k1 =
√
m min kSU ∗ V − SAk1 .
i=1
V ∈Rk×d
√
which means Vb is a m-approximation solution to problem, min kSU ∗ V − SAk1 .
V ∈Rk×d
Now, let us look into Equation (8) and Equation (9), we can obtain that
√
kU ∗ Vb − Ak1 ≤ mO(β) OPT .
Because Vbi is the optimal solution of the `2 regression problem, we have
Vbi = (SU ∗ )† SAi ∈ Rk , ∀i ∈ [d], which means Vb = (SU ∗ )† SA ∈ Rk×d .
Plugging Vb into original problem, we obtain
kU ∗ (SU ∗ )† · SA − Ak1 ≤
√
mO(β) OPT .
It means
min
rank −k
X∈Rn×m
kXSA − Ak1 ≤
√
mO(β) OPT .
(10)
If we ignore the constraint on the rank of X, we can get a bicriteria solution:
For part (I), notice that X is an n × m matrix which can be found by using a linear program,
because matrices SA ∈ Rm×d and A ∈ Rn×d are known.
For part (II), notice that X is an n × m matrix which can be found by using a linear program,
because matrices SA ∈ Rm×d and A ∈ Rn×d are known.
For part (III), notice that X is an n × m matrix which can be found by using a linear program,
when the span of rows of DA ∈ Rm×d is known. We assume that D is known in all the above
discussions. But D is actually unknown. So we need to try all the possible choices of the row span
of DA. Since D samples at most m
log k) rows of A, then the total number of choices of
= O(k
n
selecting m rows from n rows is m
= nO(k log k) . This completes the proof.
Equation (10) in the proof of our bicriteria solution implies the following result,
Corollary C.2. Given A ∈ Rn×d , there exists a rank-k matrix A0 ∈ Rn×d such that A0 ∈
rowspan(S 0 A) ⊆ rowspan(A) and kA0 − Ak1 ≤ α · min kA − Ak k1 , where S 0 ∈ Rm×n is a
rank −k Ak
sketching matrix. If S 0
√
(I). indicates the dense Cauchy transform, then α = O( k log k log d).
4.5
4.5
(II). indicates the sparse Cauchy transform, then α = O(k
√ log k log d).
(III). indicates sampling by Lewis weights, then α = O( k log k).
27
Proof. Define OPT =
min
U ∈Rn×k ,V ∈Rk×d
kU V − Ak1 .
Proof of (I). Choose S to be a dense Cauchy transform matrix with m rows, then
√
min
kU ZSA − Ak1 ≤ O( m log d) OPT,
U ∈Rn×k ,Z∈Rk×m
where m = O(k log k). Choosing A0 = U ZSA completes the proof.
Proof of (II). Choose Π = SD ∈ Rm×n where S ∈ Rm×n has each column chosen independently
and uniformly from the m standard basis vectors of Rm , and where D is a diagonal matrix with
diagonals chosen independently from the standard Cauchy distribution, then
√
min
kU ZΠA − Ak1 ≤ O( mσ log d) OPT,
U ∈Rn×k ,Z∈Rk×m
where m = O(k 5 log5 k) and σ = O(k 2 log2 k). Choosing A0 = U ZΠA completes the proof.
Proof of (III).
Choose D to be the sampling and rescaling matrix corresponding to the Lewis weights of U ∗ ,
and let it have m = O(k log k) nonzero entries on the diagonal, then
√
kU ZDA − Ak1 ≤ O( m) OPT .
min
U ∈Rn×k ,Z∈Rk×m
Choosing A0 = U ZDA completes the proof.
C.2
Input sparsity time, poly(k, log n, log d)-approximation for an arbitrary matrix A
The algorithm described in this section is actually worse than the algorithm described in the next
section. But this algorithm is easy to extend to the distributed and streaming settings (See Section K
and Section J).
Algorithm 2 Input Sparsity Time Algorithm
1:
2:
3:
4:
5:
6:
7:
8:
procedure L1LowRankApproxInputSparsity(A, n, d, k)
e 5 ), t2 ← O(k).
e
Set s ← r ← t1 ← O(k
Choose sparse Cauchy matrices S ∈ Rs×n , R ∈ Rd×r , T1 ∈ Rt1 ×n .
Choose dense Cauchy matrices T2 ∈ Rd×t2 .
Compute S · A, A · R and T1 · A · T2 .
Compute XY = arg minX,Y kT1 ARXY SAT2 − T1 AT2 kF .
return ARX, Y SA.
end procedure
. Theorem C.3
Theorem C.3. Given matrix A ∈ Rn×d , for any k ≥ 1, there exists an algorithm which takes
O(nnz(A)) + (n + d) · poly(k) time and outputs two matrices U ∈ Rn×k , V ∈ Rk×d such that
kU V − Ak1 ≤ O(poly(k) log n log2 d)
holds with probability 9/10.
28
min
rank −k Ak
kAk − Ak1
Proof. Choose a Cauchy matrix S ∈ Rs×n (notice that S can be either a dense Cauchy transform
matrix or a sparse Cauchy transform matrix). Using Corollary C.2, we have
min
U ∈Rn×k ,Z∈Rk×s
kU ZSA − Ak1 ≤ αs OPT,
where αs is the approximation by using matrix
S. If S is a dense Cauchy transform matrix, then
√
due to Part (I) of Corollary C.2, αs = O( k log k log d), s = O(k log k), and computing SA takes
O(s nnz(A)) time. If S is a sparse Cauchy transform matrix, then due to Part II of Corollary C.2,
e 4.5 log d), s = O(k
e 5 ), and computing SA takes nnz(A) time.
αs = O(k
We define U ∗ , Z ∗ = arg minU,Z kU ZSA − Ak1 . For the fixed Z ∗ ∈ Rk×s , choose a Cauchy matrix
R ∈ Rd×r (note that R can be either a dense Cauchy transform matrix or a sparse Cauchy transform
matrix) and sketch on the right of (U ZSA − A). If R is a dense Cauchy transform matrix, then
αr = O(log n), r = O(k log k), computing AR takes O(r · nnz(A)) time. If R is a sparse Cauchy
e 2 ) log n, r = O(k
e 5 ), computing AR takes O(nnz(A)) time.
transform matrix, then αr = O(k
j
j
∗
†
b = A R((Z SA)R) ∈ Rk . Then
Define a row vector U
b j Z ∗ SAR − Aj Rk2 = min kx> Z ∗ SAR − Aj Rk2 .
∀j ∈ [n], kU
x∈Rk
Recall that r is the number of columns of R. Due to Claim B.10,
n
X
n
√ X
kAj R((Z ∗ SA)R)† Z ∗ SAR − Aj Rk1 ≤ O( r)
min kU j Z ∗ SAR − Aj Rk1 ,
j=1
j=1
U j ∈Rk
which is equivalent to
√
kAR((Z ∗ SA)R)† Z ∗ SAR − ARk1 ≤ O( r) min kU Z ∗ SAR − ARk1 ,
U ∈Rn×k
where AR is an n × r matrix and SA is an s × d matrix.
Using Lemma D.8, we obtain,
√
kAR((Z ∗ SA)R)† Z ∗ SA − Ak1 ≤ O( rαr ) min kU Z ∗ SA − Ak1 .
U ∈Rn×k
We define X ∗ ∈ Rr×k , Y ∗ ∈ Rk×s ,
X ∗, Y ∗ =
arg min
kARXY SA − Ak1 .
X∈Rr×k ,Y ∈Rk×s
Then,
kARX ∗ Y ∗ SA − Ak1 ≤ kAR((Z ∗ SA)R)† Z ∗ SA − Ak1
√
≤ O( rαr ) min kU Z ∗ SA − Ak1
U ∈Rn×k
√
= O( rαr )
min
kU ZSA − Ak1
U ∈Rn×k ,Z∈Rk×s
√
≤ O( rαr αs ) OPT .
√
It means that ARX ∗ , Y ∗ SA gives an O(αr αs r)-approximation to the original problem.
Thus it suffices to use Lemma C.4 to solve
min
X∈Rr×k ,Y ∈Rk×s
kARXY SA − Ak1 ,
29
by losing an extra poly(k) log d factor in the approximation ratio.
By using a sparse Cauchy transform (for the place discussing the two options), combining the approximation ratios and running times all together, we can get poly(k) log(n) log2 (d)-approximation
ratio with O(nnz(A)) + (n + d) poly(k) running time. This completes the proof.
Lemma C.4. Given matrices A ∈ Rn×d , SA ∈ Rs×n ,RA ∈ Rr×d where S ∈ Rs×n , R ∈ Rd×r
with min(n, d) ≥ max(r, s). For any 1 ≤ k ≤ min(r, s), there exists an algorithm that takes
O(nnz(A)) + (n + d) poly(s, r, k) time to output two matrices X 0 ∈ Rr×k , Y 0 ∈ Rk×s such that
kARX 0 · Y 0 SA − Ak1 ≤ poly(r, s) log(d)
min
X∈Rr×k ,Y ∈Rk×s
kARXY SA − Ak1
holds with probability at least .999.
Proof. Choose sketching matrices T1 ∈ Rt1 ×n to sketch on the left of (ARXY SA − A) (note that
S can be either a dense Cauchy transform matrix or a sparse Cauchy transform matrix). If T1 is
a dense Cauchy transform matrix, then t1 = O(r log r), αt1 = O(log d), and computing T1 A takes
e 2 ) log d,
e 5 ), αt = O(r
O(t1 · nnz(A)) time. If T1 is a sparse Cauchy transform matrix, then t1 = O(r
1
and computing T1 A takes nnz(A) time.
Choose dense Cauchy matrices T2> ∈ Rt2 ×d to sketch on the right of T1 (ARXY SA − A) with
t2 = O((t1 + s) log(t1 + s)). We get the following minimization problem,
min
X∈Rr×k ,Y ∈Rk×s
kT1 ARXY SAT2 − T1 AT2 k1 .
(11)
Define X 0 , Y 0 to be the optimal solution of
min
X∈Rr×k ,Y ∈Rk×s
kT1 ARXY SAT2 − T1 AT2 kF .
Due to Claim B.11,
kT1 ARX 0 Y 0 SAT2 − T1 AT2 k1 ≤
√
t1 t2
min
X∈Rr×k ,Y ∈Rk×s
kT1 ARXY SAT2 − T1 AT2 k1 .
Due to Lemma D.10
kARX 0 Y 0 SA − Ak1 ≤
√
t1 t2 αt1 log t1
min
X∈Rr×k ,Y ∈Rk×s
kARXY SA − Ak1 .
It remains to solve
min
X∈Rr×k ,Y ∈Rk×s
kT1 ARXY SAT2 − T1 AT2 kF .
By using Theorem B.9 and choosing T1 to be a sparse Cauchy transform matrix, we have that
the optimal rank-k solution X 0 Y 0 is (T1 AR)† (UB UB> (T1 AT2 )VC VC> )k (SAT2 ) which can be computed
in O(nnz(A)) + (n + d) poly(s, r, k) time. Here, UB are the left singular vectors of T1 AR. VC are
the right singular vectors of SAT2 .
An alternative way of solving Equation (11) is using a polynomial system verifier. Note that
a polynomial system verifier does not allow absolute value constraints. Using Claim B.12, we are
able to remove these absolute value constraints by introducing new constraints and variables. Thus,
we can get a better approximation ratio but by spending exponential running time in k. In the
previous step, we should always use a dense Cauchy transform to optimize the approximation ratio.
30
Corollary C.5. Given A ∈ Rn×d , there exists an algorithm which takes nd · poly(k) + (n + d) · 2O(k
time and outputs two matrices U ∈ Rn×k , V ∈ Rk×d such that
e
kU V − Ak1 ≤ O(poly(k) log n log2 d)
min
rank −k Ak
2)
kAk − Ak1
holds with probability 9/10.
The poly(k) factor in the above corollary is much smaller than that in Theorem C.3.
C.3
poly(k, log d)-approximation for an arbitrary matrix A
In this section, we explain how to get an O(log d) · poly(k) approximation.
Algorithm 3 poly(k) log d-approximation Algorithm
1:
2:
3:
4:
5:
6:
7:
procedure L1LowRankApproxPolykLogd(A, n, d, k)
. Theorem C.6
e 5 ).
Set s ← O(k
Choose sparse Cauchy matrices S ∈ Rs×n and compute S · A.
Implicitly obtain B = UB VB by finding VB = SA ∈ Rs×d and UB ∈ Rn×s where ∀i ∈ [n],
row vector (UB )i gives an O(1) approximation to minx∈R1×s kxSA − Ai k1 .
U, V ←L1LowRankApproxB(UB , VB , n, d, k, s).
. Theorem C.19
return U, V .
end procedure
Intuitively, our algorithm has two stages. In the first stage, we just want to find a low rank
matrix B which is a good approximation to A. Then, we can try to find a rank-k approximation to
B. Since now B is a low rank matrix, it is much easier to find a rank-k approximation to B. The
procedure L1LowRankApproxB(UB , VB , n, d, k, s) corresponds to Theorem C.19.
Theorem C.6. Given matrix A ∈ Rn×d , for any k ≥ 1, there exists an algorithm which takes
nnz(A) + (n + d) · poly(k) time to output two matrices U ∈ Rn×k , V ∈ Rk×d such that
kU V − Ak1 ≤ poly(k) log d
min
rank −k Ak
kAk − Ak1
holds with probability 9/10.
Proof. We define
OPT :=
min
rank −k Ak
kAk − Ak1 .
The main idea is to replace the given n × d matrix A with another low rank matrix B which
also has size n × d. Choose S ∈ Rs×n to be a Cauchy matrix, where s ≤ poly(k) (note that, if S
is a dense Cauchy transform matrix, computing SA takes O(s nnz(A)) time, while if S is a sparse
Cauchy transform matrix, computing SA takes O(nnz(A)) time). Then B is obtained by taking
each row of A and replacing it with its closest point (in `1 -distance) in the row span of SA. By
using Part II of Corollary C.2, we have,
√
min
kU ZSA − Ak1 ≤ O( s poly(k) log d) OPT .
U ∈Rn×k ,Z∈Rk×s
We define B to be the product of two matrices UB ∈ Rn×s and VB ∈ Rs×d . We define VB to
be SA and UB to be such that for any i ∈ [n], (UB )i gives an O(1)-approximation to problem
31
min kxSA − Ai k1 . This can be solved using an `1 -regression solver in nnz(A) + (n + d) poly(s)
x∈R1×s
time.
By the definition of B, it is an n × d matrix. Naïvely we can write down B after finding UB and
VB . The time for writing down B is O(nd). To avoid this, we can just keep a factorization UB and
VB . We are still able to run algorithm L1LowRankApproxB. Because s = poly(k), the running
time of algorithm L1LowRankApproxB is still O(nnz(A)) + (n + d) poly(k).
By the definition of B, we have that B has rank at most s. Suppose we then solve `1 -low
rank approximation problem for rank-s matrix B, finding a rank-k g-approximation matrix U V .
Due to Lemma C.15, we have that if B is an f -approximation solution to A, then U V is also an
O(f g)-approximation solution to A,
kU V − Ak1 ≤ O(log d) · poly(k) · g OPT .
Using Theorem C.19 we have that g = poly(s), which completes the proof.
C.4
e
O(k)-approximation
for an arbitrary matrix A
e
Algorithm 4 O(k)-approximation
Algorithm
1:
2:
3:
4:
5:
6:
7:
8:
9:
procedure L1LowRankApproxK(A, n, d, k)
. Theorem C.7
r ← O(k log k), m ← t1 ← O(r log r), t2 ← O(m log m).
Guess a diagonal matrix R ∈ Rd×d with only r 1s.
. R selects r columns of A ∈ Rn×d .
n×n
Compute a sampling and rescaling matrix D ∈ R
, T1 ∈ Rn×n corresponding to the Lewis
weights of AR, and let them have m, t1 nonzero entries on the diagonals, respectively.
Compute a sampling and rescaling matrix T2> ∈ Rd×d according to the Lewis weights of
(DA)> , and let it have t2 nonzero entries on the diagonal.
Solve minX,Y kT1 ARXY DAT2 − T1 AT2 k1 .
Take the best solution X, Y over all guesses of R.
return ARX, Y DA.
end procedure
Theorem C.7. Given matrix A ∈ Rn×d , there exists an algorithm that takes poly(n) · dO(k) · 2O(k
time and outputs two matrices U ∈ Rn×k , V ∈ Rk×d such that
e
e
|U V − Ak1 ≤ O(k)
min
rank −k Ak
e
2)
kAk − Ak1
holds with probability 9/10.
Proof. We define
OPT :=
min
rank −k Ak
kAk − Ak1 .
Let U ∗ ∈ Rn×k , V ∗ ∈ Rk×d satisfy
kU ∗ V ∗ − Ak1 = OPT .
Let S > ∈ Rd×d denote the sampling and rescaling matrix corresponding to the Lewis weights
of (V ∗ )> , where the number of nonzero entries on the diagonal of S is s = r = O(k log k). Let
32
R> ∈ Rd×d denote a diagonal matrix such that ∀i ∈ [d], if Si,i 6= 0, then Ri,i = 1, and if Si,i = 0,
then Ri,i = 0. Since rowspan(R> A> ) = rowspan(S > A> ),
min
Z∈Rm×k ,V ∈Rk×d
kASZV − Ak1 =
min
Z∈Rm×k ,V ∈Rk×d
kARZV − Ak1 .
Combining with Part III of Corollay C.2, there exists a rank-k solution in the column span of AR,
which means,
√
min
kARZV − Ak1 ≤ O( r) OPT .
(12)
Z∈Rm×k ,V ∈Rk×d
e
Because the number of 1s of R is r, and the size of the matrix is d × d, there are dr = dO(k) different
choices for locations of 1s on the diagonal of R. We cannot compute R directly, but we can guess
e
all the choices of locations of 1s. Regarding R as selecting r columns of A, then there are dO(k)
choices. There must exist a “correct” way of selecting a subset of columns over all all choices. After
trying all of them, we will have chosen the right one.
For a fixed guess R, we can compute D ∈ Rn×n , which is a sampling and rescaling matrix
corresponding to the Lewis weights of AR, and let m = O(k log2 k) be the number of nonzero
entries on the diagonal of D.
By Equation (12), there exists a W ∈ Rr×k such that,
√
min kARW V − Ak1 ≤ O( r) OPT .
(13)
V ∈Rk×d
We define Vbi = arg minkDARW Vi − DAi k2 , ∀i ∈ [d], which means Vbi = (DARW )† DAi ∈ Rk . Then
Vb =
Vi ∈Rk×d
†
(DARW ) DA ∈ Rk×d .
We define V ∗ = arg minkARW V − Ak1 . Then, by Claim B.10, it has
V ∈Rk×d
√
kDARW Vb − DAk1 ≤ O( m) min kDARW V − DAk1 .
V ∈Rk×d
By applying Lemma D.11, Lemma D.8 and Equation (13), we can show
√
√
e
kARW Vb − Ak1 ≤ O( m)kARW V ∗ − Ak1 ≤ O( mr) OPT ≤ O(k)
OPT .
Plugging Vb = (DARW )† DA into kARW Vb − Ak1 , we obtain that
e
kARW (DARW )† DA − Ak1 ≤ O(k)
OPT .
and it is clear that,
min
X∈Rr×k ,Y
∈Rk×m
e
kARXY DA − Ak1 ≤ kARW (DARW )† DA − Ak1 ≤ O(k)
OPT .
Recall that we guessed R, so it is known. We can compute T1 ∈ Rn×n , which is a sampling
and rescaling diagonal matrix corresponding to the Lewis weights of AR, and t1 = O(r log r) is the
number of nonzero entries on the diagonal of T1 .
Also, DA is known, and the number of nonzero entries in D is m = O(k log2 k). We can compute
>
T2 ∈ Rd×d , which is a sampling and rescaling matrix corresponding to the Lewis weights of (DA)> ,
and t2 = O(m log m) is the number of nonzero entries on the diagonal of T2 .
33
Define X ∗ , Y ∗ =
arg min
kT1 (AR)XY (DA)T2 − T1 AT2 k1 . Thus, using Lemma D.10, we
X∈Rr×k ,Y ∈Rk×m
have
e
k(AR)X ∗ Y ∗ (DA) − Ak1 ≤ O(k)
OPT .
To find X ∗ , Y ∗ , we need to solve this minimization problem
min
X∈Rr×k ,Y ∈Rk×m
kT1 (AR)XY (DA)T2 − T1 AT2 k1 ,
which can be solved by a polynomial system verifier (see more discussion in Section B.1 and B.5).
In the next paragraphs, we explain how to solve the above problem by using a polynomial system
verifier. Notice that T1 (AR) is known and (DA)T2 is also known. First, we create r × k variables
for the matrix X, i.e., one variable for each entry of X. Second, we create k × m variables for
matrix Y , i.e., one variable for each entry of Y . Putting it all together and creating t1 × t2 variables
σi,j , ∀i ∈ [t1 ], j ∈ [t2 ] for handling the unknown signs, we write down the following optimization
poblem
min
X,Y
s.t
t1 X
t2
X
σi,j (T1 (AR)XY (DA)T2 )i,j
i=1 j=1
2
σi,j
= 1, ∀i ∈ [t1 ], j ∈ [t2 ]
σi,j · (T1 (AR)XY (DA)T2 )i,j ≥ 0, ∀i ∈ [t1 ], j ∈ [t2 ].
e 2 ), the maximum degree is O(1), and the
Notice that the number of constraints is O(t1 t2 ) = O(k
2
e ). Thus the running time is,
number of variables O(t1 t2 + km + rk) = O(k
2
(# constraints ·# degree)O(# variables) = 2O(k ) .
e
To use a polynomial system verifier, we need to discuss the bit complexity. Suppose that all entries
are multiples of δ, and the maximum is ∆, i.e., each entry ∈ {−∆, · · · , −2δ, −δ, 0, δ, 2δ, · · · , ∆}, and
e 2
e 2
∆/δ = 2poly(nd) . Then the running time is O(poly(∆/δ)) · 2O(k ) = poly(nd)2O(k ) .
Also, a polynomial system verifier is able to tell us whether there exists a solution in a semialgebraic set. In order to find the solution, we need to do a binary search over the cost C. In each
step of the binary search we use a polynomial system verifier to determine if there exists a solution
in,
t1 X
t2
X
σi,j (T1 (AR)XY (DA)T2 )i,j
i=1 j=1
2
σi,j
= 1, ∀i ∈ [t1 ], j ∈ [t2 ]
≤C
σi,j · (T1 (AR)XY (DA)T2 )i,j ≥ 0, ∀i ∈ [t1 ], j ∈ [t2 ].
In order to do binary search over the cost, we need to know an upper bound on the cost and
also a lower bound on the minimum nonzero cost. The upper bound on the cost is Cmax = O(nd∆),
and the minimum nonzero cost is Cmin = 2−Ω(poly(nd)) . Thus, the total number of steps for binary
search is O(log(Cmax /Cmin )). Overall, the running time is
2
2
ndO(k) · 2O(k ) · log(∆/δ) · log(Cmax /Cmin ) = poly(n)dO(k) 2O(k ) .
e
e
e
This completes the proof.
34
e
Instead of solving an `1 problem at the last step of L1LowRankApproxK by using a polyonomial system verifier, we can just solve a Frobenious norm minimization problem. This slighly
improves the running time and pays an extra poly(k) factor in the approximation ratio. Thus, we
obtain the following corollary,
Corollary C.8. Given matrix A ∈ Rn×d , there exists an algorithm that takes poly(n) · dO(k) time
which outputs two matrices U ∈ Rn×k , V ∈ Rk×d such that
e
|U V − Ak1 ≤ poly(k)
min
rank −k Ak
kAk − Ak1
holds with probability 9/10.
C.5
Rank-2k and O(1)-approximation algorithm for an arbitrary matrix A
In this section, we show how to output a rank-2k solution that is able to achieve an O(1)-approximation.
Algorithm 5 Bicriteria O(1)-approximation Algorithm
1:
2:
procedure L1LowRankApproxBicriteria(A, n, d, k)
UB , VB ←
min
kU V − AkF .
. Theorem C.9
U ∈Rn×k ,V ∈Rk×d
3:
4:
5:
6:
7:
8:
9:
10:
r ← O(k log k).
Guess a diagonal matrix D ∈ Rn×n with r nonzero entries.
Guess matrix DU ∈ Rr×k .
Find VA by solving minV kDU V − D(A − B)k1 .
Find UA by solving minU kU VA − (A
−B)k1 .
VA
Take the best solution UA UB ,
over all guesses.
VB
V
return UA UB , A .
VB
end procedure
Theorem C.9. Given matrix A ∈ Rn×d , for any k ≥ 1, there exists an algorithm which takes
e 2
(nd)O(k ) time to output two matrices U ∈ Rn×2k , V ∈ R2k×d ,
kU V − Ak1 .
min
rank −k Ak
kAk − Ak1 ,
holds with probability 9/10.
Proof. We define U ∗ ∈ Rn×k , V ∗ ∈ Rk×d to be the optimal solution, i.e.,
U ∗V ∗ =
kU V − Ak1 ,
arg min
U ∈Rn×k ,V ∈Rk×d
and define OPT = kU ∗ V ∗ − Ak1 .
Solving the Frobenious norm problem, we can find a factorization of a rank-k matrix
B = UB V B
√
n×k
k×d
where UB ∈ R
, VB ∈ R , and B satisfies kA − Bk1 ≤ αB OPT for an αB = nd.
Let D be a sampling and rescaling diagonal matrix corresponding to the Lewis weights of U ∗ ,
and let the number of nonzero entries on the diagonal be t = O(k log k).
By Lemma D.11 and Lemma D.8, the solution of min kDU ∗ V − D(A − B)k1 together with
U∗
V ∈Rk×d
gives an O(1)-approximation to A − B. In order to compute D we need to know U ∗ . Although
35
we do not know U ∗ , there still exists a way to figure out the Lewis weights. The idea is the same as
in the previous discussion Lemma C.10 “Guessing Lewis weights”. By Claim C.11 and Claim C.12,
the total number of possible D is nO(t) .
Lemma C.10 tries to find a rank-k solution when all the entries in A are integers at most
poly(nd). Here we focus on a bicriteria algorithm that outputs a rank-2k matrix without such a
strong bit complexity assumption. We can show a better claim C.13, which is that the total numer
e 2
of possible DU ∗ is N O(k ) where N = poly(n) is the number of choices for a single entry in U ∗ .
We explain how to obtain an upper bound on N . Consider the optimum kU ∗ V ∗ − (A − B)k1 .
We can always change the basis so assume V ∗ is an Auerbach basis (i.e., an `1 -well-conditioned
basis discussed in Section 1), so ekxk1 ≥ kxV ∗ k1 ≥ kxk1 /f , where e, f = poly(k). Then no entry of
U ∗ is larger than 2f kA − Bk1 , otherwise we could replace U ∗ with 0 and get a better solution. Also
any entry smaller than kA − Bk1 /(enkαB 100) can be replaced with 0 as this will incur additive
error at most OPT /100. So if we round to integer multiples of kA − Bk1 /(enkαB 100) we only have
O(enkαB f ) possibilities for each entry of U ∗ and still have an O(1)-approximation. We will just
refer to this rounded U ∗ as U ∗ , abusing notation.
Let U denote the set of all the matrices U that we guess. From the above discussion, we conclude
that, there exists a U ∈ U such that kU V ∗ − (A − B)k1 ≤ O(OPT).
For each guess of DU ∗ and D, we find VA , UA in the following way. We find VA by using a linear
program to solve,
min kDU ∗ V − D(A − B)k1 .
V ∈Rk×d
Given VA and A, we write down a linear program to solve this problem,
min kU VA − (A − B)k1 ,
U ∈Rn×k
which takes poly(ndk) time. Then we obtain UA .
Recall that VB , UB are the two factors of B and it is a rank-k, αB -approximation solution to
minkU V − Ak1 . Then we have
U,V
VA
UA UB
−A
VB
= UA VA − (A − UB VB )
1
= UA VA − (A − B) .
1
1
Because there must exist a pair UA, VAsatifying kUA VA − (A − B)k1 ≤ O(OPT), it follows that by
VA
taking the best solution UA UB
over all guesses, we obtain an O(1)-approximation solution.
VB
2
Overall, the running time is (nd)O(k ) .
e
Lemma C.10. Given an n × d matrix A with integers bounded by poly(n), for any k ≥ 1, there
e 3
exists an algorithm which takes (nd)O(k ) time to output two matrices U ∈ Rn×k , V ∈ Rk×d , such
that
kU V − Ak1 .
min
rank −k Ak
kAk − Ak1 ,
Proof. We define U ∗ ∈ Rn×k , V ∗ ∈ Rk×d to be the optimal solution, i.e.,
U ∗, V ∗ =
arg min
U ∈Rn×k ,V ∈Rk×d
36
kU V − Ak1 ,
and define OPT = kU ∗ V ∗ − Ak1 .
Let D denote a sampling and rescaling diagonal matrix corresponding to the Lewis weights of
U ∗ , and let the number of nonzero entries on the diagonal be t = O(k log k).
By Lemma D.11 and Lemma D.8, the solution to min kDU ∗ V − DAk1 together with U ∗ gives
V ∈Rk×d
an O(1)-approximation to A. In order to compute D, we need to know U ∗ . Although we do not
know U ∗ , there still exists a way to figure out the Lewis weights. We call the idea “Guessing Lewis
weights”. We will explain this idea in the next few paragraphs.
First, we can guess the nonzero entries on the diagonal, because the number of choices is small.
Claim C.11. The number of possibile choice of supp(D) is at most nO(t) .
Proof. The matrix has dimension
P n ×n and the number of nonzero entries is t. Thus the number
of possible choices is at most ti=1 nt = nO(t) .
Second, we can guess the value of each probability. For each probability, it is trivially at most
1. If the probability is less than 1/(poly(n)k log k), then we will never sample that row with high
probability. It means that we can truncate the probability if it is below that threshold. We can also
round each probability to 2−i which only loses another constant factor in the approximation ratio.
Thus, we have:
Claim C.12. The total number of possible D is nO(t) = nO(k) .
e
Since ∆/δ ≤ poly(n) and the entries of A are in {−∆, −∆+δ, . . . , −2δ, −δ, 0, δ, 2δ, · · · , ∆−δ, ∆},
we can lower bound the cost of kU ∗ V − Ak1 given that it is non-zero by (nd∆/δ)−O(k) (if it is zero
then A has rank at most k and we output A) using Lemma 4.1 in [CW09] and relating entrywise
`1 -norm to Frobenius norm. We can assume V is an `1 well-conditioned basis, since we can replace
U ∗ with U ∗ R−1 and V with RV for any invertible linear transformation R. By properties of such
−O(k) while preserving
basis, we can discretize the entries of U ∗ to integer multiples of (nd∆/δ)
relative error. Hence we can correctly guess each entry of DU ∗ in nO(k) time.
3
Claim C.13. The total numer of possible DU ∗ is nO(k ) .
e
In the following, let DU denote a guess of DU ∗ . Now the problem remaining is to solve
min kDU V − DAk1 . Since we already know DA can be computed, and we know DU , we can
V ∈Rk×d
solve this multiple regression problem by running linear programming. Thus the running time
of this step is in poly(nd). After we get such a solution V , we use a linear program to solve
min kDU V − DAk1 . Then we can get U .
U ∈Rn×k
After we guess all the choices of D and DU ∗ , we must find a solution U, V which gives an O(1)
e
e 3
e 3
approximation. The total running time is nO(k) · nO(k ) · poly(nd) = (nd)O(k ) .
C.6
CUR decomposition for an arbitrary matrix A
Theorem C.14. Given matrix A ∈ Rn×d , for any k ≥ 1, there exists an algorithm which takes
O(nnz(A))+(n+d) poly(k) time to output three matrices C ∈ Rn×c with columns from A, U ∈ Rc×r ,
and R ∈ Rr×d with rows from A, such that rank(CU R) = k, c = O(k log k), r = O(k log k), and
kCU R − Ak1 ≤ poly(k) log d
min
rank −k Ak
holds with probability 9/10.
37
kAk − Ak1 ,
Algorithm 6 CUR Decomposition Algorithm
1:
2:
3:
4:
5:
6:
7:
procedure L1LowRankApproxCUR(A, n, d, k)
. Theorem C.14
UB , VB ←L1LowRankApproxPolykLogd(A, n, d, k).
Let D1 ∈ Rn×n be the sampling and rescaling diagonal matrix coresponding to the Lewis
weights of B1 = UB ∈ Rn×k , and let D1 have d1 = O(k log k) nonzero entries.
Let D2> ∈ Rd×d be the sampling and rescaling diagonal matrix coresponding to the Lewis
>
weights of B2> = (D1 B1 )† D1 A ∈ Rd×k , and let D2 have d2 = O(k log k) nonzero entries.
C ← AD2 , U ← (B2 D2 )† (D1 B1 )† , and R ← D1 A.
return C, U, R.
end procedure
Proof. We define
OPT :=
min
rank −k Ak
kAk − Ak1 .
Due to Theorem C.6, we can output two matrices UB ∈ Rn×k , VB ∈ Rk×d such that UB VB gives a
rank-k, and poly(k) log d-approximation solution to A, i.e.,
kUB VB − Ak1 ≤ poly(k) log d OPT .
(14)
By Section B.3, we can compute D1 ∈ Rn×n which is a sampling and rescaling matrix corresponding
to the Lewis weights of B1 = UB in O(n poly(k)) time, and there are d1 = O(k log k) nonzero entries
on the diagonal of D1 .
Define V ∗ ∈ Rk×d to be the optimal solution of min kB1 V − Ak1 , Vb = (D1 B1 )† D1 A ∈ Rk×d ,
V ∈Rk×d
U1 ∈
Rn×k
to be the optimal solution of
min kU Vb − Ak1 , and V 0 to be the optimal solution of
U ∈Rn×k
min kD1 A − D1 B1 V k1 .
V ∈Rk×d
By Claim B.10, we have
kD1 B1 Vb − D1 Ak1 ≤
p
d1 kD1 B1 V 0 − D1 Ak1 .
Due to Lemma D.11 and Lemma D.8, with constant probability, we have
p
kB1 Vb − Ak1 ≤ d1 αD1 kB1 V ∗ − Ak1 ,
where αD1 = O(1).
Now, we can show,
kU1 Vb − Ak1 ≤ kB1 Vb − Ak1
by U1 = arg minkU Vb − Ak1
U ∈Rn×k
p
.
d1 kB1 V ∗ − Ak1
p
≤
d1 kUB VB − Ak1
≤ poly(k) log d OPT .
by Equation (14)
We define B2 = Vb , then we replace Vb by B2 ∈ Rk×d and look at this objective function,
min kU B2 − Ak1 ,
U ∈Rn×k
38
(15)
where U ∗ denotes the optimal solution. We use a sketching matrix to sketch the RHS of matrix
U B2 − A ∈ Rn×d . Let D2> ∈ Rd×d denote a sampling and rescaling diagonal matrix corresponding
to the Lewis weights of B2> ∈ Rd×k , and let the number of nonzero entries on the diagonal of D2
b = AD2 (B2 D2 )† ∈ Rn×k , U 0 ∈ Rn×k to be the optimal solution of
be d2 = O(k log k). We define U
min k(U B2 − A)D2 k1 . Recall that U1 ∈ Rn×k is the optimal of min kU B2 − Ak1 .
U ∈Rn×k
U ∈Rn×k
By Claim B.10, we have
b B2 D2 − AD2 k1 ≤
kU
p
d2 kU 0 B2 D2 − AD2 k1 .
According to Lemma D.8 and Lemma D.11, with constant probability,
p
b B2 − Ak1 ≤ αD d2 kU1 B2 − Ak1 ,
kU
2
where αB2 = O(1).
We have
b B2 − Ak1
kU
p
≤
d2 αD2 kU1 B2 − Ak1
p
d2 αD2 kU1 Vb − Ak1
=
by B2 = Vb
≤ poly(k) log(d) OPT .
by Equation (15)
b B2 = AD2 (B2 D2 )† (D1 B1 )† D1 A. Setting
Notice that U
C = AD2 ∈ Rn×d2 , U = (B2 D2 )† (D1 B1 )† ∈ Rd2 ×d1 , and R = D1 A ∈ Rd1 ×d ,
we get the desired CUR decomposition,
k AD2 · (B2 D2 )† (D1 B1 )† · D1 A −Ak1 ≤ poly(k) log(d) OPT .
| {z } |
{z
} | {z }
C
U
R
with rank(CU R) = k. Overall, the running time is O(nnz(A)) + (n + d) poly(k).
C.7
C.7.1
Rank-r matrix B
Properties
Lemma C.15. Given matrix A ∈ Rn×d , let OPT =
min
rank −k Ak
kA − Ak k1 . For any r ≥ k, if rank-r
matrix B ∈ Rn×d is an f -approximation to A, i.e.,
kB − Ak1 ≤ f · OPT,
and U ∈ Rn×k , V ∈ Rk×d is a g-approximation to B, i.e.,
kU V − Bk1 ≤ g ·
min
rank −k Bk
kBk − Bk1 ,
then,
kU V − Ak1 . gf · OPT .
39
e ∈ Rn×k , Ve ∈ Rk×d to be two matrices, such that
Proof. We define U
e Ve − Bk1 ≤ g
kU
min
rank −k Bk
kBk − Bk1 ,
and also define,
b , Vb =
U
kU V − Bk1 and U ∗ , V ∗ =
arg min
U ∈Rn×k ,V
∈Rk×d
kU V − Ak1 ,
arg min
U ∈Rn×k ,V ∈Rk×d
Then,
e Ve − Ak1 ≤ kU
e Ve − Bk1 + kB − Ak1
kU
b Vb − Bk1 + kB − Ak1
≤ gkU
by triangle inequality
by definition
≤ gkU ∗ V ∗ − Bk1 + kB − Ak1
b Vb − Bk1 ≤ kU ∗ V ∗ − Bk1
by kU
≤ gkU ∗ V ∗ − Ak1 + gkB − Ak1 + kB − Ak1
by triangle inequality
= g OPT +(g + 1)kB − Ak1
by definition of OPT
≤ g OPT +(g + 1)f · OPT
by B is f -approximation to A
. gf OPT .
This completes the proof.
Lemma C.16. Given a matrix B ∈ Rn×d with rank r, for any 1 ≤ k < r, for any fixed U ∗ ∈ Rn×k ,
choose a Cauchy matrix S with m = O(r log r) rows and rescaled by Θ(1/m). With probability .999
for all V ∈ Rk×d , we have
kSU ∗ V − SBk1 ≥ kU ∗ V − Bk1 .
Proof. This follows by definitions in Section D and Lemma D.23.
Lemma C.17. Given a matrix B ∈ Rn×d with rank r, for any 1 ≤ k < r, for any fixed U ∗ ∈
Rn×k , V ∗ ∈ Rk×d , choose a Cauchy matrix S with m rows and rescaled by Θ(1/m). We have
kSU ∗ V ∗ − SBk1 ≤ O(r log r)kU ∗ V ∗ − Bk1 ,
with probabiliaty .999.
Proof. Let U denote a well-conditioned basis of [U ∗ V ∗ , B], then U has de = O(r) columns. We have
kSU k1 =
de X
m
X
|(SUi )j |
i=1 j=1
de X
m
n
X
1 X
=
|
Sj,l Ul,i |
m
i=1 j=1
=
1
m
by Sj,l ∼ C(0, 1)
l=1
de X
m
X
|ci,j |
by ci,j ∼ C(0, kUi k1 )
i=1 j=1
de
m
1 XX
=
kUi k1 · wi+(j−1)d ,
m
i=1 j=1
40
by wi,j ∼ |C(0, 1)|
where the last step follows since each wi can be thought of as a clipped half-Cauchy random variable.
e Define event ξi to be the situation when wi < D (we will choose D later), and
Define d0 = md.
define event ξ = ξ1 ∩ ξ2 ∩ · · · ∩ ξd0 . Using a similar proof as Lemma D.12, which is also similar to
previous work [Ind06, SW11, CDMI+ 13], we obtain that
X
m
m
X
d0
log d0
Pr
+ .
kSUi k1 ≥
kUi k1 t .
t
D
i=1
i=1
Choosing t = Θ(log d0 ) and D = Θ(d0 ), we have
X
m
m
X
1
Pr
kSUi k1 ≥
kUi k1 O(log d0 ) ≤ ,
C
i=1
i=1
for a constant C. Condition on the above event. Let y = U x, for some x ∈ Rd . Then for any y,
kSyk1 = kSU xk1
0
≤
d
X
kSUj xj k1
by triangle inequality
j=1
0
=
d
X
|xj | · kSUj k1
j=1
0
0
. kxk∞ log(d )
d
X
kUj k1
j=1
. r log rkyk1 ,
where the last step follows by
completes the proof.
Pd0
j=1 kUj k1
≤ d0 and kxk∞ ≤ kU xk1 = kyk1 . Chossing C = 1000
Lemma C.18. Given a matrix M ∈ Rn×d with rank O(r), choose a random matrix S ∈ Rm×n with
each entry drawn from a standard Cauchy distribution and scaled by Θ(1/m). We have that
kSM k1 ≤ O(r log r)kM k1 ,
holds with probability .999.
Proof. Let U ∈ RO(r) be the well-condtioned basis of M . Then each column of M can be expressed
by U x for some x. We then follow the same proof as that of Lemma C.17.
C.7.2
poly(k, r)-approximation for rank-r matrix B
Theorem C.19. Given a factorization of a rank-r matrix B = UB VB ∈ Rn×d , where UB ∈
Rn×r , VB ∈ Rr×d , for any 1 ≤ k ≤ r there exists an algorithm which takes (n + d) · poly(k)
time to output two matrices U ∈ Rn×k , V ∈ Rk×d such that
kU V − Bk1 ≤ poly(r)
min
rank −k Bk
holds with probability 9/10.
41
kBk − Bk1 ,
Algorithm 7 poly(r, k)-approximation Algorithm for Rank-r matrix B
1:
2:
3:
4:
5:
6:
7:
procedure L1LowRankApproxB(UB , VB , n, d, k, r)
. Theorem C.19
e
e
e
e
Set s ← O(r),
r0 ← O(r),
t1 ← O(r),
t2 ← O(r).
0
Choose dense Cauchy matrices S ∈ Rs×n , R ∈ Rd×r , T1 ∈ Rt1 ×n , T2 ∈ Rd×t2 .
Compute S · UB · VB , UB · VB · R and T1 · UB · VB · T2 .
Compute XY = arg minX,Y kT1 UB VB RXY SUB VB T2 − T1 UB VB T2 kF .
return UB VB RX, Y SUB VB .
end procedure
Proof. We define
OPT =
min
rank −k Bk
kBk − Bk1 .
Choose S ∈ Rs×n to be a dense Cauchy transform matrix with s = O(r log r). Using Lemma C.18,
Lemma D.23, and combining with Equation (10), we have
√
kU ZSB − Bk1 ≤ sO(r log r) OPT = O(r1.5 log1.5 r) OPT .
min
U ∈Rn×k ,Z∈Rk×s
Let αs = O(r1.5 log1.5 r).
We define U ∗ , Z ∗ =
kU ZSB − Bk1 . For the fixed Z ∗ ∈ Rk×s , choose a dense
arg min
U ∈Rn×k ,Z∈Rk×d
0
Cauchy transform matrix R ∈ Rd×r with r0 = O(r log r) and sketch on the right of (U ZSB − B).
We obtain the minimization problem, min kU Z ∗ SBR − BRk1 .
U ∈Rn×k
b j = B j R((Z ∗ SB)R)† ∈ Rk , ∀j ∈ [n]. Then U
b = BR((Z ∗ SB)R)† ∈ Rn×k . Due to
Define U
Claim B.10,
n
X
n
√ X
kB j R((Z ∗ SB)R)† Z ∗ SBR − B j Rk1 ≤ O( r0 )
min kU j Z ∗ SBR − B j Rk1 ,
j=1
j=1
U j ∈Rk
which is equivalent to
√
kBR((Z ∗ SB)R)† Z ∗ SBR − BRk1 ≤ O( r0 ) min kU Z ∗ SBR − BRk1 ,
U ∈Rn×k
where BR is an n × r0 matrix and SB is an s × d matrix. Both of them can be computed in
(n + d) poly(r) time.
Using Lemma D.8, Lemma D.23, Lemma C.18, we obtain,
√
kBR((Z ∗ SB)R)† Z ∗ SB − Bk1 ≤ O( r0 αr0 ) min kU Z ∗ SB − Bk1
U ∈Rn×k
e
where αr0 = O(r).
We define X ∗ ∈ Rr×k , Y ∗ ∈ Rk×s ,
X ∗, Y ∗ =
arg min
X∈Rr0 ×k ,Y ∈Rk×s
42
kBRXY SB − Bk1 .
Then,
kBRX ∗ Y ∗ SB − Bk1 ≤ kBR((Z ∗ SB)R)† Z ∗ SB − Bk1
√
≤ O( r0 αr0 ) min kU Z ∗ SB − Bk1
U ∈Rn×k
√
0
= O( r αr0 )
kU ZSB − Bk1
min
U ∈Rn×k ,Z∈Rk×s
√
≤ O( r0 αr0 αs ) OPT .
√
It means that BRX, Y SB gives an O(αr0 αs r0 )-approximation to the original problem.
Thus it suffices to use Lemma C.20 to solve
min
X∈Rr×k ,Y ∈Rk×s
kBRXY SB − Bk1 ,
by losing an extra poly(r) approximation ratio. Therefore, we finish the proof.
0
Lemma C.20. Suppose we are given S ∈ Rs×n , R ∈ Rd×r , and a factorization of a rank-r matrix
B = UB VB ∈ Rn×d , where UB ∈ Rn×r , VB ∈ Rr×d . Then for any 1 ≤ k ≤ r, there exists an
0
algorithm which takes (n + d) poly(r, r0 , s) time to output two matrices X 0 ∈ Rr ×k , Y 0 ∈ Rk×s such
that
kBRX 0 · Y 0 SB − Bk1 ≤ poly(r)
min
X∈Rr0 ×k ,Y ∈Rk×s
kBRXY SB − Bk1
holds with probability at least .999.
Proof. Choosing dense Cauchy matrices T1 ∈ Rt1 ×n , T2> ∈ Rt2 ×d to sketch on both sides, we get the
problem
min
kT1 BRXY SBT2 − T1 BT2 k1 ,
(16)
0
X∈Rr
×k ,Y
∈Rk×s
e
e
where t1 = O(r)
and t2 = O(r).
0
0
Define X , Y to be the optimal solution of
min
X∈Rr0 ×k ,Y ∈Rk×s
kT1 BRXY SBT2 − T1 BT2 kF .
e Ye to the be the optimal solution of
Define X,
min
X∈Rr0 ×k ,Y ∈Rk×s
kBRXY SB − Bk1 .
By Claim B.11,
kT1 BRX 0 Y 0 SBT2 − T1 BT2 k1 ≤
√
t1 t2
min
X∈Rr
0 ×k
,Y ∈Rk×s
kT1 BRXY SBT2 − T1 BT2 k1 .
By Lemma D.10, Lemma C.18 and Lemma D.23, we have
√
e 2 )kBRX
e · Ye SB − Bk1 .
kBRX 0 · Y 0 SB − Bk1 ≤ t1 t2 · O(r
This completes the proof.
43
D
Contraction and Dilation Bound for `1
This section presents the essential lemmas for `1 -low rank approximation. Section D.1 gives some
basic definitions. Section D.2 shows some properties implied by contraction and dilation bounds.
Section D.3 presents the no dilation lemma for a dense Cauchy transform. Section D.4 and D.5
presents the no contraction lemma for dense Cauchy transforms. Section D.6 and D.7 contains the
results for sparse Cauchy transforms and Lewis weights.
D.1
Definitions
Definition D.1. Given a matrix M ∈ Rn×d , if matrix S ∈ Rm×n satisfies
kSM k1 ≤ c1 kM k1 ,
then S has at most c1 -dilation on M .
Definition D.2. Given a matrix U ∈ Rn×k , if matrix S ∈ Rm×n satisfies
1
∀x ∈ Rk , kSU xk1 ≥ kU xk1 ,
c2
then S has at most c2 -contraction on U .
Definition D.3. Given matrices U ∈ Rn×k , A ∈ Rn×d , let V ∗ = arg minV ∈Rk×d kU V − Ak1 . If
matrix S ∈ Rm×n satisfies
1
∀V ∈ Rk×d , kSU V − SAk1 ≥ kU V − Ak1 − c4 kU V ∗ − Ak1 ,
c3
then S has at most (c3 , c4 )-contraction on (U, A).
Definition D.4. A (c5 , c6 ) `1 -subspace embedding for the column space of an n × k matrix U is a
matrix S ∈ Rm×n for which all x ∈ Rk
1
kU xk1 ≤ kSU xk1 ≤ c6 kU xk1 .
c5
Definition D.5. Given matrices U ∈ Rn×k , A ∈ Rn×d , let V ∗ = arg minV ∈Rk×d kU V − Ak1 . Let
S ∈ Rm×n . If for all c ≥ 1, and if for any Vb ∈ Rk×d which satisfies
kSU Vb − SAk1 ≤ c · min kSU V − SAk1 ,
V ∈Rk×d
it holds that
kU Vb − Ak1 ≤ c · c7 · kU V ∗ − Ak1 ,
then S provides a c7 -multiple-regression-cost preserving sketch of (U, A).
Definition D.6. Given matrices L ∈ Rn×m1 , N ∈ Rm2 ×d , A ∈ Rn×d , k ≥ 1, let
X ∗ = arg
min
rank −k X
kLXN − Ak1 .
b ∈ Rm1 ×m2 which satisfies
Let S ∈ Rm×n . If for all c ≥ 1, and if for any rank −k X
b − SAk1 ≤ c ·
kSLXN
min
rank −k X
kSLXN − SAk1 ,
it holds that
b − Ak1 ≤ c · c8 · kLX ∗ N − Ak1 ,
kLXN
then S provides a c8 -restricted-multiple-regression-cost preserving sketch of (L, N, A, k).
44
D.2
Properties
Lemma D.7. Given matrices A ∈ Rn×d , U ∈ Rn×k , let V ∗ = arg minV ∈Rk×d kU V − Ak1 . If
S ∈ Rm×n has at most c1 -dilation on U V ∗ − A, i.e.,
kS(U V ∗ − A)k1 ≤ c1 kU V ∗ − Ak1 ,
and it has at most c2 -contraction on U , i.e.,
∀x ∈ Rk , kSU xk1 ≥
then S has at most (c2 , c1 +
1
c2 )-contraction
∀V ∈ Rk×d , kSU V − SAk1 ≥
1
kU xk1 ,
c2
on (U, A), i.e.,
1
1
kU V − Ak1 − (c1 + )kU V ∗ − Ak1 ,
c2
c2
Proof. Let A ∈ Rn×d , U ∈ Rn×k , S ∈ Rm×n be the same as that described in the lemma. Then
∀V ∈ Rk×d
kSU V − SAk1 ≥ kSU V − SU V ∗ k1 − kSU V ∗ − SAk1
≥ kSU V − SU V ∗ k1 − c1 kU V ∗ − Ak1
= kSU (V − V ∗ )k1 − c1 kU V ∗ − Ak1
=
d
X
kSU (V − V ∗ )j k1 − c1 kU V ∗ − Ak1
j=1
≥
d
X
1
kU (V − V ∗ )j k1 − c1 kU V ∗ − Ak1
c2
j=1
1
kU V − U V ∗ k1 − c1 kU V ∗ − Ak1
c2
1
1
≥ kU V − Ak1 − kU V ∗ − Ak1 − c1 kU V ∗ − Ak1
c2
c2
1
1
∗
= kU V − Ak1 − ( + c1 )kU V − Ak1 .
c2
c2
=
The first inequality follows by the triangle inequality. The second inequality follows since S has at
most c1 dilation on U V ∗ − A. The third inequality follows since S has at most c2 contraction on
U . The fourth inequality follows by the triangle inequality.
Lemma D.8. Given matrices A ∈ Rn×d , U ∈ Rn×k , let V ∗ = arg minV ∈Rk×d kU V − Ak1 . If
S ∈ Rm×n has at most c1 -dilation on U V ∗ − A, i.e.,
kS(U V ∗ − A)k1 ≤ c1 kU V ∗ − Ak1 ,
and has at most c2 -contraction on U , i.e.,
∀x ∈ Rk , kSU xk1 ≥
45
1
kU xk1 ,
c2
then S provides a (2c1 c2 + 1)-multiple-regression-cost preserving sketch of (U, A), i.e., for all c ≥ 1,
for any Vb ∈ Rk×d which satisfies
kSU Vb − SAk1 ≤ c · min kSU V − SAk1 ,
V ∈Rk×d
it has
kU Vb − Ak1 ≤ c · (2c1 c2 + 1) · kU V ∗ − Ak1 ,
Proof. Let S ∈ Rm×n , A ∈ Rn×d , U ∈ Rn×k , V ∗ , Vb ∈ Rk×d , and c be the same as stated in the
lemma.
kU Vb − Ak1 ≤ c2 kSU Vb − SAk1 + (1 + c1 c2 )kU V ∗ − Ak1
≤ c2 c min kSU V − SAk1 + (1 + c1 c2 )kU V ∗ − Ak1
V ∈Rk×d
∗
≤ c2 ckSU V − SAk1 + (1 + c1 c2 )kU V ∗ − Ak1
≤ c1 c2 ckU V ∗ − Ak1 + (1 + c1 c2 )kU V ∗ − Ak1
≤ c · (1 + 2c1 c2 )kU V ∗ − Ak1 .
The first inequality follows by Lemma D.7. The second inequality follows by the guarantee of Vb .
The fourth inequality follows since S has at most c1 -dilation on U V ∗ − A. The fifth inequality
follows since c ≥ 1.
Lemma D.9. Given matrices L ∈ Rn×m1 , N ∈ Rm2 ×d , A ∈ Rn×d , k ≥ 1, let
X ∗ = arg
min
rank −k X
kLXN − Ak1 .
If S ∈ Rm×n has at most c1 -dilation on LX ∗ N − A, i.e.,
kS(LX ∗ N − A)k1 ≤ c1 kLX ∗ N − Ak1 ,
and has at most c2 -contraction on L, i.e.,
∀x ∈ Rm1 kSLxk1 ≥ kLxk1 ,
then S provides a (2c1 c2 + 1)-restricted-multiple-regression-cost preserving sketch of (L, N, A, k),
b ∈ Rm1 ×m2 which satisfies
i.e., for all c ≥ 1, for any rank −k X
b − SAk1 ≤ c ·
kSLXN
min
rank −k X
kSLXN − SAk1 ,
it holds that
b − Ak1 ≤ c · (2c1 c2 + 1) · kLX ∗ N − Ak1 .
kLXN
b ∈ Rm1 ×m2 , X ∗ ∈ Rm1 ×m2 , N ∈ Rm2 ×d , A ∈ Rn×d , and c ≥ 1
Proof. Let S ∈ Rm×n , L ∈ Rn×m1 , X
be the same as stated in the lemma.
b − SAk1 ≥ kSLXN
b − SLX ∗ N k1 − kSLX ∗ N − SAk1
kSLXN
1
b − X ∗ N )k1 − c1 kLX ∗ N − Ak1
≥ kL(XN
c2
1
b − Ak1 − 1 kLX ∗ N − Ak1 − c1 kLX ∗ N − Ak1
≥ kLXN
c2
c2
1
b − Ak1 − ( 1 + c1 )kLX ∗ N − Ak1 .
= kLXN
c2
c2
46
The inequality follows by the triangle inequality. The second inequality follows since S has at most
c2 -contraction on L, and it has at most c1 -dilation on LX ∗ N − A. The third inequality follows by
the triangle inequality.
It follows that
b − Ak1 ≤ c2 kSLXN
b − SAk1 + (1 + c1 c2 )kLX ∗ N − Ak1
kLXN
≤ c2 c ·
kSLXN − SAk1 + (1 + c1 c2 )kLX ∗ N − Ak1
min
rank −k X
∗
≤ c2 c · kSLX N − SAk1 + (1 + c1 c2 )kLX ∗ N − Ak1
≤ cc1 c2 · kLX ∗ N − Ak1 + (1 + c1 c2 )kLX ∗ N − Ak1
≤ c · (1 + 2c1 c2 )kLX ∗ N − Ak1 .
The first inequality directly follows from the previous one. The second inequality follows from the
b The fourth inequality follows since S has at most c1 dilation on LX ∗ N − A. The
guarantee of X.
fifth inequality follows since c ≥ 1.
Lemma D.10. Given matrices L ∈ Rn×m1 , N ∈ Rm2 ×d , A ∈ Rn×d , k ≥ 1, let
X ∗ = arg
min
rank −k X
kLXN − Ak1 .
Let T1 ∈ Rt1 ×n have at most c1 -dilation on LX ∗ N − A, and at most c2 -contraction on L. Let
e = arg
X
min
rank −k X
kT1 LXN − T1 Ak1 .
e − T1 A)> , and at most c0 -contraction on N > .
Let T2> ∈ Rt2 ×d have at most c01 -dilation on (T1 LXN
2
m
×m
b ∈ R 1 2 which satisfies
Then, for all c ≥ 1, for any rank −k X
b − SA)T2 k1 ≤ c ·
kT1 (LXN
min
rank −k X
kT1 (LXN − A)T2 k1 ,
it has
b − Ak1 ≤ c · (2c1 c2 + 1)(2c01 c02 + 1) · kLX ∗ N − Ak1 .
kLXN
b ∈ Rm1 ×m2
Proof. Apply Lemma D.9 for sketch matrix T2 . Then for any c ≥ 1, any rank −k X
which satisfies
b − A)T2 k1 ≤ c ·
kT1 (LXN
min
rank −k X
kT1 (LXN − A)T2 k1 ,
has
b − A)k1 ≤ c · (2c01 c02 + 1) · kT1 (LXN
e − A)k1 .
kT1 (LXN
b ∈ Rm1 ×m2 which
Apply Lemma D.9 for sketch matrix T1 . Then for any c ≥ 1, any rank −k X
satisfies
b − A)k1 ≤ c(2c01 c02 + 1) ·
kT1 (LXN
min
rank −k X
e − A)k1 ,
kT1 (LXN
has
b − Ak1 ≤ c · (2c1 c2 + 1)(2c0 c0 + 1) · kLX ∗ N − Ak1 .
kLXN
1 2
47
Lemma D.11. Given matrices M ∈ Rn×d , U ∈ Rn×t , d ≥ t = rank(U ), n ≥ d ≥ r = rank(M ), if
sketching matrix S ∈ Rm×n is drawn from any of the following probability distributions of matrices,
with .99 probability S has at most c1 -dilation on M , i.e.,
kSM k1 ≤ c1 kM k1 ,
and S has at most c2 -contraction on U , i.e.,
∀x ∈ Rt , kSU xk1 ≥
1
kU xk1 ,
c2
where c1 , c2 are parameters depend on the distribution over S.
(I) S ∈ Rm×n is a dense Cauchy matrix: a matrix with i.i.d. entries from the standard Cauchy
distribution. If m = O(t log t), then c1 c2 = O(log d). If m = O((t + r) log(t + r)), then
c1 c2 = O(min(log d, r log r)).
(II) S ∈ Rm×n is a sparse Cauchy matrix: S = T D, where T ∈ Rm×n has each column i.i.d.
from the uniform distribution on standard basis vectors of Rm , and D ∈ Rn×n is a diagonal matrix with i.i.d. diagonal entries following a standard Cauchy distribution. If m =
O(t5 log5 t), then c1 c2 = O(t2 log2 t log d). If m = O((t + r)5 log5 (t + r)), then c1 c2 =
O(min(t2 log2 t log d, r3 log3 r)).
(III) S ∈ Rm×n is a sampling and rescaling matrix (notation S ∈ Rn×n denotes a diagonal sampling
and rescaling matrix with m non-zero entries): If S samples and reweights m = O(t log t)
rows of U , selecting each with probability proportional to the ith row’s `1 Lewis weight and
reweighting by the inverse probability, then c1 c2 = O(1).
(IV) S ∈ Rm×n is a dense Cauchy matrix with limited independence: S is a matrix with each
entry drawn from a standard Cauchy distribution. Entries from different rows of S are fully
independent, and entries from the same row of S are W -wise independent. If m = O(t log t),
e
e
and W = O(d),
then c1 c2 = O(t log d). If m = O(t log t), and W = O(td),
then c1 c2 =
O(log d).
In the above, if we replace S with σ · S where σ ∈ R\{0} is any scalar, then the relation between m
and c1 c2 can be preserved.
For (I), if m = O(t log t), then c1 c2 = O(log d) is implied by Lemma D.23 and Lemma D.13. If
m = O((t + r) log(t + r)), c1 c2 = O(r log r) is implied by [SW11].
For (II), if m = O(t5 log5 t), then c1 c2 = O(t2 log2 t log d) is implied by Corollary D.27 and
Lemma D.25. If m = O((t + r)5 log5 (t + r)), c1 c2 = O(r3 log3 r) is implied by [MM13].
For (III), it is implied by [CP15] and Lemma D.29.
For (IV), it is implied by Lemma I.4, Corollary I.5.
D.3
Cauchy embeddings, no dilation
Lemma D.12. Define U ∗ ∈ Rn×k , V ∗ ∈ Rk×d to be the optimal solution of
min
U ∈Rn×k ,V ∈Rk×d
Ak1 . Choose a Cauchy matrix S with m rows and rescaled by Θ(1/m). We have that
kSU ∗ V ∗ − SAk1 ≤ O(log d)kU ∗ V ∗ − Ak1
holds with probability at least 99/100.
48
kU V −
Proof. The proof technique has been used in [Ind06] and [CDMI+ 13]. Fix the optimal U ∗ and V ∗ ,
then
kSU ∗ V ∗ − SAk1 =
d
X
kS(U ∗ Vi∗ − Ai )k1
i=1
d X
m X
n
X
1
Sj,l · (U ∗ Vi∗ − Ai )l |
=
|
m
i=1 j=1
where Sj,l ∼ C(0, 1)
l=1
d X
m
X
1
=
|ci,j |
m
where ci,j ∼ C(0, kU ∗ Vi∗ − Ai k1 )
i=1 j=1
d
m
1 XX ∗ ∗
=
kU Vi − Ai k1 · wi+d(j−1) .
m
where wi+d(j−1) ∼ |C(0, 1)| (17)
i=1 j=1
where the last step follows since each wi can be thought of as a clipped half-Cauchy random variable.
Define d0 = md. Define event ξi to be the situation in which wi < D (we will decide upon D later),
and define event ξ = ξ1 ∩ ξ2 ∩ · · · ∩ ξd0 . Then it is clear that ξ ∩ ξi = ξ, ∀i ∈ [d0 ].
Using the probability density function (pdf) of a Cauchy and because tan−1 x ≤ x, we can lower
bound Pr[event ξi holds] in the following sense,
Pr[ξi ] =
2
2
2
tan−1 (D) = 1 − tan−1 (1/D) ≥ 1 −
.
π
π
πD
By a union bound over all i ∈ [d0 ], we can lower bound Pr[event ξ holds],
0
Pr[ξ] ≥ 1 −
d
X
Pr[ξ i ] ≥ 1 −
i=1
2d0
.
πD
(18)
By Bayes rule and ξ = ξ ∩ ξi , Pr[ξ|ξi ] Pr[ξi ] = Pr[ξ ∩ ξi ] = Pr[ξ], which implies that Pr[ξ|ξi ] =
Pr[ξ]/ Pr[ξi ]. First, we can lower bound E[wi |ξi ],
E[wi |ξi ] = E[wi |ξi ∩ ξ] Pr[ξ|ξi ] + E[wi |ξ ∩ ξ] Pr[ξ|ξi ]
≥ E[wi |ξi ∩ ξ] Pr[ξ|ξi ]
by wi ≥ 0 and Pr[] ≥ 0
= E[wi |ξ] Pr[ξ|ξi ]
by ξ = ξ ∩ ξi .
The above equation implies that
E[wi |ξi ]
Pr[ξ|ξi ]
E[wi |ξi ] Pr[ξi ]
=
Pr[ξ ∩ ξi ]
E[wi |ξi ] Pr[ξi ]
=
Pr[ξ]
E[wi |ξ] ≤
Using the pdf of a Cauchy, E[wi |ξi ] =
of E[wi |ξ],
E[wi |ξ] ≤
E[wi |ξi ] Pr[ξi ]
=
Pr[ξ]
by Bayes rule Pr[ξ|ξi ] Pr[ξi ] = Pr[ξ ∩ ξi ]
by ξ = ξ ∩ ξi .
1
π
log(1 + D2 )/ Pr[ξi ] and plugging it into the lower bound
1
π
log(1 + D2 )
≤
Pr[ξ]
49
1
π
log(1 + D2 )
1−
2d
πD
. log(D),
where the third step follows since Pr[ξ] ≥ 1 −
We can conclude
d
2d0
πD
and the last step follows by choosing D = Θ(d0 ).
m
1 XX ∗ ∗
E[kSU V − SAk1 |ξ] =
kU Vi − Ai k1 · E[wi+d(j−1) |ξ] . (log d0 ) · kU ∗ V ∗ − Ak1 . (19)
m
∗
∗
i=1 j=1
For simplicity, define X = kSU ∗ V ∗ − SAk1 and γ = kU ∗ V ∗ − Ak1 . By Markov’s inequality and
because Pr[X ≥ γt|ξ] ≤ 1, we have
Pr[X ≥ γt]
= Pr[X ≥ γt|ξ] Pr[ξ] + Pr[X ≥ γt|ξ] Pr[ξ]
≤ Pr[X ≥ γt|ξ] + Pr[ξ]
E[X|ξ]
≤
+ Pr[ξ]
γt
E[X|ξ]
2d0
≤
+
γt
πD
0
log d
2d0
.
+
t
πD
≤ .01,
by Markov’s inequality
by Equation (18)
by Equation (19)
where choosing t = Θ(log d0 ) and D = Θ(d0 ). Since k ≤ d and m = poly(k), we have t = Θ(log d),
which completes the proof.
Lemma D.13. Given any matrix M ∈ Rn×d , if matrix S ∈ Rm×n has each entry drawn from an
i.i.d. standard Cauchy distribution and is rescalsed by Θ(1/m), then
kSM k1 ≤ O(log d)kM k1
holds with probability at least 99/100.
Proof. Just replace the matrix U ∗ V ∗ − A in the proof of Lemma D.12 with M . Then we can get
the result directly.
D.4
Cauchy embeddings, no contraction
We prove that if we choose a Cauchy matrix S, then for a fixed optimal solution U ∗ of minU,V kU V −
Ak1 , and for all V , we have that with high probability kSU ∗ V −SAk1 is lower bouned by kU ∗ V −Ak1
up to some constant.
Lemma D.14. Define U ∗ ∈ Rn×k , V ∗ ∈ Rk×d to be the optimal solution of
min
U ∈Rn×k V ∈Rk×d
kU V −Ak1 ,
where A ∈ Rn×d . Let m = O(k log k), S ∈ Rm×n be a random matrix with each entry an i.i.d.
standard Cauchy random variable, scaled by Θ(1/m). Then with probability at least 0.95,
∀V ∈ Rk×d , kU ∗ V − Ak1 . kSU ∗ V − SAk1 + O(log(d))kU ∗ V ∗ − Ak1 .
Proof. This follows by Lemmas D.12, D.23, D.7.
50
D.5
Cauchy embeddings, k-dimensional subspace
The goal of this section is to prove Lemma D.23.
Before getting into the details, we first give the formal definition of an (α, β) `1 well-conditioned
basis and -net.
Definition D.15 (`1 Well-conditioned basis). [DDH+ 09] A basis U for the range of A is (α, β)condtioned if kU k1 ≤ α and for all x ∈ Rk , kxk∞ ≤ βkU xk1 . We will say U is well-conditioned if
α and β are low-degree polynomials in k, independent of n.
Note that a well-conditoned basis implies the following result.
Fact D.16. There exist α, β ≥ 1 such that,
∀x ∈ Rk ,
1
kxk1 ≤ kU xk1 ≤ αkxk1 .
kβ
Proof. The lower bound can be proved in the following sense,
kU xk1 ≥
1
11
kxk∞ ≥
kxk1 ,
β
βk
where the first step follows by the properties of a well-conditioned basis, and the second step follows
since kkxk∞ ≥ kxk1 . Then we can show an upper bound,
kU xk1 ≤ kU k1 · kxk1 ≤ αkxk1 ,
where the first step follows by kU xk1 ≤ kU k1 kxk1 , and the second step follows using kU k1 ≤ α.
Definition D.17 (-net). Define N to an -net where, for all the x ∈ Rk that kxk1 = 1, for any
vectors two x, x0 ∈ N , kx − x0 k1 ≥ , for any vector y ∈
/ N , there exists an x ∈ N such that
kx − yk1 ≤ . Then the size of N is ( 1 )O(k) = 2O(k log(1/)) .
Lemma D.18 (Lemma 6 in [SW11]). There is a constant c0 > 0 such that for any t ≥ 1 and any
constant c > c0 , if S is a t × n matrix whose entries are i.i.d. standard Cauchy random variables
scaled by c/t, then, for any fixed y ∈ Rn ,
Pr[kSyk1 < kyk1 ] ≤ 1/2t
Lemma D.19. Suppose we are given a well-conditioned basis U ∈ Rn×k . If S is a t × n matrix
whose entries are i.i.d. standard Cauchy random variables scaled by Θ(1/t), then with probability
1 − 2−Ω(t) , for all vectors x ∈ N we have that kSU xk1 ≥ kU xk1 .
Proof. First, using Lemma D.18, we have for any fixed vector y ∈ Rn , Pr[kSyk < kyk1 ] ≤ 1/2t .
Second, we can rewrite y = U x. Then for any fixed x ∈ N , Pr[kSU xk < kU xk1 ] ≤ 1/2t . Third,
choosing t & k log(1/) and taking a union bound over all the vectors in the -net N completes the
proof.
Lemma D.20 (Lemma 7 in [SW11]). Let S be a t×n matrix whose entries are i.i.d standard Cauchy
random variables, scaled by c/t for a constant c, and t ≥ 1. Then there is a constant c0 = c0 (c) > 0
such that for any fixed set of {y1 , y2 , · · · , yk } of d vectors in {y ∈ Rn : x ∈ Rk , U x = y},
k
k
X
X
Pr[
kSyi k1 ≥ c0 log(tk)
kyi k1 ] ≤
i=1
i=1
51
1
.
1000
Using Lemma D.20 and the definition of an (α, β) well-conditioned basis, we can show the
following corollary.
Corollary D.21. Suppose we are given an (α, β) `1 well-conditioned basis U ∈ Rn×k . Choose S to
be an i.i.d. Cauchy matrix with t = O(k log k) rows, and rescale each entry by Θ(1/t). Then with
probability 99/100, for all vectors x ∈ Rk ,
kSU xk1 ≤ O(αβ log(k)) · kU xk1 ,
Proof. Define event E to be the situation when
k
X
0
kSUi k1 < c log(tk)
i=1
k
X
kUi k1
i=1
holds, and U is a well-conditioned basis. Using Lemma D.20, we can show that event E holds with
probability 999/1000. We condition on Event E holding. Then for any y = U x for an x ∈ Rk , we
have
kSyk1 = kSU xk1
≤
k
X
kSUj xj k1
by triangle inequality
j=1
=
k
X
|xj | · kSUj k1
j=1
0
≤ kxk∞ c log(tk)
k
X
kUj k1
by Lemma D.20
j=1
0
= kxk∞ c log(tk)α
by kU k1 =
k
X
kUj k1 ≤ α
j=1
≤ βkU xk1 c0 log(tk)α
by kxk∞ ≤ βkU xk1
0
≤ βkyk1 c log(tk)α
by U x = y.
This completes the proof.
Lemma D.22. Given an (α, β) `1 well-conditioned basis, condition on the following two events,
1. For all x ∈ N , kSU xk1 ≥ kU xk1 . (Lemma D.18)
2. For all x ∈ Rk , kSU xk1 ≤ O(αβ log k)kU xk1 . (Corollary D.21)
Then, for all w ∈ Rk , kSU wk1 & kU wk1 .
Proof. For any w ∈ Rk we can write it as w = ` · z where ` is some scalar and z has kzk1 = 1.
Define y = arg minky 0 − zk1 .
y 0 ∈N
We first show that if U is an (α, β) well-conditioned basis for `1 , then kU (y −z)k1 ≤ αβkkU yk1 ,
kU (y − z)k1
≤ αky − zk1
by kU (y − z)k1 ≤ αky − zk1
≤ α
by ky − zk1 ≤
by kyk1 = 1
1
kyk1 .
by kU yk1 ≥
βk
= αkyk1
≤ αkU yk1 βk
52
Because < 1/(αβk c+1 ), we have
kU (y − z)k1 ≤
1
kU yk1 .
kc
(20)
Using the triangle inequality, we can lower bound kU yk1 by kU zk1 up to some constant,
kU yk1 ≥ kU zk1 − kU (y − z)k1 ≥ kU zk1 −
1
kU yk1 ,
kc
(21)
which implies
kU yk1 ≥ .99kU zk1 .
(22)
Thus,
kSU zk1
≥ kSU yk1 − kSU (z − y)k1
by triangle inequality
≥ kU yk1 − kSU (z − y)k1
by Lemma D.18
≥ kU yk1 − αβ log(k) · kU (z − y)k1
1
≥ kU yk1 − αβ log(k) · c kU yk1
k
& kU yk1
& kU zk1 ,
by Corollary D.21
by Equation (20)
by k c & αβ log(k)
by Equation (22)
by rescaling z to w, we complete the proof.
Lemma D.23. Given matrix U ∈ Rn×k , let t = O(k log k), and let S ∈ Rt×n be a random matrix
with entries drawn i.i.d. from a standard Cauchy distribution, where each entry is rescaled by Θ(1/t).
With probability .99,
∀x ∈ Rk , kSU xk1 & kU xk1 .
Proof. We can compute a well-conditioned basis for U , and denote it U 0 . Then, ∀x ∈ Rk , there
exists y ∈ Rk such that U x = U 0 y. Due to Lemma D.22, with probability .99, we have
∀y ∈ Rk , kSU 0 yk1 & kU 0 yk1 .
D.6
Sparse Cauchy transform
This section presents the proof of two lemmas related to the sparse Cauchy transform. We first
prove the no dilation result in Lemma D.24. Then we show how to get the no contraction result in
Lemma D.26.
Lemma D.24. Given matrix A ∈ Rn×d , let U ∗ , V ∗ be the optimal solutions of minU,V kU V − Ak1 .
Let Π = σ · SC ∈ Rm×n , where S ∈ Rm×n has each column vector chosen independently and
uniformly from the m standard basis vectors of Rm , where C is a diagonal matrix with diagonals
chosen independently from the standard Cauchy distribution, and σ is a scalar. Then
kΠU ∗ V ∗ − ΠAk1 . σ · log(md) · kU ∗ V ∗ − Ak1
holds with probability at least .999.
53
Proof. We define Π = σ · SC ∈ Rm×n , as in the statement of the lemma. Then,
kΠ(U ∗ V ∗ − A)k1
=
d
X
kSD(U ∗ Vi∗ − Ai )k1
i=1
S11 S12 · · ·
d
X
S21 S22 · · ·
=
··· ··· ···
i=1
Sm1 Sm2 · · ·
c1 S11 c2 S12
d
X
c1 S21 c2 S22
=
···
···
i=1
c1 Sm1 c2 Sm2
=
d X
n
X
cl S1l · (U
∗
Vi∗
S1n
c1 0
0
0
S2n
0
· 0 c2 0
· (U ∗ Vi∗ − Ai )
···
0 0 ··· 0
Smn
0 0
0 cn
· · · cn S1n
· · · cn S2n
· (U ∗ Vi∗ − Ai )
···
···
1
· · · cn Smn
− Ai )l ,
i=1
=
=
l=1
d
m
XX
n
X
i=1 j=1
l=1
|
d X
m
X
cl S2l · (U
∗
Vi∗
− Ai )l , · · · ,
l=1
d X
m X
n
X
n
X
cl Sml · (U ∗ Vi∗ − Ai )l
1
l=1
cl Sjl · (U ∗ Vi∗ − Ai )l |
|w
eij ·
i=1 j=1
=
n
X
1
n
X
|Sjl (U ∗ Vi∗ − Ai )l || where w
eij ∼ C(0, 1)
l=1
|Sjl (U ∗ Vi∗ − Ai )l | · |w
eij |
i=1 j=1 l=1
=
d X
m X
n
X
|Sjl (U ∗ Vi∗ − Ai )l | · wi+(j−1)d where wi+(j−1)d ∼ |C(0, 1)|,
i=1 j=1 l=1
where the last step follows since each wi can be thought of as a clipped half-Cauchy random variable.
Define d0 = md. Define event ξi to be the situation when wi < D (we will decide upon D later).
Define event ξ = ξ1 ∩ ξ2 ∩ · · · ∩ ξd0 . By choosing D = Θ(d0 ), we can conclude that,
E[kΠU ∗ V ∗ − ΠAk1 |ξ] =
d X
m X
n
X
|Sjl (U ∗ Vi∗ − Ai )l | · E[wi+(j−1)d |ξ]
i=1 j=1 l=1
.
d X
m X
n
X
|Sjl (U ∗ Vi∗ − Ai )l | · log(d0 )
i=1 j=1 l=1
=
d X
n
X
|(U ∗ Vi∗ − Ai )l |1 · log(d0 )
by
i=1 l=1
0
m
X
j=1
= log(d ) · kU ∗ V ∗ − Ak1 .
Thus, we can show that
Pr[kΠU ∗ V ∗ − ΠAk1 . log(d0 )kU ∗ V ∗ − Ak1 ] ≥ 0.999.
54
|Sjl | = 1, ∀l ∈ [n]
Lemma D.25. Given any matrix M ∈ Rn×d . Let Π = σ · SC ∈ Rm×n , where S ∈ Rm×n has each
column vector chosen independently and uniformly from the m standard basis vectors of Rm , where
C is a diagonal matrix with diagonals chosen independently from the standard Cauchy distribution,
and σ is a scalar. Then
kΠM k1 . σ · log(md) · kM k1
holds with probability at least .999.
Proof. Just replace the matrix U ∗ V ∗ − A in the proof of Lemma D.24 with M . Then we can get
the result directly.
We already provided the proof of Lemma D.24. It remains to prove Lemma D.26.
Lemma D.26. Given matrix A ∈ Rn×d , let U ∗ , V ∗ be the optimal solution of minU,V kU V −Ak1 . Let
Π = σ · SC ∈ Rm×n , where S ∈ Rm×n has each column vector chosen independently and uniformly
from the m standard basis vectors of Rm , and where C is a diagonal matrix with diagonals chosen
independently from the standard Cauchy distribution. Then with probability at least .999, for all
V ∈ Rk×d ,
kΠU ∗ V − ΠAk1 ≥ kU ∗ V − Ak1 − O(σ · log(md))kU ∗ V ∗ − Ak1 .
Notice that m = O(k 5 log5 k) and σ = O(k 2 log2 k) according to Theorem 2 in [MM13].
We start by using Theorem 2 in [MM13] to generate the following Corollary.
Corollary D.27. Given U ∈ Rn×k with full column rank, let Π = σ · SC ∈ Rm×n where S ∈ Rm×n
has each column chosen independently and uniformly from the m standard basis vectors of Rm , and
where C ∈ Rn×n is a diagonal matrix with diagonals chosen independently from the standard Cauchy
distribution. Then with probability .999, for all x ∈ Rk , we have
kΠU xk1 ≥ kU xk1 .
Notice that m = O(k 5 log5 k) and σ = O(k 2 log2 k) according to Theorem 2 in [MM13].
We give the proof of Lemma D.26.
Proof. Follows by Corollary D.27, Lemma D.24, Lemma D.7.
D.7
`1 -Lewis weights
In this section we show how to use Lewis weights to get a better dilation bound. The goal of this
section is to prove Lemma D.32 and Lemma D.31. Notice that our algorithms in Section C use
Lewis weights in several different ways. The first way is using Lewis weights to show an existence
result, which means we only need an existential result for Lewis weights. The second way is only
guessing the nonzero locations on the diagonal of a sampling and rescaling matrix according to the
Lewis weights. The third way is guessing the values on the diagonal of a sampling and rescaling
matrix according to the Lewis weights. The fourth way is computing the Lewis weights for a known
low dimensional matrix(n×poly(k) or poly(k)×d). We usually do not need to optimize the running
time of computing Lewis weights for a low-rank matrix to have input-sparsity time.
55
Claim D.28. Given matrix A ∈ Rn×d , let B = U ∗ V ∗ − A. For any distribution p = (p1 , p2 , . . . , pn )
define random variable X such that XP= kBi k1 /pi with probability pi . Then take any m independent
m
1
j
samples X 1 , X 2 , . . . , X m , let Y = m
j=1 X . We have
Pr[Y ≤ 1000kBk1 ] ≥ .999
Proof. We can compute the expectation of X j , for any j ∈ [m]
E[X j ] =
n
X
kBi k1
i=1
Then, E[Y ] =
1
m
Pm
j=1 E[X
j]
pi
· pi = kBk1 .
= kBk1 . Using Markov’s inequality, we have
Pr[Y ≥ 1000kBk1 ] ≤ .001
Lemma D.29. Given matrix M ∈ Rn×d , let S ∈ Rn×n be any sampling and rescaling diagonal
matrix. Then with probability at least .999,
kSM k1 . kM k1 .
Proof. Just replace the matrix B in the proof of Claim D.28 with M . Then we can get the result
directly.
Using Theorem 1.1 of [CP15], we have the following result,
Claim D.30. Given matrix A ∈ Rn×d , for any fixed U ∗ ∈ Rn×k and V ∗ ∈ Rk×d , choose D ∈ Rn×n
to be the sampling and rescaling diagonal matrix with m = O(k log k) nonzeros according to the
Lewis weights of U ∗ . Then with probability .999, for all V ,
kU ∗ V ∗ − U ∗ V k1 ≤ kDU ∗ V ∗ − DU ∗ V k1 . kU ∗ V ∗ − U ∗ V k1 .
Lemma D.31. Given matrix A ∈ Rn×d , U ∗ ∈ Rn×k , define V ∗ ∈ Rk×d to be the optimal solution of
min kU ∗ V −Ak1 . Choose a sampling and rescaling diagonal matrix D ∈ Rn×n with m = O(k log k)
V ∈Rk×d
non-zero entries according to the Lewis weights of U ∗ . Then with probability at least .99, we have:
for all V ∈ Rk×d ,
kDU ∗ V − DAk1 . kU ∗ V ∗ − U ∗ V k1 + O(1)kU ∗ V ∗ − Ak1 . kU ∗ V − Ak1 ,
holds with probability at least .99.
Proof. Using the above two claims, we have with probability at least .99, for all V ∈ Rk×d ,
kDU ∗ V − DAk1 ≤ kDU ∗ V − DU ∗ V ∗ k1 + kDU ∗ V ∗ − DAk1
∗
∗
∗
∗
by triangle inequality
∗
. kDU V − DU V k1 + O(1)kU V − Ak1
∗
∗
∗
∗
by Claim D.28
∗
. kU V − U V k1 + O(1)kU V − Ak1
∗
∗
∗
by Claim D.30
∗
∗
≤ kU V − Ak1 + kU V − Ak1 + O(1)kU V − Ak1
∗
. kU V − Ak1 .
56
by triangle inequality
Lemma D.32. Given matrix A ∈ Rn×d , define U ∗ ∈ Rn×k , V ∗ ∈ Rk×d to be the optimal solution
of
min
kU V − Ak1 . Choose a sampling and rescaling diagonal matrix D ∈ Rn×n with
U ∈Rn×k ,V ∈Rk×d
m = O(k log k) non-zero entries according to the Lewis weights of U ∗ . For all V ∈ Rk×d we have
kU ∗ V − Ak1 . kDU ∗ V − DAk1 + O(1)kU ∗ V ∗ − Ak1 ,
holds with probability at least .99.
Proof. Follows by Claim D.28, Lemma D.31, Lemma D.7.
E
`p -Low Rank Approximation
This section presents some fundamental lemmas for `p -low rank approximation problems. Using
these lemmas, all the algorithms described for `1 -low rank approximation problems can be extended
to `p -low rank approximation directly. We only state the important Lemmas in this section, due to
most of the proofs in this section being identical to the proofs in Section D.
E.1
Definitions
This section is just a generalization of Section D.1 to the `p setting when 1 < p < 2.
Definition E.1. Given a matrix M ∈ Rn×d , if matrix S ∈ Rm×n satisfies
kSM kpp ≤ c1 kM kpp ,
then S has at most c1 -dilation on M .
Definition E.2. Given a matrix U ∈ Rn×k , if matrix S ∈ Rm×n satisfies
∀x ∈ Rk , kSU xkpp ≥
1
kU xkpp ,
c2
then S has at most c2 -contraction on U .
Definition E.3. Given matrices U ∈ Rn×k , A ∈ Rn×d , denote V ∗ = arg minV ∈Rk×d kU V − Akpp . If
matrix S ∈ Rm×n satisfies
∀V ∈ Rk×d , kSU V − SAkpp ≥
1
kU V − Akpp − c4 kU V ∗ − Akpp ,
c3
then S has at most (c3 , c4 )-contraction on (U, A).
Definition E.4. A (c5 , c6 ) `p -subspace embedding for the column space of an n × k matrix U is a
matrix S ∈ Rm×n for which all x ∈ Rk
1
kU xkpp ≤ kSU xkpp ≤ c6 kU xkpp .
c5
57
Definition E.5. Given matrices U ∈ Rn×k , A ∈ Rn×d , denote V ∗ = arg minV ∈Rk×d kU V − Akpp .
Let S ∈ Rm×n . If for all c ≥ 1, and if for any Vb ∈ Rk×d which satisfies
kSU Vb − SAkpp ≤ c · min kSU V − SAkpp ,
V ∈Rk×d
it holds that
kU Vb − Akpp ≤ c · c7 · kU V ∗ − Akpp ,
then S provides a c7 -multiple-regression-cost preserving sketch of (U, A).
Definition E.6. Given matrices L ∈ Rn×m1 , N ∈ Rm2 ×d , A ∈ Rn×d , k ≥ 1, let
X ∗ = arg
min
rank −k X
kLXN − Akpp .
b ∈ Rm1 ×m2 which satisfies
Let S ∈ Rm×n . If for all c ≥ 1, and if for any rank −k X
b − SAkpp ≤ c ·
kSLXN
min
rank −k X
kSLXN − SAkpp ,
it holds that
b − Akp ≤ c · c8 · kLX ∗ N − Akp ,
kLXN
p
p
then S provides a c8 -restricted-multiple-regression-cost preserving sketch of (L, N, A, k).
E.2
Properties
Lemma E.7. Given matrices A ∈ Rn×d , U ∈ Rn×k , let V ∗ = arg minV ∈Rk×d kU V − Akpp . If
S ∈ Rm×n has at most c1 -dilation on U V ∗ − A, i.e.,
kS(U V ∗ − A)kpp ≤ c1 kU V ∗ − Akpp ,
and it has at most c2 -contraction on U , i.e.,
∀x ∈ Rk , kSU xkpp ≥
1
kU xkpp ,
c2
then S has at most (22p−2 c2 , c1 + 21−p c12 )-contraction on (U, A), i.e.,
∀V ∈ Rk×d , kSU V − SAkpp ≥
1
1
kU V − Akpp − (c1 + )kU V ∗ − Akpp ,
c2
c2
Proof. Let A ∈ Rn×d , U ∈ Rn×k , and S ∈ Rm×n be the same as that described in the lemma. Then
58
∀V ∈ Rk×d
kSU V − SAkpp ≥ 21−p kSU V − SU V ∗ kpp − kSU V ∗ − SAkpp
≥ 21−p kSU V − SU V ∗ kpp − c1 kU V ∗ − Akpp
= 21−p kSU (V − V ∗ )kpp − c1 kU V ∗ − Akpp
= 21−p
d
X
kSU (V − V ∗ )j kpp − c1 kU V ∗ − Akpp
j=1
≥ 21−p
d
X
1
kU (V − V ∗ )j kpp − c1 kU V ∗ − Akpp
c2
j=1
1
kU V − U V ∗ kpp − c1 kU V ∗ − Akpp
c2
1
1
≥ 22−2p kU V − Akpp − 21−p kU V ∗ − Akpp − c1 kU V ∗ − Akpp
c2
c2
2−2p 1
p
1−p 1
∗
p
=2
kU V − Akp − (2
+ c1 )kU V − Akp .
c2
c2
= 21−p
The first inequality follows by Fact E.14. The second inequality follows since S has at most c1
dilation on U V ∗ − A. The third inequality follows since S has at most c2 contraction on U . The
fourth inequality follows by Fact E.14.
Lemma E.8. Given matrices A ∈ Rn×d , U ∈ Rn×k , let V ∗ = arg minV ∈Rk×d kU V − Akpp . If
S ∈ Rm×n has at most c1 -dilation on U V ∗ − A, i.e.,
kS(U V ∗ − A)kpp ≤ c1 kU V ∗ − Akpp ,
and has at most c2 -contraction on U , i.e.,
∀x ∈ Rk , kSU xkpp ≥
1
kU xkpp ,
c2
then S provides a 2p−1 (2c1 c2 + 1)-multiple-regression-cost preserving sketch of (U, A), i.e., for all
c ≥ 1, for any Vb ∈ Rk×d which satisfies
kSU Vb − SAkpp ≤ c · min kSU V − SAkpp ,
V ∈Rk×d
it has
kU Vb − Akpp ≤ c · 2p−1 (2c1 c2 + 1) · kU V ∗ − Akpp ,
Proof. Let S ∈ Rm×n , A ∈ Rn×d , U ∈ Rn×k , V ∗ , Vb ∈ Rk×d , and c be the same as stated in the
lemma.
kU Vb − Akpp ≤ 22p−2 c2 kSU Vb − SAkpp + (2p−1 + 22p−2 c1 c2 )kU V ∗ − Akpp
≤ 22p−2 c2 c min kSU V − SAkpp + (2p−1 + 22p−2 c1 c2 )kU V ∗ − Akpp
2p−2
≤2
V ∈Rk×d
∗
c2 ckSU V − SAkpp + (2p−1 + 22p−2 c1 c2 )kU V ∗ − Akpp
≤ 22p−2 c1 c2 ckU V ∗ − Akpp + (2p−1 + 22p−2 c1 c2 )kU V ∗ − Akpp
≤ c · 2p−1 (1 + 2c1 c2 )kU V ∗ − Akpp .
59
The first inequality follows by Lemma E.7. The second inequality follows by the guarantee of Vb .
The fourth inequality follows since S has at most c1 -dilation on U V ∗ − A. The fifth inequality
follows since c ≥ 1.
Lemma E.9. Given matrices L ∈ Rn×m1 , N ∈ Rm2 ×d , A ∈ Rn×d , k ≥ 1, let
X ∗ = arg
min
rank −k X
kLXN − Akpp .
If S ∈ Rm×n has at most c1 -dilation on LX ∗ N − A, i.e.,
kS(LX ∗ N − A)kpp ≤ c1 kLX ∗ N − Akpp ,
and has at most c2 -contraction on L, i.e.,
∀x ∈ Rm1 kSLxkpp ≥ kLxkpp ,
then S provides a 2p−1 (2c1 c2 +1)-restricted-multiple-regression-cost preserving sketch of (L, N, A, k),
b ∈ Rm1 ×m2 which satisfies
i.e., for all c ≥ 1, for any rank −k X
b − SAkp ≤ c ·
kSLXN
p
min
rank −k X
kSLXN − SAkpp ,
it has
b − Akp ≤ c · 2p−1 (2c1 c2 + 1) · kLX ∗ N − Akp .
kLXN
p
p
b ∈ Rm1 ×m2 , X ∗ ∈ Rm1 ×m2 , N ∈ Rm2 ×d , A ∈ Rn×d and c ≥ 1
Proof. Let S ∈ Rm×n , L ∈ Rn×m1 , X
be the same as stated in the lemma.
b − SAkpp ≥ 21−p kSLXN
b − SLX ∗ N kpp − kSLX ∗ N − SAkpp
kSLXN
1
b − X ∗ N )kpp − c1 kLX ∗ N − Akpp
≥ 21−p kL(XN
c2
1
b − Akpp − 21−p 1 kLX ∗ N − Ak1 − c1 kLX ∗ N − Akpp
≥ 22−2p kLXN
c2
c2
1
b − Akp − (21−p 1 + c1 )kLX ∗ N − Akp .
= 22−2p kLXN
p
p
c2
c2
The inequality follows from the Fact E.14. The second inequality follows since S has at most c2 contraction on L, and it has at most c1 -dilation on LX ∗ N − A. The third inequality follows by
Fact E.14.
It follows that
b − Akp ≤ 22p−2 c2 kSLXN
b − SAkp + (2p−1 + 22p−2 c1 c2 )kLX ∗ N − Akp
kLXN
p
p
p
≤ 22p−2 c2 c ·
min
rank −k X
∗
kSLXN − SAkpp + (2p−1 + 22p−2 c1 c2 )kLX ∗ N − Akpp
≤ 22p−2 c2 c · kSLX N − SAkpp + (2p−1 + 22p−2 c1 c2 )kLX ∗ N − Akpp
≤ 22p−2 cc1 c2 · kLX ∗ N − Akpp + (2p−1 + 22p−2 c1 c2 )kLX ∗ N − Akpp
≤ c · 2p−1 (1 + 2c1 c2 )kLX ∗ N − Akpp .
The first inequality directly follows from the previous one. The second inequality follows from the
b The fourth inequality follows since S has at most c1 dilation on LX ∗ N − A. The
guarantee of X.
fifth inequality follows since c ≥ 1.
60
Lemma E.10. Given matrices L ∈ Rn×m1 , N ∈ Rm2 ×d , A ∈ Rn×d , k ≥ 1, let
X ∗ = arg
min
rank −k X
kLXN − Akpp .
Let T1 ∈ Rt1 ×n have at most c1 -dilation on LX ∗ N − A, and have at most c2 -contraction on L. Let
e = arg
X
min
rank −k X
kT1 LXN − T1 Akpp .
e − T1 A)> , and at most c0 -contraction on N > .
Let T2> ∈ Rt2 ×d have at most c01 -dilation on (T1 LXN
2
b ∈ Rm1 ×m2 which satisfies
Then, for all c ≥ 1, for any rank −k X
b − SA)T2 kp ≤ c ·
kT1 (LXN
p
min
rank −k X
kT1 (LXN − A)T2 kpp ,
it holds that
b − Akp ≤ c · 22p−2 (2c1 c2 + 1)(2c0 c0 + 1) · kLX ∗ N − Akp .
kLXN
p
1 2
p
b ∈ Rm1 ×m2
Proof. Apply Lemma D.9 for sketching matrix T2 . Then for any c ≥ 1, any rank −k X
which satisfies
b − A)T2 kpp ≤ c ·
kT1 (LXN
min
rank −k X
kT1 (LXN − A)T2 kpp ,
it has
b − A)kpp ≤ c · 2p−1 (2c01 c02 + 1) · kT1 (LXN
e − A)kpp .
kT1 (LXN
b ∈ Rm1 ×m2 which
Apply Lemma D.9 for sketch matrix T1 . Then for any c ≥ 1, any rank −k X
satisfies
b − A)kp ≤ c2p−1 (2c0 c0 + 1) ·
kT1 (LXN
p
1 2
min
rank −k X
e − A)kp ,
kT1 (LXN
p
it has
b − Akp ≤ c · 22p−2 (2c1 c2 + 1)(2c0 c0 + 1) · kLX ∗ N − Akp .
kLXN
p
1 2
p
Lemma E.11. Given matrices M ∈ Rn×d , U ∈ Rn×t , d ≥ t = rank(U ), n ≥ d ≥ r = rank(M ). If
sketching matrix S ∈ Rm×n is drawn from any of the following probability distributions on matrices,
with .99 probability, S has at most c1 -dilation on M , i.e.,
kSM kpp ≤ c1 kM kpp ,
and S has at most c2 -contraction on U , i.e.,
∀x ∈ Rt , kSU xkpp ≥
1
kU xkpp ,
c2
where c1 , c2 are parameters depend on the distribution over S.
(I) S ∈ Rm×n is a dense matrix with entries drawn from a p-stable distribution: a matrix with
i.i.d. standard p-stable random variables. If m = O(t log t), then c1 c2 = O(log d).
61
(II) S ∈ Rm×n is a sparse matrix with some entries drawn from a p-stable distribution: S =
T D, where T ∈ Rm×n has each column drawn i.i.d. from the uniform distribution over
standard basis vectors of Rm , and D ∈ Rn×n is a diagonal matrix with each diagonal entry
drawn from i.i.d. from the standard p-stable distribution. If m = O(t5 log5 t), then c1 c2 =
O(t2/p log2/p t log d). If m = O((t + r)5 log5 (t + r)), then
c1 c2 = O(min(t2/p log2/p t log d, r3/p log3/p r)).
(III) S ∈ Rm×n is a sampling and rescaling matrix (notation S ∈ Rn×n denotes a diagonal
sampling and rescaling matrix with m non-zero entries): If S samples and reweights m =
O(t log t log log t) rows of U , selecting each with probability proportional to the ith row’s `p
Lewis weight and reweighting by the inverse probability, then c1 c2 = O(1).
In the above, if we replace S with σ · S where σ ∈ R\{0} is any scalar, then the relation between m
and c1 c2 can be preserved.
For (I), it is implied by Lemma E.17, Lemma E.19. Also see from [SW11]1 .
For (II), if m = O(t5 log5 t), then c1 c2 = O(t2/p log2/p t log d) is implied by Corollary D.27 and
Lemma E.20 and Theorem 4 in [MM13]. If m = O((t + r)5 log5 (t + r)), c1 c2 = O(r3/p log3/p r) is
implied by [MM13].
For (III), it is implied by [CP15] and Lemma D.29.
E.3
Tools and inequalities
Lemma E.12 (Lemma 9 in [MM13], Upper Tail Inequality for p-stable Distributions). Let p ∈ (1, 2)
and m ≥ 3. ∀i ∈ [m], letPXi be m(not necessarily
random variables sampled from Dp ,
Pm independent)
p . Then for any t ≥ 1,
γ
|X
|
γ
.
Let
X
=
and let γi > 0 with γ = m
i
i=1 i
i=1 i
Pr[X ≥ tαp γ] ≤
2 log(mt)
.
t
We first review some facts about the p-norm and q-norm,
Fact E.13. For any p ≥ q > 0 and any x ∈ Rk ,
1
kxkp ≤ kxkq ≤ k q
− p1
kxkp .
We provide the triangle inequality for the p-norm,
Fact E.14. For any p ∈ (1, 2), for any x, y ∈ Rk ,
kx + ykp ≤ kxkp + kykp , and kx + ykpp ≤ 2p−1 (kxkpp + kykpp ).
Fact E.15 (Hölder’s inequality). For any x, y ∈ Rk , if
1
p
+
1
q
= 1, then |x> y| ≤ kxkp kykq .
We give the definition of a well-conditioned basis for `p ,
Definition E.16. Let p ∈ (1, 2). A basis U for the range of A is (α, β, p)-conditioned if kU kp ≤ α
and for all x ∈ Rk , kxkq ≤ βkU xkp . We will say U is well-conditioned if α and β are low-degree
polynomials in k, independent of n.
1
Full version.
62
Proof. We first show an upper bound,
kU xkp ≤ kU kp · kxkp ≤ αkxkp
Then we show a lower bound,
kU xkp ≥
1
kxkq
β
For any p and q with 1/p + 1/q = 1, by Hölder’s inequality we have
|x> y| ≤ kxkp · kykq
choosing y to be the vector that has 1 everywhere, kxk1 ≤ kxkp k 1/q
E.4
Dense p-stable transform
This section states the main tools for the dense p-stable transform. The proof is identical to that
for the dense Cauchy transform.
Lemma E.17. Given matrix A ∈ Rn×d and p ∈ (1, 2), define U ∗ ∈ Rn×k , V ∗ ∈ Rk×d to be an
kU V − Akp . Choose a p-stable distribution matrix S ∈ Rm×n ,
optimal solution of
min
rescaled by
U ∈Rn×k ,V ∈Rk×d
1/p
Θ(1/m ). Then we have
kSU ∗ V ∗ − SAkpp . log(md)kU ∗ V ∗ − Akpp
with probability at least 99/100.
Proof. Let P (0, 1) denote the p-stable distribution. Then,
kSU ∗ V ∗ − SAkpp ≤
=
d
X
kS(U ∗ Vi∗ − Ai )kpp
i=1
d X
m
X
i=1 j=1
n
X
1
|
Sj,l (U ∗ Vi∗ − Ai )l |p
m
d m
n
X
1 XX
=
|w
eij (
|(U ∗ Vi∗ − Ai )l |p )1/p |p
m
i=1 j=1
d
where Sj,l ∼ P (0, 1)
l=1
m
where w
eij ∼ P (0, 1)
l=1
n
1 XXX
|(U ∗ Vi∗ − Ai )l |p · |w
eij |p
=
m
i=1 j=1 l=1
d
m
1 XX ∗ ∗
p
=
kU Vi − Ai kpp · wi+(j−1)d
,
m
where wi+(j−1)d ∼ |P (0, 1)|
i=1 j=1
where the last step
wi can be thought of as a clipped
random variable.
Pdfollows
Pm since∗ each
Pdhalf-p-stable
Pm
p
p
∗
∗
Define X to be i=1 j=1 kU Vi − Ai kp · wi+(j−1)d and γ to be i=1 j=1 kU Vi∗ − Ai kpp . Then
applying Lemma E.12,
Pr[X ≥ tαp γ] ≤
63
2 log(mdt)
.
t
Choosing t = Θ(log(md)), we have with probability .999,
X . log(md)αp γ = log(md)αp
d
X
kU ∗ Vi∗ − Ai kpp ,
i=1
where the last steps follows by definition of γ. Thus, we can conclude that with probability .999,
kΠ(U ∗ V ∗ − A)kpp . log(md)kU ∗ V ∗ − Akpp .
Lemma E.18. Given matrix A ∈ Rn×d and p ∈ (1, 2), define U ∗ ∈ Rn×k to be the optimal solution
of
min
kU V − Akp . Choose a matrix of i.i.d. p-stable random variables S ∈ Rm×n . Then
U ∈Rn×k ,V ∈Rk×d
for all V ∈ Rk×n ,
we have
kSU ∗ V − SAkpp & kU ∗ V − Akpp − O(log(md))kU ∗ V ∗ − Akpp .
Lemma E.19. Let p ∈ (1, 2). Given an (α, β) `p well-conditioned basis, condition on the following
two events,
1. For all x ∈ N , kSU xkp & kU xkp .
2. For all x ∈ Rk , kSU xkp ≤ poly(k)kU xkp .
Then for all x ∈ Rk , kSU xkp & kU xkp .
Proof. The proof is identical to Lemma D.22 in Section D.
E.5
Sparse p-stable transform
This section states the main tools for the sparse p-stable transform. The proof is identical to that
of the sparse Cauchy transform.
Lemma E.20. Let p ∈ (1, 2). Given matrix A ∈ Rn×d with U ∗ , V ∗ an optimal solution of
minU,V kU V − Akp , let Π = σ · SC ∈ Rm×n , where S ∈ Rm×n has each column vector chosen
independently and uniformly from the m standard basis vectors of Rm , where C is a diagonal matrix
with diagonals chosen independently from the standard p-stable distribution, and σ is a scalar. Then
kΠU ∗ V ∗ − ΠAkpp . σ · log(md) · kU ∗ V ∗ − Akpp
holds with probability at least .999.
Proof. We define Π = σ · SC ∈ Rm×n as in the statement of the lemma. Then by the definition of
64
Π, we have,
kΠ(U ∗ V ∗ − A)kpp
=
d
X
kSC(U ∗ Vi∗ − Ai )kpp
i=1
S11 S12 · · ·
d
X
S21 S22 · · ·
=
··· ··· ···
i=1
Sm1 Sm2 · · ·
c1 S11 c2 S12
d
X
c1 S21 c2 S22
=
···
···
i=1
c1 Sm1 c2 Sm2
=
d X
n
X
c1 0
0
0
S1n
0
S2n
· (U ∗ Vi∗ − Ai )
· 0 c2 0
··· 0 0 ··· 0
0 0
0 cn
Smn
· · · cn S1n
p
· · · cn S2n
· (U ∗ Vi∗ − Ai )
···
···
p
· · · cn Smn
cl S1l · (U ∗ Vi∗ − Ai )l ,
i=1
=
l=1
d X
m X
n
X
n
X
cl S2l · (U ∗ Vi∗ − Ai )l , · · · ,
l=1
n
X
p
p
p
cl Sml · (U ∗ Vi∗ − Ai )l
p
l=1
p
cl Sjl · (U ∗ Vi∗ − Ai )l
by aX + bY and (|a|p + |b|p )1/p Z are indentically distributed
i=1 j=1 l=1
=
d X
m
X
n
X
|Sjl (U ∗ Vi∗ − Ai )l |p )1/p
w
eij · (
i=1 j=1
=
d X
m X
n
X
p
where w
eij ∼ P (0, 1)
l=1
|Sjl (U ∗ Vi∗ − Ai )l |p · |w
eij |p
i=1 j=1 l=1
=
d X
m X
n
X
p
|Sjl (U ∗ Vi∗ − Ai )l |p · wi+(j−1)d
, where wi+(j−1)d ∼ |P (0, 1)|
i=1 j=1 l=1
where the last step
be thought of as a clipped half-p-stable
P follows
P since
Pn each wi∗can
Pd Pm random
Pn variable.
∗ −A ) |p ·w p
∗ ∗
|S
(U
V
Define X to be di=1 m
and
γ
to
be
i
jl
l
i
j=1
l=1
i=1
j=1
l=1 |Sjl (U Vi −
i+(j−1)d
Ai )l |p . Then applying Lemma E.12,
Pr[X ≥ tαp γ] ≤
2 log(mdt)
.
t
Choosing t = Θ(log(md)), we have with probability .999,
X . log(md)αp γ = log(md)αp
d X
n
X
|(U ∗ Vi∗ − Ai )l |p ,
i=1 l=1
where the last steps follows by
γ=
d X
m X
n
X
i=1 j=1 l=1
|Sjl (U ∗ Vi∗ − Ai )l |p =
d X
m X
n
X
|Sjl |p |(U ∗ Vi∗ − Ai )l |p =
i=1 j=1 l=1
d X
n
X
|(U ∗ Vi∗ − Ai )l |p .
i=1 l=1
Thus, we can conclude that with probability .999, kΠ(U ∗ V ∗ − A)kpp . log(md)kU ∗ V ∗ − Akpp .
65
Lemma E.21. Given matrix A ∈ Rn×d with U ∗ , V ∗ an optimal solution of minU,V kU V − Akp , let
Π = σ · SC ∈ Rm×n , where S ∈ Rm×n has each column vector chosen independently and uniformly
from the m standard basis vectors of Rm , and where C is a diagonal matrix with diagonals chosen
independently from the standard p-stable distribution. Then with probability at least .999, for all
V ∈ Rk×d ,
kΠU ∗ V − ΠAkpp ≥ kU ∗ V − Akpp − O(σ log(md))kU ∗ V ∗ − Akpp .
Notice that m = O(k 5 log5 k) and σ = O((k log k)2/p ) according to Theorem 4 in [MM13].
E.6
`p -Lewis weights
This section states the main tools for `p -Lewis weights. The proof is identical to `1 -Lewis weights.
Lemma E.22. For any p ∈ (1, 2). Given matrix A ∈ Rn×d , define U ∗ ∈ Rn×k , V ∗ ∈ Rk×d to be
an optimal solution of
min
kU V − Akp . Choose a diagonal matrix D ∈ Rn×n according to
the Lewis weights of
U ∈Rn×k ,V ∈Rk×d
U ∗ . We have that
kDU ∗ V ∗ − DAkpp . kU ∗ V ∗ − Akpp ,
holds with probability at least .99.
Lemma E.23. Let p ∈ (1, 2). Given matrix A ∈ Rn×d , define U ∗ ∈ Rn×k , V ∗ ∈ Rk×d to be an
optimal solution of
min
kU V − Akp . Choose a sampling and rescaling matrix D ∈ Rn×n
U ∈Rn×k ,V ∈Rk×d
according to the Lewis weights of U ∗ . For all V ∈ Rk×d we have
kDU ∗ V − DAkpp & kU ∗ V − Akpp − O(1)kU ∗ V ∗ − Akpp ,
holds with probability at least .99.
F
EMD-Low Rank Approximation
In this section we explain how to embed EMD to `1 . For more detailed background on the EarthMover Distance(EMD) problem, we refer the reader to [IT03, AIK08, ABIW09, IP11, BIRW16] and
[SL09, LOG16]. Section F.1 introduces some necessary notation and definitions for Earth-Mover
Distance. Section F.2 presents the main result for the Earth-Mover distance low rank approximation
problem.
F.1
Definitions
[∆]2
Consider any two non-negative vectors x, y ∈ R+ such that kxk1 = kyk1 . P
Let Γ(x, y) be the
2
2
2
set of functions γ : [∆] × [∆] → R+ , such that for any i, j ∈ [∆] we have l γ(i, l) = xi and
P
l γ(l, j) = yj ; that is, Γ is the set of possible “flows” from x to y. Then we define
X
EMD(x, y) = min
γ(i, j)ki − jk1
γ∈Γ
i,j∈[∆]2
to be the min cost flow from x to y, where the cost of an edge is its `1 distance.
66
Using the EMD(·, ·) metric, for general vectors w, we define k · kEEMD distance (which is the
same as [SL09]),
kwkEEMD = min EMD(x, y) + 2∆kzk1 .
x−y+z=w
kxk1 =kyk1
x,y≥0
Using k · kEEMD distance, for general matrices X ∈ Rn×d , we define the k · k1,EEMD distance,
kXk1,EEMD =
d
X
kXi kEEMD ,
i=1
where Xi denotes the j-th column of matrix X.
F.2
Analysis of no contraction and no dilation bound
Lemma F.1. Given matrix A ∈ Rn×d and U ∗ , V ∗ =
arg min
kU V − Ak1,EEMD , there exist
U ∈Rn×k ,V ∈Rk×d
sketching matrices S ∈ Rm×n such that, with probability .999, for all V ∈ Rk×d ,
kS(U ∗ V − A)k1 ≥ kU ∗ V − Ak1,EEMD
holds.
Proof. Using Lemma 1 in [IT03], there exists a constant C > 0 such that for all i ∈ [d],
CkSU ∗ Vi − SAi k1 ≥ kU ∗ Vi − Ai kEEMD .
(23)
Then taking a summation over all d terms and rescaling the matrix S, we obtain,
d
X
kS(U ∗ Vi − Ai )k1 ≥ kU ∗ Vi − Ai kEEMD
i=1
which completes the proof.
Lemma F.2. Given matrix A ∈ Rn×d and U ∗ , V ∗ =
arg min
kU V − Ak1,EEMD , there exist
U ∈Rn×k ,V ∈Rk×d
sketching matrices S ∈ Rm×n such that
kS(U ∗ V ∗ − A)k1 ≤ O(log n)kU ∗ V ∗ − Ak1,EEMD
holds with probability at least .999.
Proof. Using Lemma 2 in [IT03], we have for any i ∈ [d],
E[kSU ∗ Vi∗ − SAi k1 ] ≤ O(log n)kU ∗ Vi∗ − Ai kEEMD .
Then using that the expectation is linear, we have
d
X
E[kSU V − SAk1 ] = E[
kSU ∗ Vi∗ − SAi k1 ]
∗
∗
=
i=1
d
X
E[kSU ∗ Vi∗ − SAi k1 ]
i=1
≤
d
X
O(log n)kU ∗ Vi∗ − Ai kEEMD
i=1
= O(log n)kU ∗ V ∗ − Ak1,EEMD .
Using Markov’s inequality, we can complete the proof.
67
by Equation (24)
(24)
Theorem F.3. Given a matrix A ∈ Rn×d , there exists an algorithm running in poly(k, n, d) time
that is able to output U ∈ Rn×k and V ∈ Rk×d such that
kU V − Ak1,EEMD ≤ poly(k) · log d · log n
min
rank −k Ak
kAk − Ak1,EEMD
holds with probability .99.
Proof. First using Lemma F.1 and Lemma F.2, we can reduce the original problem into an `1 low rank approximation problem by choosing m = poly(n). Second, we can use our `1 -low rank
approximation algorithm to solve it. Notice that all of our `1 -low rank approximation algorithms
can be applied here. If we apply Theorem C.6, we complete the proof.
Our current k · k1,EEMD is column-based. We can also define it to be row-based. Then we get a
slightly better result by applying the `1 -low rank algorithm.
Corollary F.4. Given a matrix A ∈ Rn×d , there exists an algorithm running in poly(k, n, d) time
that is able to output U ∈ Rn×k and V ∈ Rk×d such that
kU V − Ak1,EEMD ≤ poly(k) · log2 d
min
rank −k Ak
kAk − Ak1,EEMD
holds with probability .99.
G
Hardness results for Cauchy matrices, row subset selection, OSE
Section G.1 presents some inapproximability results by using random Cauchy matrices. Section G.2
is a warmup for inapproximability results for row subset selection problems. Section G.3 shows the
inapproximability results by using any linear oblivious subspace embedding(OSE), and also shows
inapproximability results for row subset selection.
G.1
Hard instance for Cauchy matrices
The goal of this section is to prove Theorem G.2. Before stating the result, we first introduce some
useful tools in our analysis.
Lemma G.1 (Cauchy Upper Tail Inequality, Lemma 3 of [CDMI+ 13]). For i ∈ [m], letPCi be m
random Cauchy
variables from C(0, 1) (not necessarily independent), and γi > 0 with γ = i∈[m] γi .
P
Let X = i∈[m] γi |Ci |. Then, for any t ≥ 1,
Pr[X > γt] ≤ O(log(mt)/t).
Theorem G.2. Let k ≥ 1. There exist matrices A ∈ Rd×d such that for any o(log d) ≥ t ≥ 1,
where c can be any constant smaller than 1/3, for random Cauchy matrices S ∈ Rt×d where each
entry is sampled from an i.i.d. Cauchy distribution C(0, γ) where γ is an arbitrary real number,
with probability .99 we have
min kU SA − Ak1 ≥ Ω(log d/(t log t))
min
rank −k A0
U ∈Rd×t
68
kA0 − Ak1 .
Proof. We define matrix A ∈ Rd×d
1
1
0
0
0
0
A=B+I =α·
· · · · · ·
0
0
1
0
0
···
0
1
1
0
0
0
+ 0
· · ·
···
0
0
···
···
···
···
···
0
1
0
···
0
0
0
1
···
0
···
···
···
···
···
0
0
0
,
· · ·
1
where α = Θ(log d).So, if we only fit the first row of A, we can get approximation cost at most
O(d).
For t > 0, let S ∈ Rt×d denote a random Cauchy matrix where Si,j denotes the entry in the ith
row and jth column. Then SA is
S1,1 S1,1 S1,1 · · · S1,1
S1,1 S1,2 S1,3 · · · S1,d
S2,1 S2,1 S2,1 · · · S2,1 S2,1 S2,2 S2,3 · · · S2,d
SA = SB + SI = α ·
S3,1 S3,1 S3,1 · · · S3,1 + S3,1 S3,2 S3,3 · · · S3,d .
··· ··· ··· ··· ··· ··· ··· ··· ··· ···
St,1 St,1 St,1 · · · St,1
St,1 St,2 St,3 · · · St,d
Since ∀γ,
min kU SA − Ak1 = min kγU SA − Ak1 ,
U ∈Rd×t
U ∈Rd×t
without loss of generality, we can let Si,j ∼ C(0, 1). Then we want to argue that, if we want to use
SA to fit the first row of A, with high probability, the cost will be Ω(d log d).
Let b ∈ Rd denote the first row of A. Then, we want to use SA to fit the first row of A, which
is a d-dimensional vector that has entry Θ(log d) on each position. The problem is equiavlent to
min k(SA)> x − bk1 .
x∈Rt
First, we want to show that for any x in Rt , if x> SA fits the first row of A very well, then the `1
and `2 norm of vector x must have reasonable size.
Claim G.3. Define A1 to be the first row of matrix A. With probability .999, for any column vector
x ∈ Rt×1 , if kx> SA − A1 k1 ≤ o(d log d), then
Property (I) : kxk2 ≥ Ω(1/t log t);
Property (II) : kxk1 ≤ O(log d).
Proof. Consider the absolute value of the i-th coordinate of x> SA. We can rewrite |h(SA)i , xi| in
the following sense,
|h(SA)i , xi| = |h(SB)i , xi + h(SI)i , xi| = |h(SB)1 , xi + h(SI)i , xi| = |h(SB)1 , xi + hSi , xi|,
(25)
where the second step follows because (SB)i = (SB)1 , ∀i ∈ [n], and the last step follows because
(SI)i = Si , ∀i ∈ [n].
We start by proving Property I. Using the triangle inequality and Equation (25),
|h(SA)i , xi|
≤ |h(SB)1 , xi| + |hSi , xi|
≤ k(SB)1 k2 kxk2 + kSi k2 kxk2
by Cauchy-Schwarz inequality
≤ k(SB)1 k1 kxk2 + kSi k1 kxk2 .
by k · k2 ≤ k · k1
69
Then, according to Lemma G.1, with probability .99999, we have k(SB)1 k1 ≤ O(t log t log d) and
for a fixed i ∈ [d], kSi k1 ≤ O(t log t). Applying the Chernoff bound, with probability 1 − 2−Ω(d) ,
there are a constant fraction of i such that kSi k1 = O(t log t). Taking the union bound, with
probability .9999, there exists a constant fraction of i such that |h(SA)i , xi| ≤ O(t log t log d)kxk2 .
Because A1,i ≥ α, ∀i ∈ [d] where α = Θ(log d), we need kxk2 ≥ Ω(1/t log t). Otherwise, the total
cost on this constant fraction of coordinates will be at least Ω(d log d).
For Property II, for a fixed x ∈ Rt , we have
|h(SA)i , xi|
=|h(SB)1 , xi + hSi , xi|
by Equation (25)
= Θ(log d)kxk1 w10 (x) + kxk1 wi0 (x) .
where wi0 (x) ∼ C(0, 1), and for different x, wi0 (x) are different. Then for a fixed x ∈ Rt with
probability at least 0.9, |wi0 (x)| = Ω(1), and with probability 0.5, w10 (x) and wi0 (x) have the same
sign. Since these two events are independent, with probability at least 0.45, we have
|h(SA)i , xi| ≥ kxk1 · Ω(1).
Applying the Chernoff bound, with probability at least 1 − 2Θ(d) , there exists a 3/10 fraction of i
such that |h(SA)i , xi| ≥ kxk1 Ω(1).
We build an -net N for x ∈ Rt on an `1 -norm unit ball, where = 1/(t2 log2 d). Thus the size
e
of the net is |N | = 2Θ(t) . Consider y to be an arbitrary vector, let y/kyk1 = x + δ, where x is the
closest point to y/kyk1 and x ∈ N . For any δ ∈ Rt ,
|h(SA)i , δi| = |h(SB)1 , δi + h(SI)i , δi|
≤ |h(SB)1 , δi| + |h(SI)i , δi|
by triangle inequality
≤ k(SB)1 k2 kδk2 + kSi k2 kδk2
by Cauchy-Schwarz inequality
≤ k(SB)1 k2 kδk1 + kSi k2 kδk1 .
by k · k2 ≤ k · k1
As we argued before, With probability .99999, we have
k(SB)1 k2 ≤ k(SB)1 k1 ≤ O(t log t log d),
(26)
and with probability 1 − 2−Θ(d) , there is a 9/10 fraction of i ∈ [d], kSi k2 ≤ kSi k1 = O(t log t).
Therefore, with probability .999, for any δ ∈ Rt , there exists a 9/10 fraction of i such that
|h(SA)i , δi| ≤ Θ(t log t log d)kδk1 .
Therefore, with probability .99, ∀y ∈ Rt , due to the pigeonhole principle, there is a 3/10 +
9/10 − 1 = 1/5 fraction of i such that
|h(SA)i , yi|
= kyk1 · |h(SA)i , y/kyk1 i|
= kyk1 · |h(SA)i , x + δi|
by y/kyk1 = x + δ
≥ kyk1 · |h(SB)i , xi + h(SI)i , xi| − |h(SB)1 , δi + h(SI)i , δi|
≥ kyk1 Ω(1) − O(t log t log d)
by triangle inequality
≥ kyk1 Ω(1).
So kyk1 should be O(log d). Otherwise, the total cost on this 1/5 fraction of coordinates is at
least Ω(d log d).
Combining Property (I) and (II) completes the proof.
70
Next, we need to show the following claim is true,
Claim G.4. For any d independent Cauchy random variables x1 , x2 , · · · , xd from C(0, 1), with
probability 1 − 1/ poly(t log d), for any j ∈ [1, 2, · · · , log d − Θ(log log(t log d))], there are Ω(d/2j )
variables belonging to (2j , 2j+1 ].
Proof. For each Cauchy random variable xi , we have for any j ∈ [1, 2, · · · , Θ(log d)],
j j+1
Pr |xi | ∈ (2 , 2 ] = Θ(1/2j ).
We define the indicator random variable zi,j
(
1 if |xi | ∈ (2j , 2j+1 ],
zi,j =
0 otherwise.
We define zj =
Pd
i=1 zi,j .
It is clear that E[zi,j ] = Θ(1/2j ). We use a Chernoff bound,
µ
e−δ
,
Pr[X < (1 − δ)µ] <
(1 − δ)1−δ
and set X = zj , δ = 1/2, µ = E[zj ]. Then, this probability is at most 2−Ω(µ) = 2−Ω(E[zj ]) . For any
j ∈ [1, log d − Θ(log log(t log d))], we have E[zj ] = d E[zij ] = Ω(d/2j ) = Ω(log(t log d)). Thus, this
probability is at most 1/ poly(t log d). Overall, we have
j
Pr zj & d/2 ≥ 1 − 1/ poly(t log d).
Taking a union bound over Θ(log d) such j, we complete the proof.
Claim G.5. Let 1 > c1 > c2 > 1/3 > 0 be three arbitrary constants. We fix the first column of
(SI)> ∈ Rd×t . All the rows are grouped together according to the value in the first column. Let
Rj be the set of rows for which the entry in the first column is ∈ (2j , 2j+1 ]. With probability at
least 1 − O(1/ poly(t log(d))), for any j ∈ [c2 log d, c1 log d], the following event holds. There exist
Ω(d/2j ) rows such that the first coordinate ∈ (2j , 2j+1 ] and all the other coordinates are at most
O(d1/3 ).
Proof. Let Rj be a subset of rows of S > , such that for any row in Rj , the first coordinate is
∈ (2j , 2j+1 ]. By Claim G.4, we have that with probability 1 − 1/ poly(t log d), for any j ∈
[c2 log d, c1 log d], |Rj | ≥ Ω(d/2j ). We want to show that for any j ∈ [c2 log d, c1 log d], there exists
a constant fraction of rows in Rj such that, the remaining coordinates are at most O(d1/3 ).
For a fixed row in Rj and a fixed coordinate in that row, the probability that the absolute value
of this entry is at least Ω(d1/3 ) is at most O(1/d1/3 ). By taking a union bound, with probability at
least 1 − O(t/d1/3 ), the row has every coordinate of absolute value at most O(d1/3 ).
By applying a Chernoff bound, for a fixed subset Rj of rows, the probability that there is a
constant fraction of these rows for which every coordinate except the first one in absolute value is
1−c
at most O(d1/3 ), is at least 1 − 2−Θ(|Rj |) ≥ 1 − 2−Θ(d 1 ) ≥ 1 − O(1/ poly(t log(d))).
After taking a union over all the j, we complete the proof.
Claim G.6. With probability at least 1 − O(1/t), the absolute value of any coordinate in column
vector (SB)1 ∈ Rt×1 is at most O(t2 log d).
71
Proof. Because each entry is sampled from a Cauchy distribution C(0, 1) scaled by O(log d), then
with probability at most O(1/t2 ), the absolute value of one coordinate is at least Ω(t2 log d). Taking
a union over all the coordinates, we complete the proof.
Because kxk1 ≤ O(log d), kxk2 ≥ Ω(1/t log t), there exists one coordinate i such that the absolute
value of it is at least Ω(1/t log t) and all the other coordinates are most O(log d). We can assume
that i = 1 for now. (To remove this assumption, we can take a union bound over all the possibilities
bj denote a subset of Rj , which is a “good”
for i ∈ [t].) According to Claim G.5 and G.6, let R
bj ⊂ Rj ⊂ [d], we have
constant fraction of Rj . Considering the `1 -norm of all coordinates l ∈ R
X
|((SA)> x)l − O(log d)|
bj
l∈R
≥
X
|h(SA)l , xi − O(log d)|
bj
l∈R
≥
X
|(SI)1,l · x1 | −
t
X
|(SI)j,l · xj | − |h(SB)l , xi| − O(log d)
j=2
bj
l∈R
X
2j
1/3
≥
Ω(
) − O(td log d) − |h(SB)1 , xi| − O(log d)
t log t
bj
l∈R
X
2j
1/3
) − O(td log d) − k(SB)1 k1 kxk1 − O(log d)
≥
Ω(
t log t
bj
l∈R
X
2j
1/3
2
≥
Ω(
) − O(td log d) − O(t log t log d) − O(log d)
t log t
bj
l∈R
&
X
2j /(t log t)
bj
l∈R
&d/2j · 2j /(t log t)
&d/(t log t).
The second inequality follows by the triangle inequality. The third inequality follows by (SB)1 =
(SB)l , |x1 | = Ω(1/t log t), (SI)1,l ∈ [2j , 2j+1 ), and ∀j 6= 1, |xj | < O(log d), (SI)j,l ≤ O(d1/3 ). The
fourth inequality follows by Cauchy-Schwarz and k · k2 ≤ k · k1 . The fifth inequality follows by
Equation (26) and Claim G.3. The sixth inequality follows by t = o(log d) where c is a constant
bj | ≥ Ω(d/2j ).
smaller than 1/3 and 2j ≥ dc2 > poly(t). The seventh inequality follows from |R
Since there are c1 − c2 different j, the total cost is Ω(d log d/(t log t)). The gap then is
Ω(log d/(t log t)). This completes the proof of Theorem G.2.
G.2
Hard instance for row subset selection
Theorem G.7. Let ∈ (0, 1). There exists a value k ≥ 1 and matrix A ∈ R(d−1)×d such that, for
any subset R of rows of A, letting B be the `1 -projection of each row of A onto R, we have
kA − Bk1 > (2 − )
min
rank −k A0
unless |R| & d.
72
kA0 − Ak1 ,
Proof. We construct the (d − 1) × d matrix A in the following way. The first column is all 1s, and
then the remaining (d − 1) × (d − 1) matrix is the identity matrix.
1
1
A=
1
· · ·
1
Note that
min
rank −1 A0
1
0
0
···
0
0
1
0
···
0
0
0
1
···
0
···
···
···
···
···
0
0
0
.
· · ·
1
kA − A0 k1 = OPT ≤ d − 1, since one could choose A0 to have the first column
all 1s and remaining entries 0. On the other hand, consider any subset of r rows. We can permute
the columns and preserve the entrywise `1 -norm so w.l.o.g., we can take the first r rows for R.
Because we are taking the first r rows, we do not pay any cost on the first rows. To minimize
the cost on the i-th row, where i ∈ {r + 1, r + 2, . . . , d}, let vi denote the row vector we use for the
i-th row. Then vi can be written as a linear combination of the first r rows {A1 , A2 , . . . , Ar },
vi =
r
X
αi,j Aj .
j=1
Then the cost of using vi to approximate the i-th row of A is:
kvi − Ai k1 = (cost in 1st col) + ( cost in 2nd,3rd, . . . , r + 1th cols) + ( cost in i + 1th col)
r
r
X
X
= | αi,j − 1| +
|αi,j | + 1
≥ |
j=1
r
X
j=1
r
X
j=1
j=1
αi,j − 1 −
αi,j | + 1 by triangle inequality
= 2.
Hence, the cost of using vi to approximate the i-th row of A is at least 2. So in total, across these
(d − 1 − r) rows, the algorithm pays at least 2(d − 1 − r) cost, which needs to be at most C(d − 1),
and therefore r ≥ (d − 1)(1 − C/2) = Ω(d). Choosing C = 2 − completes the proof.
Theorem G.8. There exists a value k ≥ 1 and matrix A ∈ R(d−1)×d such that, there is no algorithm
that is able to output a rank −k matrix B in the row span of A satisfying
1
kA − Bk1 < 2(1 − Θ( )) min 0 kA0 − Ak1 .
d rank −k A
Proof. We use the same matrix A as in the previous theorem.
1
1
0
0 ···
1
0
1
0 ···
0
0
1 ···
A= 1
· · · · · · · · · · · · · · ·
1
0
0
0 ···
73
0
0
0
.
· · ·
1
P
Let vector v be ( d−1
i=1 βi , β1 , . . . , βd−1 ). For the i-th row of A, we use αj · v to cancel the cost of
that row, where αj is a scalar. Then for any β1 , . . . , βd−1 , α1 , . . . , αd−1 , we can compute the entire
residual cost,
f (α, β) =
d−1
X
j
kA − vαj k1 =
d−1
X
j=1
d−1
X
X
|1 − αj (
βi )| + |1 − αj βj | +
|αj βi | .
j=1
i=1
i6=j
In the next few paragraphs, we will show that optimizing f (α, β) over some extra constraints
for
P
β
=
1.
α, β does not change the optimality. Without loss of generality, we can assume that d−1
i
i=1
If not we can just rescale all the αi . Consider a fixed index j. All the terms related to αj and βj
are,
|1 − αj (
d−1
X
βi )| + |1 − αj βj | +
i=1
X
|αj βi | + |αi βj |.
i6=j
We first show a simple proof by assuming βj ≥ 0. Later, we will prove the general version which
does not have any assumptions.
Handling a special case We can optimize f (α, β) in the following sense,
min f (α, β)
s.t.
d−1
X
βj = 1
j=1
αj ≥ 0, βj ≥ 0, ∀j ∈ [d − 1].
For each j, we consider three cases. Case I, if αj ≥
|1 − αj | + |1 − αj βj | +
X
1
βj ,
αj βi = αj − 1 + αj βj − 1 +
i6=j
then
X
αj βi = 2αj − 2 ≥ 2(1/βj − 1) ≥ 2(1 − βj ),
i6=j
where the last step follows by βj + 1/βj ≥ 2. Case II, if 1 ≤ αj < 1/βj , then
X
X
|1 − αj | + |1 − αj βj | +
αj βi = αj − 1 + 1 − αj βj +
αj βi = 2αj (1 − βj ) ≥ 2(1 − βj ).
i6=j
i6=j
Case III, if αj < 1, then
|1 − αj | + |1 − αj βj | +
X
αj βi = 1 − αj + 1 − αj βj +
i6=j
X
αj βi = 2(1 − αj βj ) ≥ 2(1 − βj ).
i6=j
Putting it all together, we have
f (α, β) ≥
d−1
X
2(1 − βj ) = 2(d − 1) − 2
j=1
d−1
X
j=1
74
βj = 2(d − 2).
To
handle the case where βi can be negative Without loss of generality, we can assume that
Pd−1
i=1 βi = 1. Notice that we can also assume βi 6= 0, otherwise it means we do not choose that
row. We split all βi into two disjoint sets S and S. For any i ∈ S, βi > 0 and for any
P i ∈ S, βi < 0.
As a first step, we discuss the case when all the j are in set S. Case I, if 1 − αj d−1
i=1 βi < 0 and
1
1
1 − αj βj < 0, then it means αj ≥ max( Pd−1 , β ). The cost of that row is,
i=1
= αj
d−1
X
βi
j
βi − 1 + αj βj − 1 +
i=1
X
|αj βi |
i6=j
= 2αj βj + 2αj
X
βi − 2.
i∈S\j
P
If βj ≥ 1, then αj ≥ 1. The cost is at least 2 i∈S\j βi . If βj < 1, then αj ≥ 1/βj , and the cost is
P
at least 2 β1j i∈S\j βi . If there are C such j in Case I, then the total cost of Case I is at least 0 if
C = 1, and 2(C − 1) P
if C ≥ 2.
Case II, if 1 − αj d−1
i=1 βi < 0 and 1 − αj βj > 0, then it means 1 < αj < 1/βj . The cost of that
row is,
= αj
d−1
X
βi − 1 + 1 − αj βj + |αj |
i=1
= 2αj
X
|βj |
i6=j
X
βi
i∈S\j
≥2
X
βi .
i∈S\j
Similarly to before, if there are C such j in Case II, then the total cost of Case II is at least 0 if
C = 1, and 2(C − 1) ifPC ≥ 2.
1
Case III, if 1 − αj d−1
i=1 βi > 0 and 1 − αj βj < 0, then it means 1/βj < αj < Pd−1 . The cost
i=1
of the row is,
= 1 − αj
d−1
X
βi + αj βj − 1 + |αj |
i=1
= 2|αj |
X
X
βi
|βj |
i6=j
|βi |
i∈S
X
= 2|αj |(
|βi | − 1)
i∈S
1 X
≥2
βi .
βj
i∈S\j
If there are C such j in Case III, then the total cost of Case III is at least 0 if C = 1, and 2(C ·(C −1))
if C ≥ 2.
P
1
Case IV, if 1 − αj d−1
i=1 βi > 0 and 1 − αj βj > 0, then it means αj ≤ min(1/βj , Pd−1 ). The
i=1
βi
cost for the case αj < 0 is larger than the cost for case α > 0. Thus we can ignore the case αj < 0.
75
The cost of the row is,
= 1 − αj
d−1
X
βi + 1 − αj βj + |αj |
i=1
X
|βj |
i6=j
X
= 2 − 2αj βj + 2αj
|βi |.
i∈S
P
If βj < i∈S |βi |, wePknow that the cost is at least 2. Otherwise, using αP
j ≤ 1, we have the cost
is at least 2 − 2βj + i∈S |βi |. Let T denote the set of those j with βj ≥ i∈S |βi |. If |T | = 1, we
know that cost is at least 0. If |T | ≥ 2, then the cost is at least
X
X
=
(2 − 2βj + 2
|βi |)
j∈T
i∈S
≥ 2|T | − 2
X
βj + 2|T |
j∈S
≥ 2|T | − 2 + 2(|T | − 1)
X
|βi |
i∈S
X
|βi |
i∈S
≥ 2(C − 1).
Now we discuss the case where j ∈ S, which means βj < 0.
P
P
Case V if 1 − αj d−1
0, then it means αj > 1/ d−1
i=1 βi and αj < 1/βj .
i=1 βi < 0 and 1 − αj βj < P
d−1
Notice that this case will never happen, because i=1 βi = 1 and αj < 0.
Pd−1
P
Case VI if 1 − αj d−1
i=1 βi , 1/βj ).
i=1 βi < 0 and 1 − αj βj < 0, then it means αj ≥ max(1/
Because of βj < 0, then αj ≥ 1. The cost of that is,
= αj
d−1
X
βi − 1 + 1 − αj βj + |αj |
i=1
= 2αj
X
X
|βi |
i6=j
|βi |
i∈S
≥ 2.
by αj ≥ 1 and
X
|βi | ≥ 1.
i∈S
P
Pd−1
Case VII if 1 − αj d−1
βi > 0 and 1 − αj βj < 0, then it means αj < min(1/βj , 1/ i=1
βi ).
i=1
Pd−1
Because βj < 0 and i=1 βi = 1, thus αj < 1/βj . The cost of that row is,
= 1 − αj
d−1
X
βi + αj βj − 1 + |αj |
i=1
= 2|αj |
X
X
|βi |
i6=j
|βi |
i∈S\j
≥2
1 X
|βi |.
|βj |
i∈S\j
If there are C such j in Case VII, then the total cost of Case VII is at least 0 if C = 1, and
2(C · (C − 1)) if C ≥ 2.
76
Case VIII if 1 − αj
cost of that row,
Pd−1
i=1
βi > 0 and 1 − αj βj > 0, then it means 1/βj < αj < 1/
= 1 − αj
d−1
X
βi + 1 − αj βj + |αj |
i=1
X
Pd−1
i=1
βi . The
|βi |
i6=j
= 2 − 2αj βj + 2|αj |
X
|βi |.
i∈S\j
If αj > 0, then the cost is always at least 2. If αj < 0, the cost is at least,
X
2 − 2|αj βj | + 2|αj |
|βi |.
i∈S\j
If
P
i∈S\j
|βi | > |βj |, we also have cost at least 2. Otherwise, we have
2 − 2|αj |(|βj | −
X
i∈S\j
(
P
2 − 2|βj | + 2 i∈S\j |βi |
|βi |) =
P
2 |β1j | i∈S\j |βi |
if 1/|βj | ≤ 1,
if 1/|βj | > 1.
Let C denote the number of such j with 1/|βj | ≤ 1. If C = 1, the cost is at least 0. If C ≥ 2, the
cost is at least 2C. Let C 0 denote the number of such j with 1/|βj | > 1. If C 0 = 1, the cost is at
least 0, if C 0 ≥ 2, the cost is at least 2(C 0 (C 0 − 1)). Overall, putting all the eight cases together, we
complete the proof.
G.3
Hard instance for oblivious subspace embedding and more row subset selection
The goal in this section is to prove Theorem G.31. By applying Yao’s minmax principle, it suffices
to prove Theorem G.27.
G.3.1
Definitions
We first give the definition of total variation distance and Kullback-Leibler divergence.
Definition G.9. [LPW09, Ver14] The total variation distance between two probability measures P
and Q on the measurable space (X , F) is defined as,
DTV (P, Q) = sup |P (A) − Q(A)|.
A∈F
The Kullback-Leibler( KL) divergence of P and Q is defined as,
Z
dP
dP
DKL (P ||Q) = E log
=
log
dP.
P
dQ
dQ
X
Lemma G.10. [Pin60, Tsy09, CK11] Pinsker’s inequality states that, if P and Q are two probability
distributions on a measurable (X , F), then
r
1
DTV (P, Q) ≤
DKL (P ||Q).
2
77
Lemma G.11. For any Gaussian random variable x ∼ N (0, σ 2 ), we have, for any a > 0
Pr[|x| < a] .
a
.
σ
Proof.
a
Pr[|x| < a] = erf( √ )
σ 2
−4a2
1
≤ (1 − e− 2σ2 π ) 2
by 1 − exp(−4x2 /π) ≥ erf(x)2
4a2 1
≤ ( 2 )2
2σ π
a
. .
σ
by 1 − e−x ≤ x
Lemma G.12. Let V ∈ Rn×m be a matrix with orthonormal columns. Let H0 be a distribution over
V > A, where A ∈ Rn×k is a random matrix with each entry i.i.d. Gaussian N (0, 1). Denote H as a
distribution over H ∈ Rm×k , where each entry of H is drawn from i.i.d. Gaussian N (0, 1). Then,
H and H0 are the same distribution.
Proof. It is clear that each entry of V > A is a random Gaussian variable from N (0, 1), so our goal is
to prove the entries of V > A are fully independent. Since the the j th column of V > A only depends
on Aj , the variables from different columns are fully independent. Now, we look at one column of
x = V > Aj . The density function is
1
1 > >
1
1 >
p
p
f (x) =
exp − x V V x =
exp − x x ,
2
2
(2π)m
(2π)m |V > V |
which is exactly the density function of N (0, Im ). Thus x ∈ N (0, Im ). Therefore, all the entries of
V > A are fully independent.
Lemma G.13 (Matrix Determinant Lemma). Suppose A ∈ Rn×n is an invertible matrix, and
u, v ∈ Rn . Then,
|A + uv > | = (1 + v > A−1 u)|A|.
Lemma G.14 (KL divergence between two multivariate Gaussians [PP+ 08]2 ). Given two d-dimensional
multivariate Gaussian distribution N (µ1 , Σ1 ) and N (µ2 , Σ2 ), then
1
|Σ2 |
−1
> −1
DKL (N (µ1 ||Σ1 ), N (µ2 , Σ2 )) =
log
− d + tr(Σ2 Σ1 ) + (µ2 − µ1 ) Σ2 (µ2 − µ1 ) .
2
|Σ1 |
Lemma G.15 (Sherman-Morrison formula). Suppose A ∈ Rn×n is an invertible matrix and u, v ∈
Rn . Suppose furthermore that 1 + v > A−1 u 6= 0. Then the Sherman-Morrison formula states that
(A + uv > )−1 = A−1 −
2
A−1 uv > A−1
.
1 + v > A−1 u
(27)
http://stats.stackexchange.com/questions/60680/kl-divergence-between-two-multivariate-gaussians
78
G.3.2
Main results
Lemma G.16. Let V ∈ Rn×m be a matrix with orthonormal columns, and let A ∈ Rn×k be a
random matrix with each entry drawn from i.i.d. Gaussian N (0, 1). We denote the distribution Di
over Di ∈ R(m+1)×k where
>
V A
Di =
.
Ai
If k(V > )i k22 < 12 , then
DTV (Di , G) ≤ O(kkV i k2 ) + 2−Θ(k) ,
where G is a distribution over G ∈ R(m+1)×k , where each entry of G is drawn from the i.i.d. Gaussian
N (0, 1).
Proof. Let H be a distribution over H ∈ Rm×k , where each entry of H is an i.i.d. Gaussian N (0, 1).
Let H0 be a distribution over V > A. According to Lemma G.12, H and H0 are the same distribution.
We define matrix V ¬i ∈ Rn×m as
>
V ¬i = (V 1 )> (V 2 )> · · · (V i−1 )> 0 (V i+1 )> · · · (V n )> ,
where V i denotes the ith row of V ∈ Rn×m , ∀i ∈ [n].
Let G be a distribution over G ∈ R(m+1)×k , where each entry of G is an i.i.d. Gaussian N (0, 1).
Let Pi be a distribution over Pi ∈ R(m+1)×k , where
¬i >
(V ) A
Pi =
.
Ai
bi be a distribution over Pbi ∈ Rm×k , where Pbi = (V ¬i )> A. Then we have:
Let P
bi , H0 ) = DTV (P
bi , H).
DTV (Pi , G) = DTV (P
(28)
The first equality is because (V ¬i )> A is independent from Ai . The second equality follows the
Lemma G.12.
Claim G.17. If kV i k22 ≤ 21 ,
bi ||H) = − k log(1 − kV i k2 ) + kV i k2 ≤ kkV i k2 .
DKL (P
2
2
2
2
bi , H ∼ H. Notice that different columns of Pbi are i.i.d, and all entries of H are
Proof. Let Pbi ∼ P
fully independent. We can look at one column of Pbi and H. Since Pbi = (V ¬i )> A, it is easy to see
that its column is drawn from N (0, Σ1 ) where Σ1 = Im − (V i )> V i . Since H is fully independent,
each column of H is drawn from N (0, Σ2 ) where Σ2 = Im . Let p(x) be the pdf of the column of Pbi ,
79
and let q(x) be the pdf of the column of H. We have the following calculation [PP+ 08]3 ,
k
|Σ2 |
−1
b
DKL (Pi ||H) =
log
− m + tr(Σ2 Σ1 )
2
|Σ1 |
|Im |
k
i > i
log
−
m
+
tr(I
−
(V
)
V
)
=
m
2
|Im − (V i )> V i |
k
− log |Im − (V i )> V i | − m + m − kV i k22
=
2
k
=
− log(1 − kV i k22 ) − m + m − kV i k22
2
k
= − log(1 − kV i k22 ) + kV i k22
2
k
≤ − · 2kV i k22
2
= kkV i k22 .
The first equality is due to Lemma G.14. The sixth equality follows by Σ2 = Im and Σ1 = Im −
(V i )> V i . The eighth equality follows by Lemma G.13. The first inequality follows by log(1−x)+x ≥
−2x, when 0 < x < 1/2.
According to Lemma G.10, we have
r
bi , H) ≤
DTV (P
√
1
bi ||H) ≤ kkV i k2 .
DKL (P
2
(29)
Now, we want to argue that DTV (Di , Pi ) is small, where Di is a distribution over Di ∈ R(m+1)×k
that
>
V A
Di =
.
Ai
bi (x) be a distribution over D
b i (x) ∈ Rm×k , where D
b i (x) = (V ¬i )> A+(V i )> x.
For a fixed x ∈ Rk , let D
Let p(x) be the pdf of (Ai )> , then
Z
bi (x), P
bi p(x)dx.
DTV (Di , Pi ) = DTV D
(30)
b i (x) and the j th column of Pbi . The distribution of the
Now we look at the j th column of D
i
>
previous one is over N ((V ) xj , Σ2 ), and the latter distribution as we said before is N (0, Σ2 ),
3
http://stats.stackexchange.com/questions/60680/kl-divergence-between-two-multivariate-gaussians
80
where Σ2 = Im − (V i )> V i . Now we can argue that the KL divergence between them is bounded:
DKL (N ((V i )> xj , Σ2 )||N (0, Σ2 ))
|Σ2 |
1
−1
2 i −1
i >
log
=
− m + tr(Σ2 Σ2 ) + xj V Σ2 (V )
2
|Σ2 |
1
i >
= −m + tr(Im ) + x2j V i Σ−1
(V
)
2
2
1
i >
= x2j V i Σ−1
2 (V )
2
1
= x2j V i (Im − (V i )> V i )−1 (V i )>
2
1 2 i
(V i )> V i
= xj V Im +
(V i )>
2
1 − V i (V i )>
1 2
kV i k22
i 2
= xj kV k2 +
2
1 − kV i k22
1
kV i k22
= x2j
.
2 1 − kV i k22
The first equality is due to Lemma G.14. The fourth equality follows by Σ2 = Im − (V i )> V i . The
fifth equality follows by Lemma G.15.
By summing the KL divergence on all the columns up,
i 2
bi (x)||P
bi ) = 1 kxk2 kV k2 .
DKL (D
2
2
1 − kV i k22
Applying Lemma G.10 again, we get
r
kxk2 kV i k2
bi (x)||P
bi ) = p
bi (x), P
bi ) ≤ 1 DKL (D
DTV (D
.
2
2 1 − kV i k22
Plugging it into Equation (30), we get
Z
bi (x), P
bi p(x)dx
DTV (Di , Pi ) = DTV D
Z
Z
b
b
bi (x), P
bi p(x)dx
=
DTV Di (x), Pi p(x)dx +
DTV D
kxk ≤10k
kxk >10k
Z 2
Z 2
bi (x), P
bi p(x)dx +
≤
DTV D
p(x)dx
kxk2 ≤10k
kxk2 >10k
ik
kxk kV 2
p 2
p(x)dx +
2
kxk2 ≤10k 2 1 − kV i k2
Z
10kkV i k2
≤ p
+
p(x)dx
2 1 − kV i k22
kxk2 >10k
Z
≤
Z
p(x)dx
kxk2 >10k
10kkV i k2
≤ p
+ 2−Θ(k) .
2 1 − kV i k22
The first inequality just follows from the fact that total variation distance is smaller than 1. The
second inequality is what we plugged in. The last inequality follows from the fact that x ∼ N (0, Ik )
and from the tail bounds of a Gaussian.
81
Together with Equation (28) and Equation (29), we can get
DTV (Di , G) ≤ DTV (Di , Pi ) + DTV (G, Pi )
bi , H)
= DTV (Di , Pi ) + DTV (P
√
10kkV i k2
−Θ(k)
kkV i k2
+
2
+
≤ p
2 1 − kV i k22
√
≤ 10kkV i k2 + 2−Θ(k) + kkV i k2 .
The last inequality follows by kV i k22 ≤ 21 . Then, we have completed the proof.
Lemma G.18. A ∈ Rr×k (r ≥ k) is a random matrix for which each entry is i.i.d. N (0, 1). With
√
probability at least 1 − e−Θ(r) , the maximum singular value kAk2 is at most O( r).
Proof. Since A ∈ Rk×r is a random matrix with each entry i.i.d. N (0, 1), this follows by standard
arguments (Proposition 2.4 in [RV10]). Since r ≥ k, with probability at least 1 − e−Θ(r) , kAk2 is at
√
most O( r).
Definition G.19. Let V ∈ Rn×r be a matrix with orthonormal columns, and let each entry of
b
A ∈ Rk×r be a random variable drawn from an i.i.d. Gaussian N (0, 1). Define event E(A,
V, β, γ) to
n
γ
β
be: ∀y ∈ R , kyk1 ≤ O(k ) and each coordinate of y has absolute value at most 1/k , and also AV > y
√
has at most O(k/ log k) coordinates with absolute value at least Ω(1/ log k), and kAk2 ≤ O( r).
Lemma G.20. For any k ≥ 1, and any constants c2 ≥ c1 ≥ 1, let k ≤ r = O(k c1 ), r ≤ n = O(k c2 ),
let V ∈ Rn×r be a matrix with orthonormal columns, and let each entry of A ∈ Rk×r be a random
variable drawn from i.i.d. Gaussian N (0, 1). Furthermore, if β and γ are two constants which
satisfy β > γ > 0 and β + γ < 1,
b
Pr E(A, V, β, γ) ≥ 1 − 2−Θ(k) .
√
Proof. Due to Lemma G.18, with probability at least 1 − 2−Θ(r) , kAk2 ≤ O( r), so we can restrict
√
attention to kAk2 ≤ O( r) in the following proof.
Take any y P
∈ Rn which has each non-zero coordinate with absolute value at most 1/k β , and
j
j
−j−1 , 2−j ),
write it as y = +∞
j=j0 y , where the coordinates in y have absolute value in the range [2
and the supports of different y j are disjoint. Since each coordinate of y is at most 1/k β , 2−j0 = 1/k β .
Since kyk1 ≤ O(k γ ), the support size of y j ∈ Rn is at most sj ≤ O(2j+1 k γ ), so it follows that
q
√
j
ky k2 ≤ sj · 2−2j ≤ 2−j+1 k γ .
p
√
We also know ky j k2 ≥ sj · 2−2j−2 . Then we can conclude that y j has 2-norm Θ(2−j sj ). Now
we state an ε-net for all the possible y j : let ε = O(1/(nrk 3 )) = O(1/(k c1 +c2 +3 )). Let Nj ⊂ Rn be
the following:
Nj = p ∈ Rn | ∃q ∈ Zn , s.t. p = εq, kpk1 ≤ O(k γ ), ∀i ∈ [n], either |pi | ∈ [2−j−1 , 2−j ) or pi = 0 .
Obviously, for any y j , there exists p ∈ Nj such that ky j − pk∞ ≤ ε = O(1/(k c1 +c2 +3 )), since
n ≤ O(k c2 ), ky j − pk2 ≤ ky j − pk1 ≤ nky j − pk∞ ≤ O(1/k c1 +3 ). Now let us consider the size of Nj .
If p ∈ Nj , the choice of one coordinate of p is at most 2/ε + 1. And since kpk1 ≤ O(k γ ) and each
82
coordinate of p has absolute value at least 2−j−1 , the number of supports of p is at most O(2j+1 k γ ).
Therefore,
j+1 k γ )
|Nj | ≤ (n + 1)O(2
· (2/ε + 1)O(2
j+1 k γ )
j+1 k γ (log n+log(2/ε+1))
≤ 2O(2
j+1 k γ
≤ 2O(2
log k)
.
The last inequality follows from n ≤ O(k c2 ), r ≤ O(k c1 ), ε = O(1/(rnk 3 )).
We define an event E(y j ) to be: AV > y j has k/ log2 k coordinates which are each at least 2/ log2 k
in absolute value. Now, we want to show,
Claim G.21. For any j ≥ j0 , for a fixed y j ∈ Rn ,
6
6
j−1 1−γ
j 2
j
Pr E(y ) happens ≤ 2k e−Θ(k/(ky k2 log k)) ≤ e−Θ(2 k / log k) .
Proof. We let p be the probability that the absolute value of a single coordinate of AV > y j is at least
1/ log2 k. Notice that each coordinate of AV > y j is i.i.d. Gaussian N (0, kV > y j k22 ) and because for
any Gaussian random variable g, Pr[|g| ≥ t] ≤ exp(−Θ(t2 /σ 2 )), then p ≤ exp(−1/(kV > y j k22 log4 k)),
by plugging σ 2 = kV > y j k22 and t = 1/ log2 k. So the probability AV > y j has k/ log2 k coordinates
which are each at least 1/ log2 k in absolute value is,
k
X
i
k−i
p (1 − p)
i=k/ log2 k
k
≤
i
k
X
i
k
k/2
p
i=k/ log2 k
2 /(σ 2
≤ e−Θ(kt
≤ e−Θ(k/(kV
≤ e
2
≤ 2pk/ log k 2k
log2 k)) k
by p ≤ exp(−Θ(t2 /σ 2 ))
2
> y j k2
2
−Θ(k/(ky j k22
−2j s
≤ e−Θ(k/(2
log6 k)) k
2
6
by kV > y j k22 ≤ kV k22 ky j k22 ≤ ky j k22
log k)) k
j
2
log6 k)) k
−j+1 k γ
log6 k)) k
−j+1 k γ
log6 k))
≤ e−Θ(k/(2
= e−Θ(k/(2
by sj ≤ O(2j+1 k γ )
2
j−1 k 1−γ / log6
≤ e−Θ(2
by ky j k22 ≤ 2−2j sj
2
k)
.
The first equality follows from 2−j+1 k γ log6 k ≤ 2−j0 +1 k γ log6 k ≤ 2k γ−β log6 k = o(1).
P
j
For j1 = d100(c1 + c2 + 1) log ke = Θ(log k), consider j ∈ [j1 , ∞). We have k ∞
j=j1 y k2 is at
√
most Θ(2−j1 n) ≤ 1/k 100(1+c1 ) , and so
kAV
>
∞
X
j=j1
j
y k1 ≤
√
kkAV
>
∞
X
j
y k2 ≤
√
kkAk2 kV k2 k
j=j1
∞
X
y j k2 ≤
√
k·
√
r · 1/k 100(1+c1 ) ≤ 1/k 50 .
j=j1
The last inequality follows from r = O(k c1 ). So the contribution of y j to kAV > y j k1 for all j ≥ j1
is at most 1/k 50 . Thus, if we only consider those j which contribute, i.e., j0 ≤ j ≤ j1 , we have
83
O(log k) values of j. Then we can only construct O(log k) nets Nj0 , Nj0 +1 , · · · , Nj1 . Since the size
j+1 γ
of net Nj is 2Θ(2 k log k) , by combining Claim G.21 and taking a union bound, we have
j1
j1
[
X
6
j+1 γ
j−1 1−γ
Pr ∃y j ∈
Nj , E(y j ) happens ≤
2Θ(2 k log k) · e−Θ(2 k / log k)
j=j0
j=j0
j0 −1 k 1−γ / log6
≤ e−Θ(2
≤ e−Θ(k
1+β−γ / log6
k)
k)
≤ 2−Θ(k) .
The second inequality follows since k γ = o(k 1−γ ). The third inequality follows since 2j0 ≥ k β . The
fourth inequality follows since 1 + β − γ > 1.
Then, ∀j0 ≤ j ≤ j1 , ∀e
y j 6∈ Nj , there exists a vector ybj in Nj , such that kb
y j − yej k2 ≤ 1/k 3+c1 .
We can upper bound the `∞ norm of AV > ybj − AV > yej in the following sense,
kAV > ybj − AV > yej k∞ ≤ kAV > ybj − AV > yej k2
j
by k · k∞ ≤ k · k2
j
≤ kAk2 · kb
y − ye k2
√
≤ r/k 3+c1
by Claim G.18 and kb
y j − yej k2 ≤ 1/k 3+c1
= 1/k 2
by r ≤ O(k c1 ).
We let Y = {y ∈ Rn | kyk1 ≤ O(k γ ) and each coordinate of y ≤ 1/k β }. Since 1/k 2 < 1/ log2 k we
can conclude that,
j
Pr ∃y ∈ Y, j ≥ j0 , E(y ) happens ≤ 2−Θ(k) .
(31)
P
Recalling y = j y j , by Equation (31), for any y, with probability at most 2−Θ(k) , there are
P1
at most O(log k) · k/ log2 k ≤ O(k/ log k) coordinates for which AV > jj=j
y j (the same statement
0
P
j
>
also holds for AV > ∞
j=j0 y = AV y since we argued that there is negligible “contribution” for
those j > j1 ) is at least O(log k)/ log2 k = O(1/ log k) on that coordinate. Summarizing,
b
Pr E(A, V, β, γ) ≥ 1 − 2−Θ(k) .
Lemma G.22. For any t, k ≥ 1, and any constants c2 ≥ c1 ≥ 1, let k ≤ r = O(k c1 ), r ≤ n = O(k c2 ),
let V ∈ Rn×r be a matrix with orthonormal columns, and let each entry of A ∈ Rk×r , v1 , v2 , · · · , vt ∈
Rk be an i.i.d. Gaussian N (0, 1) random variable. For a constant α ∈ (0, 0.5) which can be
b
arbitrarily small, if E(A,
V, 0.5+α/2, 0.5−α) happens, then with probability at least 1−2−Θ(tk) , there
are at least dt/10e such j ∈ [t] that ∀x ∈ Rr either kAx − vj k1 ≥ Ω(k 0.5−α ) or kV xk1 ≥ Ω(k 0.5−α )
holds.
Proof. For convenience, we define γ = 0.5 − α which can be an arbitrary constant in (0, 0.5). We
let constant β = 0.5 + α/2. Then we have β + γ < 1 and β > γ. Let v ∈ Rk be a random vector
b
with each entry drawn from i.i.d. Gaussian N (0, 1). Suppose E(A,
V, β, γ) happens.
84
For any x ∈ Rr , we can find a y ∈ Rn such that y = V x. Since V ∈ Rn×r has orthonormal
columns, x = V > V x = V > y. Then our goal is to argue that with high probability if kAV > y −vk1 ≤
O(k γ ), then kyk1 > Ω(k γ ). Take any y ∈ Rn which can be expressed as V x, and decompose it as
y = y0 +y1 , where y0 ∈ Rn and y1 ∈ Rn have disjoint supports, y0 has each coordinate with absolute
value greater than 1/k β , and y1 has each coordinate with absolute value at most 1/k β .
Now we create an ε-net for y0 : let ε = O(1/(rnk 3 )) = O(1/k c1 +c2 +3 ), we denote N ⊂ Rn as
follows:
n
o
N = p ∈ Rn | ∃q ∈ Zn , s.t. p = εq, kpk1 ≤ O(k γ ), ∀i ∈ [n], either |pi | > 1/k β or pi = 0 .
Obviously, for any y0 , there exists p ∈ N such that ky0 − pk∞ ≤ ε = O(1/(k c1 +c2 +3 )), since
n ≤ O(k c2 ), ky0 − pk2 ≤ ky0 − pk1 ≤ nky0 − pk∞ ≤ O(1/k c1 +3 ). Now let us consider the size of N .
If p ∈ N , the number of choices of one coordinate of p is at most O(k γ /ε). And since kpk1 ≤ O(k γ )
and each coordinate of p has absolute value at least 1/k β , the number of supports of p is at most
O(k γ+β ). Therefore,
|N | ≤ (n + 1)O(k
≤ 2O(k
γ+β
γ+β )
log k)
· O(k γ /ε)O(k
γ+β )
.
The last inequality follows from n ≤ O(k c2 ), r ≤ O(k c1 ), ε = O(1/(rnk 3 )).
For y0 ∈ Rn , we define event E1 (y0 ) as: ∃ valid y1 ∈ Rn , kAV > y−vk1 ≤ O(k γ ), where y = y0 +y1
is the decomposition of a possible y ∈ Rn . Here y1 is valid means that there exists y ∈ Rn such that
kyk1 ≤ O(k γ ) and the decomposition of y is y = y0 + y1 . We define event E2 (y0 ) as: ∃ valid y1 , the
absolute value of AV > y − v is at most O(1/k γ ) for at least k − O(k 2γ ) coordinates i in [k] . We
define event E3 (y0 ) as: at least k − O(k 2γ ) − O(k/ log k) coordinates of Ay0 − v have absolute value
at most O(1/ log k).
Claim G.23. For y0 ∈ Rn ,
Pr[E3 (y0 ) happens ] ≥ Pr[E2 (y0 ) happens ] ≥ Pr[E1 (y0 ) happens ].
Proof. If E1 (y0 ) happens, then there exists a valid y1 ∈ Rn such that y = y0 +y1 and kAV > y −vk1 ≤
O(k γ ). For this y, there are at least k − O(1/k 2γ ) coordinates of AV > y − v with absolute value at
most O(1/k γ ). Otherwise, kAV > y − vk1 > Ω(k γ ). Thus, E1 (y0 ) implies E2 (y0 ).
Now we want to show E2 (y0 ) implies E3 (y0 ). We suppose E2 (y0 ) happens. Then there is a valid
y1 such that there are at least k − O(1/k 2γ ) coordinates of AV > y − v with absolute value at most
b
O(1/k γ ), where y = y0 + y1 . Recall that the event E(A,
V, β, γ) happens, for any valid y1 there are
>
at most O(k/ log k) coordinates of AV y1 are at least Ω(1/ log k). Therefore,
AV > y0 − v =
AV > y − v
| {z }
≥k−O(k2γ ) coordinates
each ≤O(1/kγ )
−
AV > y
| {z }1
.
≤O(k/ log k) coordinates
each≥Ω(1/ log k)
Therefore, at least k − O(k 2γ ) − O(k/ log k) coordinates of Ay0 − v in absolute value is at most
O(1/ log k) + O(1/k γ ) = O(1/ log k)
Claim G.24. Define event F1 to be the situation for which there exists 1/2 of the coordinates of
v ∈ Rk that are at least 1/100. The probability this event F1 holds is at least 1 − 2−Θ(k) .
85
Proof. Note that F 1 means there exist k/2 of the coordinates of v ∈ Rk which are at most 1/100.
Using Lemma G.11, for each single coordinate the probability it is smaller than 1/100 is 1/200.
Then the probability more than half the coordinates are no more than 1/100 is
k
X
i=k/2
pi (1 − p)k−i
k
X
k
k
≤
pi
≤ 2pk/2 2k ≤ (1/200)Θ(k) = 2−Θ(k) .
i
k/2
i=k/2
Conditioned on F1 happening, if E3 (y0 ) happens, then due to the pigeonhole principle, there are
at least k2 − O(k 2γ ) − O(k/ log k) coordinates of AV > y0 − v that are at most O(1/ log k) and the
corresponding coordinate of v is larger than 1/100. Now, let us look at the probability on a single
coordinate of AV > y0 − v.
Claim G.25. If the ith coordinate of v ∈ Rk is at least 1/100. Then Pr[|(AV > y0 )i − (v)i | ≤
O(1/ log k)] ≤ O(1/ log k).
Proof. Since |(v)i | > 1/100 and v is independent from A ∈ Rk×r , for 0 < η < 1/100, Pr[|(AV > y0 )i −
(v)i | ≤ η] is always upper bounded by Pr[(Ax0 )i ∈ [1/100 − η, 1/100 + η]. Thus, it suffices to prove
an upper bound for Pr[(AV > x0 )i ∈ [1/100 − η, 1/100 + η]. Let f (x) be the pdf of N (0, σ 2 ), where
σ 2 = kV > y0 k22 , V ∈ Rn×r , y0 ∈ Rn . Then
Pr[(AV > y0 )i ∈ [1/100 − η, 1/100 + η]]
Z 1/100+η
=
f (x)dx
1/100−η
Z
1/100+η
≤ f (1/200)
dx
1/100−η
≤O(η).
where the last step follows since f (1/200) ≤ 200. We set η = O(1/ log k), then we get the statement.
Claim G.26. Conditioned on F1 , for a fixed y0 ∈ Rn , with probability at most 2−Θ(k) , there are
k
at least 10
coordinates of AV > y0 − v ∈ Rk which are at most O(1/ log k) and the corresponding
coordinate of v is larger than 1/100.
Proof. We look at the coordinate i ∈ [k] which has |(v)i | > 1/100. The probability k(AV > y0 )i −
(v)i k ≤ O(1/ log k) is at most O(1/ log k). Due to the independence between different coordinates of
v, since there are at least k/2 coordinates of v satisfying that they have absolute value greater than
1/100, with probability at most 2−Θ(k) , there are at least 15 · k2 = k/10 coordinates of AV > y0 − v
which are at most O(1/ log k).
Because E3 (y0 ) implies the event described in the above claim when conditioning on F1 , for a
o(1)
fixed y0 , the probability that E3 (y0 ) holds is at most 2−Θ(k) . Since γ + β < 1, the |N | ≤ 2k , we
can take a union bound over the N :
Pr[∃y0 ∈ N , E1 (y0 ) happens ] ≤ 2−Θ(k) .
86
It means that with probability
at least 1 − 2−Θ(k) , for any y = y0 + y1 with y0 ∈ N , kAV > y − vk1 >
P
j
Ω(k γ ). Let
P yej = ye0 + ye , where ye0 6∈ N . We can find yb0 ∈ N which is the closest to ye0 . Denote
yb = yb + ye . Then,
kAV > yb − vk1 ≤ kAV > ye − vk1 + kAV > ye − AV > ybk1
√
≤ kAV > ye − vk1 + kkAV > ye − AV > ybk2
√
≤ kAV > ye − vk1 + kkAk2 ke
y − ybk2
√
√
≤ kAV > ye − vk1 + k · r · ke
y0 − yb0 k2
√
√
≤ kAV > ye − vk1 + k · r · ε
√ √
= kAV > ye − vk1 + k · r · 1/k c1 +3
by triangle inequality
√
by k · k1 ≤ dimk · k2
= kAV > ye − vk1 + 1/k.
by kAk2 ≤
√
r
by ke
y0 − yb0 k2 ≤ ε
by ε0 = 1/k c1 +3
by r = O(k c1 )
and so if kAV > ye − vk1 is at most O(k γ ) then kAb
y − vk1 is at most O(k γ ).
For j ∈ [t], we now use notation E4 (vj ) to denote the event: ∃y0 ∈ Rn with ky0 k1 ≤ O(k γ ) and
each non-zero coordinate of y0 is greater than 1/k β , at least k − O(k 2γ ) − O(k/ log k) coordinates of
AV > y0 − vj in absolute value are at most O(1/ log k). Based on the previous argument, conditioned
b
on E(A,
V, β, γ),
Pr[E4 (vj )] ≤ 2−Θ(k) .
b
Also notice that, conditioned on E(A,
V, β, γ), ∀j ∈ [t], E4 (vj ) are independent. Thus, conditioned
b
on E(A, V, β, γ), due to the Chernoff bound, the probability that there are dt/10e such j that E4 (vj )
happens is at most 2−Θ(kt) . We define E5 (vj ) to be the event: ∃y ∈ Rn , kAV > y − vj k ≤ O(k γ )
b
with kyk1 ≤ O(k γ ). Similar to the proof of Claim G.23, conditioned on E(A,
V, β, γ), E5 (vj ) implies
b
E4 (vj ). Thus, conditioned on E(A, V, β, γ), the probability that there are dt/10e such j that E5 (vj )
happens is at most 2−Θ(tk) . Then, we complete the proof.
Theorem G.27. For any k ≥ 1, and any constants c1 , c2 which satisfy c2 − 2 > c1 > 1, let
r = Θ(k c1 ), n = Θ(k c2 ), and let A(k, n) denote a distribution over n × (k + n) matrices where each
entry of the first n × k matrix is i.i.d. Gaussian N (0, 1) and the next n × n matrix is an identity
b ∼ A(k, n), with probability at least
matrix. For any fixed r × n matrix S and a random matrix A
c −c
1+ 1 2 2
−Θ(k)
1 − O(k
)−2
, there is no algorithm that is able to output a matrix B ∈ Rn×r such that
b − Ak
b 1 ≤ O(k 0.5−ε )
kBS A
min
rank −k A0
b 1,
kA0 − Ak
where ε > 0 is a constant which can be arbitrarily small.
Proof. For convenience, we let γ = 0.5 − ε be a constant which can be arbitrarily close to 0.5. Since
b is an identity matrix, we can fit the first k columns of A,
b so we have
the last n columns of A
min
rank −k A0
b 1 ≤ n.
kA0 − Ak
Now, we want to argue that, for a fixed S, with high probability, for any rank-k n × r matrix B,
the cost
b − Ak
b 1 ≥ Ω(n · k γ ).
kBS A
87
Thus, the approximation gap will be at least Ω(k γ ).
We denote the SVD of S = US ΣS VS> where US ∈ Rr×r , ΣS ∈ Rr×r , VS ∈ Rn×r . Then, we can
rewrite
b − Ak
b 1 = kBUS ΣS V > A
b − Ak
b 1
kBS A
S
n
X
b −A
bl k1
=
k(BUS ΣS )l (V > A)
S
l=1
≥
X
b −A
bl k1 .
k(BUS ΣS )l (VS> A)
(32)
l:kVSl k22 ≤2r/n
The first equality follows from the SVD of S. The second equality follows from the fact that the
`1 -norm of a matrix is the sum of `1 -norms of rows. The third inequality follows since we just look
at the cost on a part of the rows.
We use βl to denote (BUS ΣS )l . We look at a fixed row l, then the cost on this row is:
b −A
bl k1
kβl (VS> A)
b [1:k] − (A
bl )[1:k] k1 + kβl (V > A)
b [k+1:k+n] − (A
bl )[k+1:n+k] k1
=kβl (V > A)
S
b [1:k]
≥kβl (VS> A)
b [1:k]
≥kβl (VS> A)
bl
− (A )[1:k] k1 +
bl )[1:k] k1 +
− (A
S
b [k+1:k+n] k1
kβl (VS> A)
b [k+1:k+n] k1
kβl (VS> A)
bl )[k+1:n+k] k1
− k(A
−1
b [1:k] − (A
bl )[1:k] k1 + kβl V > k1 − 1.
≥kβl (VS> A)
S
(33)
b [1:k] denotes the first k columns of (V > A),
b and similarly, (V > A)
b [k+1:k+n] denotes the
where (VS> A)
S
S
>
b The first equality is because we can compute the sum of `1 norms on
last n columns of (VS A).
the first k coordinates and `1 norm on the last n coordinates. The first inequality follows from the
triangle inequality. The second inequality follows since the last n columns of  form an identity, so
there is exactly one 1 on the last n columns in each row. The third inequality follows since the last
n columns of  form an identity. Let Dl be a distribution over Dl ∈ R(r+1)×k , where
#
"
b [1:k]
(VS> A)
Dl =
bl )[1:k] .
(A
Let G be a distribution over G ∈ R(r+1)×k where each entry of G is drawn from i.i.d. N (0, 1).
According to Lemma G.16, we have
DTV (Dl , G) ≤ O(kk(VS )l k2 ) + 2−Θ(k) .
(34)
b > , VS , 0.75−
Let A = G[1:r] , v = Gr+1 . Due to Lemma G.20, with probability at least 1−2−Θ(k) , E(A
>
b
γ/2, γ) happens. Then conditioned on E(A , VS , 0.75−γ/2, γ), due to Lemma G.22, with probability
at most 2−Θ(k) , there exists βl such that
kβl A − vk1 + kβl VS> k1 = o(k γ ).
Combined with Equation (34), we can get that for a fixed l, with probability at most O(kk(VS )l k2 )+
2−Θ(k) , there exists βl such that
b [1:k] − (A
bl )[1:k] k1 + kβl VS> k1 = o(k γ ).
kβl (VS> A)
88
l 2
c1 −c2 ), this probability is at most Θ(k 1+(c1 −c2 )/2 ) + 2−Θ(k) . Since
When
Pn k(VSl) k2 2 ≤ 2r/n = Θ(k
l 2
l=1 k(VS ) k2 = r, there are at most n/2 such l that k(VS ) k2 > 2r/n which means that there are
l
2
at least n/2 such l that k(VS ) k2 ≤ 2r/n. Let s be the number of l such that k(VS )l k22 ≤ 2r/n,
then s > n/2. Let t be a random variable which denotes that the number of l which satisfies
k(VS )l k22 ≤ 2r/n and achieve
b [1:k] − (A
bl )[1:k] k1 + kβl VS> k1 = o(k γ ),
kβl (VS> A)
at the same time. Then,
E[t] ≤ (O(kk(VS )l k2 ) + 2−Θ(k) )s = (O(k
p
2r/n) + 2−Θ(k) )s.
Due to a Markov inequality,
Pr[t > s/2 > n/4] ≤ O(k
p
2r/n) + 2−Θ(k) = O(k 1+(c1 −c2 )/2 ) + 2−Θ(k) .
The equality follows since r = Θ(k c1 ), n = Θ(k c2 ). Plugging it into Equation (32), now we can
conclude, with probability at least 1 − O(k 1+(c1 −c2 )/2 ) − 2−Θ(k) , ∀B ∈ Rn×r
X
b −A
bl k1 ≥ n/4 · Ω(k γ ) = Ω(n · k γ ).
b − Ak
b 1≥
kBS A
k(BUS ΣS )l (VS> A)
l:kVSl k22 ≤2r/n
Theorem G.28 (Hardness for row subset selection). For any k ≥ 1, any constant c ≥ 1, let
n = O(k c ), and let A(k, n) denote the same distribution stated in Theorem G.27. For matrix
b ∼ A(k, n), with positive probability, there is no algorithm that is able to output B ∈ Rn×(n+k) in
A
b such that
the row span of any r = n/2 rows of A
b − Bk1 ≤ O(k 0.5−α )
kA
min
rank −k A0
b 1,
kA0 − Ak
where α ∈ (0, 0.5) is a constant which can be arbitrarily small.
Proof. For convenience, we define γ = 0.5 − α which can be an arbitrary constant in (0, 0.5). Since
b is an identity matrix, we can fit the first k columns of A,
b so we have
the last n columns of A
min
rank −k A0
b 1 ≤ n.
kA0 − Ak
b
We want to argue that ∀B ∈ Rn×(k+n) in the row span of any r = n/2 rows of A,
b − Bk1 ≥ Ω(n · k γ ).
kA
b and let S ⊂ [n] be a set of indices of chosen rows of A
b
Let A> ∈ Rn×k be the first k columns of A,
S
k×n
th
S
S
with k ≤ |S| ≤ r. Let M ∈ R
with the i column Mi = Ai if i ∈ S and Mi = 0 otherwise.
S
k×r
S
c
We use M ∈ R
to be M without those columns of zeros, so it is a random matrix with each
b
entry i.i.d. Gaussian N (0, 1). Then the minimum cost of using a matrix in the span of rows of A
b is at least:
with index in S to fit A
X
minn kM S xl − Al k1 + kxl k1 − 1 .
l6∈S
xl ∈R
b and the part
The part of kM S xl − Al k1 is just the cost on the lth row of the first k columns of A,
b
of kxl k1 − 1 is just the lower bound of the cost on the lth row of the last n columns of A.
89
cS ∈ Rk×n , γ ∈ (0, 0.5),
Claim G.29. A, M
bM
cS , Ir , 0.75 − γ/2, γ)
Pr E(
b
E(A, In , 0.75 − γ/2, γ) = 1.
b
Proof. Suppose E(A,
I , 0.75 − γ/2, γ) happens. Since M S has just a subset of columns of A,
√ n
S
kM k2 ≤ kAk2 ≤ n. Notice that ∀x ∈ Rn with kxk1 ≤ O(k γ ) and each non-zero coordinate of
x is at most O(1/k 0.75−γ/2 ), M S x ≡ M S xS ≡ AxS , where xS ∈ Rn has xS i = xi if i ∈ S and
xS i = 0 otherwise. Because kxS k1 ≤ O(k γ ) and xS has each coordinate in absolute value at most
O(1/k 0.75−γ/2 ), AxS has at most O(k/ log k) coordinates in absolute value at least Ω(1/ log k). So,
M S x = AxS has at most O(k/ log k) coordinates in absolute value at least Ω(1/ log k).
We denote cost(S, l) = minxl ∈Rn kM S xl − Al k1 + kxl k1 − 1 . Since ∀l 6∈ S, Al are independent,
and they are independent from M S , due to Lemma G.22,
X
bM
cS , Ir , 0.75 − γ/2, γ) ≤ 2−Θ(rk) .
Pr
cost(S, l) ≤ O(n · k γ ) E(
(35)
l6∈S
Now we just want to upper bound the following:
X
Pr ∃S ⊂ [n], |S| ≤ r,
cost(S, l) ≤ O(n · k γ )
l6∈S
≤ Pr ∃S ⊂ [n], |S| ≤ r,
X
b
cost(S, l) ≤ O(n · k γ ) E(A,
In , 0.75 − γ/2, γ)
l6∈S
h
i
b
+ Pr ¬E(A,
In , 0.75 − γ/2, γ)
X
X
b
≤
Pr
cost(S, l) ≤ O(n · k γ ) E(A,
In , 0.75 − γ/2, γ)
S⊂[n],|S|≤r
l6∈S
h
i
b
+ Pr ¬E(A,
In , 0.75 − γ/2, γ)
X
X
b
≤
Pr
cost(S, l) ≤ O(n · k γ ) E(A,
In , 0.75 − γ/2, γ) + 2−Θ(k)
S⊂[n],|S|≤r
≤
l6∈S
X
bM
cS , Ir , 0.75 − γ/2, γ) + 2−Θ(k)
Pr
cost(S, l) ≤ O(n · k γ ) E(
X
S⊂[n],|S|≤r
l6∈S
r −Θ(rk)
≤ (n + 1) 2
+ 2−Θ(k)
≤ 2−Θ(rk) + 2−Θ(k)
≤ 2−Θ(k) .
The second inequality follows by a union bound. The third inequality follows by Lemma G.20. The
fourth inequality follows by Claim G.29. The fifth inequality is due to Equation (35). The sixth
inequality follows by n ≤ O(k c ), r = n/2. Thus, with probability at least 1 − 2−Θ(k) , ∀B ∈ Rn×(n+k)
b
which is in the span of any r ≤ n/2 rows of A,
b 1 ≥ Ω(n · k γ ).
kB − Ak
90
Then, we have completed the proof.
Definition G.30. Given a matrix A ∈ Rn×d , a matrix S ∈ Rr×n , k ≥ 1 and γ ∈ (0, 21 ), we say
that an algorithm M(A, S, k, γ) which outputs a matrix B ∈ Rn×r “succeeds”, if
kBSA − Ak1 ≤ k γ ·
min
rank −k A0
kA0 − Ak1 ,
holds.
Theorem G.31 (Hardness for oblivious embedding). Let Π denote a distribution over matrices
S ∈ Rr×n . For any k ≥ 1, any constant γ ∈ (0, 12 ), arbitrary constants c1 , c2 > 0 and min(n, d) ≥
Ω(k c2 ), if for all A ∈ Rn×d , it holds that
Pr [M(A, S, k, γ) succeeds ] ≥ Ω(1/k c1 ).
S∼Π
Then r must be at least Ω(k c2 −2c1 −2 ).
Proof. We borrow the idea from [NN14, PSW16]. We use Yao’s minimax principle [Yao77] here.
Let D be an arbitrary distribution over Rn×d , then
Pr
A∼D,S∼Π
[M(A, S, k, γ)] ≥ 1 − δ.
It means that there is a fixed S0 such that
Pr [M(A, S0 , k, γ)] ≥ 1 − δ.
A∼D
Therefore, we want to find a hard distribution Dhard that if
Pr
A∼Dhard
[M(A, S0 , k, γ)] ≥ 1 − δ,
S0 must have at least some larger poly(k) rows.
Here, we just use the distribution A(k, Ω(k c2 )) described in Theorem G.27 as our hard distribution. We can just fill zeros to expand the size of matrix to n × d. We can complete the proof by
using Theorem G.27.
Remark G.32. Actually, in Lemma G.20 and Lemma G.22, the reason we need β > γ > 0 is that
we want k β−γ = ω(poly(log k)), and the reason
we need β +γ < 1√is that we want k β+γ poly(log k) =
√
γ
k/ log20 k, and replace all the k β
o(k).
√ Thus we can replace all the k by k/ poly(log
√ k), e.g.,
by k poly(log k) with a smaller poly(log k), e.g., k log10 k. Our proofs still work. Therefore,
if
the approximation ratio in Theorem G.27, Theorem G.28, and Theorem G.31 to be
√ we replace
c
k/ log k where c is a sufficiently large constant, the statements are still correct.
H
Hardness
This section presents our hardness results. Section H.1 states several useful tools from literature.
Section H.2 shows that, it is NP-hard to get some multiplicative error. Assuming ETH is true, we
provide a stronger hardness result in Section H.3. Section H.4 extends the result from the rank-1
case to the rank-k case.
91
H.1
Previous results
Definition H.1 (kAk∞→1 ,[GV15]). Given matrix A ∈ Rn×d ,
kAk∞→1 =
min
x∈{−1,+1}n ,y∈{−1,+1}d
x> Ay.
The following lemma says that computing kAk∞→1 for matrix A with entries in {−1, +1} is
equivalent to computing a best {−1, +1} matrix which is an `1 norm rank-1 approximation to A.
Lemma H.2 (Lemma 3 of [GV15]). Given matrix A ∈ {−1, +1}n×d ,
kAk∞→1 +
min
x∈{−1,+1}n ,y∈{−1,+1}d
kA − xy > k1 = nd.
Lemma H.3 (Theorem 2 of [GV15]). Given A ∈ {−1, +1}n×d , we have
min
x∈{−1,+1}n ,y∈{−1,+1}d
kA − xy > k1 =
min
x∈Rn ,y∈Rd
kA − xy > k1 .
Combining with Lemma H.2 and Lemma H.3, it implies that computing kAk∞→1 for A ∈
{−1, +1}n×d is equivalent to computing the best `1 norm rank-1 approximation to the matrix A:
kAk∞→1 +
min
x∈Rn ,y∈Rd
kA − xy > k1 = nd.
Theorem H.4 (NP-hard result, Theorem 1 of [GV15]). Computing kAk∞→1 for matrix A ∈
{−1, +1}n×d is NP-hard.
The proof of the above theorem in [GV15] is based on the reduction from MAX-CUT problem.
The above theorem implies that computing minx∈Rn ,y∈Rd kA − xy > k1 is also NP-hard.
H.2
Extension to multiplicative error `1 -low rank approximation
The previous result only shows that solving the exact problem is NP-hard. This section presents
a stronger hardness result, which says that, it is still NP-hard even if the goal is to find a solution
that is able to achieve some multiplicative error. The proof in this section and the next section are
based on the reduction from the MAX-CUT problem. For recent progress on MAX-CUT problem,
we refer the readers to [GW95, BGS98, TSSW00, Hås01, KKMO07, FLP15].
Theorem H.5. Given A ∈ {−1, +1}n×d , computing an x
b ∈ Rn , yb ∈ Rd s.t.
kA − x
b> ybk1 ≤ (1 +
1
) min kA − x> yk1
nd x∈Rn ,y∈Rd
is NP-hard.
MAX-CUT decision problem: Given a positive integer c∗ and an unweighted graph G = (V, E)
where V is the set of vertices of G and E is the set of edges of G, the goal is to determine whether
there is a cut of G has at least c∗ edges.
Lemma H.6. MAX-CUT decision problem is NP-hard.
We give the definition for the Hadamard matrix,
92
Definition H.7. The Hadamard matrix Hp of size p × p is defined recursively :
+1 +1
with H2 =
.
+1 −1
Hp/2 Hp/2
Hp/2 −Hp/2
For simplicity, we use H to denote Hp in the rest of the proof.
Recall the reduction shown in [GV15] which is from MAX-CUT to computing k · k∞→1 for
{−1, +1} matrices. We do the same thing: for a given graph G = (V, E), we construct a matrix
A ∈ {−1, +1}n×d where n = p|E| and d = p|V |. Notice that p = poly(|E|, |V |) is a parameter
which will be determined later, and also p is a power of 2.
We divide the matrix A into |E| × |V | blocks, and each block has size p × p. For e ∈ [|E|], if the
eth edge has endpoints i ∈ [|V |], j ∈ [|V |] and i < j, let all the p × p elements of (e, i) block of A be
1, all the p × p elements of (e, j) block of A be −1, and all the (e, l) block of A be p × p Hadamard
matrix H for l 6= i, j.
Claim H.8 (Lower bound of kAk∞→1 , proof of Theorem 1 of [GV15]). If there is a cut of G with
cut size at least c,
kAk∞→1 ≥ 2p2 c − |E||V |p3/2 .
Claim H.9 (Upper bound of kAk∞→1 , proof of Theorem 1 of [GV15]). If the max cut of G has
fewer than c edges,
kAk∞→1 ≤ 2p2 (c − 1) + |E||V |p3/2 .
Remark H.10. In [GV15], they set p as a power of 2 and p > |E|2 |V |2 . This implies
∀c ∈ [|E|], 2p2 (c − 1) + |E||V |p3/2 < 2p2 c − |E||V |p3/2 .
Therefore, according to Claim H.8 and Claim H.9, if we can know the precise value of kAk∞→1 , we
can decide whether G has a cut with cut size at least c∗ .
For convenience, we use T ∗ to denote kAk∞→1 and use L∗ to denote
min
x∈Rn ,y∈Rd
Also, we use L to denote a (1 +
1
nd )
kA − x> yk1 .
relative error approximation to L∗ , which means:
L∗ ≤ L ≤ (1 +
1 ∗
)L .
nd
We denote T as nd − L.
Proof of Theorem H.5. Because L∗ ≤ L ≤ (1 +
1
∗
nd )L ,
we have:
nd − L∗ ≥ nd − L ≥ nd − (1 +
Due to Lemma H.2 and the definition of T , it has:
T∗ ≥ T ≥ T∗ −
93
1 ∗
L .
nd
1 ∗
)L .
nd
Notice that A is a {−1, +1} matrix, we have
L∗ ≤ kAk1 ≤ 2nd.
Thus,
T ∗ ≥ T ≥ T ∗ − 2.
It means
T + 2 ≥ T ∗ ≥ T.
According to Claim H.16, if G has a cut with cut size at least c, we have:
T + 2 ≥ T ∗ ≥ 2p2 c − |E||V |p3/2 .
That is
T ≥ 2p2 c − |E||V |p3/2 − 2.
According to Claim H.17, if the max cut of G has fewer than c edges,
T ≤ T ∗ ≤ 2p2 (c − 1) + |E||V |p3/2 .
Let p be a power of 2 and p > |E|3 |V |3 , we have
2p2 (c − 1) + |E||V |p3/2 < 2p2 c − |E||V |p3/2 − 2.
Therefore, we can decide whether G has a cut with size at least c based on the value of T .
Thus, if we can compute x
b ∈ Rn , yb ∈ Rd s.t.
kA − x
b> ybk1 ≤ (1 +
1
) min kA − x> yk1 ,
nd x∈Rn ,y∈Rd
in polynomial time, it means we can compute L and T in polynomial time, and we can solve
MAX-CUT decision problem via the value of T , which leads to a contradiction.
H.3
Usin the ETH assumption
The goal of this section is to prove Theorem H.13. We first introduce the definition of 3-SAT and
Exponential Time Hypothesis(ETH). For the details and background of 3-SAT problem, we refer
the readers to [AB09].
Definition H.11 (3-SAT problem). Given an r variables and m clauses conjunctive normal form
CNF formula with size of each clause at most 3, the goal is to decide whether there exists an assignment for the r boolean variables to make the CNF formula be satisfied.
Hypothesis H.12 (Exponential Time Hypothesis (ETH) [IPZ98]). There is a δ > 0 such that
3-SAT problem defined in Definition H.11 cannot be solved in O(2δr ) running time.
The main lower bound is stated as follows:
94
Theorem H.13. Unless ETH(see Hypothesis H.12) fails, for arbitrarily small constant γ > 0, given
some matrix A ∈ {−1, +1}n×d , there is no algorithm can compute x
b ∈ Rn , yb ∈ Rd s.t.
kA − x
b> ybk1 ≤ (1 +
1
log
1+γ
nd
)
min
x∈Rn ,y∈Rd
kA − x> yk1 ,
in (nd)O(1) running time.
Before we prove our lower bound, we introduce the following theorem which is used in our proof.
Definition H.14 (MAX-CUT decision problem). Given a positive integer c∗ and an unweighted
graph G = (V, E) where V is the set of vertices of G and E is the set of edges of G, the goal is to
determine whether there is a cut of G has at least c∗ edges.
Theorem H.15 (Theorem 6.1 in [FLP15]). There exist constants a, b ∈ (0, 1) and a > b, such
that, for a given MAX-CUT (see Definition H.14) instance graph G = (E, V ) which is an n-vertices
5-regular graph, if there is an algorithm in time 2o(n) which can distinguish the following two cases:
1. At least one cut of the instance has at least a|E| edges,
2. All cuts of the instance have at most b|E| edges,
then ETH(see Hypothesis H.12) fails.
Proof of Theorem H.13. We prove it by contradiction. We assume, for any given A ∈ {−1, +1}n×d ,
there is an algorithm can compute x
b ∈ Rn , yb ∈ Rd s.t.
kA − x
b> ybk1 ≤ (1 +
1
) min kA − x> yk1 ,
W x∈Rn ,y∈Rd
in time poly(nd), where W = log1+γ d for arbitrarily small constant γ > 0. Then, we show the
following. There exist constants a, b ∈ [0, 1], a > b, for a given MAX-CUT instance G = (V, E) with
|E| = O(|V |), such that we can distinguish whether G has a cut with size at least a|E| or all the
cuts of G have size at most b|E| in 2o(|V |) time, which leads to a contradiction to Theorem H.15.
Recall the reduction shown in [GV15] which is from MAX-CUT to computing k · k∞→1 for
{−1, +1} matrices. We do similar things here: for a given graph G = (V, E) where |E| = O(|V |),
we construct a matrix A ∈ {−1, +1}n×d where n = p|E| and d = p|V |. Notice that p is a parameter
which will be determined later, and also p is a power of 2.
We divide the matrix A into |E| × |V | blocks, and each block has size p × p. For e ∈ [|E|], if the
eth edge has endpoints i ∈ [|V |], j ∈ [|V |] and i < j, let all the p × p elements of (e, i) block of A be
1, all the p × p elements of (e, j) block of A be −1, and all the (e, l) block of A be p × p Hadamard
matrix H for l 6= i, j.
We can construct the matrix in nd time, which is p2 |E||V |. We choose p to be the smallest
2
number of power of 2 which is larger than 2 a−b |V |
γ
1− 10
γ
for some γ > 0. Thus, the time for construction
1−
2O(|V | 10 ) .
of the matrix A is O(nd) =
We will show, if we can compute a (1 + 1/W )-approximation to A, we can decide whether G
has a cut with size at least a|E| or has no cut with size larger than b|E|. For convenience, we use
T ∗ to denote kAk∞→1 and use L∗ to denote
min
x∈Rn ,y∈Rd
kA − x> yk1 .
95
Also, we use L to denote a (1 +
1
W)
relative error approximation to L∗ , which means:
L∗ ≤ L ≤ (1 +
We denote T as nd − L.
Because L∗ ≤ L ≤ (1 +
1
∗
W )L ,
1 ∗
)L .
W
we have:
nd − L∗ ≥ nd − L ≥ nd − (1 +
1 ∗
)L .
W
Due to Lemma H.2 and the definition of T , it has:
T∗ ≥ T ≥ T∗ −
1 ∗
L .
W
Notice that A is a {−1, +1} matrix, we have
L∗ ≤ kAk1 ≤ 2nd.
Thus,
T ∗ ≥ T ≥ T ∗ − 2nd/W.
It means
T + 2nd/W ≥ T ∗ ≥ T.
Claim H.16 (Lower bound of kAk∞→1 , proof of Theorem 1 of [GV15]). If there is a cut of G with
cut size at least c,
kAk∞→1 ≥ 2p2 c − |E||V |p3/2 .
Claim H.17 (Upper bound of kAk∞→1 , proof of Theorem 1 of [GV15]). If the max cut of G has
fewer than c edges,
kAk∞→1 ≤ 2p2 (c − 1) + |E||V |p3/2 .
According to Claim H.16, if G has a cut with cut size at least a|E|, we have:
T + 2nd/W ≥ T ∗ ≥ 2p2 a|E| − |E||V |p3/2 .
That is
T ≥ 2p2 a|E| − |E||V |p3/2 − 2nd/W.
(36)
According to Claim H.17, if the max cut of G has fewer than b|E| edges,
T ≤ T ∗ ≤ 2p2 b|E| + |E||V |p3/2 .
2
Using these conditions p ≥ 2 a−b |V |
up to some constant,
γ
1− 10
(37)
, d = p|V |, W ≥ log1+γ d, we can lower bound |W | by |V |
W ≥ log1+γ d
by W ≥ log1+γ d
= log1+γ (p|V |)
by d = p|V |
1+γ
= (log |V | + log p)
γ
2
|V |1− 10 )1+γ
≥ (log |V | +
a−b
2
≥
|V |.
a−b
2
by p ≥ 2 a−b |V |
γ
1− 10
by (1 − γ/10)(1 + γ) > 1 for γ small enough
96
(38)
Thus, we can upper bound 1/W in the following sense,
1
1
a−b
a−b
≤
≤
− p− 2 ,
W
2|V |
|V |
(39)
where the first inequality follows by Equation (38) and the second inequality follows by p ≥
2
γ
1− 10
2 a−b |V |
, γ is sufficient small, and |V | is large enough.
Now, we can conclude,
1
1
a−b
≤
− p− 2
W
|V |
3
⇐⇒ p2 |E||V |/W ≤ (a − b)p2 |E| − |E||V |p 2
by multiplying p2 |E||V | on both sides
3
⇐⇒ 2nd/W ≤ 2(a − b)p2 |E| − 2|E||V |p 2
by multiplying 2 on both sides and p2 |E||V | = nd
⇐⇒ 2p2 b|E| + |E||V |p3/2 < 2p2 a|E| − |E||V |p3/2 − 2nd/W,
(40)
by adding 2p2 b|E| + |E||V |p3/2 − 2nd/W on both sides
which implies that Equation (40) is equivalent to Equation (39). Notice that the LHS of (40) is
exactly the RHS of (37) and the RHS of (40) is exactly the RHS of (36). Therefore, we can decide
whether G has a cut with size larger than a|E| or has no cut with size larger than b|E|.
Thus, if we can compute x
b ∈ Rn , yb ∈ Rd s.t.
kA − x
b> ybk1 ≤ (1 +
1
log
1+γ
d
)
min
x∈Rn ,y∈Rd
kA − x> yk1 ,
in poly(nd) time, which means we can compute L and T in poly(nd) time. Notice that nd =
γ
1− 10
2
1−
γ
p2 |E||V |, |E| = O(|V |), p ≥ 2 a−b |V |
, it means poly(nd) = 2O(|V | 10 ) . Because we decide
whether G has a cut with size larger than a|E| or has no cut with size larger than b|E| via the value
of T , we can solve it in 2O(|V |
H.4
γ
1− 10
)
time which leads to a contradiction to Theorem H.15.
Extension to the rank-k case
This section presents a way of reducing the rank-k case to the rank-1 case. Thus, we can obtain a
lower bound for general k ≥ 1 under ETH.
Theorem H.18. For any constants c1 > 0, c2 > 0 and c3 > 0, and any constant c4 ≥ 10(c1 + c2 +
c3 + 1), given any matrix A ∈ Rn×n with absolute value of each entry bounded by nc1 , we define a
e ∈ R(n+k−1)×(n+k−1) as
block diagonal matrix A
A
0
0 ··· 0
0
B
0 ··· 0
e= 0
0
B
··· 0
A
,
· · · · · · · · · · · · · · ·
0
0
0 ··· B
b is an `1 -norm rank-k C-approximation solution to A,
e i.e.,
where B = nc4 . If A
b − Ak
e 1≤C·
kA
min
rank −k
97
b0
A
b0 − Ak
e 1,
kA
where C ∈ [1, nc3 ], then there must exist j ∗ ∈ [n] such that
>
>
c2
b[1:n]
minn kA
j ∗ v − Ak1 ≤ C · minn kuv − Ak1 + 1/n ,
u,v∈R
v∈R
b can give an `1 -norm rank-1 C-approximation to
i.e., the first n coordinates of the column j ∗ of A
A.
Proof. The first observation is that because we can use a rank-1 matrix to fit A and use a rank-(k−1)
matrix to fit other Bs, we have
min
b0
rank −k A
b0 − Ak
e 1 ≤ min kuv > − Ak1 ≤ kAk1 .
kA
n
u,v∈R
(41)
b denote the rank-k C-approximate solution to A.
e Let Z ∈ R(n+k−1)×(k−1) denote
Claim H.19. Let A
b
the rightmost k − 1 columns of A, then, rank(Z) = k − 1.
Proof. Consider the (k − 1) × (k − 1) submatrix Z [n+1:n+k−1] of Z. Each element on the diagonal of this submatrix should be at least B − CkAk1 , and each element not on the diagonal of
b − Ak
e 1 > CkAk1 which will lead to a
the submatrix should be at most CkAk1 . Otherwise kA
c
[n+1:n+k−1]
4
contradiction. Since B = n is sufficiently large, Z
is diagonally dominant. Thus
rank(Z) ≥ rank(Z [n+1:n+k−1] ) = k − 1. Because Z only has k − 1 columns, rank(Z) = k − 1.
Claim H.20. ∀x ∈ Rk−1 , i ∈ [n], ∃j ∈ {n + 1, n + 2, · · · , n + k − 1} such that,
|(Zx)j |
B
≥
.
|(Zx)i |
2(k − 1)CkAk1
Proof. Without loss of generality, we can let kxk1 = 1. Thus, there exists j such that |xj | ≥
So we have
|(Zx)n+j | = |
k−1
X
1
k−1 .
Zn+j,i xi |
i=1
≥ |Zn+j,j xj | −
X
|Zn+j,i xi |
i6=j
≥ (B − CkAk1 )|xj | −
X
|xi |CkAk1
i6=j
≥ (B − CkAk1 )/(k − 1) − CkAk1
B
.
≥
2(k − 1)
The second inequality follows because |Zn+j,j | ≥ B−CkAk1 and ∀i 6= j, |Zn+j,i | ≤ CkAk1 (otherwise
b Ak
e 1 > CkAk1 which leads to a contradiction.) The third inequality follows from |xj | > 1/(k−1)
kA−
B
1
and kxk1 = 1. The fourth inequality follows since B is large enough such that 2(k−1)
≥ CkAk
k−1 +
CkAk1 .
Now we consider any q ∈ [n]. We have
|(Zx)q | =
k−1
X
i=1
|Zq,i xi | ≤ max |Zq,i | ·
i∈[k−1]
98
k−1
X
i=1
|xi | ≤ CkAk1 .
b − Ak
e 1>
The last inequality follows that kxk1 = 1 and ∀i ∈ [k − 1], Zq,i ≤ CkAk1 . Otherwise, kA
CkAk1 which will lead to a contradiction.
B
Look at |(Zx)n+j |/|(Zx)q |, it is greater than 2(k−1)CkAk
.
1
b[1:n] , we choose i∗ , j ∗ ∈ [n] such that
We now look at the submatrix A
[1:n]
bi∗ ,j ∗ | ≥ 1/nc2 −2 .
|A
If there is no such (i∗ , j ∗ ), it means that we already found a good rank-1 approximation to A
b[1:n] k1 + kA
b[1:n] − Ak1
k0 − Ak1 ≤ k0 − A
[1:n]
[1:n]
[1:n]
b
b e
≤ k0 − A
[1:n] k1 + kA − Ak1
b − Ak
e 1
≤ 1/nc2 + kA
≤ 1/nc2 + C
min
b0
rank −k A
b0 − Ak
e 1
kA
≤ 1/nc2 + C minn kuv > − Ak1 ,
u,v∈R
[1:n]
b
where 0 is an n × n all zeros matrix. The second inequality follows since A
[1:n] − A is a submatrix
[1:n]
b − A.
e The third inequality follows since each entry of A
b
of A
should be no greater than 1/nc2 −2
[1:n]
(otherwise, we can find (i∗ , j ∗ )). The last inequality follows from equation 41.
bj ∗ is not in the column span of Z, i.e.,
Claim H.21. A
bj ∗ 6= Zx.
∀x ∈ Rk−1 , A
bj ∗ = Zx, it means (Zx)i∗ = A
bi∗ ,j ∗ ≥ 1/nc2 −2 . Due to Claim H.20,
Proof. If there is an x such that A
B
there must exist i0 ∈ {n + 1, n + 2, · · · , n + k − 1} such that (Zx)i0 ≥ 2(k−1)CkAk
c −2 . Since B is
1n 2
bi0 ,j ∗ = (Zx)i0 > CkAk1 which implies that kA
b − Ak
e 1 > CkAk1 , and so leads to
sufficiently large, A
a contradiction.
bj ∗ and the
Due to Claim H.19 and Claim H.21, the dimension of the subspace spanned by A
b
b
column space of Z is k. Since A has rank at most k, it means that each column of A can be written
bj ∗ and the columns of Z.
as a linear combination of A
th
bj of A
b for j ∈ [n]. We write it as
Now consider the j column A
bj = αj · A
bj ∗ + Zxj .
A
Claim H.22. ∀j ∈ [n], αj ≤ 2CkAk1 nc2 +2 .
Proof. Otherwise, suppose αj > 2CkAk1 nc2 +2 . We have
bi∗ ,j ∗ | − |A
bi∗ ,j |
|(Zxj )i∗ | ≥ αj · |A
1
bi∗ ,j |
≥ αj · c2 −2 − |A
n
1
1
≥ αj · c2 −2 .
2
n
99
bi∗ ,j ∗ | ≥ 1/nc2−2 . The third inequality follows from |A
bi∗ ,j | ≤
The second inequality follows from |A
1
1
kAk1 and 2 αj · nc2 −2 ≥ CkAk1 ≥ kAk1 .
B
·
Due to Claim H.20, there exists i ∈ {n+1, n+2, · · · , n+k −1} such that |(Zxj )i | ≥ 2(k−1)CkAk
1
1
2 αj
·
1
.
nc2 −2
For sufficiently large B, we can have |(Zxj )i | ≥ αj B 1/2 . Then we look at
bi,j | ≥ |(Zxj )i | − αj |A
bi,j ∗ |
|A
≥ αj (B 1/2 − CkAk1 )
1
≥ αj B 1/2 .
2
bi,j ∗ | ≤ CkAk1 , otherwise kA
b − Ak
e 1 > CkAk1 which will lead to
The second inequality follows by |A
a contradiction. The third inequality follows that B is sufficient large that 12 B 1/2 > CkAk1 .
bi,j | ≥ αj 1 B 1/2 > CkAk1 , it contradicts to the fact kA
b − Ak
e 1 ≤ CkAk1 .
Since |A
2
c
+2
Therefore, ∀j ∈ [n], αj ≤ 2CkAk1 n 2 .
Claim H.23. ∀j ∈ [n], i ∈ {n + 1, n + 2, · · · , n + k − 1}, |(Zxj )i | ≤ 4C 2 kAk21 nc2 +2
Proof. Consider j ∈ [n], i ∈ {n + 1, n + 2, · · · , n + k − 1}, we have
bi,j | + αj |A
bi,j ∗ |
|(Zxj )i | ≤ |A
≤ CkAk1 + αj · CkAk1
≤ 4C 2 kAk21 nc2 +2 .
bi,j | ≤ CkAk1 and |A
bi,j ∗ | ≤ CkAk1 , otherwise the kA
b − Ak
e 1 will
The second inequality follows by |A
be too large and leads to a contradiction. The third inequality is due to αj + 1 ≤ 4CkAk1 nc2 +2 via
Claim H.22.
[1:n]
b
Claim H.24. ∀j ∈ [n], kA
j
c2 −2
b1:n
− αj · A
j ∗ k∞ ≤ 1/n
Proof. Due to Claim H.23 and Claim H.20, ∀i, j ∈ [n], we have
|(Zxj )i | ≤
4C 2 kAk21 nc2 +2
≤ 1/B 1/2 .
B/(2(k − 1)CkAk1 )
The second inequality follows for a large enough B.
Therefore, ∀i, j ∈ [n],
bi,j − αj · A
bi,j ∗ | ≤ |(Zxj )i | ≤ 1/B 1/2 ≤ 1/nc2 −2 .
|A
The last inequality follows since B is large enough.
b1:n
Now, let us show that A
j ∗ can provide a good rank-1 approximation to A:
[1:n]
[1:n]
>
b1:n
b1:n >
b
b
kA
j ∗ α − Ak1 ≤ kAj ∗ α − A[1:n] k1 + kA[1:n] − Ak1
=
n
X
[1:n]
bj ∗ − A
b
kαj A
j
[1:n]
b
k1 + kA
[1:n] − Ak1
j=1
[1:n]
b
≤ n2 · 1/nc2 −2 + kA
[1:n] − Ak1
b − Ak
e 1
≤ 1/nc2 + kA
≤ 1/nc2 + C minn kuv > − Ak1 .
u,v∈R
100
The first inequality follows by triangle inequality. The first equality is due to the linearity of `1
b[1:n] − A is
norm. The second inequality is due to Claim H.24. The third inequality follows since A
[1:n]
b − A.
e The fourth inequality is due to the equation 41.
a submatrix of A
I
Limited independent Cauchy random variables
This section presents the fundamental lemmas with limited independent Cauchy variables, which
will be used in Section J and K. In Section I.1, we provide some notation, definitions and tools from
previous work. Section I.2 includes the main result.
I.1
Notations and Tools
For a function f : R → R and nonnegative integer `, f (`) denotes the `th derivative of f , with
f (0) = f . We also often use x ≈ y to state that |x − y| = O(). We use I[a,b] to denote the indicator
function of the interval [a, b].
To optimize the communication complexity of our ditributed algorithm, we show that instead
of using fully independent Cauchy variables, poly(k, d)-wise independent Cauchy variables suffice.
We start by stating two useful Lemmas from previous work [KNW10].
Lemma I.1 (Lemma 2.2 in [KNW10]). There exists an 0 > 0 such that the following holds. Let n
be a positive integer and 0 < < 0 , 0 < p < 2 be given. Let f : R → R satisfy kf (`) k∞ = O(α` ) for
all ` ≥ 0, for some α satisfying αp ≥ log(1/). Let k = αp . Let a ∈ RnPsatisfy kakp = O(1).
P Let
Xi be a 3Ck-independent family of p-stable random variables. Let X = i ai Xi and Y = i ai Yi .
Then E[f (x)] = E[f (Y )] + O().
Lemma I.2 (Lemma 2.5 in [KNW10]). There exist constants c0 , 0 > 0 such that for all c > 0 and
c
0 < < 0 , and for all [a, b] ⊆ R, there exists a function J[a,b]
: R → R satisfying:
c )(`) k
`
i. k(J[a,b]
∞ = O(c ) for all ` ≥ 0.
c (x)−I
ii. For all x such that a, b ∈
/ [x−, x+], and as long as c > c0 −1 log3 (1/), |J[a,b]
[a,b] (x)| <
.
I.2
Analysis of limited independent random Cauchy variables
Lemma I.3. Given a vector y ∈ Rn , choose Z to be the t × n random Cauchy matrices with
1/t rescaling and t = O(k log k). The variables from different rows are fully independent, and the
variables from the same rows are O(1)-wise independent. Then, we have
kZyk1 & kyk1
holds with probability at least 1 − 2−Ω(t) .
Proof. Let S denote the original fully independent matrix and Z denote the matrix for which
the entries in the same row are w-wise independent, and the entries from different rows are fully
independent. Notice we define the random matrices without rescaling by 1/t and it will be added
back at the end. (We will decide w later)
101
1
and X = 0 otherwise. We also
We define random variable X such that X = 1 if |(Zy)i | ≤ 50
1
define random variable Y such that Y = 1 if |(Sy)i | ≤ 50 and Y = 0 otherwise. Then, we have
1
= E I[− 1 , 1 ] ((Zy)i )
E[X] = Pr |(Zy)i | ≤
50 50
50
1
E[Y ] = Pr |(Sy)i | ≤
= E I[− 1 , 1 ] ((Sy)i )
50 50
50
The goal is to show that E[X] ≈ E[Y ]. Following the same idea from [KNW10], we need to argue
this chain of inequalities,
c
c
E[I[a,b] (X)] ≈ E[J[a,b]
(X)] ≈ E[J[a,b]
(Y )] ≈ E[I[a,b] (Y )]
Using Lemma 2.2 and Lemma 2.5 from [KNW10], choosing sufficiently small constant (which
implies w = O(1)), it follows that for each i ∈ [t], we still have
1
1
Pr |(Zy)i | > kyk1 & Pr |(Sy)i | > kyk1 ≥ 0.9
50
50
Because all rows of Z are fully independent, using the Chernoff bound we can get that
Pr kZyk1 . tkyk1 ≤ exp(−Ω(t))
as we needed for the “no contraction” part of the net argument.
For the no dilation, we need to argue that
Lemma I.4. Given a set of vectors {y1 , y2 , . . . , yd } where yi ∈ Rn , ∀i ∈ [d], choose Z to be the t × n
random Cauchy matrices with 1/t rescaling and t = O(k log k), where the variables from different
rows are fully independent, and the variables from the same rows are w-wise independent.
e
I. If w = O(dk),
we have
d
d
X
X
kZyi k1 ≤ O(log d)
kyi k1
i=1
i=1
holds with probability at least .999.
e
II. If If w = O(d),
we have
d
X
kZyi k1 ≤ O(k log d)
i=1
d
X
kyi k1
i=1
holds with probability at least .999.
Proof. Let m = t. Let S ∈ Rm×n denote the original fully independent matrix and Z denote the
matrix that for each entry in the same row are w-wise independent, where the entries from different
rows are fully independent. (We will decide onw later)
Applying matrix S to those fixed set of vectors, we have
d
X
i=1
kSyi k1 =
d m
n
d X
m X
n
X
1
1 XX X
|
Sj,l · (yi )l | =
|
Sj,l · (yi )l |
m
m
i=1 j=1
i=1 j=1
l=1
102
l=1
Applying matrix Z to those fixed set of vectors, we have a similar thing,
d
X
i=1
d m
d X
m X
n
n
X
1 XX X
1
kZyi k1 =
Zj,l · (yi )l | =
|
|
Zj,l · (yi )l |
m
m
i=1 j=1
i=1 j=1
l=1
l=1
The goal is to argue that, for any i ∈ [d], j ∈ [m],
X
X
n
n
E |
Zj,l · (yi )l | ξ . E |
Sj,l · (yi )l | ξ + δ
l=1
l=1
As long as δ is small
enough, we are in a good P
shape.
1
1 Pn
Let X = kyi k1 l=1 Zj,l (yi )l , and Y = kyi k1 m
l=1 Sj,l (yi )l . Let D be the truncating threshold of
each Cauchy random variable. Define T = O(log D). On one hand, we have
T
X
E[|X||ξ] ≤ 2 Pr[X ∈ [0, 1]|ξ] · 1 +
Pr[X ∈ (2j , 2j+1 ]|ξ] · 2j+1
j=0
≤ 2 1 +
T
X
Pr[I(2j ,2j+1 ) (X) = 1|ξ] · 2j+1
j=0
= 2 1 +
T
X
E[I(2j ,2j+1 ) (X)|ξ] · 2j+1
(42)
j=0
On the other hand, we can show
E[|Y ||ξ] ≥ 2 Pr[X ∈ [0, 1]|ξ] · 0 +
T
X
Pr[Y ∈ (2j , 2j+1 ]|ξ] · 2j
j=0
≥2
T
X
Pr[I(2j ,2j+1 ] (Y ) = 1|ξ] · 2j
j=0
≥2
T
X
E[I(2j ,2j+1 ] (Y )|ξ] · 2j
(43)
j=0
Thus, we need to show that, for each j,
E[I(2j ,2j+1 ) (X)|ξ] ≈ E[I(2j ,2j+1 ] (Y )|ξ]
(44)
Following the same idea from [KNW10], we need to argue this chain of inequalities,
c
c
E[I[a,b] (X)] ≈ E[J[a,b]
(X)] ≈ E[J[a,b]
(Y )] ≈ E[I[a,b] (Y )]
c (X)]. Notice that I
We first show E[I[a,b] (X)] ≈ E[J[a,b]
[a,b] and J[a,b] are within everywhere except
for two intervals of length O(). Also the Cauchy distribution is anticoncentrated (any length-O()
c k
interval contains O() probability mass) and kI[a,b] k∞ , kJ[a,b]
∞ = O(1), these intervals contribute
O() to the difference.
c (X)] ≈ E[J c (Y )]. This directly follows by Lemma I.1 by choosing
Second, we show E[J[a,b]
[a,b]
3
−1
α = O( log (1/)).
103
c (Y )] ≈ E[I
Third, we show E[J[a,b]
[a,b] (Y )]. The argument is similar as the first step, but we need
to show anticoncentration of Y . Suppose for any t ∈ R we had a nonnegative function f,t : R → R
symmetric about t satisfying:
(`)
I. kft, k∞ = O(α` ) for all ` ≥ 0, with α = O(1/)
II. E[ft, (z)] = O() for z ∼ D1
III. ft, (t + ) = Ω(1)
IV. ft, (x) is strictly decreasing as |x − t| → ∞
By I, II and Lemma I.1 we could have E[f,t (Y )] ≈ E[ft, (z)] = O(). Then, E[ft, (Y )] ≥ ft, (t+)·
Pr[Y ∈ [t−, t+]] = Ω(Pr[Y ∈ [t−, t+]]) by III and IV, implying anticoncentration in [t−, t+]
as desired. For the details of function ft, , we refer the readers to Section A.4 in [KNW10].
Now, combining Equation (42), (43) and (44) gives
X
X
X
n
n
T
E |
Zj,l · (yi )l | ξ . E |
Sj,l · (yi )l | ξ +
2j · · kyi k1
l=1
l=1
j=0
X
n
.E |
Sj,l · (yi )l | ξ + D · · kyi k1
l=1
e
Overall, for the fixed j, Sj,l is O(1/)-independent
family of Cauchy random variable. Choosing
D = O(dk) and = O(1/D), we can show
X
d m
n
d
X
1 XX
E |
Sj,l · (yi )l | ξ ≤ O(log d)
kyi k1
m
i=1 j=1
i=1
l=1
as before. Notice that Dkyi k1 = O(kyi k1 ). Thus, we complete the proof of the first result.
Choosing D = O(dk) and = O(1/d), the dominant term becomes Dkyi k1 = O(kkyi k1 ). Thus,
we complete the proof of second result.
Corollary I.5. Given U ∈ Rn×k , let Z ∈ Rt×n be the same as the matrix stated in the Lemma I.3,
then with probability at least .95,
∀x ∈ Rk , kZU xk1 & kU xk1 .
The proof is very similar to the proof of Lemma D.22. Without loss of generality, we can suppose
U is a well-conditioned basis. Due to Lemma I.4, with arbitrarily high constant probability kZU k1
is bounded by poly(t, k). By simply applying the net argument and using Lemma I.3 to take a
union bound over net points, we can get the above corollary.
J
Streaming Setting
Section J.1 provides some notation and definitions about row-update streaming model and the
turnstile streaming model. For some recent developments of row-update streaming and turnstile
streaming models, we refer the readers to [CW09, KL11, GP13, Lib13, KLM+ 14, BWZ16] and the
references therein. Section J.2 presents our turnstile streaming algorithm. Section J.3 presents our
row-update streaming algorithm.
104
J.1
Definitions
Definition J.1 (Row-update model). Let matrix A ∈ Rn×d be a set of rows A1 , · · · , An . In the
row-update streaming model, each row of A will occur in the stream exactly once. But the rows can
be in arbitrary order. An algorithm in this model is only allowed a single pass over these rows. At
the end of the stream, the algorithm stores some information of A. The space of the algorithm is
the total number of words required to store this information during the stream. Here, each word is
O(log(nd)) bits.
Definition J.2 (Turnstile model). At the beginning, let matrix A ∈ Rn×d be a zero matrix. In
the turnstile streaming model, there is a stream of update operations, and the ith operation has the
form (xi , yi , ci ) which means that Axi ,yi should be incremented by ci . An algorithm in this model
is only allowed a single pass over the stream. At the end of the stream, the algorithm stores some
information of A. The space complexity of the algorithm is the total number of words required to
store this information during the stream. Here, each word is O(log(nd)) bits.
J.2
Turnstile model, poly(k, log(d), log(n)) approximation
Definition J.3 (Turnstile model `1 -low rank approximation - rank-k subspace version). Given
matrix A ∈ Rn×d and k ∈ N+ , the goal is to propose an algorithm in the streaming model of
Definition J.2 such that
1. Upon termination, the algorithm outputs a matrix V ∗ ∈ Rk×d .
2. V ∗ satisfies that
min kA − U V ∗ k1 ≤ poly(k, log(d), log(n)) ·
U ∈Rn×k
min
U ∈Rn×k ,V ∈Rk×d
kA − U V k1 .
3. The space complexity is as small as possible
Theorem J.4. Suppose A ∈ Rn×d is given in the turnstile streaming model (See Definition J.2),
there is an algorithm(in Algorithm 8 without decomposition) which solves the problem in Definie
tion J.3 with constant probability. Further, the space complexity of the algorithm is poly(k) + O(kd)
words.
Proof. Correctness. The correctness is implied by (IV) of Lemma D.11, and the proof of Theob ∈ RO(k log k)×O(k log k) minimizes
rem C.3. Notice that L = T1 AR, N = SAT2 , M = T1 AT2 , so X
min
rank −k X
kT1 ARXSAT2 − T1 AT2 kF .
b
According to the proof of Theorem C.3, ARXSA
gives a `1 rank-k poly(k, log(d), log(n))-approximation
>
∗
>
b
b
b
b
b
b
to A. Because X = U ΣV , V = ΣV SA satisfies:
min kA − U V ∗ k1 ≤ poly(k, log(d), log(n)) ·
U ∈Rn×k
min
U ∈Rn×k ,V ∈Rk×d
kA − U V k1 .
e
e
Space complexity. Generating O(kd)-wise
independent random Cauchy variables needs O(kd)
2
3
3
2
2
2
bits. The size of L, N and M are k log k, k log k and k log k words separately. So the space
of maintaining them is O(k 2 log3 k) words. The size of D is O(k log k) × d, so maintaining it needs
e
O(kd log k) words. Therefore, the total space complexity of the algorithm is poly(k) + O(kd)
words.
105
It is easy to extend our algorithm to output a decomposition. The formal definition of the
decomposition problem is as the following:
Definition J.5 (Turnstile model `1 -low rank approximation - rank-k decomposition version). Given
matrix A ∈ Rn×d and k ∈ N+ , the goal is to propose an algorithm in the streaming model of
Definition J.2 such that
1. Upon termination, the algorithm outputs a matrix U ∗ ∈ Rn×k , V ∗ ∈ Rk×d .
2. U ∗ , V ∗ satisfies
kA − U ∗ V ∗ k1 ≤ poly(k, log(d), log(n)) ·
min
U ∈Rn×k ,V ∈Rk×d
kA − U V k1 .
3. The space complexity is as small as possible
Theorem J.6. Suppose A ∈ Rn×d is given by the turnstile streaming model (See Definition J.2).
There is an algorithm( in Algorithm 8 with decomposition) which solves the problem in Definition J.5
e
with constant probability. Further, the space complexity of the algorithm is poly(k) + O(k(d
+ n))
words.
Proof. Correctness. The only difference from the Algorithm 8 (without decomposition) is that the
b . Notice that U ∗ V ∗ = ARXSA,
b
algorithm maintains C. Thus, finally it can compute U ∗ = ARU
according to the proof of Theorem C.3, U ∗ V ∗ gives a `1 rank-k poly(k, log(d), log(n))-approximation
to A.
e
Space complexity. Since the size of C is O(nk log k) words, the total space is poly(k)+ O(k(d+
n)) words.
J.3
Row-update model, poly(k) log d approximation
Definition J.7 (Row-update model `1 -low rank approximation - rank-k subspace version). Given
matrix A ∈ Rn×d and k ∈ N+ , the goal is to propose an algorithm in the streaming model of
Definition J.1 such that
1. Upon termination, the algorithm outputs a matrix V ∗ ∈ Rk×d .
2. V ∗ satisfies that
min kA − U V ∗ k1 ≤ poly(k) log(d) ·
U ∈Rn×k
min
U ∈Rn×k ,V ∈Rk×d
kA − U V k1 .
3. The space complexity is as small as possible
Theorem J.8. Suppose A ∈ Rn×d is given by the row-update streaming model (See Definition J.1),
there is an algorithm( in Algorithm 9 without decomposition ) which solves the problem in Definie
tion J.7 with constant probability. Further, the space complexity of the algorithm is poly(k) + O(kd)
words.
106
Algorithm 8 Turnstile Streaming Algorithm
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
procedure TurnstileStreaming(k,S)
Construct sketching matrices S ∈ RO(k log k)×n , R ∈ Rd×O(k log k) , T1 ∈ RO(k log k)×n , T2 ∈
d×O(k
log2 k) where R, T are fully independent random Cauchy matrices, and S, T are random
R
2
1
e
Cauchy matrices with fully independent variables across different rows and O(d)-wise
independent variables from the same row.
Initialize matrices:
2
L ← {0}O(k log k)×O(k log k) , N ← {0}O(k log k)×O(k log k) .
2
M ← {0}O(k log k)×O(k log k) , D ← {0}O(k log k)×d .
if need decomposition then
C ← {0}n×O(k log k) .
end if
for i ∈ [l] do
Receive update operation (xi , yi , ci ) from the data stream S.
for r = 1 → O(k log k), s = 1 → O(k log k) do
Lr,s ← Lr,s + T1r,xi · ci · Ryi ,s .
end for
for r = 1 → O(k log k), s = 1 → O(k log2 k) do
Nr,s ← Nr,s + Sr,xi · ci · T2yi ,s .
end for
for r = 1 → O(k log k), s = 1 → O(k log2 k) do
Mr,s ← Mr,s + T1r,xi · ci · T2yi ,s .
end for
for r = 1 → O(k log k) do
Dr,yi ← Dr,s + Sr,xi · ci .
end for
if need decomposition then
for s = 1 → O(k log k) do
Cxi ,s ← Cxi ,s + ci · Ryi ,s .
end for
end if
end for
Compute the SVD of L = UL ΣL VL> .
Compute the SVD of N = UN ΣN VN> .
b = L† (UL U > M VN V > )k N † .
Compute X
L
N
b =U
bΣ
b Vb > .
Compute the SVD of X
if need decomposition then
b Vb > D, U ∗ = C U
b.
return V ∗ = Σ
else
b Vb > D.
return V ∗ = Σ
end if
end procedure
b ∈ RO(k log k)×O(k log k)
Proof. Correctness. Notice that L = T1 BR, N = SBT2 , M = T1 BT2 . Thus, X
actually minimizes
min kT1 BRXSBT2 − T1 BT2 kF .
rank −k X
107
Also notice that B is just taking each row of A and replacing it with its nearest point in the row
b
span of S 0 A. According to the proof of Theorem C.6 and (IV) of Lemma D.11 BRXSB
gives a
>
∗
>
b
b
b
b
b
b
poly(k) log d `1 norm rank-k approximation to A. Since X = U ΣV , V = ΣV SB satisfies:
min kA − U V ∗ k1 ≤ poly(k, log(d), log(n)) ·
U ∈Rn×k
min
U ∈Rn×k ,V ∈Rk×d
kA − U V k1 .
e
Space complexity. Constructing sketching matrices needs O(kd)
bits to store random seeds.
3
2
b is also O(k 2 log2 k) words.
Maintaining L, N, M needs O(k log k) words. The cost of storing X
Maintaining D needs O(kd log k) words. Therefore, the total space cost of the algorithm is poly(k)+
e
O(kd)
words.
It is easy to extend our algorithm to output a decomposition. The formal definition of the
decomposition problem is as the following:
Definition J.9 (Row-update model `1 -low rank approximation - rank-k decomposition version).
Given matrix A ∈ Rn×d and k ∈ N+ , the goal is to propose an algorithm in the streaming model of
Definition J.1 such that
1. Upon termination, the algorithm outputs matrices U ∗ ∈ Rn×k , V ∗ ∈ Rk×d .
2. U ∗ , V ∗ satisfies that
kA − U ∗ V ∗ k1 ≤ poly(k) log d ·
min
U ∈Rn×k ,V ∈Rk×d
kA − U V k1 .
3. The space complexity is as small as possible.
Theorem J.10. Suppose A ∈ Rn×d is given by the row-update streaming model (See Definition J.1),
there is an algorithm(in Algorithm 9 with decomposition ) which solves the problem in Definition J.9
e
with constant probability. Further, the space complexity of the algorithm is poly(k) + O(k(n
+ d))
words.
Proof. Correctness. The only difference is that the above algorithm maintains C. Thus, We
b in the end. Notice that U ∗ V ∗ = BRXSB,
b
can compute U ∗ = C U
according to the proof of
∗
∗
Theorem C.6, U V gives a poly(k) log d `1 norm rank-k approximation to A.
e
Space complexity. Since the size of C is nk log k words, the total space is poly(k)+ O(k(n+d))
words.
K
Distributed Setting
Section K.1 provides some notation and definitions for the Row-partition distributed model and
the Arbitrary-partition model. These two models were recently studied in a line of works such
as [TD99, QOSG02, BCL05, BRB08, MBZ10, FEGK13, PMvdG+ 13, KVW14, BKLW14, BLS+ 16,
BWZ16, WZ16]. Section K.2 and K.3 presents our distributed protocols for the Arbitrary-partition
distributed model. Section K.4 and K.5 presents our distributed protocols for the Row-partition
distributed model.
108
Algorithm 9 Row Update Streaming Algorithm
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
39:
40:
procedure RowUpdateStreaming(k, S)
Construct sketching matrices S 0 ∈ RO(k log k)×n , S ∈ RO(k log k)×n , R ∈ Rd×O(k log k) , T1 ∈
2
O(k
R log k)×n , T2 ∈ Rd×O(k log k) where R, T2 are fully independent random Cauchy variables,
and S , S 0 , T1 are random Cauchy matrices with fully independent random variables from
e
different rows and O(d)-wise
independent in the same row.
Initialize matrices:
2
L ← {0}O(k log k)×O(k log k) , N ← {0}O(k log k)×O(k log k) .
2
M ← {0}O(k log k)×O(k log k) , D ← {0}O(k log k)×d .
if need decomposition then
C ← {0}n×O(k log k) .
end if
for i ∈ [n] do
Recieve a row update (i, Ai ) from the data stream S.
0 A −A k .
Compute Yi∗ ∈ R1×O(k log k) which minimizes minY ∈R1×O(k log k) kY S:,i
i
i 1
∗
0
Compute Bi = Yi S:,i Ai .
for r = 1 → O(k log k), s = 1 → O(k log k), j = 1 → d do
Lr,s ← Lr,s + T1r,i · Bi,j · Rj,s .
end for
for r = 1 → O(k log k), s = 1 → O(k log2 k), j = 1 → d do
Nr,s ← Nr,s + Sr,i · Bi,j · T2j,s .
end for
for r = 1 → O(k log k), s = 1 → O(k log2 k), j = 1 → d do
Mr,s ← Mr,s + T1r,i · Bi,j · T2j,s .
end for
for r = 1 → O(k log k), j = 1 → d do
Dr,j ← Dr,j + Sr,i · Bi,j .
end for
if need decomposition then
for s = 1 → O(k log k), j = 1 → d do
Ci,s := Ci,s + Bi,j · Rj,s .
end for
end if
end for
Compute the SVD of L = UL ΣL VL> .
Compute the SVD of N = UN ΣN VN> .
b = L† (UL U > M VN V > )k N † .
Compute X
L
N
b =U
bΣ
b Vb > .
Compute the SVD of X
if need decomposition then
b Vb > D, U ∗ = C U
b.
return V ∗ = Σ
else
b Vb > D.
return V ∗ = Σ
end if
end procedure
109
K.1
Definitions
Definition K.1 (Row-partition model [BWZ16]).
There are s machines, and the ith machine has a
P
matrix Ai ∈ Rni ×d as input. Suppose n = si=1 ni , and the global data matrix A ∈ Rn×d is denoted
as
A1
A2
··· ,
As
we say A is row-partitioned into these s matrices distributed in s machines respectively. Furthermore,
there is a machine which is a coordinator. The model only allows communication between the
machines and the coordinator. The communication cost in this model is the total number of words
transferred between machines and the coordinator. Each word is O(log(snd)) bits.
Definition K.2 (Arbitrary-partition model [BWZ16]). There are s machines, and the ith machineP
has a matrix Ai ∈ Rn×d as input. Suppose the global data matrix A ∈ Rn×d is denoted as
A = si=1 Ai . We say A is arbitrarily partitioned into these s matrices distributed in s machines
respectively. Furthermore, there is a machine which is a coordinator. The model only allows communication between the machines and the coordinator. The communication cost in this model is the
total number of words transferred between machines and the coordinator. Each word is O(log(snd))
bits.
K.2
Arbitrary-partition model, subspace, poly(k, log(d), log(n)) approximation
Definition K.3 (Arbitrary-partition model `1 -low rank approximation - rank-k subspace version).
Given matrix A ∈ Rn×d arbitrarily partitioned into s matrices A1 , A2 , · · · , As distributed in s machines respectively, and k ∈ N+ , the goal is to propose a protocol in the model of Definition K.2 such
that
1. Upon termination, the protocol leaves a matrix V ∗ ∈ Rk×d on the coordinator.
2. V ∗ satisfies that
min kA − U V ∗ k1 ≤ poly(k, log(d), log(n)) ·
U ∈Rn×k
min
U ∈Rn×k ,V ∈Rk×d
kA − U V k1 .
3. The communication cost is as small as possible
Theorem K.4. Suppose A ∈ Rn×d is partitioned in the arbitrary partition model (See Definition K.2). There is a protocol(in Algorithm 10) which solves the problem in Definition K.3 with
e
constant probability. Further, the communication complexity of the protocol is s(poly(k) + O(kd))
words.
Proof. Correctness. The correctness is shown by the proof of Theorem C.3 and (IV) of Lemma D.11.
b ∈ RO(k log k)×O(k log k) minimizes
Notice that X
min
rank −k X
kLXN − M kF .
which is
min
rank −k X
kT1 ARXSAT2 − T1 AT2 kF .
110
Algorithm 10 Arbitrary Partition Distributed Protocol
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
procedure ArbitraryPartitionDistributedProtocol(k,s,A)
A ∈ Rn×d was arbitrarily partitioned into s matrices A1 , · · · , As ∈ Rn×d distributed in s
machines.
Coordinator
Machines i
Chooses a random seed.
Sends it to all machines.
−−−−−−−−−>
Agrees on R, T2 which are fully
independent random Cauchy matrices.
Agrees on S, T1 which are random Cauchy
matrices with fully independent entries
e
from different rows, and O(d)-wise
independent variables from the same row.
Computes Li = T1 Ai R, Ni = SAi T2 .
Computes Mi = T1 Ai T2 .
Sends Li , Ni , Mi to the coordinator.
<−−−−−−−−−
s
s
P
P
Computes L =
Li , N =
Ni .
i=1
s
P
Computes M =
i=1
Mi .
i=1
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
Computes the SVD of L = UL ΣL VL> .
Computes the SVD of N = UN ΣN VN> .
b = L† (UL U > M VN V > )k N † .
Computes X
L
N
b to machines.
Sends X
−−−−−−−−−>
b =U
bΣ
b Vb > .
Computes the SVD of X
b Vb > SAi .
Computes Vi∗ = Σ
If need decomposition
b.
Ui∗ = Ai RU
∗
Sends Ui , Vi∗ to the coordinator.
Else
Sends Vi∗ to the coordinator.
Endif
<−−−−−−−−−
If need decomposition,
P
P
return V ∗ = si=1 Vi∗ , U ∗ = si=1 Ui∗ .
Else
P
return V ∗ = si=1 Vi∗ .
Endif
end procedure
b
According to the proof of Theorem C.3, ARXSA
gives an `1 rank-k poly(k, log(d), log(n))-approximation
>
∗
>
b =U
bΣ
b Vb , V = Σ
b Vb SA satisfies:
to A. Because X
min kA − U V ∗ k1 ≤ poly(k, log(d), log(n)) ·
U ∈Rn×k
111
min
U ∈Rn×k ,V ∈Rk×d
kA − U V k1 .
e
Communication complexity. Since the random seed generates O(kd)-wise
independent rane
dom Cauchy variables, the cost of line 5 is O(skd) bits. The size of Li , Ni and Mi are k 2 log2 k, k 2 log3 k
b
and k 2 log3 k words separately. So the cost of line 15 is O(sk 2 log3 k) words. Because the size of X
2
2
2
2
is O(k log k), the cost of line 22 is O(sk log k) words. line 30 needs skd words of communication.
e
Therefore, the total communication of the protocol is s(poly(k) + O(kd))
words.
K.3
Arbitrary-partition model, decomposition, poly(k, log(d), log(n)) approximation
Definition K.5 (Arbitrary-partition model `1 -low rank approximation - rank-k decomposition
version). Given matrix A ∈ Rn×d arbitrarily partitioned into s matrices A1 , A2 , · · · , As distributed in
s machines respectively, and k ∈ N+ , the goal is to propose a protocol in the model of Definition K.2
such that
1. Upon termination, the protocol leave matrices U ∗ ∈ Rn×k , V ∗ ∈ Rk×d on the coordinator.
2. U ∗ , V ∗ satisfies that
kA − U ∗ V ∗ k1 ≤ poly(k, log(d), log(n)) ·
min
U ∈Rn×k ,V ∈Rk×d
kA − U V k1 .
3. The communication cost is as small as possible.
Theorem K.6. Suppose A ∈ Rn×d is partitioned in the arbitrary partition model (See Definition K.2). There is a protocol(in Algorithm 10 with decomposition) which solves the problem in
Definition K.5 with constant probability. Further, the communication complexity of the protocol is
e
s(poly(k) + O(k(d
+ n))) words.
Proof. Correctness. The only difference from the protocol (without decomposition) in Section K.2
b . Notice that U ∗ V ∗ =
is that the protocol sends Ui . Thus, the coordinator can compute U ∗ = ARU
b
ARXSA.
According to the proof of Theorem C.3, U ∗ V ∗ gives a `1 rank-k poly(k, log(d), log(n))approximation to A.
Communication complexity. Since the size of Ui is kn words, the total communication is
e
s(poly(k) + O(k(d
+ n))) words.
K.4
Row-partition model, subspace, poly(k) log d approximation
Definition K.7 (Row-partition model `1 -low rank approximation - rank-k subspace version). Given
matrix A ∈ Rn×d row-partitioned into s matrices A1 , A2 , · · · , As distributed in s machines respectively, and k ∈ N+ , the goal is to propose a protocol in the model of Definition K.1 such that
1. Upon termination, the protocol leaves a matrix V ∗ ∈ Rk×d on the coordinator.
2. V ∗ satisfies that
min kA − U V ∗ k1 ≤ poly(k) log d ·
U ∈Rn×k
3. The communication cost is as small as possible
112
min
U ∈Rn×k ,V ∈Rk×d
kA − U V k1 .
Theorem K.8. Suppose A ∈ Rn×d is partitioned in the row partition model (See Definition K.1).
There is a protocol(in Algorithm 11 without decomposition) which solves the problem in Definition K.7 with constant probability. Further, the communication complexity of the protocol is s(poly(k)+
e
O(kd))
words.
Algorithm 11 Row Partition Distributed Protocol
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
procedure RowPartitionDistributedProtocol(k,s,A)
A ∈ Rn×d was row partitioned into s matrices A1 ∈ Rn1 ×d , · · · , As ∈ Rns ×d distributed in s
machines.
Coordinator
Machines i
Chooses a random seed.
Sends it to all machines.
−−−−−−−−−>
Agrees on R, T2 which are fully
independent random Cauchy variables.
Generates random Cauchy matrices
2
Si0 ∈ RO(k log k)×ni , Si , T1i ∈ RO(k log k)×ni .
Computes Yi∗ =
arg min kY Si0 Ai − Ai k1 .
Y ∈Rni ×O(k log k)
Yi∗ Si0 Ai .
Computes Bi =
Computes Li = T1i Bi R, Ni = Si Bi T2 .
Computes Mi = T1i Bi T2 .
Sends Li , Ni , Mi to the coordinator.
<
−
−
−
−
−
−
−
−
−
P
P
Computes L = Psi=1 Li , N = si=1 Ni .
Computes M = si=1 Mi .
Computes the SVD of L = UL ΣL VL> .
Computes the SVD of N = UN ΣN VN> .
b = L† (UL U > M VN V > )k N † .
Computes X
L
N
b to machines.
Sends X
−−−−−−−−−>
b =U
bΣ
b Vb > .
Computes the SVD of X
b Vb > Si Bi .
Computes Vi∗ = Σ
If need decomposition
b.
Computes Ui∗ = Bi RU
∗
∗
Sends Ui , Vi to the coordinator.
Else
Sends Vi∗ to the coordinator.
Endif
<−−−−−−−−−
If need decomposition
P
P
return V ∗ = si=1 Vi∗ , U ∗ = si=1 Ui∗ .
Else
P
return V ∗ = si=1 Vi∗ .
Endif
end procedure
113
2
Proof. Correctness. For convenience, we denote matrices B ∈ Rn×d , S ∈ RO(k log k)×n and T1 ∈
2
RO(k log k)×n as
B1
B2
B=
T1 = T11 T12 · · · T1s .
· · · S = S1 S2 · · · Ss
Bs
b ∈ RO(k log k)×O(k log k) actually minimizes
Notice that L = T1 BR, N = SBT2 , M = T1 BT2 . Thus, X
min
rank −k X
kT1 BRXSBT2 − T1 BT2 kF .
Also notice that B is just taking each row of A and replacing it with its nearest point in the row
b
span of S 0 A. According to the proof of Theorem C.6, BRXSB
gives a poly(k) log d `1 norm rank-k
>
∗
>
b =U
bΣ
b Vb , V = Σ
b Vb SB satisfies:
approximation to A. Since X
min kA − U V ∗ k1 ≤ poly(k, log(d), log(n)) ·
U ∈Rn×k
min
U ∈Rn×k ,V ∈Rk×d
kA − U V k1 .
e
Communication complexity. Since the R and T2 are O(kd)-wise
independent, line 6 needs
3
2
O(sW ) bits of communication. Line 16 needs O(sk log k) words. The cost of line 23 is O(sk 2 log2 k)
words. Line 32 needs skd words of communication. Therefore, the total communication of the
e
protocol is s(poly(k) + O(kd))
words.
K.5
Row-partition model, decomposition, poly(k) log d approximation
Definition K.9 (Row-partition model `1 -low rank approximation - rank-k decomposition version).
Given matrix A ∈ Rn×d row partitioned into s matrices A1 , A2 , · · · , As distributed in s machines
respectively, and a positive integer k < rank(A), the goal is to propose a protocol in the model of
Definition K.1 such that
1. Upon termination, the protocol leaves matrices U ∗ ∈ Rn×k , V ∗ ∈ Rk×d on the coordinator.
2. U ∗ , V ∗ satisfies that
kA − U ∗ V ∗ k1 ≤ poly(k) log d ·
min
U ∈Rn×k ,V ∈Rk×d
kA − U V k1 .
3. The communication cost is as small as possible.
Theorem K.10. Suppose A ∈ Rn×d is partitioned in the row partition model (See Definition K.1).
There is a protocol(in Algorithm 11 with decomposition) which solves the problem in Definition K.9
with constant probability. Further, the communication complexity of the protocol is s(poly(k) +
e
O(k(n
+ d))) words.
Proof. Correctness. The only difference is that the above protocol sends Ui . Thus, the coordinator
b . Notice that U ∗ V ∗ = BRXSB,
b
can compute U ∗ = BRU
according to the proof of Theorem C.6,
∗
∗
U V gives a poly(k) log d `1 norm rank-k approximation to A.
Communication complexity. Since the size of Ui is kn words, the total communication is
e
s(poly(k) + O(k(n
+ d))) words.
114
4e+04
4e+04
DZHZ06
Ours
`1 -norm cost
3e+04
2e+04
2e+04
2e+04
1e+04
5e+03
0e+00
20
40
60
80
100
n
120
140
160
180
200
Figure 1: The x-axis is n where A ∈ Rn×n , and the y-axis is kA0 − Ak1 where rank(A0 ) = k. This
figure shows the performance of both our algorithm and [DZHZ06] on input matrix A defined as
Equation (45).
L
Experiments and Discussions
In this section, we provide some counterexamples for the other heuristic algorithms such that, for
those examples, the heuristic algorithms can output a solution with a very “bad” approximation
ratio, i.e., nc , where c > 0 and the input matrix has size n × n. We not only observe that heuristic
algorithms sometimes have very bad performance in practice, but also give a proof in theory.
L.1
Setup
We provide some details of our experimental setup. We obtained the R package of [KK05, Kwa08,
BDB13] from https://cran.r-project.org/web/packages/pcaL1/index.html. We also implemented our algorithm and the r1-pca algorithm [DZHZ06] using the R language. The version of the
R language is 3.0.2. We ran experiments on a machine with Intel [email protected] CPU and 24G
memory. The operating system of that machine is Linux Ubuntu 14.04.5 LTS. All the experiments
were done in single-threaded mode.
L.2
Counterexample for [DZHZ06]
The goal is to find a rank k = 1 approximation for matrix A. For any ∈ [0, 0.5), we define
A ∈ Rn×n as
1.5+
n
0
0 0
A=
+
,
(45)
0
0
0 B
where B ∈ R(n−1)×(n−1) is all 1s matrix. It is immediate that the optimal cost is at most n1.5+ .
However, using the algorithm in [DZHZ06], the cost is at least Ω(n2 ). Thus, we can conclude,
115
4e+04
4e+04
BDB13
Ours
`1 -norm cost
3e+04
2e+04
2e+04
2e+04
1e+04
5e+03
0e+00
20
40
60
80
100
n
120
140
160
180
200
Figure 2: The x-axis is n where A ∈ Rn×n , and the y-axis is kA0 − Ak1 where rank(A0 ) = k. This
figure shows the performance of both our algorithm and [BDB13] on input matrix A defined as
Equation (46).
using algorithm [DZHZ06] to solve `1 low rank approximation problem on A cannot achieve an
approximation ratio better than n0.5− .
L.3
Counterexample for [BDB13]
The goal is to find a rank k = 1 approximation for matrix A. The input matrix A ∈ Rd×d for
algorithm [BDB13] is defined to be,
1.5
n
0
0 0
A=
+
,
(46)
0 0
0 B
where B ∈ R(n−1)×(n−1) is an all 1s matrix. It is immediate that the optimal cost is at most n1.5 .
Now, let us look at the procedure of [BDB13]. Basically, the algorithm described in [BDB13] is
that they first find a rank n − 1 approximation via a best `1 -fit hyperplane algorithm, then they
rotate it based on the right singular vectors of the rank n − 1 approximation matrix, and next they
recursively do the same thing for the rotated matrix which has only n − 1 columns.
When running their algorithm on A, they will fit an arbitrary column except the first column of
A. Without loss of generality, it just fits the last column of A. After the rotation, the matrix will
be an n × (n − 1) matrix:
1.5
n
0 ··· 0
√ 0
0
√n − 2 0 · · · 0
0
n − 2 0 ··· 0
.
···
·
·
·
·
·
·
·
·
·
·
·
·
√
0
n − 2 0 ··· 0
116
8e+04
7e+04
Kwak08
Ours
`1 -norm cost
6e+04
5e+04
4e+04
3e+04
2e+04
1e+04
0e+00
20
40
60
80
100
120
(n−1)/2
140
160
180
200
Figure 3: The x-axis is n where A ∈ Rn×n , and the y-axis is kA0 − Ak1 where rank(A0 ) = k. This
figure shows the performance of both our algorithm and [Kwa08] on input matrix A defined as
Equation (47). Algorithm [Kwa08] has two ways of initialization, which have similar performance
on matrix A.
Then, after the tth iteration for t < (n − 1), they
1.5
n
√ 0
0
√n − 2
0
n−2
···
√· · ·
n−2
0
This means that their algorithm will run on an
1.5
n
0
0
···
0
will get an n × (n − t) matrix:
0 ··· 0
0 ··· 0
0 ··· 0
.
· · · · · · · · ·
0 ··· 0
n × 2 matrix:
√ 0
√n − 2
n − 2
,
·
·
·
√
n−2
√
in the last iteration. Notice that n × n − 2 < n1.5 . This means that their algorithm will fit the
first column which implies that their algorithm will output a rank-1 approximation to A by just
fitting the first column of A. But the cost of this rank-1 solution is (n − 1)2 . Since the optimal cost
is at most n1.5 , their algorithm cannot achieve an approximation ratio better than n0.5 .
117
L.4
Counterexample for [Kwa08]
We show that the algorithm [Kwa08] cannot achieve an approximation ratio better than Θ(n) on
the matrix A ∈ R(2n+1)×(2n+1) defined as,
1.5
n
0 0
A = 0 B 0,
(47)
0
0 B
where B is an n × n matrix that contains all 1s. We consider the rank-2 approximation problem for
this input matrix A. The optimal cost is at most n1.5 .
We run their algorithm. Let x0 denote the initial random unit vector. Consider the sign vector
s ∈ {±1}2n+1 where each entry is the sign of the inner product between x0 and each column of A.
There are only three possibilities,
n
n
(1, {1} , {1} )
s = (1, {1}n , {−1}n ) .
(1, {−1}n , {1}n )
P2n+1
1.5
n
n
b/kuk2 =
Case
s = (1, {1}
b =
i=1 si · Ai = (n , n, · · · , n). Let u = u
√ I, 1.5
√ , {1}
√ ). Define√ u
>
u
b/( 3n ) = (1/ 3, 1/ 3n, · · · , 1/ 3n). Define matrix D to be A−uu A. Then, we can compute
D,
D = A − uu> A
1/3 3√1 n 1> 3√1 n 1>
1
1
= A − 3√1 n 1
3n B
3n B A
1
1
1
√
1
3n B
3n B
3 n
√
1.5
n1.5 √n
n >
>
n
0 0
1
1
3
3
3
1
1
= 0 B 0 − n3 1
3B
3B
n
1
1
0
0 B
31
3B
3B
√
√
n >
n >
1.5
2n /3 − 3 1
− 3 1
n
2
= −31
− 31 B .
3B
n
1
2
−31
−3B
3B
> A. Let w denote another
Now we need to take the linear combination of columns D = A − uuP
2n+1
2n+1
sign vector {−1, +1}
. Then let v denote the basis vector, v =
i=1 Di . There are three
n
n
possibilities, Case I(a), if w = (1, {1} , {1} ), then v is the all 0 vector. Using vector v to interpolate
each column of D, the cost we obtain is at least 2n2 . Case I(b), if w = (1, {1}n , {−1}n ), then
v = (2n1.5 /3, {2/3}n , {−4/3}n ). We also obtain at least 2n2 cost if we use that v to interpolate
each column of D. Case I(c), if w = (1, {−1}n , {−1}n ), then v = (0, {−2/3}n , {−2/3}n ). The cost
is also at least 2n2 .
Case II, s = (1, {1}n , {−1}n ). Define 1 to be a length
vector.
Define
b =
√
√
√u
√ n1.5all 1s column
P2n+1
n
1.5
n
n
b/kuk2 = u
b/( 3n ) = (1/ 3, {1/ 3n} , {−1/ 3n}n ).
i=1 si ·Ai = (n , {n} , {−n} ). Let u = u
>
Define (2n + 1) × (2n + 1) matrix D to be A − uu A. Then, we can compute D,
118
D = A − uu> A
1
√
1/3
1> − 3√1 n 1>
3 n
1
1
− 3n
B
= A − 3√1 n 1
A
3n B
1
1
− 3√1 n 1 − 3n
B
B
3n
√
√
1.5
n
0 0
n1.5 /3 3n 1> − 3n 1>
1
= 0 B 0 − n3 1
− 31 B
3B
1
1
n
0
0 B
− 3 1 −3B
3B
√
√
2n1.5 /3 − 3n 1> 3n 1>
2
1
.
= − n3 1
3B
3B
n
1
2
31
3B
3B
Similarly to the previous case, we can also discuss three cases.
Case III, s = (1, {1}n , {−1}n ). Define 1 to be a length
all 1s column
vector.
u
b√=
√
√ Define
√ n 1.5
P2n+1
n , {−1/ 3n}n ).
1.5 , {−n}n , {−n}n ). Let u = u
3n
)
=
(1/
3,
{−1/
3n}
s
·A
=
(n
b
/kuk
=
u
b
/(
i i
2
i=1
Define matrix D to be A − uu> A. Then, we can compute D,
D = A − uu> A
1/3
− 3√1 n )1> − 3√1 n 1>
1
1
= A − − 3√1 n 1
3n B
3n B A
1
1
− 3√1 n 1
3n B
3n B
√
√
1.5
n
0 0
n1.5 /3 − 3n 1> − 3n 1>
1
1
= 0 B 0 − − n3 1
3B
3B
n
1
1
0
0 B
−31
3B
3B
√
√
2n1.5 /3 3n 1> 3n 1>
2
1
.
= n3 1
3B
3B
n
1
2
31
3B
3B
Similarly to the previous case, we can also discuss three cases.
L.5
Counterexample for [KK05]
We show that there exist matrices such that the algorithm of [KK05] cannot achieve an approximation ratio better than Θ(n). Their algorithm has two different ways of initialization. We provide
counterexamples for each of the initialization separately.
Random vector initialization We provide a
c
n
0
0
0
A=
· · · · · ·
0
0
counterexample matrix A ∈ Rn×n defined as,
··· 0
··· 0
+ I,
(48)
· · · · · ·
··· 0
where c ≥ 2. Consider the rank-1 approximation problem for this matrix A. The optimal cost is at
most n − 1.
119
2e+09
4e+04
KK05
Ours
KK05
Ours
4e+04
2e+09
3e+04
`1 -norm cost
`1 -norm cost
2e+04
1e+09
5e+08
2e+04
2e+04
1e+04
5e+03
0e+00
20
40
60
80
100
n
120
140
160
180
0e+00
20
200
40
60
80
100
n
120
140
160
180
200
Figure 4: The x-axis is n where A ∈ Rn×n , and the y-axis is kA0 − Ak1 where rank(A0 ) = k.
Algorithm [KK05] has two ways of initialization. The left figure shows the performance of both
our algorithm and [KK05] with random vector initialization on input matrix A defined as Equation
(48). The right figure shows the performance of both our algorithm and [KK05] with top singular
vector initialization on input matrix A defined as Equation (50).
Run their algorithm. The starting vectors are u(0) ∼ N (0, I) and v(0) ∼ N (0, 1). We define
two properties for a given vector y ∈ Rn . Property I is for all i ∈ [n], |yi | ≤ n/8, and Property II is
there exist half of the i such that |yi | ≥ 1/2. We can show that with probability 1 − 2−Ω(n) , both
u(0) and v(0) satisfy Property I and II. After 1 iteration, we can show that u(1)1 = v(1)1 = 0
Now let us use column vector u(0) ∈ Rn to interpolate the first column of matrix A. We simplify
u(0) to be u. Let ui denote the i-th coordinate of vector u, ∀i ∈ [n]. Let A1 denote the first column
of matrix A. We define α = v(1)1 = arg minα kαu(0) − A1 k1 . For any scalar α, the cost we pay on
the first column of matrix A is,
c
|α · u1 − n | +
n
X
|αui | ≥ |α · u1 − nc | +
i=2
n
1
n
|α · | ≥ |α · u1 − nc | + |α|.
2
2
4
(49)
Notice that c ≥ 2. Then |α · u1 − nc | + n4 |α| ≥ |α · n8 − nc | + n4 |α|. If α ≤ 0, then the cost is
minimized when α = 0. If α ∈ [0, 8nc−1 ], the cost is minimized when α = 0. If α ≥ 8nc−1 , the cost
is minimized when α = 8nc−1 . Putting it all together, to achieve the minimum cost, there is only
one choice for α, which is α = 0. The optimal cost is at least nc .
Then after T iterations(for any T ≥ 1), u(T )1 = v(T )1 = 0. Thus, we always pay at least nc
cost on the first entry.
Therefore, their algorithm cannot achieve any approximation ratio better than nc−1 . Because
c ≥ 2, we complete the proof.
Top singular vector initialization The counterexample input matrix A ∈ Rn×n is defined as,
n 0
A=
,
(50)
0 B
where matrix B ∈ R(n−1)×(n−1) contains all 1s. Consider the rank-1 approximation problem for this
matrix A. The optimal cost is at most n. Run their algorithm. The starting vectors u(0) and v(0)
will be set to (1, 0, · · · , 0) ∈ Rn . After T iterations(for any T > 0), the support of u(T )(resp. v(T ))
120
is the same as u(0)(resp. v(T )). Thus, the cost is at least kBk1 = (n − 1)2 . Therefore, we can
conclude that their algorithm cannot achieve any approximation ratio better than (n−1)2 /n = Θ(n).
L.6
Counterexample for all
For any ∈ (0, 0.5) and γ > 0, we construct the input
2+γ
n
0
1.5+
0
n
A=
0
0
0
0
matrix A ∈ R(2n+2)×(2n+2) as follows
0 0
0 0
,
B 0
0 B
where B is n × n all 1s matrix. We want to find a rank k = 3 solution for A. Then any of those four
heuristic algorithms [KK05, DZHZ06, Kwa08, BDB13] is not able to achieve better than nmin(γ,0.5−)
approximation ratio. We present our main experimental results in Figure 5. Both [KK05] and
[Kwa08] have two different ways of initialization. In Figure 5 we use KK05r(resp. Kwak08r) to
denote the way that uses random vector as initialization, and use KK05s(resp. Kwak08s) to denote
the way that uses top singular vector as initialization. Figure 5(a) shows the performance of all
the algorithms and Figure 5(b) presents the running time. The `1 residual cost of all the other
algorithm is growing much faster than our algorithm. Most of the algorithms (including ours) are
pretty efficient, i.e., the running time is always below 3 seconds. The running time of [BDB13, KK05]
is increasing very fast when the matrix dimension n is growing.
In Figure 5(a), the cost of KK05r at {82, · · · , 142} is in [105 , 106 ], at {162, · · · , 302} is in
6
[10 , 107 ], and at {322, · · · , 402} is in [107 , 108 ]. In Figure 5(b), the time of KK05r at {382, 482} is
64s and 160s. The running time of BDB13 at {82, · · · , 222} is between 1 minute and 1 hour. The
running time of BDB13 at {242, · · · , 322} is between 1 hour and 20 hours. The running time of
BDB13 at {342, · · · , 402} is more than 20 hours.
L.7
Discussion for Robust PCA [CLMW11]
A popular method is robust PCA [CLMW11], which given a matrix A, tries to find a matrix L for
which λkA − Lk1 + kLk∗ is minimized, where λ > 0 is a tuning parameter and kLk∗ is the nuclear
norm of L. This is a convex program, but it need not return a low rank matrix L with relative
error. As a simple example, suppose λ = 1, and the n × n matrix A is a block-diagonal matrix of
rank k and n = k2 (b + 1). Further, the first k/2 blocks are b × b matrices of all 1s, while the next
k/2 blocks are just a single value b on the diagonal.
Then the solution to the above problem may return L to be the first k/2 blocks of A. The total
cost of λkA − Lk1 + kLk∗ is (k/2)b + (k/2)b = kb.
Also, the solution to the above problem may return L to be A, which has cost 0 + kAk∗ = kb.
Because this solution has the same cost, it means that their algorithm might output a rank-k
solution, and also might output a rank-k/2 solution.
Therefore, the relative error of the output matrix may be arbitrarily bad for `1 -low rank approximation.
√
We also consider the following example. Suppose λ = 1/ n, and let n × n matrix A denote the
Hadamard
matrix
Hn . Recall
that the
Hadamard matrix Hp of size p × p is defined recursively :
√
Hp/2 Hp/2
+1 +1
with H2 =
. Notice that every singular values of A is n. We consider
Hp/2 −Hp/2
+1 −1
the objective function λkA − Lk1 + kLk∗ .
121
Then the solution to the above problem may return L to be the first n/2 rows of A. The total
√
√
cost of λkA − Lk1 + kLk∗ is (1/ n)n2 /2 + (n/2) n = n1.5 . Also, the solution to the above problem
√
may return L to be A, which has cost 0 + kAk∗ = n n = n1.5 . For any i, if the solution takes
√
√
i rows of A, the cost is (1/ n)(n − i)n + i n = n1.5 . Because this solution has the same cost,
it means that their algorithm might output a rank-n solution, and also might output a rank-n/2
solution. Therefore, the relative error of the output matrix may be arbitrarily bad for `1 -low rank
approximation.
M
Acknowledgments
The authors would like to thank Alexandr Andoni, Saugata Basu, Cho-Jui Hsieh, Daniel Hsu, Chi
Jin, Fu Li, Ankur Moitra, Cameron Musco, Richard Peng, Eric Price, Govind Ramnarayan, James
Renegar, and Clifford Stein for useful discussions. The authors also thank Jiyan Yang, Yinlam
Chow, Christopher Ré, and Michael Mahoney for sharing the code.
122
8e+04
Ours
BDB13
KK05r
KK05s
Kwak08r
Kwak08s
DZHZ06
7e+04
`1 -norm cost
6e+04
5e+04
4e+04
3e+04
2e+04
1e+04
0e+00
0
50
100
150
200
2n +2
250
300
350
400
300
350
400
(a)
50
Ours
BDB13
KK05r
KK05s
Kwak08r
Kwak08s
DZHZ06
40
Time
30
20
10
0
0
50
100
150
200
2n +2
250
(b)
Figure 5: Let A be (2n + 2) × (2n + 2) input matirx. (a) shows the performance of all the algorithms
when the matrix dimension is growing. The x-axis is n, and the y-axis is kA0 − Ak1 where A0 is
the rank-3 solution output by all the heuristic algorithms and also ours. The `1 residual cost of
all the other algorithms is growing much faster than ours, which is consistent with our theoretical
results. (b) shows the running time (in seconds) of all the algorithms when the matrix dimension
n is growing. The x-axis is n and the y-axis is time (seconds). The running time of some of the
algorithms is longer than 3 seconds. For most of the algorithms (including ours), the running time
is always less than 3 seconds.
123
| 8cs.DS
|
DLPaper2Code: Auto-generation of Code from Deep Learning Research Papers
Akshay Sethi ∗
IIIT Delhi
Anush Sankaran, Naveen Panwar, Shreya Khare, Senthil Mani
arXiv:1711.03543v1 [cs.LG] 9 Nov 2017
IBM Research, India
Abstract
With an abundance of research papers in deep learning, reproducibility or adoption of the existing works becomes a
challenge. This is due to the lack of open source implementations provided by the authors. Further, re-implementing research papers in a different library is a daunting task. To
address these challenges, we propose a novel extensible approach, DLPaper2Code, to extract and understand deep learning design flow diagrams and tables available in a research
paper and convert them to an abstract computational graph.
The extracted computational graph is then converted into execution ready source code in both Keras and Caffe, in realtime. An arXiv-like website is created where the automatically generated designs is made publicly available for 5, 000
research papers. The generated designs could be rated and
edited using an intuitive drag-and-drop UI framework in a
crowdsourced manner. To evaluate our approach, we create a
simulated dataset with over 216, 000 valid design visualizations using a manually defined grammar. Experiments on the
simulated dataset show that the proposed framework provide
more than 93% accuracy in flow diagram content extraction.
Introduction
The growth of deep learning (DL) in the field of artificial intelligence has been astounding in the last decade with about
35, 800 research papers being published since 20161 . Keeping up with the growing literature has been a real struggle
for researchers and practitioners. In one of the recent AI
conferences, NIPS 2016, the maximum number of papers
submitted (∼ 685/2500) were in the topic, “Deep Learning
or Neural Networks”. However, a majority of these research
papers are not accompanied by their corresponding implementations. In NIPS 2016, only 101/567 (∼ 18%) papers
made their source implementation available2 . Implementing
research papers takes at least a few days of effort for soft∗
Akshay Sethi interened at IBM Research, India during this
work.
Copyright c 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
1
https://scholar.google.co.in/scholar?as_
sdt=1,5\&q=\%22deep+learning\%22\&hl=en\&as_
ylo=2016\&as_vis=1
2
https://www.kaggle.com/benhamner/
nips-papers
ware engineers assuming that they have limited knowledge
in DL (Sankaran et al. 2011).
Another major challenge is the availability of various libraries in multiple programming langauges to implement
DL algorithms such as Tensorflow (Abadi et al. 2016),
Theano (Bastien et al. 2012), Caffe (Jia et al. 2014),
Torch (et al 2011), MXNet (Chen 2015), DL4J (Gibson
2015), CNTK (Seide and Agarwal 2016) and wrappers such
as Keras (Chollet and others 2015), Lasagne (Dieleman
2015), and PyTorch (Chintala 2016). The public implementations of the DL research papers are available in various libraries offering very little interoperability or communication
among them. Consider a use-case for a researcher working
in “image captioning”, where three of the highly referred research papers for the problem of image captioning3 are:
1. Show and Tell (Vinyals et al. 2015): Original implementation available in Theano; https://github.com/
kelvinxu/arctic-captions
2. NeuralTalk2 (Karpathy and Fei-Fei 2015): Original implementation available in Torch; https://github.
com/karpathy/neuraltalk2
3. LRCN (Donahue et al. 2015): Original implementation available in Caffe; http://jeffdonahue.com/
lrcn/
As the implementations are available in different libraries,
a researcher cannot directly combine the models. Also, for
a practitioner having remaining of the code-base in Java
(DL4J) directly leveraging either of these public implementations would be daunting. Thus, we highlight two highly
overlooked challenges in DL:
1. Lack of public implementation available for existing research works and thus, the time incurred in reproducing
their results
2. Existing implementations are confined to a single (or few)
libraries limiting portability into other popular libraries
for DL implementation.
We observed that most of the research paper explains the
DL model design either through a figure or a table. Thus,
3
https://competitions.codalab.org/
competitions/3221#results
Figure 1: The architecture of the proposed creative system that extracts and understands the flow diagram from a deep learning
research paper and generates an execution ready deep learning code in two differcaffent platforms: Keras and Caffe.
in this research we propose a novel algorithm that automatically parses a research paper to extract the described DL
model design. The design is represented as an abstract computational graph which is independent of the implemenation library or language. Finally, the source code is generated in multiple libraries from this abstract computational
graph of the DL design. The results are shown by automatically generating the source code of 5, 000 arXiv papers both
in CAFFE (prototxt) and Keras (python). However, evaluating the generated source code is debatable due to the lack
of ground truth. To overcome this challenge, we simulated
a large image dataset of 216, 000 valid DL model designs
in both Caffe and Keras. To generate DL visualizations, we
manually defined a grammar for DL models. As these visualizations are highly varying, they are comparable to the
figures present in research papers. Thus, the major research
contributions are:
1. a technique to automatically understand a DL model design by parsing the figures and tables in a research paper,
2. generate source code both in both Keras and Caffe from
the abstract computation graph of a DL design,
3. automatically generate design for 5, 000 arXiv papers and
build a UI system for editing them the crowdsourced way,
4. on a simulated dataset of 216, 000 DL model visualizations using a manually defined grammar, evaluate the proposed approach to achieve more than 95% accuracy.
The rest of the paper is organized as follows: Section 2
explains the entire proposed approach for auto generation of
DL source code from research paper, Section 3 talks about
the simulated dataset and its experimental performance of
the individual components of the proposed approach, Section 4 discusses the experimental results on 5, 000 arXiv DL
papers, and Section 5 concludes this paper with some discussion on our future efforts.
Proposed Approach
Consider a state-of-art paper DL paper (Szegedy et al. 2017)
published in AAAI 2017, which explains the DL design
model through figures, as shown in Figure 3(a). Similarly, in
the AAAI 2017 paper by (Parkhi et al. 2015), the DL model
design was explained using a table. Thus, given the PDF
of a research paper in deep learning, the proposed DLPaper2Code architecture consists of five major steps, as shown
in Figure 1: (i) Extract all the figures and tables from a research paper. Handling the figure and table content are done
independently, although the follow a similar pipeline, (ii) As
there could be other descriptive image and results tables in
a paper, classify each figure or table whether it contains a
DL model design. Also, perform a fine grained classification on the type of figure or table used to describe the DL
model, (iii) Extract the flow and the text information from
the figures and tables, independently, (iv) Construct an abstract computational graph which is independent of the implementation library, and (v) generate source code in both
Caffe and Keras from the computational graph.
Characterizing Research Papers
We observed that in a research paper the DL design is mostly
explained using figures or tables. Thus, it is our assertion
that by parsing the figure, as an image, and the table content, the respective novel DL design could be obtained. The
primary challenges with the figures in research papers is that
the DL design figures typically do not follow any definition
and show extreme variations. Similarly, tables can have different structures and can entail different kind of information.
We manually observed more than 30, 000 images from research papers and characterized the DL model deisgn images into five broad categories, as shown in Figure 2. The
five types are: (i) Neurons plot: the classical representation
of a neural network with each layer having circular nodes
inside them, (ii) 2D Box: each hidden layer is represented
as a 2D rectangular box, (iii) Stacked2D Box: each layer
(a) Neurons Plot
(b) 2D Box
(c) Stacked2D Box
(d) 3D Box
(e) Pipeline plot
Figure 2: Characterizing the DL model designs available in research papers and grouping them into five different categories.
is represented as a stack of 2D rectangular boxes, describing the depth of the layer, (iv) 3D Box: each hidden layer
is represented as a 3D cuboid structure, and (v) Pipeline
plot: along with the DL model design, the entire pipeline and
mostly some intermediate results of image/ text is shown as
well. Similarly, based on the representation, tables can be
classified as, (i) row-major table: where the DL model design flows along the row (Springenberg et al. 2014), and (ii)
where the DL model design flows along the column (Parkhi
et al. 2015). It is essential to account for these variations in
the proposed pipeline, as they indicate the DL design flow
represented in the paper. Following this assumption, the proposed approach does not identify a DL design flow that is
neither in a table nor in a figure.
Extracting Figures and Tables
Extracting visual figures from a PDF document, especially
a scholarly report is a well studied problem (Choudhury and
Giles 2015). Common challenges includes extracting vector
images as they are embedded in the PDF document and extracting a large figure as a whole instead of multiple figures
of its components. To this end, we have used a publicly existing tool called PDFFigures 2.04 (Clark and Divvala 2016)
for extracting a list of figures from a scholarly paper. However, none of the existing open source tools maintain the table structure that is essential for us. Thus, we built a PDF
table extraction tool using PDFMiner5 and Poppler-utils6 .
Poppler-utils provide high level information about the document such as the text dump, while using PDFMiner, certain
low level document details such as vertical line spacing are
obtained. The table structure, along with the table caption,
is retrived by building the heuristics over the horizontal and
vertical line spacing.
Figure and Table Classification
The aim is to classify and retrive only those figures and
tables in a research paper that contains a DL design flow.
Futher, a fine-grained classifier is required to classify the
figure into one of the identified five broad categories and
classify the table as a row-major or column-major flow.
In case of figures, the classifier is trained to perform the
prediction using the architecture shape and the flow. For example, figures having result graphs and showing sample images from dataset has different shape compared to an architecture flow diagram. All the figures are resized to 224×224
4
https://github.com/allenai/pdffigures2
https://euske.github.io/pdfminer/
6
https://poppler.freedesktop.org/
5
and 4, 096 features (fc2) are extracted from a fully connected
layer of a popular deep learning model VGG19 (Simonyan
and Zisserman 2014) pre-trained on ImageNet dataset. We
have two classification levels: (i) Coarse classifier: a binary
neural network (NNet) classifier trained on fc2 features to
classify if the figure contains a DL model or not, and (ii)
Fine-grained classifier: a five class neural network classifier trained on fc2 features to identify the type of DL design, only for those figures classified positive by the coarse
classifier. Having a sequence of two classifiers provided better performance as compared to a single classifier with six
classes (sixth class being no DL design flow).
In case of tables, a bag-of-words model is built using keywords from the caption text as well as the table text. A cosine distance based classifier is used to identify if there is
a DL design flow in the given table as compared to tables
containing results. Further based on the number of rows and
columns in the table, as extracted in the previous section, the
table is classified as a row-major or column-major flow.
Content Extraction from Figure
Content extraction from figures has two major steps: (i) Flow
detection to identify the nodes and the edges, and (ii) OCR
to extract the flow content. Identifying the flow is the challenges, as there is a huge variation in the type of DL design
flow diagrams. In this section, we explain the details of the
approach for a 2D Box type, as shown in Figure 3, while
similar approach could be extended to other types, as well.
Flow detection involves identifying the nodes first, followed
by the edges connecting the nodes. As the image is usually
of high resolution and quality, they are directly binarized using an adaptive Gaussian thresholding and a Canny edge detection approach is used to identify all the lines. An iterative
region grown algorithm is adopted to identify closed countours in the figure, as they represent the nodes as shown in
Figure 3(b). All the detected nodes are masked out from the
figure and the contour detection algorithm is applied again to
detection the edges, as shown in Figure 3(d). The direction
of the edge flow is obtained by analyzing the pixel distribution within each edge contour. The node and edge contours
are then sorted based on the location and direction to obtain
the flow of the entire DL model design. As shown in Figure 3, the proposed approach could also handle branchings
and forking in a design flow diagram.
Once the flow is extracted, the text in each node/ layer
is obtained through OCR using Tesseract7 .Based on our
manual observation, we assume that the a layer description
7
https://github.com/tesseract-ocr/
(a) Original Image
(b) Detected Layers
(c) Layers Masked Out
(d) Detected Edges
Abstract Computational Graph
{
"layer_type" : "Pooling2D",
"layer_name" : "MaxPool1",
"layer_params" : {
"function" : "MAX",
"trainable" : true,
"stride_row" : 2,
"kernel_col" : 2,
"kernel_row" : 2,
"stride_col" : 2,
"border_mode" : "valid"
}
“from_layers”: [“Conv1”]
Keras Python
Caffe Protobuf
layer {
name: "MaxPool1"
type: "Pooling"
bottom: "Conv1"
top: "MaxPool1"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}
MaxPooling2D(
pool_size=(2, 2),
strides=(2,2),
padding='valid',
name=’MaxPool1’
)(Conv1)
}
Figure 5: An illustration for a Pooling2D layer showing the
rule base of the inference engine, converting the abstract
JSON format into Caffe protobuf and Keras python code.
Figure 3: Illustration of the proposed flow detection approach from complex figures (Szegedy et al. 2017) (AAAI
2017) involving (b) node/ layer detection, and (d) edge detection.
Figure 4: An example table showing the DL design flow as
explained in tabular format in (Parkhi et al. 2015).
would be available within the detected node. A dictionary of
possible DL layer names is created to perform spell correction of the extracted OCR text.
Content Extraction from Table
In a row major table, every row corresponds to a layer in the
DL design flow, as shown in Figure 4. Similarly, in a column
major table, every column corresponds to a layer along with
other parameters of the layer. The layer name is extracted by
matching it with a manually created dictionary. Further, the
parameters are extracted by mapping the corresponding row
or column header with a pre-defined list of parameter names
corresponding to the layer. Thus, sequentially the entire DL
design flow is extracted from a table.
Generating Source Code
Overall, after detecting DL design flow, an abstract computational graph is represented in JSON format, as shown in
Figure 5. Two rule based converters are written to convert
the abstract computational graph extracted in the previous
step to either (i) Keras code (Python) or (ii) Caffe protobuf
format (prototxt). An inference engine acts as the convertor
to map the abstract computational graph to the grammar of
the platform. The inference engine consists of a comprehensive list of templates and dictionaries built manually for both
Keras and Caffe. Template based structures transfer each
component from the abstract representation to a platform
specific structure using the dictionary mappings. Further, another set of templates, consisting of a set of assertions, are
designed for translating each layer’s parameters. Further, the
inference engine is highly flexible allowing easy extension
and addition of new layer definitions. An example of the inference engine’s dictionary mapping for a Pooling2D layer
is shown in Figure 5.
Thus for a given research paper, by processing both the
figure and table content, the DL model design flow is obtained which is converted to execution ready code in both
Keras and Caffe.
Evaluation on Simulated Dataset
The aim of this process is to simulate and generate ground
truth deep learning designs and their corresponding flow
visualizations figures. Thus, the proposed pipeline of DL
model design could be quantitatively evaluated. To this end,
we observed that both Keras and Caffe have an in-built visualization routine for a DL design model. Further both Keras
and Caffe have their internal DL model validator and a visualization can be exported only when the simulated design is
deemed valid by the validator.
Grammar for DL Design Generation
To be able to generate meaningful and valid DL design models, we manually defined a grammar for the model flow as
well as for the hyper-parameters. We considered 10 unique
layers for our dataset simulation - {Conv2D, MaxPool2D,
AvgPool2D} for building convolution neural network like
architectures, {Embed, SimpleRNN, LSTM} for building recurrent neural network like architectures, {Dense, Flatten,
Dropout, Concat} as the core layers. The use of Concat enables our designed models to be non-sequential as well as
with a combination of recurrent and convolution architectures. This allows us to create random, complex, and highly
varying DL models. Also, RNN and LSTM layers have an additional binary parameter of return seq, which when set true
returns the output of every hidden cell, otherwise, returns the
output of only the last hidden cell in the layer. Table 1 explains the proposed grammar for generating DL design models. The grammar defines the set of all possible next layers
for a given current layer. This is determined by the shape
of the tensor flowing through each of the layer’s operation.
For example, a Dense layer strictly expect the input to be a
Current
Layer
Input
Dense
Conv2D
Flatten
Dropout
MaxPool
AvgPool
Concat
RNN
RNN
(seq)
LSTM
LSTM
(seq)
Dense Conv2D Flatten Dropout MaxPool AvgPool Concat Embed RNN RNN LSTM LSTM
(seq)
(seq)
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
Same as previous layer
4
4
4
4
4
4
4
4
4
4
4
4
If input is one dimensional, same as Dense layer; else same as previous layer
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
4
Table 1: The proposed grammar for creating valid deep learning design models defining the list of possible next layers for a
given current layer.
Layer
Dense
Dropout
Conv2D
MaxPool
AvgPool
Embed
SimpleRNN
LSTM
InputData
Hyper-parameters
#nodes - {[5:5:500]}
probability - {[0:0.1:1]}
#filters - {[16:16:256]}
filter size - {[1:2:11]}
stride - {[2:1:5]}
filter size - {[1:2:11]}
stride - {[2:1:5]}
filter size - {[1:2:11]}
embed size - {64, 100, 128, 200}
vocab - {[10000, 20000, 50000, 75000]}
#units - {[3:1:25]}
#nodes - {[3:1:25]}
MNIST - {28, 28, 1}
CIFAR-10 - {32, 32, 3}
ImageNet - {224, 224, 3}
IMDB Text
Table 2: The set of hyper-parameter options considered for
each layer in our simulated dataset generation. The parameter value [a:b:c] means a list of values from a to c in steps of
b.
vector of shape n × 1. Thus, the Dense cannot appear after a Conv2D layer without the presence of a Flatten layer.
The proposed grammar further includes the set of possible
values for each hyper-parameter of a layer, as explained in
Table 2. While hyper-parameter values beyond the defined
bounds are possible, the table values indicate the assumed
set of values in the model simulation process.
Simulated Dataset
A model simulation starts with an Input layer, where
there are four possible options - MNIST, CIFAR, ImageNet,
IMDBText. From the set of all possible next layers for the
given Input layer, a completely random layer is decided. For
the given next layer, a random value is picked for every pos-
Observation
#Points
Naive Bayes
Decision Tree
Logistic Regression
RDF
SVM (RBF Kernel)
Neural Network
Train
195, 296
98.29%
100%
100%
100%
100%
100%
Validation
48, 824
98.30%
99.57%
99.98%
99.72%
99.89%
99.93%
Test
48, 824
98.39%
99.55%
99.99%
99.68%
99.83%
99.94%
Table 3: The performance of various binary classifiers to
distinguish KerasCaffeVisulizations vs. other often occuring
images in research papers.
sible hyper-parameter. For example, for MNIST being the input layer, Conv2D could be picked as the random next layer.
Then, for Conv2D the hypar-parameters are determined randomly as 32 filters, 5 × 5 filter size with a stride of 2. The
model design always ends with a Dense layer with number
of nodes equal to the number of classes of the corresponding
Input layer.
The number of layers in between the Input layer and the
final Dense layer denotes the depth of the DL model. For
our simulation, we generated 3, 000 DL models for each
depth varied between 5 and 40, creating a total dataset of
108, 000 models. Each model contains the Keras JSON representation, Keras image visualization, Caffe protobuf files,
and Caffe image visualization, resulting in a total of 216, 000
DL model design visualizations. These models are valid by
construct since they follow a well-defined grammar. However, these models need not be the best from an execution
perspective, or with respect to their training performance.
Figure Type Classification Performance
In this experiment, a binary NNet classifier with two hidden layers of size [1024, 256] is trained on fc2 features of
VGG19 model to differentiate 216, 000 simulated DL visualizations from a set of 28, 120 other kind of diagrams often
available in research papers (scraped from PDF). The whole
dataset is split as 60% for training, 20% validation, and 20%
for testing, making it a total of 195, 296 images for training and 48, 824 images for testing. The performance of the
NNet classifier is compared with six different classifiers as
shown in Table 3. As it can be observed most of the classifier provide a classification accuracy of 100%, showing that
from a set of figures obtained from a research paper, it would
be possible to distinguish only the deep learning design flow
diagrams. All the classifiers use the default parameters as
provided the scikit-learn package.
(a) Performance on Keras Images
(ii) Edge Extraction
(i) Layer Extraction
100%
Computational Graph Extraction Performance
(b) Performance on Caffe Images
(ii) Edge Extraction
(i) Layer Extraction
100%
In this experiment, the performance of flow and content extraction from the 216, 000 Keras and Caffe visualizations is
evaluated against the ground truth. By performing OCR, on
the extracted flow, the unique layer names are obtained and
two detection accuracies are reported,
93.9%
93.3%
1. blob (or layer) detection accuracy: evaluates the performance of blob detection and layers identified using OCR
and is computed as the ratio of correct blobs detected per
model (in percent)
2. edge detection accuracy: evaluates the performance of the
detected flow and is computed as the ratio of correct arrows detected per model (in percent)
Figure 6 is the box plot showing the performance for the the
proposed figure extraction pipeline in both Keras and Caffe.
As it can be observed, the proposed pipeline provides 100%
accuracy in layer extraction and more than 93% accuracy in
extracting the edges. As the edges can be curved and can be
of any length, even connecting the first with the last layer,
the variations caused a reduction in performance. Further
details on extracting flow information Keras and Caffe visuazliation and also for additional results, kindly refer to the
supplementary material.
Results on Deep Learning Scholarly Papers
The first 5, 000 papers were downloaded from arXiv.org using “deep learning” as the input query. 30, 987 figures were
extracted from these 5, 000 downloaded papers, out of which
28, 120 figures did not contain a DL design flow while the
remaining 2, 867 contained. These represent the usual figures that are found in a deep learning research paper that
does not contain a design flow.
Figure Type Classification Accuracy
To evaluate the coarse level binary classification, a 2 hidden
layer NNet was trained on the fc2 features obtained from the
30, 987 images extracted from research papers. The whole
dataset is split as 60% for training, 20% validation, and 20%
for testing and the results are computed for seven different
classifiers as shown in Table 4.
Further, to evaluate the fine level, five-class, figure type
classification, the 2, 871 DL design flow diagrams were
manually labelled. The distribution of figures were as follows: (i) Neurons plot: 587 figures, (ii) 2D box: 1, 204, (iii)
Stacked2D box: 408, (iv) 3D box: 562, and (v) Pipeline plot:
Figure 6: Box plots showing the performance accuracy of
flow detection in Keras and Caffe visualizations.
Observation
#Points
Naive Bayes
Decision Tree
Logistic Regression
RDF
SVM (RBF Kernel)
Neural Network
Train
18, 592
77.29%
99.96%
99.96%
99.96%
99.96%
99.96%
Validation
6, 197
64.39%
76.67%
86.17%
83.72%
86.89%
87.93%
Test
6, 198
62.56%
74.35%
85.27%
82.94%
85.25%
86.25%
Table 4: The performance of coarse level binary classifier to
distinguish DL design flow figures from other figures that
usually appear in a research paper.
110. A 60 − 20 − 20 train, validation, and test split is performed to train the NNet classifier in comparison with the
six other classifier to perform this five class classification.
The results are table in Table 5. The results show that even on
highly varying DL flow design images, identifying the type
of DL flow is more than 70% accurate. For more details on
the experimental analysis, please refer to the supplementary
material.
Crowdsourced Improvement of Extracted Designs
Using the proposed DLPaper2Code framework, we extracted the DL design models for all the 5, 000 downlaoded
papers. However, quantitatively evaluating the extracted design flow would be challenging due to the lack of a ground
truth. Hence, we created an arXiv-like website, as shown in
Figure 7, where the papers, the corresponding design, and
the generated source code is available. The DL community
Observation
#Points
Naive Bayes
Decision Tree
Logistic Regression
RDF
SVM (RBF Kernel)
Neural Network
Train
1, 720
40.42%
99.65%
99.65%
99.65%
99.65%
100%
Validation
573
54.30%
50.57%
69.98%
68.72%
72.94%
74.93%
Test
574
52.84%
49.13%
68.47%
66.02%
69.68%
71.60%
Table 5: The performance of fine level, five class classifier
to identify the type of DL design flow figure obtained from
the research paper.
Figure 8: An intuitive drag-and-drop UI based framework to
edit the extracted designs and make them publicly available.
Figure 7: An arXiv-like website where DL papers along with
their extracted design, and generated source code in Caffe
and Keras is made available.
could rate the extracted designs which acts as a feedback or
performance measure of our automated approach.
Further, an intuitive drag-and-drop based UI framekwork
is generated for the community to edit the generated DL flow
design, as shown in Figure 8. Ideally the respective papers’
author or the DL community could edit the generated designs, wherever an error was found. The edited design could
be further made publicly available for other researcher to reproduce their design. Further our system could generate the
source code of the edited design in both Keras and Caffe,
in real-time. Thus, we have a two-fold advantange through
this UI system: (i) the public system could act as a one-stop
repository for any DL paper and it’s corresponding design
flow and source code, (ii) the feedback the community provides would enable us to continuously learn and improve the
system8 .
Conclusion and Discussion
Thus, researchers and developers need not struggle any further in reproducing research papers in deep learning. Using this research, the DL design explained in a research
paper could be automatedly extracted. Using an intuitive
drag-and-drop based UI editor, that we designed as a part
of this research, the extracted design could be manually
8
To adhere to the double-blind submission policy of AAAI, the
system URL is not provided in this version of the paper
edited and perfected. Further for an extracted DL design,
the source code could be generated in Keras (Python) and
Caffe (prototxt), in real-time. The proposed DLpaper2Code
framework extracts both figure and table information from
a research paper and converts it into source code. Currently,
an arXiv-like website is created that contains the DL design
and the source code for 5, 000 research papers. To evaluate our approach, we simulated a dataset of 108, 000 unique
deep learning designs validated by a proposed grammar and
their corresponding Keras and Caffe visualizations. On a total dataset of 216, 000 deep learning model visualization diagrams and 28, 120 diagrams that appeared in deep learning research papers and did not contains a model visualization, the proposed binary classification using NNet classifier obtained 99.9% accuracy. The performance of extracting
the generic computational graph figures using the proposed
pipeline is more than 93%. While this research could have a
high impact in the reproducibility of DL research, we have
planned for plenty of possible extensions for the proposed
pipeline:
1. The proposed pipeline detects only the layers (blobs) and
the edges from the diagram. It could be extended to detect
and extract the hyper-parameter values of each layer, to
make the computational graph more content rich.
2. Currently, we have two independent pipelines for generating abstract computational graphs from tables and figures. Combining the information obtained from the multimodal sources could enhance the accuracy of the extracted DL design flow.
3. The entire DLPaper2Code framework could be extended
to support additional libraries, apart from Keras and
Caffe, such as Torch, Tensorflow etc.
4. The broader aim would be to propose a definition of representating DL model design in research papers, achieving
uniformity and better readibility. Further, authors of future papers could also release their design in the created
website for easy accessibility to the community.
References
Abadi, M.; Agarwal, A.; Barham, P.; Brevdo, E.; Chen,
Z.; Citro, C.; Corrado, G. S.; Davis, A.; Dean, J.; Devin,
M.; et al. 2016. Tensorflow: Large-scale machine learning on heterogeneous distributed systems. arXiv preprint
arXiv:1603.04467.
Bastien, F.; Lamblin, P.; Pascanu, R.; Bergstra, J.; and Goodfellow, I. J. e. a. 2012. Theano: new features and speed
improvements. Deep Learning and Unsupervised Feature
Learning NIPS 2012 Workshop.
Chen, T. e. a. 2015. Mxnet: A flexible and efficient machine learning library for heterogeneous distributed systems.
arXiv preprint arXiv:1512.01274.
Chintala, S. 2016. Pytorch. https://github.com/
pytorch/pytorch.
Chollet, F., et al. 2015. Keras. https://github.com/
fchollet/keras.
Choudhury, S. R., and Giles, C. L. 2015. An architecture
for information extraction from figures in digital libraries.
In WWW (Companion Volume), 667–672.
Clark, C., and Divvala, S. 2016. Pdffigures 2.0: Mining
figures from research papers. In Digital Libraries (JCDL),
2016 IEEE/ACM Joint Conference on, 143–152.
Dieleman, S. 2015. Lasagne: First release.
Donahue, J.; Anne Hendricks, L.; Guadarrama, S.;
Rohrbach, M.; Venugopalan, S.; Saenko, K.; and Darrell, T.
2015. Long-term recurrent convolutional networks for visual recognition and description. In Computer vision and
pattern recognition, 2625–2634.
et al, R. C. 2011. Torch7: A matlab-like environment for
machine learning. In BigLearn, NIPS Workshop.
Gibson, A. 2015. Dl4j. https://github.com/
deeplearning4j/deeplearning4j.
Jia, Y.; Shelhamer, E.; Donahue, J.; Karayev, S.; Long, J.;
Girshick, R.; Guadarrama, S.; and Darrell, T. 2014. Caffe:
Convolutional architecture for fast feature embedding. In
ACM international conference on Multimedia, 675–678.
Karpathy, A., and Fei-Fei, L. 2015. Deep visual-semantic
alignments for generating image descriptions. In Computer
Vision and Pattern Recognition, 3128–3137.
Parkhi, O. M.; Vedaldi, A.; Zisserman, A.; et al. 2015. Deep
face recognition. In BMVC, volume 1, 6.
Sankaran, A.; Aralikatte, R.; Mani, S.; Khare, S.; Panwar,
N.; and Gantayat, N. 2011. DARVIZ: Deep abstract representation, visualization, and verification of deep learning
models: Nier track. In International Conference on Software
Engineering, 804–807.
Seide, F., and Agarwal, A. 2016. Cntk: Microsoft’s
open-source deep-learning toolkit. In Proceedings of the
22Nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16, 2135–2135.
New York, NY, USA: ACM.
Simonyan, K., and Zisserman, A. 2014. Very deep convolutional networks for large-scale image recognition. arXiv
preprint arXiv:1409.1556.
Springenberg, J. T.; Dosovitskiy, A.; Brox, T.; and Riedmiller, M. 2014. Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806.
Szegedy, C.; Ioffe, S.; Vanhoucke, V.; and Alemi, A. A.
2017. Inception-v4, inception-resnet and the impact of residual connections on learning. In AAAI, 4278–4284.
Vinyals, O.; Toshev, A.; Bengio, S.; and Erhan, D. 2015.
Show and tell: A neural image caption generator. In Computer Vision and Pattern Recognition, 3156–3164.
| 2cs.AI
|
Language Oriented Modularity: From Theory to Practice
Arik Hadasa and David H. Lorenza,b
*
a Open University of Israel
b Technion—Israel Institute of Technology
Abstract Language-oriented modularity (LOM) is a methodology that complements language-oriented programming (LOP) in providing on-demand language abstraction solutions during software development. It
involves the implementation and immediate utilization of domain-specific languages (DSLs) that are also
aspect-oriented (DSALs). However, while DSL development is affordable thanks to modern language workbenches, DSAL development lacks similar tool support. Consequently, LOM is often impractical and underutilized.
A challenge for LOM is making the complexity of implementing DSALs comparable to that of DSLs and the
effectiveness of programming with DSALs comparable to that of general-purpose aspect languages (GPALs).
Today, despite being essentially both domain-specific and aspect-oriented, DSALs seem to be second-class.
AspectJ development tools do not work on DSAL code. Language workbenches neither deal with the backend weaving nor handle the composition of DSALs. DSAL composition frameworks do not provide front-end
development tools. DSAL code transformation approaches do not preserve the semantics of DSAL programs
in the presence of other aspect languages.
To address this challenge we extend AspectJ with a small set of annotations and interfaces that allows DSAL
designers to define a semantic-preserving transformation to AspectJ. We present a transformation approach
that enables the use of a standard language workbench to implement DSALs and the use of standard aspect
development tools to program with these DSALs. With our approach, DSALs regain their first-class status
with respect to both DSLs and aspect languages. This, on the one hand, lowers the cost of developing DSALs
to the level of DSLs and, on the other hand, raises the effectiveness of using a DSAL to the level of a GPAL.
Consequently, LOM becomes cost-effective compared to the LOP baseline.
As validation, we modified the ajc compiler to support our approach and used it as back-end for two
different language workbenches. With Spoofax we implemented Cool to demonstrate that the non-trivial
composition of AspectJ and Cool can be accommodated using our approach. With Xtext we applied LOM to
crosscutting concerns in two open source projects (oVirt and muCommander), implementing in the process
application-specific DSALs, thus providing a sense of the decrease in the cost of developing composable DSALs
and the increase in the effectiveness of programming with them.
Crosscutting concerns remain a problem in modern real-world projects (e.g., as observed in oVirt). DSALs
are often the right tool for addressing these concerns. Our work makes LOM practical, thus facilitating the
use of DSAL solutions in the software development process.
ACM CCS 2012
Software and its engineering → Software development techniques; Development frameworks and
environments;
Keywords Domain-specific languages (DSLs), Aspect-oriented programming (AOP), Language-oriented
programming (LOP)
The Art, Science, and Engineering of Programming
Submitted
December 1, 2016
Published
April 1, 2017
doi
10.22152/programming-journal.org/2017/1/10
© Arik Hadas and David H. Lorenz
This work is licensed under a “CC BY 4.0” license.
In The Art, Science, and Engineering of Programming, vol. 1, no. 2, 2017, article 10; 37 pages.
Language Oriented Modularity: From Theory to Practice
1
Introduction
Language Oriented Modularity (LOM) [32], taking after Language Oriented Programming (LOP) [40, 8], is a programming methodology that involves the development
and use of Domain Specific Aspect Languages (DSALs) [11] on-demand during the
software modularization process. A DSAL is a programming language that is both
domain-specific and aspect-oriented. It provides not only domain-specific abstractions
and notations like a Domain Specific Language (DSL) does, but also a modularization
mechanism for the separation of domain-specific crosscutting concerns.
1.1 LOM in Theory
In principle, LOM is a special case of LOP, applied to DSALs rather than DSLs. 1 Like
LOP, LOM works middle-out. One starts with defining the DSALs and then works
outwards, combining high level programming with these DSALs in parallel to their
low level implementation. DSALs, like their DSL counterparts, simplify the definition
of (crosscutting) domain logic. Moreover, DSALs also simplify the modularization
mechanism programmers need to understand (which often makes general-purpose
aspect languages (GPALs) more complex to use than the language they extend).
LOM is especially relevant to DSAL-based software development because, in contrast
to DSLs, DSALs tend to be tightly coupled with the program for which they were
designed. This coupling comes in two forms. First, the weaving specification may rely
on the structure of the base code. Second, the crosscutting logic may rely on data
retrieved from the base code. This creates a tight coupling between the DSAL code
and the representation of the data within the base program. The more coupled the
DSAL is with the base code, the less likely it could be reused across applications, thus
encouraging ad-hoc application-specific implementation of DSALs [23].
1.2 LOM in Practice
In practice, however, LOM is not cost-effective like LOP due to the lack (or incompatibility) of supportive development tools. In terms of cost, language workbenches [13] for
LOP, such as Xtext [9], MPS [4], Spoofax [26], and Cedalion [35], provide the language
designer with a development environment for creating with reasonable effort new
DSLs. But these workbenches are not used for LOM because of the inability to express a
semantics-preserving transformation of DSALs to existing GPALs [28] (Section 3.2). In
terms of effectiveness, language workbenches provide the end-programmer with high
quality editing tools for up-to-speed programming with the new DSLs [14]. In contrast,
DSALs typically lack similar editing tools and GPAL development tools generally break
*
Work done in part while visiting the Faculty of Computer Science, Technion—Israel Institute
of Technology.
1
Unless stated otherwise, we use the term DSL to mean ordinary (non aspect-oriented)
domain-specific language rather than its broader meaning which includes DSALs.
10-2
A. Hadas and D. H. Lorenz
on DSAL code. Thus the language development and programming experience with
DSALs is neither on a par with DSLs nor on a par with GPALs.
1.3 Contribution
This work addresses two key challenges that hinder LOM adoption in practice [24, 18,
20]. The first challenge is making DSALs first-class DSLs (in the broader sense). By
this we refer to the availability of language workbenches for creating DSALs.
The second challenge is making newly created DSALs first-class aspect languages.
By this we refer to the availability of development tools for programming with DSALs
that are normally available when programming with a GPAL. For instance, when one
creates a DSAL for a Java [1] application one may expect it to work “out of the box”
not only with Java, AspectJ [27], and other DSALs, but also with Eclipse and the AspectJ
Development Tools (AJDT) [6].
Specifically, we contribute an approach in which DSALs can be implemented like
DSLs by transformation to a GPAL (annotated with metadata), without needing to
change the compiler (weaver) every time a new DSAL is introduced. This provides an
alternative to aspect composition frameworks that do require writing compiler code
as a part of the DSAL implementation. Making DSAL development more like DSL
development allows us to leverage LOP tools for LOM and push down the cost of
LOM closer to the LOP baseline. With relatively minor adjustments to GPAL tools, our
approach allows us to edit, browse, and compile DSAL code as if it were GPAL code,
thus bringing also the effectiveness of LOM closer to LOP [17].
We illustrate our approach concretely with AspectJ. By extending AspectJ with a set
of annotations and interfaces, we get a target language for DSALs that brings the costeffectiveness of the LOM process to the level of LOP. We minimize the additional cost
needed to develop DSALs, compared to DSLs, by enabling the use of standard language
workbenches also for DSALs. We make programming with DSALs more effective by
enabling the use of development tools for AspectJ also for DSALs. Consequently, LOM
becomes almost as cost-effective as LOP [19].
1.4 Outline
Section 2 motivates by example the need for LOM and the need for defining ondemand DSALs, and explains the difficulty today in applying LOM in practice. Section 3
presents our approach and the key technical idea that enables us to use a GPAL as the
target language in ordinary language workbenches. Section 4 describes a concrete
implementation of this approach via a set of annotations for AspectJ. The annotations
are used for surrendering some obliviousness in return for more control over the
visibility of join points and the ordering of advice, and for interfacing with AJDT. In
Section 5 we evaluate the improved cost-effectiveness that is achieved by our approach,
and apply LOM to crosscutting concerns found in two real-world open source projects.
10-3
Language Oriented Modularity: From Theory to Practice
Listing 1 Core methods of the FileJob class
1 public abstract class FileJob implements
,→ Runnable {
2
public FileJob(MainFrame mainFrame,
,→ FileSet files);
3
public void start();
4
public final void run();
5
protected abstract boolean
,→ processFile(AbstractFile file, Object
,→ recurseParams);
6
public void interrupt();
7
public void setPaused(boolean paused);
8 }
2
Figure 1
State diagram for a
file operation (job)
Motivation
To illustrate the LOM process and the need for creating DSALs on-demand, let us
consider the task of extending a lightweight, cross-platform, open source file manager
written in Java, called muCommander,2 with a mechanism for auditing file operations.
2.1 About muCommander
The muCommander file manager supports various operations on files, such as copy,
rename and packing, via a dual-pane interface. File operations in muCommander are
implemented according to the Command design pattern [15]. Each operation, called
job in the terminology of muCommander, is encapsulated within a class that extends
the abstract class FileJob.
Listing 1 depicts the core methods of the FileJob class. Its constructor receives a set
of files on which to operate and the dialog from which the operation was triggered.
By calling the start method, the logic in the run method is executed in a separate
thread (hence why FileJob implements the Runnable interface). The abstract method
processFile is called by the run method for each file, and needs to be implemented by
subclasses with operation-specific logic. The interrupt method stops the execution of a
job. Lastly, the setPaused method receives a boolean argument paused, and accordingly
either pauses or resumes the execution of the job.
Figure 1 depicts the permitted state transitions of a job. Initially the state of the job is
not_started. This state is changed to running once the job starts executing, and
is changed to finished upon completion. If the execution of the job is interrupted
(due to an error or by a user request) the state is changed to interrupted. If the
user asks to pause a job, its state is changed to paused. A job that is paused can be
either resumed and then its state is changed back to running or interrupted and
then its state is changed to interrupted.
2
http://www.mucommander.com
10-4
A. Hadas and D. H. Lorenz
Listing 2 Aspect with advice per job
1 public privileged aspect Logs {
2
pointcut start() : execution(void start());
3
after(CopyJob job): start() && this(job) { audit(Messages.COPY_STARTED, /* ...skipped... */ ); }
4
after(MkdirJob job): start() && this(job) { /* ...skipped... */ }
5
//... more advice for all combinations of state transitions and job types ...
6 }
2.2 Implementation Strategies
There are several approaches for implementing a mechanism that audits which jobs
were executed in muCommander (including their life-cycle and state transitions).
Java One way is to implement the audit concern in Java. This can be done by extending FileJob with a method for each transition, which concrete job classes could override
in order to generate the appropriate messages, and by adding code that persists these
messages to the relevant places in FileJob. The drawback of this approach is that it
degrades code modularity as the generation of the messages becomes scattered across
the job classes and the persistence of these messages becomes tangled among other
concerns in FileJob.
Another option is AspectJ. Both the generation and the persistence of the
messages can be placed within an aspect, separated from the existing code. Listing 2
presents an aspect with an advice that audits the execution of CopyJob (line 3), which
copies files from one directory to another. This advice invokes the audit method with
the message type Messages.copy_started and the arguments it needs. The audit
method then retrieves a translated string of the message and places the given values
in the right locations to produce a message of the form, e.g., “start copying 2 files from
/home/ to /tmp/ ([/home/a.pdf, /home/b.pdf])”.
The Logs aspect in Listing 2 may seem simple enough, however, there are two issues
with using such an aspect in real-world projects. First, it requires all developers in the
project to program in AspectJ since new advice needs to be added for every new job.
Second, as more jobs are added more pieces of advice are added. That leads to high
amount of code duplication and a larger code base that is harder to understand. Note
that while putting the common code in an abstract aspect and extending the aspect
per job class is an option, it would typically make the process of adding a new job
more cumbersome [10].
In order for developers who add jobs not to have to be familiar with AspectJ, one
could introduce a single advice per transition instead of having an advice per job class.
Inside the body of the advice, the concrete type of the job can be identified and the
message type and its values determined accordingly. Developers would only need to
modify the body of the advice to change existing audit messages or to add new ones.
However, as more jobs are added to the Logs aspect, the attractiveness of this
approach decreases. Listing 3 depicts the Logs aspect with auditing for CopyJob and
AspectJ
10-5
Language Oriented Modularity: From Theory to Practice
Listing 3 Aspect with advice per state transition
1 public privileged aspect Logs {
2
pointcut start() : execution(void start());
3
after(FileJob job): start() && this(job) {
4
if (job instanceof CopyJob)
5
audit(Message.COPY_STARTED, /* ... skipped ... */ );
6
else if (job instanceof MkdirJob)
7
if (job.mkfileMode)
8
audit(Message.MKFILE_STARTED, job.files);
9
else
10
audit(Message.MKDIR_STARTED, job.files);
11
//... handle the 'start' transition of other job types ...
12
}
13
//... more advice for other state transitions ...
14 }
MkdirJob. The MkdirJob class contains a field named mkfileMode that determines
whether a file or a directory should be created. Since different audit messages are
defined for these two cases, this field is checked by the Logs aspect (line 7). As more
jobs are being audited and as the resolution of the concrete message to be produced
gets more complicated, the Logs aspect becomes tangled and harder to maintain.
DSAL Reuse
One can consider using an off-the-shelf third-party DSAL or reusing
a DSAL developed for a similar application, like the one we have implemented for
auditing in oVirt,3 called oVirtAudit [22] (Section 5.2.1). The oVirt platform and the
muCommander tool differ in many ways. The former is for virtualization management
while the latter is a file manager. The former is a distributed client-server application
while the latter is standalone. Lastly, the former is intended for enterprise organizations
while the latter is intended for home users. Nevertheless, they both are written in Java,
following the Command design pattern, and have operations (called commands in the
terminology of oVirt) that need to be audited. It is thus tempting to reuse oVirtAudit
in muCommander.
However, one quickly discovers that oVirtAudit is not suitable for muCommander.
First, the syntax of oVirtAudit does not fit. For example, commands in oVirt cannot
be paused and thus oVirtAudit does not provide the syntactic constructs to define a
message for an operation that is being paused or resumed. Second, the semantics
is different. For example, in oVirtAudit the values placed in the audit messages are
taken from method return values while in muCommander they need to be taken from
instance variables. Third, the weaving location is different. For example, the advice
that produces a message for commands in oVirt that start executing is woven into a
method that is not called start and is not located within a class named FileJob.
3
http://www.ovirt.org
10-6
A. Hadas and D. H. Lorenz
Listing 4 Auditing two file jobs in muCommander using muAudit
1 logs for com.mucommander.job.impl.CopyJob:
2
case start log COPY_STARTED with nbFiles baseSourceFolder baseDestFolder files
3
case finish log COPY_FINISHED with nbFiles baseSourceFolder baseDestFolder
4
case interrupt log COPY_INTERRUPTED with baseSourceFolder baseDestFolder
5
case pause log COPY_PAUSED with baseSourceFolder baseDestFolder nbProcessedFiles
6
case resume log COPY_RESUMED with baseSourceFolder baseDestFolder;
7
8 logs for com.mucommander.job.impl.MkdirJob:
9
case start & mkfileMode log MKFile_STARTED with files
10
case start log MKDIR_STARTED with files
11
case finish & mkfileMode log MKFile_FINISHED with files
12
case finish log MKDIR_FINISHED with files
13
case interrupt & mkfileMode log MKFile_INTERRUPTED with files
14
case interrupt log MKDIR_INTERRUPTED with files
15
case pause & mkfileMode log MKFile_PAUSED with files
16
case pause log MKDIR_PAUSED with files
17
case resume & mkfileMode log MKFile_RESUMED with files
18
case resume log MKDIR_RESUMED with files;
New DSAL Listing 4 displays the auditing definitions for CopyJob and MkdirJob in
muAudit—a simple DSAL we can introduce to allow one to define audit messages
in the form of configuration-like case statements. The case statements are matched
top-down, i.e., the order in which they appear is significant. The first part of each
case statement specifies which condition to match based on the job state transition
(e.g., start) and the values of fields within the job class. The second part defines the
message to be produced. This includes the message type (e.g., copy_started) and
the values of fields within the job class. Clearly, muAudit is a declarative and concise
way to express the auditing concern. The question is how does one go about creating
this DSAL?
2.3 Current LOM Solutions
Now that we have established the need to create our own DSAL—muAudit—the
supportive development tools are put to the test. We review the support that is
currently available for the LOM process with respect to five capabilities (Table 1):
1. DSAL interoperability: the ability to define DSALs that can be safely used along with
other DSALs;
2. Development process: the ability to develop DSALs without requiring compiler
(weaver) modifications;
3. Editing tools: the ability to produce general editing tools for programming with the
DSAL;
4. Aspect development tools: the ability to present advice-join-point relationships when
browsing the DSAL code; and
5. Compilation: the ability to compile DSAL code from the command line.
10-7
Language Oriented Modularity: From Theory to Practice
Language Support
CF
Domainspecific
Ø
Aspectoriented
DSAL interoperability
Development process
Editing tools
Aspect development tools
Compilation
LW+
GPAL
Ø
Ø
LW+
CF
Practical
LOM
Ø
Ø
Ø
Ø
Ø
Ø
Ø
Table 1 Comparison of tool support for LOM
One can implement a DSAL like muAudit with an aspect
composition framework such as Awesome [29]. Composition frameworks focus on
the implementation of the weaving semantics for DSALs [30]. However, they neither
generate general editing tools nor provide the desired browsing and compilation capabilities. Moreover, the need to use compiler development techniques (e.g., bytecode
manipulation) results in a highly complex and inefficient development process in
terms of LOM.
Composition Framework (CF)
Language Workbench + General Purpose Aspect Language (LW+GPAL) One may try to
implement muAudit using a language workbench. Indeed, the grammar of the DSAL
can be defined, and general editing tools can be generated. The natural choice would
then be to implement a transformation from that DSAL to a GPAL. This way, we would
achieve a DSL-like development process for DSALs. However, a simple transformation
of DSALs into a GPAL does not preserve the structure of the code and therefore, in the
presence of aspects, does not preserve the meaning of the program. For instance, when
the AJAuditor aspect in Listing 5 is used along with the Logs aspect in Listing 7, it will
unintentionally expose the executions of the audit method that does not exist in the
original muAudit code (Listing 4). Such exposure of internal implementation details
may result in incorrect behavior of the generated code, e.g., even deadlock [28].4
Without a transformation, using a language workbench for the development of DSALs is
not a viable option. Other than the inability to define the weaving semantics of DSALs,
language workbenches also provide neither the desired browsing 5 or debugging
capabilities (e.g., lack the original source code location), nor the necessary standalone
compilation capabilities.
Language Workbench + Composition Framework (LW+CF)
In a sense, language workbenches and composition frameworks are complementary tools. One may consider a
simple composition of the two. An efficient way to implement the DSAL would then
be to parse it with the language workbench and transform it into the form expected
by the composition framework. The expected benefits are: (a) general editing tools
Preventing all generated methods (like audit) from being advised by modifying all other
aspects (like AJAuditor) is impractical, and sometimes impossible in real-world projects.
5
Browsing capabilities are provided only for the generated GPAL code.
4
10-8
A. Hadas and D. H. Lorenz
Listing 5 Aspect for auditing method executions
1 public aspect AJAuditor {
2
before(): call(* *(..)) && !cflow(within(AJAuditor)) { log("ENTER", thisJoinPoint); }
3
//... skipped ...
4 }
are generated by the language workbench; and (b) the DSAL weaving semantics are
implemented with a composition framework. Yet, this approach leaves much to be
desired [18]. First, it provides browsing capabilities for neither the DSAL code nor
the generated code. Second, the need to transform DSAL code before passing it to
the composition framework hinders standalone compilation. Third, the development
process is highly complex and inefficient due to the need to write compiler code.
2.4 Practical LOM
In order to improve the cost-effectiveness of LOM, we seek a solution that supports
all the capabilities listed in Table 1. With a first-class development process and DSAL
interoperability, the cost of developing DSALs in the context of LOM would become
comparable to that of DSLs in LOP. With first-class tools for editing, for aspect development, and for compilation of DSALs, the effectiveness of these DSALs relative to GPALs
would become comparable to that of DSLs relative to general-purpose languages in LOP.
Ideally, the best practical solution would be to use, when possible, standard language
workbenches available for LOP (e.g., Spoofax or Xtext) and standard development
tools available for GPALs (e.g., AJDT).
3
Approach
In pursuing LOM practicality, LOP is our baseline for comparing cost-effectiveness. In
LOP, the cost of DSLs is low thanks to language workbenches (sometimes considered
the killer-app for DSLs [13]). Language workbenches enable efficient implementation of DSLs via transformation. Developing a new DSL with a language workbench
amounts to writing a transformer (generator) from that DSL to a GPL, and writing
a transformer is much easier than writing a compiler or an interpreter. A language
workbench also provides tool support for implementing the transformation and for
effective editing of DSL code. Once the DSL code is transformed, it is compiled with
the GPL’s compiler into an executable form, allowing all development tools that are
available for the GPL to be used effectively.
To make LOM more practical we facilitate a similar approach to the implementation
of DSALs by transforming DSAL code to annotated GPAL code.
10-9
Language Oriented Modularity: From Theory to Practice
3.1 Rationale
Implementing DSALs by defining transformations rather than coding their weaving
semantics is key to making DSAL creation first-class. First, writing a transformer is
much easier than writing a weaver. It eliminates the need to implement a weaver plugin
per DSAL, a task that imposes a significant complexity in composition frameworks.
Second, when a (semantic-preserving) transformation is possible, the LOM software
development process becomes similar to that of LOP and can be completed using
existing language workbenches. Indeed, with the Spoofax [26] language workbench
one can create for some DSAL an Eclipse plugin that provides editing capabilities
(text-highlighting, auto-completion, error-checking, etc.) for writing aspects in that
DSAL. Spoofax can also assist in defining a transformation of aspects written in the
DSAL into, e.g., AspectJ.
Using a GPAL as the target language is key to also making DSAL development firstclass. The transformation of DSALs into a GPAL allows DSAL programmers to leverage,
with a one-time adjustment, development tools that exist for the GPAL. These tools
work with the transformed code, and the adjustments required for them to provide
browsing, navigation, and compilation capabilities for the DSALs are relatively minor.
3.2 Pitfalls
Unfortunately, a naive transformation to a GPAL does not work in general. Consider
AspectJ as the target language for multiple DSALs. As shown elsewhere [33] translating
aspects from different DSALs into aspects in AspectJ and compiling them with the
AspectJ compiler (ajc) may yield incorrect behavior (semantic gap). A DSAL for which
the transformation to AspectJ is not semantic-preserving becomes a second-class DSL.
Meanwhile, trying to fix this by using a different target language and you may lose
the prospects of using AJDT for your DSALs (abstraction gap), thus becoming perhaps
first-class DSL but second-class aspect language.
Semantic Gap Obliviousness [12] has traditionally been an uncompromising principle
in AspectJ. However, in the context of code transformations, complete obliviousness is
disadvantageous. In AspectJ the base code cannot refuse advisement (prevent join
points from being advised). Consequently, a code transformation that does not preserve
the join point “fingerprint” of the original code is not necessarily semantic-preserving
in the presence of foreign aspect code (cross-DSAL foreign advising [33]).
Another difficulty is weaving pieces of advice written in different DSALs at the
same join point shadow (multi-DSAL co-advising [33]). A conflict occurs when the
various pieces of advice are woven in the wrong order. AspectJ provides some control
over the ordering of advice by declaring precedence between aspects (via the declare
precedence statement). However, for programming with multiple DSALs one may need
a finer grained ordering mechanism.
Abstraction Gap
Development tools for aspect languages heavily rely on the representation of advice-join-point relationships in order to annotate the source code
10-10
A. Hadas and D. H. Lorenz
with hints on how aspects are to be woven into the base code. However, code that is
generated from DSAL code loses track of the location of advice in the original DSAL
code. Consequently, development tools cannot annotate the code, thus hindering
effective programming with the DSAL.
In addition, the ability to compile the software from the command line is of a
particular interest in real-world projects because of the use of modern tools for
continuous integration and continuous delivery. The fact that ajc cannot take DSAL
code as input, requires one to modify the compilation process significantly in order to
compile the software from the command line.
3.3 Bridging the Gap With Metadata
We use metadata to weaken AspectJ in terms of obliviousness, strengthen it in terms of
advice ordering, and enhance it in terms of bridging source code locations. The metadata is in the form of Java annotations and an interface for invoking transformations.
Semantic Gap To bridge the semantic gap, a subset of the annotations control the
visibility of join points, thus allowing the definition of the transformation to specify
where to suppress join point shadows (foreign advising). Another annotation controls
the order in which pieces of advice from different DSALs are activated at the same
join point shadow (co-advising).
Abstraction Gap
To bridge the abstraction gap, metadata can be attached within an
annotation, enabling AJDT to provide first-class browsing and navigation capabilities
for DSALs. Additionally, DSAL code transformation plugins that implement a special
interface are invoked automatically by the compiler in order to provide first-class
compilation for DSALs.
4
Implementation
We have implemented our approach by modifying the AspectJ compiler (ajc). Our
modifications to ajc are both:
Optional - when not in use, the compiler’s behavior is unaffected, thus preserving
the correctness of the weaving in ajc before the change; and
Minimal - we do the minimal changes necessary to support our extensions, thus we
expect the process of reapplying these changes to a newer version of the compiler
to be relatively straightforward.
The code changes made to ajc are available at https://github.com/OpenUniversity/ajc
and listed in part in Appendix A.
10-11
Language Oriented Modularity: From Theory to Practice
Listing 6 @Hide annotations added to AspectJ
1
2
3
4
5
6
7
8
9
10
11
12
@Target(ElementType.FIELD)
public @interface HideField {
FieldJoinpoint[] joinpoints() default { FieldJoinpoint.SET, FieldJoinpoint.GET };
}
@Target(ElementType.METHOD)
public @interface HideMethod {
MethodJoinpoint[] joinpoints() default { MethodJoinpoint.CALL, MethodJoinpoint.EXECUTION,
,→ MethodJoinpoint.WITHIN };
}
@Target(ElementType.TYPE)
public @interface HideType {
TypeJoinpoint[] joinpoints() default { TypeJoinpoint.PRE_INIT, TypeJoinpoint.INIT,
,→ TypeJoinpoint.STATIC_INIT, TypeJoinpoint.WITHIN_INIT,
,→ TypeJoinpoint.WITHIN_STATIC_INIT };
}
4.1 Forgoing Complete Obliviousness
Listing 6 defines a set of @Hide annotations for our target language that can be placed
on a code element to suppress join point shadows associated with that element:
@HideField conceals join point shadows associated with a specific field. By default
shadows of both field-set and field-get are hidden, but this can be overridden to be
any subset of them.
@HideMethod conceals join point shadows associated with a specific method. This includes ordinary methods and advice. By default shadows of method/advice-execution,
method-call and all join points that are declared within the method are hidden, but
this can be overridden to be any subset of them.
@HideType conceals join point shadows associated with the initialization of a type.
By default shadows of instance pre-initialization, instance initialization, class staticinitialization, join points declared within instance initialization, and join points
declared within static-initialization are hidden, but this can be overridden to be
any subset of them.
These @Hide annotations are useful when generating AspectJ from DSAL code in
order to hide artificial join point shadows in the generated code that do not exist in
the original code. Listing 7 illustrates the use of two such annotations (lines 1 and 11).
The support for @Hide annotations is added to ajc by modifying the BcelClassWeaver
class, which is responsible for the weaving logic. The decision whether or not to extract
a join point is made after inspecting the @Hide annotation, provided that such an
annotation exists on the program element with which the join point is associated. For
instance, let us consider method- and advice-execution and join points within a method
or advice (Listing 14). When inspecting a LazyMethodGen (which represents an ordinary
method or advice), we check for a @HideMethod annotation. If one exists, we retrieve
the kind of join points to be hidden. In case of MethodJoinpoint.execution, we skip
matching against the method/advice shadows. In case of MethodJoinpoint.within,
10-12
A. Hadas and D. H. Lorenz
Listing 7 Generated aspect from code written in muAudit
1 @HideType
2 public privileged aspect Logs {
3 @BridgedSourceLocation(line=1,
,→ file="/mucommander/src/main/java/com/mucommander/job/jobs.audit",
,→ module="jobs.audit")
4 after(com.mucommander.job.impl.CopyJob job): execution(void start()) && this(job) {
5
if (true) {
6
audit("start copying {0} files from {1} to {2} ({3})", job.nbFiles, job.baseSourceFolder,
,→ job.baseDestFolder, job.files);
7
return;
8
}
9 }
10 //... skipped ...
11 @HideMethod
12 private void audit(String msg, Object... args) {
13
//... skipped ...
14 }
15 }
Listing 8
@Order annotation added to AspectJ
1 public @interface Order {
2
double value();
3 }
we skip matching against the join point shadows within the method/advice body. But
if the method/advice is not annotated with @HideMethod, we skip nothing.
4.2 Fine-Grained Advice Ordering
Listing 8 defines the @Order annotation for ordering pieces of advice per advice rather
than per aspect. The annotation contains a value of type double that represents the
precedence of the annotated advice. The lower this value is, the higher the precedence
is.
The support for @Order annotations is added to ajc by modifying the compareTo
method in the BcelAdvice class, which is used to compare pieces of advice (Listing 15).
When the advice at hand and the advice with which it is being compared to both have
an @Order annotation, the values specified are compared. Otherwise, the comparison
defaults to the regular aspect precedence criteria.
4.3 Redirect Advice-Join-Point Relations to DSAL Code
In order to leverage the browsing and navigation capabilities of AJDT also for programming with DSALs, the @BridgedSourceLocation annotation is set during the
transformation of DSAL code to preserve the source location of advice in the original
10-13
Language Oriented Modularity: From Theory to Practice
Listing 9 @BridgedSourceLocation annotation that was added to AspectJ
1 public @interface BridgedSourceLocation {
2
public String file();
3
public int line();
4
public String module();
5 }
Listing 10 Advice in an aspect that was generated from code written in muAudit
1 @BridgedSourceLocation(line=9,
,→ file="/mucommander/src/main/java/com/mucommander/job/jobs.audit",
,→ module="jobs.audit")
2 after(com.mucommander.job.impl.MkdirJob job): execution(void start()) && this(job) {
3 //... skipped ...
4 }
DSAL code. The @BridgedSourceLocation annotation (Listing 9) cites a path to a DSAL
source code file, a line number, and a module name. When the advice-join-point
relationship mapping is returned by the weaver, the source location pointed to by
the @BridgedSourceLocation annotation is used instead of the actual location of the
advice. This way AJDT markers are shown on the DSAL code and the DSAL code is
referred to by markers on advised join points.
Listing 10 shows an example where @BridgedSourceLocation is used. An aspect
generated from code written in muAudit specifies that the original source location of
after-execution advice (line 2) is line 9 in the file /mucommander/src/main/java/com/mucommander/job/jobs.audit within the jobs.audit module. AJDT uses this information
to present the advises marker at the right location in the DSAL code and the advised-by
markers at the locations at which the advice is to be woven (Fig. 2).
The support for the @BridgedSourceLocation annotation is added to ajc by modifying the AsmRelationshipProvider class to retrieve the source location of advice
from @BridgedSourceLocation. The main change done to this class is modifying the
getHandle method to retrieve the @BridgedSourceLocation that annotates the given
advice and, if it exists, to return a handle based on the file path and line number it
specifies (Listing 16).
4.4 Internal Transformation of DSAL Code
In order to compile DSAL code from the command line like ordinary AspectJ code, we
require all DSAL-specific transformations to implement an interface called Transformation (Listing 11). This interface declares two methods: the extension method returns
the file extension of source files that need to be transformed; and the convert2java
method returns a file containing code generated by the transformation of the DSAL
code found within the given file.
An implementation of the Transformation interface enables the compiler to process
DSAL code directly. In particular, DSAL code can be compiled from the command
10-14
A. Hadas and D. H. Lorenz
Listing 11 The Transformation interface for DSAL code transformation
1 public interface Transformation {
2
String extension();
3
File convert2java(File input) throws Exception;
4 }
Listing 12 Implementation of the Transformation interface for muAudit
1 public class Main implements Transformation {
2
public static void main(String[] args) { /* ... skipped ... */ }
3
//... skipped ...
4
@Override
5
public File convert2java(File input) throws Exception {
6
Main.main(new String[] { input.getPath() });
7
return new File("src−gen/com/mucommander/job/Logs.aj");
8
}
9
@Override
10
public String extension() { return "audit"; }
11 }
line. The compiler invokes the transformation of DSAL code internally. Listing 12
demonstrates the implementation for muAudit. The extension method returns audit
as the extension of muAudit source code files (line 10), and the convert2java method
calls the transformation of muAudit (line 6) and returns the output file (line 7).
The support for Transformation plugins is added to ajc by modifying the classes
AjBuildConfig, ConfigParser, and AjBuildManager. We implemented a pluggable mechanism that uses concrete implementations of Transformation for internal code transformations. First, AjBuildConfig loads all concrete transformations that are specified
in a file called dsals.txt. Then ConfigParser uses the extension method of a loaded
transformation in order not to filter out DSAL source files. Just before the compilation process, AjBuildManager transforms the DSAL source files using the loaded
transformations (Listing 17).
4.5 Discussion
Our approach is based on a transformation of DSALs into a language that extends
a GPAL. However, the choice of GPAL is significant. While many of the crosscutting
concerns found in real-world projects can be resolved by DSALs whose weaving
semantics can be expressed in AspectJ and its join point model, there may be some that
are not. Although the join point model of AspectJ could also be extended on-demand,
this would likely require much more effort and reduce the cost-effectiveness of the
approach. Thus, we intentionally do not consider DSALs that cannot be expressed in
AspectJ, but argue that this is a reasonable choice in practice (Section 5).
The design of the @Hide and @Order annotations is based on the classification
of multi-DSAL conflicts as foreign- and co- advising [30], and mimics the solution
10-15
Language Oriented Modularity: From Theory to Practice
provided by the Awesome composition framework. Nevertheless, our approach is not
limited to these particular annotations. The specification of the proposed annotations
can be enhanced and even completely replaced with alternative metadata that one
can use to resolve multi-DSAL conflicts. Our approach will work as long as the conflict
resolution can be specified declaratively and generated during the transformation of
DSAL code (with reasonable effort), and as long as it does not impose changes that
break the compatibility with the GPAL tools and does not require writing compiler
code per DSAL.
It is up to the language designer to determine which shadows to hide using the
@Hide annotations. Generally, one would want to hide shadows of join points that do
not appear in the DSAL code (as these may be considered an internal implementation
detail). The @BridgedSourceLocation is supposed to be set on every advice in order to
present AJDT markers at the right locations for the end-programmer. The more delicate
use is that of the @Order annotation since the value assigned in the transformation
of one DSAL is affected by values assigned in the transformations of other DSALs.
Obviously, the @Order annotation allows a finer-grained ordering than AspectJ and the
use of values of type double allows us, theoretically, to introduce right values for newly
introduced DSAL without modifying existing DSALs. However, in order to provide a
practical way to use the @Order annotation, one would probably want to use a tool
like SpecTackle [34] to specify the order of all the pieces of advice from all the
DSALs that are used in the project, using a user-friendly UI with default resolution.
The tool can also help to validate the correctness of specific transformations with
business logic tests. However, the implementation of such a tool is out of the scope of
this paper.
As we shall see in Section 5, our approach is not coupled with a particular language
workbench. One can pick a language workbench that is right for the DSAL task at hand.
The only requirement that affects the selection of a language workbench is the need
to produce a standalone transformation in order to implement the Transformation
interface. However, most mainstream language workbenches provide this capability
by default.
A natural question to ask in the context of LOM is what is the complexity of
implementing and using a large number of DSALs using our approach. On the one hand,
the use of @Hide and @BridgedSourceLocation annotations and the Transformation
interface for each particular DSAL is not affected by the use of other DSALs. On the
other hand, for determining the ordering value in the @Order annotation for a given
advice, one needs to consider pieces of advice defined in other DSALs.
5
Evaluation
To evaluate our approach we compare the effort required to implement a complex
third-party DSAL with our approach to the effort of implementing the same DSAL in
the Awesome composition framework. To assess the impact our approach has on
the cost-effectivenss of LOM we present two case studies of implementing and using
DSALs in open source projects.
10-16
A. Hadas and D. H. Lorenz
Figure 2
Tool support for programming with Cool in Eclipse
5.1 Comparative Evaluation: Implementing Cool
Cool is a DSAL for thread synchronization [31]. It is a non-trivial language to implement for several reasons. First, a coordinator (aspect in Cool) may include blocks
of code in plain Java. Second, a coordinator has direct access to fields and methods
of the class that is being advised. Third, features of Cool interact with features of
AspectJ [28, 33]. In Cool, join points are not reflected in the syntax, advice comprises
distant terms and expressions, and there is no correct translation to plain AspectJ
(because the implementation-specific operations in the generated aspect might be
ill-advised by other aspects).
We compare two complete implementations of Cool defined in the Spoofax language workbench, one following our approach and the other using a composition
framework (CF) approach with Awesome as back-end. In both implementations the
grammar of the language is defined in the Syntax Definition Formalism (SDF) [38] (Listing 18) and its transformation is implemented in Stratego [39]. In our approach, a
transformation to AspectJ is implemented (with @Hide annotations). In the CF approach a transformation to Java is implemented, and a new weaver plugin is also
implemented.
To test the interaction of Cool with AspectJ we implemented a coordinator in
Cool (Listing 20) that synchronizes a bounded stack (Listing 19). Figure 2 shows a
screenshot of Eclipse that demonstrates that text-highlighting and AJDT markers are
provided in the process. In addition, we implemented an aspect in AspectJ that audits
method executions and uses the stack (Listing 21). We then ran a multi-threaded
application that reads and writes from and to a bounded stack simultaneously. The
deadlock problem reported elsewhere [28] that occurs when the coordinator is translated to plain AspectJ was not observed when the @Hide annotations were placed
during the transformation (but reproduced successfully when we removed them).
10-17
Language Oriented Modularity: From Theory to Practice
Code Transformation
EV
Other
Weaver
Plugin
Implementation
Grammar
Language
SDF
Stratego (AST)
Stratego (AST)
Java
CF Approach
Our Approach
34
34
761 (4168)
0
297 (3001)
382 (3008)
1557
0
Table 2 Number of lines of SDF, Stratego, and Java code in the implementation of Cool
Table 2 compares the implementation effort required when using the two approaches
based on #LOC written in three languages. First, the grammar definition in SDF (34
LOC) is the same in both implementations, since the language was defined the same
way in Spoofax.
Second, the code transformations are implemented in Stratego. We distinguish
between the part of the transformation that handles the resolution of external variables (EV) in Cool and everything else. For each part we compare the #LOC in
Stratego and the size of its Abstract Syntax Tree (AST) representation in Spoofax.
The implementation of the code transformation was much shorter in our approach
(382 LOC) than in the alternative approach (761+297=1058 LOC). In our approach,
the relatively complex part of the implementation that handles external variables in
Cool coordinators was eliminated because it is handled entirely by ajc. On the other
hand, the need to generate annotations in our approach slightly increased the size of
the other part of the implementation. We argue that the LOC in Stratego is highly
affected by code style and therefore the size in LOC of its AST is more representative. The AST of that part using our approach is slightly larger but the difference is
insignificant (3008-3001=7).
Third, the implementation of a weaver plugin in Java (1557 LOC) is only required
in the CF approach. This emphasizes that in the CF approach one needs to write a
relatively large amount of code that is considered to be complicated for most developers
since it requires expertise in low level bytecode manipulation tools. In contrast, this
knowledge is not needed in our approach since the weaver is not modified. However,
the language designer needs to know not only Java but also AspectJ.
5.2 Experimental Evaluation
To evaluate our approach in the context of LOM, we present two case studies. In
the first case study our approach is used for handling crosscutting concerns found in
legacy code of a real-world software. In the second case study our approach is used
for handling new requirements on-demand, which is a more typical scenario for LOM.
5.2.1 Case Study: oVirt
oVirt is an open source production-ready enterprise application for providing and
managing virtual data centers and private cloud solutions. For example, Red Hat
10-18
A. Hadas and D. H. Lorenz
Enterprise Virtualization (RHEV), a commercial competitor to VMware vSphere, is
based on oVirt and is deployed in big organizations, such as British Airways. 6
oVirt-Engine is the control center of the oVirt distributed system that manages the
different hosts that run virtual machines. Its core design is based on the Command
design pattern [15]. Each operation that is supported by oVirt-Engine is modeled by a
command class that inherits from a common root called CommandBase.
We identified three concerns, namely synchronization, auditing, and permissions,
that cross-cut many modules in the oVirt-Engine application. The synchronization
concern is about preventing conflicting commands from running simultaneously. The
auditing concern is about producing informative messages at different stages of
command execution. Lastly, the permissions concern is about ensuring that only users
with sufficient permissions on entities are able to execute commands that affect them.
These concerns are scattered across most of the command classes (as demonstrated in
Figs. 3 and 4). In addition, these concerns are tangled within the CommandBase class.
For these crosscutting concerns we implemented in Xtext three DSALs, called
oVirtSync, oVirtAudit, and oVirtPermissions, respectively [22]. Their grammars were
defined in the language grammar definition format provided by Xtext (Listings 22
to 24). Their transformations into AspectJ with our annotations were implemented in
Xtend [3], a language for code transformation provided by Xtext. The transformations
of the oVirtSync, oVirtAudit, and oVirtPermissions consisted of 259, 83, and 235 LOC,
respectively. 7 With these DSALs, we implemented aspect solutions for three commands
named MigrateVmCommand, AddDiskCommand, and ExportVmTemplateCommand (Listings 25 and 26 show aspects written in oVirtSync and oVirtAudit, respectively).
Code scattering was eliminated by encapsulating the code that was spread across
the command classes (which, for some commands, exceeded 25% of their LOC) in a
single module implemented in the corresponding DSAL. Code tangling was resolved
by extracting code that was tangled inside the CommandBase class into the DSAL
aspects (296 LOC were untangled, which is more than 12% of the overall LOC). This
illustrates that DSALs that are reducible to AspectJ using our approach were effective
in separating out the crosscutting concerns we identified in oVirt-Engine. Moreover,
the fact that these languages were implemented with Xtext using our approach,
unlike our implementation of Cool that was done with Spoofax, validates that our
approach is agnostic to the selection of the language workbench. Editing tools and
aspect development tools support provided for programming in these languages are
demonstrated in Fig. 5.
5.2.2 Case Study: muCommander
In this case study the LOM process was applied to the muCommander open source
project described in Section 2.
http://www.redhat.com/en/about/press-releases/british-airways-chooses-rhev-to-improveit-systems-to-build-internal-cloud
7
https://github.com/OpenUniversity/DSALs
6
10-19
Language Oriented Modularity: From Theory to Practice
Listing 13
1
2
3
4
5
6
7
8
9
10
Grammar definition of muAudit in Xtext
Model: (commands+=Command)*;
Command:
'logs for' type=[types::JvmDeclaredType|QualifiedName] ':' (cases+=Case)* ';'
;
Case:
'case' state=State ('&' (fields+=[types::JvmField]))*
'log' msg=[types::JvmEnumerationLiteral] ('with' (vars+=[types::JvmField])+)?
;
enum State: start | finish | interrupt | pause | resume;
QualifiedName: ID ("." ID)*;
Using Xtext we implemented a DSAL, named muAudit, for adding a missing auditing
feature for job executions in muCommander. We defined the grammar of muAudit in
the grammar definition format provided by Xtext (Listing 13). From this grammar
definition we generated a plugin for programming with muAudit in Eclipse using
Xtext. At this point we were already able to use the IDE plugin and start writing code
in muAudit in Eclipse, using general editing capabilities that are typically available
when programming with DSLs (Fig. 6). To run muAudit code we implemented a
transformation to AspectJ using our approach. This transformation comprised 110 LOC
in Xtend. Listing 7 depicts part of the aspect that is generated from the code presented
in Listing 4. Note the use of @Hide annotations to hide join point shadows associated
with the artificial type Logs and its method audit.
With the transformation implemented, we were able to use aspect development
tools that are typically available for programming with AspectJ, e.g., when writing the
code in Listing 4 (Figs. 6 and 7). Finally, we implemented the Transformation interface
for muAudit (Listing 12) and added it to the dsals.txt file. This enabled us to compile
the project not only from within Eclipse but also from the command line, with no
changes to the build process.
The LOM development process of muAudit consists of three parts that can be
compared to LOP: the grammar definition, the implementation of code transformation,
and the implementation of the Transformation interface. The first two parts were done
using Xtext, similar to how this is done for DSLs. The third part is also done using
Xtext, similar to how it is done when required for DSLs. Even if the third part is
typically not needed for DSLs, the additional effort it requires is negligible. Overall,
the cost of implementing muAudit is similar to that of a DSL in LOP.
As for the effectiveness of programming with muAudit, not only did we enjoy the
benefits of programming with a simplified and more declarative language than AspectJ,
but we also enjoyed all the development tools that are usually available for AspectJ. The
plugin for Eclipse provided us with general editing tools that are commonly provided
by IDEs nowadays. In addition, we were provided with the unique capabilities of
aspect development tools by using AJDT and we were able to compile DSAL code the
same way we compile AspectJ code. Overall, we were able to program with our DSAL
effectively (compared to AspectJ), similar to how one programs with a DSL effectively
(compared to Java) in LOP.
10-20
A. Hadas and D. H. Lorenz
5.3 Validity and Threats to Validity
The case studies illustrate the effectiveness of our approach in the process of implementing and using DSALs for crosscutting concerns in a real-world project. They
demonstrate that with a development effort comparable to that of DSL development
(the definition and implementation of the language were done in only a few hours
using an existing language workbench) and an effective programming experience
comparable to that of a GPAL (existing GPAL tools were used), the cost-effectiveness
of the LOM process using our approach is comparable to that of the LOP process.
Internal Validity
In LOM the language designer and the DSAL end-programmer are
usually different people. In the case studies presented we played both roles. Our
familiarity with the DSALs, language workbenches, and the implementation of our
extensions to AspectJ could have positively influenced the LOM process. However, to
factor out this effect we assess the cost-effectiveness of LOM relative to LOP, comparing
the process and tools used.
External Validity
One can argue that the DSALs we implemented may not be representative, e.g., that Cool is more complex than most DSALs and that the applicationspecific DSALs are simpler than most. However, the implementation of Cool is
commonly used as a benchmark test case for DSAL frameworks. The fact that the LOM
process was cost-effective even for application-specific DSALs is even more impressive
than for Cool whose development cost can be amortized across applications.
While being a first-class DSL is a direct consequence of implementing the DSALs with
a language workbnech, during the case studies we did not test all aspect development
tools available for AspectJ, and therefore it is possible that our DSALs are not first-class
aspect languages. However, the fact that the browsing capabilities of AJDT (which
are typically not available when programming with DSALs) worked in our approach,
and that our target language is based on AspectJ, reduces this risk. Yet, it is possible
that the @BridgedSourceLocation annotation would need to be enhanced in order to
preserve compatibility with future tools.
6
Related Work
Various aspect development tools aim at facilitating either the development or the use
of DSALs. The Aspect Bench Compiler (abc) [2] is more extensible than the AspectJ
compiler (ajc), allowing one to produce extensions to AspectJ and DSALs more easily.
However, abc is intended for the development of a particular extension rather than for
the composition of extensions. Moreover, abc supports only an old version of AspectJ.
Javassist [5] and similar toolkits that simplify bytecode manipulation can potentially
simplify the definition of the DSAL weaving semantics, e.g., when using a composition
framework. In contrast, our approach avoids completely the need to program a
composition framework extension per-DSAL.
10-21
Language Oriented Modularity: From Theory to Practice
Interpreter-based frameworks like Pluggable AOP [28], JAMI [25], and POPART [7]
also avoid low-level implementation of the weaving semantics. However, they achieve
simplicity at the expense of performance, since their conflict resolution is based
on interpretation. In our approach, the use of DSALs does not imply performance
degradation compared to use of GPALs.
The Awesome composition framework [29] generalizes the weaving process of
the AspectJ compiler in order to support the definition of DSAL weaving semantics.
This approach provides finer-grained constructs for the resolution of foreign advising
and co-advising conflicts than the @Hide and @Order annotations we implemented.
Our approach, in contrast, provides the ability to define the DSAL weaving semantics
without needing low-level bytecode manipulation tools.
The idea of transforming DSALs into AspectJ is found in XAspects [36]. However, a
transformation to pure AspectJ does not generally preserve the original meaning of
the program. Reflex [37] uses a low-level kernel language to which different aspect
languages are transformed. In contrast, our approach fully supports AspectJ and the
use of its development tools out-of-the-box when programming with DSALs.
Elsewhere [16, 21] we presented an improved approach to the composition of a
language workbench and a composition framework to produce first-class DSALs. In
this paper we present an alternative approach, that not only produces first-class aspect
languages like the former approach, but also achieves better first-class equality with
DSLs by making DSAL development process much more similar to that of a DSL.
SpecTackle [34] is a tool that facilitates the resolution of multi-DSAL co-advising
conflicts. It allows one to resolve co-advising conflicts per application, by presenting
the conflicts between different aspect languages and allowing programmers to resolve them. This is equivalent to making the values within @Order configurable per
application. This topic is left for future research.
7
Conclusion
This work addresses the Achilles’ heel of LOM practicality, namely that the DSAL
development process is far from being cost-effective. On the one hand, DSALs are more
costly to develop than DSLs due to the implementation of their weaving semantics. On
the other hand, due to the lack of development tools that work on DSALs, the relative
effectiveness of programming with a DSAL (relative to a GPAL) is lower than the
relative effectiveness of programming with a DSL (relative to a GPL). Consequently,
LOM is often avoided or underutilized in practice.
In contrast, LOP is practical thanks to the availability of language workbenches,
which provide tools for rapid construction of DSLs as well as for end-programmer
productivity in using these DSLs. These tools, however, were generally considered to
be inapplicable to DSLs that are aspect-oriented (DSALs). On the one hand, due to
the strong coupling between aspects and base code, code transformation may break
aspect code or change the meaning of a program as a whole. On the other hand,
developing aspect development tools and a weaver for each DSAL is a tedious and
complex task.
10-22
A. Hadas and D. H. Lorenz
Our work shows that DSALs can be produced with LOP tools (like language workbenches) and used with GPAL development tools (like AJDT). We present a transformation approach that improves the cost-effectiveness of DSAL development and
brings it to the level of DSL development. In our approach, code written in DSALs
that are reducible to a GPAL can be translated to annotated GPAL code in a semanticpreserving manner and interface with development tools intended for the GPAL. With
our approach one can implement DSALs using a standard language workbench, such
as Spoofax and Xtext. We present a concrete implementation of the approach with
AspectJ as the target language.
In a sense, our work strives to be for DSAL development what the introduction of
language workbenches was for DSL development. With our approach, LOM becomes
practical for real-world software development process, enabling the on-demand creation and use of DSALs for handling crosscutting concerns, thereby minimizing code
scattering and tangling that still prevail in modern software projects.
This research was supported in part by the Israel Science Foundation (ISF) under grant No. 1440/14.
Acknowledgment
References
[1] Ken Arnold and James Gosling. The Java Programming Language. The Java
Series. Addison-Wesley, 1996.
[2] Pavel Avgustinov, Aske Simon Christensen, Laurie Hendren, Sascha Kuzins, Jennifer Lhoták, Ondr̆ej Lhoták, Oege de Moor, Damien Sereni, Ganesh Sittampalam, and Julian Tibble. abc: an extensible AspectJ compiler. In Proceedings of the 4th International Conference on Aspect-Oriented Software Development (AOSD’05), pages 87–98, Chicago, Illinois, USA, March 2005. ACM Press.
doi:10.1145/1052898.1052906.
[3] Lorenzo Bettini. Implementing Domain-Specific Languages with Xtext and Xtend.
Packt Publishing, 2013.
[4] Fabien Campagne. The MPS Language Workbench: Volume I. Fabien Campagne,
2014.
[5] Shigeru Chiba. Javassist– a reflection-based programming wizard for Java. In
Jean-Charles Fabre and Shigeru Chiba, editors, Proceedings of the OOPSLA 1998
Workshop on Reflective Programming in C++ and Java, Vancouver, BC, Canada,
October 1998. UTCCP Report 98-4, Center for Computational Physics, University
of Tsukuba, Japan.
[6] Andy Clement, Adrian Colyer, and Mik Kersten. Aspect-oriented programming
with AJDT. In Proceedings of the ECOOP 2003 Workshop on Analysis of AspectOriented Software, Darmstadt, Germany, July 2003.
[7] Tom Dinkelaker, Mira Mezini, and Christoph Bockisch. The art of the meta-aspect
protocol. In Proceedings of the 8th International Conference on Aspect-Oriented
10-23
Language Oriented Modularity: From Theory to Practice
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
Software Development (AOSD’09), pages 51–62, Charlottesville, Virginia, USA,
March 2009. ACM Press.
Sergey Dmitriev. Language oriented programming: The next programming
paradigm. JetBrains onBoard, 1(2), 2004.
Sven Efftinge and Markus Völter. oAW xText: A framework for textual DSLs. In
Eclipse Modeling Symposium, Eclipse Summit Europe 2006, Esslingen, Germany,
October 2006.
Erik Ernst and David H. Lorenz. Aspects and polymorphism in AspectJ. In Proceedings of the 2nd International Conference on Aspect-Oriented Software Development
(AOSD’03), pages 150–157, Boston, Massachusetts, March 17-21 2003. ACM Press.
Johan Fabry, Tom Dinkelaker, Jacques Noyé, and Éric Tanter. A taxonomy of
domain-specific aspect languages. ACM Computing Surveys (CSUR), 47(3), April
2015. doi:10.1145/2685028.
Robert E. Filman and Daniel P. Friedman. Aspect-oriented programming is quantification and obliviousness. In Peri Tarr, Lodewijk Bergmans, Martin Griss, and
Harold Ossher, editors, Proceedings of the OOPSLA 2000 Workshop on Advanced
Separation of Concerns. Department of Computer Science, University of Twente,
The Netherlands, 2000.
Martin Fowler. Language workbenches: The killer-app for domain specific languages, 2005.
Martin Fowler. Domain-Specific Languages. Addison-Wesley Signature Series.
Pearson Education, 2010.
Erich Gamma, Richard Helm, Ralph Johnson, and John Vlissides. Design Patterns:
Elements of Reusable Object-Oriented Software. Professional Computing. AddisonWesley, 1995.
Arik Hadas. A language workbench for creating production-ready extensions to
AspectJ. In The 14th International Conference on Modularity (Modularity’15), Fort
Collins, CO, USA, March 2015. ACM Student Research Competition Silver Medal
Award (Graduate category).
Arik Hadas. Language oriented modularity: From theory to practice. In The 15th
International Conference on Modularity (Modularity’16), Málaga, Spain, March
2016. ACM Student Research Competition Gold Medal Award (Graduate category).
Arik Hadas and David H. Lorenz. Demanding first-class equality for domain
specific aspect languages. In Companion to the 14th International Conference on
Modularity (Modularity’15 Comp.), pages 35–38, Fort Collins, CO, USA, March
2015. ACM Press. Position paper.
Arik Hadas and David H. Lorenz. Experiencing with langauge oriented modularity (poster). In Proceedings of the 1st International Conference on the Art, Science,
and Engineering of Programming, Brussels, Belgium, 2017.
Arik Hadas and David H. Lorenz. First class domain specific aspect languages.
In Companion to the 14th International Conference on Modularity (Modularity’15
10-24
A. Hadas and D. H. Lorenz
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
Comp.), pages 29–30, Fort Collins, CO, USA, March 2015. ACM Press. Poster
Session.
Arik Hadas and David H. Lorenz. A language workbench for implementing your
favorite extension to AspectJ. In Companion to the 14th International Conference
on Modularity (Modularity’15 Comp.), pages 19–20, Fort Collins, CO, USA, March
2015. ACM Press. Demo Session.
Arik Hadas and David H. Lorenz. Application-specific language-oriented modularity: A case study of the oVirt project. In Proceedings of the Modularity’16 Workshop
on Modularity Across the System Stack (MASS’16), pages 178–183, Málaga, Spain,
March 2016. ACM Press.
Arik Hadas and David H. Lorenz. Toward disposable domain-specific aspect
languages. In Proceedings of the Modularity’16 Workshop on Foundations of AspectOriented Languages (FOAL’16), pages 83–85, Málaga, Spain, March 2016. ACM
Press.
Arik Hadas and David H. Lorenz. Toward practical language oriented modularity.
In Proceedings of the Modularity’16 Workshop on Language Modularity À La Mode
(LaMOD’16), pages 94–98, Málaga, Spain, March 2016. ACM Press.
Wilke Havinga, Lodewijk Bergmans, and Mehmet Ak¸sit. Prototyping and composing aspect languages: using an aspect interpreter framework. In Proceedings
of the 22nd European Conference on Object-Oriented Programming (ECOOP’08),
number 5142 in Lecture Notes in Computer Science, pages 180–206, Paphos,
Cyprus, July 2008. Springer-Verlag.
Lennart C.L. Kats and Eelco Visser. The Spoofax language workbench: Rules
for declarative specification of languages and IDEs. In Proceedings of the ACM
International Conference on Systems, Programming, and Applications: Software
for Humanity (SPLASH’10), pages 444–463, Reno/Tahoe, Nevada, USA, October
2010. ACM Press.
Gregor Kiczales, Erik Hilsdale, Jim Hugunin, Mik Kersten, Jeffrey Palm, and
William Griswold. An overview of AspectJ. In Proceedings of the 15th European
Conference on Object-Oriented Programming (ECOOP’01), number 2072 in Lecture Notes in Computer Science, pages 327–353, Budapest, Hungary, June 2001.
Springer-Verlag.
Sergei Kojarski and David H. Lorenz. Pluggable AOP: Designing aspect mechanisms for third-party composition. In Proceedings of the 20th Annual ACM
SIGPLAN Conference on Object-Oriented Programming Systems, Languages, and
Applications (OOPSLA’05), pages 247–263, San Diego, CA, USA, October 2005.
ACM Press.
Sergei Kojarski and David H. Lorenz. Awesome: An aspect co-weaving system for composing multiple aspect-oriented extensions. In Proceedings of the
22nd Annual ACM SIGPLAN Conference on Object-Oriented Programming Systems,
Languages, and Applications (OOPSLA’07), pages 515–534, Montreal, Canada,
October 2007. ACM Press.
10-25
Language Oriented Modularity: From Theory to Practice
[30] Sergei Kojarski and David H. Lorenz. Identifying feature interaction in aspectoriented frameworks. In Proceedings of the 29th International Conference on
Software Engineering (ICSE’07), pages 147–157, Minneapolis, MN, May 2007.
IEEE Computer Society.
[31] Cristina Videira Lopes. D: A Language Framework for Distributed Programming.
PhD thesis, Northeastern University, 1997.
[32] David H. Lorenz. Language-oriented modularity through Awesome DSALs:
summary of invited talk. In Proceedings of the 7th AOSD Workshop on DomainSpecific Aspects Languages (DSAL’12), pages 1–2, Potsdam, Germany, March 2012.
ACM Press.
[33] David H. Lorenz and Sergei Kojarski. Understanding aspect interactions, coadvising and foreign advising. In Proceedings of ECOOP’07 Second International
Workshop on Aspects, Dependencies and Interactions, pages 23–28, Berlin, Germany,
July 2007.
[34] David H. Lorenz and Oren Mishali. SpecTackle: Toward a specification-based
DSAL composition process. In Proceedings of the 7th AOSD Workshop on DomainSpecific Aspects Languages (DSAL’12), Potsdam, Germany, March 2012. ACM Press.
[35] David H. Lorenz and Boaz Rosenan. Cedalion: A language for language oriented
programming. In Proceedings of the 26th Annual ACM SIGPLAN Conference on
Object-Oriented Programming Systems, Languages, and Applications (OOPSLA’11),
pages 733–752, Portland, Oregon, USA, October 2011. ACM Press. doi:10.1145/
2048066.2048123.
[36] Macneil Shonle, Karl Lieberherr, and Ankit Shah. XAspects: An extensible system
for domain specific aspect languages. In Companion to the 18th Annual ACM
SIGPLAN Conference on Object-Oriented Programming Systems, Languages, and
Applications, pages 28–37, Anaheim, California, 2003. ACM Press.
[37] Éric Tanter. Aspects of composition in the Reflex AOP kernel. In Proceedings of
the 5th International Symposium on Software Composition (SC’06), number 4089
in Lecture Notes in Computer Science, pages 98–113, Vienna, Austria, March
2006. Springer-Verlag.
[38] Eelco Visser. Syntax Definition for Language Prototyping. PhD thesis, University
of Amsterdam, September 1997.
[39] Eelco Visser. Stratego: A language for program transformation based on rewriting strategies: system description of Stratego 0.5. In Aart Middeldorp, editor, Proceedings of the 12th International Conference on Rewriting Techniques
and Applications (RTA 2001), number 2051 in Lecture Notes in Computer Science, pages 357–362, Utrecht, The Netherlands, May 2001. Springer-Verlag.
doi:10.1007/3-540-45127-7_27.
[40] Martin P. Ward. Language-oriented programming. Software-Concepts and Tools,
15(4):147–161, 1994.
10-26
A. Hadas and D. H. Lorenz
A
Changes to the AspectJ Compiler
Listing 14 Advice-join-point matching with @Hide annotations in BcelClassWeaver
1 private boolean match(LazyMethodGen mg) {
2 //... skipped ...
3 boolean hide = false, hideWithin = false;
4 //... skipped ...
5
AnnotationAJ hideAnn = getHideMethodAnnotation(mg);
6
if (hideAnn != null) {
7
String hideStr = hideAnn.getStringFormOfValue("joinpoints");
8
hide = hideStr == null ||
,→ hideStr.contains("Lorg/aspectj/lang/annotation/MethodJoinpoint;EXECUTION");
9
hideWithin = hideStr == null ||
,→ hideStr.contains("Lorg/aspectj/lang/annotation/MethodJoinpoint;WITHIN");
10
}
11 //... skipped ...
12 if (canMatchBodyShadows && !hideWithin) {
13
for (InstructionHandle h = mg.getBody().getStart(); h != null; h = h.getNext()) {
14
match(mg, h, enclosingShadow, shadowAccumulator);
15
}
16 }
17 if (canMatch(enclosingShadow.getKind()) && !hide && /** ...skipped ... */ ) {
18
if (match(enclosingShadow, shadowAccumulator)) {
19
enclosingShadow.init();
20
}
21 }
22 //... skipped ...
23 }
Listing 15 Advice comparison with @Order annotations in BcelAdvice
1 public int compareTo(Object other) {
2 //... skipped ...
3 BcelAdvice o = (BcelAdvice) other;
4
5 try {
6
double diff = getOrder(this) − getOrder(o);
7
if (diff != 0)
8
return diff < 0 ? −1 : 1;
9 } catch (Exception e) {}
10
11 if (kind.getPrecedence() != o.kind.getPrecedence()) {
12
if (kind.getPrecedence() > o.kind.getPrecedence()) {
13 //... skipped ...
14 }
10-27
Language Oriented Modularity: From Theory to Practice
Listing 16 Retrieving source code location with @BridgedSourceLocation annotations in
AsmRelationshipProvider
1 public static String getHandle(AsmManager asm, Advice advice) {
2 if (null == advice.handle) {
3
AnnotationAJ ann = getSourceLocation(advice);
4
if (ann != null)
5
advice.handle = ann.getStringFormOfValue("file") + ann.getStringFormOfValue("line");
6
else {
7
ISourceLocation sl = advice.getSourceLocation();
8
if (sl != null) {
9
IProgramElement ipe = asm.getHierarchy().findElementForSourceLine(sl);
10
advice.handle = ipe.getHandleIdentifier();
11
}
12
}
13 }
14 return advice.handle;
15 }
Listing 17 Invoking DSAL code transformations in AjBuildManager
1 public void performCompilation(Collection<File> files) {
2 //... skipped ...
3 for (Iterator<File> fIterator = files.iterator(); fIterator.hasNext();) {
4
File f = fIterator.next();
5
for (Transformation t : AjBuildConfig.transformations)
6
if (f.getName().endsWith(t.extension())) {
7
try {
8
f = t.convert2java(f);
9
}
10
catch (Exception e) {
11
e.printStackTrace();
12
}
13
break;
14
}
15
filenames[idx++] = f.getPath();
16 }
17 //... skipped ...
18 }
10-28
A. Hadas and D. H. Lorenz
B
Grammar Definition for Cool
Listing 18 Grammar definition for Cool in the SDF format
1 module languages/cool/Main
2 imports
3 languages/java−15/Main
4 languages/cool/AspectjExtension
5
6 exports
7 sorts
8
ConditionDec CoordinatorBodyDec CoordinatorDec CoordinatorBody MethodAdditionsDec
,→ MethodSignature MutexDec SelfexDec
9
10 context−free syntax
11
CoordinatorDec −> TypeDec
12
CoordinatorDecHead CoordinatorBody −> CoordinatorDec {cons("CoordinatorDec")}
13
"coordinator" TypeName −> CoordinatorDecHead {cons("CoordinatorDecHead")}
14
"{" CoordinatorBodyDec* "}" −> CoordinatorBody {cons("CoordinatorBody")}
15
16
SelfexDec −> CoordinatorBodyDec
17
"selfex" "{" {MethodSignature ","}* "}" ";" −> SelfexDec {cons("Selfex")}
18
19
MutexDec −> CoordinatorBodyDec
20
"mutex" "{" {MethodSignature ","}* "}" ";" −> MutexDec {cons("Mutex")}
21
22
ConditionDec −> CoordinatorBodyDec
23
"condition" {Expr ","}* ";" −> ConditionDec {cons("ConditionDec")}
24
25
MethodAdditionsDec −> CoordinatorBodyDec
26
MethodSignature ":" Requires? OnEntry? OnExit? −> MethodAdditionsDec
,→ {cons("MethodAdditions")}
27
Id "(" {Type ","}* ")" −> MethodSignature {cons("MethodSignature")}
28
"requires" Expr ";" −> Requires {cons("Requires")}
29
"on_entry" Block −> OnEntry {cons("OnEntry")}
30
"on_exit" Block −> OnExit {cons("OnExit")}
31
32
FieldDec −> CoordinatorBodyDec
33
34 lexical syntax
35
"condition" −> Keyword
10-29
Language Oriented Modularity: From Theory to Practice
C
The Bounded Stack Example
Listing 19 Implementation of a non-thread-safe bounded stack
1 public class BoundedStack implements Stack {
2 protected Object[] buffer;
3 private int usedSlots = 0;
4 private static BoundedStack instance= null;
5 public BoundedStack(int capacity) { this.buffer = new Object[capacity]; instance=this; }
6 public static BoundedStack getInstance() { return instance; }
7 public Object pop() {
8
Object result = buffer[usedSlots − 1];
9
usedSlots−−;
10
buffer[usedSlots] = null;
11
return result;
12 }
13 public void push(Object obj) {
14
buffer[usedSlots++] = obj;
15 }
16 }
Listing 20 Aspect in Cool that synchronizes the bounded stack
1 coordinator base.BoundedStack {
2 selfex {push(java.lang.Object), pop()};
3 mutex {push(java.lang.Object), pop()};
4 condition full = false, empty = true;
5 int top = 0;
6 push(java.lang.Object):
7
requires (!full);
8
on_entry {top = top + 1;}
9
on_exit {
10
empty = false;
11
if (top == buffer.length) full = true;
12
}
13 pop():
14
requires (!empty);
15
on_entry {top = top − 1;}
16
on_exit {
17
full = false;
18
if (top == 0) empty = true;
19
}
20 }
Listing 21 Aspect in AspectJ that audits method calls using the bounded stack
1 public aspect AJAuditor {
2 pointcut toLog(): call(* *.*(..)) && !cflow(within(org.openu.demo.AJAuditor));
3 before(): toLog() { log(thisJoinPoint); }
4 protected void log(JoinPoint jp) {
5
BoundedStack buf = BoundedStack.getInstance();
6
try { if (buf != null) buf.add(jp); } catch(Exception e) { System.out.println(e.getMessage()); }
7 }
8 }
10-30
A. Hadas and D. H. Lorenz
D
Code Scattering in oVirt
Figures 3 and 4 illustrate scattering of synchronization, auditing, and permissions
related code in oVirt, marked in yellow, green, and red, respectively.
Figure 3 Code scattering in the MigrateVmCommand class
10-31
Language Oriented Modularity: From Theory to Practice
Figure 4 Code scattering in the AddDiskCommand class
10-32
A. Hadas and D. H. Lorenz
E
Grammar Definition for oVirtAudit, oVirtSync, and oVirtPermissions
Listing 22 Grammar definition for oVirtAudit in Xtext
1
2
3
4
5
6
7
8
Model: (commands+=Command)* ;
Command:
'logs for' type=[types::JvmDeclaredType|QualifiedName] (overrides?='(overrides)')? ':'
(cases+=Case(',' cases+=Case)* (',' 'otherwise' 'log' default=[types::JvmEnumerationLiteral])?)?
';'
;
Case:
'case' (actionState=[types::JvmEnumerationLiteral] '&')? result=Result ('&' internal?='internal')?
,→ ('&' 'state='(fields+=[types::JvmField]))* ('&' (methods+=[types::JvmOperation]))* 'log'
,→ msg=[types::JvmEnumerationLiteral] ;
9 enum Result: success|failure;
10 QualifiedName: ID ("." ID)*;
Listing 23 Grammar definition for oVirtSync in Xtext
1 Model: (commands+=Command)* ;
2 Command:
3 'locks for' type=[types::JvmDeclaredType|QualifiedName] '(' scope=Scope (wait?=('& wait'))? ')'
,→ ':'
4
(exclusiveLocks=Exclusive)? (sharedLocks=Inclusive)? (message=Message)?
5 ';'
6 ;
7 enum Scope: sync|async;
8 Exclusive:
9 {Exclusive} 'exclusively' (override?='(overrides)')? '{' (locks+=Lock(',' locks+=Lock)*)? '}' ;
10 Inclusive:
11 {Inclusive} 'inclusively' (override?='(overrides)')? '{' (locks+=Lock(',' locks+=Lock)*)? '}' ;
12 Lock: 'group: ' group=[types::JvmEnumerationLiteral] 'instance: ' id=[types::JvmOperation]
,→ (conditional?='if' condition=[types::JvmOperation])?;
13 Message: 'message: ' type=[types::JvmEnumerationLiteral] (vars+=Var)*;
14 Var: '<' key=STRING ',' value=[types::JvmOperation] '>';
15 QualifiedName: ID ("." ID)*;
Listing 24 Grammar definition for oVirtPermissions in Xtext
1
2
3
4
5
6
7
8
9
Model: commands+=Command* ;
Command:
'permissions for' type=[types::JvmDeclaredType|QualifiedName] (extends=Extends)? ':'
(permissions+=Permission (',' permissions+=Permission)*)?
';'
;
Extends: {Extends} '(extends)' | '(extends if' cond=Condition ')' ;
Permission:
'object type = ' objectType=[types::JvmEnumerationLiteral] 'object id = '
,→ objectId=[types::JvmOperation] 'action group = '
,→ actionGroup=[types::JvmEnumerationLiteral] (conditional?='if' conditions+=Condition
,→ ('and' conditions+=Condition)*)? ;
10 Condition: (not?='not')? operation=[types::JvmOperation] ;
11 QualifiedName: ID ("." ID)*;
10-33
Language Oriented Modularity: From Theory to Practice
F
Aspects for Synchronization and for Auditing in oVirt
Listing 25 Synchronizing commands for oVirt expressed in oVirtSync
1 locks for org.ovirt.engine.core.bll.ExportVmTemplateCommand (sync):
2 exclusively (overrides) { group: REMOTE_TEMPLATE instance: getVmTemplateId }
3 inclusively (overrides) { group: TEMPLATE instance: getVmTemplateId }
4 message: ACTION_TYPE_FAILED_TEMPLATE_IS_BEING_EXPORTED <"TemplateName",
,→ getVmTemplateName>;
5
6 locks for org.ovirt.engine.core.bll.MigrateVmCommand (async):
7 exclusively (overrides) { group: VM instance: getVmId }
8 message: ACTION_TYPE_FAILED_VM_IS_BEING_MIGRATED <"VmName", getVmName>;
9
10 locks for org.ovirt.engine.core.bll.AddDiskCommand (sync):
11 exclusively (overrides) { group: VM_DISK_BOOT instance: getVmId if isBootableDisk }
12 inclusively (overrides) { group: VM instance: getVmId };
Listing 26 Auditing commands for oVirt expressed in oVirtAudit
1 logs for org.ovirt.engine.core.bll.MigrateVmCommand (overrides):
2 case success & isReturnValueUp log VM_MIGRATION_DONE,
3 case success & internal log VM_MIGRATION_START_SYSTEM_INITIATED,
4 case success log VM_MIGRATION_START,
5 case failure & isHostInPrepareForMaintenance log
,→ VM_MIGRATION_FAILED_DURING_MOVE_TO_MAINTENANCE,
6 case failure log VM_MIGRATION_FAILED;
7
8 logs for org.ovirt.engine.core.bll.storage.export.ExportVmTemplateCommand:
9 case EXECUTE & success log IMPORTEXPORT_STARTING_EXPORT_TEMPLATE,
10 case EXECUTE & failure log IMPORTEXPORT_EXPORT_TEMPLATE_FAILED,
11 case END_SUCCESS & success log IMPORTEXPORT_EXPORT_TEMPLATE,
12 case END_SUCCESS & failure log IMPORTEXPORT_EXPORT_TEMPLATE_FAILED;
13
14 logs for org.ovirt.engine.core.bll.storage.disk.AddDiskCommand :
15 case EXECUTE & success & internal & isDiskStorageTypeRequiresExecuteState log
,→ ADD_DISK_INTERNAL,
16 case EXECUTE & success & isDiskStorageTypeRequiresExecuteState & isVmNameExists log
,→ USER_ADD_DISK_TO_VM,
17 case EXECUTE & success & isDiskStorageTypeRequiresExecuteState log USER_ADD_DISK,
18 case EXECUTE & failure & internal & isDiskStorageTypeRequiresExecuteState log
,→ ADD_DISK_INTERNAL_FAILURE,
19 case EXECUTE & failure & isDiskStorageTypeRequiresExecuteState & isVmNameExists log
,→ USER_FAILED_ADD_DISK_TO_VM,
20 case EXECUTE & failure & isDiskStorageTypeRequiresExecuteState log USER_FAILED_ADD_DISK,
21 case success & isVmNameExists log USER_ADD_DISK_TO_VM_FINISHED_SUCCESS,
22 case success log USER_ADD_DISK_FINISHED_SUCCESS,
23 case failure & isVmNameExists log USER_ADD_DISK_TO_VM_FINISHED_FAILURE,
24 otherwise log USER_ADD_DISK_FINISHED_FAILURE;
10-34
A. Hadas and D. H. Lorenz
G
Tool Support for Programming with oVirtSync
Figure 5 illustrates the general editing tools and aspect development tools that are
available while programming with oVirtSync in Eclipse using our approach. An IDE
plugin that is generated by Xtext provides one with general editing tools such as texthighlighting, auto-completion (line 17) and syntax-error checking (line 17). In addition,
the transformation of the DSAL into AspectJ and the use of the @BridgedSourceLocation
annotation enable one to leverage aspect development tools provided by AJDT such
as advises markers (lines 1 and 8).
Figure 5 Aspect in oVirtSync that synchronizes commands in oVirt, edited in Eclipse
10-35
Language Oriented Modularity: From Theory to Practice
H
Tool Support for Programming with muAudit
Figure 6 Aspect in muAudit that audits jobs in muCommander, editted in Eclipse
Figure 7 Code in muCommander being advised by aspect in muAudit
10-36
A. Hadas and D. H. Lorenz
About the authors
Arik Hadas is a software engineer at Red Hat and a graduate
student at the Open University of Israel, under the supervision
of Prof. David H. Lorenz. His research interests include different
aspects of programming and particularly large and distributed
software development and modularization techniques. He received
his BSc in computer science from the Ben-Gurion University in
the Negev, Israel. He is a member of the ACM. Contact him at the
Dept. of Mathematics and Computer Science, Open University of
Israel, Raanana 4310701, Israel; [email protected].
David H. Lorenz is an Associate Professor in the Department of
Mathematics and Computer Science at the Open University of Israel. He is currently a Visiting Professor at the Faculty of Computer
Science, Technion—Israel Institute of Technology. His research
interests include language-oriented software engineering, modularity, and programming, particularly involving domain specific
languages. Prof. Lorenz received his PhD in computer science from
the Technion—Israel Institute of Technology. He is a member of
the ACM and a member of the IEEE. Contact him at the Dept. of
Mathematics and Computer Science, Open University of Israel,
Raanana 4310701, Israel; [email protected].
10-37
| 6cs.PL
|
Under consideration for publication in Theory and Practice of Logic Programming
1
Stepwise Debugging of Answer-Set Programs
arXiv:1705.06564v1 [cs.AI] 18 May 2017
JOHANNES OETSCH
Technische Universität Wien,
Institut für Informationssysteme 184/3,
Favoritenstrasse 9-11, A-1040 Vienna, Austria,
(e-mail: [email protected])
JÖRG PÜHRER
Universität Leipzig,
Institut für Informatik,
Augustusplatz 10, D-04109 Leipzig, Germany,
(e-mail: [email protected])
HANS TOMPITS
Technische Universität Wien,
Institut für Informationssysteme 184/3,
Favoritenstrasse 9-11, A-1040 Vienna, Austria,
(e-mail: [email protected])
Abstract
We introduce a stepping methodology for answer-set programming (ASP) that allows for
debugging answer-set programs and is based on the stepwise application of rules. Similar
to debugging in imperative languages, where the behaviour of a program is observed
during a step-by-step execution, stepping for ASP allows for observing the effects that
rule applications have in the computation of an answer set. While the approach is inspired
from debugging in imperative programming, it is conceptually different to stepping in
other paradigms due to non-determinism and declarativity that are inherent to ASP. In
particular, unlike statements in an imperative program that are executed following a strict
control flow, there is no predetermined order in which to consider rules in ASP during a
computation. In our approach, the user is free to decide which rule to consider active in the
next step following his or her intuition. This way, one can focus on interesting parts of the
debugging search space. Bugs are detected during stepping by revealing differences between
the actual semantics of the program and the expectations of the user. As a solid formal
basis for stepping, we develop a framework of computations for answer-set programs. For
fully supporting different solver languages, we build our framework on an abstract ASP
language that is sufficiently general to capture different solver languages. To this end,
we make use of abstract constraints as an established abstraction for popular language
constructs such as aggregates. Stepping has been implemented in SeaLion, an integrated
development environment for ASP. We illustrate stepping using an example scenario and
discuss the stepping plugin of SeaLion. Moreover, we elaborate on methodological aspects
and the embedding of stepping in the ASP development process.
Under consideration in Theory and Practice of Logic Programming (TPLP).
2
J. Oetsch, J. Pührer, and H. Tompits
1 Introduction
Answer-set programming (ASP) (Niemelä 1999; Marek and Truszczyński 1999) is a
paradigm for declarative problem solving that is popular amongst researchers in artificial intelligence and knowledge representation. Yet it is rarely used by software
engineers outside academia so far. Arguably, one obstacle preventing developers
from using ASP is a lack of support tools for developing answer-set programs. One
particular problem in the context of programming support is debugging of answerset programs. Due to the fully declarative semantics of ASP, it can be quite tedious
to detect an error in an answer-set program. In recent years, debugging in ASP
has received some attention (Brain and De Vos 2005; Syrjänen 2006; Brain et al.
2007; Pührer 2007; Gebser et al. 2008; Gebser et al. 2009; Pontelli et al. 2009; Oetsch
et al. 2010; Oetsch et al. 2010; Oetsch et al. 2011; Oetsch et al. 2012b; Polleres et al.
2013; Frühstück et al. 2013; Shchekotykhin 2015). These previous works are important contributions towards ASP development support, however current approaches
come with limitations to their practical applicability. First, existing techniques and
tools only capture a basic ASP language fragment that does not include many
language constructs that are available and frequently used in modern ASP solver
languages, e.g.,, aggregates or choice rules are not covered by current debugging
strategies (with the exception of the work by Polleres et al. (2013), where cardinality constraints are dealt with by translation). Second, usability aspects are often
not considered in current approaches, in particular, the programmer is required to
either provide a lot of data to a debugging system or he or she is confronted with a
huge amount of information from the system (tackling this problem in query-based
debugging has been addressed by Shchekotykhin (2015)).
This paper introduces a stepping methodology for ASP, which is a novel technique for debugging answer-set programs that is general enough to deal with current
ASP solver languages and is intuitive and easy to use. Our method is similar in
spirit to the widespread and effective debugging strategy in imperative programming, where the idea is to gain insight into the behaviour of a program by executing
statement by statement following the program’s control flow. In our approach, we
allow for stepwise constructing interpretations by considering rules of an answer-set
program at hand in a successive manner. This method guarantees that either an
answer set will be reached, or some error will occur that provides hints why the
semantics of the program differs from the user’s expectations. A major difference to
the imperative setting is that, due to its declarativity, ASP lacks any control flow.
Instead, we allow the user to follow his or her intuition on which rule instances to
become active. This way, one can focus on interesting parts of the debugging search
space from the beginning. For illustration, the following answer-set program has
{a} as its only answer set.
a :- not b.
b :- not a.
a :- b.
Let’s step through the program to obtain explanations why this is the case. In
Stepwise Debugging of Answer-Set Programs
3
the beginning of a stepping session, no atom is considered to be true. Under this
premise, the first two rules are active. The user decides which of these rules to
apply. Choosing a rule to be applied in this manner is considered a step in our
approach. In case the user chooses the first rule, the atom a is derived. Then, no
further rule is active and one of the answer sets, {a} has been reached. If, on the
other hand, the user chooses the second rule in the first step, atom b is derived and
a is considered false. Then, the third rule becomes active. However, this rule would
derive a that is already considered false when choosing the second rule. In this case,
the user sees that no answer set can be reached based on the initial choice.
Besides single steps that allow the user to consider one rule instance at a time, we
also lay the ground for so-called jumps. The intuition is that in a jump multiple rule
instances and even multiple non-ground rules can be considered at once. Jumping
significantly speeds up the stepping process which makes our technique a usable
tool for debugging in practice. Consider the following encoding of an instance of
the three-colouring problem in the Gringo language (Gebser et al. 2011):
1{color(X,red;green;blue)}1 :- node(X).
:- edge(X,Y), color(X,C), color(X,C).
node(X):-edge(X,Y).
node(Y):-edge(X,Y).
edge(1,2). edge(1,3). edge(1,4).
edge(2,4). edge(2,5). edge(2,6).
edge(3,4). edge(3,5). edge(3,6).
edge(4,5). edge(5,6).
The user expects the program to have answer sets but it does not. Following our
approach, the reason for that can be found after two actions. First, trusting the
“instance” part of the program, the user applies a jump on all rules of this part,
and, intuitively, gets all atoms implied by these rules as an intermediate result.
Second, the user applies an arbitrary instance of the rule
1{color(X,red;green;blue)}1 :- node(X).
that is active under the atoms derived during the jump. Suppose, the user chooses
the instance
1 {color(1, red), color(1, green), color(1, blue)} 1 :- node(1).
and selects color(1, red) to be true. Then, the debugging system reveals that
the instance
:- edge(1, 2), color(1, red), color(1, red).
of the “check” constraint becomes unexpectedly active. Now, the users sees that
the second occurrence of color(X,C) in the constraint has to be replaced by
color(Y,C). Generally, bugs can be detected whenever stepping reveals differences
between the actual semantics of the program and the expectations of the user.
4
J. Oetsch, J. Pührer, and H. Tompits
In order to establish a solid formal basis for our stepping technique, we developed
a framework of computations for answer-set programs. For fully supporting current
solver languages, we were faced with several challenges. For one, the languages of
answer-set solvers differ from each other and from formal ASP languages in various
ways. In order to develop a method that works for different solvers, we need an
abstract ASP language that is sufficiently general to capture actual solver languages.
To this end, we make use of abstract constraints (Marek and Remmel 2004; Marek
and Truszczyński 2004) as an established abstraction for language constructs such
as aggregates, weight constraints, and external atoms. We rely on a semantics for
arbitrary abstract-constraint programs with disjunctions that we introduced for
this purpose in previous work (Oetsch et al. 2012a). In contrast to other semantics
for this type of programs, it is compatible with the semantics of all the ASP solvers
we want to support, namely, Clasp (Gebser et al. 2012), DLV (Leone et al. 2006),
and DLVHEX (Redl 2016). Note that our framework for computations for abstractconstraint programs differs from the one by Liu et al. (2010). We did not build on
this existing notion for three reasons. First, it does not cover rules with disjunctive
heads which we want to support. Second, steps in this framework correspond to
the application of multiple rules. Since our method is rooted in the analogy to
stepping in procedural languages, where an ASP rule corresponds to a statement in
an imperative language, we focus on steps corresponding to application of a single
rule. Finally, the semantics of non-convex literals differs from that of DLVHEX in
the existing approach. A thorough discussion on the relation of the two notions of
computations is given in Section 5.
Another basic problem deals with the grounding step in which variables are
removed from answer-set programs before solving. In formal ASP languages, the
grounding of a program consists of all rules resulting from substitutions of variables
by ground terms. In contrast, actual grounding tools apply many different types of
simplifications and pre-evaluations for creating a variable-free program. In order to
close this gap between formal and practical ASP, Pührer developed abstractions of
the grounding step together with an abstract notion of non-ground answer-set program as the base language for the stepping methodology in his PhD thesis (Pührer
2014). Based on that, stepping can easily be applied to existing solver languages
and it becomes robust to changes to these languages. As we focus on the methodological aspects of stepping in this article, we do not present these abstractions and
implicitly use grounding as carried out by actual grounding tools.
The stepping technique has been implemented in SeaLion (Oetsch et al. 2013),
an integrated development environment for ASP. We discuss how SeaLion can be
used for stepping answer-set programs written in the Gringo or the DLV language.1
1
The framework introduced in this paper subsumes and significantly extends previous versions
of the stepping technique for normal logic programs (Oetsch et al. 2010; Oetsch et al. 2011) and
DL-programs (Oetsch et al. 2012b).
Stepwise Debugging of Answer-Set Programs
5
Outline
Next, we provide the formal background that is necessary for our approach. We
recall the syntax of disjunctive abstract-constraint programs and the semantics on
which we base our framework (Oetsch et al. 2012a). Section 3 introduces a framework of computations that allows for breaking the semantics down to the level of
individual rules. After defining states and computations, we show several properties
of the framework, most importantly soundness and completeness in the sense that
the result of a successful computation is an answer set and that every answer set
can be constructed with a computation. Moreover, we study language fragments for
which a simpler form of computation suffices. In Section 4, we present the stepping
technique for debugging answer-set programs based on our computation framework.
We explain steps and jumps as a means to progress in a computation using an example scenario. Moreover, we discuss methodological aspects of stepping on the
application level (how stepping is used for debugging and program analysis) and
the top level (how stepping is embedded in the ASP development process). We
illustrate the approach with several use cases and describe the stepping interface of
SeaLion. Related work is discussed in Section 5. We compare stepping to other debugging approaches for ASP and discuss the relation of our computation framework
to that of Liu et al. (2010) and transition systems for ASP (Lierler 2011; Lierler
and Truszczyński 2016; Brochenin et al. 2014). We conclude the paper in Section 6.
In Appendix A we compile guidelines for stepping and give general recommendations for ASP development. Selected and short proofs are included in the main
text and all remaining proofs are provided in Appendix B.
2 Background
As motivated in the introduction, we represent grounded answer-set programs by
abstract-constraint programs (Marek and Remmel 2004; Marek and Truszczyński
2004; Oetsch et al. 2012a). Non-ground programs will be denoted by programs in
the input language of Gringo. Thus, we implicitly assume that grounding translates
(non-ground) Gringo rules to rules of abstract-constraint programs. For a detailed
formal account of our framework in the non-ground setting we refer the interested
reader to the dissertation of Pührer (2014).
We assume a fixed set A of ground atoms.
Definition 1
An interpretation is a set I ⊆ A of ground atoms. A ground atom a is true under
interpretation I , symbolically I |= a, if a ∈ I , otherwise it is false under I .
We will use the symbol 6|= to denote the complement of a relation denoted with the
symbol |= in different contexts.
For better readability, we sometimes make use of the following notation when
the reader may interpret the intersection of two sets I and X of ground atoms as
a projection from I to X .
Definition 2
For two sets I and X of ground atoms, I |X = I ∩ X is the projection of I to X .
6
J. Oetsch, J. Pührer, and H. Tompits
2.1 Syntax of Abstract-Constraint Programs
Rule heads and bodies of abstract-constraint programs are formed by so-called
abstract-constraint atoms.
Definition 3 (Marek and Remmel 2004; Marek and Truszczyński 2004)
An abstract constraint, abstract-constraint atom, or C-atom, is a pair A = hD, C i,
where D ⊆ A is a finite set called the domain of A, denoted by DA , and C ⊆ 2D
is a collection of sets of ground atoms, called the satisfiers of A, denoted by CA .
We can express atoms also as C-atoms. In particular, for a ground atom a, we
identify the C-atom h{a}, {{a}}i with a. We call such C-atoms elementary.
We will also make use of default negation in abstract-constraint programs. An
abstract-constraint literal, or C-literal, is a C-atom A or a default negated C-atom
notA.
Unlike the original definition, we introduce abstract-constraint programs with
disjunctive rule heads.
Definition 4
An abstract-constraint rule, or simply C-rule, is an expression of the form
A1 ∨ · · · ∨ Ak ← Ak +1 , . . . , Am , not Am+1 , . . . , not An ,
(1)
where 0 ≤ k ≤ m ≤ n and any Ai , for 1 ≤ i ≤ n, is a C-atom.
Note that if all disjuncts share the same domain they can be expressed by a single
C-atom (see Pührer 2014) but in general disjunction adds expressivity.
We identify different parts of a C-rule and introduce some syntactic properties.
Definition 5
For a C-rule r of form (1),
•
•
•
•
B(r ) = {Ak +1 , . . . , Am , not Am+1 , . . . , not An } is the body of r ,
B+ (r ) = {Ak +1 , . . . , Am } is the positive body of r ,
B− (r ) = {Am+1 , . . . , An } is the negative body of r , and
H(r ) = {A1 , . . . , Ak } is the head of r .
If B(r ) = ∅ and H(r ) 6= ∅, then r is a C-fact. For C-facts, we usually omit the
symbol “←”. A C-rule r of form (1) is normal if k = 1 and positive if m = n.
We define the domain of a default negated C-atom notA as Dnot A = DA . Then,
the domain DS of a set S of C-literals is given by
[
DS =
DL ·
L∈S
Finally, the domain of a C-rule r is
[
Dr =
DX ·
X ∈H(r )∪B(r )
Stepwise Debugging of Answer-Set Programs
7
Definition 6
An abstract-constraint program, or simply C-program, is a finite set of C-rules. A
C-program is normal, respectively positive, if it contains only normal, respectively
positive, C-rules. A C-program is elementary if it contains only elementary Catoms.
2.2 Satisfaction Relation
Intuitively, a C-atom hD, C i is a literal whose truth depends on the truth of all
atoms in D under a given interpretation. The satisfiers in C explicitly list which
combinations of true atoms in D make the C-atom true.
Definition 7
An interpretation I satisfies a C-atom hD, C i, symbolically I |= hD, C i, if I |D ∈ C .
Moreover, I |= not hD, C i iff I 6|= hD, C i.
Important criteria for distinguishing classes of C-atoms are concerned with their
semantic behaviour with respect to growing (or shrinking) interpretations. In this
respect, we identify monotonicity properties in the following.
Definition 8
A C-literal L is monotone if, for all interpretations I and I 0 , if I ⊆ I 0 and I |= L,
then also I 0 |= L. L is convex if, for all interpretations I , I 0 , and I 00 , if I ⊆ I 0 ⊆ I 00 ,
I |= L, and I 00 |= L, then also I 0 |= L. Moreover, a C-program P is monotone
(respectively, convex) if for all r ∈ P all C-literals L ∈ H(r ) ∪ B(r ) are monotone
(respectively, convex).
Next, the notion of satisfaction is extended to C-rules and C-programs in the obvious way.
Definition 9
An interpretation I satisfies a set S of C-literals, symbolically I |= S , if I |= L
for all L ∈ S . For brevity, we will use the notation I |=∃ S to denote that I |= L
for some L ∈ S . Moreover, I satisfies a C-rule r , symbolically I |= r , if I |= B(r )
implies I |=∃ H(r ). A C-rule r such that I |= B(r ) is called active under I . As
well, I satisfies a set P of C-rules, symbolically I |= P , if I |= r for every r ∈ P . If
I |= P , we say that I is a model of P .
2.3 Viewing ASP Constructs as Abstract Constraints
We want to use abstract constraints as a uniform means to represent common
constructs in ASP solver languages. As an example, we recall how weight constraints (Simons et al. 2002) can be expressed as C-atoms. In a similar fashion,
we can use them as abstractions of e.g., aggregates (Faber et al. 2004; Faber et al.
2011) or external atoms (Eiter et al. 2005). Note that the relation between abstract
constraints and ASP constructs is well known and motivated abstract constraints
in the first place (cf. Marek and Remmel 2004; Marek and Truszczyński 2004).
8
J. Oetsch, J. Pührer, and H. Tompits
Definition 10 (Simons et al. 2002)
A weight constraint is an expression of form
l [a1 = w1 , . . . , ak = wk , not ak +1 = wk +1 , . . . , not an = wn ] u ,
(2)
where each ai is a ground atom and each weight wi is a real number, for 1 ≤ i ≤ n.
The lower bound l and the upper bound u are either a real number, ∞, or −∞.
For a weight constraint to be true, the sum of weights wi of those atoms ai , 1 ≤
i ≤ k , that are true and the weights of the atoms ai , k < i ≤ n, that are false must
lie within the lower and the upper bound. Thus, a weight constraint of form (2)
corresponds to the C-atom hD, C i, where the domain D = {a1 , . . . , an } consists of
the atoms appearing in the weight constraint and
X
X
C = {X ⊆ D | l ≤ (
wi +
wi ) ≤ u} ·
1≤i≤k ,ai ∈X
k <i≤n,ai 6∈X
2.4 Semantics and Characterisations based on External Support and
Unfounded Sets
The semantics we use (Oetsch et al. 2012a) extends the FLP-semantics (Faber
et al. 2004; 2011) and coincides with the original notion of answer sets by Gelfond
and Lifschitz (1991) on many important classes of logic programs, including elementary C-programs. Similar to the original definition of answer sets, Faber et al.
make use of a program reduct depending on a candidate interpretation I for determining whether I satisfies a stability criterion and thus is considered an answer
set. However, the reduct of Faber, Leone, and Pfeifer differs in spirit from that of
Gelfond and Lifschitz as it does not reduce the program to another syntactic class
(the Gelfond-Lifschitz reduct of an elementary C-program is always positive). Instead, the so-called FLP-reduct, defined next, keeps the individual rules intact and
just ignores all rules that are not active under the candidate interpretation.
Definition 11
Let I be an interpretation, and let P be a C-program. The FLP-reduct of P with
respect to I is given by P I = {r ∈ P | r is active under I }.
The notion of answer sets for abstract-constraint programs defined next provides
the semantic foundation for the computation model we use for debugging.
Definition 12 (Oetsch et al. 2012a)
Let P be a C-program, and let I be an interpretation. I is an answer set of P if
I |= P and there is no I 0 ⊂ I such that P , I , and I 0 satisfy the following condition:
(?) for every r ∈ P I with I 0 |= B(r ), there is some A ∈ H(r ) with I 0 |= A and
I 0 |DA = I |DA .
The set of all answer sets of P is denoted by AS(P ).
Stepwise Debugging of Answer-Set Programs
9
The purpose of Condition (?) is to prevent minimisation within C-atoms: the requirement I 0 |DA = I |DA ensures that a satisfier {a, b} can enforce b to be true in an
answer set even if the same C-atom has a satisfier {a}. As a consequence answer
sets need not be subset minimal (see Pührer 2014 for details).
Our choice of semantics has high solver compatibility as its objective as we want
to support Gringo, DLV, and DLVHEX. We need an FLP-style treatment of nonconvex literals for being compatible with DLVHEX, disjunctions to support DLV and
DLVHEX, and we must allow for weight constraints in rule heads for compatibility with Gringo. Note that Gringo/Clasp treats aggregates in the way suggested
by Ferraris (2011). As a consequence, its semantics differs from our semantics in
some cases, when recursion is used through negated c-atoms, as ours is an extension
of the FLP semantics. For an in-depth comparison of FLP-semantics and Ferraris
semantics we refer to work by Truszczyński (2010). An example where the semantics differ is given by the single rule Gringo program a :- not 0{a}0 that has only
the empty set as answer set under our semantics, whereas Clasp also admits {a} as
an answer set. In practice, this difference only hardly influences the compatibility
with Gringo, as aggregates are seldom used in this way. We examined all Gringo
encodings send to the second ASP competition and could not find any such usage.
Our framework of computations for stepping is based on a characterisation of
the semantics of Definition 12 in terms of external supports. Often, answer sets are
computed following a two-step strategy: First, a model of the program is built, and
second, it is checked whether this model obeys a foundedness condition ensuring
that it is an answer set. Intuitively, every set of atoms in an answer set must be
“supported” by some active rule that derives one of the atoms. Here, it is important
that the reason for this rule to be active does not depend on the atom it derives.
Such rules are referred to as external support (Lee 2005). The extension of this
notion to our setting is the following.
Definition 13 (Oetsch et al. 2012a)
Let r be a C-rule, X a set of atoms, and I an interpretation. Then, r is an external
support for X with respect to I if
(i)
(ii)
(iii)
(iv)
I |= B(r ),
I \ X |= B(r ),
there is some A ∈ H(r ) with X |DA 6= ∅ and I |DA ⊆ S , for some S ∈ CA , and
for all A ∈ H(r ) with I |= A, (X ∩ I )|DA 6= ∅ holds.
Condition (i) ensures that r is active. Condition (ii) prevents self-support by guaranteeing the support to be “external” of X , i.e., r is also be active without the
atoms in X . In case I is a model, Items (iii) and (iv) jointly ensure that there is
some C-atom A in the head of r that is satisfied by I and derives some atom of X .
We can express the absence of an external support in an interpretation by the
concept of an unfounded set.
Definition 14 (Oetsch et al. 2012a)
Let P be a C-program, X a set of atoms, and I an interpretation. Then, X is
unfounded in P with respect to I if there is no C-rule r ∈ P that is an external
support for X with respect to I .
10
J. Oetsch, J. Pührer, and H. Tompits
Corollary 1 (Oetsch et al. 2012a)
Let P be a C-program and I an interpretation. Then, I is an answer set of P iff I
is a model of P and there is no set X with ∅ ⊂ X ⊆ I that is unfounded in P with
respect to I .
3 Computation Framework
In this section, we want to break the conceptual complexity of the semantics down
to artefacts the programmer is familiar with: the rules the user has written or,
more precisely, their ground instances. To this end, we introduce a framework of
computations that captures the semantics described in the previous section. In this
computation model, on top of which we will introduce stepping in Section 4, an
interpretation is built up step-by-step by considering an increasing number of rule
instances to be active. A computation in our framework is a sequence of states
which are structures that keep information which rules and atoms have already
been considered and what truth values were assigned to those atoms. Utilising the
framework, only one rule and the atoms it contains have to be considered at once
while building up an interpretation until an answer set is reached or a source for
the unexpected behaviour becomes apparent.
In the next two subsections, we introduce states and computations. In Section 3.3,
we define and show some properties of computations that we need later on when
we describe stepping. Section 3.4 is concerned with the existence of a stable computation which is a simpler form of computation that suffices for many popular
classes of answer-set programs. We discuss existing work related to our computation framework later in Section 5.
3.1 States
Our framework is based on sequences of states, reassembling computations, in which
an increasing number of ground rules are considered that build up a monotonically growing interpretation. Besides that interpretation, states also capture literals
which cannot become true in subsequent steps and sets that currently lack external
support in the state’s interpretation.
Definition 15
A state structure S is a tuple hP , I , I − , Υi, where P is a set of C-rules, I is an
interpretation, I − a set of atoms such that I and I − are disjoint, and Υ is a
collection of sets of atoms. We call DS = I ∪ I − the domain of S and define
PS = P , IS = I , I − S = I − , and ΥS = Υ.
A state structure hP , I , I − , Υi is a state if
(i) I |= B(r ) and I |=∃ H(r ) for every r ∈ P ,
(ii) Dr ⊆ DS for every r ∈ P , and
(iii) Υ = {X ⊆ I | X is unfounded in P with respect to I }.
We call h∅, ∅, ∅, {∅}i the empty state.
Stepwise Debugging of Answer-Set Programs
11
Intuitively, we use the first component P of a state to collect C-rules that the user
has considered to be active and satisfied. The interpretation I collects atoms that
have been considered true. Condition (i) ensures that P and I are compatible in the
sense that every C-rule that is considered active and satisfied is active and satisfied
with respect to I . Dual to I , the interpretation I − collects atoms that the user
has considered to be false. We require that all atoms appearing in a C-rule in P is
either in I or in I − which is expressed in Condition (ii). Finally, the set Υ keeps
track of unfounded subsets of I , as stated in Condition (iii). Intuitively, as we will
see later, when building a computation, the aim is to get rid of all unfounded sets
(except for the empty set) in order to compute an answer set of a C-program. If a
state does not contain such unfounded sets then we call it stable:
Definition 16
A state S is stable if IS ∈ AS(PS ).
The intuition is that when a state S is stable, no more C-rules need to be added to
PS to provide missing external support for the atoms in the current interpretation
IS . Note that a state S is stable exactly when ΥS = {∅}. For example, the empty
state is a stable state.
Example 1
Consider the C-rules
r1 : h{a, b}, {∅, {a}, {b}, {a, b}i ← not a
r2 :
b←a
and the state structures
S1 =h{r1 }, ∅, {a, b}, {∅}i,
S3 =h{r1 }, {a, b}, ∅, {∅}i,
S5 =h{r2 }, {a, b}, ∅, {{b}, {a, b}}i·
S2 =h{r1 }, {b}, {a}, {∅}i,
S4 =h{r2 }, {a, b}, ∅, {∅}i,
S1 and S2 are stable states. S3 is not a state as IS3 6|= B(r1 ). S4 is not a state as the
sets {b} and {a, b} are unfounded in PS4 with respect to IS4 but {b} 6∈ ΥS4 and
{a, b} 6∈ ΥS4 . S5 is a state but not stable.
3.2 Computations
Next, we show how we can proceed forward in a computation, i.e., which states
might follow a given state using a successor relation for state structures.
Definition 17
−
For a state S = hP , I , I − , Υi and a state structure S 0 = hP 0 , I 0 , I 0 , Υ0 i, S 0 is a
0
−
successor of S if there is a C-rule r ∈ P \ P and sets ∆, ∆ ⊆ Dr such that
(i)
(ii)
(iii)
(iv)
(v)
P 0 = P ∪ {r },
−
I 0 = I ∪ ∆, I 0 = I − ∪ ∆− , and DS ∩ (∆ ∪ ∆− ) = ∅,
Dr ⊆ DS 0 ,
I |= B(r ),
I 0 |= B(r ) and I 0 |=∃ H(r ), and
12
J. Oetsch, J. Pührer, and H. Tompits
(vi) X 0 ∈ Υ0 iff X 0 = X ∪ ∆0 , where X ∈ Υ, ∆0 ⊆ ∆, and r is not an external
support for X 0 with respect to I 0 .
We denote r by rnew (S , S 0 ).
Condition (i) ensures that a successor state considers exactly one rule more to
be active. Conditions (ii) and (iii) express that the interpretations I and I − are
extended by the so far unconsidered literals in ∆ and ∆− appearing in the new
C-rule rnew (S , S 0 ). Note that from S 0 being a state structure we get that ∆ and
∆− are distinct. A requirement for considering rnew (S , S 0 ) as next C-rule is that it
is active under the current interpretation I , expressed by Condition (iv). Moreover,
rnew (S , S 0 ) must be satisfied and still be active under the succeeding interpretation,
as required by Condition (v). The final condition ensures that the unfounded sets of
the successor are extensions of the previously unfounded sets that are not externally
supported by the new rule.
Here, it is interesting that only extended previous unfounded sets can be unfounded sets in the extended C-program P 0 and that rnew (S , S 0 ) is the only C-rule
which could provide external support for them in P 0 with respect to the new interpretation I 0 as seen next.
Theorem 1
Let S be a state and S 0 a successor of S , where ∆ = IS 0 \ IS . Moreover, let X 0 be
a set of literals with ∅ ⊂ X 0 ⊆ IS 0 . Then, the following statements are equivalent:
(i) X 0 is unfounded in PS 0 with respect to IS 0 .
(ii) X 0 = ∆0 ∪ X , where ∆0 ⊆ ∆, X ∈ ΥS , and rnew (S , S 0 ) is not an external
support for X 0 with respect to IS 0 .
The result shows that determining the unfounded sets in a computation after adding
a further C-rule r can be done locally, i.e., only supersets of previously unfounded
sets can be unfounded sets, and if such a superset has some external support then
it is externally supported by r . The result also implies that the successor relation
suffices to “step” from one state to another.
Corollary 2
Let S be a state and S 0 a successor of S . Then, S 0 is a state.
Proof
We show that the Conditions (i), (ii), and (iii) of Definition 15 hold for S 0 . Consider
some rule r ∈ PS 0 . In case r = rnew (S , S 0 ), IS 0 |= B(r ) and IS 0 |=∃ H(r ) hold
because of Item (v) of Definition 17 and Dr ⊆ DS 0 because of Item (iii) of the same
definition. Moreover, in case r 6= rnew (S , S 0 ), we have r ∈ PS . As S is a state, we
have Dr ⊆ DS . Hence, since DS ⊆ DS 0 also Dr ⊆ DS 0 . Note that IS 0 |Dr = IS |Dr
because of Item (ii) of Definition 17. Therefore, as IS |= B(r ) and IS |=∃ H(r ), also
IS 0 |= B(r ) and IS 0 |=∃ H(r ). From these two cases, we see that Conditions (i) and
(ii) of Definition 15 hold for S 0 . Finally, Condition (iii) follows from Item (vi) of
Definition 17 and Theorem 1.
Next, we define computations based on the notion of a state.
Stepwise Debugging of Answer-Set Programs
13
Definition 18
A computation is a sequence C = S0 , . . . , Sn of states such that Si+1 is a successor
of Si , for all 0 ≤ i < n. We call C rooted if S0 is the empty state and stable if each
Si is stable, for 0 ≤ i ≤ n.
3.3 Properties
We next define when a computation has failed, gets stuck, is complete, or has
succeeded. Intuitively, failure means that the computation reached a point where
no answer set of the C-program can be reached. A computation is stuck when
the last state activated rules deriving literals that are inconsistent with previously
chosen active rules. It is considered complete when there are no more unconsidered
active rules. Finally, a computation has succeeded if an answer set has been reached.
Definition 19
Let P be a C-program and C = S0 , . . . , Sn a computation such that PSn ⊆ P .
Then, C is called a computation for P . Moreover,
• C has failed for P at step i if there is no answer set I of P such that ISi ⊆ I ,
I − Si ∩ I = ∅, and PSi ⊆ P I ;
• is complete for P if for every rule r ∈ P ISn , we have r ∈ PSn ;
• is stuck in P if it is not complete for P but there is no successor Sn+1 of Sn
such that rnew (Sn , Sn+1 ) ∈ P ;
• succeeded for P if it is complete and Sn is stable.
Example 2
Let PEx 2 be the C-program consisting of the C-rules
r1
r2
r3
r4
r5
:
a
:
b
:
a
: h{c}, {∅, {c}}i
:
←
←
←
←
←
h{a, b}, {∅, {a, b}}i
a
b
c
that has {a, b} as its single answer set, and consider the sequences
•
•
•
•
•
•
•
C1 =h∅, ∅, ∅, {∅}i,
h{r4 }, {}, {c}, {∅}i,
h{r4 , r1 }, {a, b}, {c}, {{a}, {b}}i,
C2 =h∅, ∅, ∅, {∅}i, h{r4 }, {}, {c}, {∅}i, h{r4 , r1 }, {a, b}, {c}, {{a}, {b}}i,
h{r4 , r1 , r2 }, {a, b}, {c}, {{a}}i, h{r4 , r1 , r2 , r3 }, {a, b}, {c}, {∅}i,
C3 =h{r4 , r1 , r2 , r3 }, {a, b}, {c}, {∅}i,
C4 =h∅, ∅, ∅, {∅}i, h{r4 }, {c}, ∅, {∅}i,
C5 =h{r4 , r1 , r2 , r3 }, {a, b, c}, ∅, {∅}i,
C6 =h{r5 }, ∅, {c}, {∅}i,
and
C7 =h∅, ∅, ∅, {∅}i, h{r4 , r1 }, {a, b}, {c}, {{a}, {b}}i·
14
J. Oetsch, J. Pührer, and H. Tompits
C1 , C2 , C3 , C4 , and C5 are computations for PEx 2 . The sequence C6 is not a
computation, as h{r5 }, ∅, {c}, {∅}i is not a state. C7 is not a computation, as the
second state in C7 is not a successor of the empty state. C1 , C2 , and C4 are rooted.
C3 , C4 , and C5 are stable. C2 and C3 are complete and have succeeded for PEx 2 .
C1 is complete for PEx 2 \ {r2 , r3 } but has failed for PEx 2 \ {r2 , r3 } at Step 0 because
PEx 2 \ {r2 , r3 } has no answer set. C4 has failed for PEx 2 at Step 1. C5 has failed for
PEx 2 at Step 0 and is stuck in PEx 2 .
The following result guarantees the soundness of our framework of computations.
Theorem 2
Let P be a C-program and C = S0 , . . . , Sn a computation that has succeeded for
P . Then, ISn is an answer set of P .
Proof
As C is complete for P , we have P ISn ⊆ PSn . Conversely, we have PSn ⊆ P ISn
because for each r ∈ PSn we have r ∈ P and ISn |= B(r ). By stability of Sn , we get
that ISn ∈ AS(PSn ). The conjecture holds since then ISn ∈ AS(P ISn ).
The computation model is also complete in the following sense:
Theorem 3
Let S0 be a state, P a C-program with PS0 ⊆ P , and I an answer set of P with
IS0 ⊆ I and I ∩ I − S0 = ∅. Then, there is a computation S0 , . . . , Sn that has
succeeded for P such that PSn = P I and ISn = I .
As the empty state, h∅, ∅, ∅, {∅}i, is trivially a state, we can make the completeness
aspect of the previous result more apparent in the following corollary:
Corollary 3
Let P be a C-program and I ∈ AS(P ). Then, there is a rooted computation
S0 , . . . , Sn that has succeeded for P such that PSn = P I and ISn = I .
Proof
The claim follows immediately from Theorem 3 in case S0 = h∅, ∅, ∅, {∅}i.
Note, that there are states that do not result from rooted computations, e.g.,
the state h{a ← b}, {a, b}, ∅, {∅, {a, b}, {b}}i is not a successor of any other state.
However, for stable states, we can guarantee the existence of rooted computations.
Corollary 4
Let S be a stable state. Then, there is a rooted computation S0 , . . . , Sn with Sn = S .
Proof
The result is a direct consequence of Corollary 3 and Definition 16.
The next theorem lays the ground for the jumping technique that we introduce
in Section 4. It allows for extending a computation by considering multiple rules of
a program at once and using ASP solving itself for creating this extension.
Stepwise Debugging of Answer-Set Programs
15
Theorem 4
Let P be a C-program, C = S0 , . . . , Sn a computation for P , P 0 a set of C-rules
with P 0 ⊆ P , and I an answer set of PSn ∪ P 0 with ISn ⊆ I and I ∩ I − Sn = ∅. Then,
there is a computation C 0 = S0 , . . . , Sn , Sn+1 , . . . , Sm for P , such that Sm is stable,
I
PSm = PSn ∪ P 0 and ISm = I .
Proof
By Theorem 3, as PSn ⊆ PSn ∪P 0 , ISn ⊆ I , and I ∩I − Sn = ∅, there is a computation
I
Sn , . . . , Sm that has succeeded for PSn ∪ P 0 such that PSm = (PSn ∪ P 0 ) and
I
ISm = I . Then, Sm is stable and, as PSn I = PSn , we have PSm = PSn ∪ P 0 . As
0
PSm ⊆ P , we have that C = S0 , . . . , Sn , Sn+1 , . . . , Sm is a computation for P .
The following result illustrates that the direction one chooses for building up a
certain interpretation, i.e., the order of the rules considered in a computation, is
irrelevant in the sense that eventually the same state will be reached.
Proposition 1
0
Let P be a C-program and C = S0 , . . . , Sn and C 0 = S00 , . . . , Sm
computations
0
0
complete for P such that S0 = S0 . Then, ISn = ISm0 iff Sn = Sm and n = m.
Proof
The “if” direction is trivial. Let I = ISn = ISm0 . Towards a contradiction, assume
PSn 6= PSm0 . Without loss of generality, we focus on the case that there is some
r ∈ PSn such that r 6∈ PSm0 . Then, it holds that I |= r , I |= B(r ), and r ∈ P .
Consequently, r ∈ P I . By completeness of C 0 , we have r ∈ PSm0 which contradicts
our assumption. Hence, we have PSn = PSm0 .
By definition of a state, from ISn = ISm0 and PSn = PSm0 , it follows that ΥSn =
ΥSm0 . Towards a contradiction, assume I − Sn 6= I − Sm0 . Without loss of generality we
focus on the case that there is some a ∈ I − Sn such that a 6∈ I − Sm0 . Consider the
integer i where 0 < i ≤ n such that a ∈ I − Si but a 6∈ I − Si−1 . Then, by definition
of a successor, for r = rnew (Si−1 , Si ), we have a ∈ ∆− for some ∆− ⊆ Dr . As
then a ∈ Dr and, as PSn = PSm0 , we have r ∈ PSm0 , it must hold that a ∈ DSm0 by
definition of a state structure. From I ∩ I − Sn = ∅ we know that a 6∈ I . Therefore,
since a ∈ I ∪ I − Sm0 , we get that a ∈ I − Sm0 , being a contradiction to our assumption.
0
As then Sn = Sm
, PS0 = PS00 , and since in every step in a computation exactly one
rule is added it must hold that n = m.
For rooted computations, the domain of each state is determined by the atoms
in the C-rules it contains.
Proposition 2
Let C = S0 , . . . , Sn be a rooted computation. Then, ISi = ISn |DPS and I − Si =
i
I − Sn |DPS , for all 0 ≤ i ≤ n.
i
16
J. Oetsch, J. Pührer, and H. Tompits
Proof
The proof is by contradiction. Let j be the smallest index with 0 ≤ j ≤ n such that
ISj 6= ISn |DPS or I − Sj 6= I − Sn |DPS . Note that 0 < j as IS0 = I − S0 = DPS0 = ∅. As
j
j
Sj is a successor of Sj −1 , we have ISj = ISj −1 ∪ ∆ and I − Sj = I − Sj −1 ∪ ∆− , where
∆, ∆− ⊆ Drnew (Sj −1 ,Sj ) , DSj −1 ∩ (∆ ∪ ∆− ) = ∅, and Drnew (Sj −1 ,Sj ) ⊆ ISj ∪ I − Sj . As
we have ISj −1 = ISn |DPS
and I − Sj −1 = I − Sn |DPS , it holds that
j −1
j −1
ISj −1 ∪ ISn |Dδ = ISn |DPS
j −1
∪ ISn |Dδ = ISn |DPS and
j
I − Sj −1 ∪ I − Sn |Dδ = I − Sn |DPS
j −1
∪ I − Sn |Dδ = I − Sn |DPS ,
j
where Dδ = DPSj \ DPSj −1 . For establishing the contradiction, it suffices to show
that ISn |Dδ = ∆ and I − Sn |Dδ = ∆− . Consider some a ∈ ∆. Then, a ∈ Dδ because
a ∈ Drnew (Sj −1 ,Sj ) , DSj −1 ∩ (∆ ∪ ∆− ) = ∅, and DPSj −1 ⊆ DSj −1 . Moreover, a ∈ ISj
implies a ∈ ISn and therefore ∆ ⊆ ISn |Dδ . Now, consider some b ∈ ISn |Dδ . As
Drnew (Sj −1 ,Sj ) ⊆ ISj ∪ I − Sj , we have b ∈ ISj ∪ I − Sj . Consider the case that b ∈ I − Sj .
Then, also b ∈ I − Sn which is a contradiction to b ∈ ISn as Sn is a state structure.
Hence, b ∈ ISj = ISj −1 ∪ ∆. First, assume b ∈ ISj −1 . This leads to a contradiction
as then b ∈ DPSj −1 since ISj −1 = ISn |DPS . It follows that b ∈ ∆ and therefore
j −1
∆ = ISn |Dδ . One can show that ∆− = I − Sn |Dδ analogously.
3.4 Stable Computations
In this section, we are concerned with the existence of stable computations, i.e.,
computations that do not involve unfounded sets. We single out an important class
of C-programs for which one can solely rely on this type of computation and also
give examples of C-programs that do not allow for succeeding stable computations.
Intuitively, the ΣP
2 -hardness of the semantics (cf. Pührer 2014), demands for
unstable computations in the general case. This becomes obvious when considering
that for a given C-program one could guess a candidate sequence C for a stable
computation in polynomial time. Then, a polynomial number of checks whether
each state is a successor of the previous one in the sequence suffices to establish
whether C is a computation. Following Definition 17, these checks can be done in
polynomial time when we are allowed to omit Condition (vi) for unfounded sets.
Hence, answer-set existence for the class of C-programs for which every answer set
can be built up with stable computations is in NP.
Naturally, it is interesting whether there are syntactic classes of C-programs for
which we can rely on stable computations only. It turns out that many syntactically
simple C-programs already require the use of unfounded sets.
Example 3
Consider C-program PEx 3 consisting of the C-rules
r1 : a ← b
and
r2 : b ← h{a}, {∅, {a}}i·
Stepwise Debugging of Answer-Set Programs
17
We have that {a, b} is the only answer set of PEx 3 and
C =h∅, ∅, ∅, {∅}i,
h{r2 }, {a, b}, ∅, {∅, {a}}i,
h{r2 , r1 }, {a, b}, ∅, {∅}i
is the only computation that succeeds for PEx 3 : starting at the empty state, only
rule r2 is active, thus it must be the new rule in the successor. When, deciding the
truth values for the atoms in Dr2 , r2 requires b to be positive, and a must be true
as well, as otherwise the computation is stuck due to violation of r1 . The second
state of C contains the singleton {a} as unfounded set.
As Example 3 shows, unstable computations are already required for a C-program
without disjunction and a single monotone C-atom. Hence, also the use of weaker
restrictions, like convexity of C-atoms or some notion of head-cycle freeness (BenEliyahu and Dechter 1994), is not sufficient.
One can observe, that the C-program from the example has cyclic positive dependencies between atoms a and b. Hence, we next explore whether such dependencies
influence the need for computations that are not stable. To this end, we introduce
notions of positive dependency in a C-program.
Definition 20
Let S be a set of C-literals. Then, the positive normal form of S is given by
S + = {A | A ∈ S , A is a C-atom} ∪ {Ā | not A ∈ S },
where Ā = hDA , 2DA \ CDA i is the complement of A. Furthermore, the set of positive
S
atom occurrences in S is given by posOcc(S ) = A∈S + ,X ∈CA X .
Let P be a C-program. The positive dependency graph of P is the directed graph
G(P ) = hDP , {ha, bi | r ∈ P , a ∈ posOcc(H(r )), b ∈ posOcc(B(r ))}i·
We next introduce the notion of absolute tightness for describing C-programs
without cyclic positive dependencies after recalling basic notions of graph theory.
For a (directed) graph G = hV , ≺i, the reachability relation of G is the transitive
closure of ≺. Let ≺0 be the reachability relation of G. Then, G is acyclic if there is
no v ∈ V such that v ≺0 v .
Definition 21
Let P be a C-program. P is absolutely tight if G(P ) is acyclic.
One could assume that absolute tightness paired with convexity or monotonicity is sufficient to guarantee stable computations because absolute tightness forbids
positive dependencies among disjuncts and the absence of such dependencies lowers
the complexity of elementary C-programs (Ben-Eliyahu and Dechter 1994). However, as the following example illustrates, this is not the case for general C-programs.
Example 4
Consider C-program PEx 4 consisting of the C-rules
r1 : a ∨ h{a, b}, {{a}, {a, b}}i ←
r2 : b ∨ h{a, b}, {{b}, {a, b}}i ←
18
J. Oetsch, J. Pührer, and H. Tompits
We have that {a, b} is the only answer set of PEx 4 and
C1 =h∅, ∅, ∅, {∅}i,
h{r1 }, {a, b}, ∅, {∅, {b}}i,
h{r1 , r2 }, {a, b}, ∅, {∅}i
and
C2 =h∅, ∅, ∅, {∅}i,
h{r2 }, {a, b}, ∅, {∅, {a}}i,
h{r1 , r2 }, {a, b}, ∅, {∅}i
are the only computations that succeed for PEx 4 . Clearly, PEx 4 is monotone and
absolutely tight but C1 and C2 are not stable.
Nevertheless, we can assure the existence of stable computations for answer sets
of normal C-programs that are absolutely tight and convex. This is good news, as
this class corresponds to a large subset of typical answer-set programs written for
solvers like Clasp that do not rely on disjunction as their guessing device.
Theorem 5
Let C = S0 , . . . , Sn be a computation such that S0 and Sn are stable and P∆ =
PSn \ PS0 is a normal, convex, and absolutely tight C-program. Then, there is a
stable computation C0 = S00 , . . . , Sn0 such that S0 = S00 and Sn = Sn0 .
As a direct consequence of Theorem 5 and Corollary 3, we get an improved completeness result for normal convex C-programs that are absolutely tight, i.e., we
can find a computation that consists of stable states only.
Corollary 5
Let P be a normal C-program that is convex and absolutely tight, and consider
some I ∈ AS(P ). Then, there is a rooted stable computation S0 , . . . , Sn such that
PSn = P I and ISn = I .
Proof
From I ∈ AS(P ), we get by Corollary 3 that there is a rooted computation
S0 , . . . , Sn such that PSn = P I and ISn = I . Note that S0 is the empty state.
S0 and Sn are stable according to Definition 16. From Theorem 5, we can conclude the existence of another computation C0 = S00 , . . . , Sn0 such that S0 = S00 and
Sn = Sn0 that is stable. Clearly, C0 is also rooted.
4 Theory and Practice of Stepping
In this section we present our methodology for stepping answer-set programs based
on the computation model introduced in the previous section.
Step-by-step execution of a program is common practice in procedural programming languages, where developers can debug and investigate the behaviour of their
programs in an incremental way. The technique introduced in this work shows how
this popular form of debugging can be applied to ASP, despite the genuine declarative semantics of answer-set programs that lacks a control flow. Its main application
Stepwise Debugging of Answer-Set Programs
19
is debugging but it is also beneficial in other contexts such as improving the understanding of a given answer-set program or teaching the answer-set semantics to
beginners.
For stepping to be a practical support technique for answer-set programmers
rather than a purely theoretical approach, we assume the availability of a support
environment that assists a user in a stepping session. A prototype our stepping
framework has been implemented in SeaLion, an integrated development environment (IDE) for ASP (Oetsch et al. 2013). It was developed as one of the major
goals of a research project on methods and methodologies for developing answerset programs conducted at Vienna University of Technology (2009-2013). SeaLion
supports the ASP languages of Gringo and DLV and comes as a plugin of the
Eclipse platform that is popular for Java development. All the features of a stepping support environment described in this section are implemented in SeaLion, if
not stated otherwise.
To bridge the gap between theory and practical stepping, our examples use solver
syntax rather than the abstract language of c-programs. We implicitly identify
solver constructs with their abstract counterparts (cf. Section 2.3). While we use
c-programs as a formal lingua franca for different solver languages, we do not require the user to know about it. Likewise, we do not expect a user to be aware of
the specifics of the computation framework of Section 3 that provides the backbone
of stepping. For example, the user does not need to know the properties of Definition 17. The debugging environment automatically restricts the available choices a
user has when performing a step to ones that result in valid successor states.
In the following subsection, we introduce an example scenario that we use later
on. In Section 4.2, we describe the general idea of stepping for ASP. There are two
major ways for navigating in a computation in our framework: performing steps,
discussed in Section 4.3, and jumps that we describe in Section 4.4. In Section 4.5,
we describe the stepping interface of SeaLion. Building on steps and jumps, we
discuss methodological aspects of stepping for debugging purposes in Section 4.6
and provide a number of use cases.
4.1 Example Scenario - Maze Generation
Next, we introduce the problem of maze generation that serves as a running example
in the remainder of the paper. It has been a benchmark problem of the second ASP
competition (Denecker et al. 2009) to which it was submitted by Martin Brain. The
original problem description is available on the competition’s website.2
As the name of the problem indicates, the task we deal with is to generate a maze,
i.e., a labyrinth structure in a grid that satisfies certain conditions. In particular, we
deal with two-dimensional grids of cells where each cell can be assigned to be either
an empty space or a wall. Moreover, there are two (distinct) empty squares on the
edge of the grid, known as the entrance and the exit. A path is a finite sequence of
2
http://dtai.cs.kuleuven.be/events/ASP-competition/Benchmarks/MazeGeneration.shtml
20
J. Oetsch, J. Pührer, and H. Tompits
Fig. 1: Left: A grid visualising an instance of the maze generation where white
squares represent empty cells, whereas grey squares are yet undefined. Right: A
solution for the instance on the left.
cells, in which each distinct cell appears at most once and each cell is horizontally
or vertically adjacent to the next cell in the sequence.
Such a grid is a valid maze if it meets the following criteria:
1. Each cell is a wall or is empty.
2. There must be a path from the entrance to every empty cell (including the
exit).
3. If a cell is on any of the edges of the grid, and is not an entrance or an exit,
it must contain a wall.
4. There must be no 2x2 blocks of empty cells or walls.
5. No wall can be completely surrounded by empty cells.
6. If two walls are diagonally adjacent then one or other of their common neighbours must be a wall.
The maze generation problem is the problem of completing a two-dimensional grid
in which some cells are already decided to be empty or walls and the entrance and
the exit are pre-defined to a valid maze. An example of a problem instance and a
corresponding solution maze is depicted in Fig. 1.
Next, we describe the predicate schema that we use for ASP maze generation encodings. The predicates col/1 and row/1 define the columns and rows in the grid,
respectively. They are represented by a range of consecutive, ascending integers,
starting at 1. The positions of the entrance and the exit are determined by predicates entrance/2 and exit/2, respectively, where the first argument is a column
index and the second argument is a row index. In a similar manner, empty/2 and
wall/2 determine which cells are empty or contain walls. For example, the instance
of Fig. 1 can be encoded by program Π1 consisting of the following facts:
col(1..5). row(1..5).
entrance(1,2). exit(5,4). wall(3,3). empty(3,4).
Moreover, the solution in the figure could be represented by the following interpretation (projected to predicates empty/2 and wall/2):
{wall(1,1), empty(1,2), wall(1,3), wall(1,4), wall(1,5),
Stepwise Debugging of Answer-Set Programs
wall(2,1),
wall(3,1),
wall(4,1),
wall(5,1),
21
empty(2,2), empty(2,3), empty(2,4), wall(2,5),
wall(3,2), wall(3,3), empty(3,4), wall(3,5),
empty(4,2), empty(4,3), empty(4,4), wall(4,5),
wall(5,2), wall(5,3), empty(5,4), wall(5,5)}
4.2 General Idea
We introduce stepping for ASP as a strategy to identify mismatches between the
intended semantics of an answer-set program under development and its actual
semantics. Due to the declarativity of ASP, once one detects unintended semantics,
it can be a tough problem to manually detect the reason. Stepping is a method for
breaking this problem into smaller parts and structuring the search for an error. The
general idea is to monotonically build up an interpretation by, in each step, adding
literals derived by a rule that is active with respect to the interpretation obtained in
the previous step. The process is interactive in the sense that at each such step the
user chooses the active rule to proceed with and decides which literals of the rule
should be considered true or false in the target interpretation. Hereby, the user only
adds rules he or she thinks are active in an expected or an unintended actual answer
set. The interpretation grows monotonically until it is eventually guaranteed to be
an answer set of the overall program, otherwise the programmer is informed why
and at which step something went wrong. This way, one can in principle without any
backtracking direct the computation towards the interpretation one has in mind.
In debugging, having the programmer in the role of an oracle is a common scenario
as it is reasonable to assume that a programmer has good intuitions on where to
guide the search (Shapiro 1982). We use the computation model of Section 3 to
ensure that, if the interpretation specified in this way is indeed an answer set, the
process of stepping will eventually terminate with the interpretation as its result.
Otherwise, the computation will fail at some step where the user gets insight why
the interpretation is not an answer set, e.g., when a constraint becomes irrevocably
active or no further rule is active that could derive some desired literal.
4.3 Steps
By a step we mean the extension of a computation by a further state. We consider a
setting, where a programmer has written an answer-set program in a solver language
for which C-program P it the abstraction of its grounding. Moreover, we assume
that the programmer has obtained some computation for P that is neither stuck
in P nor complete for P . For performing a step, one needs to find a successor state
Sn+1 for Sn such that C 0 = S0 , . . . , Sn+1 is a computation for P .
We propose a sequence of three user actions to perform a step. Intuitively, for
quickly finding a successor state (with the help of the debugging environment), we
suggest to
1. select a non-ground rule with active ground instances, then
2. choose an active ground rule among the instances of the non-ground rule, and
22
J. Oetsch, J. Pührer, and H. Tompits
3. select for yet undefined atoms in the domain of the ground instance whether
they are considered true or false.
First, the user selects a non-ground rule ρ. In SeaLion, this can be done by directly
selecting ρ in the editor in which the program was written. The debugging system
can support this user action by automatically determining the subset of rules in
the program that have at least one instance r in their grounding that could lead to
a successor state, i.e., r = rnew (Sn , S ) for some successor S of Sn .
Then, the user selects an instance r from the grounding of ρ. As the ground
instances of ρ are not part of the original program, picking one requires a different
approach as for choosing ρ. Here, the debugging environment can display the ground
rules in a dedicated area and, as before, restrict the choice of rule groundings of ρ to
ones that lead to a successor state. Filtering techniques can be used to restrict the
amount of the remaining instances. In SeaLion, the user defines partial assignments
for the variables in ρ that determine a subset of the considered instances.
In the third user action for performing a step, the programmer chooses the truth
values for the atoms in Dr that are neither in ISn nor in I − Sn . This choice must be
made in a way such that there is a successor Sn+1 of Sn with PSn+1 = PSn ∪ {r },
ISn+1 = ISn ∪∆, and I − Sn+1 = I − Sn ∪∆− , where ∆ contains the atoms the user chose
to be true and ∆− the atoms considered false. That is, Sn , ∆, and ∆− must fulfil the
conditions of Definition 17. Here, the user needs only to ensure that Condition (v)
of Definition 17 holds, i.e., ISn+1 |= B(r ) and ISn+1 |=∃ H(r ), as the other conditions
automatically hold once all unassigned atoms have been assigned to ∆ and ∆− . In
particular, the set of unfounded sets, ΥSn+1 can always be automatically computed
following Condition (vi) of Definition 17 and does not impose restrictions on the
choice of ∆ and ∆− . The support system can check whether Condition (v) holds for
the truth assignment specified by the user. Also, atoms are automatically assigning
to ∆ or ∆− whenever their truth values are the same for all successor states that
are based on adding r .
Example 5
As a first step for developing the maze-generation encoding, we want to identify
border cells and guess an assignment of walls and empty cells. Our initial program
is Π2 , given next.
maxCol(X) :- col(X), not col(X+1).
maxRow(Y) :- row(Y), not row(Y+1).
border(1,Y) :- col(1), row(Y).
border(X,1) :- col(X), row(1).
border(X,Y) :- row(Y), maxCol(X).
border(X,Y) :- col(X), maxRow(Y).
wall(X,Y) :- border(X,Y), not entrance(X,Y), not exit(X,Y).
{ wall(X,Y) : col(X), row(Y), not border(X,Y) }.
empty(X,Y) :- col(X), row(Y), not wall(X,Y).
The first two rules extract the numbers of columns and rows of the maze from the
Stepwise Debugging of Answer-Set Programs
23
input facts of predicates col/1 and row/1. The next four rules derive border/2
atoms that indicate which cells form the border of the grid. The final three rules
derive wall/2 atoms for border cells except entrance and exit, guess wall/2 atoms
for the remaining cells, and derive empty/2 atoms for non-wall cells, respectively.
We use Π2 in conjunction with the facts in program Π1 (defined in Section 4.1)
that determine the problem instance.
We start a stepping session with the computation C0 = S0 consisting of the empty
state S0 = h∅, ∅, ∅, {∅}i. Following the scheme of user actions described above for
performing a step, we first look for a non-ground rule with instances that are active
under IS0 . As IS0 = ∅, only the facts from Π1 have active instances. We choose the
rule entrance(1,2). In this case, the only (active) instance of the rule is identical
to the rule, i.e., the fact:
entrance(1,2).
The only atom in the domain of the rule instance is entrance(1,2). Therefore,
when performing the final user action for a step one has to decide the truth value
of this atom. In order to fulfil Condition (v) of Definition 17, the rule head, i.e.,
entrance(1,2), must be true in the successor state. Thus, our first step results in
the computation C1 = S0 , S1 where
S1 = h{entrance(1,2).}, {entrance(1,2)}, ∅, {∅}i·
For the next step, we choose the rule
col(1..5).
from Π1 . The grounding3 by Gringo consists of the following instances:
col(1). col(2). col(3). col(4). col(5).
We select the instance col(5). Since the head of the rule must be true under the
successor state, as before, atom col(5) must be considered true in the successor
state of S1 . The resulting computation after the second step is C2 = S0 , S1 , S2 ,
where
S2 = h{entrance(1,2).col(5).}, {entrance(1,2), col(5)}, ∅, {∅}i·
Under IS2 a further rule in Π1 ∪ Π2 has active instances:
maxCol(X) :- col(X), not col(X+1).
That is, it has the active instance
maxCol(5) :- col(5), not col(6).
that we choose for the next step. In order to ensure that Condition (v) of Definition 17 is satisfied, we need to ensure that head and body are satisfied under the
successor state. Hence, atom maxCol(5) has to be considered true, whereas col(6)
must be considered false. We obtain the computation C3 = S0 , S1 , S2 , S3 , where
S3 = h{entrance(1,2).col(5).maxCol(5) :- col(5), not col(6).},
{entrance(1,2), col(5), maxCol(5)}, {col(6)}, {∅}i·
3
Remember that grounding refers to the translation performed by the grounding tool rather than
mere variable elimination.
24
J. Oetsch, J. Pührer, and H. Tompits
4.4 Jumps
If one wants to simulate the computation of an answer set I in a stepping session using steps only, as many steps are necessary as there are active rules in the
grounding under I . Although, typically the number of active ground instances is
much less than the total number of rules in the grounding, still many rules would
have to be considered. In order to focus on the parts of a computation that the
user is interested in, we introduce a jumping technique for quickly considering rules
that are of minor interest, e.g., for rules that are already considered correct. We say
that we jump through these rules. By performing a jump, we mean to find a state
that could be reached by a computation for the program at hand that extends the
current computation by possibly multiple states. If such a state can be found, one
can continue to expand a computation from that while it is ensured that the same
states could be reached by using steps only. Jumps can be performed exploiting
Theorem 4. In essence, jumping is done as follows.
1. Select rules that you want to jump through (i.e., the rules you want to be
considered in the state to jump to),
2. an auxiliary answer-set program is created that contains the selected rules
and the active rules of the current computations final state, and
3. a new state is computed from an answer set of the auxiliary program.
Next, we describe the items in more detail. We assume that an answer-set program
in a solver language for which C-program P it the abstraction of its grounding and
a computation S0 , . . . , Sn for P are given.
The first user action is to select a subset P 0 of P , the rules to jump through.
Hence, an implication for a stepping support environment is the necessity of means
to select the ground instances that form P 0 . In case the user wants to consider
all instances of some non-ground rules, a very user friendly way of selecting these
is to simply select the non-ground rules and the environment implicitly considers
P 0 to be their instances. SeaLion implements this feature such that this selection
can be done in the editor in which the answer-set program is written. If the user
wants to jump through only some instances of a non-ground rule, we need further
user-interface features. In order to keep memory resources and the amount of rules
that have to be considered by the user low, the system splits the selection of an
instance in two phases. First, the user selects a non-ground rule ρ, similar as in the
first user action of defining a step. Then, the system provides so far unconsidered
rules of P for selection, where similar filtering techniques as sketched for the second
user action for performing a step can be applied.
The auxiliary program of the second item can be automatically computed. That
is a C-program Paux given by Paux = PSn ∪ P 0 ∪ Pcon , where
Pcon ={← not a | a ∈ ISn } ∪
{← a | a ∈ I − Sn }
is a set of constraints that ensure that for every answer set I of Paux we have
ISn ⊆ I and I ∩ I − Sn = ∅.
After computing an answer set I of the auxiliary program, Theorem 4 ensures
Stepwise Debugging of Answer-Set Programs
25
Choose truth values for atoms in Dr
Compute answer set of auxiliary
program for jumping through P ′
obtain a new state
Initialise Computation
set S to the empty state
Set P ′ = ∅ if desired.
Add a set of ground
instances of ρ to P ′
Choose a ground
instance r of ρ
Current state S
jump
step
Choose a non-ground rule ρ
Choose a non-ground rule ρ
Fig. 2: Stepping cycle
the existence of a computation C 0 = S0 , . . . , Sn , Sn+1 , . . . , Sm for P such that Sm
I
is stable and ISm = I . Moreover, we have PSm = PSn ∪ P 0 . Then, the user can
proceed with further steps or jumps extending the computation S0 , . . . , Sm as if Sm
had been reached by steps only.
Note that non-existence of answer sets of the auxiliary program does not imply
that the overall program has no answer sets as shown next.
Example 6
Consider a program P consisting of the C-rules a ← and ← not a that has {a} as
its unique answer set. Assume we want to jump through the second rule starting
from the computation C = h∅, ∅, ∅, {∅}i consisting of the empty state. Then, Paux =
{← not a} has no answer set.
The example shows that jumping only makes sense when the user is interested in
a computation reaching an answer set of the auxiliary program. In case of multiple
answer sets of the auxiliary program, the user could pick any or a stepping environment can choose one at random. For practical reasons, the second option seems
more preferable. On the one hand, presenting multiple answer sets to the user can
lead to a large amount of information that has to be stored and processed by the
user. And on the other hand, if the user is not happy with the truth value of some
atoms in an arbitrary answer set of the auxiliary program, he or she can use steps
to define the truth of these atoms before performing the jump. In SeaLion only one
answer set of the auxiliary program is computed.
The iterative extension of a computation using steps and jumps can be described
as a stepping cycle that is depicted in Fig. 2. It summarises how a user may advance
a computation and thus provides a technical level representation of stepping.
Example 7
We continue computation C3 for program Π1 ∪ Π2 from Example 5. As we are
interested in the final three rules of Π2 that derive empty/2 and wall/2 atoms but
these rules depend on atoms of predicate border/2, entrance/2, and exit/2 that
are not yet considered in C3 , we jump through the facts from Π1 and the rules
26
J. Oetsch, J. Pührer, and H. Tompits
maxCol(X) :- col(X), not col(X+1).
maxRow(Y) :- row(Y), not row(Y+1).
border(1,Y) :- col(1), row(Y).
border(X,1) :- col(X), row(1).
border(X,Y) :- row(Y), maxCol(X).
border(X,Y) :- col(X), maxRow(Y).
of program Π2 . The resulting auxiliary program Π3 is given by the following rules
(for non-ground rules, their unconsidered instances in the grounding of Π1 ∪ Π2 ).
entrance(1,2).
col(5).
maxCol(5) :- col(5), not col(6).
col(1..5). row(1..5).
exit(5,4). wall(3,3). empty(3,4).
maxCol(X) :- col(X), not col(X+1).
maxRow(Y) :- row(Y), not row(Y+1).
border(1,Y) :- col(1), row(Y).
border(X,1) :- col(X), row(1).
border(X,Y) :- row(Y), maxCol(X).
border(X,Y) :- col(X), maxRow(Y).
::::-
not entrance(1,2).
not col(5).
not maxCol(5).
col(6).
The program Π3 has the single answer set Iaux consisting of the atoms:
col(1), col(2), col(3), col(4), col(5), maxCol(5),
row(1), row(2), row(3), row(4), row(5), maxRow(5),
empty(3,4), wall(3,3), entrance(1,2), exit(5,4),
border(1,1), border(2,1), border(3,1), border(4,1),
border(5,1), border(1,2), border(5,2), border(1,3),
border(5,3), border(1,4), border(5,4), border(1,5),
border(2,5), border(3,5), border(4,5), border(5,5),
We obtain the new state S4 = hPS4 , Iaux , DPS4 \ Iaux , {∅}i, where PS4 consists of
the following rules:
col(1). col(2). col(3). col(4). col(5).
row(1). row(2). row(3). row(4). row(5).
wall(3,3). empty(3,4). entrance(1,2). exit(5,4).
maxCol(5) :- col(5), not col(6).
maxRow(5) :- row(5), not row(6).
border(1,1) :- col(1), row(1).
border(2,1) :- col(2), row(1).
Stepwise Debugging of Answer-Set Programs
border(3,1)
border(4,1)
border(5,1)
border(1,2)
border(5,2)
border(1,3)
border(5,3)
border(1,4)
border(5,4)
border(1,5)
border(5,1)
border(5,5)
border(1,5)
border(2,5)
border(3,5)
border(4,5)
border(5,5)
:::::::::::::::::-
col(3),
col(4),
col(5),
col(1),
row(2),
col(1),
row(3),
col(1),
row(4),
col(1),
row(1),
row(5),
col(1),
col(2),
col(3),
col(4),
col(5),
27
row(1).
row(1).
row(1).
row(2).
maxCol(5).
row(3).
maxCol(5).
row(4).
maxCol(5).
row(5).
maxCol(5).
maxCol(5).
maxRow(5).
maxRow(5).
maxRow(5).
maxRow(5).
maxRow(5).
Theorem 4 ensures the existence of a computation C4 = S0 , S1 , S2 , S3 , . . . , S4 for
program Π1 ∪ Π2 .
4.5 Stepping Interface of SeaLion
In the following, we focus on the stepping functionality of SeaLion that was implemented by our former student Peter Skočovský. While it is the first implementation
of the stepping technique for ASP and hence still a prototype, it is tailored for intuitive and user-friendly usage and able to cope with real-world answer-set programs.
The stepping feature is integrated with the Kara plugin of SeaLion (Kloimüllner
et al. 2013) that can create user-defined graphical representations of interpretations.
Thus, besides visualising answer sets, it is also possible to visualise intermediate
states of a stepping session. Visualisations in Kara are defined using ASP itself, for
further information we refer to earlier work (Kloimüllner et al. 2013). A comprehensive discussion of other features of SeaLion is given in a related paper (Busoniu
et al. 2013) on the IDE. SeaLion is published under the GNU General Public
License and can be obtained from www.sealion.at
A stepping session in SeaLion can be started in a similar fashion as debugging
Java programs in Eclipse using the launch configuration framework. SeaLion launch
configurations that are used for defining which program files should be run with
which solvers can be re-used as debug configurations.
Like many IDEs, Eclipse comes with a multiple document interface in which inner
frames, in particular Eclipse editors and views, can be arranged freely by the user.
Such configurations can be persisted as perspectives. Eclipse plugins often come
with default perspectives, i.e., arrangements of views and editors that are tailored
to a specific user task in the context of the plugin. Also the stepping plugin has a
preconfigured perspective that is opened automatically once a stepping session has
28
J. Oetsch, J. Pührer, and H. Tompits
Fig. 3: SeaLion’s stepping view is devided into five areas (a-e).
been initiated. The next subsection gives an overview of the individual stepping
related subframes in this perspective.
Fig. 3 shows SeaLion in the stepping perspective. The illustration distinguishes
five regions (marked by supplementary dashed frames and labelled by letters) for
which we give an overview in what follows.
The source code editor (Fig. 3a) is the same as used for writing answer-set programs but extended with additional functionality during stepping mode for the ASP
files involved in the stepping session. In particular it indicates rules with ground
instances that are active under the interpretation of the current stepping state.
Constraints with active instances are highlighted by a red background (cf. Fig. 9),
other rules with active instances have a blue background (as, e.g., in Fig. 4). The
editor remains functional during stepping, i.e., the program can be modified while
debugging. Note, however, that the system does not guarantee that the current
computation is still a valid computation in case of a modification of the answerset program after stepping has been initiated. The source code editor is also the
starting point for performing a step or a jump as it allows for directly selecting
the non-ground rule(s) to be considered in the step or jump in the source code.
The choice of non-ground rules corresponds to the initial step in the stepping cycle
(see Section 4.4). Selecting a single rule or consecutive rules is done by directly
selecting them in the source code editor. If the rules are non-consecutive, the user
must collect rules in the jump view located in area c of Fig. 3 as the second tab.
Stepwise Debugging of Answer-Set Programs
29
Choosing a ground instance for performing a step is done in the active instances
view (Fig. 3b). It contains a list with all active ground instances (with respect to
conditional grounding) of the currently selected rule in the source editor. As these
are potentially many, the view has a textfield for filtering the list of rules. Filters
are given as dot-separated list of variable assignments of the form X=t where X is a
variable of the non-ground rule and t is the ground term that the user considers X to
be assigned to. Only ground instances are listed that obey all variable substitutions
of the entered filters.
Once a rule instance is selected in the active instances view the atoms in the rule’s
domain are displayed in three lists of the truth assignment view (Fig. 3c). The list
in the centre shows atoms whose truth value has not already been determined in
the current state. The user can decide whether they should be true, respectively
false, in the next step by putting them into the list on the left, respectively, on the
right. These atoms can be transferred between the lists by using keyboard cursors
or drag-and-drop (Fig. 4). After the truth value has been decided for all the atoms
of the rule instance and only in case that the truth assignment leads to a valid
successor state (cf. Definition 17), a button labelled “Step” appears. Clicking this
button computes the new state.
The state view (Fig. 3d) shows the current stepping state of the debugging session. Hence, it is updated after every step or jump. It comprises four areas, corresponding to the components of the state (cf. Definition 15), the list of active rules
instances, a tree-shaped representation of the atoms considered true, a tree-shaped
representation of the atoms considered false, both in a similar graphical representation as that of interpretations in the interpretation view, and an area displaying
the unfounded sets in a similar way. The sets of atoms displayed in this view can
also be visualised using Kara (via options in the context menu).
Finally, the computation view (Fig. 3e) gives an overview of the steps and jumps
performed so far. Importantly, the view implements an undo-redo mechanism. That
is, by clicking on one of the nodes displayed in the view, representing a previous step
or jump, the computation can be reset to the state after this step or jump has been
performed. Moreover, after performing an undo operation, the undone computation
is not lost but becomes an inactive branch of the tree-shaped representation of steps
and jumps. Thus, one can immediately jump back to any state that has been reached
in the stepping session by clicking on a respective node in the tree (Fig. 6).
Mismatches between the users intentions (reflected in the current stepping state)
and the actual semantics of the program can be detected in different parts of the
stepping perspective. If the user thinks a rule instance should be active but it is
not, this can already be seen in the source code editor if the non-ground version
of the rule does not have any active instance. Then, the rule is not highlighted in
the editor. If the non-ground version does have active instances but not the one
the user has in mind, this can be detected after clicking on the non-ground rule
if they are missing in the active instances view. The computation is stuck if only
rules are highlighted in the source editor that are constraints (cf. Fig. 9) or for all of
its instances, no truth assignment can be established such that the “Step” button
appears.
30
J. Oetsch, J. Pührer, and H. Tompits
Finally, if no further rule is highlighted and there is no non-empty unfounded
set visible in the state view, the atoms considered positive form an answer set of
the overall program. If there are further unfounded sets, the user sees that the
constructed interpretation is not stable. The unfounded sets indicate which atoms
would need external support.
4.6 Methodology
We identify three three conceptual levels of stepping as a methodology. The technical level corresponds to the iterative advancement of a computation covered in
Sections 4.3 and 4.4 summarised in the stepping cycle (Fig. 2). Next, we describe
stepping on the application level as a method for debugging and program analysis.
After that, we highlight how stepping is embedded in the greater context of ASP
development from a top level perspective. Finally, we illustrate our approach in
different usage scenarios. In Appendix A, we compile practical guidelines for our
methodology.
Program Analysis and Debugging Level Methodology
The main purpose for stepping in the context of this paper is its application for
debugging and analysing answer-set programs. Next, we describe how insight into
a program is gained using stepping. During stepping, the user follows his or her
intuitions on which rule(s) to apply next and which atoms to consider true or false.
In this way, an interpretation is built up that either is or is not an answer set of the
program. In both cases, stepping can be used to analyse the interplay of rules in the
program in the same manner, i.e., one can see which rule instances become active
or inactive after each step or jump. In the case that the targeted interpretation
is an answer set of the program, the computation will never fail (in the sense of
Definition 19) or get stuck and will finally succeed. It can, however, happen that
intermediate states in the computation are unstable (cf. Example 3). For debugging,
stepping towards an answer set is useful if the answer set is unwanted. In particular,
one can see why a constraint (or a group of rules supposed to have a constraining
effect) does not become active. For instance, stepping reveals other active rules
that derive atoms that make some literal in the constraint false or rules that fail do
derive atoms that activate the constraint. Stepping towards an actual answer set of
a program is illustrated in Example 8.
In the case that there is an answer set that the user expects to be different,
i.e., certain atoms are missing or unwanted, it makes sense to follow the approach
that we recommend for expected but missing answer sets, i.e., stepping towards
the interpretation that the user wants to be an answer set. Then, the computation
is guaranteed to fail at some point, i.e., there is some state in the computation
from which no more answer set of the program can be reached. In other situations,
the computation can already have failed before the bug can be found, e.g., the
computation can have failed from the beginning in case the program has no answer
sets at all. Nevertheless, the error can be found when stepping towards the intended
Stepwise Debugging of Answer-Set Programs
31
interpretation. In most cases, there will be either a rule instance that becomes active
that the user considered inactive, or the other way around, i.e., a rule instance never
becomes active or is deactivated while the computation progresses. Eventually, due
to our completeness results in the previous section, the computation will either get
stuck or ends in an unstable state S such that no active external support for a nonempty unfounded set from ΥS is available in the program’s grounding. Stepping
towards an interpretation that is not an answer set of the overall program can be
seen as a form of hypothetical reasoning: the user can investigate how rules of a
part of the program support each other before adding a further rule would cause an
inconsistency. Example 9 illustrates stepping towards an intended but non-existing
answer set for finding a bug. Another illustration of hypothetical reasoning is given
in Example 10 where a user tries to understand why an interpretation that is not
supposed to be an answer set is indeed no answer set.
Note that stepping does not provide explanation artefacts, analogous to stepping
in imperative languages, where insight in a problem is gained by simply following
the control flow and watching variable assignments. In our setting, the user only
sees which remaining rule instances are active and the current state of the computation which in principle suffices to notice differences to his or her expectations.
Nevertheless, it can be useful to combine this approach with query based debugging
methods that provide explicit explanations as will be discussed in Section 5.
As mentioned in Section 4.2, if one has a clear idea on the interpretation one
expects to be an answer set, stepping allows for building up a computation for this
interpretation without backtracking. In practice, one often lacks a clear vision on
the truth value of each and every atom with respect to a desired answer set. As a
consequence, the user may require revising the decisions he or she has taken on the
truth values of atoms as well as on which rules to add to the computation. For this
reason, SeaLion allows for retracting a computation to a previous state, i.e., let the
user select one of the states in the computation and continue stepping from there.
This way a tree of states can be built up (as in Fig. 9), where every path from the
root node to a leaf node is a rooted computation.
Top-Level Methodology
Stepping must be understood as embedded in the programming and modelling
process, i.e., the technique has to be recognised in the context of developing answerset programs. A practical consequence of viewing stepping in the big picture are
several possibilities for exploiting information obtained during the development of
a program for doing stepping faster and more accurate.
While an answer-set program evolves, the programmer will in many cases compute answer sets of preliminary versions of the program for testing purposes. If this
answer sets are persisted, they can often be used as a starting point for stepping
sessions for later versions of the program. For instance, in case the grounding P of
a previous version of a program is a subset of the current grounding P 0 it is obvious
that a successful computation C for P is also a computation for P 0 . Hence, the
user can initiate a stepping session starting from C . Also in case that P 6⊂ P 0 , a
32
J. Oetsch, J. Pührer, and H. Tompits
stepping support system could often automatically build a computation that uses
an answer set of P (or parts of it) as guidance for deciding on the rules to add and
the truth values to assign in consecutive states. Likewise, (parts of) computations
of stepping sessions for previous versions of a program can be stored and re-used
either as a computation of the current program if applicable or for computing such
a computation. The idea is that previous versions of a program often constitute
a part of the current version that one is already familiar with and “trusts” in.
By starting from a computation that already considers a well-known part of the
program, the user can concentrate on new and often more suspicious parts of the
program. Currently, there is no such feature implemented in SeaLion.
Use Cases
Next, we show application scenarios of stepping using our running example.
The first scenario illustrates stepping towards an interpretation that is an answer
set of the program under consideration.
Example 8
We want to step towards an answer set of our partial encoding of the maze generation problem, i.e., of the program Π1 ∪ Π2 . Therefore, we continue our stepping
session with computation C4 , i.e., we start stepping from state S4 that we obtained
in Example 7. In particular, we want to reach an answer set that is compatible with
the maze generation solution depicted in Fig. 1. To this end, we start with stepping
through the active instances of the rule
{wall(X,Y) : col(X), row(Y), not border(X,Y)}.
The only active instance of the rule is
{wall(2,2), wall(3,2), wall(4,2), wall(2,3), wall(3,3),
wall(4,3), wall(2,4), wall(3,4), wall(4,4)}.
Thus, we next choose a truth assignment for the atoms appearing in the instance’s
choice atom. Note that we do not need to decide for the truth value of wall(3,3)
as it is already contained in IS4 and therefore already considered true. As can be
observed in Fig. 1, among the remaining cells the rule deals with, only the one at
position (3, 2) is a wall in our example. Hence, we obtain a new state S5 from S4
by extending PS4 by our rule instance, IS4 by wall(3,2), and I − S4 by wall(2,2),
wall(4,2), wall(2,3), wall(4,3), wall(2,4), wall(3,4), wall(4,4). As in S4 ,
the empty set is the only unfounded set in state S5 . It remains to jump through
the rules
wall(X,Y) :- border(X,Y), not entrance(X,Y), not exit(X,Y).
and
empty(X,Y) :- col(X), row(Y), not wall(X,Y).
that leads to the addition of instances
33
Stepwise Debugging of Answer-Set Programs
Fig. 4: Deciding to consider atom wall(3,2) to be true by dragging it from the
middle list of atoms in the truth assignment view and dropping it in the left list. The
current state of the stepping session is S4 from Example 8 and the step prepared is
that to state S5 . Note that atom wall(3,3) is greyed in the list of positive atoms.
Greyed atoms in the positive or negative list cannot be be dragged away from there
again because their truth value is already considered positive, respectively negative,
in the current state (here S4 ). A step can only be completed once the truth value
has been decided for all atoms in the rule instance.
wall(1,
wall(2,
wall(3,
wall(4,
wall(5,
wall(5,
wall(1,
wall(5,
wall(1,
wall(1,
wall(2,
wall(3,
wall(4,
wall(5,
1)
1)
1)
1)
1)
2)
3)
3)
4)
5)
5)
5)
5)
5)
::::::::::::::-
border(1,
border(2,
border(3,
border(4,
border(5,
border(5,
border(1,
border(5,
border(1,
border(1,
border(2,
border(3,
border(4,
border(5,
1),
1),
1),
1),
1),
2),
3),
3),
4),
5),
5),
5),
5),
5),
not
not
not
not
not
not
not
not
not
not
not
not
not
not
entrance(1,
entrance(2,
entrance(3,
entrance(4,
entrance(5,
entrance(5,
entrance(1,
entrance(5,
entrance(1,
entrance(1,
entrance(2,
entrance(3,
entrance(4,
entrance(5,
1),
1),
1),
1),
1),
2),
3),
3),
4),
5),
5),
5),
5),
5),
not
not
not
not
not
not
not
not
not
not
not
not
not
not
exit(1,
exit(2,
exit(3,
exit(4,
exit(5,
exit(5,
exit(1,
exit(5,
exit(1,
exit(1,
exit(2,
exit(3,
exit(4,
exit(5,
1).
1).
1).
1).
1).
2).
3).
3).
4).
5).
5).
5).
5).
5).
34
J. Oetsch, J. Pührer, and H. Tompits
empty(1,
empty(2,
empty(4,
empty(2,
empty(4,
empty(2,
empty(3,
empty(4,
empty(5,
2)
2)
2)
3)
3)
4)
4)
4)
4)
:::::::::-
col(1),
col(2),
col(4),
col(2),
col(4),
col(2),
col(3),
col(4),
col(5),
row(2),
row(2),
row(2),
row(3),
row(3),
row(4),
row(4),
row(4),
row(4),
not
not
not
not
not
not
not
not
not
wall(1,
wall(2,
wall(4,
wall(2,
wall(4,
wall(2,
wall(3,
wall(4,
wall(5,
2).
2).
2).
3).
3).
4).
4).
4).
4).
to a new state S6 . IS6 extends IS5 by the head atoms of these rules that are not yet
in IS5 . Likewise, I − S6 extends I − S5 by the default negated atoms appearing in the
rules that are not yet in I − S5 . As ΥS6 = {∅} and no rule in Π1 ∪ Π2 has further
active instances under IS6 , the computation S0 , . . . , S6 has succeeded and hence IS6
is an answer set of the program.
In the next example, a bug is revealed by stepping towards an intended answer set.
Example 9
As a next feature, we (incorrectly) implement rules in program Π4 that should
express that there has to be a path from the entrance to every empty cell and that
2 × 2 blocks of empty cells are forbidden:
adjacent(X,Y,X,Y+1) :- col(X), row(Y), row(Y+1).
adjacent(X,Y,X,Y-1) :- col(X), row(Y), row(Y-1).
adjacent(X,Y,X+1,Y) :- col(X), row(Y), col(X+1).
adjacent(X,Y,X-1,Y) :- col(X), row(Y), col(X-1).
reach(X,Y)
:- entrance(X,Y), not wall(X,Y).
reach(XX,YY) :- adjacent(X,Y,XX,YY), reach(X,Y), not wall(XX,YY).
:- empty(X,Y), not reach(X,Y).
:- empty(X,Y), empty(X+1,Y), empty(X,X+1), empty(X+1,Y+1).
The first six rules formalise when an empty cell is reached from the entrance, and
the two constraints should ensure that every empty cell is reached and that no 2 × 2
blocks of empty cells exist, respectively.
Assume that we did not spot the bug in the second constraint—in the third body
literal the term Y+1 was mistaken for X+1. This could be the result of a typical copypaste error. It turns out that Π1 ∪ Π2 ∪ Π4 has no answer set. In order to find a
reason, one can start stepping towards an intended answer set. We assume that
the user already trusts the program Π1 ∪ Π2 from Example 8. Hence, he or she can
reuse the computation S0 , . . . , S6 for Π1 ∪ Π2 as starting point for a stepping session
because all rules in PS6 are also ground instances of rules in the extended program
Π1 ∪ Π2 ∪ Π4 . Then, when the user asks for rules with active ground instances a
stepping support environment would present the following rules:
adjacent(X,Y,X,Y+1) :- col(X), row(Y), row(Y+1).
adjacent(X,Y,X,Y-1) :- col(X), row(Y), row(Y-1).
Stepwise Debugging of Answer-Set Programs
col(1..5). row(1..5).
entrance(1,2). exit(5,4).
wall(3,3). empty(3,4).
maxCol(X) :- col(X), not col(X+1).
maxRow(Y) :- row(Y), not row(Y+1).
border(1,Y) :- col(1), row(Y).
border(X,1) :- col(X), row(1).
border(X,Y) :- row(Y), maxCol(X).
border(X,Y) :- col(X), maxRow(Y).
wall(X,Y) :- border(X,Y),
not entrance(X,Y),
not exit(X,Y).
{ wall(X,Y) : col(X), row(Y),
not border(X,Y) }.
empty(X,Y) :- col(X), row(Y),
not wall(X,Y).
adjacent(X,Y,X,Y+1) :- col(X), row(Y),
row(Y+1).
adjacent(X,Y,X,Y-1) :- col(X), row(Y),
row(Y-1).
adjacent(X,Y,X+1,Y) :- col(X), row(Y),
35
col(X+1).
adjacent(X,Y,X-1,Y) :- col(X), row(Y),
col(X-1).
reach(X,Y)
:- entrance(X,Y),
not wall(X,Y).
reach(XX,YY) :- adjacent(X,Y,XX,YY),
reach(X,Y),
not wall(XX,YY).
:- empty(X,Y), not reach(X,Y).
:- empty(X,Y), empty(X+1,Y),
empty(X,Y+1), empty(X+1,Y+1).
:- exit(X,Y), wall(X,Y).
:- wall(X,Y), wall(X+1,Y),
wall(X,Y+1), wall(X+1,Y+1).
:- wall(X,Y), empty(X+1;X-1,Y),
empty(X,Y+1;Y-1), col(X+1;X-1),
row(Y+1;Y-1).
:- wall(X,Y), wall(X+1,Y+1),
not wall(X+1,Y), not wall(X,Y+1).
:- wall(X+1,Y), wall(X,Y+1),
not wall(X,Y), not wall(X+1,Y+1).
Fig. 5: Program Π1 ∪ Π2 ∪ Π5 ∪ Π6 used in Example 10.
adjacent(X,Y,X+1,Y) :- col(X), row(Y), col(X+1).
adjacent(X,Y,X-1,Y) :- col(X), row(Y), col(X-1).
reach(X,Y)
:- entrance(X,Y), not wall(X,Y).
:- empty(X,Y), not reach(X,Y).
:- empty(X,Y), empty(X+1,Y), empty(X,X+1), empty(X+1,Y+1).
The attentive observer will immediately notice that two constraints are currently
active. There is no reason to be deeply concerned about
:- empty(X,Y), not reach(X,Y).
being active because the rule defining the reach/2 predicate—that can potentially
deactivate the constraint—has not been considered yet. However, the constraint
:- empty(X,Y), empty(X+1,Y), empty(X,X+1), empty(X+1,Y+1).
contains only atoms of predicate empty/2 that has already been fully evaluated.
Even if empty/2 was only partially evaluated, an active instance of the constraint
could not become inactive in a subsequent computation for the sole ground that it
only contains monotonic literals. When the user inspects the single ground instance
:- empty(1,2), empty(2,2), empty(1,2), empty(2,3).
of the constraint the bug becomes obvious. A less attentive observer would maybe
not immediately realise that the constraint will not become inactive again. In this
case, he or she would in the worst case step through all the other rules before the
constraint above remains as the last rule with active instances. Then, at the latest,
36
J. Oetsch, J. Pührer, and H. Tompits
one comes to the same conclusion that X+1 has to be replaced by Y+1. Moreover, a
stepping environment could give a warning when there is a constraint instance that
is guaranteed to stay active in subsequent states. This feature is not implemented
in SeaLion. We refer to the corrected version of program Π4 by Π5 .
Compared to traditional software, programs in ASP are typically very succinct
and often authored by a single person. Nevertheless, people are sometimes confronted with ASP code written by another person, e.g., in case of joint program
development, software quality inspection, legacy code maintenance, or evaluation
of student assignments in a logic-programming course. As answer-set programs can
model complex problems within a few lines of code, it can be pretty puzzling to
understand someone else’s ASP code, even if the program is short. Here, stepping
can be very helpful to get insight into how a program works that was written by
another programmer, as illustrated by the following example.
Example 10
Assume that the full encoding of the maze generation encoding is composed by the
programs Π2 ∪ Π5 and that the constraints in Π6 , given next, has been written by
another author.
:- exit(X,Y), wall(X,Y).
:- wall(X,Y), wall(X+1,Y), wall(X,Y+1), wall(X+1,Y+1).
:- wall(X,Y), empty(X+1;X-1,Y), empty(X,Y+1;Y-1), col(X+1;X-1),
row(Y+1;Y-1).
:- wall(X,Y), wall(X+1,Y+1), not wall(X+1,Y), not wall(X,Y+1).
:- wall(X+1,Y), wall(X,Y+1), not wall(X,Y), not wall(X+1,Y+1).
Note that the guess whether a cell is a wall or empty in the program Π2 ∪Π5 ∪Π6 is
realised by guessing for each non-border cell whether it is a wall or not and deriving
that a cell is empty in case we do not know that it is a wall. Moreover, observe that
facts of predicate empty/2 may be part of a valid encoding of a maze generation
problem instance, i.e., they are a potential input of the program. As a consequence,
it seems plausible that the encoding could guess the existence of a wall for a cell
that is already defined to be empty by a respective fact in the program input. In
particular, there is no constraint that explicitly forbids that a single cell can be
empty and contain a wall. The encoding would be incorrect if it would allow for
answer sets with cells that are empty and a wall according to the maze generation
problem specification. However, it turns out that the answer sets of the program
are exactly the intended ones. Let us find out why by means of stepping.
Reconsider the problem instance depicted in Fig. 1 that is encoded in the program
Π1 . It requires that cell (3, 4) is empty. If it did not, the maze shown in Fig. 7 that
contains a wall at cell (3, 4) would be a valid solution. We start a stepping session
for program Π1 ∪ Π2 ∪ Π5 ∪ Π6 , whose code is summarised in Fig. 5, and step
towards an interpretation encoding the maze of Fig. 7 to see what is happening
if we consider (3, 4) to be a wall despite the presence of fact empty(3,4). We can
reuse the computation C4 obtained in Example 7 whose final state S4 considers
already the facts describing the input instance and the rules needed for deriving
Stepwise Debugging of Answer-Set Programs
37
Fig. 6: In SeaLion, we can retract the computation to a previous state by clicking
on the node in the computation view representing the step or jump that created
the target state. The screenshot shows reverting the last two states in Example 10,
where we reused a part of the computation S1 , . . . , S6 for exploring an alternative
setting starting from S4 . The final state of the alternative computation presented
in the example is depicted in Fig. 9.
border/2 atoms. As in Example 8, we continue with a step for considering the
ground instance of the rule
{wall(X,Y) : col(X), row(Y), not border(X,Y)}.
that guesses whether non-border cells are walls. This time, instead of choosing
wall(3,2) to be true, we only add wall(3,4) to the atoms considered true. Then,
for the resulting state S50 , both empty(3,4) and wall(3,4) are contained in IS50 . A
visualisation of IS50 if given in the centre of Fig. 8. In order to derive the remaining
atoms of predicates empty/2 and wall/2 we then jump through the rules
wall(X,Y) :- border(X,Y), not entrance(X,Y), not exit(X,Y).
empty(X,Y) :- col(X), row(Y), not wall(X,Y).
to obtain state S60 , where IS60 is illustrated in the right subfigure of Fig. 8.
Now, the user sees that constraint
:- empty(X,Y), not reach(X,Y).
38
J. Oetsch, J. Pührer, and H. Tompits
Fig. 7: A valid maze—but not a solution for the instance depicted in Fig. 1 as that
requires (3, 4) to be an empty cell.
has active instances. This comes as no surprise as the rules defining reachability
between empty cells have not been considered yet. We decide to do so now and
initiate a jump through the rules
adjacent(X,Y,X,Y+1) :- col(X), row(Y), row(Y+1).
adjacent(X,Y,X,Y-1) :- col(X), row(Y), row(Y-1).
adjacent(X,Y,X+1,Y) :- col(X), row(Y), col(X+1).
adjacent(X,Y,X-1,Y) :- col(X), row(Y), col(X-1).
reach(X,Y)
:- entrance(X,Y), not wall(X,Y).
reach(XX,YY) :- adjacent(X,Y,XX,YY), reach(X,Y), not wall(XX,YY).
We obtain the new state S70 and observe that under interpretation IS70 the constraint
still has an active instance, namely
:- empty(3,4), not reach(3,4).
Obviously, the atom reach(3,4) has not been derived in the computation. When
inspecting the rules defining reach/2 it becomes clear why the answer sets of the
encoding are correct: the atom reach(X,Y) is only derived for cells that do not
contain walls. Consequently, whenever there is an empty cell which was guessed to
contain a wall it will be considered as not reachable from the entrance. As every
empty cell has to be reachable, a respective answer-set candidate will be eliminated
by an instance of the constraint
:- empty(X,Y), not reach(X,Y).
Although the encoding of the maze generation problem is correct one could consider
it to be not very well designed. Conceptually, the purpose of the constraint above
is forbidding empty cells to be unreachable from the entrance and not forbidding
them to be walls. Moreover, if one would replace the rules
reach(X,Y)
:- entrance(X,Y), not wall(X,Y).
reach(XX,YY) :- adjacent(X,Y,XX,YY), reach(X,Y), not wall(XX,YY).
by
reach(X,Y)
:- entrance(X,Y), empty(X,Y).
reach(XX,YY) :- adjacent(X,Y,XX,YY), reach(X,Y), empty(XX,YY).
Stepwise Debugging of Answer-Set Programs
39
Fig. 8: The stepping session described in Example 10: Starting from the maze
generation instance we step towards an interpretation encoding the wrong solution
of Fig. 7. After stepping through the guessing rule the resulting interpretation
contains atoms empty(3,4) and wall(3,4) stating that cell (3, 4) is both a wall
and empty.
which seem to be equivalent in the terms of the problem specification, the program
would not work. A more natural encoding would be to explicitly forbid empty cells
to contain walls either by an explicit constraint or a modified guess where nonborder cell is guessed to be either empty or contain a wall but not both.
5 Related Work
First, we describe existing approaches for debugging answer-set programs and how
they are related to the method proposed in this paper.
The first work devoted to debugging of answer-set programs is a paper by Brain
and De Vos (2005) in which they provide general considerations on the subject,
such as the discussion of error classes in the context of ASP or implications of
declarativity on debugging mentioned in the previous section. They also formulated important debugging questions in ASP, namely, why is a set of atoms subset
of a specific answer set and why is a set of atoms not subset of any answer set. The
authors provided pseudocode for two imperative ad-hoc algorithms for answering
these questions for propositional normal answer-set programs. The algorithm addressing the first question returns answers in terms of active rules that derive atoms
from the given set. The algorithm for explaining why a set of atoms is not subset of
any answer set identifies different sorts of answers such as atoms with no deriving
rules, inactive deriving rules, or supersets of the given set in which adding further
literals would lead to some inconsistency.
The goal of the work by Pontelli et al. (2009) is to explain the truth values of
literals with respect to a given actual answer set of a program. Explanations are
provided in terms of justifications which are labelled graphs whose nodes are truth
assignments of possibly default-negated ground atoms. The edges represent positive and negative support relations between these truth assignments such that every
path ends in an assignment which is either assumed or known to hold. The authors
have also introduced justifications for partial answer sets that emerge during the
40
J. Oetsch, J. Pührer, and H. Tompits
Fig. 9: The computation is stuck is state S70 from Example 10 as only a constraint
has active instances (highlighted in red in the editor). The “No entry” symbol
in the truth assignment view indicates an unsatisfied instance under the current
state. Note that the tree in the computation view has an inactive branch. The
current computation (nodes with yellow background) is an alternative branch for
the computation in Fig. 6. Clicking on the final greyed branch would set S1 , . . . , S6
to be the active computation again.
solving process (online justifications), being represented by three-valued interpretations. Pontelli et al. (2009) use sequences of three-valued interpretations (called
computations) in which monotonously more atoms are considered true, respectively,
false. The information carried in these interpretations corresponds to that of the
second and third component of a state in a computation in our framework. The
purpose of using computations in the two approaches differs however. While computations in stepping are used for structuring debugging process in a natural way,
where the choices how to proceed remains with the user, the computations of Pontelli et al. are abstractions of the solving procedure. Their goal is to allow a solver
that is compatible with their computation model to compute justifications for its
intermediate results. Thus, similar to their offline versions, online justifications are
non-interactive, i.e., they are computed automatically and used for post-mortem
debugging. As our computation model is compatible with that for online justifications, it seems very promising to combine the two approaches in practice. While
debugging information in stepping focuses on violation of rules and unfounded sets,
Stepwise Debugging of Answer-Set Programs
41
our method leaves the reasons for an atom being true or false as implicit consequences of a user’s decision. Here, online justifications could keep track of the
reasons for truth values at each state of a stepping session and presented to the
user during debugging on demand.
Syrjänen (2006) aimed at finding explanations why a program has no answer
sets. His approach is based on finding minimal sets of constraints such that their
removal yields consistency. Hereby, it is assumed that a program does not involve
circular dependencies between literals through an odd number of negations which
might also cause inconsistency. The author considers only a basic ASP language
and hence does not take further sources of inconsistency into account, caused by
program constructs of richer ASP languages, such as cardinality constraints.
Another early approach (Brain et al. 2007; Pührer 2007) is based on program
rewritings using some additional control atoms, called tags, that allow, e.g., for
switching individual rules on or off and for analysing the resulting answer sets.
Debugging requests can be posed by adding further rules that can employ tags as
well. That is, ASP is used itself for debugging answer-set programs. The translations needed were implemented in the command-line tool Spock (Brain et al. 2007;
Gebser et al. 2009) which also incorporates the translations of another approach in
which also ASP is used for debugging purposes (Gebser et al. 2008; Pührer 2007).
The technique is based on ASP meta-programming, i.e., a program over a metalanguage is used to manipulate a program over an object language (in this case, both
the meta-language and the object language are instances of ASP). It addresses the
question why some interpretation is not an answer set of the given program. Answers
are given in terms of unsatisfied rules and unfounded loops. The approach has later
been extended from propositional to disjunctive logic programs with constraints,
integer arithmetic, comparison predicates, and strong negation (Oetsch et al. 2010)
and also to programs with cardinality constraints (Polleres et al. 2013). It has been
implemented in the Ouroboros plugin of SeaLion (Frühstück et al. 2013). Moreover, Shchekotykhin (2015) developed a method on top of the meta-programming
approaches (Gebser et al. 2008; Oetsch et al. 2010) that poses questions to the user
in order to find a desired problem diagnosis while keeping the amount of required
interaction low.
In a related approach, Dodaro et al. (2015) use control atoms quite similar to
that of the tagging approach (Brain et al. 2007; Pührer 2007) to identify sets of
rules that lead to inconsistency of a program under the requirement that a given set
of atoms is true in some intended answer set. An implementation is provided that
profits from a tight integration with the ASP solver WASP (Alviano et al. 2015).
In order to reduce the possible outcomes, the debugger asks the user about the
intended truth of further atoms in an interactive session.
In another paper, Li et al. (2015) use inductive logic programming to repair
answer-set programs. The idea is that the user provides examples of desired and
undesired properties of answer sets such that the debugger can semi-automatically
revise a faulty program. The method requires a difference metric between logic programs in order to restrict repairs to programs that have the desired properties that
minimally differ from the original program. The authors propose such a measure
42
J. Oetsch, J. Pührer, and H. Tompits
in terms of number of operations required for revision. These operations are rule
removal and creation as well as addition or removal of individual body literals.
Caballero et al. (2008) developed a declarative debugging approach for datalog
using a classification of error explanations similar to that of the aforementioned
meta-programming technique (Gebser et al. 2008; Oetsch et al. 2010). Their approach is tailored towards query answering and the language is restricted to stratified datalog. However, the authors provide an implementation that is based on
computing a graph that reflects the execution of a query.
Wittocx et al. (2009) show how a calculus can be used for debugging first-order
theories with inductive definitions (Denecker 2000; Denecker and Ternovska 2008)
in the context of model expansion problems, i.e., problems of finding models of a
given theory that expand some given interpretation. The idea is to trace the proof
of inconsistency of such an unsatisfiable model expansion problem. The authors
provide a system that allows for interactively exploring the proof tree.
Besides the mentioned approaches which rely on the semantical behaviour of programs, Mikitiuk et al. (2007) use a translation from logic-program rules to natural
language in order to detect program errors more easily. This seems to be a potentially useful feature for an IDE as well, especially for non-expert ASP programmers.
We can categorise the different methods by their setting of their debugging tasks.
Here, one aspect is whether a technique works with factual or desired answer sets.
Approaches that focus on actual answer sets of the program to be debugged include
the algorithm by Brain and De Vos (2005) that aims at explaining the presence of
atoms in an answer set. Also, justifications (Pontelli et al. 2009) are targeted towards explanations in a given actual answer set, with the difference that they focus
on a single atom but can not only explain their presence but also their absence.
The approach by Caballero et al. (2008) can also be seen to target a single actual
answer set. Due to their focus on actual answer sets of the debugged program,
these methods cannot be applied on (erroneous) programs without any answer set.
The previous meta-programming based debugging technique (Gebser et al. 2008;
Pührer 2007) and follow-up works (Oetsch et al. 2010; Polleres et al. 2013) deal
with a single intended but non-actual answer set of the debugged program. In the
approach of Wittocx et al. (2009), the user can specify a class of intended semantic
structures which are not preferred models of the theory at hand (corresponding to
actual answer sets of the program to be debugged in ASP terminology). Syrjänen’s
diagnosis technique (Syrjänen 2006) is limited to the setting when a program has
no answer set at all. The same holds for the work of Dodaro et al. (2015), however
the authors demonstrate how other debugging problems can be reduced to that of
inconsistency. The method requires an intended answer set but offers the means to
generate that in an interactive way, building on the technique by Shchekotykhin
(2015). Stepping does not require actual or intended answer sets as a prerequisite,
as the user can explore the behaviour of his or her program under different interpretations that may or may not be extended to answer sets by choosing different rules
instances. In the interactive setting summarised in Fig. 2, where one can retract a
computation to a previous state and continue stepping from there that is also implemented in SeaLion, a stepping session can thus be seen as an inspection across
Stepwise Debugging of Answer-Set Programs
43
arbitrary interpretations rather than an inquiry about a concrete set of actual or
non-existent answer sets. Nevertheless, if one has a concrete interpretation in mind,
the user is free to focus on that. The ability to explore rule applications for partial
interpretations that cannot become answer sets amounts to a form of hypothetical
reasoning. A related form of this type of debugging is also available in one feature
of the tagging approach (Brain et al. 2007) that aims at extrapolating non-existent
answer sets by switching off rules and guessing further atoms. Here, the stepping
technique can be considered more focused, as the interpretation under investigation
is determined by the choices of the user in stepping but is essentially arbitrary in
the tagging approach if the user does not employ explicit restrictions.
Next, we compare the ASP languages supported by different approaches. First,
the language of theories with inductive definitions used in one of the debugging approaches (Wittocx et al. 2009) differs from the remaining approaches that are based
on logic-programming rules. Many of these works deal only with the basic ASP setting of debugging ground answer-set programs, supporting only normal rules (Brain
and De Vos 2005; Pontelli et al. 2009), disjunctive rules (Gebser et al. 2008), or simple choice rules (Syrjänen 2006). The work on tagging-based debugging (Brain et al.
2007) sketches how to apply the approach to programs with variables by means of
function symbols. The approach by Caballero et al. (2008) deals with non-ground
normal programs which have to be stratified. Explicit support for variables is also
given in an extension (Oetsch et al. 2010) of the meta-programming approach for
disjunctive programs. It was later extended to allow for weight constraints (Polleres
et al. 2013) by compiling them away to normal rules. A commonality of the previous approaches is that they target ASP languages that can be considered idealised
proper subsets of current solver languages. In this respect, stepping is the first debugging approach that overcomes these limitations as the use of C-programs and
abstract grounding (cf. Pührer 2014) make the framework generic enough to be
applied to ASP solver languages. While this does not mean that other approaches
cannot be adapted to fit a solver language, it is no always immediately clear how.
For our approach, instantiating our abstractions to the language constructs and the
grounding method of a solver is sufficient to have a ready-to-use debugging method.
Most existing debugging approaches for ASP can be seen as declarative in the
sense that a user can pose a debugging query, and receives answers in terms of
different declarative definitions of the semantics of answer-set programs, e.g., in
terms of active or inactive rules with respect to some interpretation. In particular,
the approaches do not take the execution strategy of solvers into account. This also
holds for our approach, however stepping and online justifications (Pontelli et al.
2009) are exceptional as both involve a generic notion of computation which adds
a procedural flavour to debugging. Nonetheless, the computation model we use can
be seen as a declarative characterisation of the answer-set semantics itself as it does
not apply a fix order in which to apply rules to build up an answer set.
Besides stepping, also the approaches by Wittocx et al. (2009) and Dodaro et al.
(2015) as well as Shchekotykhin (2015) can be considered interactive. While in
the approach of Wittocx et al. a fixed proof is explored interactively, the interaction in our method has influence on the direction of the computation. The other
44
J. Oetsch, J. Pührer, and H. Tompits
works (Dodaro et al. 2015; Shchekotykhin 2015) use interaction for filtering the
resulting debugging information. Also in further works (Brain et al. 2007; Gebser
et al. 2008) which do not explicitely cover interleaved communication between user
and system, user information can be used for filtering. The approaches mentioned in
this paragraph realise declarative debugging in the sense of Shapiro (1982), where
the user serves as an oracle for guiding the search for errors.
It is worth highlighting that stepping can be seen as orthogonal to the basic ideas
of all the other approaches we discussed. That is, it is reasonable to have a development kit that supports stepping and other debugging methods simultaneously.
While debugging is the main focus of this paper, we also consider the computation framework for disjunctive abstract constraint programs introduced in Section 3
an interesting theoretical contribution by itself. Here, an important related work
is that of Liu et al. (2010), who also use a notion of computation to characterise
a semantics for normal C-programs. These computations are sequences of evolving
interpretations. Unlike the three-valued ones used for online justifications (Pontelli et al. 2009), these carry only information about atoms considered true. Thus,
conceptionally, they correspond to sequences IS0 , IS1 , . . . where S0 , S1 , . . . is a computation in our sense. The authors formulate principles for characterising different
variants of computations. We will highlight differences and commonalities between
the approaches along the lines of some of these properties. One main structural
difference between their and our notion of computation is the granularity of steps:
In the approach by Liu et al. it might be the case that multiple rules must be
considered at once, as required by their revision property (R’), while in our case
computation proceeds rule instance by rule instance. The purpose of property (R’)
is to assure that every successive interpretation must be supported by the rules
active with respect to the previous interpretation. But it also requires that every
active rule in the overall program is satisfied after each step, whereas we allow rule
instances that were not considered yet in the computation to be unsatisfied. For
the purpose of debugging, rule-based computation granularity seems favourable as
rules are our primary source code artifacts. Moreover, ignoring parts of the program
that were not considered yet in a computation is essential in the stepping method,
as this breaks down the amount of information that has to be considered by the
user at once and allows for getting stuck and thereby detect discrepancies between
his or her understanding of the program and its actual semantics. Our computations (when translated as above) meet the persistence principle (P’) of Liu et al.
that ensures that a successor’s interpretation is a superset of the current one. Their
convergence principle (C’), requiring that a computation stabilises to a supported
model, is not met by our computations, as we do not enforce support in general.
However, when a computation has succeeded (cf. Definition 19), it meets this property. A further difference is that Liu et al. do not allow for non-stable computations
as required by the founded persistence of reasons principle (FPr). This explains
why the semantics they characterise treats non-convex atoms not in the FLP-way.
Besides that, the use of non-stable computations allow us to handle disjunction.
Interestingly, Liu et al. mention the support for disjunction in computations as
an open challenging problem and suspect the necessity of a global minimality re-
Stepwise Debugging of Answer-Set Programs
45
quirement on computations for this purpose. Our framework demonstrates that we
can do without such a condition: As shown in Theorem 1, unfounded sets in our
semantics can be computed incrementally “on-the-fly” by considering only the rule
instance added in a step as potential new external support. Finally, the principle
of persistence of reasons (Pr’) suggests that the “reason” for the truth value of
an atom must not change in the course of a computation. Liu et al. identify such
reasons by sets of rules that keep providing support in an ongoing computation.
We have a similar principle of persistence of reasons that is however stricter as it
is operates on the atom level rather than the rule level: Once a rule instance is
considered in a computation in our sense, the truth value of the atoms in the rule’s
domain is frozen, i.e., it cannot be changed or forgotten in subsequent steps. Persistence of reasons is also reflected in our definition of answer sets: The requirement
I 0 |DA = I |DA in Definition 12 that the stability of interpretation I is only spoiled
by I 0 if the reason for I 0 |= A is the same satisfier of C-atom A as for I |= A.
As argued above and in the introduction, our notion of computation is better
suited for stepping than that of Liu et al., yet we see potential for using the latter orthogonal to our method for debugging (for the class of programs for which
the different semantics coincide). While our jumping technique allows to consider
several rules, selected by the user, at once, a debugging system could provide proposals for jumping, where each proposal corresponds to a set of rules that result
in a next step in the sense of Liu et al. Then, the system could present the atom
assignments for each proposal such that the user has an alternative to choose a
jump based on truth of atoms rather than rules. Moreover, this can be the basis
for automated progression in a stepping session until a certain atom is assigned,
analogous to watchpoints in imperative debugging. We believe that these ideas for
(semi-)automated jumping deserve further investigation.
Another branch of research, that is related to our notion of computation, focuses
on transition systems for analysing answer-set computation (Lierler 2011; Lierler
and Truszczyński 2016; Brochenin et al. 2014). These works build on the ideas of
a transition system for the DPLL procedure for SAT solving (Nieuwenhuis et al.
2006). Transition systems are graphs whose nodes represent the state of a computation whereas the edges represent possible state transitions during solving. Typically,
a set of transition rules, depending on the answer-set program, determines the possible transitions. In ASP transition systems, nodes are represented by (ordered) sets
of literals with annotations whether a literal is a decision literal. Moreover, there
is a distinguished state F ailState for representing when a branch of computation
cannot succeed in producing an answer set. Different sets of transition rules have
been proposed that correspond to different models of computations. Typical transition rules include a unit propagation rule that derives one further literal based on
a rule of the program for which all other literals are defined in the previous state,
a decision rule that adds one further literal (annotated as decision literal), and a
transition rule for backtracking that retracts a decision literal from an inconsistent
state and replace it by its negation. Existing transition systems for ASP are intended to reflect ASP solving algorithms, including failed branches of search space
traversal. For instance, transition systems have been defined with transition rules
46
J. Oetsch, J. Pührer, and H. Tompits
for backjumping and learning as used in modern solvers (Lierler 2011). In contrast,
our framework generates ideal (possibly failed) computations without backtracking.
Another main difference is that all proposed transition systems have a transition
rule for arbitrary assignment of decision literals whereas in our framework truth
assignments are restricted to the domain of the C-rule added in the current step.
Regarding supported language constructs, to the best of our knowledge, existing
transition systems for ASP focus on elementary atoms, i.e., they do not cover aggregates. However, Lierler and Truszczyński (2016) also proposed transition systems
for multi-logic systems including ASP. There has been work on transition systems
for disjunctive programs (Brochenin et al. 2014). These are based on integrating
two sets of transition rules, one for guessing and one for checking of answer set candidates. Similarly, as in the work by Liu et al. (2010), states in transition systems
do not keep track of ASP rules as our states do. Note that our computation framework can be turned into to a transition system for disjunctive C-programs with
only two transition rules, one for propagation that is derived from the successor
relation (cf. Definition 17) and another for defining the transition of unstable final
states or states with remaining active rules but no successor to F ailState.
6 Conclusion
In this paper, we introduced the stepping technique for ASP that can be used
for debugging and analysis of answer-set programs. Like stepping in imperative
programming, where the effects of consecutive statements are watched by the user,
our stepping technique allows for monitoring the effect of rules added in a step-bystep session. In contrast to the imperative setting, stepping in our sense is interactive
as a user decides in which direction to branch, by choosing which rule to consider
next and which truth values its atoms should be assigned. On the one hand, this
breaks a general problem of debugging in ASP, namely how to find the cause for
an error, into small pieces. On the other hand, user interaction allows for focusing
on interesting parts of the debugging search space from the beginning. This is
in contrast to the imperative setting, where the order in which statements are
considered in a debugging session is fixed. Nevertheless, also in our setting, the
choice of the next rule is not entirely arbitrary, as we require the rule body to
be active first. Debuggers for procedural languages often tackle the problem that
many statements need to be considered before coming to an interesting step by
ignoring several steps until pre-defined breakpoints are reached. We developed an
analogous technique in our approach that we refer to as jumping which allows to
consider multiple rules at once. Besides developing the technical framework for
stepping, we also discussed the implementation of stepping in the SeaLion system
and methodological aspects, thereby giving guidelines for the usage of the technique,
and for setting the latter in the big picture of ASP development.
While unstable computations are often not needed, they offer great opportunities
for further work. For one, the use of unfounded sets for distinguishing states in
unstable computations is a natural first choice for expressing the lack of stability.
Arguably, when a user arrives in a state with a non-empty unfounded set, he or she
Stepwise Debugging of Answer-Set Programs
47
only knows that some external support has to be found for this set but there is no
information which atoms of the unfounded sets are the crucial ones. It might be
worthwhile to explore alternative representations for unstability such as elementary
loops (Gebser et al. 2011) that possibly provide more pinpoint information. This
would require lifting a respective notion to the full language of C-programs first.
Another issue regarding unstable computations that would deserve further attention is that in the current approach jumps can only result in stable states. Thus,
unstable states in a computation can only be reached by individual steps at present.
Here, it would be interesting to study methods and properties for computations that
allow for jumping to states that are not stable.
We next discuss functionality that could be helpful for stepping which are not yet
implemented in SeaLion. One such feature is semi-automatic stepping, i.e., the user
can push a button and then the system searches for potential steps for which no
further user interaction is required and applies them automatically until an answer
set is reached, the computation is stuck, or user interaction is required. It would
also be convenient to automatically check whether the computation of a debugging
session is still a computation for the debugged program after a program update.
In this respect, when the computation for the old version became incompatible, a
feature would be advantageous that builds up a computation for the new version
that resembles the old one as much as possible. Unlike semi-automatic stepping
and compatibility checks for computations which could be implemented without
further studies, the latter point still requires theoretical research. Further convenient
features would be functionality that highlights the truth values of atoms that cause
a rule not to be active for a given substitution and methods for predicting whether a
rule can become active in the future, i.e., in some continuation of the computation.
Acknowledgements
We thank the reviewers for their useful comments. This work was partially supported by the Austrian Science Fund (FWF) under project P21698, the German
Research Foundation (DFG) under grants BR-1817/7-1 and BR 1817/7-2, and the
European Commission under project IST-2009-231875 (OntoRule).
Appendix A Guidelines for Stepping
In what follows, we give advice on how users can exploit stepping for analysing and
debugging their code. Fig. A 1 synthesises practical guidelines for stepping from the
methodological aspects of stepping described in Section 4.6. It can be seen as a useroriented view on the stepping technique. Depending on the goals and the knowledge
of the user, this guide gives concise yet high-level suggestions on how to proceed
in a stepping session. The upper area of the figure is concerned with clarifying the
best strategy for a stepping session and for choosing the computation to start from.
The lower area, on the other hand, guides the user through the stepping process.
The diagram differentiates between four tasks a user may want to perform.
48
J. Oetsch, J. Pührer, and H. Tompits
(i) Debugging a program lacking a particular answer set: we suggest to step and
jump through rules that one thinks build up this answer set. Eventually, the
computation will get stuck when adding a rule that prevents the answer set.
(ii) Debugging a program that lacks any answer set: if an intended answer set is
known, we advise using the strategy of Item (i). Otherwise, the user should
choose rules and truth values during stepping that he or she thinks should be
consistent, i.e., lead to a successful computation. Also here, the computation
is guaranteed to fail and get stuck, indicating a reason for the inconsistency.
(iii) Debugging a program with an unintended answer set I : In case that I is
similar to an intended but missing answer set I 0 , thus if I is intuitively a wrong
version of I 0 , we recommend stepping towards I 0 , following the strategy of
Item (i). Otherwise, the user can step towards I . Unlike in the previous cases,
the computation is guaranteed to eventually succeed. Here, stepping acts as
a disciplined way to inspect how the atoms of I can be derived and why no
rule prevents I from being an answer set. If I is intended to be a model but
not stable, then the stepping process will reveal which rules provide external
support for sets of atoms that are supposed to be unfounded.
(iv) Analysing a program: In case that the user is interested in the behaviour of
the program under a particular interpretation, it is reasonable to step towards
this interpretation. Otherwise, rules and truth assignments should be chosen
that drive the computation towards states that the user is interested in.
The procedures suggested above and in Fig. A 1 are meant as rough guidelines for
the inexperienced user. Presumably, knowledge about the own source code and some
practice in stepping gives the user a good intuition on how to find bugs efficiently.
It is natural to ask how big a program can get such that it is still suitable for
stepping. Due to the vague nature of the question, answers cannot be clearly established. From a complexity theoretic point of view, the problems that need to be
solved in a stepping support environment for and after performing a step or a jump,
e.g., computing a new state from a jump, determining rules with active instances,
or checking whether a computation has failed, are not harder than computing an
answer set of the program under development. Under this observation, our technique is certainly an appropriate approach for debugging ASP. In some applications,
however, solving times of multiple minutes or even hours are acceptable. Certainly,
having waiting times of these lengths for individual debugging steps is undesirable.
On the positive side, often, following a few guidelines during the development of
an answer-set program can significantly reduce the likelihood of introducing bugs,
the amount of information the user has to deal with, and also the computational
resources required for stepping. Among these measures are best practices for ASP
development that have been discussed in a paper by Brain et al. (2009). For working
with the stepping method in particular, we give the following recommendations.
Use scalable encodings and start with small examples. Using small problem instances, also the resulting grounding as well as answer sets are typically small.
This limits the amount of information to be considered during debugging. Chances
49
Stepwise Debugging of Answer-Set Programs
STEPPING
GUIDE
What is your goal?
an unintended answer set
find
a bug
What type of bug?
analyse the program
no answer sets exist although some should
Do you want to inspect the
behaviour of your program
under a particular
interpretation?
Do you know any particular
expected answer set?
yes
no
Always choose truth values
to match the interpretation
when performing a step in
this session
no
Is the unintended answer set
a wrong version of a missing
expected answer set?
a missing answer set
no
yes
yes
Always choose truth values
to match the intended
answer set when performing
a step in this session
Follow your intuitions to
create an interesting
situation when choosing
truth values in this session
Always choose truth values
to match the unintended
answer set when performing
a step in this session
Which is the computation
you can obtain that reflects
your intended setting the
most?
a (part of a) stored computation for a trusted part of the program,
a computation generated from an answer set of a trusted part of the program,
a computation generated from an answer set of a previous version of the program, or
a computation generated from an interpretation from an external source
Start stepping from the
empty state. Then, jump
through the program’s facts
Start stepping from the
obtained computation
Store computation for later
use. Retract the final states
from the computation until
the new final state matches
your intentions
Select rules for step or jump
yes
Step
Jump
Can you already gain
satisfactory insight into the
program’s semantics?
yes
Store computation for later
use
DONE
Fig. A 1: Stepping guide
no
Do you want to continue
stepping from the current
state of the computation?
no
50
J. Oetsch, J. Pührer, and H. Tompits
that bugs are detected early, using small programs is suggested by an evaluation of
the small-scope hypothesis for ASP (Oetsch et al. 2012).
Visualise answer sets and stepping states. Tools like Kara (Kloimüllner et al. 2013)
(that is implemented in SeaLion), ASPVIZ (Cliffe et al. 2008), IDPDraw (Wittocx
2009), or Lonsdaleite (Smith 2011) allow for visualising interpretations. With
their help, one can quickly spot when an answer set differs from what is expected
and they allow to monitor the evolvement of the interpretation that is build up
during stepping. The illustrations of the maze generation problem in this section
were created using Kara. For use with stepping, we advise to specify visualisations
also for interpretations that are not supposed to be answer sets. For example, in
Fig. 8, we have visualisations for cells that are not assigned to be empty or a wall
and for cells that are assigned to be a wall and empty, despite in an expected answer
set, every cell has to be either a wall or empty.
Test often. Frequent tests allow the user to trust in large parts of the program,
hence these parts can be jumped over in a stepping session.
Appendix B Remaining Proofs
Theorem 1
Let S be a state and S 0 a successor of S , where ∆ = IS 0 \ IS . Moreover, let X 0 be
a set of literals with ∅ ⊂ X 0 ⊆ IS 0 . Then, the following statements are equivalent:
(i) X 0 is unfounded in PS 0 with respect to IS 0 .
(ii) X 0 = ∆0 ∪ X , where ∆0 ⊆ ∆, X ∈ ΥS , and rnew (S , S 0 ) is not an external
support for X 0 with respect to IS 0 .
Proof
((i)⇒(ii)) It is obvious that rnew (S , S 0 ) is not an external support for X 0 with
respect to IS 0 as otherwise X 0 cannot be unfounded in PS 0 with respect to IS 0 .
It remains to be shown that X 0 = ∆0 ∪ X for some ∆0 ⊆ ∆ and some X ∈ ΥS .
Towards a contradiction, assume X 0 6= ∆00 ∪ X 00 for all X 00 ∈ ΥS and all ∆00 ⊆ ∆.
We define X = X 0 ∩ IS .
Consider the case that X ∈ ΥS . As X 0 \ IS ⊆ ∆, and X 0 = (X 0 \ IS ) ∪ X , we
have a contradiction to our assumption. Therefore, it holds that X 6∈ ΥS . Hence,
as X ⊆ IS , by definition of a state, X is not unfounded in PS with respect to IS .
Therefore, there is some external support r ∈ PS for X with respect to IS .
In the following, we show that r is also an external support for X 0 with respect
to IS 0 . Since S 0 is a successor of S and S is a state, we get that IS and IS 0 coincide
on Dr . Consequently, from IS |= B(r ) we get that also IS 0 |= B(r ). Moreover,
because of IS \ X |= B(r ) it is also true that IS 0 \ X 0 |= B(r ). Furthermore, we know
that there is some A ∈ H(r ) with X |DA 6= ∅ and IS |DA ⊆ C , for some C ∈ CA . As
X |DA = X 0 |DA and IS |DA = IS 0 |DA we also have X 0 |DA 6= ∅ and IS 0 |DA ⊆ C . Finally,
note that for all A ∈ H(r ) with IS |= A, we have (X ∩ IS )|DA 6= ∅. Consider some
Stepwise Debugging of Answer-Set Programs
51
A ∈ H(r ) such that IS 0 |= A. From the latter we get that IS |= A and therefore
(X ∩ IS )|DA 6= ∅. As X ∩ IS ⊆ X 0 ∩ IS 0 , we also have (X 0 ∩ IS 0 )|DA 6= ∅. Hence, r
fulfils all conditions for being an external support for X 0 with respect to IS 0 , which
is a contradiction to X 0 being unfounded in PS 0 with respect to IS 0 .
((ii)⇒(i)) Towards a contradiction, assume X 0 has some external support r ∈ PS 0
with respect to IS 0 . From (ii) we know that r 6= rnew (S , S 0 ) and X 0 = ∆0 ∪ X for
some ∆0 ⊆ ∆ and some X ∈ ΥS . As r 6= rnew (S , S 0 ), we have that IS and IS 0
coincide on Dr . Therefore, from IS 0 |= B(r ) and IS 0 \ X 0 |= B(r ), it follows that
IS |= B(r ) and IS \ X 0 |= B(r ). Note that X = X 0 ∩ IS and hence IS \ X |= B(r ).
We know that there is some A ∈ H(r ) with X 0 |DA 6= ∅ and IS 0 |DA ⊆ C , for some
C ∈ CA . As X 0 |DA = X |DA we have X |DA 6= ∅. Moreover, as IS 0 |DA = IS |DA , it
holds that IS |DA ⊆ C . Finally, notice that for all A ∈ H(r ) with IS 0 |= A, we
have (X 0 ∩ IS 0 )|DA 6= ∅. Consider some A ∈ H(r ) with IS |= A. As IS 0 |DA = IS |DA ,
we also have IS 0 |= A and hence (X 0 ∩ IS 0 )|DA 6= ∅. As DA ∩ ∆ = ∅, we have
(X 0 ∩ IS 0 )|DA = (X ∩ IS )|DA . Consequently, it holds that (X ∩ IS )|DA 6= ∅. We
showed that r is an external support of X in PS with respect to IS . Therefore, we
have a contradiction to X ∈ ΥS because S is a state.
Theorem 3
Let S0 be a state, P a C-program with PS0 ⊆ P , and I an answer set of P with
IS0 ⊆ I and I ∩ I − S0 = ∅. Then, there is a computation S0 , . . . , Sn that has
succeeded for P such that PSn = P I and ISn = I .
Proof
The proof is by induction on the size of the set P I \ PS0 . Observe that from IS0 ⊆ I ,
I ∩ I − S0 = ∅, and IS0 |= B(r ) and Dr ⊆ IS0 ∪ I − S0 , for all r ∈ PS0 , we get that
I |= B(r ) for all r ∈ PS0 . Hence, as PS0 ⊆ P , we have PS0 ⊆ P I .
Consider the base case that |P I \ PS0 | = 0. From PS0 ⊆ P I we get PS0 = P I .
Consider the sequence C = hPS0 , IS0 , I − S0 , ΥS0 i. Towards a contradiction, assume
IS0 6= I . As IS0 ⊆ I this means IS0 ⊂ I . Hence, there is some a ∈ I \ IS0 . As for
all r ∈ PS0 it holds that Dr ⊆ IS0 ∪ I − S0 , and I ∩ I − S0 = ∅, we get a 6∈ DPS0 . We
have a contradiction to I ∈ AS(PS0 ) by Corollary 1, as {a} is unfounded in PS0
with respect to I . Consequently, IS0 = I must hold. As IS0 is an answer set of PS0
and S0 is a state, we have that ΥS0 = {∅} by definition of state. It follows that C
meets the criteria of the conjectured computation.
We proceed with the step case. As induction hypothesis, assume that the claim
holds whenever |P I \ PS0 | ≤ i for an arbitrary but fixed i ≥ 0. Consider some
state S0 and some I ∈ AS(PS0 ) for which the conditions in the premise hold
such that |P I \ PS0 | = i + 1. Towards a contradiction, assume there is no C-rule
r ∈ P I \PS0 such that IS0 |= B(r ). Note that there is at least one C-rule r 0 ∈ P I \PS0
because |P I \ PS0 | = i + 1. It cannot hold that I = IS0 since from r 0 ∈ P IS0
follows IS0 |= B(r 0 ). Consequently, we have IS0 ⊂ I . Consider some r 00 ∈ P I with
IS0 |= B(r 00 ). By our assumption, we get that r 00 ∈ PS0 . It follows that IS0 |= r 00 , and
consequently there is some C-atom A ∈ H(r 00 ) with IS0 |= A. As Dr 00 ⊆ DS0 , we have
52
J. Oetsch, J. Pührer, and H. Tompits
DA ⊆ IS0 ∪ I − S0 . From that, since IS0 ⊂ I and I ∩ I − S0 = ∅, we get I |DA = IS0 |DA .
We have a contradiction to I being an answer set of P by Definition 12.
So, there must be some C-rule r ∈ P I \ PS0 such that IS0 |= B(r ). From r ∈ P I
we get I |= B(r ) and I |= r . Consider the state structure S1 = hP1 , I1 , I1 − , Υ1 i,
where P1 = PS0 ∪ {r }, I1 = IS0 ∪ (I ∩ Dr ), I1 − = I − S0 ∪ (Dr \ I ), and
Υ1 = {X |X = ∆0 ∪ X 0 , where ∆0 ⊆ (I1 \ IS0 ), X 0 ∈ ΥS0 , and
r is not an external support of X with respect to I1 }·
S1 is a successor of state S0 , therefore S1 is also a state by Corollary 2. As P1 ⊆ P ,
I1 ⊆ I , I ∩ I1 − = ∅, and |P I \ P1 | = i , by the induction hypothesis, S1 , . . . , Sn is
a computation, where Sn is a stable state, PSn = P I , and ISn = I . Since S1 is a
successor of state S0 , also S0 , S1 , . . . , Sn is a computation.
For establishing Theorem 5 we make use of the following notion which reflects
positive dependency on the rule level.
Definition 22
The positive rule dependency graph of P is given by
GR (P ) = hP , {hr1 , r2 i | r1 , r2 ∈ P , posOcc(B(r1 )) ∩ posOcc(H(r2 )) 6= ∅}i·
We can relate the two notions of dependency graph as follows.
Lemma 1
Let P be a C-program. GR (P ) is acyclic iff G(P ) is acyclic.
Proof
Let ≺D denote the edge relation of G(P ) and ≺R that of GR (P ).
(⇒) Assume G(P ) is not acyclic. There must be some path a1 , . . . , an of atoms
ai such that for 1 ≤ i < n, we have ai ∈ DP , ai ≺D ai+1 , and a1 = an . Hence,
by the definition of G(P ), there must be a sequence r1 , . . . , rn−1 such that for each
1 ≤ i ≤ n − 1, ri ∈ P , ai ∈ posOcc(H(ri )), and ai+1 ∈ posOcc(B(ri )). Therefore,
for each 1 ≤ i < n − 1, we have ri+1 ≺R ri . Note that a1 ∈ posOcc(H(r1 )) and
a1 ∈ posOcc(B(rn−1 )). Consequently, we have rn−1 ≺R r1 and thus r1 , rn−1 , . . . , r1
forms a cycle in GR (P ). It follows that GR (P ) is not acyclic.
(⇐) Assume now that GR (P ) is not acyclic. There must be some path r1 , . . . , rn
of C-rules ri such that for 1 ≤ i < n we have ri ∈ P , r1 = rn , and ri ≺R ri+1 .
Hence, by the definition of GR (P ), there must be a sequence a1 , . . . , an−1 such that
for each 1 ≤ i ≤ n − 1, ai ∈ posOcc(H(ri+1 )), and ai ∈ posOcc(B(ri )). Therefore,
for each 1 ≤ i < n − 1 we have ai+1 ≺D ai . Note that an−1 ∈ posOcc(H(r1 )) and
a1 ∈ posOcc(B(r1 )). Consequently, we have an−1 ≺D a1 and thus a1 , an−1 , . . . , a1
forms a cycle in G(P ). We have that GR (P ) is not acyclic.
Lemma 2
Let P be an absolutely tight C-program. There is a strict total order ≺ on P that
extends the reachability relation of GR (P ).
Stepwise Debugging of Answer-Set Programs
53
Proof
By Definition 21, G(P ) is acyclic. Hence, by Lemma 1, GR (P ) is also acyclic. The
conjecture holds, since every directed acyclic tree has a topological ordering.
We now have the means to show Theorem 5, guaranteeing the existence of stable
computations.
Theorem 5
Let C = S0 , . . . , Sn be a computation such that S0 and Sn are stable and P∆ =
PSn \ PS0 is a normal, convex, and absolutely tight C-program. Then, there is a
stable computation C0 = S00 , . . . , Sn0 such that S0 = S00 and Sn = Sn0 .
Proof
Let ≺ be the strict total order extending the reachability relation of GR (P∆ ) that is
guaranteed to exist by Lemma 2. Let r (·) : {1, . . . , n} 7→ P∆ denote the one-to-one
mapping from the integer interval {1, . . . , n} to the C-rules from P∆ such that for
all i , j in the range of r (·), we have that i < j implies r (j ) ≺ r (i ). Consider the
sequence C0 = S00 , . . . , Sn0 , where S00 = S0 , and for all 0 ≤ i < n,
0
Pi+1
= Pi0 ∪ {r (i + 1)},
0
ISi+1
= ISi0 ∪ (ISn ∩ Dr (i+1) ),
0
I − Si+1
= I − Si0 ∪ (I − Sn ∩ Dr (i+1) ),
and
0
ΥSi+1
= {∅}·
0
|DP
Notice that Sn0 = Sn and ISi+1
S0
i
= ISi0 |DP 0 , for all 0 ≤ i < n. We show that C0
S
i
is a computation by induction on the length of a subsequence of C0 .
As base case consider the sequence C00 = S00 . As S00 = S0 and S0 is a state, C00 is
a computation. For the induction hypothesis, assume that for some arbitrary but
fixed i with 0 ≤ i < n, the sequence S00 , . . . , Si0 is a computation.
0
In the induction step it remains to be shown that Si+1
is a successor of Si0 . Clearly,
0
0
Si+1 is a state structure, and by definition of C , since C is a computation and
0
ISi+1
|DPS
i+1
= ISn |DPS
i+1
,
Conditions (i), (ii), (iii), and (v) of Definition 17 for being a successor of Si0 are
0
0
fulfilled by Si+1
. Let ∆ denote ISi+1
\ ISi0 .
Next we show that Condition (iv) holds, i.e., ISi0 |= B(r (i + 1)). Note that since
0
Condition (v) holds, we have ISi+1
|= B(r (i + 1)) and hence (iv) holds in the case
∆ = ∅. Towards a contradiction assume ∆ 6= ∅ and ISi0 6|= B(r (i + 1)). We define
∆B + = ∆ ∩ posOcc(B(r (i + 1))).
First, consider the case that ∆B + = ∅. As ISi0 6|= B(r (i + 1)), there must be some
0
C-literal L ∈ B(r (i + 1)) such that ISi0 6|= L. We know that ISi+1
|= L. Consequently,
0
0
ISi0 |DL ⊂ ISi+1
|DL and therefore ∆|DL 6= ∅. Moreover, from ISi+1
|= L we have
0
ISi+1
|DL ⊆ posOcc(B(r (i + 1)))·
It follows that ∆|DL ∩posOcc(B(r (i + 1))) 6= ∅, indicating a contradiction to ∆B + =
54
J. Oetsch, J. Pührer, and H. Tompits
∅. It holds that ∆B + 6= ∅. Note that X ⊆ ISn . From that, since Sn is a state, there
must be some C-rule r∆B + ∈ PSn such that r∆B + is an external support for ∆B +
with respect to ISn . It cannot be the case that r ∈ PS0 , since ∆B + ∩ ISi0 = ∅,
therefore, r∆B + ∈ P∆ . As r∆B + is an external support for ∆B + with respect to ISn ,
for {A} = H(r∆B + ), we have ISn |= A and ∆B + |DA 6= ∅.
Consider the case that r∆B + = r (i + 1). From that we get posOcc(H(r (i + 1))) ∩
∆B + 6= ∅. This, in turn, implies posOcc(H(r (i + 1))) ∩ posOcc(B(r (i + 1))) 6= ∅
which is a contradiction to GR (P∆ ) being acyclic. The latter is guaranteed by
absolute tightness of P∆ and Lemma 1.
Consider the case that r (i + 1) ≺ r∆B + . Then, by definition of C0 we have that
r∆B + ∈ PSi0 . Hence, from ∆B + |DA 6= ∅ follows
∆B + |DP
S0
i
6= ∅
and thus
0
ISi+1
\ ISi0 |DP
S0
i
0
The latter is a contradiction to ISi+1
|DP
S0
i
6= ∅·
= ISi0 |DP 0 .
S
i
Consider the remaining case that r∆B + ≺ r (i + 1). As ∆B + |DA 6= ∅, ∆B + ⊆ ISn ,
and ISn |DA ∈ CA , it holds that posOcc(H(r∆B + )) ∩ ∆B + 6= ∅. Therefore, we have
posOcc(H(r∆B + )) ∩ posOcc(B(r (i + 1))) 6= ∅. This implies r (i + 1) ≺ r∆B + , being
a contradiction to ≺ being a strict order as we also have r∆B + ≺ r (i + 1). Thus,
0
.
Condition (iv) of Definition 17 for being a successor of Si0 holds for Si+1
Towards a contradiction assume Condition (vi) does not hold. Hence, it must hold
that there is some ∆0 ⊆ ∆ such that ∆0 6= ∅ and r (i + 1) is not an external support
0
0
for ∆0 with respect to ISi+1
. We have ISi+1
|= B(r (i + 1)) and since we already know
0
0
0
that ISi |= B(r (i + 1)), also ISi+1 \ ∆ |= B(r (i + 1)) holds by convexity of P∆ .
0
0
|= A for H(r (i + 1)) = {A}.
|= r (i + 1), it must hold that ISi+1
Moreover, as ISi+1
Consequently, for r (i + 1) not to be an external support for ∆0 with respect to
0
ISi+1
, we have ∆0 |DA = ∅. As then ∆0 |DH(r (i+1)) = ∅ but ∆0 |Dr (i+1) 6= ∅ it must
hold that ∆0 |DB(r (i+1)) 6= ∅. Consider ∆00 = ∆0 ∩ posOcc(B(r (i + 1))) and assume
that ∆00 6= ∅. Then, as ∆00 ⊆ ISn , there must be some C-rule r∆00 that is an
external support for ∆00 with respect to ISn . Hence, posOcc(H(r∆00 )) ∩ ∆00 6= ∅
and therefore posOcc(H(r∆00 )) ∩ posOcc(B(r (i + 1))) 6= ∅. It follows that r (i +
1) ≺ r∆00 . From that we get r∆00 ∈ PSi0 . This is a contradiction as we know that
0
posOcc(H(r∆00 )) ∩ ∆00 6= ∅, posOcc(H(r∆00 )) ∩ ∆00 ⊆ ISi0 , and ∆00 ⊆ ISi+1
\ ISi0 .
0
0
Consequently, ∆ ∩ posOcc(B(r (i + 1))) = ∅ must hold. From ∆ |DB(r (i+1)) 6= ∅ we
0
get that there is some L ∈ B(r (i + 1)) with ∆0 |DL 6= ∅. As ISi+1
|= L, we have that
0
0
ISi+1
|DL ∈ C in the case L is a C-atom L = hDL , C i, and ISi+1
|DL ∈ 2D
L \ C in the
0
case L is a default negated C-atom L = not hDL , C i. In both cases, as ∆0 ⊆ ISi+1
and ∆0 |DL 6= ∅, we get a contradiction to ∆0 ∩ posOcc(B(r (i + 1))) = ∅.
References
Alviano, M., Dodaro, C., Leone, N., and Ricca, F. 2015. Advances in WASP. In Logic
Programming and Nonmonotonic Reasoning - 13th International Conference, LPNMR
2015, Lexington, KY, USA, Sept. 27-30, 2015. Proceedings, F. Calimeri, G. Ianni, and
M. Truszczyński, Eds. LNCS, vol. 9345. Springer, 40–54.
Ben-Eliyahu, R. and Dechter, R. 1994. Propositional semantics for disjunctive logic
programs. Annals of Mathematics and Artificial Intelligence 12, 1-2, 53–87.
Stepwise Debugging of Answer-Set Programs
55
Brain, M., Cliffe, O., and De Vos, M. 2009. A pragmatic programmer’s guide to
answer set programming. In Proceedings of the 2nd International Workshop on Software
Engineering for Answer-Set Programming (SEA’09), Potsdam, Germany, M. De Vos and
T. Schaub, Eds. 49–63.
Brain, M. and De Vos, M. 2005. Debugging logic programs under the answer set
semantics. In Proceedings of the 3rd International Workshop on Answer Set Programming (ASP’05), Advances in Theory and Implementation, Bath, UK, Sept. 27-29, 2005,
M. De Vos and A. Provetti, Eds. CEUR Workshop Proceedings, vol. 142. CEUR-WS.org.
Brain, M., Gebser, M., Puehrer, J., Schaub, T., Tompits, H., and Woltran, S.
2007. That is illogical Captain! The debugging support tool spock for answer-set programs – System description. In Proceeding of the 1st International Workshop on Software Engineering for Answer Set Programming (SEA’07), Tempe, AZ, USA, May 14,
2007, M. De Vos and T. Schaub, Eds. 71–85.
Brain, M., Gebser, M., Pührer, J., Schaub, T., Tompits, H., and Woltran, S. 2007.
Debugging ASP programs by means of ASP. In Proceedings of the 9th International
Conference on Logic Programming and Nonmonotonic Reasoning (LPNMR’07), Tempe,
AZ, USA, May 15-17, 2007, C. Baral, G. Brewka, and J. S. Schlipf, Eds. LNCS, vol.
4483. Springer, 31–43.
Brochenin, R., Lierler, Y., and Maratea, M. 2014. Abstract disjunctive answer
set solvers. In Proceedings of the 21st European Conference on Artificial Intelligence
(ECAI’14), Prague, Czech Republic, Aug. 18-22, 2014, T. Schaub, G. Friedrich, and
B. O’Sullivan, Eds. Frontiers in Artificial Intelligence and Applications, vol. 263. IOS
Press, 165–170.
Busoniu, P.-A., Oetsch, J., Pührer, J., Skočovský, P., and Tompits, H. 2013.
Sealion: An eclipse-based IDE for answer-set programming with advanced debugging
support. Theory and Practice of Logic Programming 13, 4-5, 657–673.
Caballero, R., Garcı́a-Ruiz, Y., and Sáenz-Pérez, F. 2008. A theoretical framework
for the declarative debugging of datalog programs. In Revised Selected Papers of the
3rd International Workshop on Semantics in Data and Knowledge Bases (SDKB’08),
Nantes, France, Mar. 29, 2008, K.-D. Schewe and B. Thalheim, Eds. LNCS, vol. 4925.
Springer, 143–159.
Cliffe, O., De Vos, M., Brain, M., and Padget, J. A. 2008. ASPVIZ: Declarative
visualisation and animation using answer set programming. In Proceedings of the 24th
International Conference on Logic Programming (ICLP’08), Udine, Italy, Dec. 9-13,
2008, M. G. de la Banda and E. Pontelli, Eds. LNCS, vol. 5366. Springer, 724–728.
Denecker, M. 2000. Extending classical logic with inductive definitions. In Proceedings
of the 1st International Conference on Computational Logic (CL’10), London, UK,
July 24-28, 2000, J. W. Lloyd, V. Dahl, U. Furbach, M. Kerber, K. Lau, C. Palamidessi,
L. M. Pereira, Y. Sagiv, and P. J. Stuckey, Eds. LNCS, vol. 1861. Springer, 703–717.
Denecker, M. and Ternovska, E. 2008. A logic of nonmonotone inductive definitions.
ACM Transactions on Computational Logic 9, 2.
Denecker, M., Vennekens, J., Bond, S., Gebser, M., and Truszczyński, M. 2009.
The second answer set programming competition. In Proceedings of the 10th International Conference on Logic Programming and Nonmonotonic Reasoning (LPNMR’09),
Potsdam, Germany, Sept. 14-18, 2009, E. Erdem, F. Lin, and T. Schaub, Eds. LNCS,
vol. 5753. Springer, 637–654.
Dodaro, C., Gasteiger, P., Musitsch, B., Ricca, F., and Shchekotykhin, K. M.
2015. Interactive debugging of non-ground ASP programs. In Logic Programming and
Nonmonotonic Reasoning - 13th International Conference, LPNMR 2015, Lexington,
KY, USA, Sept. 27-30, 2015. Proceedings, F. Calimeri, G. Ianni, and M. Truszczyński,
Eds. LNCS, vol. 9345. Springer, 279–293.
56
J. Oetsch, J. Pührer, and H. Tompits
Eiter, T., Ianni, G., Schindlauer, R., and Tompits, H. 2005. A uniform integration
of higher-order reasoning and external evaluations in answer-set programming. In Proceedings of the 19th International Joint Conference on Artificial Intelligence (IJCAI’05),
Edinburgh, Scotland, UK, July 30-Aug. 5, 2005, L. P. Kaelbling and A. Saffiotti, Eds.
Professional Book Center, 90–96.
Faber, W., Leone, N., and Pfeifer, G. 2004. Recursive aggregates in disjunctive logic
programs: Semantics and complexity. In Proceedings of the 9th European Conference
on Logics in Artificial Intelligence (JELIA’04). LNCS, vol. 3229. Springer, 200–212.
Faber, W., Pfeifer, G., and Leone, N. 2011. Semantics and complexity of recursive
aggregates in answer set programming. Artificial Intelligence 175, 1, 278–298.
Ferraris, P. 2011. Logic programs with propositional connectives and aggregates. ACM
Transactions on Computational Logic 12, 4, 25.
Frühstück, M., Pührer, J., and Friedrich, G. 2013. Debugging answer-set programs
with Ouroboros – Extending the SeaLion plugin. In Proceedings of the 12th International Conference on Logic Programming and Nonmonotonic Reasoning (LPNMR’13),
Corunna, Spain, Sept. 15-19, 2013, P. Cabalar and T. C. Son, Eds. LNCS, vol. 8148.
Springer, 323–328.
Gebser, M., Kaminski, R., König, A., and Schaub, T. 2011. Advances in gringo series
3. In Proceedings of the 11th International Conference on Logic Programming and
Nonmonotonic Reasoning (LPNMR’11), Vancouver, Canada, May 16-19, 2011, J. P.
Delgrande and W. Faber, Eds. LNCS, vol. 6645. Springer, 345–351.
Gebser, M., Kaufmann, B., and Schaub, T. 2012. Conflict-driven answer set solving:
From theory to practice. Artificial Intelligence 187-188, 52–89.
Gebser, M., Lee, J., and Lierler, Y. 2011. On elementary loops of logic programs.
Theory and Practice of Logic Programming 11, 6, 953–988.
Gebser, M., Pührer, J., Schaub, T., and Tompits, H. 2008. A meta-programming
technique for debugging answer-set programs. In Proceedings of the 23rd AAAI Conference on Artificial Intelligence (AAAI’08), Chicago, IL, USA, July 13-17, 2008, D. Fox
and C. P. Gomes, Eds. AAAI Press, 448–453.
Gebser, M., Pührer, J., Schaub, T., Tompits, H., and Woltran, S. 2009. spock: A
debugging support tool for logic programs under the answer-set semantics. In Revised
Selected Papers of the 17th International Conference on Applications of Declarative
Programming and Knowledge Management (INAP’07) and 21st Workshop on (Constraint) Logic Programming (WLP’07), D. Seipel, M. Hanus, and A. Wolf, Eds. LNCS,
vol. 5437. Springer, 247–252.
Gelfond, M. and Lifschitz, V. 1991. Classical negation in logic programs and disjunctive databases. New Generation Computing 9, 3/4, 365–386.
Kloimüllner, C., Oetsch, J., Pührer, J., and Tompits, H. 2013. Kara: A system
for visualising and visual editing of interpretations for answer-set programs. In Revised
Selected Papers of the 19th International Conference on Applications of Declarative
Programming and Knowledge Management (INAP’11) and the 25th Workshop on Logic
Programming (WLP’11), Vienna, Austria, Sept. 28-30, 2011. LNCS, vol. 7773. Springer,
325–344.
Lee, J. 2005. A model-theoretic counterpart of loop formulas. In Proceedings of the
19th International Joint Conference on Artificial Intelligence (IJCAI’05), Edinburgh,
Scotland, UK, July 30-Aug 5, 2005, L. P. Kaelbling and A. Saffiotti, Eds. Professional
Book Center, 503–508.
Leone, N., Pfeifer, G., Faber, W., Eiter, T., Gottlob, G., Perri, S., and Scarcello, F. 2006. The DLV system for knowledge representation and reasoning. ACM
Transactions on Computational Logic 7, 3, 499–562.
Stepwise Debugging of Answer-Set Programs
57
Li, T., Vos, M. D., Padget, J., Satoh, K., and Balke, T. 2015. Debugging ASP
using ILP. In Technical Communications of the 31st International Conference on Logic
Programming (ICLP 2015), Cork, Ireland, Aug. 31 - Sept. 4, 2015., M. D. Vos, T. Eiter,
Y. Lierler, and F. Toni, Eds. CEUR Workshop Proceedings, vol. 1433. CEUR-WS.org.
Lierler, Y. 2011. Abstract answer set solvers with backjumping and learning. Theory
and Practice of Logic Programming 11, 2-3, 135–169.
Lierler, Y. and Truszczyński, M. 2016. On abstract modular inference systems and
solvers. Artificial Intelligence 236, 65–89.
Liu, L., Pontelli, E., Son, T. C., and Truszczyński, M. 2010. Logic programs with
abstract constraint atoms: The role of computations. Artificial Intelligence 174, 3-4,
295–315.
Marek, V. W. and Remmel, J. B. 2004. Set constraints in logic programming. In Proceedings of the 7th International Conference on Logic Programming and Nonmonotonic
Reasoning (LPNMR’04), Fort Lauderdale, FL, USA, Jan 6-8, 2004, V. Lifschitz and
I. Niemelä, Eds. LNCS, vol. 2923. Springer, 167–179.
Marek, V. W. and Truszczyński, M. 1999. Stable models and an alternative logic programming paradigm. In In The Logic Programming Paradigm: a 25-Year Perspective,
K. R. Apt, V. W. Marek, M. Truszczyński, and D. S. Warren, Eds. Springer, 375–398.
Marek, V. W. and Truszczyński, M. 2004. Logic programs with abstract constraint
atoms. In In Proceedings of the 19th National Conference on Artificial Intelligence
(AAAI’04), San Jose, CA, USA, July 25-29, 2004, G. Ferguson and D. McGuinness,
Eds. AAAI Press, 86–91.
Mikitiuk, A., Moseley, E., and Truszczyński, M. 2007. Towards debugging of answerset programs in the language PSpb. In Proceedings of the 2007 International Conference
on Artificial Intelligence (ICAI’07), Volume II, Las Vegas, NV, USA, June 25-28, 2007,
H. R. Arabnia, M. Q. Yang, and J. Y. Yang, Eds. CSREA Press, 635–640.
Niemelä, I. 1999. Logic programs with stable model semantics as a constraint programming paradigm. Annals of Mathematics and Artificial Intelligence 25, 3-4, 241–273.
Nieuwenhuis, R., Oliveras, A., and Tinelli, C. 2006. Solving SAT and SAT modulo theories: From an abstract davis–putnam–logemann–loveland procedure to dpll(T).
Journal of the ACM 53, 6, 937–977.
Oetsch, J., Prischink, M., Pührer, J., Schwengerer, M., and Tompits, H. 2012. On
the small-scope hypothesis for testing answer-set programs. In Proceedings of the 13th
International Conference on Principles of Knowledge Representation and Reasoning
(KR’12), Rome, Italy, June 10-14, 2012, G. Brewka, T. Eiter, and S. A. McIlraith, Eds.
AAAI Press.
Oetsch, J., Pührer, J., and Tompits, H. 2010. Catching the Ouroboros: On debugging
non-ground answer-set programs. Theory and Practice of Logic Programming 10, 4-6
(July), 513–529.
Oetsch, J., Pührer, J., and Tompits, H. 2010. Let’s break the rules: Interactive
procedural-style debugging of answer-set programs. In Proceedings of the 24th Workshop on (Constraint) Logic Programming (WLP’10), Cairo, Egypt, Sept. 14-16, 2010,
S. Abdennadher, Ed. Technical Report, Faculty of Media Engineering and Technology,
German University in Cairo. 77–87.
Oetsch, J., Pührer, J., and Tompits, H. 2011. Stepping through an answer-set program. In Proceedings of the 11th International Conference on Logic Programming and
Nonmonotonic Reasoning (LPNMR’11), Vancouver, Canada, May 16-19, 2011. LNCS,
vol. 6645. Springer, 134–147.
Oetsch, J., Pührer, J., and Tompits, H. 2012a. An FLP-style answer-set semantics
for abstract-constraint programs with disjunctions. In Technical Communications of the
58
J. Oetsch, J. Pührer, and H. Tompits
28th International Conference on Logic Programming (ICLP’12), Budapest, Hungary,
A. Dovier and V. S. Costa, Eds. LIPIcs, vol. 17. Schloss Dagstuhl - Leibniz-Zentrum
fuer Informatik, 222–234.
Oetsch, J., Pührer, J., and Tompits, H. 2012b. Stepwise debugging of descriptionlogic programs. In Correct Reasoning - Essays on Logic-Based AI in Honour of Vladimir
Lifschitz, E. Erdem, J. Lee, Y. Lierler, and D. Pearce, Eds. LNCS, vol. 7265. Springer,
492–508.
Oetsch, J., Pührer, J., and Tompits, H. 2013. The SeaLion has landed: An IDE for
answer-set programming—Preliminary report. In Revised Selected Papers of the 19th
International Conference on Applications of Declarative Programming and Knowledge
Management (INAP’11) and the 25th Workshop on Logic Programming (WLP’11),
Vienna, Austria, Sept. 28-30, 2011. LNCS, vol. 7773. Springer, 305–324.
Polleres, A., Frühstück, M., Schenner, G., and Friedrich, G. 2013. Debugging
non-ground ASP programs with choice rules, cardinality constraints and weight constraints. In Proceedings of the 12th International Conference on Logic Programming
and Nonmonotonic Reasoning (LPNMR’13), Corunna, Spain, Sept. 15-19, 2013, P. Cabalar and T. C. Son, Eds. LNCS, vol. 8148. Springer, 452–464.
Pontelli, E., Son, T. C., and El-Khatib, O. 2009. Justifications for logic programs
under answer set semantics. Theory and Practice of Logic Programming 9, 1, 1–56.
Pührer, J. 2007. Master’s thesis, Vienna University of Technology, Vienna, Austria.
Pührer, J. 2014. Stepwise debugging in answer-set programming: Theoretical foundations and practical realisation. Ph.D. thesis, Vienna University of Technology, Vienna, Austria. http://repositum.tuwien.ac.at/urn:nbn:at:at-ubtuw:1-75281 [Online; accessed Dec. 14, 2016].
Redl, C. 2016. The dlvhex system for knowledge representation: recent advances (system
description). Theory and Practice of Logic Programming 16, 5-6, 866–883.
Shapiro, E. Y. 1982. Algorithmic program debugging. Ph.D. thesis, Yale University,
New Haven, CT, USA.
Shchekotykhin, K. M. 2015. Interactive query-based debugging of ASP programs. In
Proceedings of the 29th AAAI Conference on Artificial Intelligence (AAAI’15), Austin,
TX, USA, Jan. 25-30, 2015, B. Bonet and S. Koenig, Eds. AAAI Press, 1597–1603.
Simons, P., Niemelä, I., and Soininen, T. 2002. Extending and implementing the stable
model semantics. Artificial Intelligence 138, 181–234.
Smith, A. 2011. Lonsdaleite. https://github.com/rndmcnlly/Lonsdaleite. [Online;
accessed Dec. 14, 2016].
Syrjänen, T. 2006. Debugging inconsistent answer set programs. In Proceedings of the
11th International Workshop on Non-Monotonic Reasoning (NMR’06), Lake District,
UK, May 30-June 1, 2006, J. Dix and A. Hunter, Eds. Institut für Informatik, Technische
Universität Clausthal, Technical Report, 77–83.
Truszczyński, M. 2010. Reducts of propositional theories, satisfiability relations, and
generalizations of semantics of logic programs. Artificial Intelligence 174, 16-17, 1285–
1306.
Wittocx, J. 2009. IDPDraw, a tool used for visualizing answer sets. https://dtai.cs.
kuleuven.be/software/idpdraw. [Online; accessed Dec. 14, 2016].
Wittocx, J., Vlaeminck, H., and Denecker, M. 2009. Debugging for model expansion.
In Proceedings of the 25th International Conference on Logic Programming (ICLP’09),
Pasadena, CA, USA, July 14-17, 2009, P. M. Hill and D. S. Warren, Eds. LNCS, vol.
5649. Springer, 296–311.
| 6cs.PL
|
ISEC:
Iterative over-Segmentation via Edge Clustering
arXiv:1802.05816v1 [cs.CV] 16 Feb 2018
Marcelo Mendonça, Luciano Oliveira
Intelligent Vision Research Lab,
Federal University of Bahia,
Salvador-BA, Brazil
Abstract
Several image pattern recognition tasks rely on superpixel generation as a fundamental step. Image analysis based on superpixels facilitates domain-specific
applications, also speeding up the overall processing time of the task. Recent
superpixel methods have been designed to fit boundary adherence, usually regulating the size and shape of each superpixel in order to mitigate the occurrence of
undersegmentation failures. Superpixel regularity and compactness sometimes
imposes an excessive number of segments in the image, which ultimately decreases the efficiency of the final segmentation, specially in video segmentation.
We propose here a novel method to generate superpixels, called iterative oversegmentation via edge clustering (ISEC), which addresses the over-segmentation
problem from a different perspective in contrast to recent state-of-the-art approaches. ISEC iteratively clusters edges extracted from the image objects,
providing adaptive superpixels in size, shape and quantity, while preserving
suitable adherence to the real object boundaries. All this is achieved at a very
low computational cost. Experiments show that ISEC stands out from existing methods, meeting a favorable balance between segmentation stability and
accurate representation of motion discontinuities, which are features specially
suitable to video segmentation.
Keywords: Superpixels, Video object segmentation
∗ Corresponding
author
Email address: [email protected] (Luciano Oliveira)
1
1. Introduction
The result of an image over-segmentation lies in somewhere between the
pixel and the object segmentation itself. The term “superpixel” has been used
to refer to the information in such intermediate level. Indeed, many tasks that
involves image segmentation take advantage of working with superpixels rather
than pixels. Examples range from tracking [1] through 3D reconstruction [2] to
semantic image segmentation [3]. Superpixels are also applied in many domainspecific applications, such as in traffic analysis [4], and in biological [5] and
medical [6] image segmentation, just to name a few. In view of this wide applicability, many superpixel methods have been proposed in recent years (see
recent surveys in [7, 8]).
Research on superpixel started with image segmentation methods running
on over-segmentation mode. Earlier methods were not explicitly conceived to
generate image superpixels. For instance, segmentation methods such as watershed (WS) [9] and normalized cuts (NC) [10] are able to produce the oversegmentation effect by adjusting the number of seeds, in the case of the former,
or the number of graph partitions, in the latter. Later methods such as edge augmented mean shift (EAMS) [11], Felzenszwalb-Huttenlocher (FH) [12] and quick
shift (QS) [13] were specifically designed to produce image over-segmentation
by pursuing adherence to real object contours. In these latter methods, the rationale consists in automatically regulating the amount of generated superpixels
according to internal parameters, although without providing direct control over
the number of generated segments. More recently, several methods have been
groundly developed towards generating superpixels by directly providing control
over the number of segments [14, 15, 16, 17, 18, 19]; the goal is to divide the image into a generally user-defined fixed number of segments, also trying to match
the segment boundaries as accurate as possible to the real object contours. The
advantages of these methods are (i) the larger the chosen number of superpixels, the lower the probability of occurring undersegmentation errors, i.e., when
2
Figure 1: Examples of superpixel segmentation. From left to right column: SLIC [15] segmentation, with k = 200, LSC [19] segmentation, with k = 200, and ISEC segmentation, with
k = 158, where k is the number of generated superpixels; for SLIC and LSC, k is user-defined,
while for ISEC, k is image-based adaptive. Bottom images highlight the parachute man segmentation for each method: SLIC and LSC present undersegmentation failures, while ISEC
correctly segments the image object.
the same segment overlaps different objects, (ii) approaches based on a fixed
number of segments usually provide more regular superpixels regarding shape
and size, which can be useful for some applications, e.g., object recognition [20]
and labeling [21], and (iii), since superpixel methods are commonly used as a
preprocessing stage for other computer vision tasks, by keeping control over the
number of segments, one can make the computational effort in the further steps
be more predictable.
Deciding about the better strategy to generate superpixels – if the number
of segments is automatically or externally provided – gives rise to a straightforward question: “How many superpixels should an over-segmentation method
3
generate for a given image?” We could intuitively say that the answer depends
on the image. In words, if there are many objects, more segments will be
necessary; otherwise, just a small number would be enough. Despite this immediate conclusion, more recent superpixel methods have not actually addressed
the over-segmentation problem from such an adaptive perspective. Specially
because of the undersegmentation problem, most of the recent methods goes
towards generating a fixed and externally provided number of superpixels. To
illustrate that situation, Fig. 1 shows the results of superpixel generation by
two state-of-the-art methods: simple linear iterative clustering (SLIC) [15] (left
column), and linear spectral clustering (LSC) [19] (middle column). In the figure, both SLIC and LSC were set to yield 200 superpixels. With this number,
both methods provide superpixels that correctly adhere to the contours of the
big mountain in the figure, although the parachute man is not accurately segmented, causing an undersegmentation failure. Errors, like in the parachute
man segmentation, mainly take place when the generated superpixels are much
bigger than the object being segmented. An alternative to tackle this problem
could be to increase the number of superpixels, making them smaller. However,
an excessive over-segmentation implies to segment other objects in the image
into much more parts than necessary, reducing efficiency1 . Also, the grid-like
segmentation provided by methods such as SLIC and LSC does not allow to
distinguish homogeneous areas from object-filled regions in the image (e.g., the
blue sky in the background), resulting in unnecessary over-segmentation of these
areas.
1.1. Contributions
A method called iterative over-segmentation via edge clustering (ISEC) is
proposed in this paper. ISEC is a novel superpixel method that addresses the
over-segmentation problem from a different perspective in contrast to recent
1 In
this context, efficiency means capturing the maximum number of real object contours
with the minimum number of superpixels.
4
state-of-the-art approaches. The proposed method iteratively clusters edges
extracted from the image objects, providing superpixels that are, at the same
time, adaptive in size, shape and quantity, while preserving efficient adherence
to the real object boundaries. All this is achieved at a very low computational
cost. The right column of Fig. 1 illustrates the result of ISEC over-segmentation
on the same image segmented by SLIC and LSC. It is noteworthy that ISEC
is able to self-adapting to correctly segment both the big mountain and the
small parachute man, avoiding unnecessarily segmenting most of the background
comprised of the homogeneous blue sky. Thanks to the adaptiveness of the
proposed method, the resulting over-segmentation in Fig. 1 was accomplished
with comparatively fewer (158) superpixels than SLIC and LSC, turning ISEC
to be more efficient in this case. On the other hand, superpixels generated by
SLIC and LSC are much more regular in shape and size, while ISEC prioritizes
superpixels more adjusted to the real object contours. For some applications,
such as that one to segment magnetic resonance images [22], adherence to real
object contour is a crucial characteristic rather than superpixel regularity or
compactness, since superpixels with compact structure are less able to cover
complete objects in that type of image.
The proposed method iteratively groups image edges in order to form clusters. The rationale is that the borders that delimit these clusters are to be
superpixels presenting real object contour adherence efficiently (see Section 3
for details). Results from video-based evaluation metrics [23] indicate that the
superpixels generated by ISEC demonstrate promising performance regarding
video segmentation. This assessment is corroborated by the results found in a
video object segmentation experiment (see Section 4). Indeed, previous comparisons [24] have demonstrated that superpixels with characteristics similar
to ISEC perform better than grid-like, regular-shaped superpixels in video object segmentation. In [23], Neubert and Protzel make a thorough evaluation
on superpixel methods focused on video segmentation; they point out a lack of
superpixel methods able to produce segmentations that are consistent in two
criteria: segmentation stability and accurate representation of motion discon5
tinuities. Our experiments show that ISEC stands out from existing methods,
meeting a favorable balance between the aforementioned criteria.
2. Related works
Some of the existing superpixel methods are reviewed here. The presented
methods are further compared to ISEC in our experimental evaluation (results
on evaluation performance are reported in Section 4).
WS (1992) [9] – Watershed is an well-known, very fast over-segmentation algorithm. It works by iteratively growing user-defined seeds until they reach the
borders of other adjacent growing seeds. The number of superpixels is determined by the number of seeds. There is an open source C/C++ implementation
available at OpenCV2 .
NC (2000) [10] – The Normalized Cuts is a widely used graph-based algorithm,
and maybe the first applied to generate superpixels [7]. As such, NC recursively
produces a graph, in which the vertices correspond to the image pixels, while
the weights account for similarities among the pixels. Superpixels are obtained
by partitioning this graph in a user-defined number of segments. This process
takes a lot of time when compared to other methods. Our experiments using NC
have been done with resized images (160 pixels on the longer side, while keeping
the aspect ratio) to save time. An open source implementation for Matlab and
C++ is available3 .
EAMS (2001) [11] – Edge Augmented Mean Shift is a density-based algorithm
that performs an iterative mode-seeking procedure in a computed density image.
This procedure consists in finding modes in color or intensity feature space so
that each pixel is assigned to the corresponding mode where it falls into. The
superpixels are composed by the pixels that converge to the same mode. This
2 http://opencv.willowgarage.com/wiki/
3 http://www.timotheecour.com/software/ncut/ncut.html
6
method mainly differs from the k-means algorithm since it does not require a
user-defined number of clusters. Source codes in C and Matlab are available4 .
FH (2004) [12] – Felzenszwalb-Huttenlocher is a graph-based superpixel approach. The algorithm measures the evidence of a boundary between two regions by computing a graph-based representation of the image. FH tries to
preserve details when a region presents low variability, while details in image regions with high variability are ignored. This is done by using a greedy algorithm
that also satisfies global properties. FH automatically controls the number of
superpixels, and it is among the fastest superpixel methods. An open source
C++ implementation is available5 .
QS (2008) [13] – Quick Shift is also a mode-seeking segmentation method.
Instead of relying on a gradient ascent approach such as EAMS, QS uses a
medoid shift. The medoid is faster than the mean shift in Euclidean space. The
algorithm builds a tree of paths by moving each point in the feature space to the
nearest neighbor that increases the density estimative. Superpixels are obtained
by splitting the branches of the tree using a threshold. The number of segments
is automatically controlled, and the method is not fast. Implementation in C
and Matlab is available as part of the VLFeat library6.
VEK (2010) [14] – Veksler Superpixels. This method is also a graph-based
approach, but based on regular partition. The segmentation is assumed as
an energy minimization problem that explicitly encourages constant intensity
inside the superpixels. Segmentations are generated by gluing overlapping image
patches until each pixel remains only inside of an individual patch. There are
two available implementations of VEK, one for compact and other for constantintensity superpixels. In our experiments, we evaluated only the latter7 . VEK
is among the slowest algorithms, and the number of superpixels is user-defined.
4 http://www.wisdom.weizmann.ac.il/
~ bagon/matlab.html
~ pff/segment/
6 http://www.vlfeat.org/overview/quickshift.html
7 http://www.csd.uwo.ca/faculty/olga/code.html
5 http://www.cs.brown.edu/
7
SLIC (2010) [15] – Simple Linear Iterative Clustering is inspired by the k-means
algorithm. However, the strategy is modified to limit the region in which the
algorithm searches for similarities among the pixels. SLIC considers only the
pixels within a defined spatial neighborhood. The process begins by dividing
the image into a regular grid where the number of cells is defined by the user.
Color and spatial distances among the pixels are measured in order to iteratively
adjust the cells to form superpixels. SLIC is among the fastest algorithms. An
implementation in C++ and Matlab is available8 .
ERS (2011) [16] – Entropy Rate Superpixel Segmentation is formulated as a
maximization problem on a graph-based image representation. The method
relies on an objective function for superpixel segmentation that accounts for the
entropy among the graph edges and the size of the segments. ERS generates
superpixels by selecting the edges that yield the largest gain of the objective
function. ERS iterates until the user-defined number of clusters is reached,
which usually takes over a second. An implementation in C/C++ is available9 .
SEEDS (2012) [17] – Superpixels Extracted via Energy-Driven Sampling. This
method iteratively optimizes an energy function, starting from a grid-like image
segmentation and then modifying the boundaries according to the function. In
each step, pixels are exchanged between neighboring segments so that if the new
composition produces a gain of the energy function, it is retained, otherwise it is
rejected. The number of generated superpixels is user-defined, and the method
is among the fastest ones. An implementation in C and Matlab is available10 .
CWS (2014) [18] – Compact Watershed is a modified version of the original
WS method, aiming at providing control over the superpixel compactness. The
modification encompasses two aspects: grid arrangement of the seeds and incorporation of a controllable compactness constraint. This is done by means of
a single parameter that controls the level of compactness regardless the num8 http://ivrl.epfl.ch/supplementary_material/RK_SLICSuperpixels/index.html
9 https://github.com/mingyuliutw/EntropyRateSuperpixel
10 http://www.mvdblive.org/seeds/
8
ber of segments or the image size. CWS is so fast as the original WS. An
implementation in C/C++ is available11 .
LSC (2015) [19] – Linear Spectral Clustering uses kernel functions to construct
a ten-dimensional representation of the image. This high dimensionality is given
by a set of features extracted for each pixel. After initializing clusters from seeds
distributed along the image, an iterative process measures the distances between
pixels and clusters in the feature space, assigning the pixels to the cluster with
the closest distance. This process ends when the clusters become stable, which
is accomplished in a few reasonable time. The number of clusters (superpixels)
is user-defined. An implementation in C++ with Matlab interface is available12 .
3. Superpixel generation via edge clustering
Superpixels are generated from image edges, which are iteratively selected
and grouped to form clusters. Clustered pixels represent image locations where
the spatial neighborhood is densely filled by edges. Since the edges are strongly
related to the objects contours and texture, the clusters resemble the object
shapes. Superpixels are ultimately obtained by extracting the borders that separate these clusters from image regions where there are no edges. By adjusting
the edge selection procedure so that a group of edges are selected at the beginning of the process, and edges are progressively removed at each new iteration,
the input image is over-segmented from the outer contours of the objects to
their small internal parts. Figure 2 depicts the top-down view of ISEC. Next,
each part of the proposed method is described.
3.1. Gradient extraction
Given an input image (Fig. 2.a), the first step is to compute, for each
pixel, the gradient magnitude (Fig. 2.b) and orientation along the x and y axes
(Fig. 2.c). Canny edge detector is used to calculate the image gradients. The
11 https://www.tu-chemnitz.de/etit/proaut/forschung/cv/segmentation.html.en
12 https://jschenthu.weebly.com/projects.html
9
Figure 2: Top-down view of ISEC. Gradient extraction: An input image (a) is used to compute gradient magnitudes (b) and orientations (c) in the x and y axes. Cumulative contour
extraction: For each iteration, an edge set (d) is selected from the gradient; the edges are
stretched (e) by edge density filtering, being binarized to form clusters (f); a thinning operation is performed on the clusters to readjust their shapes (g); the borders of the clusters are
extracted (h) and stored. Superpixel refinement: The accumulated segments (i) are refined to
produce the final result (j). The generated superpixels are showed over the input image (k);
some parts are zoomed to highlight segmentation details.
gradients are individually computed over each RGB channel. The final gradient
magnitude is given by the highest magnitude presented by the pixel among the
channels. By doing that, color information is also encoded in the edge detection
process.
10
3.2. Cumulative contour extraction
Once the image gradients are computed, a set of edges (Fig. 2.d) is initially selected from the gradient map. Edge selection is iteratively performed
by combining non-maximum suppression and double thresholding, just as in
Canny detection13 . Initially, small thresholds are used so that a great amount
of edges are selected. As the process iterates, the thresholds become more restrictive, yielding progressively fewer edges. The idea consists in extracting the
object contours in an outside-in manner, i.e., starting from the silhouette to
inward subparts. This strategy is accomplished by handling the edge detection
thresholds. For small thresholds, many edges are selected, including those ones
related to strong contours (high gradient magnitude), or even edges originated
from smooth contours and textures (low gradient magnitude). This initial great
amount of edges tend to compose clusters with the same shape as the object that
originated the edges. As the process iterates, the thresholds are incremented,
becoming increasingly restrictive. These higher thresholds prevent many edges
to be selected from smooth contours and texture, remaining only the edges
related to stronger contours. These stronger edges tend to resemble smaller
internal subparts of bigger objects.
In principle, directly using edges to construct superpixel is tempting, since
edges are strongly related to the real object contours. However, this is not
practical because edges generally present discontinuities, resulting in unclosed
segments. To tackle this problem we perform a spatial linear filtering on the
edge image, given by
2
S
1 X
pi ,
ED = 2
S i=1
(1)
where ED denotes edge density, which is calculated by an arithmetic mean over
the neighborhood SxS of the pixel pi .
13 Different
edge detection strategies could be used instead, and Canny was chosen because
of the best results at the experiments.
11
Figure 3: Edge reconstruction. The ED filter computes the edge density in the neighborhood
SxS of each pixel pi ; the effect is to blur the edge pixels through the filter area, forming a
cluster where the gaps are filled. Next, binarization is applied to separate the edge cluster
from the empty surroundings. A thinning operation is performed (S − 1)/2 times in order to
readjust the cluster shape; the most outer pixels are removed without causing new gaps.
Although many pixels in the resulting image represent edge discontinuities,
the local ED in these pixels is not null. The effect of the filtering process is to
blur the edges through the filter area (Fig. 2.e), causing many gaps to be filled.
12
Image areas massively occupied by edges become clusters that contrasts with the
empty surroundings. After binarizing the image (Fig. 2.f), clusters are precisely
distinguished from the empty regions. ISEC follows the idea that the boundaries
that separate these two regions are convenient to compose superpixels contours.
The ED filter also causes a dilation on the clusters. That edge dilation has
the side effect of driving the edges to lose their intrinsic adherence to the real
object contours. This issue is coped by applying a thinning operation on the
edge clusters in order to readjust their shape (Fig. 2.g). The thinning is a
morphological operation that removes the most external pixels in a segment of
a binary image towards a line with the thickness of a single pixel – not allowing
for new discontinuities. Since the ED filter expands the clusters by S − 1 pixels,
the thinning operation needs to be performed (S−1)/2 times in order to readjust
the cluster borders to correctly match the real object contours (considering that
this operation performs symmetrically in both sides of the segments). Figure
3 illustrates, at the pixel level, how this filtering-binarizing-thinning process
accomplishes the edge reconstruction.
After reshaping the edge clusters, the next step is to extract the cluster
borders (Fig. 2.h). In order to save processing time, this step is done by using
look-up tables (LUT). These LUT’s allow for fast pixel-level operations in a
binary image, by replacing more expensive run-time calculations with simple
searches for precomputed values in memory. To extract the cluster borders,
we search for 1’s pixels that have at least one 0 in their 3×3 neighborhood.
Before the searching, the look-up table has been fetched with all the 512 possible
patterns that can be found in the 3×3 neighborhood of the pixel in the binary
image. At the end of each iteration, the set of extracted borders is stored.
Additionally, the edge detection thresholds are incremented and the order, S,
of the ED filter is reduced. The goal is to make the filter to better fit into the
smaller parts of the objects at the next iteration.
13
3.3. Superpixel refinement
When the segments provided by each previous iteration are assembled together, some borders eventually end up being adjacent, resulting in thick lines
(Fig. 2.i). Also, some segments can be too small or even individual pixels. The
purpose of this superpixel refinement stage is to use morphological operations
to remove such irregularities, improving the final result (Fig. 2.j). Figure 2.k
shows the generated superpixels overlapped with the input image; some parts
are zoomed to highlight segmentation details. The proposed superpixel segmentation method is summarized in Algorithm 1.
Algorithm 1 Iterative over-Segmentation via Edge Clustering
Input: Image I; size S of the filter; lower and higher thresholds [t, T ]
Output: Labeled image containing superpixels SP X
1: for each pixel pi in I do
2:
G ← argmax(gradient(pi{R} , pi{G} , pi{B} ))
3: end for
4: SP X ← 0
5: for k = t to T do
6:
E ← edgeSelection(G,k)
7:
for each pixel pi in a region S×S of E do
8:
9:
ED ←
1
S2
end for
PS 2
i=1
pi
⊲ Edge Density
10:
Binarize ED
11:
Perform thinning of ED, ((S − 1)/2) times
12:
C ← clusterContourExtraction(ED)
13:
SP X ← OR(SP X,C)
14:
if S >minimum filter order then
15:
16:
S ← S −2
end if
17: end for
18: Merge adjacent contours and remove isolated pixels in SP X
19: Label SP X
14
4. Experimental analysis
4.1. Methodology
Most of the recent methods for superpixel generation rely on user-defined
number of superpixels. That fact drove to performance metrics, which are computed as function of the number of superpixels, such as: video-based metrics – motion undersegmentation error (MUSE) and motion discontinuity error
(MDE) (see Section 4.2), and image-based metrics – undersegmentation error (UE) and boundary recall (BR) (see Section 4.3). For a given dataset, the
usual evaluation methodology consists in running the methods on all the images
repeatedly, each time generating a segmentation set with a different number of
superpixels. The evaluation metrics are computed for all the images in each
segmentation set, and averaged by the total of images. The results are usually
plotted in a graph, putting the number of superpixels against the value of the
respective metric. This methodology imposes constraints to evaluate methods
such as ISEC, whose the number of superpixels depends on the image. In order
to make ISEC be comparable to the other methods, the evaluations were run by
varying the step in which the ISEC’s edge detection thresholds are incremented.
By setting a small step, the thresholds are increased more slowly, resulting in
more edges being selected and, consequently, more superpixels. Conversely, the
bigger the steps, the fewer the superpixels. It is noteworthy that such procedure
only affects the number of superpixels indirectly. This procedure achieves an
average number of superpixels on a given data set, although this number can
considerably vary for each individual image.
Experimental evaluation also includes assessing performance of the evaluated methods based on an existing application on video object segmentation (see Section 4.4). The methods are evaluated according to the overall
performance achieved by the application to segment a video data set. For that
video segmentation task, superpixel generation works as an independent module, allowing for the evaluation of different superpixel methods compared in the
experiment.
15
4.2. Video-based evaluation metrics
In [23], a benchmark on video segmentation is proposed to evaluate superpixel methods. Primarily, the benchmark provides completely data-driven metrics, based on optical flow. The motivation is to avoid the ground truth biasing
due to eventual semantic knowledge inserted by human-made annotations. To
better understand, suppose evaluating an object segmentation provided by a
given method. Although separating the object from the background and surroundings is an obvious requirement, it is not always clear which subparts of the
object itself need to be segmented apart. As humans intrinsically rely on contextual information when doing that task, metrics based on manually annotated
data may be imprecise. Using optical flow as ground truth guarantees that only
information contained in the image itself is used in the evaluation process.
The benchmark provides two metrics, based on the following criteria:
• Stability - Evaluates the capability of the segmentation method to find
the same regions or object boundaries regardless of image changes. The
metric to assess this criterion is called MUSE, which measures the stability
of a segmentation between two frames, I1 and I2 , and is given as
1 X
M U SE =
N
F
a∈L1
X
b∈L2 :a∩b6=0
min(bin , bout ) ,
(2)
where N is the total number of pixels, LF
1 is a segmentation of I1 in which
the labels were transformed by using optical flow into the pixels of I2 ;
L2 is the segmentation of I2 . Individual superpixels belonging to these
segmentations are, respectively, a and b; bout is the number of pixels of b
that are outside a, while bin is the number of pixels that are in a ∩ b. The
rationale is to segment subsequent frames, and then use the ground-truth
optical flow to transform the first segmentation into the view of the second
image, making the segmentations comparable. MUSE measures how well
one segmentation can be reconstructed from the other.
16
• Discontinuity - Evaluates how well the generated superpixels can represent motion discontinuities in the image sequence. For instance, motion
differences between moving foreground objects and the background, or
among parts of the same object. The metric, called MDE, is defined as
XX
1
k∇F (i, j)k2 .D(B(i, j)) ,
j k∇F (i, j)k2 i
j
M DE = P P
i
(3)
where F is the ground-truth optical flow between two frames, B is the
superpixel segmentation of one of these frames, and D(B) is the distance
transform of B which contains, for each pixel, the distance to the nearest
segment boundary. MDE accumulates a penalty over all image pixels,
which is the product of the magnitude of motion discontinuity at this
pixel and its distance to the next segment border (normalized by the total
amount of motion in the image). Please, refer to [23] for more details on
the metric.
The benchmark uses the Sintel data set [25], which has 23 computer rendered
scenes, each one with 20 to 50 color images of size 1024 × 436. The ground-truth
optical flow has been directly extracted from the data used for rendering. The
usage of computer rendered images allows for the presence of denser optical
flow data in Sintel compared to data sets based on natural images. This denser
optical flow data supplies ground truth information for basically all image pixels,
even when some pixels are occluded among a sequence of frames.
We compare ISEC performance on Sintel data set with the methods: ERS
[16], QS [13], EAMS [11], FH [12], SEEDS [17], VEK [14], WS [9], NC [10], CWS
[18], SLIC [15] and LSC [19]. Also, a simple regular grid segmentation (BOX)
has been included for baseline comparison. For all the compared methods, the
implementations are based on publicly available code. The results are illustrated
in Fig. 4.
MUSE penalizes segments for which counterparts are not found in subsequent frames. Therefore an increase of the error with respect to the number
of superpixels is expected, since by raising this number, homogeneous image
17
Figure 4: Video-based evaluation results on Sintel data set. (a) MUSE measures the segmentation stability among the frames; methods that over-segment homogeneous regions in the
image perform worse in this metric. (b) MDE evaluates how well the motion discontinuities
present in the video sequence are represented by the generated superpixels; methods that provide segmentations where the superpixels are distributed by the whole image area can better
represent motion discontinuities. ISEC is the only method to perform better than the baseline
(BOX) in both MUSE and MDE, indicating that the proposed method has suitable features
for video segmentation.
regions are further segmented. Such homogeneous regions do not offer gradient
support for the segmentation process, entailing a lot of variations in the segments among the frames, which contributes for instability. That is the reason
why methods like EAMS, FH and ISEC, which do not rely on a grid-like segmentation, perform best on this stability criterion. In turn, the other methods
present results that are next or even worse than the baseline BOX, as showed
in Fig. 4.a.
For MDE, an opposite behavior is expected, that is, the greater the number
of superpixels, the lower the error. This is because by increasing the image
partitioning, more boundaries are produced. Hence it is more likely that some
of these boundaries lie on motion discontinuities regions in the image. According
to [23], MDE is complementary to MUSE. This becomes evident in the plots
of Fig. 4.b, where the methods appear in an inverted order compared to the
previous MUSE evaluation. In MDE evaluation, methods such as SLIC, ERS,
18
SEEDS, LSC, VEK, NC, CWS and QS, which provide grid-like segmentations
(and consequently more distributed boundaries along the image area), present
the best results. EAMS and FH, on the other hand, perform poorly on this
metric.
It is remarkable that on the Sintel data set, ISEC is the only method to
perform better than the BOX in both MUSE and MDE. This result tackles the
problem pointed out by [23]: methods that perform well on one criterion often
show problems with the other, that is, stability or motion discontinuity representation. In words, the results indicate that our method provides a balance
between segmentation stability and accuracy for video segmentation, filling the
gap emphasized by the benchmark. Figure 5 shows a visual comparison of Sintel
segmentations generated by ISEC and four other methods: SLIC and LSC, in
which the number of superpixels is user-defined; and EAMS and FH, in which
the number of superpixels is adaptive.
4.3. Image-based metrics
Berkeley Segmentation Database (BSD) [26] is by far the most used to evaluate superpixel methods over image-based metrics. This data set provides 500
color images split into 200 for training, 100 for validation and 200 for test, as
well as the respective human annotated ground-truth. Usually segmentation
performance on BSD is assessed from the widely known metrics BR and UE.
Although our method is geared towards video segmentation, for the sake of
completeness we have included experiments on BSD. The evaluations were run
on the 200 test images, comparing the same algorithms as the previous Sintel
test. The results are presented in Fig. 6.
Figure 6.a illustrates results using BR, which reveals that all the methods
far outperform the BOX in respect to adherence to real image boundaries. Besides, the evaluated methods appear to form two main groups according to their
performance: the best performance group, where the methods reach about 70%
of BR using a maximum of 200 superpixels; and the second best performance
group, where the algorithms need more than twice that number of superpixels
19
Figure 5: Visual comparison of superpixels generated on Sintel images. The first row depicts
the ground-truth optical flow, according to the gradient map
, where the color and intensity
represent, respectively, the orientation and magnitude. For each method named on the left
board, the rows show the generated superpixels (top row) and the MUSE (bottom row), which
is represented in heat map, according to the scale showed at the bottom of the Figure.
20
Figure 6: Image-based evaluation results on BSD data set. (a) BR measures the matching
between superpixel boundaries and the ground-truth, which is comprised of human-annotated
object contours. The plot shows that ISEC is in line with the group formed by the methods
with better performance regarding BR, only being considerably surpassed after the other
methods get an excessive number of superpixels. (b) UE penalizes superpixels that overlaps
more than one object in the image; methods that try to avoid the over-segmentation of
homogeneous regions in the image, such as ISEC and FH, are prone to be further penalized
by this metric.
to reach the same performance. ISEC is in line with the former group, only
being considerably surpassed after the other methods get an excessive number of superpixels. A visual comparison of superpixel segmentations on BSD
data set is showed in Fig. 7. Besides ISEC, the the visual comparison includes
segmentations provided by SLIC and LSC, whose the number of superpixels
is user-defined, and FH and EAMS, whose that number is automatically controlled.
Regarding the UE (see Fig. 6.b), although ISEC performance is always better than the BOX, it is not as good as most of the other methods, except for FH.
That is the drawback of trying to produce a segmentation that is fully supported
by the image content. Despite the proposed method presents strategies to deal
with eventual lacks of information (see ISEC edge reconstruction in Section 3),
eventually an object with a too long and smooth contour can lead to issues.
That is why recent superpixel methods rely on grid-like segmentation strate-
21
Figure 7: Visual comparison of superpixels generated on BSD images. For SLIC and LSC,
whose the number k of superpixel is user-defined, the presented examples were generated with
k = 500. For FH, EAMS and ISEC, whose k is adaptive, the examples were generated by
adjusting the internal parameters of these methods in order to provide k ≈ 500 (in average,
considering the entire BSD data set).
gies. However, in the case of video segmentation, sometimes it is worth dealing
with a higher UE in favor of avoiding to segment image regions unnecessarily,
as well as to obtain more stable superpixels.
Finally, results of computational time on BSD are illustrated in Fig. 8.
The plots show that our method is one of the fastest. By order: WS and CWS
come first, taking around 10 ms per image; followed by FH and SLIC (≈ 100
22
Figure 8: Computational time. WS and CWS are the fastest methods, taking less than
10 ms per image; FH, SLIC, SEEDS, ISEC and LSC come next, spending around a few
hundred milliseconds; ERS, EAMS, QS, VEK spend more than a second; and NC is not
linear, taking several minutes to generate the superpixels even for resized images. Among the
fastest methods, ISEC is the only one implemented in Matlab, while the other methods are
implemented in C/C++.
ms), SEEDS and ISEC (≈200 ms), and LSC (≈500 ms). The remaining methods
spend over a second per image, drawing attention the scalable time consumption
of NC (hundreds of seconds), even for resized images. Also note that, while all
the fastest methods were implemented in C/C++, ISEC was coded in Matlab,
taking room for speed-ups.
4.4. Application on video object segmentation
This evaluation is based on a framework for video object segmentation proposed in [27]. This framework is divided into two main steps: (i) optical flow is
used to estimate an initial foreground segmentation based on motion, and (ii)
appearance cues are extracted from the segments, relying on superpixels to obtain refined object contours. Since the superpixel generation is an independent
module in this video object segmentation framework, it is possible to evaluate
the impact of using different methods for superpixel generation.
Tests were run on the SegTrack data set [28]. This data set provides 6 video
sequences (monkeydog, girl, birdfall, parachute, cheetah and penguin), along
23
SegTrack Videos
Method
Monkey
Average
Total
Parachute
error
# Superp.
Time (s)
Birdfall
Cheetah
FH
-
1462
-
-
764
1113
49
12.6
EAMS
-
915
2536
328
19062
5710
324
274.9
345
902
2954
402
452
1011
1500
20.7
SLIC
Girl
Average
ERS
288
873
1802
1439
305
941
700
196.0
LSC
217
913
1861
350
273
723
500
62.2
ISEC
268
973
2324
390
407
872
324
32.9
Table 1: Results on the video object segmentation experiment. The columns show: (1) the
evaluated methods; (2-6) the average number of mislabeled pixels for each SegTrack video;
(7) the average error, given by the total of mislabeling divided by the number of frames; (8)
the average number of superpixels generated for each frame; and (9) the total time spend
to segment the entire dataset. The missing data (columns 2, 4, 5 for FH, and column 2 for
EAMS) are cases where the video object segmentation failed to converge due to degenerate
superpixels. The results indicate LSC and ISEC as the most efficient methods: low error rate
with a small number of superpixels. Regarding the processing time, ISEC is twice as fast as
LSC.
with the object annotation for each frame. Following [27], the penguin video
has been discarded due to inconsistencies in the ground-truth, remaining a total
of 202 frames. Table 1 presents the results of using ISEC, as well as other five
representative methods selected from experiments on Sintel and BSD. Table
columns summarizes: (1) the evaluated methods; (2-6) the average number of
mislabeled pixels for each SegTrack video; (7) the average error, given by the
total of mislabeling divided by the number of frames; (8) the average number of
superpixels generated for each frame; and (9) the total time spend to segment
the entire dataset.
The missing data (columns 2, 4, 5 for FH, and column 2 for EAMS) are
cases where the video object segmentation failed to converge due to degenerate
superpixels. In the case of methods based on a user-defined number of superpixels (SLIC, ERS and LSC), the tests were started with a small number of
superpixels, being this number increased until achieving convergence. Results
24
Figure 9: Visual comparison of object segmentation on SegTrack sequences. From the top
to the bottom, the rows show, for each method, three distinct shots of the video sequences
named birdfall, cheetah, girl, monkey and parachute. In green, is the object segmentation
achieved by [27] when using the different superpixel methods, SLIC, LSC and ISEC.
show that despite FH is very fast and generates a small number of segments,
this method is not able to support video segmentation in most of the cases. In
the cases where FH was successful, this method presented the second worst average error. EAMS, in turn, besides the highest error rate, is by far the slowest
method, spending 274 seconds to segment all the videos. Intermediate results
were reached by SLIC and ERS. While SLIC needs more than twice the number
of superpixels than the other methods to converge, ERS is also too slow (the
second in computational time). The best results were achieved by LSC and
ISEC. The former presents low error rate (only 723 pixels in average) with few
generated superpixels (500 per frame), being moderate in terms of speed. ISEC
gets low error (872 pixels in average, the second best), generating fewer superpixels per image (324) in average, being twice as fast as LSC (around 33 against
62 seconds). Figure 9 compares some examples of object segmentation provided
by the video segmentation framework when using SLIC, LSC and ISEC.
25
5. Conclusions
We introduced this paper with a discussion about whether is convenient or
not, for a superpixel method, to rely on a fixed number of superpixels. Our hypothesis was pointed in the direction that automatically controlling this number
could be suitable for some tasks, specially regarding video segmentation. To test
that hypothesis, we compared the proposed method with several state-of-the-art
methods through a set of experiments designed to evaluate different aspects of
superpixel segmentation. The evaluation using video-based metrics reinforced
the idea that ISEC approach is suitable for video segmentation; the proposed
method was the best in providing accurate representation of motion discontinuities, while preserving segmentation stability in video sequences. Indeed, these
results were corroborated by another experiment, this time involving an application for video object segmentation. By this latter experiment, ISEC proved
to be one of the most efficient methods, leading the application to accomplish
the second best segmentation performance, with the smallest number of superpixels and spending half of the time of the superpixel method that presented
the best performance in this experiment. On the other hand, the experiment
using image-based metrics indicated that approaches based on a fixed number
of superpixels are the better choice for static image segmentation, since these
types of methods are able to provide high boundary adherence, while preventing
undersegmentation failures.
References
[1] F.
ing,
Yang,
IEEE
H.
Lu,
M.
H.
Yang,
Robust
superpixel
Transactions
on
Image
Processing
(2014)
track1639–
1651doi:10.1109/TIP.2014.2300823.
[2] A. Bdis-Szomor, H. Riemenschneider, L. V. Gool, Superpixel meshes for
fast edge-preserving surface reconstruction, in: IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 2011–2020.
doi:10.1109/CVPR.2015.7298812.
26
[3] R. Gadde, V. Jampani, M. Kiefel, D. Kappler, P. Gehler, Superpixel convolutional networks using bilateral inceptions, in: European Conference on
Computer Vision (ECCV), 2016. doi:10.1007/978-3-319-46448-0_36.
[4] M. Santos, M. Linder, L. Schnitman, U. Nunes, L. Oliveira, Learning to segment roads for traffic analysis in urban images, in: IEEE Intelligent Vehicles
Symposium (IV), 2013, pp. 527–532. doi:10.1109/IVS.2013.6629521.
[5] B. P. Nguyen, H. Heemskerk, P. T. C. So, L. Tucker-Kellogg, Superpixelbased segmentation of muscle fibers in multi-channel microscopy, BMC
Systems Biology (2016) 124doi:10.1186/s12918-016-0372-2.
[6] Z. Tian, L. Liu, Z. Zhang, B. Fei, Superpixel-based segmentation for 3d
prostate mr images, IEEE Transactions on Medical Imaging (T-MI) (2016)
791–801doi:10.1109/TMI.2015.2496296.
[7] D. Stutz, A. Hermans, B. Leibe, Superpixels: An evaluation of the
state-of-the-art, Computing Research Repository (CoRR) abs/1612.01601.
doi:10.1016/j.cviu.2017.03.007.
[8] M. Wang, X. Liu, Y. Gao, X. Ma, N. Q. Soomro, Superpixel segmentation: A benchmark, Signal Processing: Image Communication (2017) 28 –
39doi:10.1016/j.image.2017.04.007.
[9] F. Meyer, Color image segmentation, in: International Conference on Image
Processing and its Applications, 1992, pp. 303–306.
[10] J. Shi, J. Malik, Normalized cuts and image segmentation, IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2000) 888–
905doi:10.1109/34.868688.
[11] P. Meer, B. Georgescu, Edge detection with embedded confidence, IEEE
Transactions on Pattern Analysis and Machine Intelligence (PAMI) 23
(2001) 1351–1365. doi:10.1109/34.977560.
27
[12] P. F. Felzenszwalb, D. P. Huttenlocher, Efficient graph-based image segmentation, International Journal of Computer Vision (IJCV) 59 (2) (2004)
167–181. doi:10.1023/B:VISI.0000022288.19776.77.
[13] A. Vedaldi, S. Soatto, Quick Shift and Kernel Methods for Mode Seeking,
2008, pp. 705–718. doi:10.1007/978-3-540-88693-8_52.
[14] O. Veksler, Y. Boykov, P. Mehrani, Superpixels and supervoxels in
an energy optimization framework, in:
Proceedings of the Euro-
pean Conference on Computer Vision (ECCV), 2010, pp. 211–224.
doi:10.1007/978-3-642-15555-0_16.
[15] R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, S. Ssstrunk, Slic superpixels compared to state-of-the-art superpixel methods, IEEE Transactions
on Pattern Analysis and Machine Intelligence (PAMI) 34 (11) (2012) 2274–
2282. doi:10.1109/TPAMI.2012.120.
[16] M. Y. Liu, O. Tuzel, S. Ramalingam, R. Chellappa, Entropy rate superpixel
segmentation, in: Conference on Computer Vision and Pattern Recognition
(CVPR), 2011, pp. 2097–2104. doi:10.1109/CVPR.2011.5995323.
[17] M. Van den Bergh, X. Boix, G. Roig, B. de Capitani, L. Van Gool,
Seeds: Superpixels extracted via energy-driven sampling, in: Proceedings
of the European Conference on Computer Vision (ECCV), 2012, pp. 13–26.
doi:10.1007/978-3-642-33786-4_2.
[18] P. Neubert, P. Protzel, Compact watershed and preemptive slic: On
improving trade-offs of superpixel segmentation algorithms, in: International Conference on Pattern Recognition (ICPR), 2014, pp. 996–1001.
doi:10.1109/ICPR.2014.181.
[19] Z. Li, J. Chen, Superpixel segmentation using linear spectral clustering, in:
IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
2015, pp. 1356–1363. doi:10.1109/CVPR.2015.7298741.
28
[20] S. Gould,
J. Zhao,
X. He,
Y. Zhang,
Superpixel Graph La-
bel Transfer with Learned Distance Metric,
2014,
pp. 632–647.
doi:10.1007/978-3-319-10590-1_41.
[21] R. Giraud, V. T. Ta, A. Bugeau, P. Coup, N. Papadakis, Superpatchmatch:
An algorithm for robust correspondences using super-
pixel patches, IEEE Transactions on Image Processing (2017) 4068–
4078doi:10.1109/TIP.2017.2708504.
[22] S. Ji, B. Wei, Z. Yu, G. Yang, Y. Yin, A new multistage medical segmentation method based on superpixel and fuzzy clustering, Computational and
Mathematical Methods in Medicine (CMMM)doi:10.1155/2014/747549.
[23] P. Neubert, P. Protzel, Evaluating superpixels in video: Metrics beyond figure-ground segmentation, in: British Machine Vision Conference
(BMVC), 2013. doi:10.5244/C.27.54.
[24] A. Khoreva, R. Benenson, F. Galasso, M. Hein, B. Schiele, Improved
Image Boundaries for Better Video Segmentation, 2016, pp. 773–788.
doi:10.1007/978-3-319-49409-8_64.
[25] D. J. Butler, J. Wulff,
G. B. Stanley, M. J. Black, A natu-
ralistic open source movie for optical flow evaluation, in:
Euro-
pean Conference on Computer Vision (ECCV), 2012, pp. 611–625.
doi:10.1007/978-3-642-33783-3_44.
[26] D. Martin, C. Fowlkes, D. Tal, J. Malik, A database of human segmented
natural images and its application to evaluating segmentation algorithms
and measuring ecological statistics, in: Proceedings of the IEEE International Conference on Computer Vision (ICCV), Vol. 2, 2001, pp. 416–423.
doi:10.1109/ICCV.2001.937655.
[27] A. Papazoglou, V. Ferrari, Fast object segmentation in unconstrained
video, in: Proceedings of the IEEE International Conference on Computer
Vision (ICCV), 2013, pp. 1777–1784. doi:10.1109/ICCV.2013.223.
29
[28] D. Tsai, M. Flagg, J. Rehg, Motion coherent tracking with multi-label mrf
optimization, in: Proceedings of the British Machine Vision Conference
(BMVC), 2010, pp. 56.1–56.11. doi:10.5244/C.24.56.
30
| 1cs.CV
|
A Generalization of Permanent Inequalities and
Applications in Counting and Optimization
Nima Anari1 and Shayan Oveis Gharan∗2
arXiv:1702.02937v1 [cs.DS] 9 Feb 2017
1
2
Stanford University, [email protected]
University of Washington, [email protected]
Abstract
A polynomial p ∈ R[z1 , . . . , zn ] is real stable if it has no roots in the upper-half complex plane.
Gurvits’s permanent inequality gives a lower bound on the coefficient of the z1 z2 . . . zn monomial
of a real stable polynomial p with nonnegative coefficients. This fundamental inequality has been
used to attack several counting and optimization problems.
Here, we study a more general question: Given a stable multilinear polynomial p with nonnegative coefficients and a set of monomials S, we show that if the polynomial obtained by
summing up all monomials in S is real stable, then we can lowerbound the sum of coefficients
of monomials of p that are in S. We also prove generalizations of this theorem to (real stable)
polynomials that are not multilinear. We use our theorem to give a new proof of Schrijver’s
inequality on the number of perfect matchings of a regular bipartite graph, generalize a recent result of Nikolov and Singh [NS16], and give deterministic polynomial time approximation
algorithms for several counting problems.
∗
This material is based on work supported by the National Science Foundation Career award.
1
Introduction
Suppose we are given a polynomial p ∈ R[z1 , . . . , zn ] with nonnegative
For a vector
Qn coefficients.
κ
n
κ ∈ Z+ we write Cp (κ) to denote the coefficient of the monomial i=1 z in p. Given a set S of
κ’s, we want to estimate
X
Cp (κ).
κ∈S
Let ∂zi be the operator that performs partial differentiation in zi , i.e., ∂/∂zi . We can rewrite the
above quantity using a differential operator. Let
q(∂z) =
n
XY
∂zκii
κ∈S i=1
Then,
X
κ!Cp (κ) = q(∂z)p(z)|z=0 .
κ∈S
More generally, given two polynomials p, q ∈ R[z1 , . . . , zn ] with nonnegative coefficients we analytically and algorithmically study the quantity
X
q(∂z)p(z)|z=0 = p(∂z)q(z)|z=0 =
κ!Cp (κ)Cq (κ).
(1)
κ∈Zn
+
for a large class of polynomials known as real stable polynomials.
Before stating our results, let us motivate this question.
Permanent:
n×n
recall that the permanent of A is
For a matrix A ∈ R+
per(A) =
n
XY
Ai,σ(i) .
σ∈Sn i=1
Then, per(A) is Cp (1) of the polynomial p,
p(z1 , . . . , zn ) =
n
n X
Y
Ai,j zj .
i=1 j=1
Matroid Intersection: Let M1 ([n], I1 ) and M2 ([n], I2 ) be two matroids on elements [n] =
{1, 2, . . . , n} with set of bases B1 , B2 respectively. Let
X Y
zi ,
p(z) =
B∈B1 i∈B
q(∂z) =
X Y
∂zi .
B∈B2 i∈B
Then, p(z)q(∂z)|z=0 is the number of bases in the intersection of M1 and M2 .
We say a polynomial p(z1 , . . . , zn ) ∈ R[z1 , . . . , zn ] is real stable if it has no roots in the upper
half complex plane. Theory of stable polynomials recently had many applications in multiple
1
areas of mathematics and computer science [Gur06, MSS13, AO15a, AO15b]. Gurvits used this
theory to give a new proof of the van der Waerden conjecture [Gur06]. Most notably he proved
that for any n-homogeneous real stable polynomial p(z1 , . . . , zn ) with nonnegative coefficients and
q = ∂z1 . . . ∂zn ,
p(z1 , . . . , zn )
p(z1 , . . . , zn )
≤ q(∂z)p(z)|z=0 ≤ inf
.
(2)
e−n inf
z>0
z>0
z1 . . . zn
z1 . . . zn
n)
In other words, the coefficient of the monomial z1 . . . zn is at least e−n inf z>0 p(zz11,...,z
...zn . One can
use this inequality to give a deterministic e−n approximation algorithm for the permanent of a
nonnegative matrix.
In this paper we prove bounds analogous to (2) for the quantity p(z)q(∂z) when p, q are real
stable polynomials. We discuss several applications in counting and optimization. We expect to
see many more applications in the future.
1.1
Our Contributions
For two vectors x, y ∈ Rn we write xy to denoteQ
the n dimensional vector where the i-th coordinate
is xi yi . For vectors x, y ∈ Rn we write xy := di=1 xyi i . We say a polynomial p ∈ R[z1 , . . . , zn ] is
multilinear if Cp (κ) = 0 for κ ∈
/ {0, 1}n . For a vector α ∈ Rn , we write α ≥ 0 to denote α ∈ Rn+ .
Theorem 1.1. For any two real stable polynomials p, q ∈ R[z1 , . . . , zn ] with nonnegative coefficients,
sup inf e−α
α≥0 y,z>0
p(y)q(z)
p(y)q(z)
≤ q(∂z)p(z)|z=0 ≤ sup inf
.
α
α
(yz/α)
α≥0 y,z>0 (yz/α)
(3)
Furthermore, if p, q are multilinear then
sup inf (1 − α)1−α
α≥0 y,z>0
p(y)q(z)
≤ q(∂z)p(z)|z=0 .
(yz/α)α
(4)
Note that when kαk∞ > 1, the term (1 − α)1−α is ill-defined in the above theorem. However
in that case we can replace it by any number, since by Fact 2.18, the LHS will be zero.
Before describing our main applications of the above theorem let us derive Gurvits’s inequality
(2). Let q(z) = z1 z2 . . . zn . Then, it follows that the supremum is obtained by α = 1. For that α
we have
p(y)
p(y)q(z)
= inf
.
inf
y>0 y1 . . . yn
y,z>0 (yz/α)α
In Section 3 we discuss several applications of the above theorem. The following two theorems are
the main consequences that we prove. Firstly, we give an algorithm to approximate q(∂z)p(z)|z=0
for two given stable polynomials p, q with nonnegative coefficients. We use this to give deterministic
approximate counting algorithms. In our results, we use convex programs involving certain polynomials. We make sure these convex programs can be approximately solved in time polynomial in the
complexity of the involved polynomials. For a polynomial p ∈ R[z1 , . . . , zn ] with nonnegative coefficients, we define its complexity as n + deg p + | log minκ:Cp (κ)6=0 Cp (κ)| + | log maxκ:Cp (κ)6=0 Cp (κ)|,
and we denote this by hpi.
Theorem 1.2. Given two real stable polynomials p, q ∈ R[z1 , . . . , zn ] with nonnegative coefficients
and an oracle that for any z ∈ Rn , returns p(z), q(z), there is a polynomial time algorithm that
returns an emin{deg p,deg q}+ǫ approximation of q(∂z)p(z)|z=0 that runs in time poly(hpi + hqi +
log(1/ǫ)).
2
Our next theorem gives a polynomial time algorithm to approximate maxκ∈{0,1}n Cp (κ)Cq (κ)
for two given multilinear stable polynomials p, q with nonnegative coefficients. We use this to
generalize a recent results of Nikolov and Singh [NS16].
Theorem 1.3. Given two real stable multilinear D-homogeneous polynomials p, q ∈ R[z1 , . . . , zn ]
with nonnegative coefficients and an oracle that for any z ∈ Rn returns p(z), q(z), there is a
polynomial time algorithm that outputs an e2D+ǫ approximation to
max Cp (κ)Cq (κ).
κ∈{0,1}n
that runs in time poly(hpi + hqi + log(1/ǫ)).
Independent of our work, Straszak and Vishnoi [SV16] studied several variants of our problem.
They show that if p, q ∈ R[z1 , . . . , zn ] are multilinear real stable polynomials with nonnegative
coefficients, there is a convex relaxation to the quantity q(∂z)p(z) with an integrality gap at most en .
In contrast, our approximation factor depends only on min{deg p, deg q} which can be significantly
smaller than n.
Structure of the paper. In Section 2 we describe many properties of real stable polynomials
that we use throughout the paper. In Section 3 we describe several applications of Theorem 1.1
in counting and optimization. Finally, in sections 4 and 5 we prove Theorem 1.1. In particular,
in Section 4 we lower bound q(∂z)p(z)|z=0 , i.e., we prove the (4) and the LHS of (3). And, in
Section 5 we prove upper bound this quantity, i.e., the RHS of (3).
2
Preliminaries
Throughout the paper, we use bold letters to denote vectors. For a vector y, we write y ≤ 1
to denote that all coordinates of y are at most 1. For two vectors x, y ∈ Rn we define xy =
(x1 y1 , . . . , xn yn ). Similarly, we define x/y = (x1 /y1 , . . . , xn /yn ). For a vector y ∈ Rn , we define
exp(x) := (ex1 , . . . , exn ).
For vectors x, y ∈ Rn we write
d
Y
y
xyi i .
x :=
i=1
Q
As a special case, for a real number c ∈ R we write cx to denote ni=1 cxi .
We use R++ = {x : x > 0} to denote the set of positive real numbers. For
an integer n ≥ 1 we
use [n] to denote the set of numbers {1, 2, . . . , n}. For any m, n, we let [m]
denote the collection
n
of subsets of [m] of size n. Throughout the paper all logs are in base e.
For a vector x ∈ Rn and an integer 1 ≤ i ≤ n we use x−i to denote the m − 1 dimensional vector
obtained from x by dropping the i-th coordinate of x. We use Z+ to denote the set of nonnegative
integers.
We say a matrix A ∈ Rn×n is doubly stochastic if all entries of A are nonnegative and all row
sums and columns sums are equal to 1,
X
Ai,j = 1 ∀i,
j
X
Ai,j = 1 ∀j.
i
3
2.1
Stable Polynomials
Stable polynomials are natural multivariate generalizations of real-rooted univariate polynomials. For a complex number z, let Im(z) denote the imaginary part of z. We say a polynomial
p(z1 , . . . , zn ) ∈ C[z1 , . . . , zn ] is stable if whenever Im(zi ) > 0 for all 1 ≤ i ≤ m, p(z1 , . . . , zn ) 6= 0. As
the only exception, we also call the zero polynomial stable. We say p(.) is real stable, if it is stable
and all of its coefficients are real. It is easy to see that any univariate polynomial is real stable if
and only if it is real rooted.
We say a polynomial p(z1 , . . . , zn ) is degree k-homogeneous, or k-homogeneous, if every monomial of p has total degree exactly k. Equivalently, p is k-homogeneous if for all a ∈ R, we have
p(a · z1 , . . . , a · zn ) = ak p(z1 , . . . , zn ).
We say a monomial z1α1 . . . znαn is square-free if α1 , . . . , αn ∈ {0, 1}. We say a polynomial is multilinear if all of its monomials are square-free. For a polynomial p, we write deg p to denote the
maximum degree of all monomials of p. One of the famous examples of real stable polynomials is
the set of elementary symmetric polynomials. For an integer k ≥ 0, the k-th elementary symmetric
polynomial in z1 , . . . , zn , ek (z1 , . . . , zn ) is defined as
X Y
ek (z1 , . . . , zn ) :=
zi .
ℓ∈S
S∈([n]
k )
Note that the above polynomial is k-homogeneous and multilinear.
Let µ be a probability distribution on subsets of [n]. The generating polynomial of µ is defined
as follows:
Y
X
zi .
µ(S)
gµ (z1 , . . . , zn ) =
S
i∈S
We say µ is a strongly Rayleigh (SR) distribution if gµ is a real stable polynomial. Borcea, Brändén,
and Liggett [BBL09] defined (SR) distributions and proved numerous properties of them. Many
interesting family of probability distributions are special cases of SR including product distributions,
random spanning trees, and determinantal point processes. Note that µ(S) ≥ 0 for all S ⊆
[n]; so gµ (S) has nonnegative coefficients. Many of theorems that we prove here have natural
probabilistic interpretations if the underlying real stable polynomial is a generating polynomial of
a SR distribution.
We say a polynomial p(y1 , . . . , yn , z1 , . . . , zn ) is bistable if p(y1 , . . . , yn , −z1 , . . . , −zn ) is a real
stable polynomial.
Fact 2.1. For any set of PSD matrices A1 , A2 , . . . , An ∈ Rd×d and a symmetric matrix B ∈ Rd×d
p(z1 , . . . , zn ) det(B + z1 A1 + · · · + zn An )
is real stable.
Fact 2.2. If p(z) and q(z) are real stable, then p(z)q(z) is real stable.
Fact 2.3. If p(y) and q(z) are real stable, then p(y)q(z) is bistable.
Fact 2.4. For any stable polynomial p(z1 , . . . , zn ), an integer 1 ≤ i ≤ n, and any number c with
nonnegative imaginary part, Im(c) ≥ 0, p(z1 , . . . , zi−1 , c, zi+1 , . . . , zn ) is stable.
4
Fact 2.5. If p(y) is real stable, then for any nonnegative vector λ ∈ Rn+ , p(λy) is real stable.
Lemma 2.6. For any real stable polynomial p(z1 , z2 , . . . , zn ) and any 1 ≤ i, j ≤ n, (1 − ∂zi ∂zj )p is
stable.
Proof. We say an operator T is stability preserver if for any real stable polynomial p(z), T p is real
stable. So, to prove the lemma we need to show that 1 − ∂zi ∂zj is a stability
operator.
P preserver
κ
Borcea and Brändén [BB10] showed that for any differential operator T = κ cκ ∂z , T is stability
preserver if the polynomial
X
cκ (−z)κ
κ∈Zn
+
is real stable. So, to prove the lemma it is enough to show that 1 − zi zj is real stable. Obviously,
if Im(zi ), Im(zj ) > 0, then zi zj 6= 1. So, 1 − zi zj is real stable.
Brändén [Brä07] proved the following characterization of multilinear stable polynomials.
Theorem 2.7 (Brändén [Brä07, Theorem 5.6]). For any multilinear polynomial p(y1 , . . . , ym ), p
is stable if and only if for all 1 ≤ i < j ≤ m,
∂yi p|y=x · ∂yj p|y=x ≥ p(x) · ∂yi ∂yj p|y=x ,
for all x ∈ Rm .
2.2
Polarization
Let p ∈ C[z1 , . . . , zn ] be a stable polynomial of degree di in zi , 1 ≤ i ≤ n. The polarization is an
operator that turns p into a multilinear “symmetric” polynomial.
For any m ≥ maxi di , The polarization πm (p) of p is a polynomial in m · n variables zi,j for
1 ≤ i ≤ n, 1 ≤ j ≤ m defined as follows: For all 1 ≤ i ≤ n, and 1 ≤ j ≤ di substitute any occurrence
of zik in p with m1 ek (zi,1 , . . . , zi,m ).
(k)
For example, if p(z1 , z2 ) = z12 + z1 z2 , then
1
1
π2 (p) = z1,1 z1,2 + (z1,1 + z1,2 ) · (z2,1 + z2,2 ).
2
2
Note that for any m and any polynomial p, πm (p) has the following properties:
• πm (p) is multilinear;
• πm (p) is symmetric in the variables zi,1 , . . . , zi,m for all 1 ≤ i ≤ n;
• if we substitute zi,j = zi for all i, j in πm (p) we recover p.
Borcea, Brändén, and Liggett used Grace-Walsh-Szego’s coincidence theorem [Gra02, Sze22, Wal22]
to prove that stability is preserved under polarization [BBL09, Thm 4.7].
Theorem 2.8 (Borcea, Brändén, Liggett [BBL09]). If p ∈ C[z1 , . . . , zn ] is stable then πm (p) is
stable for all m.
5
2.3
Stability and Convexity
Polynomial optimization problems involving real stable polynomials with nonnegative coefficients
can often be turned into concave/convex programs. Such polynomials are log-concave in the positive
orthant:
Theorem 2.9 ([Gül97]). For any real stable homogeneous polynomial p ∈ R[z1 , . . . , zn ] such that
p(z) > 0 for all z ∈ Rn++ , log p(z) is concave in the positive orthant.
Proof. Firstly, note that log p(z) is well-defined for any z ∈ Rn++ . It is enough to show that the
function is concave along any interval in the positive orthant. Let a ∈ Rn++ , b ∈ Rn , and consider
the line a + tb where for any t ∈ [0, 1], a + tb ∈ Rn++ . We show that log p(a + tb) is concave. Say
p is k-homogeneous, then
p(a + tb) = p(t(a/t + b)) = tk p(a/t + b).
We claim that p(a/t + b) is real rooted. Firstly, since a ∈ Rn++ , and p(.) is stable, p(at + b) is real
rooted. But for any root λ of p(at + b), 1/λ is a root of p(a/t + b). Note that p(at + b) has no
roots at 0, because at + b ∈ Rn++ for t ∈ [0, 1]. So, p(a/t + b) is real rooted.
Let λ1 , . . . , λk be the roots of p(a/t + b). We have
k
k
Y
Y
p(a + tb) = t p(a/t + b) = t p(a) (1/t − λi ) = p(a) (1 − tλi ).
k
k
i=1
i=1
We claim that for all 1 ≤ i ≤ k, λi < 1. Otherwise, for some t ∈ [0, 1], p(a + tb) = 0, but since
a + tb ∈ Rn++ , p(a + tb) > 0 which is a contradiction. Therefore,
log p(a + tb) = log p(a) +
k
X
log(1 − tλi ).
i=1
The theorem follows by the fact that log(1 − tλ) is a concave function of t for t ∈ [0, 1] when
λ < 1.
Corollary 2.10. For any real stable polynomial p ∈ R[z1 , . . . , zn ] with nonnegative coefficients,
log p(z) is concave in positive orthant.
Proof. The homogenization of a polynomial p ∈ R[z1 , . . . , zn ] is defined as the polynomial pH ∈
R[z1 , . . . , zn , zn+1 ] given by
deg p
pH (z1 , . . . , zn+1 ) = zn+1
p(z1 /zn+1 , . . . , zn /zn+1 ).
Borcea, Brändén and Liggett [BBL09, Thm 4.5] show that for a polynomial p with nonnegative
coefficients, p is real stable if and only pH is real stable. Since pH is real stable, by Theorem 2.9,
log pH (z1 , . . . , zn+1 ) is concave in the positive orthant. It follows that pH (z1 , . . . , zn , 1) is also
concave in the positive orthant. But, p(z1 , . . . , zn ) = pH (z1 , . . . , zn , 1). So, p(z) is concave in the
positive orthant.
It is also an immediate corollary of Hölder’s inequality that a polynomial with nonnegative
coefficients is log-convex in terms of the log of its variables (for more details on log-convex functions
see [BV06]).
6
Fact 2.11. For any polynomial p(y1 , . . . , yn ) with nonnegative coefficients, the expression
log p(exp(y)) = log p(ey1 , . . . , eyn )
is convex as a function of y.
Lemma 2.12. For any polynomial p with nonnegative coefficients, and any α ≥ 0,
inf
z>0
p(z)
p(αz)
= inf
.
z>0 zα
(z/α)α
Proof. If α > 0, the two sides become equal by a change of variable z ↔ αz. If αi = 0 for some i,
since p has nonnegative coefficients, it is obvious that one should let zi → 0 on the LHS, since the
numerator decreases monotonically as zi → 0, and the denominator does not change. In the limit
the LHS is the same as what we have on the RHS.
2.4
Jump Systems and Newton Polytope
Definition 2.13. For a polynomial p(y1 , . . . , yn ), we define supp(p) as
supp(p) := {κ ∈ Zn+ : Cp (κ) 6= 0}.
Brändén [Brä07] proved that if p is a real stable polynomial then supp(p) forms what is known
as a jump system.
Definition 2.14. For two points x, y ∈ Zn define an (x, y)-step to be any δ ∈ Zn such that
kδk1 = 1 and ky − (x + δ)k1 = ky − xk1 − 1. A set F ⊆ Zn is called a jump system iff for
every x, y ∈ F and (x, y)-step δ, either x + δ ∈ F or there exists an (x + δ, y)-step δ′ such that
x + δ + δ′ ∈ F.
Brändén [Brä07] proved that for any real stable polynomial p, supp(p) is a jump system. We
use the following immediate corollary which follows from the fact that for any polynomial p ∈
R[y1 , . . . , yn , z1 , . . . , zn ], supp(p(y, z)) = supp(p(y, −z)).
Fact 2.15. If the polynomial p(y, z) is bistable, then supp(p) is a jump system.
Lovász showed that the natural generalization of the greedy algorithm for maximizing linear
functions over matroids, also works for jump systems [Lov97]. Given a finite jump system F ⊆ Zn ,
and w ∈ Rn , the following algorithm finds maxx∈F hw, xi: Sort the coordinates in such a way that
|w1 | ≥ |w2 | ≥ · · · ≥ |wn |. Fix the coordinates of x, one by one; for i = 1, . . . , n, we will fix xi as
follows: among all members of F whose first i − 1 coordinates match x1 , . . . , xi−1 , let xi be largest
i-th coordinate if wi > 0 and let xi be the smallest otherwise.
Theorem 2.16 ([Lov97]). For any finite jump system F ⊆ Zn and w ∈ Rn , the above greedy
algorithm correctly finds a point x maximizing hw, xi.
We will be working with the convex hull of supp(p), often known as the Newton polytope.
Definition 2.17. For a polynomial p(y1 , . . . , yn ), define the Newton polytope of p as
newt(p) := conv(supp(p)) = conv{κ ∈ Zn+ : Cp (κ) 6= 0},
where conv denotes convex hull.
7
We use the following fact in multiple places in the proofs.
Fact 2.18. For a polynomial p(y) ∈ R[y1 , . . . , yn ] with nonnegative coefficients, and α ∈ Rn+ we
have
p(y)
> 0 ⇐⇒ α ∈ newt(p).
inf
y>0 yα
P
Proof. Let α ∈ newt(p); then we can write α as a convex combination of the vertices: α = i λi κi ,
where λi ’s are nonnegative and sum up to 1, and for all i, Cp (κi ) > 0. Therefore,
p(y) ≥
X
i
Cp (κi )y
κi
=
X
i
Y
Cp (κi )yκi
≥
λi
λi
i
Cp (κi )yκi
λi
λi
=y
α
Y Cp (κi ) λi
i
λi
,
where for the last inequality we used weighted AM-GM. This implies that
p(y) Y
inf α ≥
y>0 y
i
Cp (κi )
λi
λi
> 0.
Conversely, suppose α ∈
/ newt(p); we show that inf y>0 p(y)/yα = 0. Since newt(p) is convex, by
separating hyperplane theorem, there exists c ∈ Rn , such that hc, αi > hc, κi for every κ ∈ newt(p).
Now let y = exp(tc) = (etc1 , . . . , etcn ), where t > 0 will be fixed later. Then we have
P
X
Cp (κ)yκ
κ∈Zn
p(y)
+
=
≤ exp −t hc, αi − max hc, κi
·
Cp (κ).
yα
yα
κ∈newt(p))
n
κ∈Z+
It is easy to see that as t → ∞, the RHS of the above inequality becomes zero.
The convex hull of jump systems has been shown to be bisubmodular polyhedra in [BC95]. The
following is a restatement of this result.
Theorem 2.19 (Restatement of [BC95]). For a finite jump system F ⊆ Zn , define the function
f : {−1, 0, +1}n → R as
f (c) := maxhc, αi.
α∈F
Then
conv(F) =
\
{x ∈ Rn : hc, xi ≤ f (c)}.
c∈{−1,0,+1}n
Furthermore, f is a bisubmodular function.
For the definition of bisubmodular functions, see [BC95]. We will only use the following wellknown fact about bisubmodular functions. For more details see [MF08].
Fact 2.20. Suppose that F is a jump system and f is the bisubmodular function associated with
conv(F). Given an oracle that returns f (c) for any c ∈ {−1, 0, +1}n , we can construct a separation
oracle for conv(F). Each call to the separation oracle uses at most poly(n) evaluations of f .
8
Proof. To construct a separation oracle, it is enough to design a procedure that given a point
x ∈ Rn , finds a c ∈ {−1, 0, +1}n such that hc, xi > f (c) or declares that there is no such c.
The function c 7→ f (c) − hc, xi is bisubmodular, since f is bisubmodular and c 7→ hc, xi is
modular. Bisubmodular functions can be minimized in strongly polynomial time, see e.g. [MF08].
Therefore one can solve the following optimization problem using polynomially many evaluations
of f .
min
f (c) − hc, xi.
c∈{−1,0,+1}n
If the answer is nonnegative, then x ∈ conv(F). Otherwise, the optimizing c produces the
separating hyperplane.
3
Applications
3.1
Applications in Counting
In this section we prove Theorem 1.2 and we discuss several applications of our results for counting
problems.
We start with Schrijver’s inequality. The following theorem is the main technical result of
[Sch98].
Theorem 3.1 (Schrijver [Sch98]). Let A ∈ Rn×n be a doubly stochastic matrix, and let à be defined
as follows:
Ãi,j = Ai,j (1 − Ai,j )
Then,
per(Ã) ≥
Y
(1 − Ai,j ).
i,j
Schrijver [Sch98] used the above inequality to prove that any d-regular bipartite graph with 2n
d−1 n
matchings. Gurvits and [Gur11, GS14] used the above theorem to design
vertices has (d−1)
dd−2
n
a deterministic 2 -approximation algorithm for the permanent. We note that Gurvits used the
machinery of stable polynomials to prove several lower bounds on permanent, but to the best of
our knowledge his machinery could not capture Schrijver’s theorem. Very recently, Lelarge [Lel17]
used the machinery developed by Csikvári [Csi14] to reprove Schrijver’s theorem.
Here, we give a simple proof of this theorem using Theorem 1.1.
Proof. Let z = {zi,j }1≤i,j≤n be a vector of n2 variables. Define,
p(z) =
q(z) =
n X
n
Y
i=1 j=1
n
n X
Y
Ai,j zi,j ,
(1 − Ai,j )zi,j .
j=1 i=1
Then, it is not hard to see that
q(∂z)p(z) = per(Ã).
9
Qn
This follows because each monomial of
p
is
a
product
i=1 zi,σ(i) for some mapping σ : [n] → [n]
Qn
and each monomial of q is a product j=1 zσ(j),j for some σ : [n] → [n].
Let αi,j = Ai,j for all 1 ≤ i, j ≤ n. Since p, q are multilinear, by Theorem 1.1,
p(y)q(z)
.
y,z>0 (yz/α)α
per(Ã) ≥ (1 − α)1−α inf
(5)
Firstly, observe that by weighted AM-GM inequality
p(y) =
n
n X
Y
Ai,j yi,j ≥
i=1 j=1
Y
A
yi,ji,j = yα ,
i,j
where in the inequality we used that A is doubly stochastic. Secondly, by another application of
AM-GM inequality we have
n
n X
Y
Y (1 − Ai,j )zi,j Ai,j
(1 − Ai,j )zi,j
Ai,j
q(z) =
= ((1 − α)z/α)α ,
≥
Ai,j
Ai,j
j=1 i=1
i,j
where in the inequality we used that A is doubly stochastic. Plugging the above two inequalities
in (5) we get
Y
per(Ã) ≥ (1 − α)1−α (1 − α)α =
(1 − Ai,j )
i,j
as desired.
Theorem 1.2. Given two real stable polynomials p, q ∈ R[z1 , . . . , zn ] with nonnegative coefficients
and an oracle that for any z ∈ Rn , returns p(z), q(z), there is a polynomial time algorithm that
returns an emin{deg p,deg q}+ǫ approximation of q(∂z)p(z)|z=0 that runs in time poly(hpi + hqi +
log(1/ǫ)).
Proof. Firstly, we show that
p(y)q(z)
p(y)q(z)
≤ q(∂z)p(z)|z=0 ≤ sup inf
α
α
α≥0 y,z>0 (yz/α)
α≥0 y,z>0 (yz/α)
e− min{deg p,deg q} sup inf
(6)
Without loss of generality assume that deg p ≤ deg q. By Theorem 1.1, it is enough to show that
for any α ≥ 0,
p(y)q(z)
p(y)q(z)
e− deg p inf
≤ inf e−α
.
α
y,z>0 (yz/α)
y,z>0
(yz/α)α
P
To prove the above, it is enough to show that if i αi > deg p, then both sides of the above equal
0. Suppose hα, 1i > deg p. Let z = 1 and y = t1 for some parameter t. Then, as t → ∞,
p(y)/yα → 0. This is because the degree of every monomial of p(y) is less than yα . This was a
special case of Fact 2.18. In general if α ∈
/ newt(p) ∩ newt(q), then both sides of the above are
zero.
Using (6), to prove the theorem it is enough to design a polynomial time algorithm that computes
p(y)q(αz)
p(y)q(z)
supα≥0 inf y,z>0 (yz/α)
α . Firstly, by Lemma 2.12 it is enough to solve supα≥0 inf y,z>0
(yz)α . Now,
let us do a change of variable y ↔ exp(y) and z ↔ exp(z). By Fact 2.11 and Corollary 2.10,
log
p(exp(y))q(α exp(z))
(exp(y) exp(z))α
10
is concave in α and convex in y, z. Since the infimum of concave functions is also concave,
exp(z))
inf y,z log p(exp(y))q(α
is a concave function of α. Therefore,
(exp(y) exp(z))α
p(exp(y))q(exp(z))
.
α
α≥0 y,z>0 (exp(y) exp(z))
sup inf
is a convex program, or more precisely a concave-convex saddle point problem. In Subsection 3.3
we show how this convex program can be approximately solved in polynomial time. This gives an
edeg p approximation of q(∂z)p(z)|z=0 as desired.
Let us state some applications of the above result. Our first application is in the context of
recent results on Nash welfare maximization (NWM) problems. Recently the authors together with
Saberi and Singh proved a generalization of Gurvits’s inequality (2) to design an en approximation
for NWM problem [AOSS17]. Given a real stable polynomial p ∈ R[z1 , . . . , zn ] with nonnegative
coefficients and an integer 1 ≤ D ≤ n, the main technical contribution of [AOSS17] is a polynomial
time algorithm that gives an eD approximation to
X Y
∂zi p(z)|z=0 .
i∈S
S∈([n]
D)
This result immediately followsP
from Theorem
1.2. Let q be the D-th elementary symmetric polynoQ
mial on z1 , . . . , zn , i.e., q(z) = S∈([n]) i∈S ∂zi . Recall that all elementary symmetric polynomials
D
are real stable. Since deg q = D, we can approximate the above quantity up to an eD factor by
Theorem 1.2.
Our second application is a partial answer to an open problem of Kulesza and Taskar about
determinantal point processes [KT12]. A determinantal point process (DPP) is a probability distribution µ on subsets of [n] identified by a positive semidefinite ensemble matrix L ∈ Rn×n where
for any S ⊆ [n] we have
P [S] ∝ det(LS,S ),
where LS,S is the square submatrix of L corresponding to rows and columns indexed by S. We
refer interested readers to [KT12] for applications of DPPs in machine learning. For an integer
0 ≤ k ≤ n, and a DPP µ, the truncation of µ to k, µk is called a k-DPP, i.e.,
(
µ(S) if |S| = k
µk (S) ∝
0
otherwise.
Kulesza and Taskar left an open problem to estimate the quantity
X
det(LS,S )p
S⊆[n]
for some p > 1. Their motivation is to study the Hellinger distance between two DPPs.
Borcea, Brändén and Liggett [BBL09] showed that DPPs and k-DPPs are strongly Rayleigh, i.e.,
if µ is a DPP, then gµ (z1 , . . . , zn ) and gµk (z1 , . . . , zn ) are real stable polynomials (see Subsection 2.1).
Let L, L′ be ensemble matrices of two k-DPPs µ, µ′ . Let p = gµk and q = gµ′k . Note that p, q are
11
k-homogeneous, multilinear, and have nonnegative coefficients. Therefore, by the above theorem
we can estimate
X
q(∂z)p(z)|z=0 =
det(LS ) det(L′S )
S⊆[n]
up to a factor of ek . To the best of our knowledge, to this date no approximation algorithm for the
above quantity was known.
3.2
Applications in Optimization
In this section we prove Theorem 1.3. Let us restate the theorem.
Theorem 1.3. Given two real stable multilinear D-homogeneous polynomials p, q ∈ R[z1 , . . . , zn ]
with nonnegative coefficients and an oracle that for any z ∈ Rn returns p(z), q(z), there is a
polynomial time algorithm that outputs an e2D+ǫ approximation to
max Cp (κ)Cq (κ).
κ∈{0,1}n
that runs in time poly(hpi + hqi + log(1/ǫ)).
Proof. We solve the following mathematical program and we show that its optimum solution gives
an e2D -approximation of the optimum.
p(λy)q(z)
α
α,λ≥0 y,z>0 (yz/α)
X
s.t.
λi = D.
sup inf
(7)
i
Firstly, we show that we can solve the above mathematical program. The idea is very similar to
the proof Theorem 1.2. We do a change of variable y ↔ exp(y), z ↔ exp(z) and we push the α in
the denominator to q. So, instead we solve the following program:
sup inf log
α,λ≥0 y,z>0
p(λ exp(y)q(α exp(z))
(exp(y) exp(z))α
exp(y))q(α exp(z))
is convex in y, z and concave in α, λ. So, we can solve the above
As before, log p(λ(exp(y)
exp(z))α
program in polynomial time; for more details see Subsection 3.3.
Secondly, we show that the above program is a relaxation of the problem. Let
κ∗ = argmaxκ∈{0,1}n Cp (κ)Cq (κ).
P
Since every monomial of p has degree exactly D, i κ∗i = D, so λ = κ∗ is a feasible solution. Let
α = κ∗ . Then,
∗
Cp (κ∗ )yκ
p(λy)
≥ inf
= Cp (κ∗ )
inf
y>0
y>0 yα
yκ∗
Similarly,
q(αz)
q(κ∗ z)
q(z)
=
inf
=
inf
= Cq (κ∗ ),
∗
z>0 zα
z>0 zκ
z>0 (z/α)α
inf
12
The first equality follows by Lemma 2.12. Using the above two equations it follows that the optimum
of (7) is at least Cp (κ∗ )Cq (κ∗ ).
Next, we show that the optimum of (7) is within e2D factor of Cp (κ∗ )Cq (κ∗ ). Let λ∗ , α∗ be the
arguments achieving optimum value in the sup in (7)1 . Consider the following probability distribution on κ ∈ {0, 1}n . We sample D numbers X1 , X2 , . . . , XD independently (with replacement)
where for each 1 ≤ i ≤ D we let Xi = j with probability
and κi = 0 otherwise. We claim that
λ∗j
D.
We let κi = 1, if Xj = i for some i
E [Cp (κ)Cq (κ)] ≥ e−2D OPT .
(8)
Define
p̃(z) = p(λ∗ z).
Then, by Fact 2.5, p̃ is real stable. Therefore, by Theorem 1.1,
X
Cp̃ (κ)Cq (κ) ≥ e−α
κ∈{0,1}n
∗
p̃(y)q(z)
p(λ∗ y)q(z)
−D
sup inf
≥ e−D OPT .
∗ ≥ e
∗
α
α
y,z>0 (yz/α )
(yz/α)
α≥0 y,z>0
inf
The second to last inequality follows by (6) and the last inequality follows by the fact that (7) is a
relaxation of OPT.
So, to prove (8) it is enough to show that
X
E [Cp (κ)Cq (κ)] ≥ e−D
Cp̃ (κ)Cq (κ).
(9)
κ∈{0,1}n
Fix any monomial κ such that Cp (κ), Cq (κ) > 0. Then,
Y
Cp̃ (κ) = Cp (κ)
λ∗i .
i:κi =1
On the other hand,
P [κ] = D!
Y
Y λ∗
i
λ∗i ,
≥ e−D
D
i:κi =1
i:κi =1
The D! factor comes from the fact that we can sample each such κ in D! possible ways. Plugging
these two equations in (9) proves it. This completes the proof of Theorem 1.3.
One of the consequences of the above theorem is a recent result of Nikolov and Singh [NS16].
Given a PSD matrix L and a partitioning P1 , . . . , Pm of [n], among all sets S ⊆ [n] where for each
1 ≤ i ≤ m, |S ∩ Pi | = bi we want to choose one that maximizes det(LS,S ) where LS,S is the square
submatrix of L with rows and columns indexed by S. We refer interested readers to [NS16] for
applicationsP
of this problem.
Let k = i bi . Let v1 , . . . , vn be the vectors of the Cholesky decomposition of L, i.e., for all i, j,
Li,j = hvi , vj i. Let
p(y1 , . . . , yn ) = ∂xn−k det(xI + y1 v1 v1T + · · · + yn vn vnT )|x=0 .
1
Note that since we can assume the domain of α, λ is compact, the supremum is attained at some point of the
domain
13
By Fact 2.1 and Fact 2.4 the above polynomial is real-stable. Furthermore, it is not hard to see
that it is k-homogeneous and has nonnegative coefficients. In fact for any set S ⊆ [n] of size k, the
coefficient of the monomial z1S is det(LS,S ), where 1S is the indicator vector of the set S. Now, let
q(z1 , . . . , zn ) =
m
Y
ebi ({zj }j∈Pi )
i=1
be the generating polynomial corresponding to the partition matroid P1 , . . . , Pk . Again, by closure of stable polynomials under product, the above polynomial is real stable, multilinear, khomogeneous, and has nonnegative coefficients. It follows by the above theorem that we can
approximate
max
det(LS,S )
S:|S∩Pi |=bi ∀i
within an e2k factor. Note that our approximation factor is slightly worse but our theorem is
significantly more general. In particular, we can replace p, q in the above argument with any real
stable multilinear polynomial.
For example, suppose we are given a graph G = (V, E) and we have assigned a vector ve to any
edge e. We can let q be the generating polynomial of the uniform spanning tree distribution of G.
Then, using the above theorem we can approximate
X
ve veT
max det
T
e∈T
within an e2n factor.
3.3
Solving the Convex Program
In this section we show how the convex programs from the previous two sections can be approximately solved in polynomial time. We measure the complexity of our algorithm in terms of the
complexity of the involved polynomials. Recall the following definition.
Definition 3.2. For a polynomial p ∈ R[z1 , . . . , zn ] with nonnegative coefficients, we define its
complexity as n + deg p + | log minκ:Cp (κ)6=0 Cp (κ)| + | log maxκ:Cp (κ)6=0 Cp (κ)|, and we denote this
by hpi.
The main result of this section is the following.
Theorem 3.3. Given a bistable polynomial p ∈ R[y1 , . . . , yn , z1 , . . . , zn ] with nonnegative coefficients and an oracle that evaluates p at any requested point (y, z) ∈ R2n
+ , one can find a 1 + ǫ
approximation of
p(y, z)
,
(10)
sup inf
α
α≥0 y,z>0 (yz/α)
in time poly(hpi + log(1/ǫ)).
In the rest of this section, we prove this theorem. We will use the ellipsoid method to compute
the inner inf and also the outer sup. The main difficulty for computing the inf is that the variables
y, z are unbounded. In other words, we do not have a bounded outer ellipse for the ellipsoid method.
In the first step, we use the fact that supp(p) is a jump system (Fact 2.15) to show that we can
14
work with bounded y, z and only lose a small amount in the objective. The rest of the argument
consists of constructing separation oracles and lower bounding the volume of inner ellipses for the
ellipsoid methods.
Lemma 3.4. Given a polynomial p ∈ R[z1 , . . . , zn ] with nonnegative coefficients with the guarantee
that supp(p) is a jump system and α ∈ newt(p) and ǫ > 0, we have
inf
z:klog(z)k∞ <M
p(z)
p(z)
≤ (1 + ǫ) inf α ,
z>0 z
zα
for M = poly(hpi + log(1/ǫ)).
Proof. It is enough to prove that for any z > 0, we can find z∗ with klog(z∗ )k∞ < M such that
p(z)
p(z∗ )
< (1 + ǫ) α .
∗
α
(z )
z
For the convenience of notation, assume that zn+1 = 1. Without loss of generality, after renaming
coordinates, assume that
| log(z1 )| ≥ | log(z2 )| ≥ . . . | log(zn )| ≥ | log(zn+1 )| = 0.
Let ∆ be a value bounded by poly(hpi+log(1/ǫ)) that will be fixed later. Let i be the coordinate
for which | log(zi )|−| log(zi+1 )| is maximized. If this maximum value is bounded by ∆, we are done.
Otherwise we change z by subtracting a constant from | log(z1 )|, . . . , | log(zi )| to make sure that
| log(zi )| − | log(zi+1 )| is reduced to ∆ and p(z)/zα does not grow by more than a 1 + ǫ/2n factor.
We obtain z∗ by repeating this “gap-closing” procedure for no more than n times.
Define the vector c ∈ {−1, 0, +1}n in the following way:
+1 if j ≤ i and zj > 1,
ci := −1 if j ≤ i and zj < 1,
0
otherwise.
Let s = | log(zi )| − | log(zi+1 )| − ∆. Let z̃ = exp(−sc)z (i.e., z̃i = e−sci zi ) be the result of the
“gap-closing” operation. We just need to set ∆ in such a way that
p(z)
p(z̃)
≤ (1 + ǫ/2n) α .
α
z̃
z
(11)
Let us partition F = supp(p) into F1 ∪ F2 ; let
F1 := {κ ∈ F : hc, κi ≥ hc, αi},
and let F2 be the rest. Note that F1 6= ∅ because α ∈ newt(p) = conv(F). We can decompose
p = p1 + p2 :
X
X
Cp (κ)zκ +
p(z) =
Cp (κ)zκ .
κ∈F1
|
κ∈F2
{z
p1 (z)
}
15
|
{z
p2 (z)
}
To prove (11), it is enough to show the following two inequalities
p1 (z̃)
≤
z̃α
p2 (z̃) ≤
p1 (z)
,
zα
ǫ
p1 (z̃).
2n
(12)
(13)
First we prove (12). For any κ ∈ F1 , we have
zκ
zκ
z̃κ
κ−α
κ−α
=
z̃
=
(exp(−sc)z)
=
exp(−shc,
κ
−
αi)
≤
,
z̃α
zα
zα
where in the inequality we used κ ∈ F1 and s ≥ 0. Multiplying the above inequality with Cp (κ)
and summing over all κ ∈ F1 we get (12).
It remains to prove (13). We will use the following claim to prove (13).
Claim 3.5. For any κ ∈ F2 , there exists κ′ ∈ F1 such that
′
z̃κ ≤ e−∆ z̃κ .
Proof. Let κ′ be the output of the greedy algorithm (with the ordering of coordinates being 1, . . . , n)
from Theorem 2.16 for the jump system F and linear function x 7→ hc, xi. We start from κ and
move towards κ′ , ensuring that κ remains in F and that z̃κ never decreases and in at least one
step it increases by a factor of e∆ .
While κ 6= κ′ , choose the smallest j for which κj 6= κ′j . If j > i, stop. Otherwise, by the greedy
construction of κ′ , we know that cj (κ′j − κj ) > 0. Let δ ∈ Zn be such that δj = cj and δk = 0 for
k 6= j. Note that δ is a (κ, κ′ )-step, see Definition 2.14. Therefore either κ + δ ∈ F or there exists
a (κ + δ, κ′ )-step δ′ such that κ + δ + δ′ ∈ F. In the former case, we set κ to κ + δ and in the
latter case we set κ to κ + δ + δ′ . We show that after the move, z̃κ does not decrease.
Case 1 (κ + δ ∈ F). Since cj log(z̃j ) ≥ ci log(z̃i ) ≥ ∆ > 0, we get
c
z̃κ+δ = zj j z̃κ ≥ e∆ z̃κ ≥ z̃κ .
Case 2 (κ + δ + δ′ ∈ F). Recall that δ′ has only one nonzero coordinate and that coordinate is
±1. Let that be the k-th coordinate. Since the first j − 1 coordinates of κ and κ′ are the same, it
must be that k ≥ j. Therefore
c
δ′
c
z̃κ+δ+δ = zj j zkk z̃κ ≥ zj j zk−ck z̃κ ≥ z̃κ .
′
It remains to show that in one of the steps, z̃κ grows by at least e∆ . Look at the quantity hc, κi
and how it changes over the course of the above algorithm. At the beginning, since κ ∈ F2
hc, κi < hc, αi.
When the algorithm finishes hc, κi = hc, κ′ i; this is because κj = κ′j for j ≤ i and cj = 0 for
j > i. Note that hc, κ′ i ≥ hc, αi because κ′ ∈ F1 . Therefore, at one point the quantity hc, κi must
increase. It is easy to see that when hc, κi increases, z̃κ grows by at least a factor of e∆ . In case
1, z̃κ always increases by a factor of e∆ . In case 2, hc, κi increases exactly when either δk′ = ck or
k > i, and in both situations z̃κ also increases by a factor of e∆ .
16
Now let us finish the proof of (13). We have
X
Cp (κ)z̃κ ≤ (deg p + 1)n · max Cp (κ) · max z̃κ ,
p2 (z̃) =
κ∈F2
κ∈F2
κ∈F2
where we used the fact that the number of terms in p is bounded by (deg p + 1)n . Similarly for p1 ,
we have
X
Cp (κ)z̃κ ≥ min Cp (κ) · max z̃κ .
p1 (z̃) =
κ∈F1
κ∈F1
κ∈F1
Putting the above two inequalities and using Claim 3.5, we get
p2 (z̃)
(deg p + 1)n · maxκ∈F2 Cp (κ) maxκ∈F2 z̃κ
≤
·
≤ epoly(hpi)−∆ .
p1 (z̃)
minκ∈F1 Cp (κ)
maxκ∈F1 z̃κ
It is easy to see that by making ∆ = poly(hpi)+ log(2n/ǫ) = poly(hpi+ log(1/ǫ)), we get the desired
inequality (13).
In the remaining part of this section, we first show that we can construct separation oracles, and then we bound the volume of the inner ellipses for the two ellipsoid methods. Let
us construct the separation oracles for the two ellipsoid methods. For the ellipsoid computing
inf y,z>0 p(y, z)/(yz/α)α , the separation oracle just needs to compute the gradient of the (log of
the) objective with respect to log(y), log(z). It is easy to see that this gradient can be computed
from the partial derivatives of p with respect to y, z. The latter can be computed using the following
lemma.
Lemma 3.6. Given oracle access to a polynomial p ∈ R[y1 , . . . , yn ], for any i, the quantity ∂yi p
can be computed at any given point using O(deg p) oracle calls.
Proof. Given a point y, consider the polynomial
q(t) = p(y1 , . . . , yi−1 , yi + t, yi+1 , . . . , yn ).
This is a univariate polynomial of degree at most deg p. Therefore we can compute all of its
coefficients by evaluating p at deg p + 1 many points. Then we can simply output the coefficient of
t.
Now we construct the separation oracle for the ellipsoid method computing
p(y, z)
.
α
α≥0 y,z>0 (yz/α)
sup inf
p(y,z)
Recall that if α ∈
/ newt(p), then inf y,z>0 (yz/α)
α = 0, see Fact 2.18. Given some α ≥ 0, we first
check whether α ∈ newt(p) using Fact 2.20, and in case it is not, we return the promised separating
hyperplane. So let us assume that α ∈ newt(p).
p(y,z)
In its current form, the expression (yz/α)
α is not concave in α. So we need to use the same trick
as in Lemma 2.12 to transform it into a concave form. We do this transformation after finding
p(y,z)
an approximate optimum for the inf. We compute an approximate optimum of inf y,z>0 (yz/α)
α
by restricting the domain of the variables y, z using Lemma 3.4 and running the inner ellipsoid
method to desired accuracy.
17
Next, we find a separating hyperplane at α, by calculating the gradient of
p(β(y/α), z)
(yz/α)β
with respect to β evaluated at β = α. This can again be done using Lemma 3.6.
It remains to lower bound the volume of the inner ellipses in the ellipsoid methods. First
consider the ellipsoid method computing the inf. It is enough to lower bound the volume of the
following set for any given ǫ > 0, y∗ , z∗ > 0
p(y∗ , z∗ )
p(y, z)
≤ (1 + ǫ) ∗ ∗
.
(log(y), log(z)) |
(yz/α)α
(y z /α)α
p(y,z)
To see this, it is enough to note that log( (yz/α)
α ) is Lipschitz with respect to log(y), log(z) with
Lipschitz constant bounded by poly(deg p, n) = poly(hpi).
Now consider the ellipsoid method for computing the sup. To bound the volume of the inner
ellipse we are going to use a similar idea as before, except we prove a slightly weaker form of
smoothness compared to being Lipschitz. We only consider the set of α’s in newt(p). We are
going to assume newt(p) is full-dimensional. If newt(p) is lower-dimensional we can use subspaceidentification methods to restrict the ellipsoid method to the affine subspace spanned by newt(p),
and proceed similarly to the full-dimensional case.
Let us start by considering the function
gy,z (α) = log(
p(y, z)
).
(yz)α
It is easy to see that for any y, z > 0, gy,z is Lipschitz with respect to α with Lipschitz constant
bounded by poly(log(y), log(z), n). So the function
gM (α) =
inf
y,z:kyk∞ ,kzk∞ <M
gy,z (α)
is also Lipschitz with respect to α with Lipschitz constant bounded by poly(log(M ), n). Now
consider the function h(α) = gM (α) + α log(α). This function is not Lipschitz in α, but it satisfies
a weaker form of smoothness. In particular if α is perturbed by ǫ2 , one can show that h(α) changes
by at most Lǫ for some L = poly(log(M ), n). The function h provides a uniform approximation of
p(y,z)
inf y,z>0 log( (yz/α)
α ) over newt(p). Using this fact and that newt(p) has only integral vertices of
bounded norm, one can construct the desired inner ellipse.
4
The Lower Bound
In this section we prove the LHS of (3) and (4). However, instead of lower bounding the quantity
P
Cp (κ)Cq (κ) for two real stable polynomials p(y), q(z), we treat p(y) · q(z) as a new bistable
κ∈Zn
+
polynomial on 2n variables y, z. Recall that by Fact 2.3 the product of any two stable polynomials
is a bistable polynomial. This allows us to prove the lower bound using an inductive argument
with a stronger statement. The following theorem implies the LHS of (3).
18
Theorem 4.1. For any bistable polynomial p(y1 , . . . , yn , z1 , . . . , zn ) with nonnegative coefficients
and any vector α ∈ Rn+ ,
X
p(y, z)
.
(14)
κ! · Cp (κ, κ) ≥ e−α inf
y,z>0
(yz/α)α
n
κ∈Z+
Instead of directly proving the above theorem, we reduce the general case to the multilinear case.
This reduction allows us to exploit Brändén characterization of real stable multilinear polynomials
(see Theorem 2.7).
Theorem 4.2. For any bistable multilinear polynomial p(y1 , . . . , yn , z1 , . . . , zn ) with nonnegative
coefficients and any vector α ∈ Rn+ ,
X
p(y, z)
y,z>0 (yz/α)α
Cp (κ, κ) ≥ (1 − α)1−α inf
κ∈{0,1}n
Note that not only do we use the above theorem to prove Theorem 4.1, but we also use it to
prove (4). In the rest of this section, we prove Theorem 4.1 using the above theorem; then in
Subsection 4.1 we prove the above theorem. Observe that if p is a multilinear bistable polynomial,
then Theorem 4.1 follows from the above theorem (this is because e−x ≤ (1 − x)1−x for any 0 ≤
x ≤ 1). So, we just need to reduce the case of a general p to the multilinear case.
Let m be a very large integer that we fix later in the proof. For now, we assume m is larger than
the degree of all variables in p. Throughout the proof we use y, z to denote n-dimensional (positive)
vectors and we use ỹ, z̃ ∈ Rn×m to denote the variables of the polarization of p. Let q(ỹ, z̃) := πm (p)
n×m
where for all 1 ≤ i ≤ n and 1 ≤ j ≤ m,
and r(ỹ, z̃) := q(ỹ, mz̃). Also, let α̃ ∈ R+
α̃i,j =
1
αi .
m
By applying Theorem 4.2 to r(ỹ, z̃) and α̃ we get
X
r(ỹ, z̃)
.
ỹ,z̃>0 (ỹz̃/α̃)α̃
Cr (κ, κ) ≥ (1 − α̃)1−α̃ inf
κ∈{0,1}n×m
(15)
First, we show that the RHS of the above is at least the RHS of (14) for any m. Then, we show
that as m → ∞ the LHS of the above converges to the LHS of (14).
Lemma 4.3.
(1 − α̃)1−α̃ inf
ỹ,z̃>0
p(y, z)
r(ỹ, z̃)
≥ e−α inf
.
α̃
y,z>0
(ỹz̃/α̃)
(yz/α)α
Proof. Firstly, note that by the definition of α̃, α̃α̃ = (α/m)α . Also, using the inequality (1 −
x)1−x ≥ e−x , we have
(1 − α̃)1−α̃ ≥ e−α̃ = e−α .
Therefore, to prove the lemma it is enough to show that
r(ỹ, z̃)
p(y, z)
≥ inf
.
α̃
y,z>0 (yz)α
ỹ,z̃>0 (mỹz̃)
inf
19
(16)
It follows by the change of variable z̃ ↔ z̃/m that
r(ỹ, z̃)
r(ỹ, z̃/m)
q(ỹ, z̃)
= inf
= inf
,
α̃
α̃
ỹ,z̃>0 (mỹz̃)
ỹ,z̃>0
ỹ,z̃>0 (ỹz̃)α̃
(ỹz̃)
inf
where in the last identity we used r(ỹ, z̃/m) = q(ỹ, mz̃/m) = q(ỹ, z̃). So, to prove (16), it is enough
to show that
p(y, z)
q(ỹ, z̃)
≥ inf
.
(17)
inf
α̃
y,z>0 (yz)α
ỹ,z̃>0 (ỹz̃)
Q
We use the AM-GM inequality to prove this. Fix an arbitrary ỹ, z̃ ∈ Rn×m , let yi = ( j ỹi,j )1/m
Q
p(y,z)
and zi = ( j z̃i,j )1/m for all i. We show that q(ỹ,z̃)
(ỹz̃)α ≥ (yz)α which proves (17).
Firstly, ỹα̃ = yα and similarly z̃α̃ = zα . Therefore, all we need to show is that p(y, z) ≤
q(ỹ, z̃). Recall that q = πm (p); this means that we have substituted each term yik in p with
1
e (y , . . . , yi,m ). Since all coefficients of p are nonnegative, to show p(y, z) ≤ q(ỹ, z̃) it is
(mk) k i,1
enough to show that for all i, k, m
1
m ek (ỹi,1 , . . . , ỹi,m )
k
or equivalently,
X
S∈([m]
k )
≤ yik ,
1/m
m
Y
Y
1
ỹi,j
ỹi,j ≥
.
m
k
j=1
j∈S
The above inequality is just an application of AM-GM. This proves (17) and completes the proof
of Lemma 4.3.
To finish the proof of Theorem 4.1 it remains to show that the LHS of (15) converges to the
LHS of (14) as m → ∞. We prove this statement termQby term.
Fix κ ∈ Zn+ and consider the monomial Cp (κ, κ) ni=1 yiκi ziκi of p. The contribution of this
monomial to the LHS of (14) is κ! · Cp (κ, κ). This monomial is substituted with
Cp (κ, κ)
=
n
Y
eκi (yi,1 , . . . , yi,m )eκi (mzi,1 , . . . , mzi,m )
m 2
i=1
n
Y
κi
κ
m i eκi (yi,1 , . . . , yi,m )eκi (zi,1 , . . . , zi,m )
Cp (κ, κ)
.
m 2
i=1
κi
in r. The contribution of this to the LHS of (15) is
n
n
Y
Y
mκ i
m
mκ i
Cp (κ, κ)
.
·
=
κ!
·
C
(κ,
κ)
·
p
m 2
m(m − 1) . . . (m − κi + 1)
κi
i=1
i=1
κi
As m → ∞ the ratio in the RHS converges to 1, so we get the same contribution. This completes
the proof of Theorem 4.1.
20
4.1
Multilinear Case
In this section we prove Theorem 4.2. Our first observation is the following useful identity:
X
Cp (κ, κ) =
n
Y
(1 + ∂yi ∂zi )p(y, z)
i=1
κ∈{0,1}n
y=z=0
.
To see this identity, note that any monomial of p which is not of the form yκ yκ for some κ is
mapped to zero in the RHS. Furthermore, any monomial
of the form Cp (κ, κ)yκ zκ is mapped to
Qn
Cp (κ, κ). So, throughout the proof we lower bound i=1 (1+∂yi ∂zi )p(y, z) y=z=0 . The main reason
that we use this reformulation is that it allows us to use induction. In the following lemma we show
that for any bistable polynomial p, (1 + ∂yi ∂zi )p is also bistable. So, we can apply the operator
Q
n
i=1 (1 + ∂yi ∂zi ) inductively.
Lemma 4.4. For any multilinear bistable polynomial p(y1 , . . . , yn , z1 , . . . , zn ) and any 1 ≤ i ≤ n,
(1 + ∂yi ∂zi )p|yi =zi =0 is also bistable.
Proof. Let q(y−i , z−i ) = (1 + ∂yi ∂zi )p(y, z)|yi =zi =0 . We need to show that q(y−i , −z−i ) is stable.
We will prove that
(18)
q(y−i , −z−i ) = (1 − ∂yi ∂zi )p(y, −z)|yi =zi =0 .
Then, the claim follows from the fact the the polynomial on the RHS is stable. In particular, since
p(y, −z) is stable, by Lemma 2.6, (1 − ∂yi ∂zi )p(y, −z) is stable. Furthermore, by Fact 2.4 when we
substitute yi = zi = 0 in the latter we still have a stable polynomial.
It remains to show (18). We can write p as follows:
p(y, z) = yi zi p1 (y−i , z−i ) + yi p2 (y−i , z−i ) + zi p3 (y−i , z−i ) + p4 (y−i , z−i ).
So,
q(y−i , −z−i ) = (1 + ∂yi ∂zi )p(y, −z1 , . . . , −zi−1 , zi , −zi+1 , . . . , zn )|yi =zi =0
= p1 (y−i , −z−i ) + p4 (y−i , −z−i )
On the other hand,
(1 − ∂yi ∂zi )p(y, −z)|yi =zi =0 = (1 − ∂yi ∂zi ) − yi zi p(y−i , −z−i ) + yi p2 (y−i , −z−i )
− zi p3 (y−i , −z−i) + p4 (y−i , −z−i ) |yi =zi =0
= p1 (y−i , −z−i ) + p4 (y−i , −z−i )
This proves (18).
We prove Theorem 4.2 by induction. See Lemma 4.5 for the base case of the induction. Suppose the statement of the theorem holds for any bistable multilinear polynomial with nonnegative
coefficients and at most 2n − 2 variables. Given p(y1 , . . . , yn , z1 , . . . , zn ) and some ǫ > 0 we show
that there exists y, z > 0 such that
n
Y
i=1
(1 + ∂yi ∂zi )p
y=z=0
+ ǫ ≥ (1 − α)1−α
21
p(y, z)
.
(yz/α)α
(19)
Let
q(y1 , . . . , yn−1 , z1 , . . . , zn−1 ) = (1 + ∂yn ∂zn )p|yn =zn =0 .
Note that by definition, q has nonnegative coefficients and is multilinear. Also, by Lemma 4.4, q is
a bistable polynomial.
Now, by the induction hypothesis, there exists y∗ , z∗ ∈ Rn−1 such that
n−1
Y
(1 + ∂yi ∂zi )q
i=1
y=z=0
+ ǫ/2 ≥ (1 − α−n )1−α−n
q(y∗ , z∗ )
.
(y∗ z∗ /α−n )α−n
Also, observe that by the definition of q, the LHS of the above is equal to the LHS of (19). So to
prove (19), it is enough to show that there exist yn , zn such that
(1 − α−n )1−α−n
p(y∗ , yn , z∗ , zn )
q(y∗ , z∗ )
1−α
.
+
ǫ/2
≥
(1
−
α)
(y∗ z∗ /α−n )α−n
(y∗ z∗ /α−n )α−n (yn zn /αn )αn
Let f (yn , zn ) = p(y∗ , yn , z∗ , zn ). To prove the above it is enough to show that there exists yn , zn
such that
f (yn , zn )
.
(20)
(1 + ∂yn ∂zn )f (yn , zn )|yn =zn =0 + ǫ/2 ≥ (1 − αn )1−αn
(yn zn /αn )αn
We prove this using Lemma 4.5. All we need to apply this lemma is that f (yn , zn ) is a multilinear bistable polynomial with nonnegative coefficients. Since p is multilinear with nonnegative
coefficients, and we substitute y1 , . . . , yn−1 , z1 , . . . , zn−1 with positive numbers, we get that f is
multilinear with nonnegative coefficients. Furthermore, since p(y1 , . . . , yn , −z1 , . . . , −zn ) is stable,
by Fact 2.4,
∗
∗
f (yn , −zn ) = p(y1∗ , . . . , yn−1
, yn , z1∗ , . . . , zn−1
, −zn )
is stable. So, f is a bistable polynomial. Therefore, (20) simply follows by Lemma 4.5. This
completes the proof of Theorem 4.2.
Lemma 4.5. Let p(y, z) be a multilinear bistable bivariate polynomial with nonnegative coefficients.
Then, for any α ≥ 0,
p(y, z)
(1 + ∂y ∂z )p y=z=0 ≥ (1 − α)1−α inf
y,z>0 (yz/α)α
Proof. Since p is multilinear, if α > 1, then the RHS of the above is zero and we are done. In
particular, for a fixed z, as y → ∞, p(y, z)/y α → 0. So, throughout the proof we assume that
α ≤ 1.
Since p is multilinear we can write it as
p(y, z) = ayz + by + cz + d,
for a, b, c, d ≥ 0. To prove the lemma’s conclusion, it is enough to show
ayz + by + cz + d
.
y,z>0
(yz/α)α
a + d ≥ (1 − α)1−α inf
(21)
When α = 1, the RHS of (21) is at most a since the ratio converges to a as y, z → ∞. When
α = 0, the RHS of (21) is at most d since the ratio converges to d as y, z → 0. In both cases the
RHS is at most a + d and we are done. So for the rest of the proof assume that 0 < α < 1.
22
Since p(x, −y) is stable, by Theorem 2.7 we have
−b · c = ∂y p(y, −z)|y=z=0 · ∂z p(y, −z)|y=z=0 ≥ p(0, 0) · ∂y ∂z p(y, −z)|y=z=0 = −a · d,
or, in other words b · c ≤ a · d. If this inequality is not tight, we can increase b, c for it to become
tight. In other words, let b′ ≥ b and c′ ≥ c be such that b′ c′ = ad. It is easy to see that such b′ , c′
can always be found. Then, to prove (21), it is enough to show that
ayz + b′ y + c′ z + d
,
y,z>0
(yz/α)α
a + d ≥ (1 − α)1−α inf
Consider the following matrix:
M :=
a b′
c′ d
.
Since det(M ) = ad − b′ c′ = 0, we have rank(M ) ≤ 1, and in particular we can write M as
uv ⊺ for u, v ∈ R2 . Since the entries of M are nonnegative we can further assume that u, v ∈ R2+ .
Therefore we have
e
a b′
=
M=
· g h ,
′
c d
f
where e, f, g, h ≥ 0. This implies the factorization ayz + b′ y + c′ z + d = (ey + f )(gz + h). So to
prove the conclusion of the lemma it is enough to show the following inequality:
(ey + f )(gz + h)
ey + f
gz + h
= αα (1 − α)1−α inf
· inf
. (22)
α
α
y,z>0
y>0
z>0
(yz/α)
y
zα
a + d = eg + f h ≥ (1 − α)1−α inf
To prove the above inequality we will use the following claim.
Claim 4.6. For any 0 < α < 1 and β, γ ≥ 0,
β α γ 1−α
βt + γ
=
.
t>0
tα
αα (1 − α)1−α
inf
Before proving the claim, let us use it to prove (22). Because of Claim 4.6, we can write the
RHS of (22) as
αα (1 − α)1−α ·
gα h1−α
(eg)α (f h)1−α
eα f 1−α
·
=
.
αα (1 − α)1−α αα (1 − α)1−α
αα (1 − α)1−α
Applying Claim 4.6 once more, with β = eg, γ = f h, we get
egt + f h
egt + f h
(eg)α (f h)1−α
=
inf
≤
t>0
αα (1 − α)1−α
tα
tα
t=1
= eg + f h,
which is the LHS of (22).
It only remains to prove Claim 4.6.
Proof. By the weighted AM-GM inequality we have
α
1−α
βt
βt
γ
γ
β α γ 1−α
βt + γ = α ·
+ (1 − α) ·
≥
.
·
= tα α
α
1−α
α
1−α
α (1 − α)1−α
23
Therefore
β α γ 1−α
βt + γ
≥
.
t>0
tα
αα (1 − α)1−α
inf
αγ
Note that we have equality in AM-GM when βt/α = γ/(1 − α). So it is enough to set t = (1−α)β
to get equality, which is possible if β, γ > 0.
It remains to prove the claim when either β = 0 or γ = 0. In both cases the RHS is 0 so we
just need to show that the LHS is 0 as well. When β = 0, it is enough to let t → ∞, and when
γ = 0, it is enough to let t → 0 to get
βt + γ
= 0.
t>0
tα
inf
This completes the proof of Lemma 4.5.
5
The Upper Bound
In this section we prove the RHS of (3). This will complete the proof of Theorem 1.1.
Theorem 5.1. For any two real stable polynomials p(y1 , . . . , yn ) and q(z1 , . . . , zn ) with nonnegative
coefficients,
X
p(y)q(z)
.
(23)
κκ · Cp (κ)Cq (κ) ≤ sup inf
α
α≥0 y,z>0 (yz/α)
n
κ∈Z+
Note that κκ ≥ κ! for any κ, so the above theorem immediately implies the RHS of (3).
Note that unlike Theorem 4.1 here we only prove our theorem for product of stable polynomials (as
opposed to bistable polynomials). This is because the above statement fails for (multilinear) bistable
polynomials. For example, the polynomial p(y, z) = 1 + yz is bistable; however by Claim 4.6, for
any fixed 0 < α < 1 we have
p(y, z)
1 + yz
= inf
= (1 − α)−(1−α) .
α
y,z>0 (yz/α)
y,z>0 (yz/α)α
inf
It is an easy exercise to see that (1 − α)−(1−α) ≤ e1/e . But the quantity
P
κ∈Zn
+
κκ · Cp (κ, κ) is 2
which is strictly more than e1/e .
In order to prove Theorem 5.1 we will crucially use the following minimax result about the RHS
of (23).
Lemma 5.2. For any bistable polynomial p(y, z) with nonnegative coefficients,
p(y, z)
p(y, z)
= inf sup inf
.
α
α
α≥0 y,z>0 (yz/α)
y>0 α≥0 z>0 (yz/α)
sup inf
Before proving Lemma 5.2, let us show how it can be used to prove Theorem 5.1.
24
Proof of Theorem 5.1. Because of Lemma 5.2, we have
p(y)q(z)
p(y)q(z)
= inf sup inf
.
α
α
α≥0 y,z>0 (yz/α)
y>0 α≥0 z>0 (yz/α)
sup inf
So to prove (23), we just need to show that for any fixed y > 0, the following inequality holds:
X
p(y)q(z)
≥
κκ · Cp (κ) · Cq (κ).
α
(yz/α)
α≥0 z>0
n
sup inf
κ∈Z+
Note that p(y) is fixed on the LHS. So
p(y)q(z)
α
α≥0 z>0 (yz/α)
sup inf
q(z)
α
α≥0 z>0 (yz/α)
= p(y) sup inf
=
X
q(z)
.
α
α≥0 z>0 (yz/α)
Cp (κ)yκ sup inf
κ∈Zn
+
(24)
But for any fixed κ ∈ Zn+ , we have
q(z)
sup inf
α
α≥0 z>0 (yz/α)
≥
q(z)
inf
z>0 (yz/α)α
q(z)
z>0 (yz/κ)κ
= inf
α=κ
Cq (κ)zκ
Cq (κ)
≥ inf
= κκ ·
.
κ
z>0 (yz/κ)
yκ
Plugging back into (24) gives us
X
Cq (κ) X κ
p(y)q(z)
≥
κ · Cp (κ) · Cq (κ).
Cp (κ)yκ κκ κ =
α
y
α≥0 z>0 (yz/α)
n
κ
sup inf
κ∈Z+
This concludes the proof of Theorem 5.1, assuming Lemma 5.2.
5.1
Duality
In this section we prove Lemma 5.2. From basic properties of inf and sup, it is immediate that
p(y, z)
p(y, z)
≤ inf sup inf
.
α
α
α≥0 y,z>0 (yz/α)
y>0 α≥0 z>0 (yz/α)
sup inf
So to prove equality, it is enough to prove that the LHS is greater than or equal to the RHS. We
would ideally like to apply one of the classical minimax theorems, such as von Neumann’s minimax
theorem, which often require the function inside inf and sup to be convex-concave. However in its
current form p(y, z)/(yz/α)α does not satisfy this property. Instead we transform both sides to a
form suitable for minimax theorems.
Because of Fact 2.18, we can replace supα≥0 by inf α∈[0,deg p]n . This is crucial for applying
Theorem 5.3, since the domain of α has to be compact. Additionally, by Lemma 2.12 we can move
25
α from the denominator to the numerator, and inside of p. Finally, we apply the change of variables
y ↔ exp(ỹ) and z ↔ exp(z̃). Therefore, to prove the above inequality, it is enough to show that
sup
inf
α∈[0,deg p]n ỹ,z̃∈Rn
p(exp(ỹ), α exp(z̃))
p(exp(ỹ), α exp(z̃))
inf
sup
≥ inf
.
α
α
α
α
exp(ỹ) exp(z̃)
ỹ∈Rn α∈[0,deg p]n z̃∈Rn exp(ỹ) exp(z̃)
(25)
We use Sion’s minimax theorem. We state a weaker form of this theorem below, which will be
enough for our purposes; for the original stronger version, look at [Sio58].
Theorem 5.3 (Immediate corollary of [Sio58]). Let A be a compact convex subset of Rn and B
a (not necessarily compact) convex subset of Rm . Suppose that f : A × B → R has the following
properties:
• For each a ∈ A, the function f (a, ·) : B → R is continuous and quasi-convex.
• For each b ∈ B, the function f (·, b) : A → R is continuous and quasi-concave.
Then we have
sup inf f (a, b) = inf sup f (a, b).
a∈A b∈B
b∈B a∈A
Recall that a real-valued function f is quasi-convex if f −1 (−∞, t] is convex for any t ∈ R and
f is quasi-concave if −f is quasi-convex.
We will apply Theorem 5.3 to the following function.
f (α, (ỹ, z̃)) :=
p(exp(ỹ), α exp(z̃))
.
exp(ỹ)α exp(z̃)α
For a fixed α ≥ 0, the function f is log-convex, and therefore quasi-convex, in ỹ, z̃ because of
Fact 2.11. For a fixed ỹ, z̃, the function f is log-concave, and therefore quasi-concave, in α because
of Corollary 2.10. It is easy to see that f is continuous in each variable as well. Therefore, by
Theorem 5.3 we have
sup
inf
α∈[0,deg p]n ỹ,z̃∈Rn
p(exp(ỹ), α exp(z̃))
exp(ỹ)α exp(z̃)α
=
inf
sup
ỹ,z̃∈Rn α∈[0,deg p]n
≥
inf
sup
p(exp(ỹ), α exp(z̃))
exp(ỹ)α exp(z̃)α
inf
ỹ∈Rn α∈[0,deg p]n z̃∈Rn
p(exp(ỹ), α exp(z̃))
.
exp(ỹ)α exp(z̃)α
This proves (25) and finishes the proof of Lemma 5.2.
References
[AO15a]
Nima Anari and Shayan Oveis Gharan. Effective-Resistance-Reducing Flows, Spectrally
Thin Trees, and Asymmetric TSP. In FOCS, pages 20–39, 2015. 2
[AO15b]
Nima Anari and Shayan Oveis Gharan. The kadison-singer problem for strongly rayleigh
measures and applications to asymmetric tsp. 2015. 2
[AOSS17] Nima Anari, Shayan Oveis Gharan, Amin Saberi, and Mohit Singh. Nash social welfare,
matrix permanent, and stable polynomials. In ITCS, 2017. to appear. 11
26
[BB10]
J. Borcea and P. Brädén. Multivariate Pólya-Schur classification problems in the Weyl
algebra. Proceedings of the London Mathematical Society, 101(3):73–104, 2010. 5
[BBL09]
Julius Borcea, Petter Branden, and Thomas M. Liggett. Negative dependence and the
geometry of polynomials. Journal of American Mathematical Society, 22:521–567, 2009.
4, 5, 6, 11
[BC95]
André Bouchet and William H. Cunningham. Delta-matroids, jump systems, and bisubmodular polyhedra. SIAM J. Discrete Math., 8(1):17–32, 1995. 8
[Brä07]
Petter Brändén. Polynomials with the half-plane property and matroid theory. Advances
in Mathematics, 216(1):302–320, 2007. 5, 7
[BV06]
S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University Press, 2006.
6
[Csi14]
Péter Csikvári. Lower matching conjecture, and a new proof of schrijver’s and gurvits’s
theorems. arXiv preprint arXiv:1406.0766, 2014. 9
[Gra02]
JH Grace. The zeros of a polynomial. In Proc. Cambridge Philos. Soc, volume 11, pages
352–357, 1902. 5
[GS14]
Leonid Gurvits and Alex Samorodnitsky. Bounds on the permanent and some applications. In FOCS, pages 90–99. IEEE Computer Society, 2014. 9
[Gül97]
Osman Güler. Hyperbolic polynomials and interior point methods for convex programming. MOR, 22(2):350–77, 1997. 6
[Gur06]
Leonid Gurvits. Hyperbolic polynomials approach to van der waerden/schrijver-valiant
like conjectures: Sharper bounds, simpler proofs and algorithmic applications. In STOC,
STOC ’06, pages 417–426, 2006. 2
[Gur11]
L. Gurvits. Unleashing the power of schrijvers permanental inequality with the help of
the bethe approximation. arXiv:1106.2844, 2011. 9
[KT12]
Alex Kulesza and Ben Taskar. Determinantal point processes for machine learning. arXiv
preprint arXiv:1207.6083, 2012. 11
[Lel17]
Marc Lelarge. Counting matchings in irregular bipartite graphs. In SODA, 2017. to
appear. 9
[Lov97]
László Lovász. The membership problem in jump systems. J. Combin. Theory Ser. B,
70(1):45–66, 1997. 7
[MF08]
S. Thomas McCormick and Satoru Fujishige. Strongly polynomial and fully combinatorial algorithms for bisubmodular function minimization. In Proceedings of the Nineteenth
Annual ACM-SIAM Symposium on Discrete Algorithms, pages 44–53. ACM, New York,
2008. 8, 9
[MSS13]
Adam Marcus, Daniel A Spielman, and Nikhil Srivastava. Interlacing Families II: Mixed
Characteristic Polynomials and the Kadison-Singer Problem. 2013. 2
27
[NS16]
Aleksandar Nikolov and Mohit Singh. Maximizing determinants under partition constraints. In Proceedings of the 48th Annual ACM SIGACT Symposium on Theory of
Computing, STOC 2016, Cambridge, MA, USA, June 18-21, 2016, pages 192–201, 2016.
1, 3, 13
[Sch98]
Alexander Schrijver. Counting 1-factors in regular bipartite graphs. Journal of Combinatorial Theory, Series B, 72(1):122–135, 1998. 9
[Sio58]
Maurice Sion. On general minimax theorems. Pacific J. Math., 8:171–176, 1958. 26
[SV16]
Damian Straszak and Nisheeth K. Vishnoi. Real stable polynomials and matroids: Optimization and counting. 2016. 3
[Sze22]
G Szegö. Bemerkungen zu einem satz von jh grace über die wurzeln algebraischer gleichungen. Mathematische Zeitschrift, 13(1):28–55, 1922. 5
[Wal22]
JL Walsh. On the location of the roots of certain types of polynomials. Transactions of
the American Mathematical Society, 24(3):163–180, 1922. 5
28
| 8cs.DS
|
arXiv:1604.01757v1 [math.GR] 6 Apr 2016
ON SEMIGROUPS WITH PSPACE-COMPLETE SUBPOWER
MEMBERSHIP PROBLEM
MARKUS STEINDL
Abstract. Fix a finite semigroup S and let a1 , . . . , ak , b be tuples in a direct
power S n . The subpower membership problem (SMP) for S asks whether b
can be generated by a1 , . . . , ak . For combinatorial Rees matrix semigroups we
establish a dichotomy result: if the corresponding matrix is of a certain form,
then the SMP is in P; otherwise it is NP-complete. For combinatorial Rees
matrix semigroups with adjoined identity, we obtain a trichotomy: the SMP
is either in P, NP-complete, or PSPACE-complete. This result yields various semigroups with PSPACE-complete SMP including the 6-element Brandt
monoid, the full transformation semigroup on 3 or more letters, and semigroups
of all n by n matrices over a field for n ≥ 2.
1. Introduction
In this paper we continue the investigation of the subpower membership problem
(SMP) for semigroups started in [1] and [10]. At the Conference on Order, Algebra,
and Logics in Nashville, 2007, Ross Willard proposed the SMP as follows [12]: Fix a
finite algebraic structure S with finitely many basic operations. Then the subpower
membership problem for S is the following decision problem:
SMP(S )
Input:
{a1 , . . . , ak } ⊆ S n , b ∈ S n
Problem: Is b in the subalgebra of S n generated by a1 , . . . , ak ?
The SMP occurs in connection with the constraint satisfaction problem (CSP) [4].
In the algebraic approach to the CSP, each constraint relation is considered to be
a subalgebra of a power (subpower) of a certain finite algebra whose operations
are the polymorphisms of the constraint language. Checking whether a given tuple
belongs to a constraint relation represented by its generators is precisely the SMP
for the polymorphism algebra.
The input size of SMP(S) is essentially (k + 1)n. We can always decide the
problem using a straightforward closure algorithm in time exponential in n. Thus
SMP(S) is in EXPTIME for every algebra S. However, the following questions
arise:
• How does the algebra S affect the computational complexity of SMP(S)?
• For which algebras S can SMP(S) be solved in time polynomial in k and
n?
Date: March 19, 2018.
2000 Mathematics Subject Classification. Primary: 20M99; Secondary: 68Q25.
Key words and phrases. subalgebras of powers, membership test, computational complexity,
PSPACE-complete, NP-complete.
Supported by the Austrian Science Fund (FWF): P24285.
1
2
MARKUS STEINDL
• When is the problem complete in NP, PSPACE, or EXPTIME? Can it also
be complete in a class other than these?
Mayr [6] proved that the SMP for Mal’cev algebras is in NP. He also showed that
for certain generalizations of groups and quasigroups the SMP is in P. Kozik [5]
constructed a finite algebra with EXPTIME-complete SMP.
For semigroups the SMP is in PSPACE. This was shown in [1] by Bulatov,
Mayr, and the author of the present paper. We also proved that the SMP of the
full transformation semigroup on five letters is PSPACE-complete. It was the first
algebra known to have a PSPACE-complete SMP. In the same paper a dichotomy
result for commutative semigroups was established: if a commutative semigroup S
embeds into a direct product of a Clifford semigroup and a nilpotent semigroup,
then SMP(S) is in P; otherwise it is NP-complete.
Another dichotomy for idempotent semigroups was established in [10]: if an
idempotent semigroup S satisfies a certain pair of quasiidentities, then SMP(S) is
in P; otherwise it is NP-complete.
The first result of the current work is a condition for a semigroup S under which
SMP(S) is NP-hard:
Theorem 1.1. Let r, s, t be elements of a finite semigroup S such that s does not
generate a group and rs = st = s. Then SMP(S) is NP-hard.
We will prove this result in Section 2 by reducing the Boolean satisfiability problem
SAT to SMP(S).
A semigroup is called combinatorial if every subgroup has one element. Combinatorial Rees matrix semigroups are of the form M0 ({1}, I, Λ, P ) (see [3, Theorem
3.2.3]). We give the following alternative notation: For nonempty sets I, Λ and a
matrix P ∈ {0, 1}Λ×I we let SP := (I × Λ) ∪ {0} and define a multiplication on SP
by
(
[i, µ] if P (λ, j) = 1,
[i, λ] · [j, µ] :=
0
if P (λ, j) = 0,
0 · [i, λ] := [i, λ] · 0 := 0 · 0 := 0.
It is easy to see that SP is indeed a combinatorial semigroup. We say the matrix
P ∈ {0, 1}Λ×I has one block if there exist J ⊆ I, ∆ ⊆ Λ such that for i ∈ I, λ ∈ Λ,
P (λ, i) = 1 if and only if
(λ, i) ∈ ∆ × J.
( 10 01 )
For P =
we call B2 := SP the Brandt semigroup, and for P = ( 11 10 ) we denote
SP by A2 .
In Section 3 we establish the following two results:
Theorem 1.2. Let SP be a finite combinatorial Rees matrix semigroup. If the matrix P has one block, then SMP(SP ) is in P. Otherwise SMP(SP ) is NP-complete.
Corollary 1.3. The SMP for the Brandt semigroup B2 and for the semigroup A2
is NP-complete.
In Section 4 we state a condition for semigroups S under which SMP(S) is PSPACEcomplete:
Theorem 1.4. Let S be a finite semigroup and s, t, 1 ∈ S such that
(a) sts = s,
ON SEMIGROUPS WITH PSPACE-COMPLETE SMP
3
(b) s does not generate a group,
(c) s1 = s and t1 = t.
Then SMP(S) is PSPACE-complete.
In the proof we will reduce quantified 3SAT to SMP(S). It follows that adjoining
an identity to B2 or A2 already results in a PSPACE-complete SMP:
Theorem 1.5. The SMP for the Brandt monoid B21 and for the monoid A12 is
PSPACE-complete.
This result is part of Corollary 4.9. Both B21 and A12 embed into T3 , the full transformation semigroup on three letters. Thus SMP(T3 ) is also PSPACE-complete.
So Theorem 1.5 generalizes the result from [1] that SMP(T5 ) is PSPACE-complete.
In addition, B2 and A2 are the first groupoids known to have an NP-complete SMP
where adjoining an identity yields a groupoid with PSPACE-complete SMP. Further examples of semigroups with PSPACE-complete SMP are listed in Section 4.
In Section 5 we will consider Rees matrix semigroups with adjoined identity and
prove the following trichotomy result:
Theorem 1.6. Let SP be a finite combinatorial Rees matrix semigroup.
(a) If all entries of the matrix P are 1, then SMP(SP1 ) is in P.
(b) If P has one block and some entries are 0, then SMP(SP1 ) is NP-complete.
(c) Otherwise SMP(SP1 ) is PSPACE-complete.
2. Semigroups with NP-hard SMP
In this section we will prove Theorem 1.1 by reducing the Boolean satisfiability
problem SAT to SMP(S). It follows that the SMP for a semigroup S is already
NP-hard if S has a D-class that contains both group and non-group H-classes.
We denote Green’s equivalences by L, R, J , H, D [3, p. 45]. For the definition
of the related preorders ≤L , ≤R , ≤J see [3, p. 47]. We write [n] := {1, . . . , n} for
n ∈ N and set [0] := ∅. We consider a tuple a in a direct power S n to be a function
a : [n] → S. This means the ith coordinate of this tuple is denoted by a(i) rather
than ai . The subsemigroup generated by a set A = {a1 , . . . , ak } may be denoted
by hAi or ha1 , . . . , ak i.
Lemma 2.1. Let s belong to a finite semigroup S. Then s generates a group if
and only if s2 J s.
Proof. If s generates a group, then sk = s for some k ≥ 2. Thus s2 J s.
For the converse let s2 J s. First assume the J -class Js is the minimal ideal
of S. Then Js is a finite simple semigroup by [3, Proposition 3.1.4]. By the Rees
Theorem for finite simple semigroups s generates a group.
Now assume Js is not the minimal ideal of S. Let
J(s) := {r ∈ S | r ≤J s},
I(s) := {r ∈ S | r <J s}.
By [3, Proposition 3.1.4] the principal factor J(s)/I(s) is either null or 0-simple.
Since s2 J s, the second case applies. By the Rees Theorem for finite 0-simple
semigroups s generates a group.
Lemma 2.2. Let r, s, t be elements of a finite semigroup S such that s does not
generate a group and rs = st = s. Then there are idempotents e, f ∈ S such that
es = sf = s and every product a1 · · · ak in s, e, f in which s occurs at least twice
does not yield s.
4
MARKUS STEINDL
Proof. First assume s is regular, i.e. sus = s for some u ∈ S. Let e and f be the
idempotent powers of su and us respectively. Clearly es = sf = s. Let a1 · · · ak
be a product in s, e, f , and i < j such that ai = aj = s. Let ℓ ∈ {i + 1, . . . , j}
be maximal such that ai+1 = . . . = aℓ−1 = f . Then ai · · · aℓ−1 = s, and thus
ai · · · aℓ ∈ {s2 , se}. Note that se = s(su)m for some m ∈ N. Now a factor s2 occurs
in the product ai · · · aℓ . Since s does not generate a group, Lemma 2.1 implies that
s2 <J s. Thus a1 · · · aℓ <J s, and the result follows.
Now assume s is not regular. By [3, Theorem 3.1.6] the principal factor J(s)/I(s)
is null. Let e and f be the idempotent powers of r and t respectively. Let a1 · · · ak
be a product in s, e, f , and let i < j such that ai = aj = s. Then a1 · · · ai ≤J s
and ai+1 · · · ak ≤J s. Since J(s)/I(s) is null, it follows that a1 · · · ak <J s.
Proof of Theorem 1.1. Let S satisfy the assumptions. We reduce the Boolean satisfiability problem SAT to SMP(S). SAT is NP-complete [2], and we give the
following definition:
SAT
Input:
Clauses C1 , . . . , Cm ⊆ {x1 , . . . , xk , ¬x1 , . . . , ¬xk }.
Problem: Do truth values for
W x1 , . . . , xk exist
W for which the Boolean formula
Φ(x1 , . . . , xk ) := ( C1 ) ∧ . . . ∧ ( Cm ) is true?
For all j ∈ [k] we may assume that xj or ¬xj occurs in some clause Ci . We define
an SMP(S) instance
A := {a01 , . . . , a0k , a11 , . . . , a1k } ⊆ S k+m , b ∈ S k+m .
Let e, f ∈ S be idempotents with the properties from Lemma 2.2. Let g be the
idempotent power of se. Observe that e and g form a two-element semilattice with
g < e.
Let j ∈ [k] and z ∈ {0, 1}. For i ∈ [k] let
f if i < j,
azj (i) := s if i = j,
e if i > j,
and for i ∈ [m] let
a0j (k
a1j (k
+ i) :=
(
g
e
if ¬xj ∈ Ci ,
otherwise,
+ i) :=
(
g
e
if xj ∈ Ci ,
otherwise.
Let
b(i) := s
b(k + i) := g
for i ∈ [k],
for i ∈ [m].
We claim that
(1)
the Boolean formula Φ is satisfiable if and only if b ∈ hAi.
For the (⇒) direction let z1 , . . . , zk ∈ {0, 1} such that Φ(z1 , . . . , zk ) = 1. We show
that
(2)
b = az11 · · · azkk .
ON SEMIGROUPS WITH PSPACE-COMPLETE SMP
5
For i ∈ [k] we have az11 · · · azkk (i) = ei−1 sf k−i = s = b(i). For i ∈ [m] the clause
W
Ci is satisfied under the assignment x1 7→ z1 , . . . , xk 7→ zk . Thus there is a j ∈ [k]
z
such that xj ∈ Ci and zj = 1, or ¬xj ∈ Ci and zj = 0. In both cases aj j (k + i) = g,
zk
z1
and thus a1 · · · ak (k + i) = g = b(k + i). This proves (2) and the (⇒) direction of
(1).
For the (⇐) direction of (1) assume b = azj11 · · · azjℓℓ for some ℓ ∈ N, j1 , . . . , jℓ ∈ [k],
and z1 , . . . , zℓ ∈ {0, 1}. We show that j1 , . . . , jℓ are distinct. Suppose jp = jq for
p < q. The factors of the product azj11 · · · azjℓℓ (jp ) are given by s, e, f . The factor s
occurs at least twice since ajp (jp ) = ajq (jp ) = s. By Lemma 2.2 this product does
not yield s, contradicting our assumption. We define an assignment
θ : xj1 7→ z1 , . . . , xjℓ 7→ zℓ ,
xj 7→ 0
for j ∈ [k] \ {j1 , . . . , jℓ },
and show that θ satisfies the formula Φ. Let i ∈ [m]. Since azj11 . . . azjℓℓ (k + i) is a
z
product in e, g that yields g, some factor ajpp (k + i) must be g. From the definition
zp
of ajp we see that either zp = 0 and ¬xjp ∈ Ci , or zp = 1 and xjp ∈ Ci . This means
W
the formula Ci is satisfied under the assignment θ. Since i was arbitrary, Φ is
also satisfied. The equivalence (1) and the theorem are proved.
Corollary 2.3. If a J -class of a finite semigroup S contains both group and nongroup H-classes, then SMP(S) is NP-hard.
Proof. Let s ∈ S such that Hs is not a group and Js contains group H-classes.
From Green’s Theorem [3, Theorem 2.2.5] we know that s does not generate a
group. Since Js contains an idempotent and Js = Ds , the element s is regular
by [3, Proposition 2.3.1]. That is, there is a u ∈ S such that sus = s. Now su, s,
and us fulfill the hypothesis of Theorem 1.1.
3. Combinatorial Rees matrix semigroups
In this section we will establish a P/NP-complete dichotomy for the SMP for
combinatorial Rees matrix semigroups by proving Theorem 1.2. After that we apply
this result to combinatorial 0-simple semigroups.
Combinatorial Rees matrix semigroups have the following property:
Lemma 3.1 (cf. [9, Lemma 2.2]). Let k ≥ 2 and a1 , . . . , ak be elements of a
combinatorial Rees matrix semigroup SP .
(a) We have a1 · · · ak = 0 if and only if aj aj+1 = 0 for some j ∈ [k − 1].
(b) If a1 · · · ak 6= 0, then there are i, j ∈ I and λ, µ ∈ Λ such that a1 = [i, λ],
ak = [j, µ], and a1 · · · ak = [i, µ].
Proof. Straightforward.
The next two results will allow us to show that the SMP for a combinatorial
Rees matrix semigroup is in NP.
Lemma 3.2 (cf. [9, Theorem 4.3]). Let f := y1 · · · yk and g := z1 · · · zℓ be words
over an alphabet X such that
(a) {yi yi+1 | i ∈ [k − 1]} = {zj zj+1 | j ∈ [ℓ − 1]},
(b) y1 = z1 and yk = zℓ .
Then every combinatorial Rees matrix semigroup SP satisfies f ≈ g.
6
MARKUS STEINDL
Proof. Let SP be a combinatorial Rees matrix semigroup, and let α : X + → SP
be a homomorphism from the free semigroup over X to SP . By item (a) we have
{y1 , . . . , yk } = {z1 , . . . , zℓ }. We claim that
(3)
α(y1 · · · yk ) = 0
if and only if α(z1 · · · zℓ ) = 0.
Assume α(y1 · · · yk ) = 0. Then α(yi )α(yi+1 ) = 0 for some i ∈ [k−1] by Lemma 3.1 (a).
By item (a) yi yi+1 = zj zj+1 for some j ∈ [ℓ − 1]. Thus α(zj )α(zj+1 ) = 0, and hence
α(z1 · · · zℓ ) = 0. This proves (3).
If α(y1 · · · yk ) = 0, then α(y1 · · · yk ) = α(z1 · · · zℓ ) by (3). Assume α(y1 · · · yk ) 6=
0. Then also α(z1 · · · zℓ ) 6= 0, and Lemma 3.1 (b) implies
α(y1 ) · · · α(yk ) = α(z1 ) · · · α(zℓ ).
This proves the lemma.
Lemma 3.3. Let f be a word over x1 , . . . , xk . Then there is a word g such that
(a) the length of g is at most k(k 2 + 1), and
(b) every combinatorial Rees matrix semigroup satisfies f ≈ g.
Proof. Let f = y1 · · · yℓ for y1 , . . . , yℓ ∈ {x1 , . . . , xk }. We show that there is a word
g such that item (b) holds and in which each variable xi occurs at most k 2 + 1
times. Fix i ∈ [k]. Let j1 , . . . , jm ∈ [ℓ] be the positions of xi in y1 · · · yℓ . Let
v1 := y1 · · · yj1 ,
vr := yjr−1 +1 · · · yjr
for r ∈ {2, . . . , m},
vm+1 := yjm +1 · · · yℓ .
Note that f = v1 · · · vm+1 . Now for every word h := z1 · · · zn over x1 , . . . , xk let
E(h) := {zj zj+1 | j ∈ [n − 1]}.
It is not hard to see that
E(v1 · · · vr ) = E(v1 ) ∪ E(xi v2 ) ∪ . . . ∪ E(xi vr )
for r ∈ {2, . . . , m + 1}.
We define
R := {r ∈ {2, . . . , m} | E(xi vr ) 6⊆ E(v1 · · · vr−1 )}
and let
Y
g := v1 ( vr )vm+1 .
r∈R
Apparently g is a concatenation of subwords of f , and f and g start with the same
letter. We show that
(4)
f and g also end with the same letter.
If vm+1 is nonempty, then (4) is clear. If vm+1 is empty, then yℓ = xi , and g ends
with a subword vr for some r ∈ [m]. Since vr and f both end with xi , (4) is proved.
We have
m
[
E(f ) = E(v1 · · · vm+1 ) = E(v1 ) ∪
E(xi vr ) ∪ E(xi vm+1 )
r=2
= E(v1 ) ∪
[
E(xi vr ) ∪ E(xi vm+1 ) = E(g).
r∈R
Now Lemma 3.2 implies item (b).
ON SEMIGROUPS WITH PSPACE-COMPLETE SMP
7
Next observe that |R| ≤ k 2 by the definitions of R and E. This means xi occurs
at most k 2 + 1 times in g. Since xi was arbitrary, we can reduce the number of
occurrences of each variable in f to at most k 2 + 1. Item (a) is proved.
Lemma 3.4. The SMP for a finite combinatorial Rees matrix semigroup is in NP.
Proof. Let S be such a semigroup, and let {a1 , . . . , ak } ⊆ S n , b ∈ S n be an instance of SMP(S). If b ∈ ha1 , . . . , ak i, then there is a term function f such that
f (a1 , . . . , ak ) = b. By Lemma 3.3 there is a word g which induces f and whose
length is polynomial in k. Now g witnesses the positive answer.
For the following result note that the all-0 matrix has one block.
Lemma 3.5. Let SP be a finite combinatorial Rees matrix semigroup such that
P ∈ {0, 1}Λ×I has one block. Then Algorithm 1 decides SMP(SP ) in polynomial
time.
Algorithm 1 Decides SMP(SP ) in polynomial time if P has one block.
Input: A ⊆ SP n , b ∈ SP n ,
m ∈ {0, . . . , n} such that b(i) 6= 0 iff i ∈ [m],
J ⊆ I, ∆ ⊆ Λ such that P (λ, i) = 1 iff (λ, i) ∈ ∆ × J for i ∈ I, λ ∈ Λ.
Output: true if b ∈ hAi, false otherwise.
1: if b ∈ A then
2:
return true
3: end if
Q
4: d :=
{a ∈ A | a([m]) ⊆ J × ∆} (some order)
5: return ∃a1 , a2 ∈ A : a1 da2 = b
Proof. Fix an input A ⊆ SP n , b ∈ SP n . We may assume that there is an m ∈
{0, . . . , n} such that
b(i) 6= 0 for i ∈ [m],
b(i) = 0 for i ∈ {m + 1, . . . , n}.
Correctness of Algorithm 1. If Algorithm 1 returns true, then clearly b ∈ hAi.
Conversely assume b ∈ hAi. We show that true is returned. Let g1 , . . . , gk ∈ A such
that b = g1 · · · gk . If k = 1 then true is returned in line 2. Assume k ≥ 2. We have
(5)
g1 (i) ∈ I × ∆,
gk (i) ∈ J × Λ, and
g2 (i), . . . , gk−1 (i) ∈ J × ∆ for all i ∈ [m];
otherwise we obtain the contradiction g1 · · · gk (i) = 0 for some i ∈ [m]. Let d have
a value assigned by line 4. We claim that
(6)
g1 dgk = b.
For i ∈ [m] we have d(i) ∈ J × ∆. The multiplication rule and (5) imply
b(i) = g1 · · · gk (i) = g1 dgk (i).
Now let i ∈ {m + 1, . . . , n}. Since b(i) = 0, there are three cases: g1 (i) ∈
/ I × ∆,
gk (i) ∈
/ J × Λ, or gj (i) ∈
/ J × ∆ for some j ∈ {2, . . . , k − 1}. In the first two cases
g1 dgk (i) = 0 = b(i) holds. In the third case a := gj occurs as factor in line 4. Thus
8
MARKUS STEINDL
d(i) ∈
/ J × ∆, and hence g1 dgk (i) = 0. This proves (6). So the algorithm returns
true in line 5.
Complexity of Algorithm 1. The product in line 4 can be computed in O(|A|n)
time. Checking the condition in line 5 requires O(|A|2 n) time. Altogether Algorithm 1 runs in O(|A|2 n) time.
Now we prove Theorem 1.2 and Corollary 1.3.
Proof of Theorem 1.2. Assume P ∈ {0, 1}Λ×I . If P has one block, then SMP(SP )
is in P by Lemma 3.5. Assume P does not have one block. Then there are i, j ∈ I
and λ, µ ∈ Λ such that
P (λ, i) = P (µ, j) = 1
and P (µ, i) = 0.
Let r := [i, λ], s := [i, µ], and t := [j, µ]. Then rs = st = s, and s does not
generate a group. By Theorem 1.1 SMP(SP ) is NP-hard. NP-easiness follows from
Lemma 3.4.
Proof of Corollary 1.3. The result is immediate from Theorem 1.2.
Next we restate the Rees Theorem (see [3, Theorem 3.2.3]) for the case of finite
combinatorial 0-simple semigroups:
Theorem 3.6 (Rees Theorem). Let P be a finite 0-1 matrix such that each row
and each column has at least one 1. Then SP is a finite combinatorial 0-simple
semigroup.
Conversely, every finite combinatorial 0-simple semigroup is isomorphic to one
constructed in this way.
Proof. See [3].
Lemma 3.7. Let SP be a finite combinatorial 0-simple semigroup. Then the matrix
P has one block if and only if SP has no zero divisors, i.e. for s, t ∈ SP , st = 0
implies that s = 0 or t = 0.
Proof. Assume P ∈ {0, 1}Λ×I . If P has one block, then all entries of P are 1. Thus
SP has no zero divisors. If P does not have one block, then P (λ, i) = 0 for some
λ ∈ Λ, i ∈ I. Now [i, λ] is a zero divisor since [i, λ]2 = 0.
Corollary 3.8. If a finite combinatorial 0-simple semigroup S has no zero divisors,
then SMP(S) is in P. Otherwise SMP(S) is NP-complete.
Proof. The result is immediate from Theorem 1.2 and Lemma 3.7.
4. Semigroups with PSPACE-complete SMP
In [1] an upper bound on the complexity of the SMP for semigroups was established:
Theorem 4.1 ([1, Theorem 2.1]). The SMP for a finite semigroup is in PSPACE.
Proof. Let S be a finite semigroup. We show that
(7)
SMP(S) is in nondeterministic linear space.
To this end, let A ⊆ S n , b ∈ S n be an instance of SMP(S). If b ∈ hAi, then there
exist a1 , . . . , am ∈ A such that b = a1 · · · am .
ON SEMIGROUPS WITH PSPACE-COMPLETE SMP
9
Now we pick the first generator a1 ∈ A nondeterministically and start with
c := a1 . Pick the next generator a ∈ A nondeterministically, compute c := c · a,
and repeat until we obtain c = b. Clearly all computations can be done in space
linear in |A|n. This proves (7). By a result of Savitch [8] this implies that SMP(S)
is in deterministic quadratic space.
In [1] it was shown that the SMP for the full transformation semigroup on 5
letters is PSPACE-complete by reducing Q3SAT to SMP(T5 ). We adapt the proof
of this result and show that under the following conditions the SMP for a semigroup
is PSPACE-complete.
Lemma 4.2. Let S be a finite semigroup and s, t, 1 ∈ S such that
(a) sts = s, tst = t,
(b) s2 , t2 <J s,
(c) s1 = s and t1 = t.
Then SMP(S) is PSPACE-complete.
Proof. From item (a) we know that s, t, st, ts are in the same J -class. Observe that
s 6= st; otherwise s2 = sts = s, which is impossible. We consider s and st as states
and let 1, s, t, st, ts act on these states by multiplication on the right. This yields
the partial multiplication table
t st ts
S 1 s
s s ∞ st ∞ s
st st s ∞ st ∞
(8)
where ∞ means that this entry is <J s.
SMP(S) is in PSPACE by Theorem 4.1. For PSPACE-hardness we reduce
Q3SAT to SMP(S). Q3SAT is PSPACE-complete [7] and can be defined as follows.
Q3SAT
Input:
triples C1 , . . . , Cm over
{x1 , . . . , xn , ¬x1 , . . . , ¬xn , y1 , . . . , yn , ¬y1 , . . . , ¬yn }
Problem: Is the Boolean formula W
W
Φ := ∀x1 ∃y1 . . . ∀xn ∃yn ( C1 ) ∧ . . . ∧ ( Cm ) true?
Let C1 , . . . , Cm be a Q3SAT instance and Φ be the corresponding Boolean formula.
We refer to x1 , . . . , xn as universal variables, to y1 , . . . , yn as existential variables,
and to C1 , . . . , Cm as clauses.
We define the corresponding SMP(S) instance
G ⊆ S 3n+m+1 , f ∈ S 3n+m+1
where
G := {a} ∪ B ∪ C ∪ D ∪ E,
B := {b1 , . . . , bn },
C := {ckj | j ∈ [m], k ∈ {+, −, 0}},
D := {djk | j ∈ [m], k ∈ [3]},
E := {e1 , . . . , en }.
10
MARKUS STEINDL
The coordinates of these tuples will have the following meaning. The first n positions encode truth values assigned to x1 , . . . , xn , and the second n positions truth
values assigned to y1 , . . . , yn . Truth values are represented by
0 := s
and 1 := st.
The positions 2n + 1 to 3n control the order in which the given tuples are
multiplied. The m positions after that indicate the status of W
the clauses C1 , . . . , Cm .
In particular, for i ∈ [m] position 3n + i encodes whether Ci is satisfied by the
assignment given by the first 2n positions. We encode
“unsatisfied” by s, and “satisfied” by st.
The last position ensures that the first generator of the target tuple f is the tuple
a. The generators are explicitly given as follows. For an overview see Figure 1.
• We define our target tuple f by
f (i) := st for i ∈ [3n + m],
f (3n + m + 1) := s.
• The tuple a will be the first generator of f . It encodes the all-zero assignment for the variables x1 , . . . , xn , y1 , . . . , yn . Let
a(i) := s
for i ∈ [3n + m + 1].
The idea is that a is multiplied only on the right by elements of G. The
components of this product are the states s and st. Each multiplication on
the right modifies the states according to the multiplication table (8).
• For j ∈ [n] let bj change the assignment for the universal variables
from (x1 , . . . , xj−1 , 0, 1, . . . , 1) to (x1 , . . . , xj−1 , 1, 0, . . . , 0).
For j ∈ [n] define
1
bj (i) := t
s
(
st
bj (2n + i) :=
s
if i ∈ [j − 1] or n + 1 ≤ i ≤ 2n,
if i = j,
if j < i ≤ n,
if i ∈ [j − 1],
if j ≤ i ≤ m + n,
bj (3n + m + 1) := ts.
−
• For j ∈ [n] let c+
j and cj change the assignment for the existential variable
yj from 0 to 1 and from 1 to 0, respectively. Let c0j leave the variables
unchanged. For j ∈ [n] and i ∈ [3n + m + 1] let
t if i = 2n + j,
0
cj (i) := st if i = 3n + m + 1,
1 otherwise.
−
0
The tuples c+
j and cj differ from cj only in the following positions:
c+
j (n + j) := t,
c−
j (n + j) := s.
ON SEMIGROUPS WITH PSPACE-COMPLETE SMP
11
• For j ∈ [m] and k ∈ [3] the tuple djk evaluates the kth literal Cjk of the jth
clause. If this literal is satisfied by the assignment encoded in the first 2n
components, then multiplying by djk changes the status of the clause Cj
to “satisfied”. This will be more formally stated in Claim 4.3. For i ∈ [n]
define
st if Cjk = xi ,
djk (i) := ts if Cjk = ¬xi ,
1 otherwise,
st if Cjk = yi ,
djk (n + i) := ts if Cjk = ¬yi ,
1 otherwise,
djk (2n + i) := st.
For i ∈ [m] let
djk (3n + i) :=
(
t
1
if i = j,
otherwise,
djk (3n + m + 1) := ts.
• Let j ∈ [n]. After each assignment was succesfully evaluated, the tuple ej
sets position n + j to st in order to match the target tuple if necessary. We
define
ej (i) := st, for i ∈ [n],
(
t if i = j,
ej (n + i) :=
1 if i ∈ [2n] \ {j},
ej (3n + i) := st
for i ∈ [m],
ej (3n + m + 1) := ts.
The tuples e1 , . . . , en will only occur as final generators of f .
Now we state what we already mentioned in the definition of djk .
Claim 4.3. Let h ∈ hGi such that
h(i) ∈ {s, st} for all i ∈ [2n],
h(2n + i) = st
for all i ∈ [n],
h(3n + j) = s
for some j ∈ [m],
h(3n + m + 1) = s.
Let ρ : {x1 , . . . , xn , y1 , . . . , yn } → {0, 1} be the assignment encoded by h|[2n] , i.e.
ρ(xi ) := h(i)
for i ∈ [n],
ρ(yi ) := h(n + i) for i ∈ [n].
Let k ∈ [3], Cjk be the kth literal of the jth clause, and h′ := hdjk .
(a) If ρ(Cjk ) = 1, then h and h′ differ only in the following position:
h′ (3n + j) = st.
(b) Otherwise h′ (i) <J s for some i ∈ [2n].
12
MARKUS STEINDL
a = ( s ···
b1 = ( t
..
(
.
bn = (
s
..
.
s
s ···
s
s
s
t
c+
1 =(
..
(
.
+
cn = (
t
c−
1 =(
..
(
.
−
cn = (
s
..
e1 = (
..
(
.
en = (
s
s ···
s
s ···
st . . .
st st
s
..
.
s ···
.. . .
.
.
s ···
s ts )
.. ..
. . )
s ts )
t
.
..
..
t
.
..
∗ ···
.. . .
.
.
∗ ···
∗
..
.
∗
st · · · st
.. . .
.
. ..
.
st · · · st
∗ ···
.. . .
.
.
∗ ···
t
..
..
s )
t
ts )
..
. )
ts )
t
ts )
..
. )
ts )
t
ts )
..
. )
ts )
.
s
t
s
.
t
c01 = (
..
(
.
c0n = (
d1k = (
..
(
.
dmk = (
s ···
.
∗ st · · · st
.
.. .. . .
. ..
. .
∗ st · · · st
t
..
.
t
ts )
..
. )
ts )
st · · · st ts )
.. . .
. .
. .. .. )
.
st · · · st ts )
.
t
f = ( st · · · st st · · · st st · · · st st · · · st s )
| {z } | {z } | {z } | {z }
n
n
n
m
Figure 1. Generators and target tuple of the SMP(T ) instance
for k ∈ [3]. Empty positions encode the element 1, and the symbol
“∗” indicates that this entry depends on the Q3SAT instance.
Proof. The literal Cjk is of the form z or ¬z for some variable z. Let ℓ ∈ [2n] be
the position of z in (x1 , . . . , xn , y1 , . . . , yn ). First assume Cjk is of the form z. Then
djk (ℓ) = st.
(a) Assume ρ(Cjk ) = 1. This means h(ℓ) = st. Thus h′ (ℓ) = st · st = st = h(ℓ),
and h′ (3n + j) = s · t = st. It is easy to see that h(i) = h′ (i) for the remaining
positions i ∈ [3n + m + 1] \ {ℓ, 3n + j}.
(b) Assume ρ(Cjk ) = 0. This means h(ℓ) = s. Thus h′ (ℓ) = s · st <J s, and (b)
is proved.
If Cjk is of the form ¬z, then (a) and (b) are proved in a similar way.
Note that if ρ satisfies Cjk , then multiplying h by djk changes the status of the
jth clause from “unsatisfied” to “satisfied”. Otherwise the target tuple f cannot be
reached by further multiplying hdjk with elements of G.
In the remainder of the proof we show the following.
ON SEMIGROUPS WITH PSPACE-COMPLETE SMP
13
Claim 4.4. Φ holds if and only if f ∈ hGi.
(⇒) direction of Claim 4.4. Assume Φ is true. This means that for every
i ∈ [n] there is a function ψi : {0, 1}i → {0, 1} such that for every assignment
ϕ : {x1 , . . . , xn } → {0, 1} the assignment
ρϕ := ϕ ∪ {yi 7→ ψi (ϕ(x1 ), . . . , ϕ(xi )) | i ∈ [n]}
satisfies all the clauses C1 , . . . , Cm .
We prove by induction on assignments ϕ in lexicographic order that for each ϕ
the following tuple gϕ belongs to hGi:
gϕ (i) := ϕ(xi )
for i ∈ [n],
gϕ (n + i) := ρϕ (yi ) for i ∈ [n],
gϕ (2n + i) := st
for i ∈ [n + m],
gϕ (3n + m + 1) := s.
For the base case let ϕ(xi ) := 0 for all i ∈ [n]. For i ∈ [n] we define
(
c+
if ρϕ (yi ) = 1,
′
i
ci :=
0
ci otherwise.
Apparently a · c′1 · · · c′n |[2n] encodes ρϕ . For each j ∈ [m] there is a kj ∈ [3] such
that the literal Cjkj is satisfied by ρϕ . By Claim 4.3 (a) it is straightforward to
verify that
gϕ = a · c′1 · · · c′n · d1k1 · · · dmkm .
Now let ϕ be an assignment with successor ϕ′ in lexicographical order such that
gϕ ∈ hGi. Let j ∈ [n] be maximal such that ϕ(xj ) = 0. Then
ϕ(xi ) = ϕ′ (xi )
for i < j,
′
ϕ(xj ) = 0, ϕ (xj ) = 1,
ϕ(xi ) = 1, ϕ′ (xi ) = 0
To adjust the assignment for the
+
ci
c′i := c−
i
0
ci
for j < i ≤ n.
existential variables, for j ≤ i ≤ n set
if ρϕ (yi ) = 0, ρϕ′ (yi ) = 1,
if ρϕ (yi ) = 1, ρϕ′ (yi ) = 0,
otherwise.
For h := gϕ · bj · c′j · · · c′n we have
h(i) = ϕ′ (xi )
for i ∈ [n],
h(n + i) = ρ (yi ) for i ∈ [n],
ϕ′
h(2n + i) = st
for i ∈ [n],
h(3n + i) = s
for i ∈ [m + 1].
For each j ∈ [m] the clause Cj is satisfied by ρϕ′ . Thus there is a kj ∈ [3] such that
ρϕ′ satisfies the literal Cjkj . From Claim 4.3 (a) follows that
gϕ′ = h · d1k1 · · · dmkm .
This completes the induction argument.
14
MARKUS STEINDL
Finally let ϕ be such that ϕ(xi ) = 1 for all i ∈ [n], and gϕ as defined above.
Denote the positions i ∈ [n] where gϕ (n + i) = s by i1 , . . . , ip . Then we have
f = gϕ · ei1 · · · eip .
Thus f ∈ hGi. The (⇒) direction of Claim 4.4 is proved.
We give another description of the product that yields f . For each assignment
ϕ 6= 0 for x1 , . . . , xn let jϕ := max{j ∈ [n] | ϕ(xj ) = 1}. From our argument above
we see that f is of the form
Y
(9)
f = ac∗1 · · · c∗n d1† · · · dm† · ( bjϕ c∗jϕ · · · c∗n d1† · · · dm† ) · ei1 · · · eip ,
ϕ6=0
where i1 , . . . , ip ∈ [n] are distinct, each ∗ belongs to {+, −, 0}, and each † to [3].
The product is taken over all assignments ϕ 6= 0 to x1 , . . . , xn in lexicographical
order.
(⇐) direction of Claim 4.4. Assume f ∈ hGi. Let k ∈ N be minimal such that
f = u1 · · · uk for some u1 , . . . , uk ∈ G, and let vi := u1 · · · ui for i ∈ [k].
Claim
(a)
(b)
(c)
4.5. Let i ∈ {2, . . . , k} and j ∈ [3n + m + 1]. Then
u1 = a and ui 6= a,
vi (j) ∈ {s, st}.
If ui (j) ∈ {1, st, ts}, then vi (j) = vi−1 (j).
Proof. (a) If ui = a, then ui−1 ui (3n + m + 1) ∈ {ts2 , s2 }, which yields the contadiction vk (3n + m + 1) <J s. If u1 6= a, then we obtain the contradiction
vk (3n + m + 1) = ts.
(b) We use induction on i. By (a) we have v1 = a. Thus v1 (j) ∈ {s, st}. Now
assume vi−1 (j) ∈ {s, st}. Since ui (j) ∈ {1, s, st, t, ts}, either vi (j) ∈ {s, st}, or a
factor s2 or t2 occurs in vi (j). As s2 , t2 <J f (j), the first case applies.
(c) is immediate from item (b) and the multiplication table (8).
The next claim states that the product u1 · · · uk is of a similar form to the one
given in (9).
Claim 4.6. Let i ∈ [k]. Let either j ∈ [n] and ui = bj , or j = 1 and ui = a. Then
for i1 := i + n − j + 1 and i2 := i1 + m the following holds:
p
(a) {ui+1 , . . . , ui1 } = {cj j , . . . , cpnn } for some p1 , . . . , pm ∈ {+, −, 0}.
(b) {ui1 +1 , . . . , ui2 } = {d1k1 , . . . , dmkm } for some k1 , . . . , km ∈ [3].
(c) If there is a greatest j ′ ∈ [n] such that vi (j ′ ) = s, then ui2 +1 = bj ′ ;
(d) otherwise ui2 +1 , . . . , uk are distinct and form the set
{eℓ | ℓ ∈ [n], vi2 (n + ℓ) = s}.
Proof. First let j ∈ [n] and ui = bj .
(a) From Claim 4.5 we know that i ≥ 2 and that every coordinate of vi−1 is
either s or st. From the definition of bj and the multiplication table (8) we know
that
(
st for ℓ ∈ [j − 1],
vi (2n + ℓ) =
s for ℓ ∈ {j, . . . , n}.
Thus the only choice for the n − j + 1 generators subsequent to ui is given by
p
cj j , . . . , cpnn for some pj , . . . , pn ∈ {+, −, 0}, where the order does not matter; otherwise we would obtain a factor s2 or t2 in a position 2n + ℓ for some ℓ ∈ [n + m],
which is impossible.
ON SEMIGROUPS WITH PSPACE-COMPLETE SMP
15
(b) From (a) we know that
vi1 (2n + ℓ) = st for ℓ ∈ [n],
vi1 (3n + ℓ) = s
for ℓ ∈ [m].
Thus the m generators subsequent to ui1 are given by d1k1 , . . . , dmkm for some
k1 , . . . , km ∈ [3] where the order does not matter; otherwise we would obtain a
factor t2 in vk (2n + ℓ) for some ℓ ∈ [n], or s2 in vk (3n + ℓ) for some ℓ ∈ [m]. Both
cases contradict the fact that vk = f .
(c) Let j ′ ∈ [n] be maximal such that vi (j ′ ) = s. From (a) and (b) we know that
vi2 (j ′ ) = s
and
vi2 (2n + ℓ) = st for ℓ ∈ [n + m].
If ui2 +1 ∈ E, then vi2 +1 (j ′ ) = s2 t. If ui2 +1 ∈ C ∪ D, then vi2 +1 (2n + ℓ) = st2 for
some ℓ ∈ [n + m]. Thus ui2 +1 = bℓ for some ℓ ∈ [n]. If ℓ < j ′ , then vi2 +1 (j ′ ) = s2 .
If ℓ > j ′ , then vi2 +1 (ℓ) = st2 . Therefore ℓ = j ′ .
(d) Assume vi (ℓ) = st for all ℓ ∈ [n]. Suppose some generator among ui2 +1 , . . . , uk
belongs to B. Let i3 ∈ {i2 + 1, . . . , k} be minimal such that ui3 ∈ B. By (a), (b),
and Claim 4.5 (c) we have
vi |[n] = . . . = vi3 −1 |[n] .
2
/ B.
Thus vi3 (ℓ) = st for some ℓ ∈ [n], which is impossible. Hence ui2 +1 , . . . , uk ∈
This together with (a) and (b) implies
vi2 (2n + ℓ) = . . . = vk (2n + ℓ) = st for all ℓ ∈ [n + m].
Thus ui2 +1 , . . . , uk ∈
/ C ∪ D. Otherwise we would have a factor t2 in vk (n + ℓ). So
ui2 +1 , . . . , uk ∈ E. If ui2 +1 , . . . , uk were not distinct, then we had a factor t2 in
vk (n + ℓ) for some ℓ ∈ [n]. Finally observe that for each ℓ ∈ [n] with vi2 (n + ℓ) = s
we have eℓ ∈ {ui2 +1 , . . . , uk }; otherwise vk (n + ℓ) = s which is impossible. We
proved (d).
For j = 1 and ui = a, items (a) to (d) are proved in a similar manner.
In the following we define assignments to the variables using the first 2n positions
of the tuples v1 , . . . , vk . For i ∈ [k] and j ∈ [n] let
ϕi : {x1 , . . . , xn } → {0, 1}, ϕi (xj ) := vi (j),
θi : {y1 , . . . , yn } → {0, 1},
θi (yj ) := vi (n + j).
These assignments fulfill the following conditions.
Claim 4.7.
(a) For i ∈ [k − 1] we have ϕi 6= ϕi+1 if and only if ui+1 ∈ B.
(b) ϕ1 , . . . , ϕk is a list of all assignments for x1 , . . . , xn (possibly with repetitions) in lexicographic order.
(c) Let i ∈ [k − 1] such that ui ∈ D and ui+1 ∈
/ D. Then ϕi ∪ θi satisfies all
the clauses C1 , . . . , Cm .
Proof. (a) follows from the definitions of the generators, Claim 4.5 (c), and Claim 4.6.
(b) Let i ∈ [k − 1] such that ϕi 6= ϕi+1 . By (a) and Claim 4.6 ui+1 = bj for the
greatest j ∈ [n] for which ϕi (xj ) = 0. It is easy to see that ϕi+1 (xℓ ) = ϕi (xℓ ) for
ℓ < j, ϕi+1 (xj ) = 1, and ϕi+1 (xℓ ) = 0 for ℓ > j. Thus ϕi+1 is the successor of ϕi in
lexicographic order. By Claim 4.5 (a) ϕ1 is the all-zero assignment for x1 , . . . , xn .
16
MARKUS STEINDL
Since vk (i) = st for all i ∈ [n], ϕk is the all-one assignment. Hence ϕ1 , . . . , ϕk is a
list of all assignments.
(c) By Claim 4.6 {ui−m+1 , . . . , ui } = {d1k1 , . . . , dmkm } for some k1 , . . . , km ∈ [3].
Thus ϕi−m ∪ θi−m = . . . = ϕi ∪ θi . Suppose ϕi ∪ θi does not satisfy some clause
Cj for j ∈ [m]. Then its kth literal Cjkj is also unsatisfied. Claim 4.3 (b) implies
vi (ℓ) <J s for some ℓ ∈ [2n], which is impossible.
For ϕ : {x1 , . . . , xn } → {0, 1} let
iϕ := max{i ∈ [k] | ϕi = ϕ, ui ∈ D}.
From Claim 4.7 we know that for every assignment ϕ to x1 , . . . , xn the assignment
ϕ ∪ θiϕ satisfies all the clauses of Φ. It only remains to prove that θiϕ (yi ) only
depends on ϕ(x1 ), . . . , ϕ(xi ).
Claim 4.8. Let i ∈ [n]. For all ϕ, χ : {x1 , . . . , xn } → {0, 1} the equations
(10)
ϕ(x1 ) = χ(x1 ), . . . , ϕ(xi ) = χ(xi )
imply θiϕ (yi ) = θiχ (yi ).
Proof. We consider ϕ as fixed and prove the implication for all χ ≥ ϕ by induction
in lexicographical order. The base case χ = ϕ is clear.
Now let χ ≥ ϕ be an assignment for which the implication holds, and assume
its successor χ′ fulfills (10). Since ϕ ≤ χ < χ′ , the assignment χ also fulfills (10).
From Claim 4.6 we know that
(11)
p
{uiχ +1 , . . . , uiχ′ } = {bj , cj j , . . . , cpnn , d1k1 , . . . , dmkm }
for some j ∈ [n], pj , . . . , pn ∈ {+, −, 0}, and k1 , . . . , km ∈ [3]. If j ≤ i was true,
then χ′ would not fulfill (10). Thus j > i. From (11) follows viχ′ (n+ i) = viχ (n+ i).
Thus θiχ′ (yi ) = θiχ (yi ) = θiϕ (yi ), and Claim 4.8 is proved.
We complete the proof of Claim 4.4. By Claims 4.7 and 4.8, for each assignment
ϕ for the universal variables there is an assignment θiϕ for the existential variables
such that ϕ ∪ θiϕ satisfies the conjunctive normal form in Φ. For all i ∈ [n] the
value θiϕ (yi ) depends only on ϕ(x1 ), . . . , ϕ(xi ). Thus Φ is true. Claim 4.4 and
Lemma 4.2 are proved.
Proof of Theorem 1.4. Let s′ := (s, tst), t′ := (tst, s), and 1′ := (1, 1) be elements
of S 2 := S × S. Apparently s′ t′ s′ = s′ and t′ s′ t′ = t′ . Both s′ and t′ do not
generate groups. By Lemma 2.1 s′2 <J s′ and t′2 <J t′ . Since t′ J s′ , we have
t′2 <J s′ . Now s′ , t′ , 1′ fulfill the hypothesis of Lemma 4.2. Thus SMP(S 2 ) is
PSPACE-complete. As SMP(S 2 ) reduces to SMP(S) and conversely, the result
follows.
Now we are able to list several “naturally occuring” semigroups with PSPACEcomplete SMP:
Corollary 4.9. The SMP for the following semigroups is PSPACE-complete:
(a) the Brandt monoid B21 and the monoid A12 ;
(b) for n ≥ 2 and a finite ring R with identity 1 6= 0, the semigroup of all n × n
matrices over R;
(c) the full transformation semigroup Tn on n ≥ 3 letters;
(d) the symmetric inverse semigroup In on n ≥ 2 letters.
ON SEMIGROUPS WITH PSPACE-COMPLETE SMP
17
Proof. We apply Theorem 1.4.
(a) For B21 let s := [1, 2] and t := [2, 1]. For A12 let s := [2, 2] and t := [1, 1].
(b) Define n × n matrices s, t over R by
(
(
1 if (i, j) = (1, 2),
1 if (i, j) = (2, 1),
sij :=
tij :=
0 otherwise,
0 otherwise
for i, j ∈ [n]. Let 1 be the identity matrix.
(c) Let 1 be the identity mapping on [n], and s, t : [n] → [n],
(
(
2 if x = 1,
1 if x = 2,
s(x) :=
t(x) :=
3 otherwise,
3 otherwise.
(d) Let 1 be the identity mapping, s : 1 7→ 2, and t : 2 7→ 1.
For monoids we can now generalize Corollary 2.3:
Corollary 4.10. If a J -class of a finite monoid S contains both group and nongroup H-classes, then SMP(S) is PSPACE-complete.
Proof. Let S be as above. Similar to the proof of Corollary 2.3, there is a t ∈ S such
that sts = s. Now s, t, and the identity fulfill the hypothesis of Theorem 1.4.
5. Proof of Theorem 1.6
Lemma 5.1. If the 0-1 matrix P of a finite combinatorial Rees matrix semigroup
SP has one block, then SMP(SP1 ) is in NP.
Proof. Assume P ∈ {0, 1}Λ×I , and let J ⊆ I and ∆ ⊆ Λ such that P (λ, i) = 1 if
and only if (λ, i) ∈ ∆ × J for i ∈ I, λ ∈ Λ. Let T := SP1 and A ⊆ T n , b ∈ T n be an
instance of SMP(T ) such that b ∈ hAi. Let a1 , . . . , ak ∈ A such that b = a1 · · · ak .
If b = (1, . . . , 1) or k = 1, then clearly b ∈ A. In this case the position of b in the
list A is a witness. Assume b 6= (1, . . . , 1) and k ≥ 2.
We claim that for i ∈ [n] with b(i) = 0 there are ℓi , ri ∈ [k], ℓi < ri such that
(12)
aℓi ari (i) = 0 and aℓi +1 (i) = . . . = ari −1 (i) = 1.
This follows from Lemma 3.1 (a). For i ∈ [n] with b(i) ∈ I × Λ let
ℓi := min{j ∈ [k] | aj (i) 6= 1},
ri := max{j ∈ [k] | aj (i) 6= 1}.
Now define an index set N ⊆ [k] by
N := {ℓi | i ∈ [n], b(i) 6= 1} ∪ {ri | i ∈ [n], b(i) 6= 1}.
Note that N 6= ∅; otherwise b = (1, . . . , 1) which contradicts our assumption.
For i ∈ [n] we claim that
Y
(13)
aj (i) = b(i),
j∈N
where the indexes j of the factors are in ascending order. If b(i) = 1, then aj (i) = 1
for all j ∈ [k], and (13) follows. Assume b(i) = 0. We have ℓi , ri ∈ N . By (12)
all factors in (13) between aℓi (i) and ari (i) are equal to 1. This and (12) imply
18
MARKUS STEINDL
(13). Finally assume b(i) ∈ I × Λ. For ℓi < j < ri we have aj (i) ∈ {1} ∪ (J × ∆);
otherwise we obtain the contradiction b(i) = 0. Thus
Y
Y
aj (i) =
aj (i).
j∈N
ℓi ≤j≤ri
ℓi ≤j≤ri
Since aj (i) = 1 for j < ℓi and j > ri , (13) follows.
The length of the product in (13) is |N | and thus at most 2n. Thus this product
is a valid witness for b ∈ hAi, and the lemma is proved.
Proof of Theorem 1.6. Assume P ∈ {0, 1}Λ×I .
(a) If P is the all-1 matrix, then SP1 is a band (idempotent semigroup) with
J -classes {0}, I × Λ, and {1}. We show that SP1 is a regular band, that is, SP1
satisfies the identity
(14)
xyxzx = xyzx.
SP1 .
Let x, y, z ∈
If one of the variables is 0 or 1, then (14) clearly holds. If
x, y, z ∈ I × Λ, then xyxzx = x = xyzx by the definition of the multiplication.
Thus SP1 is a regular band. By [10, Corollary 1.7] the SMP for every regular band
is in P.
(b) Assume P has one block and some entries are 0. Let i ∈ I and λ ∈ Λ such
that P (λ, i) = 0. Let s := [i, λ] and r := t := 1. Since s does not generate a group,
SMP(SP1 ) is NP-hard by Theorem 1.1. NP-easiness follows from Lemma 5.1.
(c) In this case P does not have one block. Thus there are i, j ∈ I and λ, µ ∈ Λ
such that
P (λ, i) = P (µ, j) = 1 and P (λ, j) = 0.
Let s := [j, λ] and t := [i, µ]. Then s does not generate a group, sts = s, s1 = s,
and t1 = t. By Theorem 1.4 SMP(S) is PSPACE-complete.
6. Conclusion
In Section 3 we established a P/NP-complete dichotomy for combinatorial Rees
matrix semigroups. The next goal is to investigate the complexity for the more
general case of Rees matrix semigroups. For Rees matrix semigroups without 0
a polynomial time algorithm for the SMP is known [11]. However, the following
questions are open:
Problem 6.1. Is the SMP for finite Rees matrix semigroups (with 0) in NP? In
particular, is there a P/NP-complete dichotomy?
In Section 4 we saw the first example of a semigroup with NP-complete SMP
where adjoining an identity results in a PSPACE-complete SMP. This leads to the
following question:
Problem 6.2. How hard is the SMP for finite Rees matrix semigroups with adjoined identity?
The answer is not even known for the completely regular case. E.g. the complexity
for the following 9-element semigroup is open:
Problem 6.3. Let 1, c be the elements of the cyclic group Z2 such that c2 = 1, and
define a Rees matrix semigroup S := M(Z2 , [2], [2], ( 11 1c )). How hard is SMP(S 1 )?
ON SEMIGROUPS WITH PSPACE-COMPLETE SMP
19
References
[1] A. Bulatov, P. Mayr, and M. Steindl. The subpower membership problem for semigroups,
submitted. Available at http://arxiv.org/pdf/1603.09333v1.pdf.
[2] S. A. Cook. Characterizations of pushdown machines in terms of time-bounded computers.
J. Assoc. Comput. Mach., 18:4–18, 1971.
[3] J. M. Howie. Fundamentals of semigroup theory, volume 12 of London Mathematical Society
Monographs. New Series. The Clarendon Press, Oxford University Press, New York, 1995.
Oxford Science Publications.
[4] P. Idziak, P. Marković, R. McKenzie, M. Valeriote, and R. Willard. Tractability and learnability arising from algebras with few subpowers. SIAM J. Comput., 39(7):3023–3037, 2010.
[5] M. Kozik. A finite set of functions with an EXPTIME-complete composition problem. Theoretical Computer Science, 407(1–3):330–341, 2008.
[6] P. Mayr. The subpower membership problem for Mal’cev algebras. Internat. J. Algebra Comput., 22(7):1250075, 23, 2012.
[7] C. H. Papadimitriou. Computational complexity. Addison-Wesley Publishing Company, Reading, MA, 1994.
[8] W. J. Savitch. Relationships between nondeterministic and deterministic tape complexities.
J. Comput. System. Sci., 4:177–192, 1970.
[9] S. Seif and C. Szabó. Computational complexity of checking identities in 0-simple semigroups
and matrix semigroups over finite fields. Semigroup Forum, 72(2):207–222, 2006.
[10] M. Steindl. The subpower membership problem for bands, submitted. Available at
http://arxiv.org/pdf/1604.01014v1.pdf.
[11] M. Steindl. Computational Complexity of the Subpower Membership Problem for Semigroups.
PhD thesis, Johannes Kepler University Linz, Austria, 2015. Available at
http://epub.jku.at/obvulihs/download/pdf/893649?originalFilename=true.
[12] R. Willard. Four unsolved problems in congruence permutable varieties. Talk at International
Conference on Order, Algebra, and Logics, Vanderbilt University, Nashville, June 12–16, 2007.
Institute for Algebra, Johannes Kepler University Linz, Altenberger St 69, 4040
Linz, Austria
Department of Mathematics, University of Colorado Boulder, Campus Box 395,
Boulder, Colorado 80309-0395
E-mail address: [email protected]
| 4math.GR
|
1
Ultra-Reliable and Low Latency Communication in
mmWave-Enabled Massive MIMO Networks
Trung Kien Vu, Student Member, IEEE, Chen-Feng Liu, Student Member, IEEE,
Mehdi Bennis, Senior Member, IEEE, Mérouane Debbah, Fellow, IEEE, Matti Latva-aho, Senior Member, IEEE,
and Choong Seon Hong, Senior Member, IEEE
arXiv:1705.06907v1 [cs.IT] 19 May 2017
Abstract—Ultra-reliability and low-latency are two key components in 5G networks. In this letter, we investigate the problem
of ultra-reliable and low-latency communication (URLLC) in millimeter wave (mmWave)-enabled massive multiple-input multipleoutput (MIMO) networks. The problem is cast as a network
utility maximization subject to probabilistic latency and reliability
constraints. To solve this problem, we resort to the Lyapunov technique whereby a utility-delay control approach is proposed, which
adapts to channel variations and queue dynamics. Numerical
results demonstrate that our proposed approach ensures reliable
communication with a guaranteed probability of 99.99%, and
reduces latency by 28.41% and 77.11% as compared to baselines
with and without probabilistic latency constraints, respectively.
Index Terms—5G, massive MIMO, mmWave, ultra-reliable low
latency communications (URLLC).
I. I NTRODUCTION
C
URRENTLY, millimeter wave (mmWave) and massive
multiple-input multiple-output (MIMO) techniques are
investigated to provide reliable communication with an overthe-air latency of few milliseconds and extreme throughput [1].
While massive MIMO with large degrees of freedom provides
high energy and spectral efficiency [2], mmWave frequency
bands provide large bandwidth [3]. In addition, due to the short
wavelength of mmWaves, large antenna array can be packed
into highly directional beamforming, which makes massive
MIMO practically feasible [4]. Thus far, most of existing works
on mmWave-enabled massive MIMO systems focus mainly
on providing capacity improvement [4], while latency and
reliability are not addressed. Although latency and reliability are
applicable to many scenarios (e.g. mission-critical applications),
in this work, we are interested in the integration of mmWave
communication and massive MIMO techniques, which holds the
promise of providing great enhancements of the overall system
performance [1], [2], [4]. Specifically, this letter is concerned
with addressing the fundamental question in mmWave-enabled
massive MIMO systems: “how to simultaneously provide order
of magnitude capacity improvements and latency reduction?”
Manuscript received March 24, 2017; revised May 05, 2017; accepted May
12, 2017. Date of publication May 17, 2017; This work was supported in part
by the Finnish Funding Agency for Technology and Innovation (Tekes), Nokia,
Huawei, Anite, in part by the Academy of Finland CARMA project, in part
by the Academy of Finland funding through the grant 284704, and in part
by the ERC Starting Grant 305123 MORE (Advanced Mathematical Tools for
Complex Network Engineering).
T. K. Vu, C.-F. Liu, M. Bennis, and M. Latva-aho are with the Centre for Wireless Communications, University of Oulu, Oulu 90014, Finland
(e-mail: [email protected]; [email protected]; [email protected];
[email protected]).
M. Debbah is with the Large Networks and System Group (LANEAS),
CentraleSupélec, Université Paris-Saclay, Gif-sur-Yvette, France and is with
the Mathematical and Algorithmic Sciences Laboratory, Huawei France R&D,
Paris, France (e-mail: [email protected]).
C. S. Hong is with the Department of Computer Engineering, Kyung Hee
University, Yongin 446-701, South Korea (email: [email protected]).
Citation information: DOI 10.1109/LCOMM.2017.2705148, IEEE Communications Letters
By invoking the Lyapunov optimization framework, an utilityoptimal solution is obtained to maximize network throughput
subject to queuing stability [5]. This solution establishes a
utility-delay tradeoff, which achieves utility-optimality at the
price of large queuing delays [5]. To cope with this shortcoming, in this letter the Lyapunov framework is extended
to incorporate probabilistic latency and reliability constraints,
which takes into account queue length, arrival rate, and channel
variations with a guaranteed probability. To do so, the problem is formulated as a network utility maximization (NUM).
By applying the drift-plus-penalty technique, the problem is
decoupled into a dynamic latency control and rate allocation.
Here, the latency control problem is a difference of convex
(DC) programming problem, which is solved efficiently by the
convex-concave procedure (CCP) [6]. Finally, a performance
evaluation is carried out to show the latency reduction and the
tradeoff between reliability, traffic intensity, and user density.
II. S YSTEM M ODEL
Consider the downlink (DL) transmission of a single cell
massive MIMO system1 consisting of one macro base station
(MBS) equipped with N antennas, and a set, M = {1, . . . , M },
of single-antenna user equipments (UEs). We assume that
N ≥ M and N ≫ 1. Further, the co-channel time-division
duplexing (TDD) is considered in which the MBS estimates
channels via the uplink phase. We denote the propagation
chan√
1/2
nel between the MBS and the mth UE as hm = N Θm h̃m ,
where Θm ∈ CN ×N depicts the antenna spatial correlation, and the elements of h̃m ∈ CN ×1 are independent and
identically distributed (i.i.d.) with zero mean and variance
1/N . In addition, channels experience flat and block fading,
and imperfect channel state information (CSI) is assumed.
As per [9], the estimated
channel can be modeled as ĥm =
p
√
1/2
N ×1
2h +τ
1 − τm
m
m N Θm zm , ∀ m ∈ M. Here, zm ∈ C
denotes the estimated noise vector which has i.i.d. elements
with zero mean and variance 1/N , and τm ∈ [0, 1] reflects the
estimation accuracy; in case of perfect CSI, τm = 0. Given
the estimated channel matrix Ĥ = [ĥ1 , · · · , ĥM ] ∈ CN ×M ,
the MBS utilizes the regularized zero-forcing2 (RZF) precoder
with the precoding matrix, V = [v1 , · · · , vM ] ∈ CN ×M ,
−1 †
Ĥ . Note that
which is given by V = Ĥ† Ĥ + N αIN
vm is the precoding vector for UE m. In V, † denotes the
conjugate transpose, and the regularization parameter α > 0
is scaled by N to ensure the matrix Ĥ† Ĥ + N αIN is wellconditioned as N → ∞ [7]. Further, transmit power pm ≥ 0
is allocated to UE m. Denoting all allocated powers in the
diagonal matrix P = diag(p1 , · · · , pM ), we have the constraint
1 Our model can be extended to multi-cell massive MIMO systems in
which the problem of inter-cell interference can be addressed by designing a
hierarchical precoder at the MBS [7], to mitigate both intra-cell and inter-cell
interference, or by applying an interference coordination approach [8].
2 Other hybrid beamforming designs are left for future works.
1089-7798 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
2
Tr PV† V ≤ P , with P the maximum transmit power of the
MBS. With the aid of the results in [9, Theorem 1], the transmit
power allocation constraint can be expressed as
1
N
M
P
m=1
pm
Ωm
m
≤ P, and pm ≥ 0, ∀ m ∈ M,
(1)
where the parameter Ωm is the solution to Ωm =
PM
−1
Θm
1
1
. By designing the prem=1 α+Ωm + IN
N Tr Θm N
coding matrix V and transmit power P, the ergodic DL rate of
UE m ∈ M is
i
h
p |h† v |2
(2)
rm (P) = E log2 1 + 1+PM m m pm|h† v |2 .
k=1,k6=m
k
m
k
Here, we invoke results from random matrix theory in order
to get the deterministic equivalence for (2) [9]. In particular,
as N ≥ M and N ≫ 1, for small α, the ergodic DL rate (2)
almost surely converges to
a.s.
2
) , ∀ m ∈ M,
(3)
rm (P) −−→ log2 1 + pm (1 − τm
a.s.
where −−→ denotes almost sure convergence [7], [9, Theorem
2]. Moreover, we assume that the MBS has queue buffers to
store UE data [5]. In this regard, we first index the coherence
time block by slot t ∈ Z+ . At the beginning of each slot t, the
queue length for UE m is denoted by Qm (t) which evolves as
follows:
Qm (t + 1) = [Qm (t) − rm (t)]+ + am (t), ∀ m ∈ M,
(4)
where [x]+ , max{x, 0}, and am (t) is the data arrival rate
of UE m. Further, we assume that am (t) is i.i.d. over time slots
with mean arrival rate λm and upper bounded by amax
[5].
m
III. P ROBLEM F ORMULATION
According to Little’s law
PT [10], the average delay is proportional to limT →∞ T1 t=1 E[Qm (t)]/λm . Thus, we use
Qm (t)/λm as a delay measure and enforce an allowable upper
bound dth
m . We further note that the delay (or queue length)
bound violation is related to reliability. Thus, taking into
account the latency and reliability requirements, we characterize the delay bound violation with a tolerable probability.
Specifically, we impose a probabilistic constraint on the queue
size length for UE m ∈ M as follows:
n
o
Pr Qλmm(t) ≥ dth
(5)
m ≤ ǫm , ∀ t.
In (5), dth
m reflects the UE delay requirement. Here, ǫm ≪ 1 is
the target probability for reliable communication.
In order to reduce latency, the intuitive way is to send
as many data as possible. However, this might over-allocate
resources to UEs, i.e., rm (t) ≫ Qm (t). To handle this issue,
max
we enforce a maximum rate constraint rm
for each UE m.
On the other hand, we enforce the MBS to guarantee for all
UEs a certain level of QoS, i.e., the minimum rate requirement
min
rm
, ∀ m ∈ M.
PM
We define the network utility as m=1 ωm f (r̄m ) where the
PT
time average expected rate r̄m = limT →∞ T1 t=1 E[rm (t)]
and the non-negative weight ωm for each UE m. Additionally,
we assume that f (·) is a strictly concave, increasing, and twice
continuously-differentiable function. Taking into account these
constraints presented above yields the following network utility
maximization:
M
P
OP :
max
ωm f (r̄m )
(6a)
P(t)
subject to
Our main problem involves a probabilistic constraint (5), which
cannot be addressed tractably. To overcome this challenge,
we apply Markov’s inequality [11] to linearize (5) such that
E[Qm (t)]
Pr Qλmm(t) ≥ dth
m ≤ λm dth . Then, (5) is satisfied if
m=1
min
rm
max
≤ rm (t) ≤ rm
, ∀ m ∈ M, ∀ t, (6b)
(1) and (5).
E[Qm (t)] ≤ λm dth
m ǫm , ∀ m ∈ M, ∀ t.
(7)
Thereafter, we consider (7) to represent the latency and reliability constraint. Assuming that {am (t)|∀ t ≥ 1} is aPPoisson art
rival process [11], we have that E[Qm (t)] = tλm − τ =1 rm (τ )
which is plugged into (7). Subsequently, we obtain
rm (t) ≥ tλm − λm dth
m ǫm −
t−1
P
τ =1
rm (τ ), ∀ m ∈ M, ∀ t, (8)
which represents the minimum rate requirement in slot t for
UE m for reliable communication. Here, we transform the
probabilistic latency and reliability constraint (5) into one linear
constraint (8) of instantaneous rate requirements, which helps to
analyse and optimize the URLLC problem. In particular, if the
delay requirement/reliability constraint is looser (i.e., larger dth
m
or ǫm ), the instantaneous rate requirement is reduced. In contrast, if we have a tighter constraint for reliable communication
or delay requirement, then the instantaneous rate requirement
is higher. Combining (6b) and (8), we rewrite OP as follows:
max
P(t)
subject to
M
P
ωm f (r̄m )
m=1
0
rm
(t)
(9a)
max
≤ rm (t) ≤ rm
, ∀ m ∈ M, ∀ t,
(9b)
and (1),
min
0
, tλm − λm dth
(t) = max{rm
with rm
m ǫm −
Pt−1
τ =1 rm (τ )}.
IV. LYAPUNOV O PTIMIZATION F RAMEWORK
To tackle (9), we resort to Lyapunov optimization techniques
[5]. Firstly, for each DL rate rm (t), we introduce the auxiliary
variable vector ϕ(t) = (ϕm (t)|∀ m ∈ M) that satisfies
T
P
1
E ϕm (t) ≤ r̄m , ∀ m
T
T →∞
t=0
max
ϕ0m (t) ≤ ϕm (t) ≤ rm
, ∀ m ∈ M, ∀ t,
ϕ̄m = lim
∈ M,
(10)
(11)
Pt−1
min
with ϕ0m (t) = max{rm
, tλm − λm dth
m ǫm −
τ =1 ϕm (τ )}.
Incorporating the auxiliary variables, (9) is equivalent to
LP :
max
P(t),ϕ(t)
subject to
T P
M
P
1
ωm E[f (ϕm (t))]
T
T →∞
t=1 m=1
lim
(1), (10), and (11).
In order to ensure the inequality constraint (10), a virtual queue
vector Y(t) = (Ym (t)|∀ m ∈ M) is introduced, where each
element evolves according to
Ym (t + 1) = [Ym (t) + ϕm (t) − rm (t)]+ , ∀ m ∈ M.
(12)
Subsequently, we express the conditional Lyapunov drift-pluspenalty for each time slot t as:
M h
P 1
1
2
2
E
2 Ym (t + 1) − 2 Ym (t)
m=1
i
− νm (t)wm f (ϕm (t)) Y(t) . (13)
In (13), νm (t) is the control parameter which affects the utilityqueue length tradeoff. This control parameter is conventionally
chosen to be static and identical for all UEs [5]. However,
this setting does not hold for system dynamics (e.g., instantaneous data arrivals) and the diverse system configuration
3
(i.e., different delay and QoS requirements). Thus, we dynamically design these control parameters. From the analysis
in the Lyapunov optimization framework [5], we can find
with πm being the largest firstYm (t) ≤ νm (t)ωm πm + amax
m
order derivative of f (x). Letting ωm = 1, ∀ m ∈ M, we have
0
the lower bound πm νm (t) ≥ νm
(t), ∀ m ∈ M, for selecting
0
the control parameters, where νm (t) = max{Ym (t) − amax
m , 1}.
Subsequently, following the straightforward calculations of the
Lyapunov drift-plus-penalty technique which are omitted for
space, we obtain
M
P
(13) ≤ E
Ym (t)ϕm (t) − νm (t)ωm f ϕm (t)
(14a)
m=1
M
P
m=1
i
Ym (t)rm P(t) + C Y(t) . (14b)
Due to space limitation, we omit the details of the constant
value C which does not influence the system performance [5].
We note that the solution to LP is acquired by minimizing
the right-hand side (RHS) of (14a) and (14b) in every slot t.
Further, (14a) is related to the reliability and QoS requirements
while (14b) reflects optimal power allocation to UEs.
A. Auxiliary Variable and Control Parameter Selection
ϕm (t), νm (t)
0
πm νm (t) ≥ νm
(t),
0
max
rm (t) ≤ ϕm (t) ≤ rm
.
(15b)
(15c)
Before proceeding with problem (15), we rewrite
−νm (t) log(ϕm (t)) in (15a), for any ϕm (t) > 0 and
νm (t) > 0, as
νm (t)
νm (t) log
− νm (t) log νm (t) ,
ϕm (t)
|
{z
}
|
{z
}
g0 (νm )
h0 (ϕm , νm )
in which both h0 (ϕm , νm ) (i.e., relative entropy function) and
g0 (νm ) (i.e., negative entropy function) are convex functions.
Since (15a) is the difference of convex functions while constraints (15b) and (15c) are affine functions, problem (15)
belongs to DC programming problems [12], which can be
efficiently and iteratively addressed by the CCP [6]. The CCP
algorithm to obtain the solution to problem (15) is detailed in
Algorithm 1. We note that the CCP provably converges to the
local optima of DC programming problems [6]. However, due to
space limitation, we omit the convergence proof of Algorithm 1
(please refer to [6] for the formal proof).
B. Power Allocation
The optimal transmit power in (14b) is computed by
min
P(t)
−
M
P
22
Proposed algorithm
Baseline 1
Baseline 2
20
18
16
14
12
10
dth=10 ms, ǫ = 5%
8
6
Considering the logarithmic fairness utility function, i.e.,
f (x) = log(x), minimizing the RHS of (14a) for each m ∈ M
is formulated as:
min
Ym (t)ϕm (t) − νm (t) log ϕm (t)
(15a)
subject to
ϕm ,νm
subject to (15b) and (15c),
(i)⋆
(i)⋆
Find the optimal ϕm and νm .
(i)⋆
(i+1)
:= νm and i := i + 1.
Update νm
until Convergence
end while
Average latency [ms]
−
Algorithm 1 CCP algorithm for solving sub-problem (15).
while m ∈ M do
(i)
Initialize i = 0 and a feasible point νm in (15b).
repeat
(i)
(i)
(i)
Convexify ĝ0 (νm , νm ) = g0 (νm ) + ∇g0 (νm − νm ).
Solve:
(i)
h0 (ϕm , νm ) − ĝ0 (νm , νm ) + Ym ϕm
min
Ym (t)rm (P(t))
m=1
subject to (1).
Here, the objective function is strictly convex for pm (t) ≥
0, ∀ m ∈ M, and the constraints are compact. Therefore, the
optimal solution of P⋆ (t) exists and is efficiently reached by
numerical methods.
After obtaining the optimal auxiliary variable and transmit
power, we update the queues Qm (t + 1) and Ym (t + 1) as per
(4) and (12), respectively.
4
2
1
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
Mean arrival rate λ [Gbps]
Fig. 1. Average latency versus mean arrival rates, M = 16 per km2 .
V. N UMERICAL R ESULTS
We consider a single-cell massive MIMO system3 in which
the MBS, with N = 32 antennas and P = 38 dBm, is located
at the center of the 0.5 × 0.5 km2 square area. UEs (from 8 to
60 UEs per km2 ) are randomly deployed within the MBS’s
coverage with a minimum MBS-UE distance of 35 m. Data
arrivals follow a Poisson distribution with different means, and
max
min
the rate requirements are specified as rm
= 1.2λm , rm
=
0.8λm , ∀ m ∈ M. The system bandwidth is 1 GHz. The path
loss is modeled as a distance-based path loss with the line-ofsight (LOS) model4 for urban environments at 28 GHz [13].
The maximum delay requirement dth and the target reliability
probability ǫ are set to 10 ms and 5%, respectively. The numerical results are obtained via Monte-Carlo simulations over
10000 realizations with different channel realizations and UE
locations. Furthermore, we compare our proposed scheme with
the following baselines:
• Baseline 1 refers to the Lyapunov framework in which the
probabilistic latency constraint (5) is considered.
• Baseline 2 is a variant of Baseline 1 without the probabilistic latency constraint (5).
A. Impact of Arrival Rate
In Fig. 1, we report the average latency versus the mean
arrival rates λ = E[a(t)] for M = 16. At low λ, all
schemes do not violate latency constraints, and our proposed
algorithm outperforms other baselines with a small gap. At
higher λ, the average delay of baseline 2 increases dramatically
as λ > 1.8 Gbps, since baseline 2 does not incorporate the
3 The multi-cell scenario raises a problem of additional delay due to the
need of information exchange among base stations, which is required by either
the coordination scheme or distributed approach. This problem is also a very
interesting open topic for future work.
4 We assume that the probability of LOS communication is very high, while
the impact of other channel models is left for future works.
1
3.5
Proposed algorithm
Baseline 1
Baseline 2
Proposed algorithm
Baseline 1
Baseline 2
Proposed algorithm
Baseline 1
Baseline 2
λ = 2 Gbps
0.8
CCDF
0.7
λ = 2.4 Gbps
0.6
λ = 2.6 Gbps
0.5
0.4
0.3
Proposed: Pr(delay > 7.4) < 1e-4
0.2
BL1: Pr(delay > 9.4) < 1e-4
0.1
0.05
0.0001
5
8 10
15
20
25
30
35
40
45
Latency [ms]
Fig. 2. Tail distribution (CCDF) of latency.
delay constraint, whereas our proposed scheme reduces latency
by 28.41% and 77.11% as compared to baselines 1 and 2,
respectively, when λ = 2.4 Gbps. When λ > 2.4 Gbps, the
average delay of all schemes increases, violating the delay
requirement of 10 ms. It can be observed that under limited
maximum transmit power, at very high traffic demand, the
latency requirement could not be guaranteed. This highlights
the tradeoff between the mean arrival rate and latency. In
Fig. 2, we report the tail distribution (complementary cumulative distribution function (CCDF)) of latency to showcase
how often the system achieves a delay greater than target
delay levels. In particular, at λ = 2.4 Gbps, by imposing
the probabilistic latency constraint (5), our proposed approach
and baseline 1 ensure reliable communication with better
guaranteed probabilities, i.e, Pr(delay > 7.5ms) < 10−4 and
Pr(delay > 9.4ms) < 10−4 , respectively. In contrast, baseline
2 violates the latency constraint with a high probability, where
Pr(delay > 10ms) = 74.75%.
B. Impact of User Density
In Fig. 3, we compare the average user throughput (avgUT)
and average latency of our proposed approach with the two
baselines under the impact of user density. Additionally, we
consider the weighted sum rate maximization (WSRM) case
without considering queue dynamics, i.e., problem (6) without
the constraints (5) and (6b). The WSRM case is the conventional way to find the system throughput limit but suffers from
higher latency. Since all users share the same resources, the
average delay (“solid lines”) increases with the number of users
M, whereas the avgUT (“dash lines”) decreases. Fig. 3 further
shows that when M > 24, the delay of all schemes increases
dramatically and is far-above the latency requirement. Hence,
only a limited number of users can be served to guarantee the
delay requirement, above which, a tradeoff between latency
and network density exists. Our proposed approach achieves
better throughput and higher latency reduction than baselines
1 and 2, while the WSRM case has the worst delay performance as expected. Compared with WSRM, our proposed
approach maintains at least 87% of the throughput limit, while
achieving up to 80% latency reduction. Moreover, our proposed
approach reaches Gbps capacity, which represents the capacity
improvement brought by the combination of mmWave and
massive MIMO techniques. Numerical results show that our
approach simultaneously provides order of magnitude capacity
improvements and latency reduction.
VI. C ONCLUSION
In this letter, we have investigated the problem of mmWaveenabled massive MIMO networks from a latency and reliability
Average user throughput [Gbps]
BL2: Pr(delay > 10) = 74.75%
0.9
3
WSRM
Proposed algorithm
Baseline 1
Baseline 2
Throughput
2.5
2
1.5
Latency
1
0.5
8
12
16
20
24
28
32
36
40
Number of nodes per km 2
44
48
52
56
160
150
140
130
120
110
100
90
80
70
60
50
40
30
20
10
0
60
Average latency [ms]
4
Fig. 3. Average latency and avgUT versus number of users per km2 , λ = 2
Gbps.
standpoint. Specifically, the problem is modeled as a NUM
problem subject to the probabilistic latency/reliability constraint
and QoS/rate requirement. By incorporating these constraints,
we have proposed a dynamic Lyapunov control approach, which
adapts to channel variations and system dynamics. Numerical
results show that our proposed approach reduces the latency by
28.41% and 77.11% as compared to current baselines.
R EFERENCES
[1] “2020: Beyond 4G radio evolution for the gigabit experience,” White
Paper, Nokia Siemens Networks, 2011.
[2] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive
MIMO for next generation wireless systems,” IEEE Commun. Mag.,
vol. 52, no. 2, pp. 186–195, Feb. 2014.
[3] O. Semiari, W. Saad, and M. Bennis, “Downlink cell association and
load balancing for joint millimeter wave-microwave cellular networks,”
in Proc. IEEE Global Commun. Conf., Washington, D.C., USA, Dec.
2016, pp. 1–6.
[4] T. K. Vu, M. Bennis, S. Samarakoon, M. Debbah, and M. Latva-aho,
“Joint in-band backhauling and interference mitigation in 5G heterogeneous networks,” in Proc. 22th European Wireless Conf., Oulu, Finland,
May 2016, pp. 1–6.
[5] M. J. Neely, Stochastic Network Optimization with Application to Communication and Queueing Systems. Morgan and Claypool Publishers,
Jun. 2010.
[6] T. Lipp and S. Boyd, “Variations and extension of the convex–concave
procedure,” Optimization and Eng., pp. 1–25, 2015.
[7] A. Liu and V. Lau, “Hierarchical interference mitigation for massive
MIMO cellular networks,” IEEE Trans. Signal Process., vol. 62, no. 18,
pp. 4786–4797, Sep. 2014.
[8] W. Feng, Y. Wang, D. Lin, N. Ge, J. Lu, and S. Li, “When mmWave
communications meet network densification: A scalable interference coordination perspective,” IEEE J. Sel. Areas Commun., 2017, to be published.
[9] S. Wagner, R. Couillet, M. Debbah, and D. T. M. Slock, “Large system
analysis of linear precoding in correlated MISO broadcast channels under
limited feedback,” IEEE Trans. Inf. Theory, vol. 58, no. 7, pp. 4509–4537,
Jul. 2012.
[10] J. D. Little and S. C. Graves, “Little’s law,” in Building intuition.
Springer, 2008, pp. 81–100.
[11] A. Mukherjee, “Queue-aware dynamic on/off switching of small cells in
dense heterogeneous networks,” in Proc. IEEE Global Commun. Conf.
Workshops, Atlanta, GA, USA, Dec. 2013, pp. 182–187.
[12] T. H. A. Le and D. T. Pham, “The DC (difference of convex functions)
programming and DCA revisited with DC models of real world nonconvex
optimization problems,” Ann. Operations Research, vol. 133, no. 1, pp.
23–46, Jan. 2005.
[13] M. R. Akdeniz, Y. Liu, M. K. Samimi, S. Sun, S. Rangan, T. S. Rappaport,
and E. Erkip, “Millimeter wave channel modeling and cellular capacity
evaluation,” IEEE J. Sel. Areas Commun., vol. 32, no. 6, pp. 1164–1179,
Jun. 2014.
| 7cs.IT
|
Local weak form meshless techniques based on the radial point
interpolation (RPI) method and local boundary integral equation (LBIE)
method to evaluate European and American options
Jamal Amani Rada , Kourosh Paranda, Saeid Abbasbandyb,∗
arXiv:1412.6063v1 [cs.CE] 29 Oct 2014
a Department
of Computer Sciences, Faculty of Mathematical Sciences, Shahid Beheshti University, Evin, P.O. Box
198396-3113,Tehran,Iran
b Department of Mathematics, Imam Khomeini International University, Ghazvin 34149-16818, Iran
Abstract
For the first time in mathematical finance field, we propose the local weak form meshless methods for
option pricing; especially in this paper we select and analysis two schemes of them named local boundary
integral equation method (LBIE) based on moving least squares approximation (MLS) and local radial point
interpolation (LRPI) based on Wu’s compactly supported radial basis functions (WCS-RBFs). LBIE and
LRPI schemes are the truly meshless methods, because, a traditional non-overlapping, continuous mesh
is not required, either for the construction of the shape functions, or for the integration of the local subdomains. In this work, the American option which is a free boundary problem, is reduced to a problem
with fixed boundary using a Richardson extrapolation technique. Then the θ-weighted scheme is employed
for the time derivative. Stability analysis of the methods is analyzed and performed by the matrix method.
In fact, based on an analysis carried out in the present paper, the methods are unconditionally stable for
implicit Euler (θ = 0) and Crank-Nicolson (θ = 0.5) schemes. It should be noted that LBIE and LRPI
schemes lead to banded and sparse system matrices. Therefore, we use a powerful iterative algorithm named
the Bi-conjugate gradient stabilized method (BCGSTAB) to get rid of this system. Numerical experiments
are presented showing that the LBIE and LRPI approaches are extremely accurate and fast.
Keywords: Option Pricing; American option; Meshless weak form; LBIE; MLS; LRPI; Richardson
Extrapolation; BCGSTAB; Stability analysis.
AMS subject classification: 91G80; 91G60; 35R35.
1. Introduction
Over the last thirty years, financial derivatives have raised increasing popularity in the markets. In
particular, large volumes of options are traded everyday all over the world and it is therefore of great
importance to give a correct valuation of these instruments.
Options are contracts that give to the holder the right to buy (call) or to sell (put) an asset (underlying)
at a previously agreed price (strike price) on or before a given expiration date (maturity). The majority of
options can be grouped in two categories: European options, which can be exercised only at maturity, and
American options, which can be exercised not only at maturity but also at any time prior to maturity.
Options are priced using mathematical models that are often challenging to solve. In particular, the
famous Black-Scholes model [1] yields explicit pricing formulae for some kinds of European options, including
vanilla call and put, but the modeling of American options is quite complicated. Hence, an analytical solution
is impossible. Therefore, to solve this problem, we need to have a powerful computational method. To this
aim, the most common approaches are the finite difference/finite element/finite volume methods/fast Fourier
transform (see, e.g., [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19]) and the binomial/triniomial
∗ Corresponding
author
Email addresses: [email protected];[email protected] (Jamal Amani Rad), [email protected] (Kourosh
Parand), [email protected] (Saeid Abbasbandy)
Preprint submitted to .
December 19, 2014
trees (see, e.g., [20, 21, 22, 23]), nevertheless some authors have also proposed the use of meshless algorithms
based on radial basis functions [24, 25, 26, 27, 28, 29] and on quasi radial basis functions [30].
Recently, a great attention has been paid to the development of various meshless formulations for solution
of boundary value problems in many branches of science and engineering. Meshless methods are becoming
viable alternatives to either finite element method (FEM) and boundary element method (BEM). Compared
to the FEM and the BEM, the key feature of this kind of method is the absence of an explicit mesh, and
the approximate solutions are constructed entirely based on a group of scattered nodes. Meshless methods
have been found to possess special advantages on problems to that the conventional mesh-based methods
are difficult to be applied. These generally include problems with complicated boundary, moving boundary
and so on [31, 32, 33, 34, 35, 36, 37]. A lot of meshless methods are based on a weak-form formulation on
global domain or a set of local sub-domains.
In the global formulation background cells are needed for the integration of the weak form. Strictly
speaking, these meshless methods are not truly meshless methods. It must be realized that integration is
completed only those background cells with a nonzero shape function.
In the financial literature global meshless method (or Kansa method) has been proposed for pricing
options under various models such as the Black-Scholes model, stochastic volatility models and Merton
models with jumps. In particular, in the case of Black-Scholes model, we mention two papers by Hon and
his co-author [30, 26] where the global RBFs and quasi-radial basis functions are developed. Moreover, as
regards the cases of stochastic volatility models and BlackScholes model on two underlying assets, new global
meshless method is presented in Ballestra and Pacelli [25]. The techniques presented in [25] is combined
of the Gaussian radial basis functions with a suitable operator splitting scheme. Also a numerical method
has recently been presented by Saib et al. [38]. In particular, in this latter work, a differential quadrature
radial basis functions is used to reduce the American option pricing problem under Merton’s jump-diffusion
model to a system of ordinary differential equations. The interested reader can also see [39, 40].
In methods based on local weak-form formulation no cells are needed and therefore they are often known
as truly meshless methods. By using a simple form for the geometry of the sub-domains, one can use a
numerical integration method, easily. Recently, two family of meshless methods, on the basis of the local
weak form for arbitrary partial differential equations with moving least-square (MLS) and radial basis
functions (RBFs) approximation have been developed [41, 42, 43, 44, 45]. Local boundary integral equation
method (LBIE) with moving least squares approximation and local radial point interpolations (LRPI) with
radial basis functions have been developed by Zhu et al. [46] and Liu et al. [47, 48], respectively. Both
methods (LBIE and LRPI) are meshless, as no domain/boundary traditional non-overlapping meshes are
required in these two approaches. Particularly, the LRPI meshless method reduces the problem dimension
by one, has shape functions with delta function properties, and expresses the derivatives of shape functions
explicitly and readily. Thus it allows one to easily impose essential boundary and initial (or final) conditions.
Though the LBIE method is an efficient meshless method, it is difficult to enforce the essential boundary
conditions for that the shape function constructed by the moving least-squares (MLS) approximation lacks
the delta function property. Some special techniques have to be used to overcome the problem, for example,
the Lagrange multiplier method and the penalty method [49]. In this paper, meshless collocation method
is applied to the nodes on the boundaries. The papers of Zhu et al. [46, 50, 51] in linear and non-linear
acoustic and potential problems, and for heat conduction problems, the works of Sladek brothers [52, 53]
by meshless LBIE are useful for researchers. This method has now been successfully extended to a wide
rang of problems in engineering. For some examples of these problems, see [54, 55, 56] and other references
therein. The interested reader of meshless methods can also see [57, 58].
The objective of this paper is to extend the LRPI based on Wu’s compactly supported radial basis
functions (WCS-RBFs) with C 4 smoothness [59] and LBIE method based on moving least squares with
cubic spline weight function to evaluate European and American options. To the best of our knowledge, the
local weak form of meshless method has not yet been used in mathematical finance. Therefore, it appears
to be interesting to extend such a numerical technique also to option valuation, which is done in the present
manuscript. In particular, we develop a local weak form meshless algorithm for pricing both European and
American options under the Black-Scholes model.
In addition, in this paper the infinite space domain is truncated to [0, Smax ] with a sufficiently large
Smax to avoid an unacceptably large truncation error. The options’ payoffs considered in this paper are
non-smooth functions, in particular their derivatives are discontinuous at the strike price. Therefore, to
2
reduce as much as possible the losses of accuracy the points of the trial functions are concentrated in a
spatial region close to the strike prices. So, we employ the change of variables proposed by Clarke and
Parrott [60].
As far as the time discretization is concerned, we use the θ-weighted scheme. Stability analysis of
the method is analyzed and performed by the matrix method in the present paper. Furthermore, in this
paper we will see that the time semi-discretization is unconditionally stable for implicit Euler (θ = 0) and
Crank-Nicolson (θ = 0.5) schemes.
Finally, in order to solve the free boundary problem that arises in the case of American options is computed by Richardson extrapolation of the price of Bermudan option. In essence the Richardson extrapolation
reduces the free boundary problem and linear complementarity problem to a fixed boundary problem which
is much simpler to solve.
The paper is organized as follows: In Section 2 a detailed description of the Black-Scholes model for
European and American options is provided. Section 3 is devoted to presenting the LBIE and LRPI
approaches and the application of such a numerical technique to the option pricing problems considered is
shown in this section. The numerical results obtained are presented and discussed in Section 4 and finally,
in Section 5, some conclusions are drawn.
2. The Black-Scholes model for European and American options
For the sake of simplicity, from now we restrict our attention to options of put type. However the reader
will note that the numerical method presented in this paper can be used with little modifications also to
price call options.
Let us consider a put option with maturity T and strike price E on an underlying asset s that follows
(under the risk-neutral measure) the stochastic differential equation [61]:
ds = r s dt + σ s dW,
(2.1)
where r and σ are the interest rate and the volatility, respectively. Moreover let V (s, t) denote the option
price, and let us define the Black-Scholes operator [61]:
LV (s, t) = −
∂
σ2 2 ∂ 2
∂
V (s, t) − rs V (s, t) + rV (s, t).
V (s, t) −
s
∂t
2 ∂s2
∂s
(2.2)
2.1. European option
The option price V (s, t) satisfies, for s ∈ (0, +∞) and t ∈ [0, T ), the following partial differential problem
[61, 30, 26]:
LV (s, t) = 0,
(2.3)
V (s, T ) = ς(s),
(2.4)
with final condition:
and boundary conditions:
V (0, t) = E exp{−r(T − t)},
lim V (s, t) = 0,
s→+∞
(2.5)
where ς is the so-called option’s payoff:
ς(s) = max(E − s, 0),
(2.6)
which is clearly not differentiable at s = E.
An exact analytical solution to the problem (2.3)-(2.5), i.e. the famous Black-Scholes formula, is available.
3
2.2. American option
The option price V (s, t) satisfies, for s ∈ [0, +∞) and t ∈ [0, T ), the following partial differential problem
[30, 26]:
LV (s, t) = 0,
s > B(t),
(2.7)
V (s, t) = E − s,
∂V (s, t)
= −1,
∂s
s=B(t)
0 ≤ s < B(t),
(2.8)
(2.9)
V (B(t), t) = E − B(t),
(2.10)
with final condition:
V (s, T ) = ς(s),
(2.11)
lim V (s, t) = 0,
(2.12)
and boundary condition:
s→+∞
where B(t) denotes the so-called exercise boundary, which is unknown and is implicitly defined by (2.7)(2.12). The above free-boundary partial differential problem does not have an exact closed-form solution,
and thus some numerical approximation is required.
Problem (2.7)-(2.12) can be reformulated as a linear complementarity problem [10]:
LV (s, t) ≥ 0,
(2.13)
V (s, t) − ς(s) ≥ 0,
(LV (s, t)) · (V (s, t) − ς(s)) = 0,
(2.14)
(2.15)
which holds for s ∈ (0, +∞) and t ∈ [0, T ), with final condition:
V (s, T ) = ς(s),
(2.16)
and boundary conditions:
V (0, t) = E,
lim V (s, t) = 0.
s→+∞
(2.17)
In this work, the price of American option is computed by Richardson extrapolation of the price of
Bermudan option which is also employed in other works, such as for example [8, 9, 23, 25, 62]. In essence
the Richardson extrapolation reduces the free boundary problem and linear complementarity problem to a
fixed boundary problem which is much simpler to solve. Thus, instead of describing the aforementioned linear
complementarity problem or penalty method, we directly focus our attention onto the partial differential
equation satisfied by the price of a Bermudan option which is faster and more accurate than other methods.
Let us consider in the interval [0, T ], M + 1 equally spaced time levels t0 = 0, t1 , t2 , ..., tM = T . Let VM (s, t)
denote the price of a Bermudan option with maturity T and strike price E. The Bermudan option is an
option that can be exercised not on the whole time interval [0, T ], but only at the dates t0 , t1 , . . ., tM . That
is we consider the problems
(
LVM (s, t) = 0,
(2.18)
VM (0, t) = E,
lims→∞ VM (s, t) = 0,
which hold in the time intervals (t0 , t1 ), (t1 , t2 ), . . ., (tM−1 , tM ). By doing that also the relation (2.15) is
automatically satisfied in every time interval (tk , tk+1 ), k = 0, 1, . . . , M − 1. Moreover, the relation (2.14)
is enforced only at times t0 , t1 , . . ., tM−1 , by setting
VM (s, tk ) = max( lim+ VM (s, t), ς(s)),
t→tk
4
k = 0, 1, . . . , M − 1.
(2.19)
Note that the function VM (·, tk ) computed according to (2.19) is used as the final condition for the problem
(2.18) that holds in the time interval (tk−1 , tk ), k = 1, 2, . . . , M − 1. Instead, the final condition for the
problem (2.18) that holds in the time interval (tM−1 , tM ), according to the relation (2.16), is prescribed as
follows:
VM (s, tM ) = ς(s).
(2.20)
That is, in summary, problems (2.18) are recursively solved for k = M − 1, M − 2, . . . , 0, starting from the
condition (2.20), and at each time tM−1 , tM−2 , . . ., t0 the American constraint (2.19) is imposed.
The Bermudan option price VM (s, t) tends to become a fair approximation of the American option price
V (s, t) as the number of exercise dates M increases. In this work the accuracy of VM (s, t) is enhanced by
Richardson extrapolation which is second-order accurate in time.
3. Methodology
To evaluate the option, the LBIE and LRPI methods are used in the present work. The methods are
based on local weak forms over intersecting sub-domains. For ease of exposition, we focus our attention
onto American option, as the application of the method to European option is substantially analogous and
requires only minor modifications. At first we discuss a time-stepping method for the time derivative.
3.1. Time discretization
First of all, we discretize the Black-Scholes operator (2.2) in time. For this propose, we can apply
the Laplace transform or use a time-stepping approximation. Algorithms for the numerical inversion of a
Laplace transform lead to a reduction in accuracy. Then, we employ a time-stepping method to overcome
the time derivatives in this equation.
Let V k (s) denote a function approximating VM (s, tk ), k = 0, 1, ..., M − 1. Note that the subscript M
has been removed from V k (s) to keep the notation simple. According to (2.20) we set V M (s) = ς(s). Let
us consider the following θ-weighted scheme:
LV k (s) =
d
1
θσ 2 2 d2 k+1
V
(s) + θrs V k+1 (s) + (−θr +
s
)V k+1 (s) −
2
ds2
ds
∆t
d
(θ − 1)σ 2 2 d2 k
V (s) − (θ − 1)rs V k (s) −
s
2
2
ds
ds
1
k
)V (s).
(−(θ − 1)r +
∆t
Therefore, the American option problems are rewritten as follows:
(
LV k (s) = 0,
V k (0) = E,
lims→∞ V k (s) = 0,
and also, the relations (2.19) and (2.20) are rewritten as follows:
k = 0, 1, . . . , M − 1,
V k (s) = max
lim V k (s), ς(s) ,
t→t+
k+1
(3.1)
(3.2)
(3.3)
V M (s) = ς(s).
3.2. Spatial variable transformation
Now, the infinite space domain is truncated to [0, Smax ] with a sufficiently large Smax to avoid an
unacceptably large truncation error. However, in [63] shown that upper bound of the asset price is three
or four times of the strike price, so we can set Smax = 5E. The options’ payoffs considered in this paper
are non-smooth functions, in particular their derivatives are discontinuous at the strike price. Therefore, to
reduce the losses of accuracy the points of the trial functions are concentrated in a spatial region close to
5
s = E. So, we employ the following change of variables which is not new but has already been proposed by
[60] and has also been employed for example in [25, 39, 40]:
x(s) =
or
sinh−1 (ξ(s − E)) + sinh−1 (ξE)
,
sinh−1 (ξ(Smax − E)) + sinh−1 (ξE)
!
1
s(x) = sinh x sinh−1 (ξ(Smax − E)) − (1 − x) sinh−1 (ξE) + E.
ξ
(3.4)
(3.5)
Note that the Eq. (3.4) maps the [0, Smax ] to the [0, 1]. We define
U (x, t) =
e k (x)
LU
=
V (s(x), t),
d2
d
θα(x) 2 U k+1 (x) + θβ(x) U k+1 (x) + γ1 U k+ (x) −
dx
dx
d
d2 k
(θ − 1)α(x) 2 U (x) − (θ − 1)β(x) U k (x) − γ2 U k (x),
dx
dx
(3.6)
(3.7)
where
2
1 2 s(x)
,
α(x) = σ
2
s′ (x)
2 ′′
s (x)
s(x)
s(x)
1
+r ′
,
β(x) = − σ 2 ′
′
2
s (x)
s (x)
s (x)
1
,
γ1 = − θr +
∆t
1
γ2 = − (θ − 1)r +
.
∆t
Using the change of variable (3.4), the relations (3.2) are rewritten as follows:
(
e k (x) = 0,
LU
U k (0) = E,
U k (1) = 0,
(3.8)
and also, the relations (3.3) are rewritten as follows:
U k (x) = max( lim U k (x), ςe(x)),
t→t+
k+1
where
k = 0, 1, . . . , M − 1,
(3.9)
U M (x) = ςe(x),
ςe(x) = max E − s(x), 0 .
(3.10)
3.3. The local weak form
In this section, we use the local weak form instead of the global weak form. The LBIE and LRPI methods
construct the weak form over local sub-domains such as Ωs , which is a small region taken for each node in
the global domain Ω = [0, 1]. The local sub-domains overlap each other and cover the whole global domain
Ω. This local sub-domains could be of any size. Therefore the local weak form of the approximate equation
(3.8) for x ∈ Ωis can be written as
!
Z
d2 k+1
d k+1
θα(x) 2 U
(x) + θβ(x) U
(x) + γ1 U k+1 (x) u∗ dΩ
dx
dx
Ωis
!
Z
d2 k
d k
k
(θ − 1)α(x) 2 U (x) + (θ − 1)β(x) U (x) + γ2 U (x) u∗ dΩ,
(3.11)
=
dx
dx
Ωis
6
which Ωis is the local domain associated with the point i, i.e. it is a interval centered at x of radius rQ . In
LRPI, u∗ is the Heaviside step function
(
1, x ∈ Ωis ,
∗
u (x) =
0, otherwise,
as the test function in each local domain, but in LBIE method, u∗ is the fundamental solution for the
one-dimensional Laplace operator defined by the equation
∂2 ∗
u (x, ξ) = δ(x, ξ),
∂x2
(3.12)
where x and ξ are a field point and a source point, respectively, and δ is Dirac delta function. The
fundamental solution and its derivative are given as follows
1
|x − ξ|,
2
1
∂ ∗
u (x, ξ) = sgn(x − ξ),
∂x
2
u∗ (x, ξ) =
(3.13)
where the symbol sgn denotes the signum function. Using the integration by parts, we have
!
Z
d j
d2 j
∗
′
j
α(x) 2 U (x)u dΩ = α(x) U (x) − α (x)U (x) u∗
dx
dx
Ωis
∂Ωis
Z
Z
Z
∗
d j
∂u
∂u∗
+
U j (x)α′′ (x)u∗ dΩ +
U j (x)α′ (x)
dΩ −
U (x)α(x)
dΩ,
∂x
∂x
Ωis
Ωis
Ωis dx
(3.14)
and
Z
d
β(x) U j (x)u∗ dΩ = β(x)U j (x)u∗
dx
i
Ωs
Z
∂u∗
−
U j (x)β(x)
dΩ,
∂x
Ωis
7
−
∂Ωis
Z
U j (x)β ′ (x)u∗ dΩ
Ωis
(3.15)
for j = k, k + 1. In above relations, ∂Ωis is the boundary of Ωis . Using relations (3.11), (3.12), (3.14) and
(3.15) for LBIE scheme, we get
!
d k+1
θ
i
′
k+1
|x − x | α(x) U
(x) − α (x)U
(x)
2
dx
i
∂Ωs
Z
θ
U k+1 (x)α′′ (x)|x − xi |dΩ + |x − xi |β(x)U k+1 (x)
2
Ωis
∂Ωis
Z
θ
−
U k+1 (x)β ′ (x)|x − xi |dΩ
2 Ωis
Z
Z
θ
U k+1 (x)β(x)sgn(x − xi )dΩ
+γ1
U k+1 (x)dΩ −
2
i
i
Ωs
Ω
Z s
Z
θ
θ
d k+1
k+1
′
i
+
U
(x)α (x)sgn(x − x )dΩ −
U
(x)α(x)sgn(x − xi )dΩ
2 Ωis
2 Ωis dx
!
d
θ−1
|x − xi | α(x) U k (x) − α′ (x)U k (x)
=
2
dx
i
θ
2
+
Z
θ−1
2
∂Ωs
θ−1
|x − xi |β(x)U k (x)
2
Ωis
∂Ωis
Z
Z
θ−1
U k (x)β ′ (x)|x − xi |dΩ + γ1
U k (x)dΩ
−
2
Ωis
Ωis
Z
Z
θ−1
θ−1
U k (x)β(x)sgn(x − xi )dΩ +
U k (x)α′ (x)sgn(x − xi )dΩ
−
2
2
Ωis
Ωis
Z
θ−1
d k
−
U (x)α(x)sgn(x − xi )dΩ.
2
dx
i
Ωs
+
U k (x)α′′ (x)|x − xi |dΩ +
(3.16)
Also, because the derivative of the Heaviside step function u∗ is equal to zero, then for LRPI scheme using
relations (3.14) and (3.15) the local weak form equation (3.11) is transformed into the following simple local
integral equation
!
d k+1
′
k+1
(x) − α (x)U
(x)
θ α(x) U
dx
i
+θ
∂Ωs
Z
U
k+1
′′
(x)α (x)dΩ + θβ(x)U
k+1
(x)
Ωis
+γ1
Z
−θ
∂Ωis
Z
U k+1 (x)β ′ (x)dΩ
Ωis
U k+1 (x)dΩ
Ωis
!
d k
′
k
= (θ − 1) α(x) U (x) − α (x)U (x)
dx
+(θ − 1)
+γ2
Z
Z
k
∂Ωis
k
′′
− (θ − 1)
U (x)α (x)dΩ + (θ − 1)β(x)U (x)
Ωis
∂Ωis
U k (x)dΩ.
Z
U k (x)β ′ (x)dΩ
Ωis
(3.17)
Ωis
It is important to observe that in relations (3.16) and (3.17) exist unknown functions, we should approximate
these functions. To this aim the local integral equations (3.16) and (3.17) are transformed in to a system
of algebraic equations with real unknown quantities at nodes used for spatial approximation, as described
in the next section.
8
3.4. Spatial approximation
Instead of using traditional non-overlapping, contiguous meshes to form the interpolation scheme, the
LBIE and LRPIM methods use a local interpolation or approximation to represent the trial or test functions
with the values (or the fictitious values) of the unknown variable at some randomly located nodes. There
are a number of local interpolation schemes for this purpose. The moving least squares approximation and
radial point interpolation are two of them. In this section, the fundamental idea of these approximations is
reviewed.
3.4.1. The MLS approximation
Consider a sub-domain Ωx of Ω in the neighborhood of a point x for the definition of the MLS approximation of the trial function around x. To approximate the distribution of function U k (x) in Ωx , over a
e k (x) of
number of randomly located nodes {xi }, i = 1, 2, ..., n, the moving least squares approximation U
k
U (x) for each x ∈ Ωx , can be defined by [49, 50, 51, 52, 54, 64, 65]
e k (x) = pT (x)ak (x), ∀x ∈ Ωx ,
U
(3.18)
where pT (x) = [p1 (x) p2 (x) ... pm (x)] is a complete monomial basis of order m. In the present work, both
the constant and the linear monomials are used, i.e. we set m = 2. Also ak (x) is an undetermined coefficient
vector with components akj (x), j = 1, 2, ..., m, which is decided by minimizing the evaluation function Jk (x)
shown below [49, 50, 51, 52, 54, 64, 65]
k
J (x)
=
n
X
i=1
=
"
"
bik
wi (x) p (x )a (x) − U
T
bk
P.a (x) − U
k
i
#T
k
"
#2
#
k
b
.W. P.a (x) − U ,
k
(3.19)
where wi (x) is a positive weight function associated with the node i, which decreases as |x − xi | increases.
It always takes unit value at the sampling point x and vanishes outside the support domain Ωx of x. The
cubic spline weight function is considered in the present work. The weight function corresponding to node
i for a one-dimensional domain may be written as [49, 50, 51]
2
2
3
ri ≤ 0.5,
3 − 4ri + 4ri ,
4
4
2
3
(3.20)
wi (x) = 3 − 4ri + 4ri − 3 ri , 0.5 < ri ≤ 1,
0,
ri > 1,
where ri = |x − xi |/rw is the distance from node xi to x, while rw is the size of support for the weight
function wi (x).
b k are the fictitious nodal
In relation (3.19), n is the number of nodes in the support domain Ωx and U
i
nodes. The matrices P and W are defined as [49, 50, 51, 52, 54, 64, 65]
T 1
p (x )
pT (x2 )
P=
,
W = Diag(w1 (x), w2 (x), . . . , wn (x)).
(3.21)
..
.
pT (xn )
n×m
The stationarity condition of Jk (x) in Eq. (3.19) with respect to ak (x) i.e. ∂Jk (x)/∂ak (x), leads to the
b k (see [49, 50, 51, 52, 54, 64, 65] for more details)
following linear relation between ak (x) and U
from which
b k,
A(x)ak (x) = B(x)U
(3.22)
b k,
ak (x) = A−1 (x)B(x)U
(3.23)
9
where the matrices A(x) and B(x) are defined by
B(x) = PT W,
A(x) = PT WP = B(x)P.
The MLS is well-defined only when A is non-singular or the rank of P equals m or at least m weight
functions are non-zero i.e. n > m for each x ∈ Ω. Substituting Eq. (3.23) into Eq. (3.18), the approximate
e k (x) can be expressed in terms of the shape functions as
function U
e k (x) = ϕT (x)U
b k (x) =
U
n
X
i=1
where the nodal shape function ϕi (x) is given by
bik ,
ϕi (x)U
∀x ∈ Ωx ,
(3.24)
ϕT (x) = pT (x)A−1 (x)B(x),
or
ϕi =
m
X
pj (x)[A−1 (x)B(x)]ji .
(3.25)
j=1
Derivatives of ϕi (x) are obtained as (see [54, 64, 65] for more details)
ϕi,x =
m
X
pj,x [A−1 B]ji + pj [A−1 Bx + A−1
x B]ji ,
j=1
m
X
ϕi,xx =
−1
−1
pj,xx [A−1 B]ji + 2pj,x [A−1 Bx + A−1
Bxx + A−1
x B]ji + pj [2Ax Bx + A
xx B]ji ,
j=1
−1
−1
= (A−1 )xx represent the one and two derivatives of the inverse of A with
)x and Axx
where A−1
x = (A
respect to x, respectively, which are given by
−1
A−1
Ax A−1 ,
x = −A
−1
−1
A−1
− A−1 Ax A−1 − A−1 Ax A−1
xx = −Ax Ax A
x ,
where (.)x denotes
d(.)
dx .
3.4.2. Local radial point interpolation
According to the local point interpolation [48], the value of point interpolation approximation of U k (x)
at any (given) point x ∈ Ω = [0, 1] is approximated by interpolation at n nodes x1 , x2 , . . ., xn (centers)
laying in a convenient neighborhood of x i.e. Ωx . The domain in which these nodes are chosen, whose
shape may depend on the point x, is usually referred to as local support domain. Various different local
point interpolation approaches can be obtained depending on the functions used to interpolate U k (x). In
this paper we focus our attention onto the so-called local radial point interpolation method (LRPI), which
employs a combination of polynomials and radial basis functions.
k
The function that interpolate U k (x) for each x ∈ Ωx , which we denote by URBP
I (x), is obtained as
follows:
k
URBP
I (x) =
n
X
Ri (x)aki +
m
X
Pj (x)bkj ,
(3.26)
j=1
i=1
where P1 , P2 , . . ., Pm denote the first m monomials in ascending order (i.e. P1 = 1, P2 = x, . . ., Pm = xm−1 )
and R1 , R2 , . . ., Rn are n radial functions centered at x1 , x2 , . . ., xn , respectively. Moreover ak1 , ak2 , . . .,
akn , bk1 , bk2 , . . ., bkm are n + m real coefficients that have to be determined.
As far as the radial basis functions R1 , R1 , . . ., Rn are concerned, several choices are possible (see,
for example, [66]). In this work we decide to use the Wu’s compactly supported radial basis functions
10
(WCS-RBFs) with C 4 smoothness [59], as they do not involve any free shape parameter (which is not
straightforward to choose, see [67, 68, 69, 39, 40]). WCS-RBFs are as follows:
Ri (x) = (1 − ri )6+ (6 + 36ri + 82ri2 + 72ri3 + 30ri4 + 5ri5 ), i = 1, 2, . . . , n,
(3.27)
where ri = |x − xi |/rw is the distance from node xi to x, while rw is the size of support for the radial
function Ri (x). Also, (1 − ri )6+ is (1 − ri )6 for 0 ≤ ri < 1 and zero otherwise.
Note that the monomials P1 , P2 , . . . , Pm are not always employed (if bki = 0, i = 1, 2, . . . , m, pure RBF
approximation is obtained). In the present work, both the constant and the linear monomials are used to
augment the RBFs (i.e. we set m = 2).
k
By requiring that the function URBP
I interpolate U at x1 , x2 , . . ., xn , we obtain a set of n equations in
k
the n + m unknown coefficients a1 , ak2 , . . ., akn , bk1 , bk2 , . . ., bkm :
n
X
Ri (xp )aki +
m
X
j=1
i=1
b k (xp ),
Pj (xp )bkj = U
p = 1, 2, . . . , n,
(3.28)
b k are the fictitious nodal nodes. Moreover, in order to uniquely determine U k
where U
RBP I , we also impose:
n
X
Pj (xi )aki = 0,
j = 1, 2, . . . , m.
(3.29)
i=1
That is we have the following system of linear equations:
k
bk
a
U
G
=
,
bk
0
where
bk =
U
h
bk
U
1
bk
U
2
bk
... U
n
iT
G=
R=
P=
R1 (x1 )
R1 (x2 )
..
.
=
h
b (x1 )
U
R
PT
P
0
R2 (x1 )
R2 (x2 )
..
.
...
...
..
.
b (x2 )
U
b (xn )
... U
iT
,
(3.30)
,
Rn (x1 )
Rn (x2 )
..
.
R1 (xn ) R2 (xn )
. . . Rn (xn )
P1 (x1 )
P1 (x2 )
..
.
P2 (x1 )
P2 (x2 )
..
.
...
...
..
.
P1 (xn )
P2 (xn )
. . . Pm (xn )
Pm (x1 )
Pm (x2 )
..
.
,
,
ak = [ak1
ak2
. . . akm ]T ,
(3.31)
bk = [bk1
bk2
. . . bkm ]T .
(3.32)
Unique solution is obtained if the inverse of matrix R exists, so that
k
bk
a
U
−1
.
=G
bk
0
11
Accordingly, (3.26) can be rewritten as
k
URBP
I (x) =
or, equivalently,
k
URBP
I (x)
=
RT (x)
T
PT (x)
T
R (x)
P (x)
Let us define the vector of shape functions:
Φ(x) = [ ϕ1 (x)
ϕ2 (x)
ak
bk
−1
G
bk
U
0
,
.
(3.33)
p = 1, 2, . . . , n,
(3.34)
. . . ϕn (x)] ,
where
ϕp (x) =
n
X
Ri (x)G−1
i,p +
i=1
m
X
−1
Pj (x)Gn+j,p
,
j=1
−1
and G−1
. Using (3.34) relations (3.33) are rewritten in the more
i,p is the (i, p) element of the matrix G
compact form:
k
bk
URBP
I (x) = Φ(x)U ,
or, equivalently,
k
URBP
I (x) =
n
X
i=1
(3.35)
b k ϕi (x).
U
i
(3.36)
It can be easily shown that the shape functions (3.34) satisfy the so-called Kronecker property, that is
ϕi (xj ) = δij ,
(3.37)
where δij is the well-known Kronecker symbol, so that essential boundary and final conditions such as those
considered in Section 2 (e.g., (2.4), (2.5), (2.16), (2.17)) can be easily imposed. Note also that the derivatives
k
of URBP
I (of any order) with respect to x are easily obtained by direct differentiation in (3.36).
3.5. Discretized equations
For implementation of the LBIE and LRPI methods, N + 1 regularly nodes xj , j = 0, 1, ..., N are chosen
in the interval [0, 1] where x0 = 0 and xN = 1. We define X = {x0 , x1 , ..., xN }. Distance between two
nodes is defined by h = xi+1 − xi , for i = 0, 1, ..., N − 1. It is important to observe that U k+1 (x) must be
considered as known quantities, since it is approximated at the previous iteration. We want to approximate
U k (x) using radial point interpolation and MLS approximations. Substituting the displacement expression
in Eqs. (3.24) and (3.36) into the local weak forms (3.16) and (3.17), respectively, the discrete equation for
each interior node is obtained as follows
Z
n
n
X
X
k+1
b k+1 (β(x) − α′ (x))ϕj (x)
b
U
Uj
(θα′′ (x) − θβ ′ (x) + γ1 )ϕj (x)dΩ + θ
j
Ωis
j=1
+θ
n
X
j=1
=
n
X
j=1
j=1
b k+1 α(x)ϕj,x (x)
U
j
bk
U
j
+(θ − 1)
Z
∂Ωis
∂Ωis
((θ − 1)α′′ (x) − (θ − 1)β ′ (x) + γ2 )ϕj (x)dΩ
Ωis
n
X
j=1
b k (β(x) − α′ (x))ϕj (x)
U
j
+ (θ − 1)
∂Ωis
12
n
X
j=1
b k α(x)ϕj,x (x)
U
j
(3.38)
∂Ωis
and
#
θ ′
θ ′′
′
i
i
(α (x) − β (x))|x − x | + (α (x) − β(x))sgn(x − x ) + γ1 ϕj (x)dΩ
2
Ωis 2
j=1
"
#
Z
n
n
X
θ
θ X b k+1
k+1
i
′
i
b
Uj
−
|x − x |(β(x) − α (x)) ϕj (x)
U
α(x)sgn(x − x ) ϕj (x)dΩ +
2 j=1 j
Ωis 2
j=1
∂Ωis
"
#
n
θ X b k+1
|x − xi |α(x) ϕj,x (x)
U
+
2 j=1 j
∂Ωis
"
#
Z
n
X
θ
−
1
θ
−
1
k
′′
′
i
′
i
bj
U
=
(α (x) − β (x))|x − x | +
(α (x) − β(x))sgn(x − x ) + γ1 ϕj (x)dΩ
2
2
Ωis
j=1
"
#
Z
n
n
X
X
θ
−
1
θ
−
1
k
i
′
i
k
b |x − x |(β(x) − α (x)) ϕj (x)
bj
U
α(x)sgn(x − x ) ϕj (x)dΩ +
U
−
2
2 j=1 j
Ωis
j=1
∂Ωis
"
#
n
θ − 1 X bk
+
U |x − xi |α(x) ϕj,x (x)
.
(3.39)
2 j=1 j
i
n
X
b k+1
U
j
"
Z
∂Ωs
The matrix forms of the relations (3.38) and (3.39) are respectively as follows
!
n
n
X
X
b k+1 =
(θ − 1)Aij + γ2 Bij
θAij + γ1 Bij + θCij + θDij U
j
j=1
bk,
+(θ − 1)Cij + (θ − 1)Dij U
j
and
(3.40)
!
n
X
e ij − θC
e ij + θD
e ij + θE
e ij U
e ij + γ1 B
b k+1 =
θA
j
j=1
where
j=1
!
!
e
e
e
bjk ,
−(θ − 1)Cij + (θ − 1)Dij + (θ − 1)Eij U
Aij =
Z
(α′′ (x) − β ′ (x))ϕj (x)dΩ,
Bij =
Ωis
e ij =
B
Z
Z
"
,
(3.41)
Z
ϕj (x)dΩ,
Dij = α(x)ϕj,x (x)
∂Ωis
′′
′
,
∂Ωis
i
′
i
#
(α (x) − β (x))|x − x | + (α (x) − β(x))sgn(x − x ) ϕj (x)dΩ,
Ωis
ϕj (x)dΩ,
Ωis
e ij
e ij + γ2 B
(θ − 1)A
j=1
Ωis
Cij = (β(x) − α′ (x))ϕj (x)
e ij = 1
A
2
n
X
e ij = 1
C
2
#
(3.42)
Z
"
1
e ij =
|x − xi |(β(x) − α′ (x)) ϕj (x)
D
2
α(x)sgn(x − xi )ϕj,x (x)dΩ,
Ωis
,
∂Ωis
"
#
1
i
e ij =
|x − x |α(x) ϕj,x (x)
E
2
.
∂Ωis
In the LBIE method, it is difficult to enforce the boundary conditions (3.8) for that the shape function
constructed by the MLS approximation lacks the delta Kronecker property. In this work, we use the
13
collocation method to the nodes on the boundary. In fact, at the boundary x0 and xN , U k (x) will be
obtained by
e k (x0 ) =
E=U
n
X
j=1
bjk ,
ϕj (x0 )U
e k (xN ) =
0=U
n
X
j=1
bjk ,
ϕj (xN )U
(3.43)
but in LRPIM, using the delta Kronecker property, the boundary conditions (3.8) can be easily imposed.
In fact, we have
bjk = 0,
U
bjk = δ1j E,
U
j = 1, 2, . . . , n.
(3.44)
Therefore, we can rewrite the relations (3.41) and (3.43) in LBIE scheme, and also (3.40) and (3.44) in
LRPI scheme in the compact forms
and
In these systems, we have
b k = QU
b k+1 + H
b k,
PU
(3.45)
b k = GV
b k+1 .
FV
(3.46)
bk = [ U
b0k
U
bk = [ U
bk
V
1
b1k
U
bk
U
2
b2k
U
bk
U
3
b k ]T1×(N +1) ,
... U
N
T
bk
... U
N −1 ]1×(N −1) ,
b k is an 1 × (N + 1) vector
also in the linear system (3.45), H
bk = [ E
H
0 0
. . . 0 ]T1×(N +1) ,
(3.47)
and Q is the (N + 1) × (N + 1) banded matrix with bandwidth bw = 2⌊rw ⌋ + 1 whose first and last rows
contain only zero elements and the other rows are
where
e i − θC
e i + θD
e i + θE
e i,
e i + γ1 B
Qi = θ A
e ij =
A
i = 1, . . . , N − 1,
(
integration obtained in relation (3.42), xj ∈ X ∩ Ωis ,
0,
otherwise.
e i, C
e i, D
e i and E
e i are defined analogously. Also P is the (N + 1) × (N + 1) banded matrix
The quantities B
with bandwidth bw (See Figure 1) whose first and last rows given by the first and second of relations (3.43),
respectively, and the other rows are
Pi
=
e i − (θ − 1)C
e i + (θ − 1)D
ei
e i + γ2 B
(θ − 1)A
e i , i = 1, . . . , N − 1.
+(θ − 1)E
(3.48)
On the other hand, in the linear system (3.46), we have two banded matrices F and G of size (N −1)×(N −1)
with bandwidth bw = 2⌊rw ⌋ + 1, so that
Gi = θAi + γ1 Bi + θCi + θDi ,
Fi = (θ − 1)Ai + γ2 Bi + (θ − 1)Ci + (θ − 1)Di ,
for i = 1, 2, ..., N − 1. Finally, in LBIE scheme, combining Eqs. (3.9) and (3.45) leads to the following
system:
(
b k = QU
b k+1 + H
b k,
PΞ
(3.49)
k
k b
b
b
U = max{Ξ , Π},
14
to be recursively solved for k = M − 1, M − 2, . . . , 0, starting from
b M = Π,
b
U
(3.50)
b are obtained from MLS approximation to option’s payoff (3.10).
where Π
And for LRPI scheme, we find the following iterative system
(
b k = GV
b k+1 ,
FΞ
k
b = max{Ξ
b k , Π},
V
to be recursively solved for k = M − 1, M − 2, . . . , 0, starting from
b M = Π = [ ςe(x1 )
V
ςe(x2 )
. . . ςe(xN −1 ) ]T .
(3.51)
Remark 1: Note that the numerical methods proposed in this work require solving at every time step
a system of linear equations (systems (3.45) and (3.46)). These highly sparse linear equations can be
solved using efficient solution techniques for sparse matrices based on either direct methods or iterative
methods. Direct methods, like LU factorization method, can be applied to any non-singular matrix and
are well adapted to matrix inversion and solution of linear systems. These methods are especially well
suited to solving dense systems with O(N 3 ) computational complexity. Direct methods can be impractical
if coefficient matrix is large and sparse, because the triangular factors of a sparse matrix usually have many
more nonzero elements than itself. So a considerable amount of memory is required and even the solution of
the triangular system costs many floating point operations. This necessitates the use of iterative algorithms
to preserve the sparsity of the coefficient matrix. The most powerful iterative algorithm of these types is
the Bi-conjugate gradient stabilized method (BCGSTAB) developed by Van de Vorst [70] for solving sparse
linear systems. Suppose that the final system of equations has the form Ax = b. This method generates
a sequence of approximate solutions {x(k) } and essentially involve the coefficient matrix A only in the
context of matrix-vector multiplication which is relatively inexpensive for a sparse matrix O(N (2bw + 1))
as compared to a full matrix O(N 2 ). It should also be noted that the complexity of BCGSTAB method
O(2N (2bw + 1)K), where K is the number of iteration of algorithm. We simply observe that complexity of
this algorithm is very lower than complexity of direct methods. Note that in all the cases considered, the
tolerance number tolerance is selected as 10−10 . Convergence was achieved after less than 5 iterations in all
cases tested.
Remark 2: The key note in applying the local weak form schemes is computing the local integrals using
an accurate numerical integration rule. In the following we consider the well-known numerical integration
15
4
e ij and C
e ij :
rule named Simpson’s rule which has a truncate error of order O(rQ
) to Aij , Bij , A
Aij =
Z
′′
′
(α (x) − β (x))ϕj (x)dΩ =
Z
xi +rQ
(α′′ (x) − β ′ (x))ϕj (x)dx
xi −rQ
Ωis
"
rQ
(α′′ (xi − rQ ) − β ′ (xi − rQ ))ϕj (xi − rQ ) + 4(α′′ (xi ) − β ′ (xi ))ϕj (xi )
=
2
#
4
+ (α′′ (xi + rQ ) − β ′ (xi + rQ ))ϕj (xi + rQ ) + O(rQ
),
Bij =
Z
ϕj (x)dΩ =
Ωis
e ij = 1
A
2
Z
Ωis
rQ
=
4
(
+
e ij = 1
C
2
"
"(
Z
xi +rQ
xi −rQ
"
#
rQ
4
ϕj (x)dx =
ϕj (xi − rQ ) + 4ϕj (xi ) + ϕj (xi + rQ ) + O(rQ
),
2
#
(α′′ (x) − β ′ (x))|x − xi | + (α′ (x) − β(x))sgn(x − xi ) ϕj (x)dΩ
)
(α′′ (xi − rQ ) − β ′ (xi − rQ ))rQ − (α′ (xi − rQ ) − β(xi − rQ )) ϕj (xi − rQ )
)
#
4
(α (x + rQ ) − β (x + rQ ))rQ + (α (x + rQ ) − β(x + rQ )) ϕj (x + rQ ) + O(rQ
),
′′
Z
i
′
i
′
i
i
i
α(x)sgn(x − xi )ϕj,x (x)dΩ
Ωis
"
#
rQ
4
− α(xi − rQ )ϕj,x (xi − rQ ) + α(xi + rQ )ϕj,x (xi + rQ ) + O(rQ
),
=
4
The Bi-conjugate gradient stabilized method is presented in the following pseudo code:
begin
Set x(0) ← 0
Compute r(0) = b − Ax(0)
Set p(0) ← r(0) , q(0) ← 0, v(0) ← 0, ω
b (0) ← 1 , β (0) ← 1 and α(0) ← 1
(i)
repeat i = 1, 2, ... until kr k2 < tolerancekr(0) k2 ;
tolerance = 10−10
βb(i) =< p(i−1) , r(i−1) >,
b(i) (i−1)
ω (i) = ββ(i−1)ωbα(i−1) ,
β (i−1) = βb(i) ,
q(i) = r(i−1) + ω
b (i) (q(i−1) − α(i−1) v(i−1) ),
(i)
(i)
v = Aq ,
b(i)
ω
b (i) = <p(i)β ,v(i) > ,
s(i) = r(i−1) − ω
b (i) v(i) ,
t(i) = As(i) ,
(i) (i)
,s >
,
α(i) = <t
<t(i) ,t(i) >
x(i) = x(i−1) + ω
b (i) q(i) + α(i) s(i) ,
(i)
(i)
(i) (i)
r =s −α t
end repeat
x = x(i) ,
end
16
3.6. Stability analysis
In this section, we present an analysis of the stability of the LBIE and LRPI schemes. Initially, we
consider the LBIE scheme. In this scheme, the solution at any time level can be obtained using Eqs. (3.24)
and (3.49)
e k = φ max{P−1 Qφ−1 U
e k+1 + P−1 H
b k , Π},
e
U
(3.52)
where φ is the (N + 1) × (N + 1) banded matrix with bandwidth bw which is easily obtained using the
relation (3.24). Then we have
e k = φU
b k,
U
and also we have
e k = φΠ
b k.
Π
e l . Assume that
By choosing k = l and using (3.52), we get U
bl = [ U
b0l
U
b1l
U
b2l
U
b l ]T .
... U
N
Also, let Ule be the exact solution at the lth time level with the following components
l
Ule = [ Ue0
l
Ue1
l
Ue2
l
. . . UeN
]T .
e l is either smaller than U l or greater than it i.e.
It is well-known that for any i = 0, 1, ..., N , U
i
ei
l
eil < Uei
U
,
l
eil ≥ Uei
U
,
or
∀ i = 0, 1, . . . , N.
e l and Ul which have the following property
Case 1: Firstly, we consider the vector components of U
e
l
eil ≥ Uei
U
,
e l and 1 Ul as follow
let us define the vectors 1 U
e
(
e l, U
el ≥ Ul ,
U
i
i
ei
el
1 Ui =
0,
otherwise,
l
1 Uei
=
(
l
el ≥ Ul ,
Uei
, U
i
ei
0,
otherwise.
e l as follows
Eq. (3.52) can be rewritten using the vector 1 U
el
1U
e l+1 + MP−1 H
b l , MΠ},
e
= φ max{P−1 Qφ−1 1 U
(3.53)
where M is a (N + 1) × (N + 1) matrix
Mij =
(
el ≥ Ul ,
1, i = j, and U
i
ei
0, otherwise.
(3.54)
The error El1 at the lth time level is given by
e l − 1 Ul .
El1 = 1 U
e
(3.55)
el
1U
(3.56)
It is important to observe that all components of El1 are positive values, also we conclude
= El1 + 1 Ule .
Using the relations (3.53) and (3.56), we get
e
b l + P−1 Qφ−1 El+1 , MΠ},
El1 + 1 Ule = φ max{P−1 Qφ−1 1 Ul+1
+ MP−1 H
e
1
17
(3.57)
we can easily see that the relation (3.57) is transformed to the following equation using the maximum
function properties:
El1 + 1 Ule
≤
b l , MΠ}
e
φ max{P−1 Qφ−1 1 Ul+1
+ MP−1 H
e
+φ max{P−1 Qφ−1 El+1
1 , O},
(3.58)
where O is the zero vector. Also we know that
l
1 Ue
b l , MΠ}.
e
= φ max{P−1 Qφ−1 1 Ul+1
+ MP−1 H
e
(3.59)
Therefore, using (3.58) and (3.59) we can write
El1 ≤ φ max{P−1 Qφ−1 El+1
1 , O}.
(3.60)
Finally, we obtain [71]
||El1 || ≤
≤
||φ max{P−1 Qφ−1 El+1
1 , O}||
||φP−1 Qφ−1 El+1
||
≤
||φP−1 Qφ−1 ||||El+1
1
1 ||,
(3.61)
or, equivalently
||El1 || ≤ ||φP−1 Qφ−1 ||||El+1
1 ||.
(3.62)
e l and Ul which have the following property
Case 2. Now, we consider the vector components of U
e
el < Ul ,
U
i
ei
e l and 2 Ul are two vectors defined by
suppose that 2 U
e
(
el < Ul ,
e l, U
U
ei
i
i
l
e
2 Ui =
0,
otherwise,
l
2 Uei
=
(
l
el < Ul ,
, U
Uei
ei
i
0,
otherwise.
e l only
Eq. (3.52) can be rewritten such to contain 2 U
el
2U
e l+1 + NP−1 H
b l , NΠ},
e
= φ max{P−1 Qφ−1 2 U
where N is a (N + 1) × (N + 1) matrix defined by
(
el < Ul ,
1, i = j, and U
i
ei
Nij =
0, otherwise.
(3.63)
(3.64)
In this case we propose the error El2 at the lth time level
It is clear that El2 is hold as
e l − 2 Ul .
El2 = 2 U
e
(3.65)
El2 ≥ 0.
(3.66)
= 2 Ule − El2 .
(3.67)
By using relation (3.65), we obtain
el
2U
Therefore, the relation (3.63) converted to the following equation
l
2 Ue
b l − P−1 Qφ−1 El+1 , NΠ}.
e
+ NP−1 H
− El2 = φ max{P−1 Qφ−1 2 Ul+1
e
2
18
(3.68)
Moreover, using the maximum function property, we have
l
2 Ue
− El2
≥
b l , NΠ}
e
φ max{P−1 Qφ−1 2 Ul+1
+ NP−1 H
e
−φ max{P−1 Qφ−1 El+1
2 , O},
(3.69)
0 ≤ El2 ≤ φ max{P−1 Qφ−1 El+1
2 , O}.
(3.70)
or
Then, it follows from the norm and maximum property that [71]
−1
||El2 || ≤ ||φ max{P−1 Qφ−1 El+1
Qφ−1 El+1
2 , O}|| ≤ ||φP
2 ||
≤ ||φP−1 Qφ−1 ||||El+1
2 ||,
(3.71)
or, equivalently
||El2 || ≤ ||φP−1 Qφ−1 ||||El+1
2 ||.
(3.72)
The numerical scheme will be stable if l → ∞, the error ||El1 || → 0 and ||El2 || → 0 [71]. This can be
guaranteed provided ρ(φP−1 Qφ−1 ) ≤ 1 or ρ(P−1 Q) ≤ 1 (because P−1 Q and φP−1 Qφ−1 are similar
matrices), where ρ denotes the spectral radius of the matrix.
For the analysis, we need a simple version of the matrix P. It is given by
e − (θ − 1)C
e + (θ − 1)D
e + (θ − 1)E
e + Z,
e + γ2 B
P = (θ − 1)A
e B,
e C,
e D
e and E
e are the (N + 1) × (N + 1) banded matrix with bandwidth bw whose first and last
where A,
rows of them contain only zero elements and the other rows are obtain using relation (3.48). Also in the
above relation, Z is an (N + 1) × (N + 1) matrix
ϕj (x0 ), i = 0,
Zij = ϕj (xN ), i = N,
0,
otherwise,
where x0 = 0 and xN = 1.
Lemma 1. In the MLS approximation, if we put
m = 2,
rQ = 0.51h,
rw = 4rQ ,
then the following relations hold:
ϕj (x0 ) = δ0j ,
ϕj (xN ) = δN j ,
where δ is the Kronecker delta.
Proof. We can write the relation (3.25) as follows
ϕj (x0 ) =
2
X
pk (0)[A−1 (0)B(0)]kj = [A−1 (0)B(0)]1j .
k=1
Also using section 3.4.1, we have
B(x) =
and
−1
A
w1 (x)
x1 w1 (x)
w2 (x)
x2 w2 (x)
...
wn (x)
. . . xn wn (x)
,
Pn
x2i wi (x)
(x) = "
n
#"
# "
#2 Pi=1
i=1 xi wi (x)
Pn
Pn
Pn
2 w (x) −
w
(x)
x
x
w
(x)
i=1 i
i=1 i i
i=1 i i
1
19
P
− Pni=1 xi wi (x)
.
n
i=1 wi (x)
Then, one has
"
wj (0)
−1
[A
(0)B(0)]1j = "
Pn
i=1
Pn
2
i=1 xi wi (0)
#"
wi (0)
Pn
i=1
− xj
#
Pn
i=1
x2i wi (0) −
"
#
xi wi (0)
Pn
i=1
#2 .
(3.73)
xi wi (0)
It is exactly clear that to boundary points x0 and xN , we have n = ⌊rw ⌋ + 1, then here n = 3. Indeed, we
have xj = (j − 1)h, j = 1, 2, 3. Eq. (3.73) is calculated as follows
"
#
P3
P3
2 2
wj (0)
i=1 (i − 1)hwi (0)
i=1 (i − 1) h wi (0) − (j − 1)h
[A−1 (0)B(0)]1j = "
P3
i=1
#"
wi (0)
P3
i=1 (i
#
− 1)2 h2 wi (0) −
"
P3
i=1 (i
#2 .
− 1)hwi (0)
With a little computation, we conclude that
"
#
wj (0) (2 − j)w2 (0) + 2(3 − j)w3 (0)
−1
[A
(0)B(0)]1j = "
#,
w1 (0)w2 (0) + w2 (0)w3 (0) + 4w1 (0)w3 (0)
on the other hand, we know w3 (0) = 0, then
[A−1 (0)B(0)]1j =
(2 − j)wj (0)
.
w1 (0)
Therefore, we obtain
ϕj (x0 ) = [A−1 (0)B(0)]1j = δ1j .
Now, we want to show ϕj (xN ) = δN j . Using relation (3.25), we have
ϕj (xN ) = [A−1 (1)B(1)]1j + [A−1 (1)B(1)]2j ,
similar to the above discussion we can claim the following formula
[A−1 (1)B(1)]1j =
wj (1)[(1 − 2h)(1 − j)w1 (1) + (1 − h)(2 − j)w2 (1) + (3 − j)w3 (1)]
,
hw3 (1)(4w1 (1) + w2 (1))
and
[A−1 (1)B(1)]2j =
wj (1)[(j − 1)w1 (1) + (j − 2)w2 (1) + (j − 3)w3 (1)]
.
hw3 (1)(4w1 (1) + w2 (1))
Also, we get
ϕj (xN ) =
wj (1)[2(j − 1)w1 (1) + (j − 2)w2 (1)]
.
w3 (1)(4w1 (1) + w2 (1))
It is well-known that w1 (1) = 0, therefore we have
ϕj (xN ) = ϕj (1) = δjN .
20
According to Lemma 1 we can write
δ1j ,
Zij = δN j ,
0,
then we obtain
i = 0,
i = N,
otherwise,
P = (θ − 1)S +
Q = θS +
where
1 e
B,
∆t
(3.74)
1 e
B + Z,
∆t
(3.75)
e − rB
e −C
e +D
e + E.
e
S=A
We simply observe that the first and last rows of Q contain only zero elements, so using Gershgorin theorem
all the eigenvalues of Q are in Q, i.e.
0 0 ... 0 0
,
Q(x) =
Q
0 0 ... 0 0
where
Q = [Q2 Q3 ... QN ]T .
Also, using the relation (3.74) and (3.75), we can see that
1 0 ... 0
P(x) =
P
0 0 ... 0
0
1
where
,
P = [P2 P3 ... PN ]T .
It is also worth noticing that one of the eigenvalues of P is λ = 1 i.e.
|λ1 − 1| ≤ 0,
→
|λN +1 − 1| ≤ 0,
λ1 = 1,
→
λN +1 = 1.
This is obtained using the Gershgorin theorem on the first and last rows of P. Therefore, the other
−1
eigenvalues of P are in matrix P. However, we can consider the ρ(P Q) instead of ρ(P−1 Q), we have
P = (θ − 1)S +
Q = θS +
or
e
B
on the other hand, we know that
P
−1
Q=P
e
B
−1
−1
−1
1 e
B,
∆t
e
P = (θ − 1)B
e
Q = θB
e B
e
B
−1
−1
−1
S+
e
Q = (B
21
1 e
B,
∆t
S+
1
I,
∆t
(3.76)
1
I,
∆t
−1
P)
−1
e
(B
−1
Q),
let us define
e
Σ=B
−1
P,
e
Γ=B
−1
Q,
therefore, we can rewrite relation (3.77) as follows
e
Υ=B
−1
S,
1
I,
∆t
(3.77)
θ∆tλ(Υ) + 1
λ(Γ)
=
≤ 1,
λ(Σ)
(θ − 1)∆tλ(Υ) + 1
(3.78)
Σ = (θ − 1)Υ +
Γ = θΥ +
1
I.
∆t
Now by applying Cayley-Hamilton theorem, we have
λ(P
−1
Q) = λ(Σ−1 Γ) =
where λ are eigenvalues of the matrices. We can easily see that for implicit Euler (θ = 0) and CrankNicolson (θ = 0.5) schemes, the inequality (3.78) is always satisfied and the scheme will be unconditionally
stable if ρ(Υ) ≤ 0. Figure 2 shows numerically how the spectral radius of Υ varies as a function of N and
M . Recollect that the stability condition is satisfied only when ρ(Υ) ≤ 0. It can be seen from Figure 2
that this condition is satisfied in the present numerical methods i.e. LBIE and LRPI. In this work, we use
Crank-Nicolson scheme which is second-order accuracy with respect to time variable.
Remark 3: Note that the analysis of the stability of the LRPI method is similar to the above analysis.
It should also be noted that the stability analysis for European option is similar to the American option.
4. Numerical results and discussions
The numerical simulations are run on a PC Laptop with an Intel(R) Core(TM)2 Duo CPU T9550 2.66
GHz 4 GB RAM and the software programs are written in Matlab. Let U , ULBIE and ULRP I respectively
denote the option price (either European or American) and its approximation obtained using the LBIE and
LRPI methods developed in the Section 3. To measure the accuracy of the ULBIE and ULRP I methods at
the current time (t = 0), the discrete maximum norm and the mean square norm have been used with the
following definitions:
MaxErrorLBIE =
MaxErrorLRP I =
RMSErrorLBIE
RMSErrorLRP I
max |ULBIE (xi , 0) − U (xi , 0)| ,
i=0,1,...,8
(4.1)
max |ULRP I (xi , 0) − U (xi , 0)| ,
i=0,1,...,8
v
u 8
X
1 u
2
t
=
(ULBIE (xi , 0) − U (xi , 0)) ,
8 + 1 i=0
v
u 8
X
1 u
t
(ULRP I (xi , 0) − U (xi , 0))2 .
=
8 + 1 i=0
(4.2)
In MaxErrorLBIE , MaxErrorLRP I , RMSErrorLBIE and RMSErrorLRP I , xi , i = 0, 1, .., 8 are nine different
points that will be chosen in a convenient neighborhood of the strike E, i.e. xi ∈ ( 54 E, 65 E). For simplicity,
in European and American options we set xi = 8 + 0.5i, i = 0, 1, .., 8 and xi = 80 + 5i, i = 0, 1, .., 8,
respectively. Note that in the case of the American option the exact value of U is not available. Therefore,
in (4.1), (4.2) we use instead a very accurate approximation of it, which is obtained using the LBIE method
with a very large number of nodes and time steps (precisely we set N = 1024 and M = 1024).
In the following analysis, the radius of the local sub-domain is selected rQ = 0.51h, where h is the
distance between the nodes. The size of rQ is such that the union of these sub-domain must cover the whole
global domain i.e. ∪Ωis ⊂ Ω. It is also worth noticing that the MLS approximation is well-defined only when
A is non-singular or the rank of P equals m or at least m weight functions are non-zero i.e. n > m for each
22
x ∈ Ω. Therefore, to satisfy these conditions, the size of the support domain rw should be large enough
to have sufficient number of nodes covered in Ωx for every sample point (n > m). In all the simulations
presented in this work we use rw = 4rQ . Also, in Test cases 1 and 2, we use ξ = 1 and ξ = 0.1, respectively.
These values are chosen by trial and error such as to roughly minimize the errors on the numerical solutions.
To show the rate of convergence of the new schemes when h → 0 and ∆t → 0, the values of ratio with
the following formula have been reported in the tables
)
(
MaxErrorLBIE in the previous row
,
RatioLBIE = log2
MaxErrorLBIE in the current row
(
)
MaxErrorLRP I in the previous row
RatioLRP I = log2
.
MaxErrorLRP I in the current row
Finally, the computer time required to obtain the option price using the numerical method described in
previous section is denoted by CP U T ime.
4.1. Test case 1
To test the proposed numerical schemes, our first test problem is the same test problem as is presented
in [30] and [72]. To be precise, let us consider an European option with the parameters and option data
which are chosen as in Table 1. Moreover, we set Smax = 5E.
Of special interest to us is testifying the numerical convergence of the solution by this problem. Hence,
applying the suggested schemes in this paper together with different choice of N and M , we get the consequences tabulated in Tables 2 and 3. The number of time discretization steps is set equal to nodes
distributed in the domain. As we have experimentally checked, this choice is such that in all the simulations
performed the error due to the time discretization is negligible with respect to the error due to the LBIE
and LRPI discretization (note that in the present work we are mainly concerned with the LBIE and LRPI
spatial approximation).
From Tables 2 and 3, it can be seen that the LBIE and LRPI methods provide very accurate, stable and
fast approximation for option pricing. We observe that the accuracy grows as the number of basis increases
gradually, then the option price can be computed with a small financial error in a small computer time. In
fact, for example, in LBIE scheme the option pricing is computed with at least 3 correct significant digits.
note that, in Table 2, as well as in the following ones, the fact that MaxErrorLBIE is approximately 10−3
means that ULBIE (x, 0) is up to at least the 3th significant digit, equal to U (x, 0). Moreover, the computer
times necessary to perform these simulations are extremely small. In fact, using 32 nodal points in domain
and 32 time discretization steps, European option is computed with an error of order 10−4 in only 7 × 10−3
second, instead, using 128 nodal points and 128 time discretization steps, the option price is computed with
an error of order 10−5 in 0.13 second and using 256 nodal points and 256 time discretization steps, the
option price is computed with an error of order 10−6 in 0.44 second. On the other hand, by looking at
Table 3, it can also be seen that in LRPI scheme error of orders 10−4 , 10−5 and 10−6 are computed using
64, 256 and 1024 nodal points and time discretization steps, respectively in 0.01, 0.42 and 7.07 seconds. We
can observe that the option price computed by LBIE on sub-domains requires approximately 0.44 second
to reach an accuracy of 10−6 compared to the 0.42 second required by LRPI method to reach an accuracy
of 10−5 on the same number of sub-domains. Therefore, we can simply conclude that in this case LBIE
method is more accurate than LRPI method. As the final look at the numerical results in Tables 2 and 3,
MaxErrorLRP I
we can see MaxError
≈ 4, i.e. LBIE scheme is approximately 4 times accurate than LRPI scheme. Also
LBIE
rate of convergence of LBIE is 2 but in LRPI, it is approximately 2.
Remark 4: The error estimates of MLS approximation are given in [73] and [74] using different strategies. They have proved that the error of this approximation is of order O(∆xm ). According to theoretical
bounds, the ratios should be approximately m. As we can see in Ratio of Table 2, the numerical results confirm the analytical bounds. Anyway, the presented schemes have nearly the order 2 (Note that in previous
section we consider m = 2).
Comparison with previous works: We can compare the prices obtained from the two meshless
methods proposed in this paper with a solution computed in [26]. Hon and his co-author [26] proposed
Kansa collocation method combined with method of lines (MOL) for the pricing of European options.
23
Kansa collocation method is a well-known strong form meshless method based on radial basis functions
(especially multi-quadratic radial basis function which is used in [26]) and also the method of lines is
a general way of viewing a time-dependent partial differential equation (PDE) as a system of ordinary
differential equations (ODEs). The partial derivatives with respect to the space variables are discretized to
obtain a system of ODEs in the variable t. This system could be obtained using an explicit fourth order
backward time integration scheme (RK4). However, according to what was reported in [26], the numerical
proposed therein allows us to obtain the European option price with the RM SError = 3 × 10−4 using
N = 81 and M = 30. In [26], the authors say no any things about the computer time of algorithm but in
the other paper of these authors [30], they said that the computer time of Qusi-RBF method for European
option pricing which is 12 seconds, is much less than method proposed in [26]. The main reasons of why the
our numerical methods are more efficient than the one proposed by [26] are the following: (1) In contrast to
the sparse and banded matrices associated with local weak form meshless methods proposed in this paper,
the matrices associated to system of the method in [26] usually leads to a fully populated matrices. (2)
The approach followed in [26] is based on the MOLs, which allows one to reduce the model to a system
of ordinary differential equations (ODEs). However the system of ODEs is solved using RK4 which is a
iterative method, in particular, to reach convergence, such an iterative procedure requires us to perform at
each time step, a long number of iterations, thus reducing the efficiency of the numerical scheme. However,
according to what is reported in this paper, the numerical methods proposed do not have any requirement
to solve a system of ODEs. (3) Looking at Tables 2 and 3 and reported result in [26], we can see that
the solutions of [26] are slower and more inaccurate than our results. Putting all these things together
(comment 2 and this comment), we conclude that the numerical method proposed in this paper is very fast
and accurate than the method developed in [26]. (4) More importantly, in [26] there aren’t any stability
analysis of scheme. (5) The convergence error of approximated solution obtained by [26] is the same as
reported by Remark 4, i.e. O(∆x2 ), but it can be increased using our method by increasing value of m.
(please see Remark 4).
Let us compare the numerical schemes proposed in this work with that proposed in [30]. Again, in [30],
Hon et al. solved this problem using combined MOL method and a strong form meshless method named as
Qusi-RBF method. The method is more better than the proposed method in [26], because, the coefficient
matrix of the final system is sparse and banded. In [30], the most significant result obtained in Test case 1 is
the following: the RM SError is 4 × 10−5 in a computer time 12 seconds using M = 100 and N = 2000. Let
us investigate the reasons why the numerical technique proposed in this paper performs significantly better
than the method presented in [30]. (1) The final linear system of our work and [30] is a system using banded
coefficient matrix with bandwidth 2P + 1, but value of P in our work and [30] is 5 and 20, respectively.
Therefore, the coefficient matrix obtained using LBIE and LRPI schemes is very more sparse than [30].
(2) In this work, we solve at every time step a sparse and banded linear system using a powerful iterative
algorithm named Bi-conjugate gradient stabilized method with complexity O(2N (2bw + 1)K), whereas in
[30], the authors solved the final linear system with a direct method named banded LU factorization method
with partial pivoting with complexity O(2N P (P + 2)). In fact, for example, suppose that in [30] we get
P = 20 and in our meshless methods we get bw = 5 and K = 5. We can easily see that the complexity
of [30] is more bigger that our schemes. (3) Authors of [30] did not present the stability analysis for their
method. (4) Looking at Tables 2 and 3 and also reported numerical result in [30], we observed that solution
of [30] is slower and more inaccurate than our results. In fact, for example, in [30] option price is founded
with RMSError 4 × 10−5 in a computer time 12 seconds using N = 2000 and M = 100, whereas we compute
the option price with RMSErrorLBIE 3.45 × 10−5 in a computer time 0.13 in LBIE using N = 128 and
M = 128 and also RMSErrorLRP I 4.37 × 10−5 in a computer time 0.42 in LRPI second using N = 256 and
M = 256. Putting all these things together we conclude that the numerical methods proposed in this paper
are at least thirty of times faster and more efficient than the method of [30].
4.2. Test case 2
For the second test problem consider an American option with the useful data which are provided in
the Table 4. As done in Test case 1, we suppose Smax = 5E. The results of implementing the problem by
utilizing the present methods with various number of sub-domains and time discretization steps are shown
in Tables 5 and 6. Overall, as already pointed out, following the numerical findings in the different errors,
convinces us that the error of the proposed techniques decrease very rapidly as the number of sub-domains in
24
domain and time discretization steps increase. In fact, for example, in LBIE and LRPI schemes the price of
the American option can be computed with 3 correct significants in only 0.87 and 0.92 second, respectively
which is excellent and very fast. We emphasize that the ratio shown in Tables 5 and 6 are approximately
1.5 order accuracy instead of second order. This is thanks to the fact that this model is a free boundary
problem, therefore its solution is non-smoothness which spoils the accuracy of the spatial approximation
(again we have experimentally checked that the time discretization alone is second-order accurate and thus
the low rate of convergence is due to the error of the spatial approximation).
Remark 5: We believe that, in this option model we can do a comparison between our schemes and
numerical methods proposed in [30] and [26]. Firs of all, we emphasize that approximately this comparison
is substantially analogous with presented comparison between theirs in Test case 1 and requires only minor
changes. In particular, looking at Tables 5 and 6 and what is reported in [30] and [26] we simply observe that
the numerical methods proposed in this paper perform much better than the numerical methods proposed
in [30] and [26]. In fact, for example, the LBIE and LRPI allow us to compute the American option price
with the errors 2.08 × 10−3 and 2.02 × 10−3 in 5.13 and 6.49 seconds, respectively (Tables 5 and 6) using
N = 512 and M = 512, whereas, it takes 47 seconds for the quasi-RBF scheme to compute the option price
with the error 3.4 × 10−3 using N = 2000 and M = 500. Putting all these things together, we conclude
that the numerical method proposed in this paper is approximately seven times faster than the method
developed in [30].
5. Conclusions
In this paper, the weak form meshless methods namely local boundary integral equation method (LBIE)
based on moving least squares approximation (MLS) and local radial point interpolation (LRPI) based on
Wu’s compactly supported radial basis functions (WCS-RBFs), were formulated and successfully applied to
price European and American options under the Block-Scholes model.
Overall the numerical achievements that should be highlighted here are as given in the following:
(1) The price of American option is computed by Richardson extrapolation of the price of Bermudan
option. In essence the Richardson extrapolation reduces the free boundary problem and linear complementarity problem to a fixed boundary problem which is much simpler to solve. Thus, instead of describing the
aforementioned linear complementarity problem or penalty method, we directly focus our attention onto the
partial differential equation satisfied by the price of a Bermudan option which is faster and more accurate
than other methods.
(2) The infinite space domain is truncated to [0, Smax ] with a sufficiently large Smax to avoid an unacceptably large truncation error. The options’ payoffs considered in this paper are non-smooth functions, in
particular their derivatives are discontinuous at the strike price. Therefore, to reduce as much as possible
the losses of accuracy the points of the trial functions are concentrated in a spatial region close to the strike
prices. So, we employ the change of variables proposed by Clarke and Parrott [60].
(3) We used the θ-weighted scheme. Stability analysis of the method is analyzed and performed by the
matrix method in the present paper. In fact, according to an analysis carried out in the present paper, the
time semi-discretization is unconditionally stable for implicit Euler (θ = 0) and Crank-Nicolson (θ = 0.5)
schemes.
(4) Up to now, only strong form meshless methods based on radial basis functions (RBFs) and quasiRBF have been used for option pricing in mathematical finance field. These techniques yield high levels
of accuracy, but have of a very serious drawback such as produce a very ill-conditioned systems and very
sensitive to the select of collocation points. Again we do emphasize that in the new methods presented in this
manuscript, coefficient matrix of the linear systems are sparse and banded with bandwidth bw = 2⌊rw ⌋ + 1.
It should be noted that these coefficient matrices are positive definite.
(5) The present methods are truly a meshless method, which do not need any element or mesh for both
field interpolation and background integration.
(6) Meshless methods using global RBFs such as Gaussian and multiquadric RBFs have a free parameter
known as shape parameter. Despite many research works which are done to find algorithms for selecting
the optimum values of ǫ [67, 68, 69, 75, 76], the optimal choice of shape parameter is an open problem
which is still under intensive investigation. Cheng et al. [67] showed that when ǫ is very small then the
RBFs system error is of exponential order but the solution breaks down when ǫ is smaller than the limiting
25
value. In general, as the value of the shape parameter ǫ decreases, the matrix of the system to be solved
becomes highly ill-conditioned. To overcome this drawback of the global RBFs, the local RBFs such as Wu,
Wendland and Buhmann compactly supported radial basis functions which are local and stable functions,
are proposed which are applied in this work.
(7) In the LBIE method, it is difficult to enforce the boundary conditions for that the shape function
constructed by the MLS approximation lacks the delta Kronecker property but in LRPIM, using the delta
Kronecker property, the boundary conditions can be easily imposed. In LBIE, we use the collocation method
to the nodes on the boundary.
(8) A crucial point in the weak form meshless methods is an accurate evaluation of the local integrals.
Based on MLS and RBFs, the nodal trial functions are highly complicated, hence an accurate numerical
integration of the weak form is highly difficult. In this work, the numerical integration procedure used is
Simpson’s rule, which is well-known to be fourth-order accurate.
(9) It should be noted that LBIE and LRPI schemes lead to banded and sparse system matrices. Therefore, we use a powerful iterative algorithm named the Bi-conjugate gradient stabilized method (BCGSTAB)
to get rid of this system. Numerical experiments are presented showing that the LBIE and LRPI approaches
are extremely accurate and fast.
(10) To demonstrate the accuracy and usefulness of these methods, some numerical examples were
presented. For all test cases, the RMSError, MaxError, Ratio and CPU Time of schemes to various nodal
points in domain and time discretization steps, were reported. A good agreement between the results for
the LBIE and LRPI techniques, and the solutions obtained by other numerical schemes in literature [30]
and [26], was observed clearly.The results of our numerical experiments, confirm the validity of the new
techniques.
26
References
References
[1] F. Black, M. Scholes, The pricing of options and corporate liabilities, J. Polit. Econ. 81 (1973) 637–659.
[2] M. Brennan, E. Schwartz, Finite difference methods and jump processes arising in the pricing of
contingent claim: A synthesis, J. Financ. Quant. Anal. 13 (1978) 461–474.
[3] C. Vazquez, An upwind numerical approach for an American and European option pricing model, Appl.
Math. Comput. 97 (1998) 273–286.
[4] X. Wu, W. Kong, A highly accurate linearized method for free boundary problems, Comput. Math.
Appl. 50 (2005) 1241–1250.
[5] A. Arciniega, E. Allen, Extrapolation of difference methods in option valuation, Appl. Math. Comput.
153 (2004) 165–186.
[6] M.Yousuf, A. Q. M. Khaliq, B. Kleefeld, The numerical approximation of nonlinear Black-Scholes
model for exotic path-dependent American options with transaction cost, I. J. Comput. Math. 89
(2012) 1239–1254.
[7] R. Zvan, P. A. Forsyth, K. R. Vetzal, A general finite element approach for PDE option pricing models,
Ph.D. thesis, University of Waterloo, Waterloo (1998).
[8] L. V. Ballestra, C. Sgarra, The evaluation of American options in a stochastic volatility model with
jumps: An efficient finite element approach, Comput. Math. Appl. 60 (2010) 1571–1590.
[9] L. V. Ballestra, L. Cecere, A numerical method to compute the volatility of the fractional Brownian
motion implied by American options, Int. J. Appl. Math. 26 (2013) 203–220.
[10] P. A. Forsyth, K. R. Vetzal, Quadratic convergence for valuing American options using a penalty
method, SIAM J. Sci. Comput. 23 (2002) 2095–2122.
[11] R. Zvan, P. A. Forsyth, K. Vetzal, A Finite Volume approach for contingent claims valuation, IMA J.
Numer. Anal. 21 (2001) 703–731.
[12] S. J. Berridge, J. M. Schumacher, An irregular grid approach for pricing high-dimensional American
options, J. Comput. Appl. Math. 222 (2008) 94–111.
[13] A. Q. M. Khaliq, D. A. Voss, S. H. K. Kazmi, Adaptive θ-methods for pricing American options, J.
Comput. Appl. Math. 222 (2008) 210–227.
[14] A. Q. M. Khaliq, D. A. Voss, S. H. K. Kazmi, A fast high-order finite difference algorithm for pricing
American options, J. Comput. Appl. Math. 222 (2008) 17–29.
[15] B. F. Nielsen, O. Skavhaug, A. Tveito, Penalty methods for the numerical solution of American multiasset option problems, J. Comput. Appl. Math. 222 (2008) 3–16.
[16] J. Zhao, M. Davison, R. M. Corless, Compact finite difference method for American option pricing, J.
Comput. Appl. Math. 206 (2007) 306–321.
[17] D. Tangman, A. Gopaul, M. Bhuruth, Numerical pricing of options using high-order compact finite
difference schemes, J. Comput. Appl. Math. 218 (2008) 270–280.
[18] B. Hu, J. Liang, L. Jiang, Optimal convergence rate of the explicit finite difference scheme for American
option valuation, J. Comput. Appl. Math. 230 (2009) 583–599.
[19] S. Zhang, L. Wang, A fast numerical approach to option pricing with stochastic interest rate, stochastic
volatility and double jumps, Commun. Nonlinear Sci. Numer. Simulat. 18 (2013) 1832–1839.
27
[20] J. C. Cox, S. A. Ross, M. Rubinstein, Option pricing: A simplified approach, J. Financ. Econ. 7 (1979)
229–263.
[21] M. Broadie, J. Detemple, American option valuation: New bounds, approximations, and a comparison
of existing methods, Rev. Financ. Stud. 9 (1996) 1211–1250.
[22] M. Gaudenzi, F. Pressacco, An efficient binomial method for pricing American put options, Decis.
Econ. Finance 4 (2003) 1–17.
[23] S. L. Chung, C. C. Chang, R. C. Stapleton, Richardson extrapolation techniques for the pricing of
American-style options, J. Futures Markets 27 (2007) 791–817.
[24] A. Khaliq, G. Fasshauer, D. Voss, Using meshfree approximation for multi-asset American option
problems, J. Chinese Institute Engineers 27 (2004) 563–571.
[25] L. V. Ballestra, G. Pacelli, Pricing European and American options with two stochastic factors: A
highly efficient radial basis function approach, J. Econ. Dyn. Cont. 37 (2013) 1142–1167.
[26] Y. C. Hon, X. Mao, A radial basis function method for solving options pricing models, Financ. Eng. 8
(1999) 31–49.
[27] A. Golbabai, D. Ahmadian, M. Milev, Radial basis functions with application to finance: American
put option under jump diffusion, Math. Comp. Modelling 55 (2012) 1354–1362.
[28] Z. Wu, Y. C. Hon, Convergence error estimate in solving free boundary diffusion problem by radial
basis functions method, Eng. Anal. Bound. Elem. 27 (2003) 73–79.
[29] M. D. Marcozzi, S. Choi, C. S. Chen, On the use of boundary conditions for variational formulations
arising in financial mathematics, App. Math. Comput. 124 (2003) 197–214.
[30] Y. C. Hon, A quasi-radial basis functions method for American options pricing, Comput. Math. Appl.
43 (2002) 513–524.
[31] A. Shokri, M. Dehghan, A Not-a-Knot meshless method using radial basis functions and predictorcorrector scheme to the numerical solution of improved Boussinesq equation, Comput. Phys. Commun.
181 (2010) 1990–2000.
[32] M. Dehghan, R. Salehi, A boundary-only meshless method for numerical solution of the Eikonal equation, Comput. Mech. 47 (2011) 283–294.
[33] M. Dehghan, A. Shokri, Numerical solution of the nonlinear Klein-Gordon equation using radial basis
functions, J. Comput. Appl. Math. 230 (2009) 400–410.
[34] K. Parand, J. A. Rad, Kansa method for the solution of a parabolic equation with an unknown
spacewise-dependent coefficient subject to an extra measurement, Comput. Phys. Commun., (2012)
184 (2013) 582–595.
[35] S. Kazem, J. A. Rad, K. Parand, Radial basis functions methods for solving Fokker-Planck equation,
Eng. Anal. Bound. Elem. 36 (2012) 181–189.
[36] S. Kazem, J. A. Rad, K. Parand, A meshless method on non-Fickian flows with mixing length growth
in porous media based on radial basis functions, Comput. Math. Appl. 64 (2012) 399–412.
[37] K. Rashedi, H. Adibi, J. Rad, K. Parand, Application of meshfree methods for solving the inverse
one-dimensional Stefan problem, Eng. Anal. Bound. Elem. 40 (2014) 1–21.
[38] A. A. E. F. Saib, D. Y. Tangman, M. Bhuruth, A new radial basis functions method for pricing
American options under Merton’s jump-diffusion model, I. J. Comput. Math. 89 (2012) 1164–1185.
[39] L. V. Ballestra, G. Pacelli, A radial basis function approach to compute the first-passage probability
density function in two-dimensional jump-diffusion models for financial and other applications, Eng.
Anal. Bound. Elem. 36 (2012) 1546–1554.
28
[40] L. V. Ballestra, G. Pacelli, Computing the survival probability density function in jump-diffusion
models: A new approach based on radial basis functions, Eng. Anal. Bound. Elem. 35 (2011) 1075–
1084.
[41] M. Dehghan, A. Ghesmati, Numerical simulation of two-dimensional sine-gordon solitons via a local
weak meshless technique based on the radial point interpolation method (RPIM), Comput. Phys.
Commun. 181 (2010) 772–786.
[42] D. Mirzaei, M. Dehghan, New implementation of MLBIE method for heat conduction analysis in
functionally graded materials, Eng. Anal. Bound. Elem. 36 (2012) 511–519.
[43] R. Salehi, M. Dehghan, A generalized moving least square reproducing kernel method, J. Comput.
Appl. Math. 249 (2013) 120–132.
[44] R. Salehi, M. Dehghan, A moving least square reproducing polynomial meshless method, Appl. Numer.
Math. 69 (2013) 34–58.
[45] M. Dehghan, A. Shokri, Implementation of meshless LBIE method to the 2D non-linear SG problem,
, volume 79 (2009) pages 1662-1682 ., I. J. Numer. Meth. Eng. 79 (2009) 1662–1682.
[46] T. Zhu, J. D. Zhang, S. N. Atluri, A local boundary integral equation (LBIE) method in computational
mechanics, and a meshless discretization approach, Comput. Mech. 21 (1998) 223–235.
[47] J. Wang, G. Liu, A point interpolation meshless method based on radial basis functions, Int. J. Numer.
Meth. Eng. 54 (2002) 1623–1648.
[48] G. Liu, Y. Gu, An Introduction to Meshfree Methods and Their Programing, Springer, Netherlands,
2005.
[49] S. N. Atluri, H. G. Kim, J. Y. Cho, A critical assessment of the truly meshless local Petrov-Galerkin
(MLPG), and local boundary integral equation (LBIE) methods, Comput. Mech. 24 (1999) 348–372.
[50] T. Zhu, J. D. Zhang, S. N. Atluri, A meshless local boundary integral equation (LBIE) for solving
nonlinear problems, Comput. Mech. 22 (1998) 174–186.
[51] T. Zhu, J. D. Zhang, S. N. Atluri, A meshless numerical method based on the local boundary integral
equation (LBIE) to solve linear and non-linear boundary value problems, Eng. Anal. Bound. Elem. 23
(1999) 375–389.
[52] J. Sladek, V. Sladek, C. Zhang, Transient heat conduction analysis in functionally graded materials by
the meshless local boundary integral equation method, Comput. Mat. Sci. 28 (2003) 494–504.
[53] J. Sladek, V. Sladek, J. Krivacek, C. Zhang, Local BIEM for transient heat conduction analysis in 3-D
axisymmetric functionally graded solids, Comput. Mech. 32 (2003) 169–176.
[54] M. Dehghan, D. Mirzaei, Meshless local boundary integral equation (LBIE) method for the unsteady
magnetohydrodynamic (MHD) flow in rectangular and circular pipes, Comput. Phys. Commun. 180
(2009) 1458–1466.
[55] A. Shirzadi, V. Sladek, J. Sladek, A local integral equation formulation to solve coupled nonlinear
reaction-diffusion equations by using moving least square approximation, Eng. Anal. Bound. Elem. 37
(2013) 8–14.
[56] S. M. Hosseini, V. Sladek, J. Sladek, Application of meshless local integral equations to two dimensional
analysis of coupled non-fick diffusion-elasticity, Eng. Anal. Bound. Elem. 37 (2013) 603–615.
[57] V. Sladek, J. Sladek, Local integral equations implemented by MLS-approximation and analytical
integrations, Eng. Anal. Bound. Elem. 34 (2010) 904–913.
[58] X. Li, Meshless galerkin algorithms for boundary integral equations with moving least square approximations, Appl. Numer. Math. 61 (2011) 1237–1256.
29
[59] Z. Wu, Compactly supported positive definite radial functions, Adv. Comput. Math. 4 (1995) 283–292.
[60] N. Clarke, K. Parrott, Multigrid for American option pricing with stochastic volatility, Appl. Math.
Finance 6 (1999) 177–195.
[61] J. C. Hull, Options, Futures, Other Derivatives, 7th ed., Prentice Hall, University of Toronto, 2002.
[62] C. C. Chang, J. B. Lin, W. C. Tsai, Y. H. Wang, Using Richardson extrapolation techniques to price
American options with alternative stochastic processes, Review of quantitative finance and accounting
39 (2012) 383–406.
[63] P. Wilmott, J. Dewynne, S. Howison, Option Pricing: Mathematical Models and Computation, Oxford
Financial Press, 1996.
[64] M. Dehghan, D. Mirzaei, Meshless local Petrov-Galerkin (MLPG) method for the unsteady magnetohydrodynamic (MHD) flow through pipe with arbitrary wall conductivity, Appl. Numer. Math 59
(2009) 1043–1058.
[65] M. Dehghan, D. Mirzaei, The meshless local Petrov-Galerkin MLPG method for the generalized twodimensional non-linear Schrodinger equation, Eng. Anal. Bound. Elem. 32 (2008) 747–756.
[66] M. D. Buhmann, Radial Basis Functions: Theory and Implementations, Cambridge University Press,
New York, 2004.
[67] A. H. D. Cheng, M. A. Golberg, E. J. Kansa, Q. Zammito, Exponential convergence and H-c multiquadric collocation method for partial differential equations, Numer. Meth. Part. D. E. 19 (2003)
571–594.
[68] R. E. Carlson, T. A. Foley, The parameter r2 in multiquadric interpolation, Comput. Math. Appl. 21
(1991) 29–42.
[69] G. Fasshauer, J. Zhang, On choosing “optimal” shape parameters for RBF approximation, Numer.
Algorithms 45 (2007) 346–368.
[70] H. V. der Vorst, BCGSTAB: a fast and smoothly converging variant of BCG for the solution of nonsymmetric linear systems, SIAM J.Sci. Stat. Comp. 18 (1992) 631–634.
[71] B. N. Datta, Numerical Linear Algebra and Applications, 2th ed., SIAM, 2010.
[72] P. Wilmott, S. Howison, J. Dewynne, The Mathematics of Financial Derivatives, Cambridge University
Press, 1995.
[73] H. Wendland, Scattered Data Approximation, Cambridge University Press, New York, 2005.
[74] D. Mirzaei, R. Schaback, M. Dehghan, On generalized moving least squares and diffuse derivatives,
IMA J. Numer. Anal. 32 (2012) 983–1000.
[75] S. Rippa, An algorithm for selecting a good parameter c in radial basis function interpolation, Advan.
Comp. Math. 11 (1999) 193–210.
[76] A. E. Tarwater, A parameter study of Hardy’s multiquadric method for scattered data interpolation,
Report UCRL-53670, Lawrence Livermore National Laboratory, 1985.
30
Table 1: Test case 1, model parameters and data
Volatility
0.2 year−0.5
Interest rate
0.05 year−1
Strike price
10
Maturity
0.5 year
Table 2: Test case 1, efficiency of the LBIE scheme
N
M
16
32
64
128
256
512
1024
16
32
64
128
256
512
1024
N
M
16
32
64
128
256
512
1024
16
32
64
128
256
512
1024
RMSErrorLBIE
3.17 × 10−3
2.42 × 10−4
5.90 × 10−5
1.46 × 10−5
3.61 × 10−6
8.78 × 10−7
2.09 × 10−7
MaxErrorLBIE
6.94 × 10−3
5.55 × 10−4
1.38 × 10−4
3.45 × 10−5
8.54 × 10−6
2.10 × 10−6
5.06 × 10−7
RatioLBIE
−
3.64
2.01
2.00
2.01
2.02
2.05
CPU Time (s)
RatioLRP I
−
1.15
1.97
1.97
1.93
1.89
1.79
CPU Time (s)
0.005
0.007
0.02
0.13
0.44
3.79
6.34
Table 3: Test case 1, efficiency of the LRPI scheme
RMSErrorLRP I
2.17 × 10−3
1.11 × 10−3
2.82 × 10−4
7.08 × 10−5
1.77 × 10−5
4.40 × 10−6
1.09 × 10−6
MaxErrorLRP I
5.70 × 10−3
2.57 × 10−3
6.54 × 10−4
1.67 × 10−4
4.37 × 10−5
1.18 × 10−5
3.42 × 10−6
31
0.003
0.005
0.01
0.14
0.42
4.11
7.07
Sub−figure 3
Sub−figure 2
Sub−figure 1
0
0
0
20
20
20
40
40
40
60
60
60
80
80
80
100
100
100
120
120
120
0
50
nz = 639
100
0
50
nz = 891
0
100
50
nz = 1141
100
Figure 1: Plot of sparsity for N = 128, rQ = 0.51h and rw = 4rQ (sub-figure 1), rw = 6rQ (sub-figure 2), rw = 8rQ (sub-figure
3)
−0.064
−0.066
Spectral radius
−0.068
−0.07
−0.072
−0.074
−0.076
LRPI
LBIE
−0.078
−0.08
0
200
400
600
N, M
800
Figure 2: Spectral radius of Υ in LRPI and LBIE methods
32
1000
1200
Table 4: Test case 2, model parameters and data
Volatility
0.3 year−0.5
Interest rate
0.1 year−1
Strike price
Maturity
100
1 year
Table 5: Test case 2, efficiency of the LBIE scheme
N
M
64
128
256
512
64
128
256
512
RMSErrorLBIE
5.79 × 10−3
2.74 × 10−3
1.19 × 10−3
4.13 × 10−4
N
64
128
256
512
M
64
128
256
512
RMSErrorLRP I
8.13 × 10−3
3.49 × 10−3
1.37 × 10−3
4.50 × 10−4
MaxErrorLBIE
3.86 × 10−2
1.67 × 10−2
6.11 × 10−3
2.08 × 10−3
RatioLBIE
−
1.21
1.44
1.55
CPU Time (s)
RatioLRP I
−
1.01
1.44
1.51
CPU Time (s)
0.02
0.21
0.92
6.49
0.05
0.22
0.87
5.13
Table 6: Test case 2, efficiency of the LRPI scheme
MaxErrorLRP I
3.15 × 10−2
1.56 × 10−2
5.75 × 10−3
2.02 × 10−3
33
| 5cs.CE
|
1
When Vehicles See Pedestrians with Phones:
A Multi-Cue Framework for Recognizing
Phone-based Activities of Pedestrians
arXiv:1801.08234v1 [cs.CV] 24 Jan 2018
Akshay Rangesh, Member, IEEE, and Mohan M. Trivedi, Fellow, IEEE
Abstract—The intelligent vehicle community has devoted considerable efforts to model driver behavior, and in particular to
detect and overcome driver distraction in an effort to reduce
accidents caused by driver negligence. However, as the domain
increasingly shifts towards autonomous and semi-autonomous
solutions, the driver is no longer integral to the decision making
process, indicating a need to refocus efforts elsewhere. To this
end, we propose to study pedestrian distraction instead. In
particular, we focus on detecting pedestrians who are engaged in
secondary activities involving their cellphones and similar handheld multimedia devices from a purely vision-based standpoint.
To achieve this objective, we propose a pipeline incorporating
articulated human pose estimation, followed by a soft object
label transfer from an ensemble of exemplar SVMs trained on
the nearest neighbors in pose feature space. We additionally
incorporate head gaze features and prior pose information
to carry out cellphone related pedestrian activity recognition.
Finally, we offer a method to reliably track the articulated pose
of a pedestrian through a sequence of images using a particle
filter with a Gaussian Process Dynamical Model (GPDM), which
can then be used to estimate sequentially varying activity scores
at a very low computational cost. The entire framework is fast
(especially for sequential data) and accurate, and easily extensible
to include other secondary activities and sources of distraction.
Index Terms—Pedestrian activity recognition, exemplar SVMs,
articulated pose tracking, panoramic surround behavior analysis,
highly autonomous vehicles, deep learning, computer vision.
I. I NTRODUCTION
W
ITH the explosion of hand-held device usage globally,
smart phones have made their way into most hands.
This trend is expected to continue as devices get cheaper and
find more utility in our day to day lives. As of 2011, there
were more phones than people in the USA, and internationally,
the number of mobile phone subscriptions is an estimated 5.9
billion. Though such devices are extremely useful and even
indispensable for many, it is this very dependence that is
a major cause of pedestrian distraction, and possible injury.
From here on-wards, we shall make use of the term cellphone
as a placeholder for any hand-held multimedia device that a
pedestrian may interact with.
Distracted walking, like distracted driving, is likely to increase in parallel with the penetration of electronic devices into
the consumer market. Although driver distraction has received
abundant attention since the turn of the century, distraction
among pedestrians is a relatively nascent area of research.
The authors are with the Laboratory for Intelligent and Safe Automobiles,
University of California, San Diego, CA 92092, USA.
email - arangesh, [email protected]
Fig. 1: Odds of failing to display optimal crossing behavior
for different activities [2], along with their 95% confidence
intervals.
This is surprising given that pedestrians are in fact prone to
acting less cautiously when distracted. Furthermore, a recent
report by the Governors Highway Safety Association (GHSA)
reveals a disturbing trend - between the mid-1970s and early
2000s, pedestrian deaths steadily declined, eventually dipping
to around 11 percent of all motor vehicle fatalities. But
since 2009, pedestrian fatalities have actually increased by 15
percent, climbing to 4,735 in 2013. Meanwhile, the percentage
of pedestrians killed while using cell phones has risen, from
less than 1 percent in 2004 to more than 3.5 percent in 2010,
according to [1]. Also, the study shows that the number of
pedestrians injured while on their cellphones has more than
doubled since 2005.
The severity of this phenomenon is further reflected by
the number of studies conducted over the last few years,
each of which arrive at similar conclusions. In a recent study
conducted by Thompson et al. [2], they conclude that nearly
one-third (29.8%) of all pedestrians performed a distracting
activity while crossing, with text messaging associated with
the highest risk among different technological and social factors (Figure 1). Meanwhile, Nasar et al. [1] found that mobilephone related injuries among pedestrians increased relative to
total pedestrian injuries, and paralleled the increase in injuries
for drivers, and in 2010 exceeded those for drivers. The study
by Byington et al. [3] confirms this by a virtual street based
simulation, stating that - while distracted, participants waited
longer to cross the street, missed more safe opportunities to
cross, took longer to initiate crossing when a safe gap was
available, looked left and right less often, spent more time
2
looking away from the road, and were more likely to be hit
or almost hit by an oncoming vehicle. Moreover, it is noted
that the demographic of individuals between ages 18 and 29
is more susceptible to exhibit such behavior. For a detailed
report on the global nature of the pedestrian safety problem
and the inadequacy of current systems in ensuring it, we refer
the reader to [4].
It is also interesting to note that as the emphasis of automobile manufacturers gradually shifts towards more automated
vehicles, so must the emphasis placed on preventing pedestrian
distraction related injuries. In such scenarios, the intelligent
vehicle must be able to gauge the risk associated with each
pedestrian, and demonstrate more caution in avoiding those
with larger risks.
In this study, we focus only on distraction due to technological factors, particularly the use of cellphones for different
tasks, and ignore social impacts such as talking or walking in
a group. To summarize, we propose to classify each of many
pedestrians in an image, into one of 3 activity classes- none,
texting and handheld phone call. We additionally extend this
approach to work on sequences of images, where knowledge
about temporal dynamics can enable faster and more efficient
operation.
The rest of the paper is organized as follows - Section II
briefly outlines some related work in the field. Section III
describes the data we are working with, and the semantic
annotations that are available for use. Section IV details
the proposed methodology to estimate confidence scores of
cellphone based activities for a pedestrian using a single
image, and section V extends this model to predict a score
at every instance for a sequence of pedestrian images. Section
VI lists the experiments carried out, and tabulates each of their
results. Finally, section VII concludes this work.
II. R ELATED W ORK
There is an abundance of work related to human activity
recognition and classification from the last decade. However,
these studies pertain to generic human activities and are
not of much use in studying pedestrian distraction. Even
though there have been quite a few studies that deal with
driver distraction and activity modeling [5]–[8], these models
are not directly applicable to pedestrians because the forms
of distraction and the activities of interest are considerably
different. Nonetheless, there have been ample efforts devoted
to studying pedestrians in the context of path prediction, intent
analysis and action/activity recognition. We briefly go over
these tasks, highlighting how they differ from the goal of this
work. This study may appear similar to our previous work [9]
in terms of the end goal, however, the proposed methodology
is entirely different. We also use a significantly larger dataset
and provide a more exhaustive evaluation in comparison to
[9]. For a more detailed list of studies conducted on humans
around vehicles, we refer the reader to [10].
Path prediction and gait analysis: There have been numerous studies on predicting the trajectories of pedestrians
to prevent collisions and improve surround vehicle safety.
These methods generally ignore high-level semantics (such as
TABLE I: Related work in image based pedestrian safety.
Study
Path
Prediction
Intent
Analysis
Activity
Recognition
(Output Classes)
Moeslund et al. [11]
Gandhi et al. [12]
Goldhammer et al. [13]
Köhler et al. [16]
Madrigal et al. [17]
3
3
3
3
3
3
3
3
-
Schulz et al. [18]
-
3
cross, turn into
road, stop
Bandyopadhyay et al. [19]
Keller et al. [20]
Kooij et al. [21]
3
3
3
3
walking, stopping
walking, stopping
Kataoka et al. [22]
-
3
crossing, walking,
standing, riding
a bicycle
Quintero et al. [23]
3
-
walking, starting,
standing, stopping
Choi et al. [24]
-
-
collective activities
of pedestrians
like crossing,
waiting, queuing,
walking, talking
Rangesh et al. [9]
-
-
using phone, none
This work
-
-
texting, handheld
phone call, none
pedestrian intent) and predict the paths based on low level
cues alone [11]–[15].
Intent analysis: The aim of such studies is to make an
estimate of the pedestrians’ intention in the near future, so
as to take appropriate measures to reduce risk of collision.
These studies are commonly carried out in conjunction with
path prediction, in a manner that benefits both tasks. Recent
examples in this domain are [16]–[21].
Action/Activity recognition: The terms action and activity
have been used quite loosely in the context of pedestrians.
In most cases, these terms allude to the different stages in
the trajectory of a pedestrian [22], [23], e.g. walking, waiting,
crossing etc. This notion of activity has also been extended
to groups of people, where portions of a crowd are assigned
a common activity based on context and collective behavior
[24]. In this study, we use the term activity to refer to the
secondary activity of a pedestrian being performed in addition
to walking/crossing.
Although, the tasks listed above are focused on modeling
pedestrians and their behavior, none of them consider pedestrian distraction due to secondary activities like cellphone
usage (see Table I for reference). Moreover, this study could
be complementary to existing studies on pedestrian intention
and path prediction, and could result in a more holistic
understanding of pedestrian behavior.
III. DATASET D ESCRIPTION & S EMANTIC A NNOTATIONS
Since pedestrian distraction due to cellphone usage is more
common among a young demographic, we mounted 4 GoPro
cameras, each facing a different direction, on an intelligent
3
(a) Histogram of pedestrian bounding box
heights
(b) Pedestrian activities
(c) Objects in pedestrian hands
(d) Pedestrian samples from the dataset. The joints obtained after articulated pose estimation have been overlaid for reference.
Fig. 2: Details pertaining to the proposed dataset. The dataset is demonstrably diverse in viewpoints, pedestrian size, activity
and object interactions.
vehicle testbed parked at an intersection in the UC San Diego
campus. By capturing different viewpoints on each camera,
we ensure that pedestrians are not predisposed to appear in a
particular location or facing a certain direction. Most of the
data is captured on afternoons and evenings, on both sunny
and overcast days to ensure diverse illumination conditions
and reasonable foot traffic. Since the proposed methodology
carries out fine-grained analysis of pedestrians, we avoid night
time situations where it is hard to identify small objects
and features even for humans. Furthermore, pedestrians are
captured holding a variety of objects in addition to cellphones,
such as bags, drinks, food and other miscellaneous items. To
facilitate the finer analysis of each pedestrian, videos were
captured at 2.7k resolution, resulting in pedestrians as large
as 1000 pixels in height in a few cases. Figure 2 visualizes
certain key statistics of our dataset, and shows a few sample
pedestrians chosen at random.
The dataset comprises of a total of 1586 cropped pedestrians, each with annotated activities and objects. These pedestrians are then divided into train and test sets using a 75-25
split, while making sure that the fraction of occurrences of
each activity is retained in both sets.
Additionally, we annotate 7 sequences of pedestrians (3
for training, 4 for testing), each approximately 10 seconds
in duration (≈ 300 frames). In this case, the pedestrian is
assigned an activity for each frame to account for temporal
dynamics. In addition to this, the upper body joints (listed in
section IV) are annotated for each frame, to enable evaluation
of the proposed articulated pose tracker.
IV. S INGLE F RAME ACTIVITY C LASSIFICATION
Figure 3 depicts the flow diagram of the proposed activity
classification framework. The pipeline takes in an image patch
corresponding to a pedestrian, and outputs the corresponding
activity. We detail each processing block in the subsections
that follow.
Image of
pedestrian
Pose
estimation
query
Hand
localization
Training
exemplars
Head
localization
nearest
neighbors
Hand
analysis
Pose
analysis
Head
analysis
Activity
classification
Output class
Fig. 3: Flow diagram of proposed methodology for single
frame activity classification.
A. Articulated Pose Estimation
The articulated pose of a pedestrian can be an invaluable
cue in estimating the activity he/she is involved in. Recent
advances in pose estimation using deep convolutional neural
networks (ConvNets) have led to state of the art results on
challenging benchmarks. We make use of one such architecture, called the Convolutional Pose Machines [25] proposed
by Wei et al. This is a multi-stage ConvNet, where each
subsequent stage operates both on image evidence as well
as belief maps from preceding stages, gradually refining the
pose estimate. This setup offers us great flexibility while
choosing the number of stages, with the trade-off being speed
versus accuracy. The network has been trained on the MPII
4
TABLE II: PCK scores [32] of the pose estimation module
[25] on the test set before and after fine-tuning on the train set.
Values close to 1 indicate near-perfect keypoint localization.
TABLE III: Fraction of pedestrian hands falling within predicted wrist and hand centered windows for different values
of window scale factor α. These evaluations were carried on
a separate validation set.
right
right right
left
left
left
head neck shoulder elbow wrist shoulder elbow wrist
α
Before
fine-tuning 1.00 1.00
0.99
1.00
0.96
1.00
1.00
0.94
After
fine-tuning 1.00 1.00
0.99
1.00
0.97
1.00
1.00
0.96
dataset comprising of 25K images containing over 40K people,
involved in 410 different activities, and outputs the locations
of 16 joints corresponding to the articulated pose of a human
body. We use this pre-trained network and fine-tune it on
our own dataset. This gives us marginal improvements in
performance compared to an out-of-the-box implementation
(see Table II). Additionally, we only make use of the upper
body joints for any further processing, as these are the most
informative in our application. The framework can easily
accommodate the full body pose instead, if necessary. The
final set of keypoint locations retained are - head, neck, left
shoulder, left elbow, left wrist, right shoulder, right elbow and
right wrist. See Figure 2 for some visual results of the pose
estimation module on the proposed dataset.
Most human pose estimation algorithms require the rough
location and scale of the human in the image plane. In this
study, we assume that such information is available beforehand, and focus our attention on analyzing each pedestrian
in finer detail. However, if desired, the location and scale
of pedestrians may be obtained easily from any generic
pedestrian detector. We would also like to point out that many
recent studies like [26], [27] demonstrate state-of-the-art multiperson pose estimation in real time, without prior information
on pedestrian locations and scales. This makes our approach
viable for time critical applications like pedestrian safety and
path planning.
The pose estimation module is used in our pipeline for three
specific purposes. First, it allows us to localize the head and
hands of each pedestrian for further examination. Second, it is
used to identify similar training exemplars in the pose space.
Third, the pose alone may be used as a informative prior over
all activities. In each of the following subsections, we make
use of the articulated pose in a manner mentioned above.
B. Hand Analysis using Exemplar SVMs
An important cue for predicting the activity of a pedestrian
are the objects they interact with. To identify the objects held
in the hands of a pedestrian, we look at local image patches
around the location of each hand. To do so, we first regress
to the approximate location of the hands of a pedestrian,
assuming that it is collinear with the joints corresponding to
the elbow and wrist. Let (xe , ye ) and (xw , yw ) denote the
image plane coordinates of the elbow and wrist respectively.
Using the assumption above, the approximate location of the
hand (xh , yh ) is obtained as follows:
Wrist-centered Hand-centered
window
window
0.05
0.07
0.10
0.12
0.15
0.20
xh = xe +
0.0825
0.3400
0.9125
0.9850
0.9975
0.9975
0.7250
0.9275
0.9900
0.9950
0.9975
0.9975
xw − xe
yw − ye
, yh = ye +
,
r
r
(1)
where r is a a parameter that depends on the ratio of distances
of the elbow from the wrist and hand respectively. In our
experiments, r = 5/6 seemed to generate the best results.
Once we have the rough locations of both hands in the
image plane, we crop out a local image window around these
locations. The window size is chosen to be α·h for a pedestrian
parametrized by (x, y, w, h). Here α is a hyper-parameter
that ensures that the local window scales with the size of
the pedestrian. In our experiments, α is set to 0.1 to extract
training patches. α is chosen to ensure that the hand almost
always falls into the window, and also that the window is
small enough to capture only the object of interest and nothing
more. As demonstrated in Table III, α = 0.1 offers the best
results, beyond which increasing α does not improve hand
localization by much. Examples of such local patches for
windows centered around both the wrist and the hand can
be found in Figure 4. It is obvious that inferring the hand
location, even if approximate, helps in centering the object of
interest with respect to the window.
With a collection of such training patches centered around
the hand, we proceed to build an object classifier. Our experiments demonstrated that traditional one-versus-all classifiers
severely overfit the training data and failed to generalize well
to new object instances. Moreover, training a separate classifier
for each object class, as well as the intra-class variance
(cellphones come in a variety of shapes and sizes) makes
the classification task an especially hard one, considering the
limited availability of training data.
We bypass all these limitations by training an ensemble
of exemplar-SVMs (ESVMs) [28]. The method is based on
training a separate linear SVM classifier for every exemplar
in the training set. Each ESVM is thus defined by a single
positive instance and millions of negatives, obtained by hard
negative mining. In our case, an ESVM is trained to represent
a rigid HOG template from an image patch around each hand
of every pedestrian in the training set. At test time, the ESVM
that results in the highest score is considered to provide the
best match, and the object label associated with the exemplar
is transferred to the new test instance. Figure 5 shows a few
examples of matched hand-object instances.
5
∠x̄(i,j,k) =
(a)
(b)
Fig. 4: Image patches obtained when the local window is
centered around the (a)wrist versus the (b)hand.
∠x(i,j,k)
.
π
(3)
The final feature vector x ∈ R23 is obtained as a simple
concatenation of the set of normalized joint locations and
angles. Our experiments indicated this to be much more stable
in terms of closest neighbors in comparison to using either just
the joint locations, or just the joint angles. With a set of pose
features gathered from the pedestrians in the training set, we
train a simple K-nearest neighbor classifier using a K-d tree
structure for fast neighbor retrieval.
E. Pedestrian Activity Classification
C. Gaze Analysis
The rough gaze direction of a pedestrian can be very
effective in separating out instances where pedestrians are just
holding a phone, versus when they are actually engaged in its
use.
In this study, we use the gaze pathway from the GazeFollow
Convnet proposed in [29]. The gaze pathway takes in an
image patch of the head along with its normalized location
in the image plane (obtained from the articulated pose), and
returns a 13 × 13 heat-map (Figure 5) that encodes the rough
gaze direction of the pedestrian. This sub-network has five
convolutional layers followed by three fully-connected layers,
the final output of which is a single channel heat-map. Finally,
we reshape this output to produce a 169-length feature vector
that encodes the gaze.
D. Querying Nearest Neighbor Pose Exemplars
The main intuition behind our approach is that pedestrians
with similar body poses tend to interact with objects in a
similar form, and are likely to be involved in analogous
activities. To have such a notion of similarity, it is necessary
to construct a suitable feature representation of the articulated
pose, and to enforce a reasonable distance metric that ensures
that similar poses are close by.
We make use of a combination of the normalized joint
locations and the normalized joint angles as the feature descriptor. Consider a pedestrian bounding box parametrized as
(x, y, w, h). Here, x and y correspond to image coordinates
of the top left corner of the bounding box, and w and h
describe the dimensions of the box. For the pedestrian under
consideration, the pose estimation network outputs a set of
image locations {xi = (xi , yi )}i=1,··· ,8 corresponding to each
joint in the upper body. The set of normalized joint locations
{x̄i = (x̄i , ȳi )}i=1,··· ,8 are then found as follows:
yi − y
xi − x
, ȳi =
, ∀i = 1, · · · , 8.
(2)
w
h
Next, consider the set of joint triplets that are connected
consecutively in the articulated pose tree. For each such triplet
(xi , xj , xk ), let the angle subtended (in radians) at xj by the
line segment joining points xi and xk be denoted by ∠x(i,j,k) .
We have 7 such joint angles in the upper body pose. The
normalized joint angle at xj is then obtained as follows:
x̄i =
Having set up the individual parts, we now focus on
integrating the cues from the different modalities to predict
a final class probability score. For this study, the possible
output classes for activity classification are none, texting and
handheld phone call, which we encode as y = 0, 1 and 2
respectively.
Consider a new pedestrian with pose features xo calculated
in the manner described above. The aim now is to predict
a class label yo , and estimate the probability associated with
this prediction. Let NK denote the set of K nearest neighbor
pose exemplars obtained from the trained classifier in IV-D.
We denote this set as follows NK = {(eli , oli , eri , ori , hi , yi )}i=1,··· ,K ,
(4)
where eli and eri denote the trained ESVMs on the left and
right hands (from IV-B), oli and ori denote the object labels
associated with the left and right hands, hi represents the
gaze features obtained as mentioned in IV-C, and yi denotes
the activity label associated with the ith nearest neighbor
exemplar.
Let E denote the image evidence available for the pedestrian
whose activity is to be predicted. The desired predictive
distribution may then be expressed as P(yo |E; NK ) ∝ P(E|yo ; NK ) P(yo ; NK ).
(5)
Decomposing the image evidence into individual head and
hand based evidences Ehead and Ehand , and making use of
conditional independence yields P(yo |E; NK ) ∝ P(Egaze |yo ; NK ) P(Ehand |yo ; NK ) P(yo ; NK ).
(6)
Each term in the equation above is described below
P(Egaze |yo ; NK ) := max
1≤i≤K
ho · hi
||ho ||2 · ||hi ||2
· 1{yo } (yi ),
(7)
where ho and hi are the gaze descriptors, and 1A (·) is
the indicator function for set A. This is simply the cosine
similarity between the gaze features within the same class.
Next, let us denote the maximum match score obtained for
ESVM eli on the left hand image patch as pli , and that for
eri on the right hand image patch as pri . This probabilistic
6
Gaze
ConvNet
Gaze
ConvNet
Gaze
ConvNet
ESVM w
ESVM w
Match
Match
ESVM w
Match
Fig. 5: Illustration of head and hand related cues described in sections IV-B and IV-C. In each of the three examples above,
for a pedestrian in the test set (left), the gaze heatmap obtained from the gaze ConvNet is shown on top and the best hand
(object) exemplar match with a pedestrian from the train set is shown below. The matched pedestrian (right) and exemplar
weights are shown in addition to the matched hand patch. Best viewed in color.
score is obtained by testing each ESVM on the corresponding
image patch, and then re-scaling the match score using the
parameters determined by carrying out the Platt calibration
for each ESVM offline. Further, only matches with an overlap
score greater than 0.4 with the test patch are retained as done
in [30]. We can now define the hand evidence likelihood as
follows -
P(Ehand |yo ; NK ) :=
min
max pli · 1{0} (yi ),
i
r
max
p
·
1
(y
)
,
{0} i
i
i
if yo = 0
l
max
pi · 1{1} (oli )+
i
pri · 1{1} (ori ) · 1{yo } (yi ) , if yo = 1, 2,
(8)
where 1
equals 1 if the object associated with the left
hand is a cellphone, else it equals 0. The same is true for
1{1} (ori ) and the right hand.
Finally, the term P(yo ; NK ) acts as a prior over the activities, given just the articulated pose of a pedestrian. This is
defined to be l
{1} (oi )
K
P
P(yo ; NK ) :=
i=1
1{yo } (yi )
K
V. ACTIVITY C LASSIFICATION FOR S EQUENTIAL DATA
In our proposed framework, the major bottleneck in terms of
speed is the pose estimation network described in IV-A. Even
though it is possible to reliably estimate the pose for a few
pedestrians in real time using a GPU (for a reasonable number
of stages in the network), the network can no longer operate
at a desired frequency when the number of pedestrians in the
scene are considerably large. This issue can be alleviated by
tracking the articulated pose of pedestrians for the duration
between successive outputs from the pose estimation network.
This also ensures that the pose estimated by the network makes
reasonable transitions between successive instances, thereby
reducing single frame errors.
A. GPDM-based Particle Filter for Articulated Pose Tracking
In this sub-section, we briefly describe the proposed particle
filter based tracking framework with a Gaussian Process
Dynamical Model (GPDM) [31]. Let xit ∈ R23 be the state
of particle i at time t, which represents the normalized pose
features of a pedestrian as described in IV-D. Let lit ∈ R2
denote the latent space projection of xit using a Gaussian
Process Latent Variable Model (GPLVM) as described below
xt = f (lt ; A) + nx,t .
.
(9)
Using the equations 6-9, the final predicted activity for the
pedestrian is then chosen to be the MAP estimate -
In addition to this, a GPDM enforces an auto-regressive
dynamical model in the latent space lt = g(lt−1 ; B) + nl,t .
yo∗
= arg max P(y|E; NK ).
(10)
y∈{0,1,2}
Since the probability terms on the right hand side of
equation 6 are not calibrated to provide compatible scores,
we propose a second method based on late fusion of these
scores. To do so, we create 9-length score vectors made up of
the terms P(Egaze |yo ; NK ), P(Ehand |yo ; NK ) and P(yo ; NK )
for yo ∈ {0, 1, 2}. These vectors are created by performing a
5 fold cross validation split on the training set. Multi-class
classification is carried out in a one versus all manner to
predict the final activity of the pedestrian.
(11)
(12)
Here, nx,t and ny,t are zero-mean, white Gaussian processes,
f and g are nonlinear mappings parametrized by A and B
respectively. Using small training sequences (x0 , x1 , · · · , xT ),
we can solve for both the corresponding latent space projections (l0 , l1 , · · · , lT ), and the necessary hyperparameters
in closed form [31]. Despite the use of small data sets, the
GPDM learns an effective representation of the highly nonlinear dynamics associated with articulated pose tracking.
At any instant t, the particle filter functions by propagating
a set of particles in the latent space {lit }i , by sampling
(with noise) from the dynamical model in equation 12. This
7
TABLE IV: Per class and overall accuracies for four sets of
cues - hands alone, pose alone, pose and hands, and pose,
hands and gaze.
Accuracy Accuracy
Accuracy
Overall
for None for Texting for Phone Call Accuracy
Cues
(a)
hand only
pose only
pose+hands
pose+hands+gaze
(b)
(c)
(d)
Fig. 6: Latent space projections of articulated pose trajectories
for four different viewpoints.
results in an updated set of particles {lit+1 }i . To determine
the likelihood of each particle, it is necessary to project the
latent particles back into the observation space using the
learned GPLVM mapping (equation 11), where they may
be evaluated against available measurements. This results in
a corresponding set of particles {xit+1 }i in the observation
space. To evaluate the likelihood of each particle, we note that
most pose estimation networks output a heatmap for each joint
location, which can be interpreted as a probabilistic score for
its location in the image. With this in mind, let hj (·, ·) denote
the heatmap for the j th joint; the function takes in the x and
y coordinates of any location on the image plane, and returns
the score associated with the location, encoded in the heatmap.
The likelihood of a particle lit+1 is then considered to be L(lit+1 ) = L(xit+1 ) :=
8
Y
hj (xij , yji ),
(13)
j=1
where (xij , yji ) denotes the co-ordinates of the j th joint obtained from the pose features xit+i .
In practice, we train a set of GPDMs for different activities
and viewpoints (e.g. walking towards the camera, walking
away from the camera, walking sideways etc.). During test
time, particles are initialized in latent space by locating the
latent point (across all GPDMs) whose mapping in observation
space yields the best match with the current measurement.
Figure 6 shows the learnt latent space mappings for 4 different
viewpoints.
B. Avenues for Speedup
The tracking framework considerably reduces the burden
on the pose estimation network in ensuring near real time
operation. When a large number of pedestrians are present in
the scene, we can simply run the networks on a subset of
pedestrians, while the poses of the rest are updated based on
state updates alone. When the network is finally run on a given
pedestrian, the heatmaps are used as measurements to update
the state of each particle in the filter. Alternatively, one can
0.94
0.90
0.93
0.97
0.58
0.65
0.71
0.88
0.20
0.67
0.81
0.89
0.810
0.858
0.916
0.946
make use of more recent real-time algorithms for multi-person
pose estimation [26], [27], which enables faster operation due
to reduced overhead from having separate detection and pose
estimation modules.
Additionally, by only running the ESVMs associated to the
K nearest neighbor exemplars, we bypass the computational
drawbacks associated with ESVMs. We can further reduce the
computational burden required to predict class labels for a
pedestrian at every instant in case of sequential data. Once the
hand evidence term P(Ehand |yo ; NK ) in equation 8 is reliably
estimated for all output classes, we need only update the head
evidence and prior terms in equation 6 at every instant. This
removes the need to run the ensemble of ESVMs at every
frame.
VI. E XPERIMENTAL A NALYSIS
A. Single Frame Activity Classification
The critical hyperparamater that needs to be set for the
proposed pipeline is the neighborhood size K. We experiment
with a set of different values - K ∈ {25, 50, 100, 200}.
Figure 7 shows confusion matrices for different values of K,
for both the MAP estimation scheme and the SVM based
late fusion. The MAP estimation scheme is seen to perform
relatively poorly. This can be attributed to the fact that the pose
prior P(yo ; NK ) is far too dominant in comparison to the other
two probability terms. This leads to predictions that are overly
influenced by the pose term, and hence the considerable false
positives for all values of K. In comparison, the SVM weights
each individual cue accordingly and predicts a more balanced
output. This leads to a much better overall accuracy in comparison to the MAP estimate. The best performance is seen
for K = 100, which results in 94.6% overall accuracy. We
notice that most mis-classifications in the output correspond
to pedestrians who are considerably small in size (≤ 200
pixels in height), which makes it relatively harder to infer
object labels and gaze information. Some other error modes are
observed in cases where the pedestrian is holding objects that
are considerably different to those observed in the training set.
This issue however may be alleviated by collecting more data
for training. Finally, there are cases where the correct label is
ambiguous even to human annotators. This occurs when it is
hard to infer the exact direction of gaze or the correct object
label. For examplar results on the test dataset, we refer the
reader to Figure 8.
To understand the contribution and utility of each individual
cue while making a prediction, we consider K = 100
corresponding to the best performing method. We train four
8
(a)
K = 25
(b)
K = 50
(c)
K = 100
(d)
K = 200
(e)
K = 25
(f)
K = 50
(g)
K = 100
(h)
K = 200
Fig. 7: Confusion matrices for MAP estimation (top row), and SVM based late fusion (bottom row). K = 100 with SVM
based late fusion results in the best overall accuracy.
separate SVM based fusion models, for four different sets
of cues - hands only, pose only, pose and hands, and pose,
hands and gaze. For each set of such cues, only features
based on those cues are used for training the fusion SVM.
Table IV shows the per class and overall accuracies for each of
these configurations. Pose alone is seen to perform reasonably
well, indicating that it is the strongest of the three cues.
However, it tends to be too harsh in its assignment, depending
too heavily on the nearest neighbors. Adding hand related
cues from the ensemble of ESVMs considerably improves the
classification accuracy, especially for the texting and phone
call classes as these are more reliant on recognition of handobject interactions. Finally, adding gaze information further
enhances the performance, noticeably for the texting class
since this requires the pedestrian to look at the phone directly.
As far as the processing time is concerned, our algorithm
with K = 100 nearest neighbors takes about 4ms on average
for each pedestrian on a 6th generation i7 CPU. This does not
include the time for running the articulated pose estimation
module, which we run independently on a Titan X GPU.
As mentioned in V-B, state-of-the-art pose estimation for
multiple persons is almost nearing real time operation, and
the processing times for other operations in our framework
are minimal in comparison. This indicates feasibility for real
world applications.
B. Pose Tracking
As tracking the pose enables us to achieve speedups during
runtime, it is important to validate its reliability on pedestrian
sequences. We do this by training the proposed GPDM based
particle filter on three pedestrian sequences. The tracker is then
evaluated on 4 separate sequences from the test set using the
PCK metric [32]. The results are shown in Table V. While
tracking pedestrians in videos captured at 30Hz, it is seen
that providing pose measurements even once every 6 frames
(5Hz) is more than enough to result in very suitable tracks.
TABLE V: PCK scores of proposed articulated pose tracking
for different measurement rates, evaluated on 4 different
pedestrian sequences. The tracking stays consistent and reliable even for relatively infrequent measurements obtained
from the pose ConvNet.
Measureright
right right
left
left
left
ment rate head neck shoulder elbow wrist shoulder elbow wrist
untracked
30 Hz
15 Hz
10 Hz
5 Hz
1.000
1.000
0.980
0.980
0.980
1.000
1.000
0.990
0.990
0.990
0.990
1.000
0.981
0.981
0.981
1.000
1.000
0.990
0.908
0.970
0.960
0.980
0.955
0.940
0.940
1.000
1.000
0.975
0.975
0.975
1.000
1.000
0.930
0.925
0.920
0.940
0.940
0.910
0.905
0.895
Furthermore, tracking with a measurement at every frame
provides an improvement over the tracks obtained by running
the pose ConvNet alone. This illustrates that tracking gives
us robust estimates of the pose in addition to making our
algorithm run faster.
C. Activity Classification for Sequential Data
Next, we evaluate our proposed framework on 4 test sequences, each captured at 30 Hz and approximately 10 seconds
in duration (280 - 310 frames per sequence). Each frame in a
test sequence is annotated with the correct activity class. Our
activity classification framework (with K = 100) is run with
the proposed articulated pose tracker, and predicts an output
class for each frame. Additionally, we run the ensemble of
ESVMs once every 50 frames as described in section V-B to
reduce the computational burden. We plot the predicted and
ground truth classes as a function of the frame number for
each of the 4 test sequences in Figure 9. It is clearly observable
that the activity classification framework, along with the pose
tracker result in class labels that are quite consistent with
the ground truth, even under frequent changes in the activity
dynamics.
9
y* = 1
y* = 0
y* = 0
y* = 1
y* = 2
y* = 1
y* = 1
y* = 0
y* = 0
y* = 0
y* = 1
y* = 2
y* = 1
y* = 1
y* = 1
yGT = 0
y* = 0
y* = 0
y* = 2
yGT = 0
y* = 1
yGT = 0
Fig. 8: Examples of pedestrians from the test set along with their predicted activity classes (y ∗ ). Correctly predicted classes
are enclosed in green boxes, and incorrectly predicted classes are enclosed in red boxes along with the corresponding ground
truth class y GT .
Ground Truth
Output Class
Phone Call
Ground Truth
Output Class
Phone Call
Ground Truth
Output Class
Phone Call
Ground Truth
Output Class
Phone Call
Texting
Texting
Texting
Texting
None
None
50
100
150
200
250
300
None
50
100
150
200
250
300
None
50
100
150
200
50
100
150
200
Frame Number
Frame Number
Frame Number
Frame Number
(a) Sequence 1
(b) Sequence 2
(c) Sequence 3
(d) Sequence 4
250
300
Fig. 9: Plot of ground truth and predicted output class as a function of frame number for 4 test sequences.
VII. C ONCLUDING R EMARKS
In this paper, we studied pedestrian distraction caused by
cellphone usage in an effort to reduce growing number of
pedestrian fatalities. To this end, a multi-cue pipeline to
recognize pedestrian activity is proposed. A pedestrian is
classified to be either texting, in a phone call, or be involved
in no secondary activities based on cues from the articulated
pose, hands and gaze. ESVMs trained offline are used to
encode hand-object labels, whereas gaze features are obtained
from a pre-trained ConvNet. Each cue is then used to propose
scores based on the K neighboring pedestrians from the
training set. Finally, these scores are combined effectively
using an SVM based late fusion scheme. In addition to this,
we propose a GPDM based particle filter that operates based
on measurements obtained from a pose estimation ConvNet
in order to improve pose estimation, and speedup operation.
Both the proposed methodology and the tracking framework
are trained and evaluated on a unique pedestrian distraction
dataset, which provides rich semantic annotations to facilitate
a more detailed study of pedestrians.
Although the results are promising, there are still many
issues to be addressed. Pedestrian activities are rich in variety,
and so are the objects they interact with. However, it must
be noted that out proposal is highly scalable. Since it works
on similarity based metrics obtained from pedestrians in the
training set, as more diverse pedestrians are added to the
training process, the performance would only improve, and
without any evident drop in computational speed. Future
work encompasses going beyond phone based distraction, and
studying other sources of pedestrian distraction (e.g. talking,
walking in a group, listening to music etc.), and integrating all
such factors to predict a combined distraction score for each
pedestrian.
VIII. ACKNOWLEDGMENTS
We would like to thank all our colleagues at the LISA
lab, UCSD for their assistance in collecting and annotating
the dataset. We would also like to express our gratitude to
the reviewers and the editor for their valuable comments and
suggestions.
10
R EFERENCES
[1] J. L. Nasar and D. Troyer, “Pedestrian injuries due to mobile phone use
in public places,” Accident Analysis & Prevention, vol. 57, pp. 91–95,
2013.
[2] L. L. Thompson, F. P. Rivara, R. C. Ayyagari, and B. E. Ebel, “Impact
of social and technological distraction on pedestrian crossing behaviour:
an observational study,” Injury prevention, vol. 19, no. 4, pp. 232–237,
2013.
[3] K. W. Byington and D. C. Schwebel, “Effects of mobile internet use on
college student pedestrian injury risk,” Accident Analysis & Prevention,
vol. 51, pp. 78–83, 2013.
[4] T. Gandhi and M. M. Trivedi, “Pedestrian protection systems: Issues,
survey, and challenges,” Transactions on Intelligent Transportation Systems, vol. 8, no. 3, pp. 413–430, 2007.
[5] M. Roth, F. Flohr, and D. M. Gavrila, “Driver and pedestrian awarenessbased collision risk analysis,” in Intelligent Vehicles Symposium (IV),
2016 IEEE. IEEE, 2016, pp. 454–459.
[6] T. Hoang Ngan Le, Y. Zheng, C. Zhu, K. Luu, and M. Savvides, “Multiple scale faster-rcnn approach to driver’s cell-phone usage and hands
on steering wheel detection,” in Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition Workshops, 2016, pp. 46–53.
[7] E. Ohn-Bar, S. Martin, A. Tawari, and M. M. Trivedi, “Head, eye, and
hand patterns for driver activity recognition,” in Pattern Recognition
(ICPR), 2014 22nd International Conference on. IEEE, 2014, pp. 660–
665.
[8] A. Tawari, S. Sivaraman, M. M. Trivedi, T. Shannon, and M. Tippelhofer,
“Looking-in and looking-out vision for urban intelligent assistance:
Estimation of driver attentive state and dynamic surround for safe
merging and braking,” in Intelligent Vehicles Symposium Proceedings,
2014 IEEE. IEEE, 2014, pp. 115–120.
[9] A. Rangesh, E. Ohn-Bar, K. Yuen, and M. M. Trivedi, “Pedestrians
and their phones-detecting phone-based activities of pedestrians for
autonomous vehicles,” in Intelligent Transportation Systems (ITSC),
2016 IEEE 19th International Conference on. IEEE, 2016, pp. 1882–
1887.
[10] E. Ohn-Bar and M. M. Trivedi, “Looking at humans in the age of
self-driving and highly automated vehicles,” Transactions on Intelligent
Vehicles, vol. 1, no. 1, pp. 90–104, 2016.
[11] A. Mogelmose, M. M. Trivedi, and T. Moeslund, “Trajectory analysis
and prediction for improved pedestrian safety: Integrated framework and
evaluations,” in Intelligent Vehicles Symposium (IV). IEEE, 2015, pp.
330–335.
[12] T. Gandhi and M. M. Trivedi, “Image based estimation of pedestrian
orientation for improving path prediction,” in Intelligent Vehicles Symposium. IEEE, 2008, pp. 506–511.
[13] M. Goldhammer, M. Gerhard, S. Zernetsch, K. Doll, and U. Brunsmann,
“Early prediction of a pedestrian’s trajectory at intersections,” in 16th
International Conference on Intelligent Transportation Systems (ITSC).
IEEE, 2013, pp. 237–242.
[14] A. Prioletti, A. Møgelmose, P. Grisleri, M. M. Trivedi, A. Broggi,
and T. B. Moeslund, “Part-based pedestrian detection and featurebased tracking for driver assistance: real-time, robust algorithms, and
evaluation,” IEEE Transactions on Intelligent Transportation Systems,
vol. 14, no. 3, pp. 1346–1359, 2013.
[15] A. Bera, S. Kim, T. Randhavane, S. Pratapa, and D. Manocha, “Glmprealtime pedestrian path prediction using global and local movement
patterns,” in Robotics and Automation (ICRA), 2016 IEEE International
Conference on. IEEE, 2016, pp. 5528–5535.
[16] S. Köhler, M. Goldhammer, S. Bauer, K. Doll, U. Brunsmann, and
K. Dietmayer, “Early detection of the pedestrian’s intention to cross the
street,” in 15th International Conference on Intelligent Transportation
Systems (ITSC). IEEE, 2012, pp. 1759–1764.
[17] F. Madrigal, J.-B. Hayet, and F. Lerasle, “Intention-aware multiple
pedestrian tracking,” in 22nd International Conference on Pattern
Recognition (ICPR). IEEE, 2014, pp. 4122–4127.
[18] A. T. Schulz and R. Stiefelhagen, “Pedestrian intention recognition
using latent-dynamic conditional random fields,” in Intelligent Vehicles
Symposium (IV). IEEE, 2015, pp. 622–627.
[19] T. Bandyopadhyay, C. Z. Jie, D. Hsu, M. H. Ang Jr, D. Rus, and
E. Frazzoli, “Intention-aware pedestrian avoidance,” in Experimental
Robotics. Springer, 2013, pp. 963–977.
[20] C. G. Keller and D. M. Gavrila, “Will the pedestrian cross? a study on
pedestrian path prediction,” Transactions on Intelligent Transportation
Systems, vol. 15, no. 2, pp. 494–506, 2014.
[21] J. F. P. Kooij, N. Schneider, F. Flohr, and D. M. Gavrila, “Contextbased pedestrian path prediction,” in European Conference on Computer
Vision. Springer, 2014, pp. 618–633.
[22] H. Kataoka, Y. Aoki, Y. Satoh, S. Oikawa, and Y. Matsui, “Finegrained walking activity recognition via driving recorder dataset,” in 18th
International Conference on Intelligent Transportation Systems (ITSC).
IEEE, 2015, pp. 620–625.
[23] R. Quintero, I. Parra, D. Llorca, and M. Sotelo, “Pedestrian path
prediction based on body language and action classification,” in 17th
International Conference on Intelligent Transportation Systems (ITSC).
IEEE, 2014, pp. 679–684.
[24] W. Choi, K. Shahid, and S. Savarese, “What are they doing?: Collective
activity classification using spatio-temporal relationship among people,”
in 12th International Conference on Computer Vision Workshops (ICCV
Workshops). IEEE, 2009, pp. 1282–1289.
[25] S.-E. Wei, V. Ramakrishna, T. Kanade, and Y. Sheikh, “Convolutional
pose machines,” in Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, 2016, pp. 4724–4732.
[26] Z. Cao, T. Simon, S.-E. Wei, and Y. Sheikh, “Realtime multiperson 2d pose estimation using part affinity fields,” arXiv preprint
arXiv:1611.08050, 2016.
[27] G. Papandreou, T. Zhu, N. Kanazawa, A. Toshev, J. Tompson, C. Bregler,
and K. Murphy, “Towards accurate multi-person pose estimation in the
wild,” arXiv preprint arXiv:1701.01779, 2017.
[28] T. Malisiewicz, A. Gupta, and A. A. Efros, “Ensemble of exemplarsvms for object detection and beyond,” in Computer Vision (ICCV),
2011 IEEE International Conference on. IEEE, 2011, pp. 89–96.
[29] A. Recasens, A. Khosla, C. Vondrick, and A. Torralba, “Where are they
looking?” in Advances in Neural Information Processing Systems, 2015,
pp. 199–207.
[30] A. Shrivastava, T. Malisiewicz, A. Gupta, and A. A. Efros, “Datadriven visual similarity for cross-domain image matching,” in ACM
Transactions on Graphics (TOG), vol. 30, no. 6. ACM, 2011, p. 154.
[31] J. M. Wang, D. J. Fleet, and A. Hertzmann, “Gaussian process dynamical
models,” in NIPS, vol. 18, 2005, p. 3.
[32] Y. Yang and D. Ramanan, “Articulated human detection with flexible
mixtures of parts,” IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. 35, no. 12, pp. 2878–2890, 2013.
Akshay Rangesh is currently working towards his
PhD in electrical engineering from the University
of California at San Diego (UCSD), with a focus
on intelligent systems, robotics, and control. His
research interests span computer vision and machine
learning, with a focus on object detection and tracking, human activity recognition, and driver safety
systems in general. He is also particularly interested
in sensor fusion and multi-modal approaches for real
time algorithms.
Mohan Manubhai Trivedi is a Distinguished
Professor at University of California, San Diego
(UCSD) and the founding director of the UCSD
LISA: Laboratory for Intelligent and Safe Automobiles, winner of the IEEE ITSS Lead Institution
Award (2015). Currently, Trivedi and his team are
pursuing research in intelligent vehicles, machine
perception, machine learning, human-robot interactivity, driver assistance, active safety systems. Three
of his students have received ”best dissertation”
recognitions. Trivedi is a Fellow of IEEE, ICPR and
SPIE. He received the IEEE ITS Society’s highest accolade ”Outstanding
Research Award” in 2013. Trivedi serves frequently as a consultant to industry
and government agencies in the USA and abroad.
| 1cs.CV
|
All-Transfer Learning for Deep Neural Networks and
its Application to Sepsis Classification
arXiv:1711.04450v1 [cs.CV] 13 Nov 2017
Yoshihide Sawada1 and Yoshikuni Sato2 and Toru Nakada2 and Kei Ujimoto2 and Nobuhiro Hayashi3
Abstract. In this article, we propose a transfer learning method for
deep neural networks (DNNs). Deep learning has been widely used
in many applications. However, applying deep learning is problematic when a large amount of training data are not available. One of
the conventional methods for solving this problem is transfer learning
for DNNs. In the field of image recognition, state-of-the-art transfer
learning methods for DNNs re-use parameters trained on source domain data except for the output layer. However, this method may
result in poor classification performance when the amount of target domain data is significantly small. To address this problem, we
propose a method called All-Transfer Deep Learning, which enables
the transfer of all parameters of a DNN. With this method, we can
compute the relationship between the source and target labels by the
source domain knowledge. We applied our method to actual twodimensional electrophoresis image (2-DE image) classification for
determining if an individual suffers from sepsis; the first attempt
to apply a classification approach to 2-DE images for proteomics,
which has attracted considerable attention as an extension beyond
genomics. The results suggest that our proposed method outperforms
conventional transfer learning methods for DNNs.
1 Introduction
Deep learning has been widely used in the fields of machine learning and pattern recognition [4, 11, 16, 22, 24, 42, 43] due to its advanced classification performance. Deep learning is used to train a
large number of parameters of a deep neural network (DNN) using
a large amount of training data. For example, Le et al. [24] trained 1
billion parameters using 10 million videos, and Krizhevsky et al. [22]
trained 60 million parameters using 1.2 million images. They collected training data via the web. On the other hand, original data,
such as biomedical data, cannot be easily collected due to privacy
and security concerns. Therefore, researchers interested in solving
the original task are unable to collect a sufficient amount of data to
train DNNs. Conventional methods address this problem by applying
transfer learning.
Transfer learning is a method that re-uses knowledge of the source
domain to solve a new task of the target domain [18, 30, 31, 34, 43]. It
has been studied in various fields of AI, such as text classification [8],
natural language processing [20], and image recognition [35]. Transfer learning for DNN can be divided into three approaches, supervised, semi-supervised, and unsupervised. Recent researches focus
on unsupervised domain adaptation [14, 26]. Unsupervised and semisupervised approach assume that the target domain labels equal to
1
2
Advanced Research Division, Panasonic Corporation
Department of Life Science, Tokyo Institute of Technology
the source domain label. However, in the biomedical field, it is difficult to collect target domain data having the same label as the source
domain. Therefore, we focus on the supervised transfer learning approach, which allows the labels of the source/target domain to be
different.
The state-of-the-art supervised transfer learning [1, 10, 29] construct the first (base) model based on the source domain data by using
the first cost function. They then construct the second model based
on the target domain data by re-using the hidden layers of the first
model as the initial values and using the second cost function. This
approach outperforms non-transfer learning when the source and target domains are similar. However, these methods faced with the problem that causes poor classification performance and overfitting when
the output layer has to be trained on a significantly small amount of
target domain data. Oquab et al. [29] and Agrawal et al. [1] used the
Pascal Visual Object Classes [13] and Donahue et al. [10] used ImageNet [7]. The amount of target domain data of their studies was
over 1,000 data points. On the other hand, the amount of original
biomedical target domain data may be less than 100 data points. To
prevent this problem, it is necessary to re-use all layers including
the output layer. However, the method for effectively transferring the
knowledge (model) including the output layer has yet to be proposed.
In addition to the above problem, these methods are not structured
to avoid negative transfer. Negative transfer is a phenomenon that degrades classification accuracy when we transfer the knowledge of the
source domain/task. It is caused by using parameters computed using
the data of the source domain/task irrelevant to the target task. Although Pan et al. [31] considered the avoidance of this phenomenon
as a “when to transfer” problem, little research has been published
despite this important issue. For example, Rosenstein et al. [34] proposed a hierarchical naı̈ve Bayes to prevent this problem. However,
they did not use DNNs, and few articles have been devoted to research pertaining to DNNs.
In this article, we propose a novel method based on the transfer
learning approach, which uses two cost functions described above.
By using this approach, we can prepare the first model in advance.
It is difficult to upload the target domain data outside a hospital and
prepare a sufficient computer environment, especially for small and
medium sized hospitals. Therefore, we argue that this approach fits
the clinical demand.
The main difference is that our proposed method re-uses all parameters of a DNN trained on the source domain data and seamlessly
links two cost functions by evaluating the relationship between the
source and target labels on the basis of the source domain knowledge (Section 3). By using this relationship, our method regularizes
all layers including the output layer. This means that it can reduce
the risk of falling into the local-optimal solution caused by the ran-
domness of the initial values. We call our method All-Transfer Deep
Learning (ATDL).
We applied ATDL to actual two-dimensional electrophoresis (2DE) image [32, 33] classification for determining if an individual
suffers from sepsis. Sepsis is a type of disease caused by a dysregulated host response to infection leading to septic shock, which affects
many people around the world with a mortality rate of approximately
25% [6, 9, 37]. Therefore, high recognition performance of this disease is important at clinical sites. We use 2-DE images of proteomics
to determine sepsis, which is currently attracting considerable attention in the biological field as the next step beyond genomics. In addition, we also show that there is a correlation between the relationship described above and classification performance. This means that
ATDL is possible to reduce the risk of negative transfer. We explain
2-DE images in Section 2 and explain experimental results in Section 4.
The contributions of this article are as follows:
• We propose ATDL for a significantly small amount of training
data to evaluate the relationship between the source and target labels on the basis of the source domain knowledge.
• The experimental results from actual sepsis-data classification
and open-image-data classifications show that ATDL outperforms
state-of-the-art transfer learning methods for DNNs, especially
when the amount of target domain data is significantly small.
• We argue that there is a correlation between the relationship described above and classification performance.
• This is the first attempt to apply machine learning by using DNNs
to 2-DE images. An actual sepsis-data classification accuracy of
over 90% was achieved.
2 Two-dimensional Electrophoresis Images
Two-dimensional electrophoresis images represent the difference between the isoelectric points and molecular weights of proteins [32,
33]. Figure 1 shows an overview of the process by which 2-DE images are produced, and Figure 2 illustrates examples of 2-DE images
showing sepsis and non-sepsis. Such images are produced by first
extracting and refining proteins from a sample. After that, the proteins are split off on the basis of the degree of isoelectric points and
molecular weights. Therefore, the X-axis of 2-DE images represents
the degree of molecular weights, Y-axis represents the degree of isoelectric points, and black regions represent the protein spots [28].
Normally, 2-DE images are analyzed for detection of a specific
spot corresponding to a protein as a bio-marker, using computer assistance [5]. However, many diseases, such as sepsis, are multifactorial, which cause minute changes at many spots and unexpected
spots in some cases. Therefore, when the polymerase chain reaction (PCR) method [3], which amplifies specific genes, is applied,
we must guess the target genes, and testing of each gene must be carried out. If the number of biomarkers increases, the labor will also
increase. On the other hand, if we directly use 2-DE images, this
problem can be solved because we can consider the comprehensive
changes of proteins at one time. From this situation, we try to use
2-DE images for diagnostic testing instead of using spot analysis.
Figure 3 shows an overview of our system for detecting diseases
by using 2-DE images. First, a doctor puts a sample of the blood of
a patient on a micro-tip, then insert it into a device that can generate 2-DE images. Then, our system detects diseases and display the
results to doctors. The main point with our system is to detect diseases, such as sepsis, with complex electrophoresis patterns of 2-DE
STUVWXY
!"
#$%&'()*+ ,-./0123 45 678 9:;<= >?
@AB CDEFGHIJK LMNOPQR
Figure 1. Overview of production process of 2-DE images. After extracting
and refining proteins from sample, proteins are split off by degree of isoelectric points and molecular weights (SDS-PAGE [32, 33]).
Z[\ ]^_`ab
cde fghijklmno
Figure 2. Examples of 2-DE images. X- and Y-axes represent degrees of
molecular weights and isoelectric points, respectively, and black regions represent protein spots.
images by using DNNs. It is a matter of course that current devices
for generating 2-DE images are not suitable for this concept due to
issues such as low-throughput ability and low reproducibility. A few
groups [2, 17] have developed techniques to generate 2-DE images
with high sensitivity, high throughput ability, and high reproducibility. However, even if they can solve these problems in generating
2-DE images, collecting 2-DE images produced from patients is difficult due to privacy and security concerns. This clearly indicates that
the need for a classification method, such as ATDL, for a significantly
small amount of training data is increasing.
3.2 Training Process
pqrst
uvwxyz
3.2.1 Construction of Deep Neural Network for Source Task
{|}
~
Figure 3. Overview of our system for detecting diseases. We focused on
classification step involving analysis using DNNs.
NOPQRS TUVWXY
µ¶·
¹̧º
hi = s(W i h̃i−1 + bi ).
ÜÝÞ ÷øù
èéê
ßàá úûü ôõö ëìí
åæç
òóñ
âãä ýþÿ
îïð
åæçèéêëìíîï ðñòóô õö÷ø
yz{|}~
Z[\]^_ `abcde fghi
()*
¦§¨
©ª« ²³´
£¤¥ °̄±
¡¢ ¬®
§¨©
ª«¬
"#$
ÅÆÇ ÈÉÊËÌ
%&'
!
jklm
¿
À
Á
¤
¥
¦
®¯
»¼½¾
¿ÀÁÂ
ÇÈÉ
ÃÄÅÆ
ÙÚÛ
It should be noted that layer i = 0 represents an input layer, that
is, h̃0 = x̃s . The weight and bias are computed by minimizing a
de-noising reconstruction error [38].
At the output layer ((L + 1)-th layer), we apply a regression function f (.) and the cost function of Ds as follows,
nopq
v
w
x
rstu
"
¡¢£
ÊËÌÍ
ÎÏÐÑ
Ö×Ø
ÒÓÔÕ
s
l({y s , xs }) =
N
1 X s
||y j − f (hL |xsj )||2 ,
s
N j
(2)
°±²³
´µ¶·
¼
½
¾
¸¹º»
+,-./012 3456789
ÕÖ×ØÙÚ ÛÜÝÞßà áâãä
(1)
ùúûüýþÿ
ÍÎÏ ÐÑÒÓÔ
We first explain the SdA for constructing Ds . Let xs ∈ RDx denote
a Dx dimensional source vector and W i and bi denote a weight
matrix and bias vector of the i-th hidden layer (i = 1, 2, · · · , L),
respectively. Let x̃ denote a corrupting vector drawn from corruption
process q(x̃|x), and s(.) denote a sigmoid function. Then, the i-th
hidden vector of Ds is as follows,
ÂÃÄ
where N s is the amount of source domain data, y sj is a label vector
of xsj (yjs (k) = {0, 1}, k = 1, 2, · · · , Dy s ), Dy s is the dimension
of y s , and
f (hL |xsj ) = hL+1 = W L+1 hL + bL+1 .
(3)
The parameters of all layers are simultaneously fine-tuned using
a stochastic gradient descent. In this article, we use {W i , bi |i =
1, 2, · · · , L + 1} as the initial parameters of Dt .
:;<=>?@ABCD EFGHIJKLM
Figure 4. Outline of ATDL. (A): Training DNN for source task, (B): computing relation vectors of each target label, (C): tuning all parameters to transfer Ds to Dt .
3 All-Transfer Deep Learning
3.2.2 Computation of Relation Vectors
Relation vectors represent the characteristics of the target labels in
the Dy s dimensional feature space computed by the source domain
knowledge, Ds . Let rl ∈ RDys denote the l-th relation vector (l =
1, 2, · · · , Dy t ), Dy t denote the number of target labels, and xtl ∈
RDx denote a target vector corresponding to the l-th target label.
Then, rl is computed using the following equation.
3.1 Overview
rl = arg max p(hL+1 |xtl ),
hL+1
An outline of the ATDL training process is shown in Figure 4. First,
ATDL trains a DNN, Ds , to solve the task of the source domain (Figure 4 (A)). In this study, we constructed Ds on the basis of stacked
de-noising autoencoder (SdA) [12, 38]. Second, ATDL computes the
output vectors of each target vector by inputting them into the DNN
trained on the source domain data (Figure 4 (B)). It then computes
the relation vectors of each target label. A relation vector denotes a
vector representing the relationship between the source and target labels on the basis of the source domain knowledge, Ds , by using the
output vectors. Finally, we fine-tune all parameters in such a way that
the variance between the output and relation vectors is sufficiently
small (Figure 4 (C)). By using the steps corresponding to Figures 4
(B) and (C), we can transfer Ds to the DNN for the target task Dt ,
regularizing all parameters including the output layer. This means
that ATDL provides Dt , which can avoid the local-optimal solution
caused by the randomness of the initial values.
(4)
where p(hL+1 |xtl ) is the probability distribution of hL+1 given xtl .
We assume p(hL+1 |xtl ) obeys a Gaussian distribution. Therefore, rl
is equal to the average vector of f (hL |xtl ).
t
N (l)
X
1
rl = t
f (hL |xtl,j ),
N (l) j
(5)
where xtl,j and N t (l) are the j-th target domain vector and amount
of target domain data corresponding to the l-th target label, respectively. The k-th variable rl (k) means the strength of the relationship
between the k-th source label and l-th target label. Therefore, by confirming the values of relation vectors, we can understand which labels
of the source domain data are similar to those of the target domain
data.
3.2.3 Fine-tuning
After computing rl , we set r l as the l-th label vector of the target
task, and all parameters including W L+1 and bL+1 are fine-tuned
by minimizing the following main cost function using a stochastic
gradient descent. It should be noted that this equation represents the
variance of the target domain data.
D
l({r, xt }) =
Table 1.
CPU
Memory
GPU
Experimental environment.
Intel (R) Core (TM) i7-4930K
64.0GB
GeForce GTX 760
t
yt N (l)
1 X X
||r l − f (hL |xtl,j )||2 ,
Nt
j
(6)
l
where N t = N t (1) + N t (2) + · · · + N t (Dy t ). By using this algorithm, we can have Dt regularizing all parameters by using Ds .
3.3 Classification Process
In the classification process, Dt predicts the label l̂ of the test vector
x on the basis of the following equation.
l̂ = arg min(rl − f (hL |x))⊤ Σl (rl − f (hL |x)),
(7)
Figure 5. Examples of 2-DE images that differ in extraction and refining
protocol of protein. Label number of source domain data is from 1 (top left)
to 9 (bottom right) in order.
l
where Σl is a covariance matrix. It should be noted that classification
performance does not improve if r l̂ ≈ rl̂′ (l̂ 6= l̂′ ). This means that
the source domain/task is not suitable for transfer.
4 Experimental Results
We conducted experiments on 2-DE image classification for determining if an individual suffers from sepsis. We compared the classification performance of five methods: non-transfer learning, simple semi-supervised learning (SSL), transfer learning by Agrawal et
al. [1], that by Oquab et al. [29], and ATDL.
The SSL is a method to construct a mixture model that computes
hi using xs and xt and fine-tunes using only xt . In addition, this
method is a special case of that by Weston et al. [40], which embeds
the regularizer to the output layer, when the parameter to balance
between the object function and regularizer is zero.
Agrawal’s method removes the output layer of Ds and adds a new
output layer. In addition to these two steps, Oquab’s method contains
an additional adaptation layer to compensate for the different statistics of the source and target domain data. Then, Agrawal’s method
fine-tunes all layers including the hidden layers [41], and Oquab’s
method fine-tunes only the adaptation and output layers.
In our study, we used a soft max function as the output layer for
constructing Ds of the above transfer learning, SSL, and non-transfer
learning methods.
To investigate the difference in classification performance, we
changed the source domain data and evaluated classification performance. We used 2-DE images that were given different labels from
the target domain data of sepsis or non-sepsis, MNIST [25], and
CIFAR-10 [21], as the source domain data. In addition, to investigate the generality of our method, we applied it to a convolutional
neural network (CNN) [25] and a different task of open-image-data
classifications. For open-image-data classifications, we investigated
the effectiveness of our method for two open image data classifications. Finally, we investigated the correlation coefficients between
the classification performance and Mahalanobis distance of rl .
4.1 Environment and Hyperparameter Settings
We used the computer environment summarized in Table 1 and
pylearn2 [15] to minimize (2) and (6).
In this study, we set the learning rate to λ/(1.00004 × t), where t
is the iteration. Momentum gradually increased from 0.5 to µ when
t increased. We selected the initial learning rate λ from {1.0 ×
10−3 , 5.0×10−3 , 1.0×10−2 , 5.0×10−2 }, final momentum µ from
{0.7, 0.99}, and size of minibatches from {10, 100}.
4.2 Actual Sepsis-Data Classification
Table 2. List of source 2-DE images. These images represent different extraction and refining protocols of proteins.
# of source 2-DE images
Type of protocol
N s (1) = 25
Change amount of protein
N s (2) = 4
Change concentration protocol
N s (3) = 30
Unprocessed
N s (4) = 49
Removal of only top-2 abundant proteins
N s (5) = 11
Focus on top-2 abundant proteins
N s (6) = 15
Focus on 14 abundant proteins
N s (7) = 12
Plasma sample instead of serum
N s (8) = 19
Removal of Sugar chain
N s (9) = 15
Other protocols
For actual sepsis-data classification, we collected the following
number of target 2-DE images N t = 98, sepsis data of N t (1) = 30
and non-sepsis data of N t (2) = 68. The size of the 2-DE images was
53 × 44 pixels (Dx = 2, 332), which was determined to save the information of the large spots. We evaluated classification performance
on the basis of two-fold cross validation. As the source domain data,
we first used 2-DE images with different labels from the target domain sepsis or non-sepsis data. These images were generated from
patients which were diagnosed as being normal. The source task was
to classify the differences between the extraction and refining protocols of proteins [39] shown in Table 2 and Figure 5. As shown in
Figure 6.
Example of relation vectors of actual sepsis-data classification.
this table, we set N s = 180 and Dy t = 9. On the other hand, the
target 2-DE images were generated using serum and by removing 14
abundant proteins. These data were generated from actual patients at
Juntendo University Hospital and were judged by infectious disease
tests and SOFA/SIRS score [37]. This study was approved by the institutional review board, and written informed consent was obtained
from patients.
4.2.1 Comparison with Conventional Methods
Table 3. Classification performance of actual
function of the number of hidden layers.
PPV
NPV
PCA + logistic regression
0.875
0.805
Non-transfer (L=1)
0.725
0.983
Non-transfer (L=2)
0.718
0.967
Non-transfer (L=3)
0.644
0.981
Non-transfer (L=4)
0.644
0.981
SSL (L=1)
0.682
1
SSL (L=2)
0.644
0.981
SSL (L=3)
0.592
0.980
SSL (L=4)
0.558
0.978
Oquab et al. [29] (L=1)
0.732
1
Oquab et al. [29] (L=2)
0.771
0.952
Oquab et al. [29] (L=3)
0.702
0.934
Oquab et al. [29] (L=4)
0.658
0.947
Agrawal et al. [1] (L=1)
0.750
1
Agrawal et al. [1] (L=2)
0.744
0.983
Agrawal et al. [1] (L=3)
0.690
0.982
Agrawal et al. [1] (L=4)
0.667
1
ATDL (L=1)
0.844
0.955
ATDL (L=2)
0.871
0.955
ATDL (L=3)
0.875
0.970
ATDL (L=4)
0.958
0.905
sepsis-data classification as
MCC
0.545
0.755
0.726
0.676
0.676
0.736
0.676
0.620
0.580
0.783
0.753
0.670
0.648
0.800
0.796
0.722
0.720
0.812
0.834
0.859
0.806
F1
0.609
0.829
0.811
0.773
0.773
0.811
0.773
0.734
0.707
0.845
0.831
0.776
0.761
0.857
0.841
0.806
0.8
0.871
0.885
0.903
0.852
ACC
0.816
0.878
0.867
0.827
0.827
0.857
0.827
0.786
0.755
0.888
0.888
0.847
0.827
0.898
0.888
0.857
0.847
0.918
0.929
0.939
0.918
We compared the classification performances, including that of
ATDL, with respect to the changing number of hidden layers L =
1, 2, 3 and 4. We set the dimension of the 1st hidden layer to D1 =
188 by PCA using xs and xt (cumulative contribution of 188 features is over 99.5%), and D2 , D3 , and D4 were set to the same dimensions.
Table 3 lists the classification accuracies (ACCs) of six methods
including the baseline, PCA + logistic regression (used 188 features).
It also lists the positive predictive values (PPVs), negative predictive
values (NPVs), Matthews correlation coefficients (MCCs), and F1scores (F1s) as reference. It should be noted that PPV and NPV are
Figure 7. Example of relation vectors when source domain data are from
CIFAR-10.
used in diagnostic tests, MCC is used for evaluating performance
considering the unbalance-ness of N t (1) and N t (2), and F1 is the
harmonic value computed by precision (=PPV) and recall.
As shown in this table, classification accuracy improved by using transfer learning. In addition, the classification accuracy of
ATDL (L = 3) outperformed those of the other transfer learning
methods. For example, the classification accuracy of ATDL improved
at least 4 percentage points compared to that of Agrawal’s method of
L = 1. These results suggest that ATDL is effective for performing
actual sepsis-data classification.
Figure 6 shows an example of relation vectors of sepsis and nonsepsis (L = 3). The red bars represent the relation vector of sepsis,
whereas the green bars represent that of non-sepsis. The numbers on
the X-axis correspond to the source label indices listed in Table 2. As
shown in this figure, the relation vectors of sepsis and non-sepsis differed. These results suggest that rl can represent the characteristics
of the target label in the feature space computed by Ds .
4.2.2 Comparison of Various Source Tasks
Table 4. Classification performance of actual sepsis-data classification for
different source tasks.
Non-transfer (Di = 188)
Non-transfer (Di = 500)
Non-transfer (Di = 1, 000)
CIFAR-10 (Di = 188)
CIFAR-10 (Di = 500)
CIFAR-10 (Di = 1, 000)
MNIST (Di = 188)
MNIST (Di = 500)
MNIST (Di = 1, 000)
2-DE image (Di = 188)
2-DE image (Di = 500)
2-DE image (Di = 1, 000)
PPV
0.718
0.644
0.7
0.657
0.923
0.690
0.778
0.839
0.828
0.875
0.844
0.824
NPV
0.967
0.981
0.966
0.889
0.912
0.982
0.968
0.940
0.913
0.970
0.955
0.969
MCC
0.726
0.676
0.709
0.568
0.804
0.722
0.780
0.786
0.735
0.859
0.812
0.818
F1
0.811
0.773
0.8
0.708
0.857
0.806
0.849
0.852
0.813
0.903
0.871
0.875
ACC
0.867
0.827
0.857
0.806
0.918
0.857
0.898
0.908
0.888
0.939
0.918
0.918
We compared classification performance with respect to changing
the source domain data, which were obtained from MNIST, CIFAR10, and 2-DE images. The number of images extracted from MNIST
and CIFAR-10 were N s = 50, 000. The CIFAR-10 images were
converted to gray-scale, and the MNIST and CIFAR-10 images were
resized to Dx = 53 × 44 = 2, 332 to ensure they were aligned with
the 2-DE images. In addition, we set L = 3 and D1 = D2 = D3 . We
Figure 8. Example of relation vectors when source domain data are from
MNIST.
Figure 9. CNN structure.
also evaluated classification performance with respect to changing
the dimension of each hidden layer Di = 188, 500, and 1, 000.
Table 4 lists the classification accuracies, and the PPVs, NPVs,
MCCs, and F1s as reference. The classification accuracy based on
the use of 2-DE images as the source domain data was higher than
those obtained with MNIST and CIFAR-10, although the number of
2-DE images was smaller (N s = 180).
Figure 7 shows an example of the relation vectors using CIFAR10 (Di = 500) and Figure 8 shows them using MNIST (Di = 500).
Compared to Figure 6, the relation vector of sepsis was considerably
closer to that of non-sepsis.
These results show that information on the differences between
the extraction and refining protocols of proteins is useful for classifying sepsis, rather than using CIFAR-10 and MNIST. Namely, if we
collect the source domain data, we have to consider the relationship
between the source and target domain data.
4.2.3 Applying ATDL to Convolutional Neural Network
Table 5. Classification performance when ATDL was applied to CNN.
PPV
NPV
MCC
F1
ACC
Non-transfer
0.717
0.966
0.726
0.812
0.867
ATDL
0.829
0.984
0.845
0.892
0.929
To investigate the effectiveness of our method regarding other
DNNs, we applied it to a CNN and evaluated its classification performance by using 2-DE images as the source domain data. Figure 9
shows the structure of the CNN, which was determined on the basis
of two-fold cross validation with respect to changing the hyperparameters shown in this figure.
Table 5 lists the classification performances. The ATDL performed
better than the non-transfer learning method and approximately equal
to the SdA of ATDL (L = 3) shown in Table 3. Thus, these results
suggest that ATDL is applicable to CNNs as well as SdAs. The CNN
Figure 10. Relation vectors when ATDL was applied to CNN. Numbers on
X-axis correspond to source label indices.
is widely used in image recognition and achieves high classification
accuracy on several standard data [19, 23, 24]. Therefore, we consider that ATDL is possible to be applied to various image recognition problems.
Figure 10 shows the relation vectors. Sepsis had a relationship to
the 4th source label (removal of only top-2 abundant proteins), which
is the same as in Figure 6. This result suggests that there are biological relationships between them. In the future, we plan to examine
this result from a biological point of view.
4.3 Open-Image Data Experiment
Table 6. Accuracy of automobile and pedestrian crossing.
# of target images N t
400
1,500
Non-transfer
0.724
0.753
SSL
0.750
0.789
Oquab et al. [29]
0.763
0.782
Agrawal et al. [1]
0.753
0.781
ATDL
0.789
0.797
Table 7. Accuracy of MNIST.
# of target images N t
1,000
5,000
Non-transfer
0.854
0.926
SSL
0.844
0.928
Oquab et al. [29]
0.773
0.875
Agrawal et al. [1]
0.844
0.923
ATDL
0.887
0.928
10,000
0.945
0.951
0.887
0.951
0.932
To investigate the generalization of our method, we first applied
our method to two open image data classifications: (1) CIFAR10 [21] as the source domain and images of an automobile and
pedestrian crossing from ImageNet [22] as the target domain, and (2)
SVHN [27] as the source domain and MNIST [25] as the target domain. Task (1) is an example in which the source/target domain data
consist of color images, and (2) is an example of multiclass classification.
For task (1), we constructed Ds on the basis of the SdA and set
L = 3, Dy s = 10, Di = 1, 000 (i = 1, 2, 3), and N s = 50, 000.
As the target test data, we used 750 images of automobile and 750
images of pedestrian crossing. For task (2), we also constructed Ds
on the basis of the SdA and set L = 3, Dy s = 10, Di = 100 (i =
Figure 11.
Example of relation vectors of task (1).
89:
567
F
G
234
H
I
/01
[ ,-.
Figure 13. Relationship between Mahalanobis distance and classification
performance. Blue lines represent classification performance of non-transfer
learning method.
)*+
&'(
%
!#$
;
<
=
>
?
@
A
B
JKLMNO PQRST UVWXYZ
C
DE
Figure 12. Example of relation vectors of task (2) (N t = 1, 000).
Table 8. Correlation coefficient R and p-value of each classification performance.
PPV
NPV
MCC
F1
ACC
R
0.627
0.521
0.663
0.657
0.665
p-value
< 0.01
< 0.01
< 0.01
< 0.01
< 0.01
1, 2, 3), N s = 73, 257, and SVHN images were converted to grayscale. As the target test data, we used 10, 000 images from MNIST.
Test images of task (1) and (2) were not included in the training target
domain data.
Table 6 and 7 list the classification accuracies of the five methods for different amounts of target domain data. It should be noted
that we could not conduct actual sepsis-data classification in this
experiment because collecting sepsis data is difficult. As shown in
these tables, our method outperformed other methods when N t =
400, 1, 500 for task (1) and N t = 1, 000 for task (2). These results
suggest that our method is effective when the amount of target domain data is significantly small.
Figure 11 and 12 show examples of the relation vectors of each
task. As shown in these figures, the target automobile showed a relationship with the source automobile (2nd source label) and source
truck (10th source label). In addition, the highest relation of the character “6” of MNIST was the character “6” of SVHN. On the other
hand, the relation vectors of the target automobile/pedestrian crossing differed. These results suggest that r l enabled the representation
of the target label characteristics in the feature space computed by
Ds , the same as 2-DE images.
transfer.
Let M denote the number of source domain sub-groups, Nas (a =
1, 2, · · · , M ) denote the amount of a-th source domain sub-group
data, and X a = {y sb , xsb |b = 1, 2, · · · , Nas } denote the a-th source
domain sub-group data sampled from all source domain data X. We
constructed Dsa by using X a and computed the Mahalanobis distance dm (a) of r l by inputting the a-th DNN Dsa . Then, we finetuned to transfer from Dsa to Dt .
We used MNIST as X and set M = 100 and Nas = 5, 000 (a =
1, 2, · · · , M ). Sub-groups were randomly selected from X, and
Dsa was constructed on the basis of the SdA. The target task was
sepsis-data classification, and we set L = 1, D1 = 188, and
N t = 49 (N t (1) = 15, N t (2) = 34). As the target test data, we
used 15 images of sepsis and 34 images of non-sepsis. It should be
noted that all hyperparameters were fixed.
To evaluate the relationship between dm (a) and classification performance t(a), we computed the correlation coefficient R as follows.
4.4 Correlation of Performance and Distance
As described above, the classification performance of ATDL depends
on the distance of rl . In this subsection, we discuss the investigation of the correlation between classification performance and Mahalanobis distance dm . If they correlate, ATDL can be used to select
Ds before the fine-tuning process and reduce the risk of negative
PM
(dm (a) − d¯m )(t(a) − t̄)
R = qP a
,
M
¯ 2 PM (t(a) − t̄)2
a (dm (a) − dm )
a
(8)
where d¯m and t̄ are the averages of dm and t. Figure 13 shows dm (a)
and the corresponding classification performances, and Table 8 lists
the correlation coefficients and p-values. The blue lines in Figure 13
represent the performance of non-transfer learning. The Mahalanobis
distance and classification performances correlated, suggesting that
higher classification performance than that of non-transfer learning
is possible by using Dsa with large dm (a). This means that we can
select Dsa effectively before the fine-tuning process.
5 Conclusion
We proposed ATDL, a novel transfer learning method for DNNs, for
a significantly small amount of training data. It computes the relation
vectors that represent the characteristics of target labels by the source
domain knowledge. By using the relation vectors, ATDL enables the
transfer of all knowledge of DNNs including the output layer.
We applied ATDL to actual sepsis-data classification. The experimental results showed that ATDL outperformed other methods. We
also investigated the generality of ATDL with respect to changing
the DNN model and target task, and compared the classification performance with respect to changing the source domain data. From the
results, we argue that our method is applicable to other tasks, especially when the amount of target domain data was significantly small,
and classification performance improves when we use the source domain data that are similar to the target domain data. Furthermore,
we showed the possibility of selecting an effective DNN before the
fine-tuning process.
To the best of our knowledge, this work involved the first trial in
which 2-DE images were analyzed using the classification approach,
which resulted in over 90% accuracy. Thus, this article will be influential not only in machine learning, but also medical and biological
fields.
In the future, we will collect source domain 2-DE images that can
be uploaded easily, apply ATDL to a deeper network, predict classification performance more accurate, and analyze the relation vectors
from a biological point of view.
6 Acknowledgments
We would like to acknowledge Dr. Iba, Professor of Juntendo University School of Medicine, for his help in collecting samples to generate 2-DE images.
REFERENCES
[1] Pulkit Agrawal, Ross Girshick, and Jitendra Malik, ‘Analyzing the performance of multilayer neural networks for object recognition’, in Proceedings of the European Conference on Computer Vision, 329–344,
(2014).
[2] Stephen Barnes, Helen Kim, and Victor M Darley-Usmar, ‘High
throughput two-dimensional blue-native electrophoresis: a tool for
functional proteomics of mitochondria and signaling complexes’, Proteomics, 2, 969–977, (2002).
[3] John MS Bartlett and David Stirling, ‘A short history of the polymerase
chain reaction’, PCR protocols, 3–6, (2003).
[4] Yoshua Bengio, ‘Learning deep architectures for AI’, Foundations and
trends in Machine Learning, 2(1), 1–127, (2009).
[5] Matthias Berth, Frank Michael Moser, Markus Kolbe, and Jörg Bernhardt, ‘The state of the art in the analysis of two-dimensional gel electrophoresis images’, Applied microbiology and biotechnology, 76(6),
1223–1243, (2007).
[6] R Phillip Dellinger, Mitchell M Levy, Andrew Rhodes, Djillali Annane,
Herwig Gerlach, Steven M Opal, Jonathan E Sevransky, Charles L
Sprung, Ivor S Douglas, Roman Jaeschke, et al., ‘Surviving sepsis campaign: international guidelines for management of severe sepsis and
septic shock, 2012’, Intensive care medicine, 39(2), 165–228, (2013).
[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei,
‘Imagenet: A large-scale hierarchical image database’, in Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition,
pp. 248–255, (2009).
[8] Chuong Do and Andrew Y Ng, ‘Transfer learning for text classification’, in Proceedings of the Advances in Neural Information Processing
Systems, pp. 299–306, (2005).
[9] Viktor Y Dombrovskiy, Andrew A Martin, Jagadeeshan Sunderram,
and Harold L Paz, ‘Rapid increase in hospitalization and mortality rates
for severe sepsis in the united states: A trend analysis from 1993 to
2003’, Critical care medicine, 35(5), 1244–1250, (2007).
[10] Jeff Donahue, Yangqing Jia, Oriol Vinyals, Judy Hoffman, Ning Zhang,
Eric Tzeng, and Trevor Darrell, ‘Decaf: A deep convolutional activation
feature for generic visual recognition’, in Proceedings of the International Conference on Machine Learning, pp. 647–655, (2014).
[11] Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang, ‘Learning a deep convolutional network for image super-resolution’, in Proceedings of the European Conference on Computer Vision, 184–199,
Springer, (2014).
[12] Dumitru Erhan, Yoshua Bengio, Aaron Courville, Pierre-Antoine Manzagol, Pascal Vincent, and Samy Bengio, ‘Why does unsupervised pretraining help deep learning?’, The Journal of Machine Learning Research, 11, 625–660, (2010).
[13] Mark Everingham, Luc Van Gool, Christopher KI Williams, John
Winn, and Andrew Zisserman, ‘The pascal visual object classes (VOC)
challenge’, International journal of computer vision, 88(2), 303–338,
(2010).
[14] Yaroslav Ganin and Victor Lempitsky, ‘Unsupervised domain adaptation by backpropagation’, in ICML, pp. 1180–1189, (2015).
[15] Ian J. Goodfellow, David Warde-Farley, Pascal Lamblin, Vincent Dumoulin, Mehdi Mirza, Razvan Pascanu, James Bergstra, Frédéric
Bastien, and Yoshua Bengio, ‘Pylearn2: a machine learning research
library’, arXiv preprint arXiv:1308.4214, (2013).
[16] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean, ‘Distilling the knowledge in a neural network’, in Proceedings of the Deep Learning and
Representation Learning Workshop, (2014).
[17] Atsunori Hiratsuka, Hideki Kinoshita, Yuji Maruo, Katsuyoshi Takahashi, Satonari Akutsu, Chie Hayashida, Koji Sakairi, Keisuke Usui,
Kisho Shiseki, Hajime Inamochi, et al., ‘Fully automated twodimensional electrophoresis system for high-throughput protein analysis’, Analytical chemistry, 79(15), 5730–5739, (2007).
[18] Junlin Hu, Jiwen Lu, and Yap-Peng Tan, ‘Deep transfer metric learning’, in Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pp. 325–333, (2015).
[19] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev,
Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell, ‘Caffe: Convolutional architecture for fast feature embedding’, in
Proceedings of the ACM International Conference on Multimedia, pp.
675–678, (2014).
[20] Jing Jiang and ChengXiang Zhai, ‘Instance weighting for domain adaptation in nlp’, in Proceedings of the Annual Meeting of the Association
for Computational Linguistics, pp. 264–271, (2007).
[21] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of
features from tiny images. master’s thesis, university of tronto, 2009.
[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton, ‘Imagenet
classification with deep convolutional neural networks’, in Proceedings
of the Advances in neural information processing systems, pp. 1097–
1105, (2012).
[23] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton, ‘Imagenet
classification with deep convolutional neural networks’, in NIPS, volume 1, pp. 1–9, (2012).
[24] Quoc V Le, ‘Building high-level features using large scale unsupervised learning’, in Proceedings of the IEEE International Conference
on Acoustics, Speech and Signal Processing, pp. 8595–8598, (2013).
[25] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner,
‘Gradient-based learning applied to document recognition’, in Proceedings of the IEEE, volume 86, pp. 2278–2324, (1998).
[26] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan,
‘Learning transferable features with deep adaptation networks’, in
ICML, pp. 97–105, (2015).
[27] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu,
and Andrew Y Ng, ‘Reading digits in natural images with unsupervised
feature learning’, In NIPS workshop on deep learning and unsupervised
feature learning, 1–9, (2011).
[28] Patrick H O’Farrell, ‘High resolution two-dimensional electrophoresis of proteins.’, Journal of biological chemistry, 250(10), 4007–4021,
(1975).
[29] Maxime Oquab, Leon Bottou, Ivan Laptev, and Josef Sivic, ‘Learning
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
and transferring mid-level image representations using convolutional
neural networks’, in Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pp. 1717–1724, (2014).
Sinno Jialin Pan, Ivor W Tsang, James T Kwok, and Qiang Yang, ‘Domain adaptation via transfer component analysis’, IEEE Transactions
on Neural Networks, 22(2), 199–210, (2011).
Sinno Jialin Pan and Qiang Yang, ‘A survey on transfer learning’, IEEE
Transactions on Knowledge and Data Engineering, 22(10), 1345–1359,
(2010).
Thierry Rabilloud, Mireille Chevallet, Sylvie Luche, and Cécile Lelong, ‘Two-dimensional gel electrophoresis in proteomics: past, present
and future’, Journal of proteomics, 73(11), 2064–2077, (2010).
Thierry Rabilloud and Cécile Lelong, ‘Two-dimensional gel electrophoresis in proteomics: a tutorial’, Journal of proteomics, 74(10),
1829–1841, (2011).
Michael T Rosenstein, Zvika Marx, Leslie Pack Kaelbling, and
Thomas G Dietterich, ‘To transfer or not to transfer’, in Proceedings
of the NIPS’05 Workshop on Inductive Transfer: 10 Years Later, volume 2, pp. 7–10, (2005).
Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell, ‘Adapting
visual category models to new domains’, in Proceedings of the European Conference on Computer Vision, 213–226, (2010).
Yoshihide Sawada, Yoshikuni Sato, Toru Nakada, Kei Ujimoto, and
Nobuhiro Hayashi, ‘All-transfer learning for deep neural networks and
its application to sepsis classification’, in ECAI, pp. 1586–1587, (2016).
Mervyn Singer, Clifford S Deutschman, Christopher Warren Seymour,
Manu Shankar-Hari, Djillali Annane, Michael Bauer, Rinaldo Bellomo,
Gordon R Bernard, Jean-Daniel Chiche, Craig M Coopersmith, et al.,
‘The third international consensus definitions for sepsis and septic
shock (sepsis-3)’, Jama, 315(8), 801–810, (2016).
Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, and
Pierre-Antoine Manzagol, ‘Stacked denoising autoencoders: Learning
useful representations in a deep network with a local denoising criterion’, The Journal of Machine Learning Research, 11, 3371–3408,
(2010).
John M Walker, The proteomics protocols handbook, Springer, 2005.
Jason Weston, Frédéric Ratle, Hossein Mobahi, and Ronan Collobert,
‘Deep learning via semi-supervised embedding’, in Neural Networks:
Tricks of the Trade, 639–655, Springer, (2012).
Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson, ‘How
transferable are features in deep neural networks?’, in Advances in Neural Information Processing Systems, pp. 3320–3328, (2014).
Bolei Zhou, Agata Lapedriza, Jianxiong Xiao, Antonio Torralba, and
Aude Oliva, ‘Learning deep features for scene recognition using places
database’, in Proceedings of the 27th Advances in Neural Information
Processing Systems, pp. 487–495, (2014).
Fuzhen Zhuang, Xiaohu Cheng, Ping Luo, Sinno Jialin Pan, and Qing
He, ‘Supervised representation learning: transfer learning with deep autoencoders’, in Proceedings of the International Conference on Artificial Intelligence, pp. 4119–4125, (2015).
| 1cs.CV
|
arXiv:1504.01148v2 [math.AC] 12 May 2016
On The Cohomological Dimension of Local Cohomology Modules
Vahap Erdoğdu and Tuğba Yıldırım∗
Dedicated to the memory of Alexander Grothendieck
Department of Mathematics
Istanbul Technical University
Maslak, 34469, Istanbul, Turkey
Abstract
Let R be a Noetherian ring, I an ideal of R and M an R-module
with cd(I, M ) = c. In this article, we first show that there exists a
descending chain of ideals I = Ic ) Ic−1 ) · · · ) I0 of R such that for
each 0 ≤ i ≤ c − 1, cd(Ii , M ) = i and that the top local cohomology
module HiIi (M ) is not Artinian. We then give sufficient conditions
for a non-negative integer t to be a lower bound for cd(I, M ) and
use this to conclude that in non-catenary Noetherian local integral
domains, there exist prime ideals that are not set theoretic complete
intersection. Finally, we set conditions which determine whether or
not a top local cohomology module is Artinian.
Keywords: Top local cohomology modules, Cohomological dimensions, Set
theoretic complete intersections.
2000 Mathematics Subject Classification. 13D45, 13E10.
1
Introduction
Throughout, R denote a commutative Noetherian ring with unity, I an ideal
of R. For an R-module M, the i-th local cohomology module of M with
support in I is defined as
HiI (M) = lim ExtiR (R/I n , M).
→
∗
Corresponding author
1
For details about the local cohomology modules, we refer the reader to [4]
and [7].
One of the important invariant related to local cohomology modules is the
cohomological dimension of M with respect to I, denoted by cd(I, M), and
defined as:
cd(I, M) = sup {i ∈ N | HiI (M) 6= 0}.
If M = R, we write cd(I) instead of cd(I, R).
There are two interesting questions related to local cohomology modules, the
first one is to determine the lower and upper bounds for cd(I, M) and the
second one is to determine whether or not HiI (M) is Artinian(see e.g. [1], [5],
[6], [8], [9] and [10]).
Our results in this regard are as follows:
In section 2, we show that for an R-module M with cd(I, M) = c, there is a
descending chain of ideals
I = Ic ) Ic−1 ) · · · ) I0
of R such that for each 0 ≤ i ≤ c, cd(Ii , M) = i, and that the top local
cohomology module HiIi (M) is not Artinian.
In section 3, we prove a result that gives a sufficient condition for an integer
to be a lower bound for the cohomological dimension, cd(I, M), of M at
I. One of the important conclusion of this result is that over a Noetherian
local ring (R, m), for a finitely generated R-module M of dimension n and an
ideal I of R with dim(M/IM) = d ≥ 1, n − d is a lower bound for cd(I, M)
and if moreover cd(I, M) = n − d, then Hdm (HIn−d (M)) ∼
= Hnm (M). As an
application of this result, we show that in non-catenary Noetherian local
integral domains, there exist prime ideals that are not set theoretic complete
intersection.
In section 4, we examine the Artinianness and non-Artinianness of top local
cohomology modules.
2
Descending Chains With Successive Cohomological Dimensions
In this section, we prove the existence of descending chains of ideals and
locally closed sets with successive cohomological dimensions. The main result
of this section is the following:
2
Theorem 2.1. Let R be a Noetherian ring, I an ideal of R and M an Rmodule with cd(I, M) = c > 0. Then there is a descending chain of ideals
I = Ic ) Ic−1 ) · · · ) I0
such that cd(Ii , M) = i for all i, 0 ≤ i ≤ c. Moreover HiIi (M) is not Artinian
for all i, 0 ≤ i ≤ c − 1.
Proof. Consider the set
S = { J ( I | cd(J, M) < c }.
Clearly, the zero ideal belongs to S and so S is a non-empty subset of ideals
of R. Since R is Noetherian, S has a maximal element, say Ic−1 . We claim
that cd(Ic−1, M) = c − 1. To prove this, let x ∈ I \ Ic−1 and so Ic−1 + Rx ⊆ I.
But then it follows from the maximality of Ic−1 in S and Remark 8.1.3 of [4]
that
c ≤ cd(Ic−1 + Rx, M) ≤ cd(Ic−1 , M) + 1 < c + 1.
Hence cd(Ic−1 + Rx, M) = c. Now consider the exact sequence
···
/
(Hc−1
Ic−1 (M))x
/
HcIc−1 +Rx (M)
/
HcIc−1 (M) = 0 .
Since HcIc−1 +Rx (M) is nonzero, it follows that (Hc−1
Ic−1 (M))x is nonzero, then
c−1
so is HIc−1 (M). Therefore the claim follows.
Iterating this argument, one can obtain a descending chain of ideals, as desired.
For the second part, let x ∈ Ii+1 \ Ii and consider the ideal Ii + Rx. Then
it follows from the construction of the ideal Ii that cd(Ii + Rx, M) = i + 1.
Now HiIi (M) is non-Artinian follows from Corollary 4.1 of [5].
Recall that a subspace Z of a topological space X is said to be locally
closed, if it is the intersection of an open and a closed set. Let X be a
topological space, Z ⊆ X be a locally closed subset of X and let F be an
abelian sheaf on X. Then the ith local cohomology group of F with support
in Z is denoted by HZi (X, F ). We refer the reader to [7] and [11] for its
definition and details.
If, in particular, X = Spec(R) is an affine scheme, where R is a commutative
Noetherian ring, and F = M ∼ is the quasi coherent sheaf on X associated
to an R-module M, we write HZi (M) instead of HZi (X, M ∼ ).
The following corollary may be considered as an easy application of our result
above.
3
Corollary 2.2. Let R be a Noetherian ring, M an R-module and I an ideal
of R such that cd(I, M) = c > 1. Then there is a descending chain of locally
closed sets
Tc−1 ) Tc−2 ) · · · ) T1
in Spec(R) such that cd(Ti , M) = i for all 1 ≤ i ≤ c − 1.
Proof. Let I be an ideal of R with cd(I, M) = c > 1. Then it follows from
Theorem 2.1 that there is a descending chain of ideals
I = Ic ) Ic−1 ) · · · ) I1 ) I0
such that cd(Ii , M) = i for all 0 ≤ i ≤ c. Let now Ui = V (Ii ) and define the
locally closed sets Ti := U1 \ Ui+1 . Then it is easy to see that
Tc−1 ) Tc−2 ) · · · ) T1 .
On the other hand, it follows from Proposition 1.2 of [11] that there is a long
exact sequence,
···
/
HUj 1 (M)
/
HTj i (M)
/ H j+1 (M)
Ui+1
/
HUj+1
(M)
1
/
···
j
As HUj i (M) ∼
= HIi (M) for all 1 ≤ i ≤ c − 1 and for all j ≥ 0, it follows from
the above long exact sequence that cd(Ti , M) = i.
3
Lower Bound For Cohomological Dimension
The main purpose of this section is to establish a lower bound for cohomological dimension and, in this regard, we prove the following theorem which
gives a sufficient condition for an integer t to be a lower bound for cd(I, M).
Theorem 3.1. Let R be a Noetherian ring, M an R-module (not necessarily
finitely generated) and I an ideal of R with dim(R/I + Ann M ) = d. Let
t ≥ 0 be an integer. If there exists an ideal J of R such that Hd+t
I+J (M) 6= 0,
then t is a lower bound for cd(I, M). Moreover, if cd(I, M) = t, then
HdJ (HtI (M)) ∼
= Hd+t
I+J (M)
and dim Supp(HtI (M)) = d.
4
Proof. Consider the Grothendieck’s spectral sequence
E2p,q = HpJ (HqI (M)) =⇒ Hp+q
I+J (M)
and look at the stage p + q = n. Since Supp(HqI (M)) ⊆ V (I) ∩ Supp(M) ⊆
V (I + Ann M), dim Supp(HqI (M)) ≤ d for all q. Therefore it follows from
Grothendieck’s vanishing theorem that for all p > d, E2p,d+t−p = 0. But then
since from the hypothesis Hd+t
I+J (M) does not vanish, there is at least one
p ≤ d such that
E2p,d+t−p = HpJ (Hd+t−p
(M)) 6= 0.
I
Hence Hd+t−p
(M) 6= 0 and so cd(I, M) ≥ d + t − p ≥ t.
I
If, in particular, cd(I, M) = t, then E2p,q = 0 for all q > t. Now from the
subsequent stages of the spectral sequence
Ekd−k,t+k−1
/
Ekd,t
/
Ekd+k,t−k+1
d,t
and the fact that Ekd−k,t+k−1 = Ekd+k,t−k+1 = 0 for all k ≥ 2, we have E∞
=
d,t
d
d+t
t
∼
E2 . Hence HJ (HI (M)) = HI+J (M).
Since HdJ (HtI (M)) 6= 0, it follows from Grothendieck’s vanishing theorem
that dim Supp(HtI (M)) ≥ d. On the other hand, since dim Supp(HtI (M)) ≤
dim(R/I + Ann M ) = d, we conclude that dim Supp(HtI (M)) = d.
So far, for a finitely generated R- module M, the best known lower bound
for cd(I, M) is htM (I) = ht I(R/AnnM). As an immediate consequence of
Theorem 3.1, we sharpen this bound to dim(M) − dim(M/IM) ≥ htM (I).
Corollary 3.2. Let (R, m) be a Noetherian local ring, M a finitely generated
R-module of dimension n and I an ideal of R such that dim(M/IM) = d.
Then n − d is a lower bound for c = cd(I, M). Moreover, if c = n − d, then
Hdm (HIn−d (M)) ∼
= Hnm (M)
and dim Supp(HIn−d (M)) = d.
Proof. This follows from Theorem 3.1 and the fact that Hmn (M) 6= 0.
For an ideal I of R, it is a well-known fact that ht(I) ≤ cd(I) ≤ ara(I),
where ara(I) denotes the smallest number of elements of R required to generate I up to radical. If, in particular, ara(I) = cd(I) = ht(I), then I is
called a set-theoretic complete intersection ideal. Determining set-theoretic
5
complete intersection ideals is a classical and long-standing problems in commutative algebra and algebraic geometry. Many questions related to an ideal
I to being a set-theoretic complete intersection are still open, see [12] for
more details. Varbaro in [13] show that under certain conditions there exists
ideals I satisfying the property that cd(I) = ht(I), knowing the existence of
ideals with such properties, we have the following:
Corollary 3.3. Let (R, m) be a Noetherian local ring of dimension n and
I an ideal of R with d = dim(R/I) such that cd(I) = ht(I) = h. Then
dim(R) = ht(I) + dim(R/I) and
n−h
Hm
(HhI (R)) ∼
= Hnm (R).
Proof. It follows from Corollary 3.2 that dim(R)−dim(R/I) ≤ cd(I) = ht(I),
while the other side of the inequality always holds. Therefore dim(R) =
ht(I) + dim(R/I). Now the required isomorphim follows from Corollary
3.2.
We end this section with the following conclusion :
Corollary 3.4. Let (R, m) be a non-catenary Noetherian local domain of
dimension n. Then there is at least one prime ideal of R that is not a set
theoretic complete intersection.
Proof. Since R is non-catenary, there is a prime ideal p of R such that ht(p) <
n − dim(R/p). Then it follows from Corollary 3.3 that cd(p) 6= ht(p) and
therefore p can not be a set theoretic complete intersection ideal.
4
Artinianness and Non-Artinianness of Top
Local Cohomology Modules
Let (R, m) be a Noetherian local ring and M an R-module of dimension n.
Recall that if M is a coatomic or a weakly finite (in particular, finitely generated, I-cofinite, or a balanced big Cohen Macaulay) module, then Hnm (M)
is nonzero and Artinian [[2], [3]].
In light of this information, we have the following results the first of which
is the generalization of Theorem 7.1.6 of [4]:
6
Theorem 4.1. Let (R, m) be a Noetherian local ring and M an R-module of
dimension n such that Hnm (M) is Artinian. Then HnI (M) is Artinian for all
ideals I of R.
Proof. We use induction on d = dim(R/I). If d = 0, then I is an m-primary
ideal and so HnI (M) ∼
= Hnm (M) is Artinian.
Let now d > 0 and suppose the hypothesis is true for all ideals J of R with
dim(R/J) < d.
Choose an element x ∈ m such that dim(R/(I + Rx)) = d − 1 < d. Then
by induction hypothesis, HnI+Rx (M) is Artinian. Now consider the exact
sequence
···
/
HnI+Rx (M)
/
HnI (M)
/
HnI (Mx )
/
···
Since HnI+Rx (M) is Artinian and HnI (Mx ) = 0 (dim(Mx ) < n, as x ∈ m), it
follows from the above exact sequence that HnI (M) is Artinian.
Recall that a class S of R-modules is a Serre subcategory of the category
of R-modules, C(R), when it is closed under taking submodules, quotients
and extensions. The main result of this section is the following:
Theorem 4.2. Let R be a Noetherian ring, M an R-module (not necessarily
finitely generated) and let S be a Serre subcategory of C(R). Let I and J be
c−i
two ideals of R such that Ht+i
J (HI (M)) ∈ S for all 0 < i ≤ c = cd(I, M)
and Ht+c
/ S for some positive integer t. Then HtJ (HcI (M)) ∈
/ S.
I+J (M) ∈
Proof. Consider the Grothendieck’s spectral sequence
E2p,q = HpJ (HqI (M)) =⇒ Hp+q
I+J (M)
and look at the stage p + q = c + t. Let now 0 < i ≤ c = cd(I, M). Since
t+i,c−i
E∞
= Ert+i,c−i for sufficiently large r and Ert+i,c−i is a subquotient of
t+i,c−i
t+i,c−i
E2
∈ S, E∞
∈ S for all 0 < i ≤ c = cd(I, M).
On the other hand, since E2t,c = HtJ (HcI (M)) =⇒ Ht+c
I+J (M), there exists a
finite filtration
0 = Φt+c+1 Ht+c ⊆ Φt+c Ht+c ⊆ · · · ⊆ Φ1 Ht+c ⊆ Φ0 Ht+c = Ht+c
p,q
p t+c
of Ht+c = Ht+c
/Φp+1 Ht+c for all p + q = t + c.
I+J (M) such that E∞ = Φ H
p,q
Since for all p < t, E∞
= 0, we have that Φt Ht+c = · · · = Φ1 Ht+c =
7
t+i,c−i
Φ0 Ht+c = Ht+c . But then since E∞
= Φt+i Ht+c /Φt+i+1 Ht+c ∈ S for all
t+1 t+c
0 < i ≤ c, Φ H ∈ S and so it follows from the short exact sequence
0
/ Φt+1 Ht+c
|
{z
∈S
}
/
Ht+c
(M)
| I+J{z }
/
t,c
E∞
/
0
∈S
/
t,c
t,c
t,c
that E∞
∈
/ S. Since E∞
is a subquotient of E2t,c and E∞
∈
/ S, it follows that
E2t,c = HtJ (HcI (M)) ∈
/ S.
Corollary 4.3. Let (R, m) be a Noetherian local ring, M an R-module
of dimension n such that Hnm (M) 6= 0 with dim(R/I + AnnM ) = d. If
cd(I,M )
n−i
HI
(M) is Artinian, then either cd(I, M) = n or Hm
(HiI (M)) 6= 0 for
some n − d ≤ i < cd(I, M).
Proof. We prove the contrapositive of the statement. Let S be the category of
n−i
zero module and suppose that c = cd(I, M) < n and Hm
(HiI (M)) = 0 ∈ S
for all n − d ≤ i < c. But then since Hnm (M) ∈
/ S, it follows from Theorem
n−c
4.2 that Hm
(HcI (M)) 6= 0. Hence dim Supp(HcI (M)) > 0 and so HcI (M) is
not Artinian.
The following results determine the Artinianness and non-Artinianness
cd(I,M )
of the top local cohomology module, HI
(M), for the ideals of small
dimension.
Theorem 4.4. Let (R, m) be a Noetherian local ring, M an R-module of
dimension n such that Hnm (M) is nonzero and Artinian and let I be an ideal
cd(I,M )
of R such that dim(R/I + Ann M ) = 1. Then HI
(M) is Artinian if and
only if cd(I, M) = n.
Proof. Since dim(R/I + Ann M ) = 1 and Hnm (M) 6= 0 , it follows from Theorem 3.1 that either cd(I, M) = n − 1 or cd(I, M) = n. If cd(I, M) = n, then
by Theorem 4.1, HnI (M) is Artinian. If, on the other hand, cd(I, M) = n − 1,
then it follows from Theorem 3.1 that dim Supp(HIn−1 (M)) = 1 and so
HIn−1 (M) is non-Artinian.
Theorem 4.5. Let (R, m) be a Noetherian local ring, M an R-module of
dimension n such that Hnm (M) is nonzero and Artinian and let I be an ideal
cd(I,M )
of R with dim(R/I + Ann(M)) = 2. If HI
(M) is Artinian, then either
cd(I, M) = n, or cd(I, M) = n − 1 and H2m (HIn−2 (M)) 6= 0.
8
Proof. Since dim(R/I + Ann M ) = 2 and Hnm (M) 6= 0, it follows from Theorem 3.1 that n − 2 is a lower bound for cd(I, M). If cd(I, M) = n − 2,
cd(I,M )
then again by Theorem 3.1, dim Supp(HIn−2 (M)) = 2 and so HI
(M)
is non-Artinian. If, on the other hand, cd(I, M) = n, then from Theorem
cd(I,M )
4.1, HI
(M) is Artinian. Finally, if cd(I, M) = n − 1 and HIn−1 (M) is
Artinian, then the result follows from Corollary 4.3.
References
[1] M. Aghapournahr and L. Melkersson, Artinianness of local cohomology modules, Ark.
Mat., 52(2014), 1-10.
[2] M. Aghapournahr and L. Melkersson, Finiteness properties of minimax and coatomic
local cohomology modules, Arch. Math. 94 (2010) 519–528.
[3] A. Bagheri, ‘A Non-Vanishing Theorem for Local Cohomology Modules’, Bull. Malays.
Math. Sci. Soc. (2) 37 (2014) 65–72.
[4] M. P. Brodmann and R. Y. Sharp, Local Cohomology, 2nd ed., Cambridge Studies in
Advanced Mathematics, Vol. 136, Cambridge University Press, Cambridge, 2013. An
algebraic introduction with geometric applications.
[5] M. T. Dibaei and A. Vahidi, Artinian and Non-Artinian Local Cohomology Modules,
Canad. Math. Bull., 54(2011), 619-629.
[6] K. Divaani-Aazar, R. Naghipour and M. Tousi , Cohomological dimension of certain
algebraic varieties , Proc. Amer. Math. Soc, 130(12) (2002), 3537 3544.
[7] A. Grothendieck (notes by R. Hartshorne), ‘Local cohomology’ , Lect. Notes in Math.
41 Springer-Verlag, 1967.
[8] R. Hartshorne, Cohomological dimension of algebraic varieties, Annals of Math.,
88(1968), 403-450.
[9] C. Huneke, Problems in local cohomology, Free Resolutions in Commutative Algebra
and Algebraic Geometry Sundance 90, Res. Notes in Math., Jones and Barlett, 2(1992),
93-108.
[10] C. Huneke and G. Lyubeznik, On the vanishing of local cohomology modules, Inv.
Math., 102(1990), 73-93.
[11] G. Lyubeznik, A partial survey of local cohomology, local cohomology and its applications,Lect. Notes Pure Appl. Math.,226 (2002), 121-154.
[12] G. Lyubeznik , A survey of problems and results on the number of defining equations.
Commutative Algebra, Math. Sci. Res. Inst. Publ., 15 (1989), 375-390.
[13] M. Varbaro, Cohomological and Projective Dimensions , Compositio Mathematica,
149(2013), 1203-1210.
9
| 0math.AC
|
Score matching estimators for directional distributions
Kanti V Mardia, John T Kent, and Arnab K Laha
arXiv:1604.08470v1 [math.ST] 28 Apr 2016
University of Leeds and University of Oxford, University of Leeds,
Indian Institute of Management Ahmedabad
Abstract
One of the major problems for maximum likelihood estimation in the wellestablished directional models is that the normalising constants can be difficult to
evaluate. A new general method of “score matching estimation” is presented here
on a compact oriented Riemannian manifold. Important applications include von
Mises-Fisher, Bingham and joint models on the sphere and related spaces. The estimator is consistent and asymptotically normally distributed under mild regularity
conditions. Further, it is easy to compute as a solution of a linear set of equations
and requires no knowledge of the normalizing constant. Several examples are given,
both analytic and numerical, to demonstrate its good performance.
Some key words: Exponential family, Fisher-Bingham distribution, Riemannian
manifold, sphere, torus, von Mises distribution.
1
Introduction
A novel “score matching estimator” was proposed in Hyvärinen (2005, 2007) as an
alternative to the maximum likelihood estimator. For exponential family models,
a key advantage of this new estimator is that it avoids the need to work with
awkward normalizing constants. A minor limitation is that it requires the underlying
distributions to have sufficiently smooth probability densities.
The score matching estimator was originally developed for densities on Euclidean
space. The extension to Riemannian manifolds was sketched in Dawid and Lauritzen
(2005) and Parry et al. (2012), but without a detailed analysis. Here we give a
systematic investigation on a compact oriented Riemannian manifold M . The main
focus is on exponential family models. Key applications include the Fisher-Bingham
and multivariate von Mises distributions, which lie on spheres and related spaces.
1
2
Background on Riemannian manifolds
A p-dimensional Riemannian manifold M , say, is characterized by a metric tensor
G = G(x) = (gij (x)), where G(x) is a p × p symmetric positive definite matrix.
Here x is a p-dimensional vector representing a typical element of M in local coordinates. Let G−1 (x) = (g ij (x)) denote the inverse matrix. A uniform measure on
M can be defined in local coordinates by µ(dx) = {det(G(x))}1/2 dx. See, e.g., Jost
(2005); Rosenberg (1997); Stein and Weiss (1971) for a background on Riemannian
manifolds.
Next, let u = u(x), v = v(x), x ∈ M , be two real-valued functions on M . Let
∇u = (∂u/∂xj , j = 1, . . . , p)T denote the gradient in local coordinates, i.e. the
column vector of partial derivatives. An inner product on gradient vectors can be
defined by
hu, vi = hu, vi(x) = (∇u)T G−1 (∇v),
(1)
treated as a function of the local coordinate vector x.
The Laplace-Beltrami operator, acting on u, is defined in local coordinates by
∆M u =
p
X
h
i
{det(G(x))}−1/2 ∂/∂xi {det(G(x))}1/2 g ij (x)∂u/∂xj .
(2)
i,j=1
treated as a function of the local coordinate vector x. Although the gradient vector
∇u depends on the choice of local coordinates, the uniform measure µ(dx), the
gradient inner product (1) and the Laplace-Beltrami operator (2) are invariant under
a change of local coordinates.
Stokes’ Theorem, also known as the divergence theorem, connects the gradient
inner product and the Laplace-Beltrami operator. Further, if M is assumed compact
and oriented, there is no need for boundary conditions.
Theorem 1 (Stokes’ Theorem). If M is a compact oriented Riemannian manifold
and u(x), v(x) are twice continuously differentiable functions on M , then
Z
Z
hu, vi(x) µ(dx) = −
(∆M u)v µ(dx).
(3)
M
M
If M is isometrically embedded as a p-dimensional surface in a Euclidean space,
M ⊂ Rq , then a point in M can be represented either as a p-dimensional vector x
in local coordinates or as a q-dimensional vector z = (z1 , . . . , zq )T , say, in Euclidean
coordinates. In this setting it is possible to give simpler representations of the
gradient inner product and the Laplace-Beltrami operators.
Given a point z in M , the tangent plane to M at z is a p-dimensional hyperplane.
Let P (z) = P = (pij ) denote the q × q orthogonal projection matrix onto this
hyperplane. Then P = P T , P = P 2 and P has rank p. Suppose the functions u
and v have been extended to a neighbourhood N of M as functions ũ(z), ṽ(z) of
z ∈ N ⊂ Rq , and define the usual Euclidean gradient
∇E ũ = (∂ ũ/∂zj , j = 1, . . . , q)T .
2
The gradient inner product and the Laplace-Beltrami operator were given by (1)
and (2) in local coordinates. They can also be written in Euclidean coordinates as
hu, vi = (∇E ũ)T P (∇E ṽ) = (P ∇E ũ)T (P ∇E ṽ)
and
(4)
q
X
∆M u = tr P ∇TE (P ∇E ũ) =
pij ∂/∂zi (pjk ∂ ũ/∂zk ) .
i,j,k=1
Note the derivatives in (4) are q-dimensional whereas the derivatives in (1) are
p-dimensional; however, the inner product is the same.
The simplest example is the sphere S2 = {z ∈ R3 : z T z = 1}. In polar coordinates, x = (θ, φ)T with θ = colatitude and φ = longitude, the metric tensor
becomes
"
#
1
0
G=
,
0 sin2 θ
and the uniform measure becomes sin θdθdφ. The Laplace-Beltrami operator becomes
∆M u = ∂ 2 u/∂θ2 + cot θ∂u/∂θ + (sin θ)−2 ∂ 2 u/∂φ2 .
The Euclidean embedding takes the form z1 = cos θ, z2 = sin θ cos φ, z3 = sin θ sin φ,
and the projection matrix is P = I3 − zz T . It is straightforward to check that gradient inner products in (1) and (4) are two ways of writing the same function.
3 The score matching criterion on a compact
oriented Riemannian manifold
One way to motivate a statistical estimator is through the minimization of a divergence between two probability distributions. The most common example is the
Kullback-Liebler divergence, which leads to the maximum likelihood estimator. In
this paper we use a divergence due to Hyvärinen, which leads to the score matching
estimator.
Let f and f ∗ be two probability densities on a Riemannian manifold M , defined
with respect to the uniform measure µ, where f and f ∗ are assumed to be everywhere
nonzero and twice continuously differentiable. Define the Hyvärinen divergence
(Hyvärinen, 2005, 2007) between the two densities in terms of an integrated gradient
inner product for the log ratio,
Z
1
∗
Φ(f ; f ) =
hlog(f /f ∗ ), log(f /f ∗ )i f ∗ (x)µ(dx)
2 M
Z
1
=
{hlog f, log f i − 2hlog f, log f ∗ i + hlog f ∗ , log f ∗ i} f ∗ (x) µ(dx).
2 M
(5)
3
Hyvärinen proposed this divergence in the setting where M is a Euclidean space.
In this setting the metric tensor G = I is hardly needed, and the the gradient inner
P
product (1) simplifies to hu, vi = pj=1 (∂u/∂xj )(∂v/∂xj ).
For the rest of the paper we limit attention to the setting where M is compact
and oriented. In some ways this setting is more complicated than the Euclidean
case. The gradient inner product h·i is now given by (1); it can also be written as
(4) when the manifold M can be embedded in a Euclidean space. However, in other
ways this setting is simpler. Stokes’ Theorem (3) holds automatically without the
need for boundary conditions and all continuous functions on M are automatically
bounded and integrable.
Two simple properties ensure that minimizing (5) over f is an identifiable approach to finding f ∗ .
Property 1. If f = f ∗ , then Φ(f ; f ∗ ) = 0.
Property 2. If f 6= f ∗ , then Φ(f ; f ∗ ) > 0.
To prove these properties note by inspection in (5) that Φ(f ; f ∗ ) ≥ 0 for all
choices of densities and that Φ(f ∗ ; f ∗ ) = 0. Conversely, if f 6= f ∗ , then we claim
there must be at least one point x in some set of local coordinates such that ∇f (x) 6=
∇f ∗ (x); hence the integral (5) must be strictly positive. For if there were no such
x, a contradiction would arise: the gradients would be everywhere equal and so the
densities would be the same up to a proportionality constant; further, since both
densities integrate to 1, the proportionality constant would have to equal 1.
For fitting purposes, let f ∗ (x) be regarded as the “true” distribution of the data
and let f (x) = f (x; π) denote a parametric model, where π is an m-dimensional
vector of parameters. Then a “best-fitting” model can be defined by minimizing
Φ(f ; f ∗ ) over the parameters π.
Since the final term in (5) does not depend on f , it can be dropped from the
minimization criterion. Further, by Stokes’ Theorem (3) and the simple result,
∂ log f ∗ (x)/∂xj = {f ∗ (x)}−1 ∂f ∗ (x)/∂xj , the middle term becomes
Z
Z
∗ ∗
−
hlog f, log f if (x) µ(dx) = −
hlog f, f ∗ i(f ∗ )−1 f ∗ (x) µ(dx)
M
M
Z
=−
hlog f, f ∗ i µ(dx)
M
Z
=
(∆M log f )f ∗ (x) µ(dx).
M
The final term in (5) does not depend on f . Hence minimizing Φ(f ; f ∗ ) over the
parameters π in f is equivalent to minimizing
Z
1
∗
Ψ(f ; f ) =
{hlog f, log f i + 2(∆M log f )} f ∗ (x)µ(dx)
2 M
1
= E {hlog f, log f i + 2(∆M log f )} ,
(6)
2
4
where E denotes expectation under the measure f ∗ (x)µ(dx). At this stage it is no
longer necessary to impose any regularity conditions on f ∗ ; indeed it can be replaced
by any probability measure F ∗ , say, in which case we write Ψ(f, F ∗ ).
For the rest of the paper we specialize to the case where f (x; π) forms a canonical
exponential family on a compact oriented Riemannian manifold M with density
f (x) = f (x; π) ∝ exp{π T t(x)}
(7)
with respect to the uniform measure µ(dx), where π is an m-vector of natural
parameters and t(x) = (t1 (x), . . . , tm (x) is a vector of sufficient statistics. The
sufficient statistics are assumed to satisfy the following regularity conditions.
A1 The constant function 1 and the functions t` (x) (` = 1, . . . , m) are linearly
independent on M .
A2 The functions t` (x) (` = 1, . . . , m) are twice continuously differentiable with
respect to x.
The first assumption ensures the identifiability of π; different values of π correspond to different distributions. The second assumption justifies the use of Stokes’
Theorem.
In this exponential family setting (6) simplifies to
1
Ψ(f ; F ∗ ) = π T W π − π T d,
2
where W = (w`1 `2 ) and d = (d` ) have elements
w`1 `2 = E{ht`1 , t`2 i(x)},
d` = −E{∆M t` (x)} (`, `1 , `2 = 1, . . . , m).
(8)
(9)
Minimizing (8) yields the moment equation W π − d = 0. In particular, if the
“true” distribution F ∗ (dx) = f (x; π0 ) lies in the parametric family, it follows that
π0 can be recovered from the moments in W and d by
π0 = W −1 d.
The identifiability property, Property 2 mentioned below (5), implies that W must
be nonsingular. In later sections we illustrate how to calculate W and d in particular
cases.
In many examples the components of t(x) are eigenfunctions of the LaplaceBeltrami operator, so that ∆M t` (x) = −λ` t` (x), for suitable constants λ` > 0 (` =
1, . . . , m) See, e.g., Patrangenaru and Ellingson (2016, Ch 3.3) for the eigenfunctions of various manifolds arising in Statistics; the specific example of the sphere is
discussed below.
4 The score matching estimator and its properties
Let F ∗ = Fn denote the empirical distribution for a set of data {xh , h = 1, . . . , n}.
The value π̂ minimizing Ψ(f ; Fn ) is called the score matching estimator. More
5
explicitly,
π̂SME = Wn−1 dn ,
where Wn and dn are obtained from (9) after replacing the expectation under F ∗
by a sample average over the n data points.
To discuss the asymptotic sampling properties of the score matching estimator,
suppose the data comprise a random sample from the exponential family model
f (x; π) with π = π0 . Since W is nonsingular in the population case, it follows from
the strong law of large numbers, that Wn must be positive definite with probability
1 for sufficiently large n. This point has not generally been emphasized in the
literature.
Further, in many models this nonsingularity statement about Wn can be strengthened to conclude that there is a fixed value n0 , depending on the model but not on
the data, such that Wn must be positive definite with probability 1 for n ≥ n0 . An
analogous result for the p-dimensional multivariate normal distribution states that
the sample covariance matrix is guaranteed to be positive definite with probability
1 for n ≥ p + 1.
The limiting behaviour of π̂ is straightforward to describe. By the central limit
theorem, dn and Wn are asymptotically jointly normally distributed with population
means d and W . Hence by the delta method, it follows that
n1/2 {π̂SME − π0 } ∼ Nm (0, Ω)
for some limiting covariance matrix Ω. Further, by the asymptotic optimality of
maximum likelihood, Ω ≥ I −1 under the usual ordering for positive semi-definite
matrices, where I denotes the Fisher information matrix.
The discussion here is limited to the exponential family case. Forbes and Lauritzen (2014) note that extra regularity conditions are needed for consistency and
asymptotic normality when looking at score matching estimation for more general
densities.
The term “score” has several distinct connotations in estimation. (a) Conventionally, the “score” refers to a derivative of the log likelihood with respect to
the parameters; it has close connections to maximum likelihood estimation. (b)
However, in the context of the score matching estimator, the “score” refers to the
derivative of the log likelihood with respect to the state variable x. (c) In addition,
the term “scoring rule” (e.g., Forbes and Lauritzen, 2014; Parry et al., 2012) refers
to a more general function of x and a distribution. Each scoring rule determines a
divergence, and the minimization of the divergence leads to an estimator. A scoring
rule is different from the scores in (a) and (b), though suitable choices for scoring
rules lead to both the maximum likelihood and score matching estimators.
5
Details for the sphere
For this paper the most important choice for the manifold M is the unit sphere
P 2
Sp = {z ∈ Rq :
zj = 1}, a p-dimensional manifold embedded in Rq , q = p + 1.
6
There are two natural coordinate systems: embedded Euclidean coordinates z and
local coordinates x, i.e. polar coordinates in this case.
First we set out the key steps for the derivative calculations. Let ũ(z) denote a
scalar-valued function in Euclidean coordinates. The projected Euclidean gradient
vector becomes
P ∇E ũ = (I − zz T )∇E ũ, P = Iq − zz T .
The eigenfunctions of ∆M on Sp , p ≥ 1, are known as the spherical harmonics.
A spherical harmonic of degree k ≥ 0 has eigenvalue −λk where λk = k(k + p − 1) =
k(k + q − 2); see, e.g., Chavel (1984, p. 35), Patrangenaru and Ellingson (2016, p.
125). The action of ∆M on the linear and quadratic spherical harmonics, expressed
in Euclidean coordinates, can be summarized as follows,
∆M zj = −λ1 zj ,
∆M (zi2 − zj2 ) = −λ2 (zi2 − zj2 ),
(10)
∆M (zi zj ) = −λ2 (zi zj ),
where i 6= j ∈ {1, . . . , q}. A systematic construction of higher order spherical
harmonics is given in Stein and Weiss (1971, pp. 137–152), but they will not be
needed here. Statistical models involving spherical harmonics of degree greater
than two are straightforward in principle (e.g. Beran, 1979), but in practice the
components of t(z) will usually consist of linear and quadratic functions of z.
To illustrate these calculations, consider the Fisher-Bingham density on Sp ,
f (z) ∝ exp bT z + z T Az ,
(11)
where b is a q-vector, and A is a q × q symmetric matrix, q = p + 1. To ensure
P
identifiability, the side condition qj=1 ajj = 0 is imposed.
The density can be recast as
2
f (z) ∝ exp{b1 z1 + · · · + bq zq + a11 (z12 − zq2 ) + · · · + aq−1,q−1 (zq−1
− zq2 )+
a12 (2z1 z2 ) + · · · + a1q (2z1 zq ) + · · · + aq−1,q (2zq−1 zq )}
(m
)
X
= exp
π` t` (z) .
`=1
Here the vector π denotes the m = q + q − 1 + q(q − 1)/2 parameters in b and
A in the order listed, and the vector t = t(z) denotes the corresponding functions
of z; that is, the linear terms, the diagonal quadratic terms, and the cross-product
quadratic terms, respectively.
Next we gather the information needed to evaluate W and d in (9). For each
sufficient statistic t` = t` (z), create a vector-valued function u` = u` (z) = ∇E t` (z)
by taking its Euclidean gradient, and create a scalar-valued function v` = v` (z) =
z T u` . Then create an m × m matrix W with entries
w`1 `2 = E{uT`1 u`2 − v`1 v`2 }.
7
(12)
Table 1: Gradient details for the Fisher-Bingham density (11) on Sq . Here t` (z) and v` (z)
are scalars; u` (z) is a q-dimensional vector. Also ej represents a unit vector along the jth
coordinate axis (j = 1, . . . , q).
t` (z)
zj
2
zj − zp2
2zi zj
u` (z)
ej
2(zj ej − zq eq )
2(zi ej + zj ei )
v` (z)
zj
2
2(zj − zq2 )
4zi zj .
Table 1 gives the entries for u` and v` . Also create a vector d with entries
d` = −E{∆M t` (z)},
(13)
using equation (10) to evaluate the Laplace-Beltrami operator.
6
Hybrid estimators and reduced models
The parameters for directional distributions can typically be split into two parts:
the concentration parameters and the orientation parameters, where the normalizing
constant depends just on the concentration parameters. Further it is often possible
to estimate the orientation parameters explicitly using sample moments. Often
this orientation estimator can be viewed as an exact or an approximate maximum
likelihood estimator, and it can be computed without needing estimates of the
concentration parameters.
Further, if the orientation parameters are known, then the distribution of the
data becomes a natural exponential family for the concentration parameters. We
call this latter model reduced because the number of concentration parameters is
smaller than the number of original parameters.
Hence, the following hybrid strategy provides a tractable estimation procedure:
(a) Split the parameters for the full model into orientation and concentration
parameters.
(b) Estimate the orientation parameters for the original data and the transform
the data to a standardized form.
(c) After standardization, the estimated orientation parameters take a simple
canonical form. Treating them as known for the standardized data, the concentration parameters become natural parameters in a reduced exponential
family.
(d) Use the score matching estimator to estimate the concentration parameters of
the reduced model for the standardized data.
8
Table 2: Log densities for various standard directional distributions in full and reduced
form. In full form, the models include both orientation and concentration parameters.
In reduced form there are just concentration parameters. The models named in square
brackets are too general to have a useful reduced form, but the submodels listed below
them do have a useful reduced form. The first four models lie on the sphere Sp . The final
two models are multivariate (MV) von Mises models lying on the torus (S1 )k .
Name
Full
Reduced
von Mises-Fisher
µT z
Bingham
z T Az
[Fisher-Bingham]
µT z + z T Az
—
Kent
Aµ = 0; λ1 = 0, λ2 = −λ3 = β
κz1 + β(z22 − z32 )
[ MV von Mises]
k
X
r=1
MV von Mises sine
z T Λz =
X
µ(r)T z (r) +
z (r)T Ω(r,s) z (s)
κz1
Pq−1
j=1
λj (zj2 − zp2 )
—
r<s
P
P
k
(rs) 0 0
(r) 0
sr ss ,
+
κ
c
r
r<s λ
r=1
c0r = cos θ(r)0 , s0r = sin θ(r)0 ,
θ(r)0 = θ(r) − θ(r)
(r)
Set θ0 = 0.
0
The resulting estimator can be called the hybrid score matching estimator and
denoted by π̂SME,hybrid .
The next section gives several examples from directional statistics to illustrate
this estimation strategy. Table 2 summarizes the form of each density. See, e.g.
Jammalamadaka and Sengupta (2001); Mardia and Jupp (2000) for background
information on these distributions. The first three examples lie on the sphere M =
Sp ; the last example lies on the torus, a direct product of k circles, M = (S1 )k . In
each case the details (a)–(d) are specified explicitly.
In all cases we assume a sample of size n from the stated distribution. On the
sphere the data are represented by an n × q matrix Z whose rows zhT (h = 1, . . . , n),
say, are q-dimensional unit vectors in Euclidean coordinates. The models are special
cases of the Fisher-Bingham density (11) and Table 1 gives the details needed for the
gradient calculations. On the torus, it is more convenient to use polar coordinates
to represent the data as an n × k matrix Θ of angles lying in [0, 2π), with gradient
calculations carried out directly. In each case the key step is to derive the formulas
for the matrix Wn and the vector dn in the reduced model.
9
7
7.1
Directional distributions
von Mises-Fisher distribution
(a) The density for the full von Mises-Fisher distribution takes the form in (7) with
t(z) = (z1 , . . . , zq )T being the vector of linear functions in z. This model forms
a canonical exponential family. The parameter vector can be written π = κµ0
where κ ≥ 0 is a scalar concentration parameter and µ0 is a unit orientation
vector. On the circle, it is sometimes convenient to write µ0 = (cos θ0 , sin θ0 )T
in polar coordinates.
(b) For the data matrix Z(n×q), the sufficient statistic is the q-dimensional sample
mean vector z. The maximum likelihood estimate of µ0 is the unit vector
µ̂0,MLE = z/||z|| and is also the hybrid score matching estimator. Let R be a
q × q orthogonal matrix such that RT µ̂0,MLE = e1 , where e1 is a unit vector
along the first coordinate axis in Rq , and let Y = ZR, i.e. yh = RT zh (h =
1, . . . , n), denote the standardized data.
(c) As shown in Table 2, the reduced model for Y involves just a single concentration parameter κ.
(d) For the reduced model, Wn and dn in (12)–(13) are one-dimensional,
Wn = 1 −
1X 2
yh1 ,
n
dn = (q − 1)
X
yh1 ,
so that the hybrid score matching estimator of κ becomes
κ̂SME,hybrid = dn /Wn ,
(14)
expressed in terms of the first two sample moments of the standardized data.
All sums here and below range over h = 1, . . . , n.
The score matching estimator for the von Mises distribution can also be derived
from the two trigonometric moments,
E(cos νθ) = Iν (κ)/I0 (κ),
ν ≥ 0,
for ν = 1, 2. Using the Bessel function identity
Iν+1 (κ) = Iν−1 (κ) −
2ν
Iν (κ)
κ
(15)
with ν = 1 and simplifying yields the population version of (14) for the circle,
q = 2. Analogous results using Legendre polynomials yield the population
version of (14) for larger values of q.
If q = 2 and the data are represented in polar coordinates, (zh1 , zh2 ) =
(cos θh , sin θh )(h = 1, . . . , n), then the score matching estimators can be recast in polar coordinates. In particular, the estimated orientation angle for
the hybrid score matching estimator is
10
θ̂0,SME,hybrid = θ̂0,MLE = atan2(S̄, C̄) = θ̂0 , say,
(16)
where atan2(.) is defined so that atan2(y, x) = θ if and only if (cos θ, sin θ)T ∝
(x, y)T for x2 + y 2 > 0, and where
C̄ =
1X
cos θh ,
n
S̄ =
1X
sin θh ,
n
R̄ =
p
S̄ 2 + C̄ 2 .
The hybrid score matching estimator of κ can be re-expressed as
P
cos(θh − θ̂0 )
nR̄
κ̂SME,hybrid = P 2
=P 2
.
sin (θh − θ̂0 )
sin (θh − θ̂0 )
(17)
The full score matching estimator on the circle is also straightforward to derive,
with m = 2 and sufficient statistics t1 (θ) = cos θ and t2 (θ) = sin θ. Set
P
P
C̄2 = n1
cos 2θh , S̄2 = n1
sin 2θh , R̄2 = (C̄22 + S̄22 )1/2 . Then Wn and dn
have elements
1
(n)
w11 = (1 − C̄2 ),
2
(n)
1
(n)
w12 = − S̄2 ,
2
1
(n)
w22 = (1 + C̄2 ),
2
(n)
and d1 = C̄, d2 = S̄. Since |Wn | = (1 − R̄22 )/4, the full score matching
estimator becomes
θ̂0,SME = atan2{C̄ S¯2 + S̄(1 − C¯2 ), C̄(1 + C¯2 ) + S̄ S¯2 }
2
2
κ̂SME = 2{R̄2 (1 + R̄2 ) + 2(C̄ 2 − S̄ 2 )C¯2 + 4C̄ S̄ S¯2 }1/2 /(1 − R̄2 ).
After rotating the data so that S̄ = 0, the full score matching estimate of κ
turns out to be the same as the hybrid estimate if S̄2 = 0.
7.2
Bingham distribution
(a) The parameter matrix A for the Bingham distribution in Table 2 is a symmetric
q × q matrix, where without loss of generality, the trace of A may be taken to
be 0. If A = ΓΛΓT is the spectral decomposition of A, then the orthogonal
matrix Γ = γ(1) , . . . , γ(q) , whose columns are eigenvectors of A, represents
the orientation parameters and the eigenvalues Λ = diag(λj ) represent the
P
concentration parameters, with qj=1 λj = tr(A) = 0.
(b) Given the n × q original data matrix Z calculate the moment of inertia matrix
T (Z) = (1/n)Z T Z and find its spectral decomposition T (Z) = GLGT . Then
the maximum likelihood estimate of Γ is G. Define the standardized data
matrix by Y = ZG.
(c) In the reduced model, the matrix A simplifies to the diagonal matrix Λ, with
P
Pq
2
2
y T Λy = q−1
j=1 λj (yj − yq ) since
j=1 λj = 0.
11
(d) For the reduced model π becomes the parameters λj (j = 1, . . . , q − 1) and the
estimates Wn and dn have entries for i, j = 1, . . . , q − 1,
( P
4
2 − (y 2 − y 2 )2 },
{y 2 + yhq
i=j
(n)
hi
hq
wij = n P hi
4
2 − (y 2 − y 2 )(y 2 − y 2 )},
{yhq
i 6= j
hi
hq
hj
hq
n
and
(n)
di
7.3
=
2q X 2
2
(yhi − yhq
).
n
Kent distribution
The Fisher-Bingham distribution in Table 2 has q 2 +3q −2 parameters which can be
estimated by the score matching estimator. However, this distribution has too many
parameters to be of much interest in practice. Instead it is more useful to consider
sub-families of this distribution. One such sub-family for S2 is the 5-parameter FB5
distribution, also known as the Kent distribution, which forms a curved exponential
family.
(a) For the Kent distribution, µ = κγ(1) is assumed to be an eigenvector of
A = ΓΛΓT with eigenvalue λ1 = 0; the other two eigenvalues are assumed
to be of equal size with opposite signs, λ2 = −λ3 = β. The orthogonal matrix Γ contains the orientation parameters and there are two concentration
parameters, κ ≥ 0 and β ≥ 0.
(b) Kent (1982) describes a moment estimator Γ̂ for the orientation matrix, which
will also be used for the hybrid score matching estimator. After standardizing
to Y = Z Γ̂, the sample mean vector becomes y = (R, 0, 0)T and the moment
of inertia matrix T (Y ) = (1/n)Y T Y satisfies (T (Y ) )23 = (T (Y ) )32 = 0 and
(T (Y ) )22 − (T (Y ) )33 ≥ 0.
(c) The reduced form of the distribution is given in Table 2 with π = (κ, β)T
having two components.
(d) The estimates Wn and dn have entries
X
2X
(n)
(n)
(n)
2
2
2
w11 = 1 −
yh1
, w12 = w21 =
yh1 (yh3
− yh2
),
n
4X 2
(n)
2 2
2
2
) }
− (yh2
− yh3
w22 =
{yh2 + yh3
n
and
2X
6X 2
(n)
(n)
2
d1 =
yh1 , d2 =
(yh2 − yh3
).
n
n
7.4
Multivariate von Mises sine model on the torus
Q
A general model on a product manifold M = kr=1 M (r) = M (1) ⊗ · · · ⊗ M (k) ,
involving first order interaction terms, takes the form
f (x; π) ∝ exp{
k
X
π (r)T t(r) (x(r) ) +
X
r<s
r=1
12
t(r)T (x(r) )Ω(r,s) t(s) (x(s) )}
(Jupp and Mardia, 1980). Different sufficient statistics, of possibly different dimensions, are allowed each manifold M (r) . In principle it is also possible to include
higher order interactions, and in some applications it may be desirable to consider
submodels by setting some of the parameters to 0.
An important example of this construction is the general multivariate von Mises
(r) (r)
distribution on the torus M = (S1 )k , for which t(r) (z (r) ) = (z1 , z2 )T =
(cos θ(r) , sin θ(r) )T comprises the two Euclidean coordinates on each circle. See e.g.
(Mardia, 1975; Mardia and Patrangenaru, 2005) for the torus case. Kume et al.
(2013) give an extension to a product of higher dimensional spheres, q > 2.
Even on the torus, this model has too many interaction parameters to be easily
interpretable, so simplifications are often considered, including a sine model and
two versions of a cosine model (e.g., Mardia, 2013, p. 501). For this paper we
limit attention to the sine model on the torus; the cosine versions can be analyzed
similarly.
(r)
(a) The density for the sine model takes the form in Table 2 where θ(r)0 = θ(r) −θ0
(r)
(Mardia et al., 2008; Singh et al., 2002). Here θ0 (r = 1, . . . , k), denote a
set of orientation or centering parameters. The sine model forms a curved
exponential family.
(b) The centering parameters can be estimated marginally by the sample mean
(r)
(r)
(r)
(r)
directions θ̂0 (r = 1, . . . , k) on each circle separately. Let φh = θh − θ̂0
denote the standardized angles for the n data points arranged as an n × k
matrix Φ. In this example, it is simpler to work in polar coordinates than in
Euclidean coordinates.
(c) The reduced model takes the form
f (φ) = exp{
k
X
κ(r) cos φ(r) +
X
λ(rs) sin φ(r) sin φ(s) }
(18)
r<s
r=1
and forms a canonical exponential family with m = k + k(k − 1)/2 parameters.
The m-dimensional sufficient statistic, denoted t(φ), say, can be split into two
blocks, cos φ(r) (r = 1, . . . , k) and sin φ(r) sin φ(s) (r < s).
(d) Working in polar coordinates, the m × k matrix of partial derivatives ∇T t(φ)
can be similarly be partitioned into two blocks, where the nonzero elements
are
(∇ cos φ(r) )j = − sin φ(r) , j = r
(
n
o
cos φ(r) sin φ(s) , j = r,
(r)
(s)
∇(sin φ sin φ ) =
j
sin φ(r) cos φ(s) , j = s,
for j, r, s = 1, . . . , k, r < s. Then
Wn =
1X T
∇ t(φh ){∇T t(φh )}T ,
n
13
where φh = (φh1 , . . . , φhk )T denotes the k-vector of angles for row h of the
standardized data matrix.
The functions cos φ(r) are eigenfunctions of the Laplace-Beltrami operator with
eigenvalue -1. Similarly, the product functions sin φ(r) sin φ(s) are eigenfunctions with eigenvalue -2. Hence the elements of dn have entries
1X
(r)
cos φh
(r = 1, . . . , k),
n
2X
(r)
(s)
=
sin φh sin φh
(r < s).
n
d(n)
r =
0
1
2
^
κ
3
4
d(n)
rs
0.0
0.2
0.4
0.6
0.8
R
Figure 1: Comparison of two estimators of κ for the von Mises distribution with n = 2:
maximum likelihood estimator (solid line) and score matching estimator (dashed line).
8 Efficiency study for the von Mises distribution
For the von Mises distribution on the circle, it is possible to study the behaviour of
the score matching estimator in more detail, both empirically and analytically. The
score matching estimators were described in (16) and (17). The maximum likelihood
estimator of θ0 is the same as hybrid score matching estimator. The maximum
likelihood estimator of κ is κ̂M LE = A−1
1 (R̄), where Aν (κ) = Iν (κ)/I0 (κ), ν ≥ 0.
14
1.00
0.90
0.80
ARE
0
2
4
6
8
10
κ
Figure 2: Asymptotic relative efficiency (ARE), as a ratio of mean squared errors, for the
estimation of κ for the von Mises distribution, comparing the score matching estimator
to the maximum likelihood estimator.
Table 3 gives the relative efficiency, as a ratio of mean squared errors, comparing the score matching estimator of log κ to the maximum likelihood estimator.
Each entry is based on 100,000 simulated datasets. The log transformation is used to
improve the numerical stability of the results, though in the final column for asymptotic relative efficiency, the use of a transformation makes no difference. The simulation from von Mises distribution was carried out in R (R Core Team, 2013) using
the package CircStat (Jammalamadaka and Sengupta, 2001; Lund and Agostinelli,
2012). For all cases, the relative efficiency is at least 78% and is generally much
closer to 100%.
The case n = 2 is interesting because it is possible to write the estimators in
closed form and so compare their behaviour in detail. Further, the hybrid and
full score matching estimators are identical in this setting. After centering, the
standardized data take the form ±θ for a single value of θ, 0 < θ < π/2, where for
simplicity we exclude the extreme possibilities θ = 0, π/2. Then
κ̂SME = cos θ/ sin2 θ,
κ̂MLE = A−1
1 (cos θ).
(19)
The two estimators are compared in Figure 1 as a plot of the estimated κ vs.
R = arccos θ. Although (19) does not provide enough information to compute
the relative efficiency, even in the extreme setting n = 2 the two estimators are
15
Table 3: Relative efficiency, as a ratio of mean squared errors, comparing the score matching estimator of log κ to the maximum likelihood estimator. The final column ARE gives
the asymptotic relative efficiency as n → ∞, taken from (20).
κ n=2 n = 10 n = 20 n = 100 ARE
0.5 100
89
93
95
95
1
98
88
86
85
85
2
98
92
85
79
78
10 100
100
99
99
99
Table 4: Great Whin Sill data on the sphere S2 , with sample size n = 34, and fitted by
the Kent distribution. Estimates of κ and β are given by four different methods described
in the text.
Method
κ̂
β̂
Hybrid SME
42.13 9.34
Hybrid MLE
42.16 9.27
Hybrid approximate MLE 41.76 8.37
Full MLE
42.41 9.28
reasonably similar, with the difference tending to 0 as R → 0 and with the relative difference tending to 0 as R → 1. The maximum difference between the two
estimators is about κ̂MLE − κ̂SME = 2.46 − 1.80 = 0.66 when R = 0.76, i.e. θ = 41o .
In the limiting case n → ∞, it is possible to compute analytically the asymptotic relative efficiency of the hybrid score matching estimator for κ, relative to the
maximum likelihood estimator,
ARE =
A2 (κ)
1
.
{2κ − 3A1 (κ)} κ − κA21 (κ) − A1 (κ)
(20)
A plot of (20) is given in Figure 2 and a proof is given in the Appendix.
9
9.1
Numerical examples
Kent distribution
The Great Whin Sill dataset was analyzed in Kent (1982) and is presented here
to illustrate various estimates of the concentration parameters for the Kent distribution; see Table 4. The first three methods of estimation are hybrid estimators.
Hence, as discussed in Section 7.3, the moment estimator is used for the orthogonal matrix representing the orientation parameters. After rotation of the data, the
reduced model involves just the two concentration parameters.
16
Table 5: Isoleucine data, component 1, on the torus (S1 )4 , with sample size n = 23,
and fitted using the multivariate von Mises sine model. Estimates of the concentration
parameters are given by three different methods: the hybrid score matching estimator
(SME), the hybrid composite likelihood estimator and the hybrid approximate maximum
likelihood estimator (AMLE).
SME
4.3 -6.3 3.4 2.5
* 45.4 24.5 6.03
*
* 59.3 -4.6
*
*
* 2.2
composite
4.9 -7.3 4.0 3.4
* 48.4 26.3 7.1
*
* 61.6 -5.6
*
*
* 2.5
6.6
*
*
*
AMLE
-5.3 3.1
47.1 24.4
* 60.1
*
*
3.1
6.0
-4.3
3.6
Here are further details about the estimators in Table 4. The hybrid score
matching estimator for κ and β was described in Section 7.3. The “hybrid maximum likelihood estimate” is the maximum likelihood estimate for the concentration
parameters after rotating the data using the moment estimate of orientation (eqn.
(4.8) in Kent (1982)). The “hybrid approximate maximum likelihood estimate” is
the same, but using a normal approximation for the normalizing constant (eqn. (4.9)
in Kent (1982)). The “full maximum likelihood estimate” involves maximizing the
5-parameter likelihood over both the orientation and concentration parameters; the
estimated orientation is negligibly different from the moment estimator and is not
reported here. All the estimates of the concentration parameters are close together.
9.2
Torus
In Mardia et al. (2012), k = 4 angles from the amino acid isoleucine were modelled
by a mixture of multivariate von Mises sine models and grouped into 17 clusters.
Here we look at just one of those clusters, Cluster 1, and look at the fits to the
concentration parameters from three estimation methods. All the methods are
hybrid methods. Thus in each case location is estimated using the moment estimator
given by the sample mean directions for each of the k = 4 angles. After rotation
the reduced model involves just the k(k + 1)/2 = 10 concentration parameters.
The hybrid score matching estimator was described in Section 7.4. The hybrid
approximate maximum likelihood estimator uses a high concentration normal approximation for the normalizing constant in the reduced model; the numerical values
were given in Mardia et al. (2012). The hybrid composite likelihood estimator is
based on the conditional von Mises distribution in the reduced model for each angle
given the remaining angles; the methodology is summarized in Mardia et al. (2009).
The estimated parameters are given in Table 5. The estimates are presented as a
symmetric matrix: the diagonal elements are the κ(r) and the off-diagonal elements
are the λ(rs) in (18), given for clarity just in the upper triangle. In general all the
17
estimates match reasonably closely.
10
Discussion
Methods of estimation for exponential families on manifolds can be divided into at
least three broad categories:
• Maximum likelihood estimators, both the exact version and approximate versions. Although the exact version is preferred in principle, there may be problems in practice evaluating the normalizing constant. Hence approximations
may be used, such as (a) saddlepoint (Kume et al., 2013), (b) holonomic (Sei
and Kume, 2015), and (c) approximate normality under high concentration.
• Composite maximum likelihood estimation. Suppose a point on the manifold
can represented as a set of variables, such that the conditional distribution of
each variable given the rest is tractable. Then the composite likelihood is the
product of the conditional densities. In some cases this method can be very
efficient (Mardia et al., 2009).
• Score matching estimators. As shown in this paper these estimators often
reduce to the solution to a set of linear equations based on sample moments
of the data. Hence the method is easy to implement and straightforward to
apply to large datasets, including streaming data.
Strictly speaking, an approximate maximum likelihood estimator, at a fixed level
of approximation, will not be consistent as the sample size n → ∞. The score
matching estimator is always consistent under the mild regularity conditions (A1)–
(A2) on f in Section 3, with an asymptotic variance at least as large as the maximum
likelihood estimator. The the numerical examples here suggest the efficiency of the
hybrid score matching estimator, compared to the maximum likelihood estimator,
will often be close to 1.
Score matching estimators can also be developed for distributions on noncompact manifolds, including Euclidean spaces. The simplest example is the multivariate normal distribution (Hyvärinen, 2005) where it turns out that the score
matching estimator is identical to the maximum likelihood estimator. Many directional distributions are approximately normal under high concentration. Hence we
expect high efficiency of the score matching estimator in this setting. See Table 3
for confirmation in the von Mises case.
In the directional setting, the score matching estimator can often be interpreted
as a “double moment estimator”. For example, for the von Mises-Fisher distribution
the sufficient statistic is a linear function of z, but the matrix Wn in the score
matching estimator involves quadratic functions of z. Similarly, for the Bingham
distribution the sufficient statistic is a quadratic function of z, but the matrix Wn
in the score matching estimator involves quartic functions of z.
This paper has emphasized the setting where M is a sphere or a product of
spheres. Work is in progress to investigate the score matching estimator for models
18
on other important manifolds such as Stiefel and Grassmann manifolds.
Acknowledgements
The authors are grateful to Derek Harland, John Wood and Peter Kim for helpful
discussions about the differential geometry, and to Peter Forbes for helpful discussions of the von Mises case. The first author also thanks the Indian Institute of
Management Ahmedabad for hospitality during his visits to carry out some of this
work.
11
Appendix
This section gives a proof of the asymptotic relative efficiency for the score matching
estimator of κ in (20). For simplicity the proof focuses on the estimator in the
reduced model.
First, it can be shown that, asymptotically,
nvar(κ̂SME,hybrid ) =
κ
{2κ
A21 (κ)
− 3A1 (κ)}.
(A1)
To verify this equation note that, treating θ as random from the von Mises distribution, var(cos θ) = 1 − A21 (κ) − A1 (κ)/κ, var(sin2 θ) = 81 {3 + A4 (κ) − 4A2 (κ)} −
A21 (κ)/κ2 and cov(cos θ, sin2 θ) = 41 {A1 (κ) − A3 (κ)} − A21 (κ)/κ. Using the delta
rule gives
var(κ̂SME,hybrid ) =
κ
κ(8 + 3κ2 )I0 − 8I1 − 4κ2 I1 − 4κ3 I2 + 4κ2 I3 + κ3 I4 ,
8nI1 A1 (κ)
with the shorthand notation Iα = Iα (κ). Repeated use of (15) leads to (A1).
For the maximum likelihood estimator it can be shown that, asymptotically,
n var(κ̂MLE ) =
1−
A21 (κ)
1
.
− A1 (κ)/κ
(A2)
For both estimators the asymptotic variance converges to 2 as κ → 0 and is asymptotic to 2κ2 as κ → ∞. Hence for small and large κ , the asymptotic relative
efficiency tends to 1. Combining (A1) and (A2) yields (20).
The discussion here has emphasized the hybrid score matching estimator and the
maximum likelihood estimator for the full model with two unknown parameters. But
in fact there are other models and estimators to consider, including three versions of
the score matching estimator and two versions of the maximum likelihood estimator.
The hybrid score matching estimator and the full score matching estimator are
defined for the full model with two unknown parameters θ0 and κ. The score
matching estimator can also be defined under the reduced model with just one
unknown parameter. It turns out that all three estimators of κ have the same
asymptotic variance. Similarly, the maximum likelihood estimator of κ can be
19
defined in the setting of the full model or the reduced model. Again both estimators
have the same asymptotic variance. Hence the asymptotic relative efficiency takes
the same value for all these possibilities.
References
Beran, R. (1979). Exponential models for directional data. Ann. Statist., 7:1162–
1178.
Chavel, I. (1984). Eigenvalues in Riemannian Geometry. Academic Press, Orlando.
Dawid, A. P. and Lauritzen, S. L. (2005). The geometry of decision theory. In
Proceedings of the Decond International Symposium on Information Geometry
and Applications, pages 22–28. Univ. Tokyo.
Forbes, P. G. M. and Lauritzen, S. (2014). Linear estimating equations for exponential families with applications to Gaussian linear concentrations models. Linear
Algebra and its Applications, 473:261–283.
Hyvärinen, A. (2005). Estimation of non-normalized statistical models by score
matching. Journal of Machine Learning Research, 6:695–709.
Hyvärinen, A. (2007). Some extensions of score matching. Computational Statistics
and Data Analysis, 51:2499–2512.
Jammalamadaka, S. R. and Sengupta, A. (2001). Topics in Circular Statistics.
World Scientific.
Jost, J. (2005). Riemannian Geometry and Geometric Analysis, fourth edition.
Springer, Berlin.
Jupp, P. E. and Mardia, K. V. (1980). A general correlation coefficient for directional
data and related regression problems. Biometrika, 67:163–73.
Kent, J. T. (1982). The Fisher–Bingham distribution on the sphere. J. Roy. Statist.
Soc. B, 44:71–80.
Kume, A., Preston, S. P., and Wood, A. T. A. (2013). Saddlepoint approximations for the normalizing constant of Fisher–Bingham distributions on products
of spheres and Stiefel manifolds. Biometrika, 100:971–984.
Lund, U. and Agostinelli, C. (2012). CircStats. R package version 0.2-4.
Mardia, K. V. (1975). Statistics of directional data (with discussion). J. Roy. Statist.
Soc. B, 37:349–393.
Mardia, K. V. (2013). Statistical approaches to three key challenges in protein
structural bioinformatics. Applied Statistics, 62:487–514.
20
Mardia, K. V., Hughes, G., Taylor, C. C., and Singh, H. (2008). A multivariate von
Mises distribution with applications to bioinformatics. The Canadian Journal of
Statistics, 36:99–109.
Mardia, K. V. and Jupp, P. E. (2000). Directional Statistics. Wiley.
Mardia, K. V., Kent, J. T., Hughes, G., and Taylor, C. C. (2009). Maximum
likelihood estimation using composite likelihoods for closed exponential families.
Biometrika, 96:975–982.
Mardia, K. V., Kent, J. T., Zhang, Z., Taylor, C. C., and Hamelryck, T. (2012). Mixtures of concentrated multivariate sine distributions with applications to bioinformatics. J. Appl. Statist., 39:2475–2492.
Mardia, K. V. and Patrangenaru, V. (2005). Directions and projective shapes. Ann.
Statist., 33:1666–1699.
Parry, M., Dawid, A. P., and Lauritzen, S. (2012). Proper local scoring rules. Ann.
Statist., 40:561–592.
Patrangenaru, V. and Ellingson, L. (2016). Nonparametric Statistics on Manifolds
and Their Applications to Object Data Analysis. CRC Press, Boca Raton.
R Core Team (2013). R: A Language and Environment for Statistical Computing.
R Foundation for Statistical Computing, Vienna, Austria.
Rosenberg, D. (1997). The Laplacian on a Riemannian Manifold. Cambridge University Press.
Sei, T. and Kume, A. (2015). Calculating the normalising constant of the Bingham
distribution on the sphere using the holonomic gradient method. Statistics and
Computing, 25:321–332.
Singh, H., Hinizdo, V., and Demchuk, E. (2002). Probabilistic model for two dependent circular variables. Biometrika, 89:719–723.
Stein, E. M. and Weiss, G. (1971). Introduction to Fourier Analysis on Euclidean
Spaces. Princeton University Press.
21
| 10math.ST
|
arXiv:1711.08295v1 [math.GR] 22 Nov 2017
SCALING LIMITS OF CAYLEY GRAPHS WITH POLYNOMIALLY GROWING
BALLS
ROMAIN TESSERA AND MATTHEW C. H. TOINTON
Abstract. Benjamini, Finucane and the first author have shown that if (Γn , Sn ) is a sequence of
d
Cayley graphs such that |Snn | ≪ nD |Sn |, then the sequence (Γn , Snn ) is relatively compact for the
Gromov–Hausdorff topology and every cluster point is a connected nilpotent Lie group equipped
with a left-invariant sub-Finsler metric. In this paper we show that the dimension of such a cluster
point is bounded by D, and that, under the stronger bound |Snn | ≪ nD , the homogeneous dimension
of a cluster point is bounded by D. Our approach is roughly to use a well-known structure theorem
for approximate groups due to Breuillard, Green and Tao to replace Snn with a coset nilprogression
of bounded rank, and then to use results about nilprogressions from a previous paper of ours to
study the ultralimits of such coset nilprogressions. As an application we bound the dimension of
the scaling limit of a sequence of vertex-transitive graphs of large diameter. We also recover and
effectivise parts of an argument of Tao concerning the further growth of single set S satisfying the
bound |S n | ≤ M nD |S|.
Contents
1. Introduction
2. Background on approximate groups
3. GH-limits and ultralimits
4. The abelian case
5. Progressions in free nilpotent groups
6. Overview of the general case
7. Reduction to simply connected nilpotent Lie groups
8. Reduction to ultralimits of normed Lie algebras
9. Marked Lie algebras
10. Volume growth and Hausdorff dimension of the limit
Appendix A. Detailed growth of nilprogressions
References
1
5
7
9
12
15
17
26
27
29
32
36
1. Introduction
We start by recalling Gromov’s famous polynomial growth theorem.
The first author is supported by grant ANR-14-CE25-0004 “GAMME”. The second author is supported by grant
FN 200021 163417/1 of the Swiss National Fund for scientific research; for earlier periods of this project he was
supported by ERC grant GA617129 ‘GeTeMo’ and a Junior Research Fellowship from Homerton College, University
of Cambridge.
1
2
ROMAIN TESSERA AND MATTHEW C. H. TOINTON
Theorem ([14, 27]). Let Γ be a group generated by a finite symmetric subset S. If there exists an
unbounded sequence of integers mn , and constants C and D such that |S mn | ≤ CmD
n for all n, then
Γ is virtually torsion-free nilpotent.
Conversely, we know from Bass and Guivarc’h [1, 16] that for every torsion-free nilpotent group
Γ equipped with a symmetric finite generating subset S there exists a constant C = C(Γ, S) ≥ 1
such that for every n ∈ N we have
C −1 nhdim (Γ) ≤ |S n | ≤ Cnhdim (Γ) ,
where hdim (Γ) is the homogeneous dimension of Γ (it is in particular an integer).
More recently, Shalom and Tao [22] showed that for every D there exists n0 such that for every
group Γ generated by a finite symmetric subset S, if there exists just one integer n1 ≥ n0 such that
(1.1)
|S n1 | ≤ nD
1
then Γ is virtually nilpotent. Breuillard, Green and Tao [8, Corollary 11.5] subsequently showed
that (1.1) could be weakened to a “relative” polynomial-growth condition of the form
(1.2)
|S n1 | ≤ nD
1 |S|.
One can actually go further and show that if one takes a “large” group Γ satisfying (1.2) then
when “zooming out” at scale n1 it “looks like” a connected nilpotent Lie group. To make this
idea precise, one can use the Gromov–Hausdorff topology, or GH topology, which gives a way to
measure how “close” two metric spaces are to one another; see §3 for a precise definition. To study
Γ “zoomed out” at scale n1 , one can consider the rescaled Cayley graph (Γ, dnS1 ) with respect to the
GH topology.
The first author obtained the following result in this direction in joint work with Benjamini and
Finucane.
Theorem 1.1 ([2, Theorem 3.2.2.]). Let (Γn , Sn ) be a sequence of Cayley graphs such that |Snn | ≪
d
nD |Sn |. Then for every sequence mn ≫ n the sequence (Γn , mSnn ) is GH-relatively compact, and every
cluster point is a connected nilpotent Lie group equipped with a left-invariant sub-Finsler metric.
Remarks. The authors of [2] use the terminology Carnot–Carathéodory metric to mean a subFinsler metric (we caution that sometimes Carnot–Carathéodory is used more specifically to mean
sub-Riemmannian). Moreover, whilst [2, Theorem 3.2.2.] is stated for finite graphs this is in fact a
misprint and the proof is also valid for infinite graphs. Finally, [2, Theorem 3.2.2.] has the stronger
hypothesis that |Snn | = O(nD ), but the same proof yields the same result under the hypothesis of
Theorem 1.1 that |Snn | = O(nD |Sn |).
Main new results. Our first result in this paper bounds by D the dimension of any cluster point
arising from Theorem 1.1.
Theorem 1.2. Let (Γn , Sn ) be a sequence of Cayley graphs such that |Snn | ≪ nD |Sn |. Then for
d
every sequence mn ≫ n the sequence (Γn , mSnn ) is GH-relatively compact, and every cluster point is
a connected nilpotent Lie group of dimension at most D equipped with a left-invariant sub-Finsler
metric.
Remark 1.3. In particular, the homogeneous dimension of every cluster point coming from Theorem
1.2 is at most (D − 1)D/2 + 1.
SCALING LIMITS OF CAYLEY GRAPHS WITH POLYNOMIALLY GROWING BALLS
3
In light of the Bass–Guivarc’h formula, it is tempting to wonder whether even the homogeneous
dimension should be bounded by D. However, the following example shows that Theorem 1.2 and
1ZZ
Remark 1.3 are both sharp. Consider the subset S of the discrete Heisenberg group H(Z) = 0 1 Z
0 0 1
defined by
1 [−n,n] [−n3 ,n3 ]
,
S= 0 1
[−n,n]
0
0
1
which also appears in [23, Example 1.11]. This set Sn satisfies |Snn | ≤ n3 |Sn |, while one easily checks
that the scaling limit of (H(Z), dSn /n) is the real Heisenberg group, whose homogeneous dimension
is 4 but whose dimension is indeed 3 (note that, although S is not symmetric, this can be fixed by
considering the set S ∪ S −1 in its place; we leave the details to the reader).
Nonetheless, replacing the relative growth hypothesis (1.2) with the stronger absolute growth
hypothesis (1.1) we can indeed bound the homogeneous dimension by D, as follows.
Theorem 1.4. Let (Γn , Sn ) be a sequence of Cayley graphs such that |Snn | ≪ nD . Then for every
d
sequence mn ≫ n the sequence (Γn , mSnn ) is GH-relatively compact, and every cluster point is a
connected nilpotent Lie group of homogeneous dimension at most D equipped with a left-invariant
sub-Finsler metric.
We emphasise that the proofs of Theorems 1.2 and 1.4 (and, indeed, Theorem 1.1) build on the
fundamental structure theorem for approximate groups proved by Breuillard, Green and Tao in
[8]. Our new results also crucially rely on more specific results about approximate groups that we
established in our previous paper [24].
The dimension and the homogeneous dimension coincide for a connected abelian Lie group.
Accepting for a moment that if the groups Γn appearing in Theorem 1.2 are abelian then every
cluster point naturally comes with the structure of an abelian group, it follows that in the abelian
setting we do have Theorem 1.4 under the weaker hypothesis of Theorem 1.2, as follows.
Theorem 1.5. Let (Γn , Sn ) be a sequence of abelian Cayley graphs such that |Snn | ≪ nD |Sn |. Then
d
for every sequence mn ≫ n the sequence (Γn , mSnn ) is GH-relatively compact, and every cluster point
is a connected abelian Lie group of dimension at most D equipped with an invariant Finsler metric.
In fact, as well as a stronger statement, the abelian case admits a much shorter proof than the
general case that nonetheless captures many of the ideas we use in the general case and so may
serve as a useful introduction to the general case for first-time readers. We therefore give this short
proof in §4.
Application to scaling limits of vertex-transitive graphs of large diameter. As
a corollary of Theorem 1.5 we obtain the following refinement of [2, Theorems 1 & 2]. In [2,
Theorem 1], Benjamini, Finucane and the first author show that if (Xn ) is a sequence of vertexdSn
)
transitive graphs satisfying the large-diameter condition |Xn | ≪ diam(Xn )D then (Xn , diam(X
n)
is GH-relatively compact and every cluster point is a torus equipped with an invariant Finsler
metric. In [2, Theorem 2] they give a sharp bound on the dimension of the limiting torus under the
assumption that the degree of the sequence Xn is uniformly bounded. We are now able to lift this
condition on the degree. Indeed, combining Theorem 1.5 with [2, Propositions 3.1.1 & 3.3.1] as in
the proof of [2, Theorem 2], we obtain the following result.
4
ROMAIN TESSERA AND MATTHEW C. H. TOINTON
Corollary 1.6. Let (Xn ) be a sequence of finite vertex-transitive graphs satisfying the large-diameter
condition
(1.3)
|Xn | ≪ deg(Xn )diam(Xn )D ,
d
Sn
)
where deg(Xn ) is the degree of Xn (potentially going to infinity). Then the sequence (Xn , diam(X
n)
is GH-relatively compact, and every cluster point is a torus of dimension at most D equipped with
an invariant Finsler metric.
The Hausdorff dimension of the scaling limit. It is interesting to note that in both Theorems 1.2 and 1.4 we use D to control some algebraic invariant of the limit. Indeed, both the
dimension and the homogeneous dimension depend only on the limiting Lie group, not on the limiting metric. It is therefore tempting to try to refine Theorems 1.2 by relating D to the Hausdorff
dimension of the limit, which always lies between the dimension and the homogeneous dimension
(see e.g. the proof of [4, Theorem 13]). The following statement shows that this is in fact impossible
in quite a strong sense.
Proposition 1.7. For every positive non decreasing function f such that limt→∞ f (t) = ∞, there
exists a sequence of symmetric generating subsets Sn of the discrete Heisenberg group H(Z) such
that |Sn |n ≪ f (n)n3 |Sn |, whilst (H(Z), dSn /n) converges to the real Heisenberg group equipped with
the left-invariant Carnot–Carathéodory metric associated to the ℓ∞ -norm on R2 , whose Hausdorff
dimension coincides with the homogeneous dimension (which is 4).
We refer to §10 for the necessary background on Carnot–Carathéodory metrics.
It turns out that this proposition itself is sharp, in the sense that the condition |Snn | ≪ f (n)n3 |Sn |
cannot be improved to |Sn |n ≪ n3 |Sn |. This follows from the following general fact.
Proposition 1.8. Suppose that the assumptions of Theorem 1.2 are satisfied, and that the sequence
d
(Γn , mSnn ) converges to a connected Lie group G of dimension exactly D equipped with some leftinvariant sub-Finsler metric d. Then d is Finsler and (G, d) has Hausdorff dimension D.
Further growth of locally polynomially growing sets. A key ingredient in the proof
of Theorem 1.1 was a result of Breuillard and the second author [9, Theorem 1.1] implying, for
example, that there exists a constant CD such that if (1.2) holds for sufficiently large n1 (depending
k |S m | for every k ∈ N and m ≥ n . Tao subsequently gave a more precise
on D) then |S km | ≤ CD
1
description of the further growth of S m , as follows.
Theorem 1.9 (Tao [23, Theorem 1.9]). Given D > 0 there exists N = NM,D such that if n ≥ N
and S is a finite symmetric generating set for a group G such that
|S n | ≤ M nD |S|
then there exists a non-decreasing continuous piecewise-linear function f : [0, ∞) → [0, ∞) with
f (0) = 0 and at most OD (1) distinct linear pieces, each with a slope that is a natural number at
most OD (1), such that
log |S mn | = log |S n | + f (log m) + OD (1)
for every m ∈ N.
Tao phrases his argument in the language of non-standard analysis, both to streamline the presentation and to give easier access to some results of [8] that are also phrased in that language. It
turns out that some of the techniques we have developed in the present paper and its predecessor
SCALING LIMITS OF CAYLEY GRAPHS WITH POLYNOMIALLY GROWING BALLS
5
[24] allow us to translate Tao’s argument to a finitary setting (we emphasise that it is still essentially Tao’s argument). Since the details are brief, we take this opportunity to present them in the
appendix.
We caution that Theorem 1.9 is still ineffective, since the argument uses [8, Theorem 1.6] as a
black box (see Remark A.8 for details). Nonetheless, this is now the only source of ineffectiveness
in the argument, and so Theorem 1.9 is effective for any class of group in which there is an effective
version of [8, Theorem 1.6] (this includes residually nilpotent groups [26] and certain linear groups
[7, 12], for example). In particular, an effective version of [8, Theorem 1.6] in general would yield
an effective version of Theorem 1.9 in general.
Notation. We follow the standard convention that if X, Y are real quantities and z1 , . . . , zk are
variables or constants then the expressions X ≪z1 ,...,zk Y and Y ≫z1 ,...,zk X each mean that there
exists a constant C > 0 depending only on z1 , . . . , zk such that X is always at most CY . We write
X ≍z1 ,...,zk Y to mean that X ≪z1 ,...,zk Y ≪z1 ,...,zk X.
Moreover, the notation Oz1 ,...,zk (Y ) denotes a quantity whose absolute value is at most a certain
constant (depending on z1 , . . . , zk ) multiple of Y , whilst Ωz1 ,...,zk (X) denotes a quantity that is at
least a certain positive constant (depending on z1 , . . . , zk ) multiple of X. Thus, for example, the
meaning of the notation X ≤ O(Y ) is identical to the meaning of the notation X ≪ Y .
Here, and throughout this paper, we use the standard notation AB = {ab : a ∈ A, b ∈ B},
−1
An = {a1 · · · an : ai ∈ A} and A−n = {a−1
1 · · · an : ai ∈ A}. We also adopt the additional convention
that given two subsets A, B of a group we write A ≈z1 ,...,zk B to mean that A ⊂ B Oz1,...,zk (1) and
B ⊂ AOz1 ,...,zk (1) . If A, B are subsets of a Lie algebra, we adopt the same convention but with
additive notation, writing A ≈z1 ,...,zk B to mean that A ⊂ Oz1 ,...,zk (1)B and B ⊂ Oz1 ,...,zk (1)A.
Definition 1.10. Given C ≥ 1 and K ≥ 0 and metric spaces X, Y , a map f : X → Y is said to be
a (C, K)-quasi-isometry if
C −1 d(x, y) − K ≤ d(f (x), f (y)) ≤ Cd(x, y) + K,
for every x, y ∈ X, and every y ∈ Y lies at distance at most K from f (X).
Overview of the paper. In Section 2 we recall some notation and background material on
approximate groups and progressions from our previous paper. In Section 3 we define Gromov–
Hausdorff limits, explaining how to reduce their study to the study of ultralimits, and then in Section
4 we give a direct proof of Theorem 1.5.
We give an overview of the general argument for Theorems 1.2 and 1.4 in Section 6, reducing
their proofs to two results labelled Proposition 6.1 and Theorem 6.2. We prove Proposition 6.1 in
Section 7, and Theorem 6.2 in Sections 8 and 9.
Finally, we prove Propositions 1.7 and 1.8 in Section 10, and Theorem 1.9 in the appendix.
Acknowledgements. The authors are grateful to Itai Benjamini for stimulating discussions.
2. Background on approximate groups
The strenghtening of Gromov’s theorem due to Breuillard, Green and Tao described in the
1/2
introduction relies on the observation that if S satisfies (1.2) then there is some n ≥ n1 such that
S n is a so-called approximate group. A finite subset A of a group G is said to be a K-approximate
6
ROMAIN TESSERA AND MATTHEW C. H. TOINTON
subgroup of G, or simply a K-approximate group, if it is symmetric and contains the identity and
there exists X ⊂ G with |X| ≤ K such that A2 ⊂ XA.
The central result of approximate group theory is a remarkable theorem of Breuillard, Green
and Tao [8, Corollary 2.11] describing the algebraic structure of an arbitrary approximate group in
terms of certain objects called progressions, versions of which we now define. This result underpins
the approach of the present paper.
Following [25], we define the ordered progression on generators u1 , . . . , ud ∈ G with lengths
L1 , . . . , Ld is to be
Pord (u; L) := {uℓ11 · · · uℓdd : |li | ≤ Li }.
If P is an ordered progression and H is a finite subgroup normalised by P , then we say that HP is
an ordered coset progression.
Following [8], we say that the tuple (u; L) = (u1 , . . . , ud ; L1 , . . . , Ld ) is in C-upper-triangular form
if, whenever 1 ≤ i < j ≤ d, for all four choices of signs ± we have
CLj+1
CLd
±1
,
.
.
.
,
(2.1)
[u±1
,
u
]
∈
P
u
,
.
.
.
,
u
;
j+1
ord
d Li Lj
i
j
Li Lj .
We say that an ordered progression is in C-upper-triangular form if the corresponding tuple is. We
say that a coset ordered coset progression HP is in C-upper-triangular form if the corresponding
tuple is in C-upper-triangular form modulo H.
Given m > 0, an ordered progression P on the tuple (u; L) = (u1 , . . . , ud ; L1 , . . . , Ld ) is said to be
m-proper with respect to a homomorphism π : hP i → N if the elements π(uℓ11 · · · uℓdd ) are all distinct
as the ℓi range over those integers with |ℓi | ≤ mLi . The progression P is said to be m-proper
with respect to a subgroup H ⊳ hHP i if P is m-proper with respect to the quotient homomorphism
hHP i → hHP i/H. In this case we also say that the coset ordered coset progression HP is m-proper.
If a coset ordered coset progression HP is m-proper for every m < 0 then we say it is infinitely
proper.
Having made these definitions, we can now state the main the result of Breuillard, Green and
Tao. This result is essentially [8, Corollary 2.11], although we state a version of it from our earlier
paper [24, Theorem 1.4] in order to be compatible with our notation.
Theorem 2.1 (Breuillard–Green–Tao). Let A be a K-approximate group. Then there exist an
ΩK (1)-proper ordered coset progression HP ⊂ A4 , of rank and step at most OK (1) and in OK (1)upper-triangular form, and a set X ⊂ hAi with |X| ≪K 1 such that A ⊂ XHP .
This result enters the present paper via Proposition 7.5.
We adapt some of the above terminology for use in a Lie group G. In particular, given elements
R (u; L) via
u1 , . . . , ud ∈ G and positive reals L1 , . . . , Ld we define the real ordered progression Pord
R
(u; L) = {uℓ11 · · · uℓdd : ℓi ∈ [−Li , Li ]},
Pord
calling d the rank of this progression. We say that (u; L) is in C-upper-triangular form over R if
for every i < j ≤ d and for all four choices of signs ± we have
CLj+1
CLd
±1
R
,
.
.
.
,
u
,
.
.
.
,
u
;
[u±1
,
u
]
⊂
P
j+1
d Li Lj
ord
i
j
Li Lj .
Finally, given group elements u1 , . . . , ud and positive reals L1 , . . . , Ld , and given integers i1 <
. . . < ik ∈ [1, d], we often abbreviate (ui1 , . . . , uik ; Li1 , . . . , Lik ) to simply (ui1 , . . . , uik ; L).
SCALING LIMITS OF CAYLEY GRAPHS WITH POLYNOMIALLY GROWING BALLS
7
3. GH-limits and ultralimits
In this section we describe two different ways of defining a limit of a sequence of metric spaces:
a Gromov–Hausdorff limit and an ultralimit. Underpinning the entire approach of this paper is the
standard fact, stated below as Lemma 3.9, that in the setting we are concerned with these two types
of limit coincide.
We start by describing Gromov–Hausdorff limits, or GH-limits as we call them from now on.
Definition 3.1. Given a sequence Xn of compact metric spaces we will say that Xn GH-converges
to X if the Xn have bounded diameter and if there exist maps φn : Xn → X such that for all ε,
then for n large enough,
• every point of X is at ε-distance of a point of φn (Xn );
• (1 − ε)d(x, y) − ε ≤ d(φn (x), φn (y)) ≤ (1 + ε)d(x, y) + ε for all x, y ∈ Xn .
GH-convergence also extends naturally to (not necessarily compact) locally compact pointed
metric spaces [15, §3], as follows.
Definition 3.2. Given a sequence (Xn , on ) of locally compact pointed metric spaces, (Xn , on ) is
said to converge to the locally compact pointed metric space (X, o) if for every R > 0, the sequence
of balls B(on , R) GH-converges to B(o, R).
Gromov gave the following useful criterion for GH-precompactness.
Lemma 3.3 (Gromov’s compactness criterion, [10, Theorem 5.41]). A sequence of metric spaces
(Xn ) is relatively GH-compact if and only if the Xn have bounded diameter, and are “equi-relatively
compact”: for every ε > 0, there exists N ∈ N such that for all n ∈ N, Xn can be covered by at
most N balls of radius ε.
This fact readily extends to pointed metric spaces as follows.
Corollary 3.4. A sequence of pointed metric spaces (Xn , on ) is relatively GH compact if and only
if for every ε > 0, there exists N ∈ N such that for all n ∈ N, B(on , 1/ε) can be covered by at most
N balls of radius ε.
The following standard observation is particularly useful in light of Corollary 3.4 (recall that a
metric space is called homogeneous if its group of isometries acts transitively).
Lemma 3.5. Let (X, d) be a homogeneous discrete metric space, and assume that for all r > 0 and
x ∈ X we have |B(x, 2r)| ≤ C|B(x, r)|. Then B(x, 2r) is covered by OC (1) balls of radius r.
Proof. Let (xi )i∈I be a maximal r-separated (i.e. elements are at pairwise distance at least r) subset
of B(x, 2r). Since the balls B(xi , r/3) are disjoint and contained in B(x, 3r), we deduce that I has
cardinality OC (1). On the other hand we deduce from maximality that the balls B(xi , r) cover
B(x, 2r).
Corollary 3.6. Let (Xn , on ) be a sequence of pointed discrete metric spaces such that for every n,
the metric space Xn is homogeneous and satisfies some uniform doubling property: there exists a
sequence εn → 0 and a constant C such that for all n ∈ N, r ≥ εn , |B(on , 2r)| ≤ C|B(on , r)|. Then
the sequence of pointed metric spaces (Xn , on ) is GH-relatively compact.
In [15, §3], Gromov discusses functoriality of these notions of convergence. In particular, he
mentions a notion of equivariant GH-convergence. In [13, §3], Fukaya and Yamaguchi provide a
8
ROMAIN TESSERA AND MATTHEW C. H. TOINTON
precise notion of GH-convergence for sequences of triplets (Xn , Gn , on ), where (Xn , on ) are locally
compact pointed metric spaces, and Gn is a group of isometries of Xn . They prove the important fact
that if (Xn , on ) GH-converges to (X, o) and if Gn is a group of isometries of Xn acting transitively,
then in some suitable sense, a subsequence of the triplet (Xn , Gn , on ) converges to (X, G, o) where
G is a subgroup of isometries of X acting transitively. We will recover this fact in a special case,
through the convenient notion of ultralimits.
Before we can define ultralimits, we need to define ultrafilters. An ultrafilter is a map from
ω : P(N) → {0, 1}, such that ω(N) = 1, and which is “additive” in the sense that ω(A ∪ B) =
ω(A) + ω(B) for all A and B disjoint subsets of N [11]. It is easy to check that this is equivalent to
the definition used in [8], for example.
Given an ultrafilter ω, we say that a statement holds for ω-almost every n ∈ N if the set A of
n for which the statement holds satisfies ω(A) = 1. We write xn =ω yn to mean that xn = yn
for ω-almost every n; we write xn ≪ω yn to mean that there exists some real number C such that
xn ≤ Cyn for ω-almost every n; and we write Oω (xn ) for a sequence that is ≪ω xn , write Ωω (xn )
for a sequence that is ≫ω xn , and write oω (xn ) for a sequence that is not in Ωω (xn ).
Let us highlight a slightly subtle distinction between the notation xn ≪ω yn and the asymptotic
notation xn ≪ yn . The notation xn ≪ yn implies the existence of some universal constant C such
that xn ≤ Cyn for all n, whereas the notation xn ≪ω yn merely requires that for the sequence
in question there exists some constant C such that xn ≤ Cyn for ω-almost every n. Thus, for
example, if we let m be an arbitrary integer, and then defined f (n) = m for every n, we would have
f (n) ≪ω 1, but we could not write f (n) ≪ 1 because for any choice of C we would violate this
assertion by choosing m = ⌈C + 1⌉.
Ultrafilters are used to “force” convergence of bounded sequences of real numbers. Given such a
sequence an , its limit is defined to be the unique real number a such that for every ε > 0 we have
|an − a| < ε for ω-almost every n. In this case we denote limω an = a, or write an →ω a. A trivial
observation that is nonetheless extremely useful in this paper is the following.
Lemma 3.7. Let (an ) be a sequence of real numbers taking only finitely many different values.
Then an =ω limω an .
An ultrafilter is called non-principal if it vanishes on finite subsets of N. Non-principal ultrafilters
are known to exist, although this requires the axiom of choice [11]. From now on we fix some nonprincipal ultrafilter ω.
Definition 3.8. Given a sequence of pointed metric spaces (Xn , on ), its ultralimit (X∞ , o∞ ) with
respect to ω is the quotient of
Xω = {(xn ) ∈ Πn Xn : d(xn , on ) ≪ω 1}
by the equivalence relation xn ∼ yn if d(xn , yn ) =ω o(1). It is equipped with a distance defined by
d∞ ((xn ), (yn )) = limω d(xn , yn ).
It is a basic fact that a sequence an ∈ R converging to a ∈ R in the usual sense also converges to
a with respect to ω, in the sense that limω an = a. A similar fact holds for ultralimits of pointed
metric spaces, as follows.
Lemma 3.9 ([10, Exercise 5.52]). If a sequence of pointed metric spaces converges in the pointed
GH sense to X, then its ultralimit with respect to ω is isometric to X.
Assume from now that the sequence (Xn , on ) is a sequence of groups (Gn , en , dn ) where en is the
identity element and dn is some left-invariant metric. Then Gω is a group that acts transitively on
SCALING LIMITS OF CAYLEY GRAPHS WITH POLYNOMIALLY GROWING BALLS
9
G∞ . The stabiliser of the identity is
Gω,0 = {(xn ) ∈ Πn Gn : d(xn , en ) =ω o(1)}.
If the Gn are abelian groups, then G∞ naturally comes with an abelian group structure as well.
More generally, Gω,0 is a normal subgroup of Gω if and only if for all sequences (gn ) and (hn ) such
that d(gn , en ) =ω O(1) and d(hn , en ) =ω o(1) we have d([gn , hn ], en ) =ω o(1). In this case (G∞ , d∞ )
naturally comes with a metric group structure.
We close this section by recording the following basic fact about maps between metric spaces that
behave well under ultra-limits. The proof is straightforward from the definition of an ultralimit, so
we leave it as an exercise.
Proposition 3.10. Let ω be a non-principal ultrafilter on N. Assume that (Xn , on ) and (Xn′ , o′n )
are sequences of pointed metric spaces, and let φn : Xn → Xn′ be a sequence of maps. We denote by
′ , o′ ) the respective ultralimits of (X , o ) and (X ′ , o′ ).
(X∞ , o∞ ) and (X∞
n n
n n
∞
(i) Assume that there exists C ≥ 1 such that for every sequence un , vn ∈ B(on , Oω (1)) we have
d(φ(un ), φ(vn )) ≤ω Cd(un , vn ) + oω (1).
Then there exists a C-Lipschitz map φ∞ : X∞ → Y∞ , defined as φ∞ (limω un ) = limω φ(un )
for all un ∈ B(on , Oω (1)).
(ii) Assume that in addition to (i) that there exist α, β > 0 such that for every sequence u′n ∈
B(o′n , β) there exist un ∈ B(on , α) such that d(u′n , φn (un )) = oω (1). Then φ∞ (B(o∞ , α))
contains B(o′∞ , β).
Corollary 3.11. Let φn : (Gn , dn ) → (Hn , δn ) be a sequence of morphisms between two groups
equipped with left-invariant metrics such the assumptions of Proposition 3.10 (ii) are satisfied.
Assume, moreover, that Gω,0 and Hω,0 are normal subgroups of respectively Gω and Hω , and hence
that the ultralimits G∞ and H∞ naturally come with group structures (this is the case, for example,
if Gn and Hn are abelian). Then φ∞ is a Lipschitz open homomorphism, and if H∞ is connected
then φ∞ is surjective, and therefore is an isomorphism.
We end this section with the following standard observation.
Remark 3.12. Recall that an ultralimit of a sequence of geodesic metric spaces is geodesic (see for
instance [20, §7.5]). On the other hand if (Xn ) is a sequence of connected graphs equipped with their
usual geodesic distances dn , mn is a sequence going to infinity, and ω is a non-principal ultrafilter,
the mn -rescaled ultralimit of (Xn , dn ) is isometric to the mn -rescaled ultralimit of the 0-skeletons of
Xn equipped with the induced metric. In particular the latter is geodesic. Applied to our setting,
if (Gn , dSn ) is a sequence of groups equipped with word metrics, then limω (Gn , dSn /mn ) is geodesic
(hence connected).
4. The abelian case
In this section we provide the short proof of Theorem 1.5. This provides a gentle introduction to
the methods of the rest of the paper, but is completely independent of the subsequent sections; in
particular, the reader interested only in Theorem 1.4 could easily skip the present section.
Given an abelian group G, a coset progression of rank d is a subset of G of the form P H,
where H < G is a finite subgroup and P = P (u, L), with u = (u1 , . . . , ud ) for some ui ∈ G and
L = (L1 , . . . , Ld ) for some Li ≥ 1, is defined via
o
n
P = uℓ11 · · · uℓdd : |ℓj | ≤ Lj .
10
ROMAIN TESSERA AND MATTHEW C. H. TOINTON
Given m > 0, this coset progression is called m-proper if the morphism φ : Zd → G/H mapping the
canonical basis x = (x1 , . . . , xi ) to u is injective in restriction to P (x, mL).
The following lemma is both standard and trivial.
Lemma 4.1. Let P be a coset progression of rank at most d inside an abelian group G. Then for
every k ∈ N there exists Xk ⊂ G with |Xk | ≪d 1 such that P 2k ⊂ Xk P k .
We recall the following result from our previous paper.
Theorem 4.2 ([24, Theorem 1.16]). Let A be a K-approximate group such that hAi is abelian.
Then for every m > 0 there exists an m-proper coset progression HP of rank at most K O(1) such
that
A ⊂ HP ⊂ AOK,m (1) .
Lemma 4.3. Under the assumptions of Theorem 1.5, for all but finitely many n ∈ N there exists
a sequence kn ∈ N with n1/4 ≤ kn ≤ n1/2 and a 2-proper coset progression Hn Pn of rank dn ≪D 1
such that
(4.1)
Snkn ⊂ Hn Pn ⊂ SnOD (kn ) ,
and such that, writing Pn = Pord (u(n) , L(n) ), the rank dn is minimal in the sense that for each
(n)
(n)
(n)
(n)
1 ≤ i ≤ dn , if L̄(n) = (L1 , . . . , Li−1 , 0, Li+1 , . . . , Ldn ) then Snkn is not contained in Hn P (u(n) , L̄(n) ).
Proof. Applying [24, Lemma 8.2] with ⌊n1/2 ⌋ in place of n gives, for all but finitely many n, an
integer ℓn satisfying n1/4 ≤ ℓn ≤ 21 n1/2 such that |Sn5ℓn |/|Snℓn | ≪D 1. It then follows from [9, Lemma
2.2] that Sn2ℓn is an OD (1)-approximate group. Setting kn = 2ℓn , the existence of a 2-proper coset
progression Hn Pn satisfying (4.1) then follows from Theorem 4.2. We obtain the minimality of dn
simply by deleting dimensions whenever this is possible without violating (4.1).
We will eventually apply the following lemma to the coset progressions coming from Lemma 4.3.
Lemma 4.4. Let Γ be an abelian group with finite symmetric generating set S containing the
identity, let k ∈ N, and let HP = HP (u; L) be a 2-proper coset progression of rank d such that
S k ⊂ HP but such that
(4.2)
S k 6⊂ HP (u; L1 , . . . , Li−1 , 0, Li+1 , . . . , Ld )
for every i. Then Li ≥ k for every i and S ⊂ HP (u; L/k).
Proof. Let x1 , . . . xd be a basis of Zd , write Q = P (x, L), and let ϕ : Zd → Γ be the homomorphism
mapping xi to ui for each i. The properness of HP implies that there exists a unique subset U ⊂ Q
such that ϕ(U ) = S modulo H.
We first claim that U k ⊂ Q. Indeed, let ℓ ∈ N be minimal such that there exists z1 , . . . , zℓ ∈ U
with z = z1 · · · zℓ ∈
/ Q, noting that by minimality we have z1 · · · zℓ−1 ∈ Q and hence z ∈ Q2 \ Q. If
ℓ ≤ k we would have ϕ(z) ∈ S k ⊂ HP , and so there would exist z ′ ∈ Q with ϕ(z) = ϕ(z ′ ) mod H,
contradicting the 2-properness of P . Thus ℓ > k, and in particular
(4.3)
U k ⊂ Q,
as claimed.
This immediately implies that S ⊂ HP (u; L/k), as required. Combined with (4.2) and the
properness of HP , it also implies that for every i satisfying 1 ≤ i ≤ d there exists z ∈ U such that
the ith coordinate of z is non-zero. It then follows from (4.3) that z k , whose ith coordinate is at
least k, belongs to Q, and so Li ≥ k, as required.
SCALING LIMITS OF CAYLEY GRAPHS WITH POLYNOMIALLY GROWING BALLS
11
Lemma 4.5. Let d ∈ N, and let (Γn , Vn ) be a sequence of abelian Cayley graphs such that for every
n the generating set Vn := P (u(n) , K (n) )Hn is a coset progression of rank at most d. Then for every
sequence mn → ∞ the sequence of (pointed) metric spaces (Γn , dVn /mn ) is GH-relatively compact.
Moreover, for every non-principal ultrafilter ω the ultralimit limω (Γn , dVn /mn ) is a connected abelian
Lie group equipped with a Finsler metric.
Proof. It follows from Corollary 3.4 and Lemma 4.1 that (Γn , dVn /mn ) is relatively compact. Observe that since Hn is contained in a ball of radius 1 = o(mn ), the limit is not affected by modding
out by Hn , so we can simply assume that it is trivial. Note also that by Lemma 3.7 we may assume
that the rank of Vn is equal to d for every n. Consider then the morphism πn : Zd → Γn mapping
the canonical basis x = (x1 , . . . , xd ) to u(n) , and write Wn := P (x(n) , K (n) ). Being a length space,
limω (Γn , dVn /mn ) is in particular connected, therefore Corollary 3.11 and Remark 3.12 implies that
limω φn : limω (Zd , dWn /mn ) → limω (Γn , dVn /mn ) exists and is a surjective morphism. It is therefore
sufficient to prove that limω (Zd , dWn /mn ) is a normed real vector space of dimension d. Note that
WnR = P R (x(n) , K (n) ) is the convex hull of Wn in Rd , and by [2, Lemma 2.2.5.], for all k ∈ N
(WnR )k ⊂ Wnk (WnR )Od (1) .
Hence, (Zd , Wn ) ֒→ (Rd , WnR ) is a (Od (1), Od (1))-quasi-isometry. We deduce from Corollary 3.11
that limω (Zd , dWn /mn ) is bi-Lipschitz isomorphic to limω (Rd , dWnR /mn ). Observe that (Rd , WnR ) is
isometric to (Rd , W R ), where W R = P R (x(n) , 1)), and one easily checks that limω (Rd , dW R /mn ) is
isometric to (Rd , k · k), where the norm k · k has W R as unit ball. In conclusion, limω (Zd , dWn /mn )
is bi-Lipschitz isomorphic to Rd , and since its metric is invariant and geodesic (by Remark 3.12), it
actually comes from a norm (see the proof of [2, Theorem 2.2.4]).
Lemma 4.6. Let d ∈ N, let ω be a non-principal ultrafilter, and let (Γn , Vn ) be a sequence of abelian
Cayley graphs such that for ω-almost every n the generating set Vn := P (u(n) , K (n) )Hn is a coset
(n)
progression of rank d. Then, on permuting the indices of ui for every such n if necessary, there
(n)
(n)
exists d′ ≤ d such that the coset progression Vn′ := P (u, K1 , . . . , Kd′ , 0, . . . , 0))Hn is Ωω (mn )′
′
′
′
proper and, writing Γn = hVn i the inclusion (Γn , Vn ) ֒→ (Γn , Vn ) is a 1-Lipschitz (1, oω (mn ))-quasiisometry.
Proof. The lemma is trivial if d = 0, so we may assume that d ≥ 1 and, by induction, that the
lemma holds with d − 1 in place of d.
If Vn is Ωω (mn )-proper the lemma is trivially satisfied, so we may assume that there exists a
sequence rn = oω (mn ) and, for ω-almost every n, an element z (n) ∈ P (x(n) , K (n) )rn such that,
writing x1 , . . . xd for the standard basis of Zd and ϕn : Zd → Γn for the homomorphism mapping xi
(n)
to ui for each i, we have ϕn (z (n) ) = 1. On permuting the coordinates of z (n) we may assume that
(n)
(n)
(n)
(n)
|zd |/Kd = maxi |zi |/Ki .
(n)
(n)
Let Λn be the subgroup of Γn generated by Wn = P (u, K1 , . . . , Kd−1 , 0)). We claim that, for
every r ≥ 0, every element v of length r in (Γn , Vn ) lies at distance at most rn + 1 (with respect
to dVn ) from an element y ∈ Λn of length r + rn with respect to dWn . By definition, such a v is of
(n)
the form v = uj11 . . . ujdd h with maxi |ji |/Ki = r, and h ∈ Hn . Let q and t be integers such that
(n)
(n)
(n)
jd = qzd + t, with |t| < |zd | and 0 ≤ q ≤ |jd |/|zd |. Now, using that ϕn (z (n) ) = 1, we obtain
v = yutd h,
12
ROMAIN TESSERA AND MATTHEW C. H. TOINTON
j −qz
(n)
jd−1 −qz
(n)
(n)
(n)
(n)
d−1
where y = u11 1 . . . ud−1
. In particular, v lies at distance at most |t|/Kd +1 ≤ |zd |/Kd +
1 ≤ rn + 1 from y. On the other hand, y lies in the ball of radius
(n)
max |ji |/Ki
i
(n)
(n)
+ q max |zi |/Ki
i
(n)
(n)
= r + q|zd |/Kd
≤ r + rn ,
which proves the claim. This implies that the inclusion (Λn , Wn ) ֒→ (Γn , Vn ) is a 1-Lipschitz
(1, oω (mn ))-quasi-isometry. We conclude by applying the induction hypothesis to (Λn , Wn ).
Proof of Theorem 1.5. Let kn , Hn , Pn and dn be as given by Lemma 4.3 for all but finitely many
(n)
n, and note that by Lemma 4.4 we may assume that for every such n we have Li ≥ω kn for every
(n)
(n)
i and Sn ⊂ Hn P (u(n) ; L(n) /kn ). Combined with (4.1), this implies that setting Ki = Li /kn ≥ 1
for every i and writing Vn = P (u(n) , K (n) )Hn we have
(4.4)
Vnkn ≈D Snkn .
m /k
This implies in particular that the word metrics on Γn associated respectively to Sn n n and to
m /k
Vn n n are OD (1)-bi-Lipschitz equivalent. Lemma 4.5 therefore implies that (Γ, dSn /mn ) is GHprecompact, as required.
Lemma 3.9 implies that given a cluster point (X, dX ) of (Γ, dSn /mn ) we may fix a non-principal
ultrafilter ω and assume that (Γ, dSn /mn ) →ω (X, dX ). We claim that we may also assume that
Sn =ω Vn . We first consider the ultralimits. On the one hand, observe that by Corollary 3.11 the
identity map (Γ, dSn /mn ) → (Γ, dSnkn /(mn /kn )) (resp. Γ, dVn /mn ) → (Γ, dVnkn /(mn /kn ))) induces an
isometric isomorphism between the ultralimits. Moreover, since the identity (Γ, dSnkn /(mn /kn )) →
(Γ, dVnkn /(mn /kn )) is OD (1)-bi-Lipschitz by (4.4), by Corollary 3.11 it induces a bi-Lipschitz isomorphism between the ultralimits, which are both geodesic by Remark 3.12. Composing these maps,
we deduce that the identity map (Γ, dSn /mn ) → (Γ, dVn /mn ) induces a bi-Lipschitz isomorphism
between the ultralimits. Since a geodesic invariant metric on an abelian connected Lie group is
Finsler [3] (or see the proof of [2, Theorem 2.2.4.]), we deduce that if one ultralimit has a Finsler
metric then so does the other.
Second, observe that as Sn ⊂ Vn , we have |Sn | ≤ |Vn |. On the other hand, since kn = o(n), (4.4)
implies that there exists c ≫D 1 such that Vncn ⊂ Snn , which combines with Lemma 4.1 to imply
that |Vnn | ≪ |Snn |. The assumption that |Snn | ≪ nD |Sn | therefore implies that |Vnn | ≪ nD |Vn |, and
so we may assume that Sn = Vn , as claimed.
Since dn ≪D 1, Lemma 3.7 implies that there exists d ≪D 1 such that dn =ω d. Lemma 4.6
then reduces matters to the case where Sn is Ωω (mn )-proper. One consequence is that Snn ≫ω
nd |Sn |, so that d ≤ D. A second consequence is that, writing Wn = P (x, K (n) ), the morphism
limω (Zd , dWn /mn ) → limω (Γn , dSn /mn ), which is surjective by Corollary 3.11 is injective in restriction to a ball of radius Ωω (1). By the end of the proof of Lemma 4.5, we have that limω (Zd , dWn /mn )
is a normed vector space of dimension d. Hence the dimension of any cluster point of (Γn , dSn /mn )
equals d ≤ D, as required.
5. Progressions in free nilpotent groups
Throughout this paper N = Nd,s denotes the free s-nilpotent group on d generators x =
R be the Malcev
(x1 , . . . xd ), and L = (L1 , . . . , Ld ) is a tuple of positive integers. Let N R = Nd,s
completion of N , namely the free s-nilpotent Lie group of rank d, and let n be its Lie algebra.
SCALING LIMITS OF CAYLEY GRAPHS WITH POLYNOMIALLY GROWING BALLS
13
We recall that the Baker–Campbell–Hausdorff formula states that for elements X, Y in a Lie
algebra we have
(5.1)
exp(X) exp(Y ) = exp(X + Y + 21 [X, Y ] +
1
12 [X, [X, Y
]] + · · · ).
The precise values of the rationals appearing later in the series (5.1) are not important for our
arguments; all that matters is that in a nilpotent Lie group the series is finite and depends only on
the nilpotency class of the group.
For each i we set fi = log xi . Following [17, §11.1] and [6, §1], we extend f1 , . . . , fd to a list
f = f1 , . . . , fr of so-called basic commutators in the fi . We first define general commutators in
the fi recursively, starting by declaring f1 , . . . , fd themselves to be commutators. As part of the
recursive definition, we assign to each commutator α a weight vector χ(α) = (χ1 (α), . . . , χr (α)); for
f1 , . . . , fd , these weight vectors are simply given by χi (fi ) = 1 and χj (fi ) = 0 for i 6= j. Now, given
commutators α and β whose weight vectors have already been defined, the Lie bracket [α, β] is also
defined to be a commutator, with weight vector χ(α) + χ(β). We define the total weight |χ(α)| of
a commutator α to be kχ(α)k1 .
We also now declare f1 , . . . , fd to be basic commutators. Having defined the basic commutators
f1 , . . . , fm of total weight less than k, we define the basic commutators of total weight k to be
those commutators of the form [fi , fj ] with i > j and |χ(fi )| + |χ(fj )| = k, and such that if fi
is of the form [fs , ft ] then j ≥ t. We order these arbitrarily subject to the constraint that basic
commutators with the same weight vector are consecutive, and abbreviate χ(i) = χ(fi ). Note that
the arbitrariness of the order implies that the list of basic commutators is not uniquely defined. We
caution that the basic commutators in the fi are not in general equal to the logarithms of the basic
commutators in the xi . It is well known that (f1 , . . . , fr ) is a basis of n, and more generally that
the basic commutators of total weight at least i form a basis for ni , the ith term of the lower central
series of n; see [17, §11] or [18].
As in our previous paper [24], given L ∈ Rd and χ ∈ Nd we use the P
notation Lχ to denote the
χ1
χd
χ
tuple L1 · · · Ld . Thus BR (f ; L ) is the set of linear combinations of i λi fi where |λi | ≤ Lχ(i) .
We also write Lf = (Lχ(1) f1 , . . . , Lχ(r) fr ). We remark that BR (f ; Lχ ) is the continuous version of
the nilbox B(f1 , . . . , fd ; L) appearing in [6].
Lemma 5.1. For every L ∈ Nd we have [BR (f ; Lχ ), BR (f ; Lχ )] ⊂ Od,s (1)BR (f ; Lχ ).
Proof. Since BR (f ; Lχ ) = BR (Lf ; 1), we may assume that L = 1. The lemma then follows from the
fact that BR (f ; 1) is a compact neighbourhood of the origin.
Lemma 5.2. For every n ∈ N we have
R (x, L)n ≈
χ
(i) Pord
d,s exp(BR (f ; (nL) ) ≈d,s exp(BR (f ; nL)),
R
R (x, L)n ≈
(ii) Pord
d,s Pord (x, nL), and
χ
(iii) exp(BR (f ; (nL) ) ⊂ exp(BR (f ; Lχ )Od,s,n (1) .
R (x, nL) = P R (xL , n), B (f ; (nL)χ ) = B (Lf ; nχ ),
Proof. Observe that for all n ∈ N we have Pord
R
R
ord
and BR (f ; nL) = BR (Lf ; n), which reduces the the proof of these statements to the case where
L = 1. Then (iii) and the n = 1 case of (i) simply follow from the fact that all terms are compact
generating sets (explicit constants can also be obtained through the Baker–Campbell–Hausdorff
formula); the general case of (i) follows from the proof of [16, Theorem II.1]. Finally, (i) implies
χ
R
n
R (x, nL) ≈
that Pord
d,s exp(BR (f ; (nL) ) ≈d,s Pord (x, L) , which implies (ii).
14
ROMAIN TESSERA AND MATTHEW C. H. TOINTON
Proposition 5.3. For all k ∈ N we have
(5.2)
R
R
R
Pord
(x; L)k ⊂ Pord (xL ; 1)Od,s (k) Pord
(x; L)Od,s (1) ⊂ Pord (x; L)Od,s (k) Pord
(x; L)Od,s (1) .
Moreover,
(5.3)
R
Pord
(x; L)k ∩ N ⊂ Pord (x; L)Od,s (k) .
Proof. The second inclusion of (5.2) is trivial. To prove the first inclusion, note first that applying
i
the automorphism of NR that maps xL
i to xi reduces the statement to the case where Li = 1, and
so it suffices to prove that
R
R
(x; 1)Od,s (1) .
(x; 1)k ⊂ Pord (x; 1)Od,s (k) Pord
Pord
Let D be a compact subset such that N D = N R . Observe that the map φ : NR → N that sends
g ∈ NR to the unique γ ∈ N such that g ∈ γD is a left-inverse of the inclusion N → NR . Since
R (x; 1) and P
the latter is a quasi-isometry for the word metric associated to Pord
ord (x; 1), we deduce
k
R
that φ itself is a quasi-isometry. In particular, the ball Pord (x; 1) of radius k must me contained in
the preimage by φ of a ball of radius Od,s (k). This implies that
R
(x; 1)k ⊂ Pord (x; 1)Od,s (k) D,
Pord
R (x; 1) generates N R . The inclusion
and so the first inclusion of (5.2) follows from the fact that Pord
R
(5.3) then follows from the fact that Pord (x; L) ∩ N = Pord (x; L).
Note that ordered progressions are “almost” symmetric in the sense that for group elements
x1 , . . . , xd and integers L1 , . . . , Ld the ordered progression Pord (x, L) satisfies
(5.4)
Pord (x, L)−1 ⊂ Pord (x, L)d
(a similar statement holds for real ordered progressions). We deduce the following statement.
R (x; L) ∪ P R (x; L)−1 Then for
Corollary 5.4. Denote P = Pord (x; L) ∪ Pord (x; L)−1 and PR = Pord
ord
all k ∈ N we have
O (1)
P k ⊂ PRk ⊂ P Od,s (k) PR d,s .
In particular, the inclusion (N, P ) → (NR , PR ) is a (Od,s (1), Od,s (1))-quasi-isometry.
Let us also note the equally trivial fact that for m ∈ N we have
(5.5)
Pord (x, mL) ⊂ Pord (x, L)dm .
We end this section with an application of Proposition 5.3. Breuillard and the second author
[9] have shown that doubling of a Cayley graph at some sufficiently large scale implies uniform
doubling at all subsequent scales, as follows.
Theorem 5.5 ([9, Theorem 1.1]). For every K ≥ 1 there exist n0 = n0 (K) ∈ N and θ(K) ≥ 1,
such that if S is a finite symmetric set inside some group, and if there exists n ≥ n0 for which
(5.6)
|S 2n+1 | ≤ K|S n |,
then for every m ≥ n and every c ∈ N we have |S cm | ≤ θ(K)c |S m |.
Using the main result of [8], they reduce Theorem 5.5 to the following statement.
SCALING LIMITS OF CAYLEY GRAPHS WITH POLYNOMIALLY GROWING BALLS
15
Proposition 5.6 ([9, Theorem 2.8]). Let K ≥ 1 and s ≥ 1. Suppose that S is K-approximate
subgroup in some s-step nilpotent group G. Then for every m ≥ 1 the set S m is an OK,s (1)approximate subgroup.
The proof of Proposition 5.6 in [9] requires some fairly involved computations with nilprogressions.
We take the opportunity here to note that Proposition 5.3 yields an alternative proof of Proposition
5.6. We make use of the following well-known lemma, the proof of which is left as an exercise.
Lemma 5.7. Let r0 ≥ 1, j ∈ N and φ : (X, dX ) → (Y, dY ) be a (C, K)-quasi-isometry between two
metric spaces, and assume that for all all r ≥ r0 , all balls of radius 2r in Y are covered by at most
j balls of radius r. Then there exists j ′ = Oj,C,K (1) and r0′ = Oj,C,K (r0 ) such that for all r ′ ≥ r0′ ,
all balls of radius 2r ′ in X are covered by at most j ′ balls of radius r ′ .
Lemma 5.8. We keep the notation of Corollary 5.4. For all r ≥ 1 there exists X ⊂ NR of
cardinality Os,d (1) such that
PR2r ⊂ XPRr .
i
Proof. Applying the automorphism of NR that maps xL
i to xi reduces the statement to the case
where Li = 1. In that case, it follows from [16, Theorem II.1] that there exists some K = Od,s (1)
such that |PR2r | ≤ K|PRr | for every r, and so the result follows from Lemma 3.5.
Proof of Proposition 5.6. By the first paragraph of the proof of [9, Theorem 2.8], it is sufficient to
find a positive integer M0 depending only on K and s such that S M0 m is an OK,s (1)-approximate
subgroup for all m ≥ 1.
We may assume that S generates G, in which case it follows from [25, Theorem 1.5] that there
exists a normal subgroup H ⊳ G, as well as u1 , . . . , ud ∈ G and L1 , . . . , Ld ∈ N with d ≤ K Os (1)
such that
(5.7)
S ⊂ H(Pord (u; L) ∪ Pord (u; L)−1 ) ⊂ S K
Os (1)
.
Write N = Nd,s for the free s-step nilpotent group on d generators x1 , . . . , xd , let π : N → G be the
homomorphism taking xi to ui for every i, and write P = Pord (x; L) ∪ Pord (x; L)−1 . It follows from
Corollary 5.4, Lemma 5.7 and Lemma 5.8 that there exists r0 ≪K,s 1 such that for every r ≥ r0
there exists Xr ∈ G of cardinality OK,s (1) such that π(P )2r ⊂ Xr π(P )r . It also follows from (5.7)
that there exists M0 ≪K,s 1 such that S ⊂ Hπ(P )r0 ⊂ S M0 , and hence
S 2M0 m ⊂ Hπ(P )2M0 r0 m
2M0
⊂ Xmr
Hπ(P )r0 m
0
as required.
2M0 M0 m
⊂ Xmr
S
,
0
6. Overview of the general case
In this section we give an overview of the proofs of Theorems 1.2 and 1.4. The majority of the
argument is contained in the following two results, which we prove in the next few sections.
Proposition 6.1. Let (Γn , Sn ) and (mn ) be as in Theorem 1.2. Then there exists a natural number
d ≪D 1 and, for ω-almost every n, a simply connected nilpotent Lie group Gn of dimension d and
homogeneous dimension at most D such that there is a basis e(n) of the Lie algebra of Gn and a tuple
L(n) ∈ Ndn such that (e(n) ; L(n) ) is in OD (1)-upper-triangular form. There also exists a sequence
16
ROMAIN TESSERA AND MATTHEW C. H. TOINTON
(n)
(n)
R (u(n) ; L(n) ) with u
of integers jn → ∞ such that, denoting Qn = Pord
= exp ei , the sequence
i
(Gn , dQn /jn ) is relatively compact for the GH topology, limω (Gn , dQn /jn ) is a connected nilpotent
Lie group with a geodesic metric, and
dim lim(Γn , dSn /mn ) ≤ dim lim(Gn , dQn /jn ).
ω
ω
Moreover, under the stronger assumption of Theorem 1.4, we have in addition
hdim lim(Γn , dSn /mn ) ≤ hdim lim(Gn , dQn /jn ).
ω
ω
Theorem 6.2. Let Gn be a sequence of simply connected nilpotent Lie groups of dimension d, and
let e(= e(n) ) be a basis of the Lie algebra gn of Gn , and let (jn ) be a sequence of integers tending to
(n)
(n)
∞. Suppose that for each n the convex hull Ωn of {±e1 , . . . , ±ed } satisfies [Ωn , Ωn ] ⊂ Oω (1)Ωn .
R (u; 1). Then (G , d
Write ui = exp(ei ), and let Qn = Pord
n Qn /jn ) is relatively compact for the GH
topology, limω (Gn , dQn /jn ) is a connected nilpotent Lie group with a geodesic metric,
dim lim(Gn , dQn /jn ) = d,
ω
and if hdim (Gn ) ≤ω D, then
hdim lim(Gn , dQn /jn ) ≤ D.
ω
Proof of Theorems 1.2 and 1.4. It follows from Theorem 1.1 that the sequence (Γn , dSn /mn ) is relatively compact and that every cluster point is a connected nilpotent Lie group with a left-invariant
sub-Finsler metric. Given such a cluster point (G, d), on restricting to a subsequence we may assume
d
d
that (Γn , mSnn ) → (G, d), and hence by Lemma 3.9 that (Γn , mSnn ) →ω (G, d). The theorems then follow from Proposition 6.1 and Theorem 6.2, noting that upon replacing the pairs (e(n) , L(n) ) coming
from Proposition 6.1 with (L(n) e(n) , 1) they indeed satisfy the hypotheses of Theorem 6.2.
Proposition 6.1 and Theorem 6.2 each feature a statement about the relative compactness of a
sequence (Gn , dQn /jn ) and the form of its ultralimit. Underpinning this aspect of these results is
the following statement.
Proposition 6.3. Let (jn ) be a sequence of integers going to infinity, and let (Γn , Pn ) be a sequence
of Cayley graphs such that either Γn is a finitely generated nilpotent group of bounded step and Pn
is an ordered progression of bounded rank, or Γn is a connected nilpotent Lie group of bounded step
and Pn is a real ordered progression of bounded rank. Let dPn denote the word metric associated to
Pn . Then the sequence (Γn , j1n dPn ) is relatively compact for the GH topology and limω (Γn , j1n dPn ) is
a connected nilpotent Lie group equipped with a geodesic metric.
R (u, L), where u =
Proof. We first treat the connected case. Assume therefore that Pn = Pord
(n)
(n)
(u1 , . . . , ud ) ∈ Γdn and L(n) = (L1 , . . . , Ld ) ∈ (1, ∞)d , and we assume that Γn has step at most
s. Let NR be the free connected nilpotent Lie group of step s and rank d, let x1 , . . . xr be elements
inducing a basis of NR /[NR , NR ] and let πn : NR → Γn be the canonical projection mapping xi to ui .
R (x, L(n) )).
To prove the relative compactness of (Gn , Pn ) it is then sufficient to prove it for (NR , Pord
L
(n)
Applying the automorphism of NR that maps xi i to xi reduces this statement to the case where
(n)
Li = 1. However, as in the proof of Lemma 5.8, it follows from [16, Theorem II.1] and the
R (x; 1)) has the doubling property, and so its relative
argument of [9, Lemma 2.2] that (NR , Pord
compactness is a consequence of Corollary 3.4.
SCALING LIMITS OF CAYLEY GRAPHS WITH POLYNOMIALLY GROWING BALLS
17
The ultralimit
π∞ : lim(NR , dP R
ω
ord (x;L
(n) )
/jn ) ≃ NR → lim(Γn , dP R
ω
ord (u;L
(n) )
/jn )
of the sequence πn induces a transitive action by isometries of NR on limω (Γn , dP R (u;L(n) ) /jn ).
ord
Therefore, to conclude it is enough to show that the limit is a group, or equivalently that for all
sequences gn , hn ∈ Γn such that d(gn , en ) = Oω (1), and d(hn , en ) = oω (1), then d([gn , hn ], en ) =
oω (1). Once again, up to lifting these sequences in NR , we are reduced to proving this statement
R (x, 1)), for which it follows by an easy computation based on Lemma 5.2.
for sequences in (NR , Pord
The discrete case similarly reduces to the case of N = Nr,s , the free step s nilpotent group
generated by x1 , . . . xr , equipped with the word metric associated to Pord (x; L). Reasoning as in
R (x, L(n) )) is a
the proof of Proposition 5.3, we see that the injection (N, Pord (x; L(n) )) → (NR , Pord
(O(1), O(1))-quasi-isometry, so that the proof reduces to the connected case.
7. Reduction to simply connected nilpotent Lie groups
The aim of this section is to prove Proposition 6.1. An important tool in the proof is the notion
of a Lie progression.
Definition 7.1 (Lie coset progression). Let m ∈ N ∪ {∞}, let Γ be a group, and let y1 , . . . , yd ∈ Γ.
An ordered coset progression HPord (y; L) is said to be an m-proper Lie coset progression of rank
d and homogeneous dimension D in C-upper-triangular form if there exists a connected, simply
connected nilpotent Lie group G with Lie algebra g of homogeneous dimension D, with basis
e1 , . . . , ed such that (e; L) is in C-upper-triangular form and exphe1 , . . . , ed i is a subgroup of G,
and such that, writing ui = exp ei for each i, there exists a map ϕ : hu1 , . . . , ud i → hy1 , . . . yd i that
is a homomorphism modulo H such that ϕ(ui ) = yi for each i and such that Pord (u; L) is m-proper
with respect to ϕ modulo H. We also define the injectivity radius of this Lie coset progression to be
the supremum of those j ∈ R such that ϕ is injective modulo H on Pord (u; L)j . If H is trivial then
we say simply that Pord (y; L) is an m-proper Lie progression of rank d and homogeneous dimension
D in C-upper-triangular form.
Remark 7.2. It follows from [24, Proposition 4.1] and that an m-proper Lie coset progression of
rank d in C-upper-triangular form is m-proper and in OC,d (1)-upper-triangular form as an ordered
coset progression.
Remark 7.3. Adopting the notation from Definition 7.1, it follows from Remark 7.2 and [24, Lemma
2.1] that if HP is an m-proper Lie coset progression then it has injectivity radius ΩC,d (m1/d ).
The following result essentially reduces the proof of Proposition 6.1 to the study of Lie progressions.
Theorem 7.4. Let D > 0. Let (Γn , Sn ) be a sequence of Cayley graphs such that
(7.1)
|Snn | ≪ω nD |Sn |.
Then there exist integers t ≪D 1 and a sequence of positive integers kn = oω (n) such that for
ω-almost every n there exists a Lie coset progression Hn Pn ⊂ Γn of rank at most D in OD (1)upper-triangular form with injectivity radius Ωω (n/kn ), and a finite subset Xn ⊂ Snt such that
|Xn | ≪D 1 and such that for every r ≥ 1 we have
Xn Hn Pnr ⊂ Snrkn ⊂ Xn Hn PnOD (r)
Moreover, if |Snn | ≪ω nD then ω-almost every Pn has homogeneous dimension at most D.
18
ROMAIN TESSERA AND MATTHEW C. H. TOINTON
We spend the first part of this section proving Theorem 7.4. The starting point is the following
result from our first paper.
Proposition 7.5 ([24, Proposition 1.10]). Let M, D ′ > 0, and let S be a finite symmetric generating
k
D ′ |S| for
′
set for a group G such that 1 ∈ S. Then there exists N ′ = NM,D
′ such that if |S | ≤ M k
some k ≥ N ′ then there exist integers t, η ≪D′ 1, a set X ⊂ S t such that |X| ≪D′ 1, and a 1-proper
Lie coset progression HP of rank at most OD′ (1) in OD′ (1)-upper-triangular form such that
(7.2)
for every r ∈ N.
XHP r ⊂ S rk ⊂ XHP ηr ,
Remarks on the proof. The proposition here is slightly stronger than [24, Proposition 1.10] in that
that proposition does not specify that the progression should be a Lie progression. To obtain this
extra conclusion requires two minor modifications of the proof. More precisely, it requires two minor
modifications of the proof of [24, Proposition 8.1], which is a more detailed statement from which
we deduce [24, Proposition 1.10].
The first modification is to replace [24, Theorem 1.15] with [24, Theorem 7.1] in the proof ([24,
Theorem 7.1] mentions the Lie structure explicitly, whereas this is suppressed in the statement of
[24, Theorem 1.15]). The second modification is to remove from the proof of [24, Proposition 8.1]
the stage where we delete generators that are not necessary for [24, (8.5)] to hold.
The next result allows us to modify the progression coming from Proposition 7.5 to make it
proper in some sense.
Proposition 7.6. Let (Γn ) be a sequence of groups, let η ∈ N, and suppose that for ω-almost every
n we have a finite symmetric generating set Sn for Γn , a finite subset Xn of Γn , and a Lie coset
progression Hn Pn in Γn of rank d in C-upper-triangular form such that
(7.3)
Xn Hn Pn ⊂ Snkn r ⊂ Xn Hn Pnηr
for every r ∈ N. Suppose that the injectivity radius of Hn Pn is oω (n/kn ). Then for ω-almost every n
there exists kn′ = oω (n) and a Lie coset progression Hn′ Pn′ of rank strictly less than d and injectivity
radius Ωω (n/kn′ ) in Od (1)-upper-triangular form such that
(7.4)
for every r ∈ N.
′
Xn Hn′ (Pn′ )r ⊂ Snkn r ⊂ Xn Hn′ (Pn′ )OC,d (ηr)
There are two main ingredients in our proof of Proposition 7.6. The first is one of the key results
from our previous paper, as follows.
Proposition 7.7 ([24, Proposition 7.2]). Let H0 P0 be a Lie coset progression of rank d in C-uppertriangular form that is not m-proper. Then there exists an m-proper Lie coset progression HP of
rank strictly less than d and in Od (1)-upper-triangular form such that
OC,d,m (1)
H0 P0 ⊂ HP ⊂ H0 P0
.
The second ingredient is an important property of Lie progressions in upper-triangular form,
namely that a power of such a progression is itself roughly equal to such a progression, as follows.
Proposition 7.8. Let G be a nilpotent Lie group with Lie algebra g with basis e1 , . . . , ed such that
Λ = he1 , . . . , ed i satisfies [Λ, Λ] ⊂ Λ, and let L1 , . . . , Ld ∈ N. Then for each k ∈ N there exists a
SCALING LIMITS OF CAYLEY GRAPHS WITH POLYNOMIALLY GROWING BALLS
19
basis e′1 , . . . , e′d for g and L′1 , . . . , L′d ∈ N such that (e′ , L′ ) is in 1-upper triangular form and such
that, writing ui = exp ei and u′i = exp e′i for each i, we have
Pord (u; L)k ≈d Pord (u′ , L′ ).
In proving Proposition 7.8 we make use of the following lemma. We recall from our previous
paper that we say a convex body Ω in a vector space V is strictly thick with respect to a lattice Λ
if there exists some λ < 1 such that λΩ ∩ Λ generates Λ.
Lemma 7.9. In the set-up of Proposition 7.8, for all k ∈ N there exists a symmetric convex body
R (u; L)k ≈ exp(Ω).
Ω ⊂ g, strictly thick with respect to he1 , . . . , ed i, such that [Ω, Ω] ⊂ Ω and Pord
d
Proof. Note that the step s of G satisfies s ≪d 1, so that any bound that in principle depends on
d and s may in fact be taken to depend only on d. Write π : N R → G for the homomorphism
mapping xi to ui for each i, write πn : n → g for the corresponding homomorphism of Lie algebras,
write fi = log xi for each i, and extend (fi ) to a list of basic commutators as in Section 5. It follows
R (x; L)k ≈
from Lemmas 5.1 and 5.2 that [BR (f ; (kL)χ ), BR (f ; (kL)χ )] ⊂ Od (1)BR (f ; (kL)χ ) and Pord
χ
exp(BR (f ; (kL) )). The same argument used in the proof of [24, Proposition 5.1] to put the tuple
(e, L) into 1-upper-triangular form then implies that there exists some L′ ∈ Zr satisfying (kL)χ(i) ≤
L′i ≤ Od ((kL)χ(i) ) such that
[BR (f ; L′ ), BR (f ; L′ )] ⊂ BR (f ; L′ ),
and then Lemma 5.2 (iii) implies that exp(BR (f ; L′ )) ≈d exp(BR (f ; (kL)χ )). The desired conclusion
therefore follows by taking Ω = πn (BR (f ; L′ )).
Proof of Proposition 7.8. Let Ω be given by Lemma 7.9, so that
(7.5)
R
(u; L)k ≈d exp(Ω).
Pord
By [24, Proposition 5.1] there exists a basis e′ of the lattice he1 , . . . , ed i ⊂ g, and L′ ∈ Zd such that
(e′ ; L′ ) is in 1-upper-triangular form and such that
Ω ⊂ BR (e′ ; L′ ) ⊂ Od (1)Ω.
The Baker–Campbell–Hausdorff formula then implies that
(7.6)
whilst Lemma 5.2 (i) implies that
(7.7)
exp(Ω) ≈d exp(BR (e′ ; L′ )),
R
exp(BR (e′ ; L′ )) ≈d Pord
(u′ , L′ ).
Proposition 7.8 then follows from (7.5), (7.6) and (7.7).
Proof of Proposition 7.6. Since a Lie coset progression of rank 0 has infinite injectivity radius, by
induction on d it suffices to prove the proposition without the conclusion that Hn′ Pn′ has injectivity
radius Ωω (n/kn′ ).
The fact that Hn Pn has injectivity radius oω (n/kn ) implies in particular that if we define jn to
jn
be the minimum integer such that ϕ−1
n (Hn ) ∩ P̃n 6= {e} then
(7.8)
jn = oω (n/kn ).
Proposition 7.8 and Remark 7.3 then imply that there exist Lie coset progressions Hn Qn of rank d
in OC,d (1)-upper triangular form such that
(7.9)
Hn Qn ≈C,d Hn Pnjn
20
ROMAIN TESSERA AND MATTHEW C. H. TOINTON
for ω-almost every n, and ξ > 0 depending only on C and d such that for ω-almost every n the Lie
coset progression Hn Qn is not ξ-proper. Proposition 7.7 therefore implies that there exist Lie coset
progressions Hn′ Pn′ of rank strictly less than d and in Od (1)-upper-triangular form such that
Hn′ Pn′ ≈C,d Hn Qn
(7.10)
for ω-almost every n. It then follows from (7.3), (7.9) and (7.10) that (7.4) holds for some kn′ ≪ω
jn kn , which by (7.8) is in oω (n), as required.
Lemma 7.10. Let d, η, k, t be positive integers. Let S be a finite symmetric generating subset for a
group Γ, let HP ⊂ Γ be a 1-proper Lie coset progression of rank at most d, and let X ⊂ S t be such
that
XHP r ⊂ S rk ⊂ XHP ηr
for every r ∈ N. Then, writing T = S 3t ∩ HP (d+1)η , we have
HP ⊂ (S 2t+1 ∩ HP (d+1)η )k ⊂ T k .
In particular, for all r ≥ 1 we have
(7.11)
S rk ⊂ S t T ηrk .
Proof. Let z ∈ HP . Since HP ⊂ S k , we can write z = sk . . . s1 , with si ∈ S. Let g0 = 1, and for
every 1 ≤ i ≤ k write gi = si si−1 . . . s1 , observing that therefore z = gk . Hence for all 1 ≤ i ≤ k
−1
there exists xi ∈ X and zi ∈ HP η such that gi = xi zi . We therefore have that zi zi−1
= xi si x−1
i−1 ,
2t+1
(d+1)η
which belongs to S
∩ HP
by (5.4).
Lemma 7.11. Under the assumptions of Lemma 7.10, and writing q = max{3tη, (d+ 1)η}, we have
|X|
T k ⊂ HP (d+3) q .
Proof. Since T k ⊂ S 3tk by definition, we have
(7.12)
T k ⊂ XHP 3tη .
We first claim that, writing q = max{3tη, (d + 1)η}, there exist X ′ ⊂ X and m ∈ N satisfying
q ≤ m ≤ (d + 3)|X| q such that
(7.13)
XHP 3tη ⊂ X ′ HP m
and
(7.14)
xHP m ∩ HP 2m = ∅
for every x ∈ X ′ \{1}.
To prove this claim, we first take X ′ = X and m = q. These choices certainly satsify (7.13).
If they also satisfy (7.14) then the claim is proved. If not, there exists x ∈ X ′ \{1} such that
xHP m ∩ HP 2m 6= ∅, which by (5.4) means that
xHP m ⊂ HP (d+3)m .
If we replace X ′ by X ′ \{x} and m by (d + 3)m then (7.13) is therefore still satisfied, and so we
check again whether (7.14) is satisfied, and repeat if necessary. This process terminates after at
most |X| steps, and so the claim is proved.
SCALING LIMITS OF CAYLEY GRAPHS WITH POLYNOMIALLY GROWING BALLS
21
We now claim that T j ⊂ HP m for j = 1, . . . , k, which of course implies in particular that
|X|
⊂ HP (d+3) q , as required. The case j = 1 follows from the definition of T and the fact that
m ≥ q, so by induction we may assume that T j−1 ⊂ HP m . This then implies that
Tk
T j ⊂ HP m+(d+1)η ∩ X ′ HP m
⊂ HP
m
(by (7.12) and (7.13))
(by definition of X ′ and m),
and the claim is proved.
In order to use the growth bounds in the hypothesis of Theorem 7.4, we use two lower bounds
on the growth of sets in nilpotent groups of given dimension. The first of these gives a lower bound
on the growth in terms of the homogeneous dimension, as follows.
Lemma 7.12. Let Λ be a torsion-free nilpotent group of rank r and step s, and let G be its Mal’cev
completion. Then for every symmetric generating subset Σ of Λ we have
for n ∈ N.
|Σn | ≫s,r nhdim G
±1
Proof. Note that Σ contains a set of the form (s±1
1 , . . . , sr ), where the si generate Λ/[Λ, Λ]. Let
s
q = dim C (Λ) ⊗ Q. As iterated commutators ci of length s form a generating set of C s (Λ) ⊗ Q, we
can extract a basis (ci1 , . . . , ciq ). Let ci = [sj1 , . . . , sjs ]. For all (k1 , . . . , ks ) ∈ Ns , we have
[sjk11 , . . . , skjss ] = cik1 k2 ...ks .
Hence the ball of radius n contains Ωr,s (ns ) elements of the form cki . We deduce that it contains
Ωr,s(nqs ) elements of C s (Λ). Hence the lemma follows by induction on s.
The second such result gives a lower bound on the relative growth in terms of the dimension.
Proposition 7.13. Let d ∈ N, there exists c = c(d) > 0 such that the following holds. Let Λ be a
torsion-free nilpotent group with Mal’cev completion G of dimension d. Then for every symmetric
generating subset Σ of Λ we have
|Σn | ≥ cnd |Σ|
for n ∈ N.
The proof is, unsurprisingly, by induction on d. The induction step is based on the following
standard result from additive combinatorcs.
Lemma 7.14. Let A and B be finite subsets of Z. Then |A + B| ≥ |A| + |B| − 1. In particular, by
induction on n we have |nA| ≥ n(|A| − 1).
Proof. Label the elements of A as a1 < . . . < ap and the elements of B as b1 < . . . < bq , and note
that the elements
a1 + b1 , a2 + b1 , . . . , ap + b1 , ap + b2 , . . . , ap + bq
are strictly increasing, and therefore, in particular, distinct.
Proof of Proposition 7.13. Let Λ′ be a subgroup of Λ such that Λ/Λ′ ∼
= Z and such that the Mal’cev
completion of Λ′ has codimension 1 in G, and write π : Λ → Λ/Λ′ for the quotient homomorphism.
Let R be a set of representatives in Λ of the cosets of Λ′ that have non-trivial intersection with Σ,
and note that by symmetry of Σ we have Σ ⊂ R(Σ2 ∩ Λ′ ). Since |R| = |π(Σ)|, this implies that
(7.15)
|Σ| ≤ |π(Σ)||Σ2 ∩ Λ′ |.
22
ROMAIN TESSERA AND MATTHEW C. H. TOINTON
On the other hand, given n ∈ N let Rn be a set of representatives in Λ of the cosets of Λ′ that have
non-trivial intersection with Σn . Since |Rn | = |π(Σ)n |, Lemma 7.14 implies that |Rn | ≫ n|π(Σ)|,
and by induction we may assume that |(Σ2 ∩ Λ′ )n | ≫d nd−1 |Σ2 ∩ Λ′ |. Since Rn (Σ2 ∩ Λ′ )n ⊂ Σ3n ,
this combines with (7.15) to imply that
which implies the required bound.
|Σ3n | ≫d nd |Σ|,
D′
Proof of Theorem 7.4. For ω-almost every n we may apply Proposition 7.5 with
= 2D to obtain
t
integers t, η depending only on D, and for ω-almost every n a subset Xn ⊂ Sn with |Xn | ≪D 1
and a 1-proper√Lie coset progression Hn Pn of rank OD (1) in OD (1)-upper triangular form such that
r n
Xn Hn Pnr ⊂ Sn ⊂ Xn Hn Pnηr for all r. By Lemma 3.7, on passing to an appropriate subsequence
we may assume that
√ each of these Lie coset progressions has the same rank d ≪D 1. By Proposition
7.6, on replacing n with some sequence kn = oω (n) we may assume that Hn Pn has injectivity
radius Ωω (n/kn ). It remains to show that d ≤ D, and that if |Snn | ≪ω nD then the homogeneous
dimension of ω-almost every Pn is at most D.
By the definition of a Lie coset progression, for ω-almost every n we have a connected, simply
(n)
(n)
connected nilpotent Lie group Gn with Lie algebra with basis e1 , . . . , ed such that (e(n) ; L(n) )
(n)
(n)
is in C-upper-triangular form and exphe1 , . . . , ed i is a subgroup of Gn , and such that, writing
(n)
(n)
(n)
(n)
(n)
(n)
ui = exp ei for each i and Λn = hu1 , . . . , ud i, there exists a map ϕn : Λn → hy1 , . . . yd i that
(n)
(n)
is a homomorphism modulo Hn such that ϕn (ui ) = yi for each i, such that Pn = Pord (y (n) ; L(n) ),
and such that ϕn is injective on Pord (u(n) ; L(n) )Ωω (n/kn ) modulo Hn .
(d+1)η
Writing Tn = Sn3t ∩ Hn Pn
for every n where this is defined, Lemmas 7.10 and 7.11 imply
that there exists ρ ∈ N depending only on D such that
(7.16)
Hn Pn ⊂ T kn ⊂ Hn Pnρ .
The fact that the injectivity radius tends to infinity with respect to ω implies that for ω-almost
every n there exists a unique subset T̃n ⊂ Pord (u(n) ; L(n) )(d+1)η such that ϕn (T̃n )Hn = Tn Hn . The
second containment of (7.16) then implies that ϕn is injective modulo Hn in restriction to T̃nqn for
some sequence qn ≫ω n, which implies in particular that
(7.17)
|T̃nqn ||Hn | =ω |Tnqn Hn |.
o (n)
We may assume that qn ≤ω n/4t, which combined with the fact that Tn ⊂ Sn3t and Hn ⊂ Snω
implies that Tnqn Hn ⊂ω Snn , and hence
(7.18)
|Tnqn Hn | ≤ω |Snn |.
Since qn ≥ω kn , it follows from (7.16) that T̃nqn contains Pord (u(n) ; L(n) ), and hence generates Λn ,
and so Lemma 7.12 implies that
(7.19)
and Proposition 7.13 implies that
(7.20)
|T̃nqn | ≫ω nhdim Gn
|T̃nqn | ≫ω nd |T̃n |.
Finally, the fact that Sn ⊂ Xn Hn Pnη and Xn ⊂ Snt implies that Sn ⊂ Xn (Snt+1 ∩ Hn Pnη ) ⊂ Xn Tn ,
and hence
(7.21)
|Tn | ≥ |Sn |/|Xn | ≫D |Sn |.
SCALING LIMITS OF CAYLEY GRAPHS WITH POLYNOMIALLY GROWING BALLS
23
Combining these inequalities gives
|Snn | ≥ω |Tnqn Hn |
(by (7.18))
=ω |T̃nqn ||Hn |
≫ω nd |T̃n ||Hn |
≥ω nd |Tn |
(by (7.17))
(by (7.20))
(by definition of T̃n )
d
≫ω n |Sn |
(by (7.21)),
and so (7.1) implies that d ≤ω D, as required. Moreover, (7.17), (7.18) and (7.19) imply that if
|Snn | ≪ω nD then hdim Gn ≤ω D, as required.
Having reduced the proof of Proposition 6.1 to the study of Lie progressions via Theorem 7.4,
we now study these progressions a little.
Proposition 7.15. Let G be a connected, simply connected nilpotent Lie group with Lie algebra g
with basis e1 , . . . , ed , and let L1 , . . . , Ld ∈ N be such that (e; L) is in C-upper-triangular form. Write
ui = exp ei for each i, and write Λ = hu1 , . . . , ud i. Then for every k ∈ N we have
R
Pord (u, L)k ⊂ Pord
(u; L)k ∩ Λ
and
R
Pord
(u; L)k ∩ Λ ⊂ Pord (u, L)Od,C (k) .
R (u; L)n ⊂ P R (u; O
Lemma 7.16. In the setting of Proposition 7.15 we have Pord
C,d,n (L)) for every
ord
n ∈ N.
Proof. The argument of [24, Proposition 4.1] implies that (u; L) is in OC,d (1)-upper-triangular form
over R. The result then follows from using this in place of the upper-triangular form over Z in the
proof of [24, Lemma 2.1].
Proof of Proposition 7.15. The first conclusion is trivial. Write s for the nilpotency class of Λ,
noting that s ≪d 1, and write π : N = Nd,s → Λ for the homomorphism mapping xi to ui for each
i. Then note that
R
R
(u; L)Od (1)
(u; L)k ⊂ Pord (u; L)Od (k) Pord
Pord
R
(u; Od,C (L))
⊂ Pord (u; L)Od (k) Pord
(by Proposition 5.3)
(by Lemma 7.16),
and hence that
R
Pord
(u; L)k ∩ Λ ⊂ Pord (u; L)Od (k) Pord (u; Od,C (L))
⊂ Pord (u; L)Od,C (k)
(by (5.5)),
which gives the second conclusion.
Corollary 7.17. Under the assumptions of Proposition 7.15 the embedding of Cayley graphs
R
(u; L))
(Λ, Pord (u, L)) ֒→ (G, Pord
is an (Od,C (1), Od,C (1))-quasi-isometry (see Definition 1.10).
24
ROMAIN TESSERA AND MATTHEW C. H. TOINTON
Corollary 7.18. Let (Gn ) be sequence of connected, simply connected nilpotent Lie groups of di(n)
(n)
mension d. For each n write gn for the Lie algebra of Gn , and let e(n) = e1 , . . . , ed be a basis
(n)
(n)
for gn and L(n) = (L1 , . . . , Ld ) ∈ Nd a d-tuple of integers such that (e(n) ; L(n) ) is in C-upper(n)
(n)
(n)
(n)
triangular form. Write ui = exp ei for each i, and write Λn = hu1 , . . . , ud i. Let Sn be a
generating set for Λn , and suppose that (kn ) is a sequence of positive integers such that kn = o(n)
and such that Snkn ≈ω Pord (u(n) , L(n) ). Then for every
sequence (mn ) of integers such that mn ≫ n
the sequences (Λn , dSn /mn ) and
Gn ,
dP R
dS
lim Λn , n
ω
mn
ord
(u(n) ;L(n) )
mn /kn
are precompact and
∼
= lim Gn ,
dP R
ord (u
(n) ;L(n) )
!
.
mn /kn
dP R (u(n) ;L(n) )
ord
Proof. Proposition 6.3 implies precompactness of the sequence Gn ,
, which by Corolmn /kn
ω
lary 7.17 implies that of (Λn , dSn /mn ). The isomorphism of the limits then follows from Corollary
7.17.
(d+1)η
Lemma 7.19. Let (Γn , Sn ) and (mn ) be as in Theorem 1.1, and write Tn = Sn3t ∩ Hn Pn
for the ω-almost every n for which this is therefore defined by Theorem 7.4. Then the sequence
(hHn Pn i, dTn /mn ) is precompact, and for ω-almost every n there exists a surjective continuous morphism from limω (hHn Pn i, dTn /mn ) to limω (Γn , dSn /mn ), both being connected nilpotent Lie groups
(equipped with left-invariant sub-Fisnler metrics).
Proof. To see that (hHn Pn i, dTn /mn ) is precompact, simply note that |Tnn | ≤ |Sn |3t ≪ω n3tD and
apply Theorem 1.1. To prove the existence of the surjective morphism, first note that the inclusion
ψn : (hHn Pn i, dTn ) → (Γn , dSn ) is 3t-Lipschitz, whilst Lemma 7.10—specifically (7.11)—implies that
for all r ≥ 1 and u ∈ Snrkn there exists v ∈ Tn±ηrkn such that dSn (u, v) ≤ t. The desired morphism
is therefore given by Corollary 3.11.
Proof of Proposition 6.1. Applying Theorem 7.4, we obtain a sequence
(7.22)
kn = o(n),
positive integers d, t, η depending only on D, and, for ω-almost every n, a Lie coset progression
Hn Pn ⊂ Γn of rank d and injectivity radius Ωω (n/kn ) in OD (1)-upper-triangular form, and, if
|Snn | ≪ nD , of homogeneous dimension at most D. We also obtain a finite subset Xn ⊂ Snt such
that |Xn | ≪D 1 and such that for every r ≥ 1 we have
(d+1)η
Defining Tn = Sn3t ∩ Hn Pn
Xn Hn Pnr ⊂ Snrkn ⊂ Xn Hn Pnηr .
, it follows from Lemmas 7.10 and 7.11 that
Hn Pn ⊂ Tnkn ⊂ Hn PnOD (1) ,
(7.23)
and from Lemma 7.19 that (hHn Pn i, dTn /mn ) is precompact and
(7.24)
dim lim(Γn , dSn /mn ) ≤ dim lim(hHn Pn i, dTn /mn )
ω
ω
and
(7.25)
hdim lim(Γn , dSn /mn ) ≤ hdim lim(hHn Pn i, dTn /mn ).
ω
ω
SCALING LIMITS OF CAYLEY GRAPHS WITH POLYNOMIALLY GROWING BALLS
25
By the definition of a Lie coset progression, for each n for which Pn is defined there exists a
connected, simply connected nilpotent Lie group Gn —of homogeneous dimension at most D if
(n)
(n)
(n)
(n)
|Snn | ≪ nD —with Lie algebra gn with basis e1 , . . . , ed and positive integers L1 , . . . , Ld such
(n)
(n)
that (e(n) ; L(n) ) is in OD (1)-upper-triangular form, such that exphe1 , . . . , ed i is a subgroup of
(n)
(n)
(n)
(n)
Gn , and such that, writing ui = ei and Λn = hu1 , . . . , ud i for each i, there exists a map
(n)
(n)
(n)
(n)
ϕn : Λn → hy1 , . . . yd i that is a homomorphism modulo Hn such that ϕn (ui ) = yi for each i
R (u(n) , L(n) ), we claim that (G , Q ) satisfies
and such that Pn = Pord (y (n) ; L(n) ). Writing Qn = Pord
n
n
o(n)
′
the proposition with jn = mn /kn . It follows from (7.22) that Hn ⊂ Tn , writing Γn = hHn Pn i/Hn ,
and writing ψn : hHn Pn i → Γ′n for the quotient homomorphism, we have
(7.26)
lim(hHn Pn i, dTn /mn ) ∼
= lim(Γ′n , dψn (Tn ) /mn ).
ω
ω
Since the injectivity radius of Hn Pn tends to infinity with respect to ω, and since ψn (Tn ) ⊂
O (1)
ψn (Pn D ), for ω-almost every n there exists a unique subset T̃n ⊂ Pord (u(n) ; L(n) )OD (1) such
that ψn ◦ ϕn (T̃n ) = ψn (Tn ). It also follows from (7.23) and the increasing injectivity radius that for
ω-almost every n we have
T̃nkn ≈D Pord (u(n) ; L(n) ),
and so Corollary 7.18 implies that taking jn = mn /kn , which tends to infinity by (7.22), the
sequences (Λn , dT̃n /mn ) and (Gn , dQn /jn ) are both precompact and
lim(Λn , dT̃n /mn ) ∼
= lim(Gn , dQn /jn ).
(7.27)
ω
ω
Since (Γ′n , dψn (Tn ) /mn ) is the image of (Λn , dT̃n /mn ) under ψn ◦ ϕn , we have
(7.28)
dim lim(Γ′n , dψn (Tn ) /mn ) ≤ dim lim(Λn , dT̃n /mn )
ω
ω
and
(7.29)
hdim lim(Γ′n , dψn (Tn ) /mn ) ≤ hdim lim(Λn , dT̃n /mn ).
ω
ω
It therefore follows from combining (7.24), (7.25), (7.26), (7.27), (7.28) and (7.29) that (Gn , Qn )
satisfies the proposition, as claimed.
We close this section by noting that the proof of Theorem 7.4 actually gives a more precise result
that we record here for potential future use.
Proposition 7.20. Let M, D > 0. There exists N = NM,D ∈ N such that whenever n ≥ N and
(Γ, S) is a Cayley graph such that |S n | ≤ M nD |S|, the following holds. For every δ > 0 there exist
j = OD,δ (1), an increasing sequence nδ = k1 < k2 < . . . < kj < kj+1 = n such that for every
i = 1 . . . j, there exists a Lie coset progression Hi Pi ⊂ Γ of rank at most OD,δ (1) in OD,δ (1)-uppertriangular form, and a finite subset Xi ⊂ S OD,δ (1) such that |Xi | ≪D,δ 1 and such that
• for all r ≥ 1, Xi Hi Pir ⊂ Siki r ⊂ XHP OD,δ (r) ;
• Hi Pi has injectivity radius ΩD,δ (ki+1 /ki ).
26
ROMAIN TESSERA AND MATTHEW C. H. TOINTON
8. Reduction to ultralimits of normed Lie algebras
In this section we give an overview of the proof of Theorem 6.2. The basic strategy is to bound
the homogeneous dimension (resp. the dimension) of the ultralimit by considering its Lie algebra as
the ultralimit of the normed Lie algebras (gn , k · kn ), where k · kn is a suitable sequence of norms.
There are three main steps. The first step is to define the ultralimit of the (gn , k · kn ), as follows.
Before being a Lie algebra, or even a real vector space, (gn , k · kn ) is an abelian group equipped with
an invariant distance dn (associated to its norm). Hence we can define (g∞ , d∞ ) as the ultralimit
limω (gn , dn ) as in Definition 3.8. It is easy to see that g∞ naturally comes with a real vector space
structure, and that d∞ is associated to a norm k · k∞ . Note moreover that if the dimension of gn
is in Oω (1), then the dimension of g∞ is finite and actually equals limω dim(gn ). Indeed, up to
passing to a subsequence, we can assume that the dimension is fixed, but then all the (gn , k · kn ) are
uniformly bi-Lipschitz equivalent to one another, and hence to g∞ , which therefore has the same
dimension.
In order for the Lie bracket to converge as well along the ultrafilter, we need k · kn to be in
some way compatible with the Lie bracket. This compatibility condition turns out to translate
geometrically into a “triangular form” property for the unit ball, as follows.
Lemma 8.1. Let (gn , k·kn ) be a sequence of normed Lie algebras of bounded dimension, and suppose
that for ω-almost every n, the unit ball Bn = Bk·kn (0, 1) satisfies [Bn , Bn ] ⊂ Oω (1)Bn . Then the
bracket operation on g∞
[lim un , lim vn ] = lim[un , vn ]
ω
ω
ω
is well defined and makes g∞ into a Lie algebra. Moreover, if the Lie algebras gn are nilpotent, then
so is g∞ .
The next step is to show that the Lie algebras gn appearing in Theorem 6.2 satisfy the conditions
of Lemma 8.1, as follows.
Lemma 8.2. Under the assumptions of Theorem 6.2 there exists a sequence of symmetric convex
bodies Bn ⊂ gn such that, denoting k · kn the norm whose unit ball is Bn ,
(i) [Bn , Bn ] ⊂ Bn ;
O(j )
(ii) for every sequence gn ∈ Gn we have gn ∈ Qn n if and only if k log gn kn = O(1);
o(j )
(iii) for every sequence gn ∈ Gn we have gn ∈ Qn n if and only if k log gn kn = o(1).
The final step is to bound the homogeneous dimension of the ultralimit Lie algebra, as follows.
Theorem 8.3. Let (gn , k · kBn ) be a sequence of normed nilpotent Lie algebras of dimension d such
that the unit balls Bn satisfy [Bn , Bn ] ⊂ O(1)Bn . Then
hdim (g∞ ) ≤ lim hdim (gn ).
ω
The proof of Theorem 8.3 is somewhat involved, and so we defer it until the next section. The
proofs of Lemmas 8.1 and 8.2 are more straightforward, and we present them shortly. First, however,
let us put all of these results together to prove Theorem 6.2.
Proof of Theorem 6.2. Lemma 8.2 (i) implies that (g∞ , k · kn ) satisfies the assumptions of Lemma
8.1, making g∞ a Lie algebra. Then Lemma 8.2 (ii) and (iii) together with the Baker–Campbell–
Hausdorff formula imply that exp and log commute with the ultralimits, and so g∞ canonically
identifies with the Lie algebra of limω (Gn , dQn /jn ) (which exists and is a connected nilpotent Lie
SCALING LIMITS OF CAYLEY GRAPHS WITH POLYNOMIALLY GROWING BALLS
27
group equipped with a geodesic metric by Proposition 6.3). The desired bound on the homogeneous
dimension therefore follows from Theorem 8.3 and the fact that hdim (Gn ) = hdim (gn ).
Note that the dimension is GH-continuous among normed vector spaces, and therefore that
dim limω (Gn , dQn /jn ) = d, as required. Indeed vector spaces of dimension d are Od (1)-bilipschitz
equivalent to the euclidean space of dimension d. Therefore given a sequence of vector spaces of
dimension d, its ultralimit is bilipschitz equivalent to the ultralimit of the constant sequence equal
to the euclidean space of dimension d.
Proof of Lemma 8.1. The assumption that [Bn , Bn ] ⊂ O(1)Bn implies that the set of sequences
un such that kun kBn =ω o(1) is an ideal in the Lie algebra {un : kun kBn =ω O(1)}, and this in
turn implies that the Lie bracket is well defined, as claimed. Assume that the Lie algebras gn are
nilpotent. Since their dimension is bounded, this implies that they are s-nilpotent for some s ∈ N.
This clearly implies that g∞ is s-nilpotent as well.
Proof of Lemma 8.2. Let N = Nd,d be the free d-step nilpotent Lie group of rank d, let n be its Lie
algebra, let f = (f1 , . . . , fd ) be a basis of a complement of [n, n], and let xi = exp(fi ). We consider
R (u, L) = π (P R (x, L).
the projection πn : N → Gn mapping x to u. As we already noted, Qn = Pord
n ord
Now the lemma follows from Lemma 5.2. Indeed, let Ωn = B(f, (jn L)χ ), and let Bn = πn (Ωn ). The
first statements is an obvious consequence of Lemma 5.2, and moreover we deduce that exp(Bn ) ≈
o(j )
Qjnn . The proofs of (ii) and (iii) being similar, we focus on (iii). Let gn ∈ Gn be such that gn ∈ Qn n .
R (x, L)o(jn ) be such that π(g̃ ) = g . By Lemma 5.2, log g̃ ∈ B(f, (o(j L))χ ). we have
Let g˜n ∈ Pord
n
n
n
n
that log g̃n ∈ Ωo(jn ) ⊂ o(1)Ωn . Hence projecting back to gn , we see that log gn ∈ o(1)Bn , or in
other words that k log gn kBn = o(1). Conversely, assume that wn ∈ o(1)Bn , that we lift via πn to
an element w˜n ∈ o(1)Ωn . Multilinearity of the basic commutators implies that for all 0 ≤ ε ≤ 1,
we have εB(f, (jn L)χ ) ⊂ B(f, (ε1/d jn L)χ ). Hence w˜n ∈ B(f, (o(jn )L)χ ), which by Lemma 5.2,
R (x, L)o(jn ) , from which we deduce that g := exp(w ) ∈ Qo(jn ) , as
implies that g̃n := exp(w̃n ) ∈ Pord
n
n
n
required.
9. Marked Lie algebras
In this section we prove Theorem 8.3. We fix s, d ∈ N, let n = ns,d be the free s-nilpotent Lie
algebra of rank d, and let x = (x1 , . . . , xd ) be a basis of a complement of [n, n]. The basic objects in
consideration in this section are pairs (g, π), where g is an s-nilpotent Lie algebra of rank at most
d, and π : n → g is a surjective morphism. Such a pair is equivalent to a pair (g, e), where e is a
family of d vectors of g that generate g as a Lie algebra. We call such a pair a marked Lie algebra.
An isomorphism of marked Lie algebras (g, π) → (g′ , π ′ ) is an isomorphism between g and g′ that
commutes with π and π ′ . Alternatively, an isomorphism of marked Lie algebras (g, e) → (g′ , e′ ) is
an isomorphism between g and g′ that maps e to e′ . Isomorphism classes of marked Lie algebras
are in one-to-one correspondence with ideals of n.
P
WePdefine a relation in (g, e) toPbe a linear combination i λi ei of the basic commutators ei such
that i |λi | = 1, and satisfying i λi ei = 0. Alternatively, it is an element of u ∈ ker π satisfying
kvkΩ = 1 for the norm whose unit ball Ω is the convex hull of the ±ei .
We now define a distance on the set of isomorphism classes of marked Lie algebras.
Definition 9.1. Given ε > 0 we say that d((g, e), (g′ , e′ )) ≤ ε if for every relation u ∈ ker π of
norm 1 there exists a relation u′ ∈ ker π ′ of norm 1 such that ku − u′ kΩ ≤ ε, and for every relation
v ′ ∈ ker π ′ of norm 1 there exists a relation v ∈ ker π of norm 1 such that kv − v ′ kΩ ≤ ε.
28
ROMAIN TESSERA AND MATTHEW C. H. TOINTON
Remark 9.2. Observe that this distance coincides with the Hausdorff distance between the compact
subsets ker π ∩ ∂Ω and ker π ′ ∩ ∂Ω, where ∂Ω is the unit sphere for k · kΩ . In particular, we deduce
that the space of isomorphism classes of marked groups is compact.
Let (gn , πn ) be a sequence of marked Lie algebras of rank at most d and step at most s. We
now have two natural ways to define an ultralimit of this sequence. First, we may consider the
limit (ḡ, π̄) along ω with respect to the topology defined in Definition 9.1. Note that in this case
ḡ = n/ ker π̄, where ker π̄ denotes the subspace of n spanned by all those u that appear as ω-limits
of sequences un with un ∈ ker πn ∩ ∂Ω.
On the other hand, we may consider the ultralimit of a sequence of normed Lie algebras as in
the previous section. Indeed, since [Ω, Ω] ⊂ Ω we also have [π(Ω), π(Ω)] ⊂ π(Ω), and so Lemma 8.1
implies that g∞ is a well-defined Lie algebra. Moreover, the constant sequence (n, k · kΩ ) satisfies
limω (n, k · kΩ ) = (n, k · kΩ ), and it is then easy to verify that πn converges along ω to some surjective
morphism π∞ : n → g∞ in the sense that for every bounded sequence wn ∈ (n, k · kΩ ) we have
π∞ (wn ) = limω πn (wn ).
Conveniently, these two definitions give the same limit, as follows.
Proposition 9.3. The ultralimits (ḡ, π̄) and (g∞ , π∞ ) are isomorphic.
P (n) (n)
(n)
(n)
Proof. Let
λi ei be a sequence of relations in gn . Then since the kei kBn ≤ 1 and |λi | ≤ 1,
(n)
(n)
(∞)
the limits limω ei = e∞
= λi exist, and we have
i and limω λi
X (n) (n) X (∞) (∞)
λi ei =
λi ei ,
0 = lim
ω
(n) (n)
P
from which we deduce that
limω λi ei is a relation in g∞ .
P
(∞)
Conversely, if i λi ei is a relation in g∞ , then this means that there exists a sequence un such
P
(n)
that kun kBn =ω o(1) such that i λi ei − un = 0. Note that the condition on un means that
P (n) (n)
(n)
un = µi ei , with µi =ω o(1). Therefore we have for ω-a.e. n,
X
(n) (n)
(λi − µi )ei = 0.
i
Dividing this sum by
in gn .
P
i |λi
(n)
− µi | (which tends to 1) shows that
P
(∞)
i
λi ei
is a limit of relations
Proposition 9.3 reduces the proof of Theorem 8.3 to the following elementary (and probably
well-known) fact.
Proposition 9.4. The homogeneous dimension of marked Lie algebras is lower semicontinuous
with respect to the topology implied by Definition 9.1.
We start with two preliminary lemmas. We let g be a nilpotent Lie algebra, and write g =
C1 (g) ⊂ C2 (g) ⊂ . . . for the lower central series. For each u ∈ g we write ξ(u) for the maximal k
such that u ∈ Ck (g). The following lemma trivially follows from the fact that the Ci (g) are closed.
Lemma 9.5. Given a finite-dimensional Lie algebra g, the map u → ξ(π(u)) is upper semicontinuous.
Write Z(g) for the centre of g, and for all k write ak (g) = dim Ck (g)/Ck+1 (g).
SCALING LIMITS OF CAYLEY GRAPHS WITH POLYNOMIALLY GROWING BALLS
29
Lemma 9.6. Let u ∈ Z(g). Then for k 6= ξ(u) we have ak (g/hui) = ak (g), and and for k = ξ(u)
we have ak (g/hui) = ak (g) − 1. In particular,
hdim (g/hui) = hdim (g) − ξ(u).
Proof. These statements are obvious for k ≤ ξ(u), and so we may assume that k > ξ(u), and need
to prove that ak (g/hui) = ak (g). Write p : g → g/hui. For all j we have p(Cj (g)) = Cj (p(g)), and so
ak (p(g)) ≤ ak (g). If this inequality were strict then there would exist v ∈ Ck (g) \ Ck+1 (g) such that
v + tu ∈ Ck+1 (g) for some t 6= 0. However, this would imply that tu ∈ Ck (g), and hence u ∈ Ck (g),
contradicting the assumption that k > ξ(u).
Proof of Proposition 9.4. First, we claim that given a marked Lie algebra (g, π), we can always find
a finite set of generators of ker π (as a vector space) u1 , . . . , ul such that for all 1 ≤ i ≤ l the element
ui is central modulo the ideal generated by u1 , . . . ui−1 and such that kui kΩ = 1 (ui is a relator). To
see that, start with an element in ker π, and if it is not central in n, take a non-trivial commutator
with a unit vector of n, normalise it and repeat this procedure until we get a non-trivial central unit
vector in ker π. This will give us the first vector u1 . Then do the same replacing n by n/hu1 i, and
so on until (u1 , . . . , ul ) generates the kernel. This proves the claim.
(n)
(n)
For all n, we let (u1 , . . . , uln ) be a generating set of ker πn satisfying the claim. Note that since
(n)
ln is bounded by the dimension of n, we can assume that it is constant = l. Let ūi = limω ui .
Clearly, (ū1 , . . . , ūl ) also satisfies the above property. We claim that for all 1 ≤ i ≤ l,
(n)
(n)
hdim (n/hū1 , . . . , ūi i) ≤ lim hdim (n/hu1 , . . . , ui i).
ω
The proposition follows by applying this to i = l. This statement follows by induction on i. The
case i = 0 (corresponding to taking n all along) is obvious. And if we have proved it for i, then one
easily sees that
(n)
n/hū1 , . . . , ūi+1 i = lim n/hū1 , . . . , ūi , ui+1 i.
ω
(n)
Hence the statement follows by Lemmas 9.5 and 9.6, using that ui+1 is central in n/hū1 , . . . , ūi i.
Proof of Theorem 8.3. This follows from Propositions 9.3 and 9.4.
10. Volume growth and Hausdorff dimension of the limit
In this section we examine to what extent the Hausdorff dimension of a cluster point arising from
Theorem 1.2 can be related to the exponent D. We have seen in the introduction that nothing can
be said in general, unless the dimension of the limit equals D, in which case we shall show that it
coincides with the Hausdorff dimension (or equivalently that the limiting metric is Finsler by [4,
Theorem 13]).
We start recalling a few well-known facts about Carnot–Carathéodory metrics on simply connected nilpotent Lie groups (see for instance [5]). We often abbreviate the term ‘Carnot–Carathéodory
metric’ to simply ‘cc-metric’.
Let n be the Lie algebra of a simply connected nilpotent Lie group N , and let m be a vector
subspace complementary to [n, n] equipped with a norm k · k. A smooth path γ : [0, 1] → N is said
to be horizontal if γ(t)−1 · γ ′ (t) belongs to m for all t ∈ [0, 1]. The length of γ with respect to k · k
is then defined by
Z 1
(10.1)
l(γ) =
kγ(t)−1 · γ ′ (t)kdt.
0
30
ROMAIN TESSERA AND MATTHEW C. H. TOINTON
The Carnot–Carathéodory distance associated to k · k on N is then defined by
(10.2)
d(x, y) = inf {l(γ) : γ(0) = x, γ(1) = y},
γ
where the infimum is taken over all piecewise-horizontal paths (i.e. concatenations of finitely many
horizontal paths). Note that if N = Rm , so that m = n, then the Carnot–Carathéodory metric is
just the distance associated to the norm k · k (one easily checks that up to isometry, this distance
indeed only depends on the norm k · k).
Recall that the real Heisenberg group H(R) is defined as the matrix group
1 u w
H(R) = 0 1 v ; u, v, w ∈ R ,
0 0 1
and that the discrete Heisenberg H(Z) sits inside H(R) as the cocompact discrete subgroup consisting
of unipotent matrices with integral coefficients. The group H(R) (resp. H(Z)) is 2-step nilpotent;
indeed, its centre, which coincides with its derived subgroup, is isomorphic to R (resp. Z), and
consists in matrices whose only non-zero coefficient is the top right coefficient. It follows that
H(R)/[H(R), H(R)] ∼
= R2 (and similarly H(Z)/[H(Z), H(Z)] ∼
= Z2 ). The group H(R) comes with a
one-parameter group of automorphisms (δt )t∈R defined as follows:
1 tu t2 w
1 u w
δt 0 1 v = 0 1 tv .
0 0 1
0 0
1
Given a norm k · k on R2 , there exists a unique left-invariant Carnot–Carathéodory metric dcc on
H(R) that projects to k · k and is scaled by δt , i.e. such that dcc (e, δt (g)) = tdcc (e, g) for all t ∈ R∗+
and all g ∈ H(R). Normalise the Haar measure on H(R) so that the ball of radius 1 has volume
1. It follows from the formula defining δt that for all r > 0, the ball of radius r for this metric
has measure equal to r 4 . In particular, the Hausdorff dimension of H(R) equals 4 (see [21, Lemma
C.3]).
For every j, k, l ∈ N write
1 [−j,j]∩Z [−k,k]∩Z
1 [−j,j] [−k,k]
P (j, k) = 0
and PR (j, k) = 0 1 [−l,l] ,
1
[−l,l]∩Z
0
0
1
0
0
1
and write S(j, k) = P (j, k) ∪ P (j, k)−1 ⊂ H(Z) and SR (j, k) = PR (j, k) ∪ PR−1 (j, k) ⊂ H(R). A
straightforward calculation shows that P −1 (j, k) ⊂ P (j, j 2 + k), and so if k ≥ j 2 we have
(10.3)
P (j, k) ⊂ S(j, k) ⊂ P (j, 2k),
and similarly
PR (j, k) ⊂ SR (j, k) ⊂ PR (j, 2k).
The following lemma is an easy consequence of a celebrated result of Pansu [19] (strictly speaking,
we use [5, Theorem 1.4 ] which is a slight generalisation due to Breuillard).
Lemma 10.1. Let jn and qn be sequences of positive integers such that qn → ∞, and let Sn =
S(jn , jn2 ). Then (H(Z), dSn /qn ) GH-converges to H(R) endowed with the cc-metric associated to the
ℓ∞ -norm on R2 .
SCALING LIMITS OF CAYLEY GRAPHS WITH POLYNOMIALLY GROWING BALLS
31
Proof. We consider a relatively compact subset D ⊂ H(R) such that H(R) is the disjoint union
of left H(Z)-translates of D. For every left-invariant distance d on H(Z), define a H(Z)-invariant
pseudo distance dR on H(R), defined as dR (g, h) = d(γ, λ) where g ∈ ΓD and h ∈ λD. Note that
the embedding (H(Z), d) → (H(R), dR ) is an isometry (in particular it is essentially surjective, in
the sense that every point of H(R) lies at distance zero from a point of H(Z)). For every j ∈ N, we
denote
1 ±j 0
Pj = 0 1 ±j ,
0 0
1
Pj−1 .
and Sj = Pj ∪
Note that we have
R
dR
2 ) ≤ dS(j ,j 2 ) ≤ dSR (jn ,j 2 ) .
S(jn ,jn
n
n n
Applying the automorphism δjn−1 to H(R), we obtain
R
dR
2 ) (δj −1 (·), δj −1 (·)) ≤ dSR (1,1) .
S(1,1) ≤ dS(jn ,jn
n
n
But then it follows from [5, Theorem 1.4] that (H(R), dR
2 ) (δj −1 (·), δj −1 (·))/qn ) GH-converges
S(jn ,jn
n
n
∞
to H(R) endowed with the cc-metric associated to the ℓ -norm on R2 . Now δjn defines an isometric
R
isomorphism from (H(R), dR
2 ) (δj −1 (·), δj −1 (·))/qn ) to (H(R), dS(j ,j 2 ) /qn ). And we have seen
S(jn ,jn
n n
n
n
at the beginning of this proof that the injection (H(Z), dS(jn ,jn2 ) /qn ) → (H(R), dR
2 ) /qn ) is an
S(jn ,jn
isometry as well. So we deduce that (H(Z), dS(jn ,jn2 ) /qn ) converges to H(R) endowed with the
cc-metric associated to the ℓ∞ -norm on R2 , and the lemma follows.
Proof of Proposition 1.7. Pick a sequence an such that nan ≈ f (n). Let Sn = S(n, n3−an ). Using
(10.3), one easily checks that |Snn | ≍ n8 and |Sn | ≍ n5−an , and so |Snn |/|Sn | ≪ n3 nan ≍ f (n)n3 .
We now prove that the limit is isometric to the cc-metric associated to the ℓ∞ -norm. This
follows from Lemma 10.1 together with the fact that there exists tn going to zero such that for all
tn n ≤ k ≤ n
S(n, n3−an )k = S(n, n2 )k .
Indeed, a direct calculation (left to the reader), shows that for all positive integers i, j, m such that
(mi)2 ≥ 10mj,
S(i, j)m = S(i, i2 )m .
Therefore, in order to find tn , one has to solve in t the equation t4 n8 = 10t3 n8−an , giving t =
10n−an .
We now move on to the proof of Proposition 1.8, starting with an immediate consequence of
Proposition 7.13.
Lemma 10.2. Let Λ be a torsion-free nilpotent group with Mal’cev completion G of dimension d.
Let Σ be a symmetric finite generating set of Λ, let n ≥ 0 and C be such that
|Σn | ≤ Cnd |Σ|.
Then for every i, j with 0 < i < j ≤ n we have
c(j/i)d |Σi | ≤ |Σj | ≤ (C/c2 )(j/i)d |Σi |,
where c is the constant coming from Proposition 7.13.
32
ROMAIN TESSERA AND MATTHEW C. H. TOINTON
Proof of Proposition 1.8. We first reduce to the case where Γn has no torsion and its Mal’cev completion has dimension D. Indeed, combining Theorem 7.4 and Lemma 7.19, we are reduced to
the case where Γn is a torsion-free nilpotent group. In that case, the connected Lie group Gn of
Proposition 6.1 is simply its Mal’cev completion (this is Corollary 7.18). Finally we deduce from
Theorem 6.2 that the dimension of Gn equals the dimension of the ultralimit, i.e. D.
It then follows from Lemma 10.2 that in the ultralimit G∞ (with mn = n) we have
(c2 /C)r D |B(e, 1)| ≤ |B(e, r)| ≤ (1/c)r D |B(e, 1)|
for every r ≤ 1. This is well known to imply that the Hausdorff dimension equals D (see [21, Lemma
C.3]). Finally we deduce from [4, Theorem 13] that the metric is Finsler.
Appendix A. Detailed growth of nilprogressions
The purpose of this appendix is to prove Theorem 1.9. We essentially reproduce Tao’s original
proof of the theorem in [23, §4], but substitute in various results from the present paper and its
predecessor [24] to make the argument finitary.
Proposition 7.5 essentially reduces Theorem 1.9 to the study of the growth of Lie progressions.
The key result on that topic is the following.
Proposition A.1. Let P be a Lie progression of rank d in C-upper-triangular form for some C.
Then there exists a non-decreasing continuous piecewise-linear function f : [0, ∞) → [0, ∞) with
f (0) = 0 and at most Od (1) distinct linear pieces, each with a slope that is a natural number at
most Od (1), such that
log |P m | = log |P | + f (log m) + Od (1)
for every m ∈ N.
Remark A.2. In fact, we do not need the Lie progression P appearing in Proposition A.1 to be
in upper-triangular form; it would be enough for the lattice Λ generated in the Lie algebra by the
basis defining P to satisfy [Λ, Λ] ⊂ Λ as in Proposition 7.8.
We prove Proposition A.1 are using material from Section 7 and the following results, which are
almost identical in spirit to part of Tao’s argument in [23, §4].
Proposition A.3. Let G be a connected, simply connected nilpotent Lie group with Lie algebra g
with basis e1 , . . . , ed , and write Λ = he1 , . . . , ed i. Suppose that exp Λ is a subgroup of G. Write
ui = exp ei for each i, let L1 , . . . , Ld be positive integers, and write P = Pord (u; L). Then there
exists a polynomial f of degree Od (1) with positive coefficients such that
|P m | ≍d f (m).
Lemma A.4. Let f be a polynomial of degree k with no negative coefficients. Then there exists a
continuous, piecewise-monomial function h of increasing degree with at most k + 1 pieces, each of
which has a positive coefficient, such that
f (x) ≍k h(x)
for x > 0.
Proof. Following Tao [23, §4], write f (x) = α0 + α1 x + . . . + αk xk , and note that for every a, b > 0
we have a + b ≍ max{a, b}, so that
f (x) ≍k max αi xi .
i
We may therefore take h(x) = maxi αi
xi .
SCALING LIMITS OF CAYLEY GRAPHS WITH POLYNOMIALLY GROWING BALLS
33
Before we prove Proposition A.3, let us see how it combines with Lemma A.4 to imply Proposition
A.1.
Proof of Proposition A.1. We proceed by induction on d, taking the trivial case d = 0 as the base.
If P is infinitely proper then the result follows from Proposition A.3 and Lemma A.4. If not then
P has finite injectivity radius, say k ∈ N, and Proposition A.3 and Lemma A.4 imply that there
exists a non-decreasing continuous piecewise-linear function f : [0, ∞) → [0, ∞) with f (0) = 0 and
at most Od (1) distinct linear pieces, each with a slope that is a natural number at most Od (1), such
that
(A.1)
log |P m | = log |P | + f (log m) + Od (1)
for every m ≤ k. Let c > 0 be a constant to be determined by depending only on d. Proposition
7.8 implies that there exists a Lie progression P0 of rank d in 1-upper-triangular form and with
injectivity radius Od (1) such that
(A.2)
P ck ⊂ P0 ⊂ P Od (ck) .
Remark 7.3 implies that there exists q ≪d 1 such that P0 is not q-proper. Proposition 7.7 therefore
implies that there exists a Lie coset progression HP1 of rank strictly less than d in Od (1)-upperO (1)
triangular form such that P0 ⊂ HP1 ⊂ P0 d , and then (A.2) implies that P ck ⊂ HP1 ⊂ P Od (ck) .
We may thus choose c so that
(A.3)
Od (r)
HP1r ⊂ P rk ⊂ HP1
for every r ∈ N.
Following Tao [23, §4], write P̂ for the image of P1 in the quotient hHP1 i/H and note that
(A.4)
|HP1r | = |H||P̂ r |
for every r ∈ N. By induction we may assume that there exists a non-decreasing continuous
piecewise-linear function h : [0, ∞) → [0, ∞) with h(0) = 0 and at most Od (1) distinct linear pieces,
each with a slope that is a natural number at most Od (1), such that
log |P̂ r | = log |P̂ | + h(log r) + Od (1)
for every r ∈ N. This implies in particular that |P̂ ℓr | ≪ℓ,r |P̂ r | for every ℓ, r ∈ N, and hence,
combined with (A.3) and (A.4), that |P rk | ≍d |H||P̂ r | for every r ∈ N. It follows that
log |P rk | = log |P k | + h(log r) + Od (1)
for every r ∈ N. The monotonicity of log |P m | in m therefore implies that
m
m
k
log |P k | + h(log⌊ m
k ⌋) − Od (1) ≤ log |P | ≤ log |P | + h(log⌈ k ⌉) + Od (1)
for every m ≥ k, and so the fact that the slope of h is bounded in terms of d implies that log |P m | =
log |P k | + h(log m
k ) + Od (1) for every m ≥ k. In light of (A.1), this implies that
log |P m | = log |P | + f (log k) + h(log m − log k) + Od (1)
for every m ≥ k. Since h(0) = 0, this combines with (A.1) to give the desired result for all m.
We start our proof of Proposition A.3 with a general lemma from linear algebra, which was
essentially present in [23, §4].
34
ROMAIN TESSERA AND MATTHEW C. H. TOINTON
Lemma A.5. Let r ≥ d, x1 , . . . xr to span Rd , and M1 , . . . , Mr > 0. There exist i1 < . . . < id such
that
BR (x; M ) ⊂ r · BR (xi1 , . . . , xid ; Mi1 , . . . , Mid ).
Proof. Pick the i1 < . . . < id that maximise
vol(BR (xi1 , . . . , xid ; Mi1 , . . . , Mid ))
(where the volume is defined with respect to the canonical basis of Rd ). On reordering the xi ’s, we
can assume that ij = j for all 1 ≤ j ≤ d. It suffices to show that for a given xk we have
(A.5)
Mk xk ∈ BR (x1 , . . . , xd ; M1 , . . . , Md ).
View each xi as a column vector with coordinates with respect to the basis x1 , . . . , xd , and write A
for the d × d matrix with columns M1 x1 . . . , Md xd . Cramer’s rule implies that the solution y ∈ Rd
to the equation Ay = Mk xk satisfies
|yj | =
vol(BR (x1 , . . . , xj−1 , xk , xj+1 , . . . , xd ; M1 , . . . , Mj−1 , Mk , Mj+1 , . . . , Md ))
vol(BR (x1 , . . . , xd ; M1 , . . . , Md ))
(i.e. with xj replaced by xk in the numerator). By maximality, this implies in particular that
|yj | < 1, which gives (A.5), as required.
From now on in this section we adopt the notation of Proposition A.3. We also extend e =
(e1 , . . . , ed ) to the list e = (e1 , . . . , er ) of basic commutators, as defined in Section 5. Given a
measurable subset B of g we write vol(B) for the measure of B, normalised so that the lattice Λ
has determinant 1.
Proposition A.6. For every m ∈ N we have
|P m | ≍d vol(BR (e; (mL)χ )).
Proof. A result of van der Corput [28] states that for any convex body K ⊂ Rd we have |Zd ∩ K| ≥
1
vol(K), which in this case implies that
2d
vol(BR (e; (mL)χ )) ≤ 2d |Λ ∩ BR (e; (mL)χ )|
for every m ∈ N. Since
R
exp Λ ∩ BR (e; (mL)χ ) ⊂ exp Λ ∩ Pord
(u; L)m
⊂ P Od (m)
(by Lemma 5.2 (i))
(by Proposition 5.3),
this implies that for some constant Bd > 0 we have
(A.6)
for every m ∈ N. Since
vol(BR (e; (mL)χ )) ≪d |P Bd m |
P
vol(BR (e; (αmL)χ )) ≪ α
i
|χ(i)|
vol(BR (e; (mL)χ ))
for every α ≥ 1 and every m, on replacing m by Bd−1 m in (A.6) we obtain
vol(BR (e; (mL)χ )) ≪d |P m |,
which is one direction of what we need to prove.
SCALING LIMITS OF CAYLEY GRAPHS WITH POLYNOMIALLY GROWING BALLS
35
For the other direction, write Q = [0, 1]d and note that
R
(u; L)m
P m ⊂ exp Λ ∩ Pord
(since exp Λ is a group)
χ
⊂ exp Λ ∩ BR (e; (Od (mL)) )
and hence
(by Lemma 5.2 (i)),
|P m | ≤ |Λ ∩ BR (e; (Od (mL))χ )|
= vol
Λ ∩ BR (e; (Od (mL))χ ) + Q
≤ vol BR (e; (Od (mL))χ )
χ
≪d vol BR (e; (mL) ),
(since Q ⊂ BZ (e; L))
as required.
Lemma A.7. There exists a polynomial f of degree Od (1) with no negative coefficients such that
vol(BR (e; (mL)χ )) ≍d f (m).
Proof. We continue to follow Tao [23, §4]. Set
X
f (m) =
vol(BR (ei1 , . . . , eid ; (mL)χ )),
1≤i1 <...<id ≤r
and note that
f (m) =
X
vol(BR (ei1 , . . . , eid ; Lχ ))m
Pd
j=1
|χ(ij )|
,
1≤i1 <...<id ≤r
which is certainly of the required form. It follows from Lemma A.5 that
vol(BR (e; (mL)χ )) ≤ r d f (m),
whilst the fact that BR (ei1 , . . . , eid ; (mL)χ ) ⊂ BR (e; (mL)χ ) for every i1 , . . . , id implies that
f (m) ≤ dr vol(BR (e; (mL)χ )).
Proof of Theorem 1.9. It follows from Proposition 7.5 that there exist |X| ≪D 1, and a Lie coset
progression HP of rank at most OD (1) in OD (1)-upper-triangular form such that
(A.7)
HP m ⊂ S mn ⊂ XHP OD (m) ,
for every m ∈ N. Write P̂ for the image of P in the quotient hHP i/H, and note that |HP m | =
|H||P̂ m | for every m ∈ N. Since Proposition A.1 implies that |P̂ rm | ≪r,D |P̂ m | for every r, m ∈ N,
this combines with (A.7) to imply that
|S mn | ≍D |H||P̂ m |
for every m ∈ N, and so the theorem follows from Proposition A.1.
Remark A.8. The ineffectiveness in Theorem 1.9 arises from the single use of Proposition 7.5 at the
beginning of the proof; the ineffectiveness of Proposition 7.5 in turn arises from a single use of [8,
Theorem 1.6] (see also the remark immediately after the proof of [24, Proposition 8.3]).
36
ROMAIN TESSERA AND MATTHEW C. H. TOINTON
References
[1] H. Bass, The degree of polynomial growth of finitely generated nilpotent groups, Proc. London Math. Soc. (3)
25 (1972), 603–614.
[2] I. Benjamini, H. Finucane and R. Tessera. On the scaling limit of finite vertex transitive graphs with large
diameter. Combinatorica 36 (2016), 1–41.
[3] V. N. Berestovskiı́. Homogeneous manifolds with an intrinsic metric. I, Sibirsk. Mat. Zh. 29(6) (1988), 17–29.
[4] V. N. Berestovskiı́. Locally Compact Homogeneous Spaces with Inner Metric, J. Gen. Lie Theory Appl. 9(1)
(2015), 6 pp.
[5] E. Breuillard. Geometry of locally compact groups of polynomial growth and shape of large balls. arXiv:0704.0095.
[6] E. Breuillard and B. J. Green. Approximate groups. I. The torsion-free nilpotent case, J. Inst. Math. Jussieu
10(1) (2011), 37–57.
[7] E. Breuillard and B. J. Green. Approximate groups, II: the solvable linear case, Q. J. Math. 62(3) (2011), 513–521.
[8] E. Breuillard, B. J. Green and T. C. Tao. The structure of approximate groups, Publ. Math. IHES. 116(1) (2012),
115–221.
[9] E. Breuillard and M. C. H. Tointon. Nilprogressions and groups with moderate growth, Adv. Math. 289 (2016),
1008–1055.
[10] M. Bridson and A. Hefliger, Metric Spaces of Non-Positive Curvature, Grundl. der Math. Wiss. 319, Springer
Verlag (1999).
[11] W. W. Comfort and S. Negrepontis. The theory of ultrafilters, Berlin, New York: Springer-Verlag (1974).
[12] N. Gill and H. A. Helfgott. Growth in solvable subgroups of GLr (Z/pZ), Math. Ann. 360(1) (2014) 157–208.
[13] K. Fukaya and T. Yamaguchi. The fundamental groups of almost nonnegatively curved manifolds, Math. Ann.
136 (1992), 253–333.
[14] M. Gromov. Groups of polynomial growth and expanding maps, Publ. Math. IHES 53 (1981), 53–73.
[15] M. Gromov. Metric structures for Riemannian and non-Riemannian spaces, volume 152 of Progress in Mathematics. Birkhäuser Boston Inc., Boston, MA (1999). Based on the 1981 French original, With appendices by M.
Katz, P. Pansu and S. Semmes.
[16] Y. Guivarc’h. Croissance polynomiale et périodes des fonctions harmoniques, Bull. Sc. Math. France 101 (1973),
333–379.
[17] M. Hall. The theory of groups, Amer. Math. Soc./Chelsea, Providence, RI (1999).
[18] M. Hall. A basis for free Lie rings and higher commutators in free groups, Proc. Amer. Math. Soc. 1 (1950),
575–581.
[19] P. Pansu. Croissance des boules et des géodésiques fermées dans les nilvariétés. Ergodic Theory Dyn. Syst. 3
(1983), 415–445.
[20] J. Roe, Lectures on Coarse Geometry, Univ. Lecture Ser., vol. 31, American Mathematical Society, Providence,
RI, 2003
[21] S. Semmes, Finding curves on general spaces through quantitative topology, with applications to Sobolev and
Poincaré inequalities, Selecta Math. 2 (1996), 155–295.
[22] Y. Shalom and T. Tao, A finitary version of Gromov’s polynomial growth theorem. Geom. Funct. Anal. 20 (2010),
no. 6, 1502–1547.
[23] T. C. Tao. Inverse theorems for sets and measures of polynomial growth, Q. J. Math. 68(1) (2017), 13–57.
[24] R. Tessera and M. C. H. Tointon. Properness of nilprogressions and the persistence of polynomial growth of given
degree, arXiv:1612.05152.
[25] M. C. H. Tointon. Freiman’s theorem in an arbitrary nilpotent group, Proc. London Math. Soc. (3) 109 (2014),
318–352.
[26] M. C. H. Tointon. Approximate subgroups of residually nilpotent groups, arXiv:1509.03876.
[27] L. van den Dries and A .J. Wilkie. Gromov’s theorem on groups of polynomial growth and elementary logic, J.
Alg. 89 (1984), 349–374.
[28] J. G. van der Corput. Verallgemeinerung einer Mordellschen Beweismethode in der Geometrie der Zahlen II, Acta
Arith. 2 (1936), 145–146.
Laboratoire de Mathématiques d’Orsay, Univ. Paris-Sud, CNRS, Université Paris-Saclay, 91405 Orsay, France
E-mail address: [email protected]
SCALING LIMITS OF CAYLEY GRAPHS WITH POLYNOMIALLY GROWING BALLS
37
Insitut de Mathématiques, Université de Neuchâtel, Rue Emile-Argand 11, CH-2000 Neuchâtel,
Switzerland
E-mail address: [email protected]
| 4math.GR
|
Pseudo-Centroid Clustering
Fred Glover
Department of Electrical, Computer and Energy Engineering
School of Engineering & Science
University of Colorado
Boulder, Colorado 80309, USA
[email protected]
October 2016
Abstract
Pseudo-centroid clustering replaces the traditional concept of a centroid expressed as a center of
gravity with the notion of a pseudo-centroid (or a coordinate free centroid) which has the
advantage of applying to clustering problems where points do not have numerical coordinates (or
categorical coordinates that are translated into numerical form). Such problems, for which
classical centroids do not exist, are particularly important in social sciences, marketing,
psychology and economics, where distances are not computed from vector coordinates but rather
are expressed in terms of characteristics such as affinity relationships, psychological preferences,
advertising responses, polling data and market interactions, where distances, broadly conceived,
measure the similarity (or dissimilarity) of characteristics, functions or structures.
We formulate a K-PC algorithm analogous to a K-Means algorithm and focus on two key types
of pseudo-centroids, MinMax-centroids and (weighted) MinSum-centroids, and describe how
they, respectively, give rise to a K-MinMax algorithm and a K-MinSum algorithm which are
analogous to a K-Means algorithm. The K-PC algorithms are able to take advantage of problem
structure to identify special diversity-based and intensity-based starting methods to generate
initial pseudo-centroids and associated clusters, accompanied by theorems for the intensity-based
methods that establish their ability to obtain best clusters of a selected size from the points
available at each stage of construction. We also introduce a regret-threshold PC algorithm that
modifies the K-PC algorithm together with an associated diversification method and a new
criterion for evaluating the quality of a collection of clusters.
Keywords: Clustering – centroids – K-Means – K-Medoids – Advanced starting methods –
Metaheuristics.
Published online in Soft Computing, ( ), pp. 1-22, Springer, 13 October 2016 at
http://link.springer.com/article/10.1007/s00500-016-2369-6, content sharing version:
http://rdcu.be/k1kY.
1
1. Introduction.
Consider a set X of n points, given by X = {xr, r N = {1, …, n}}. We are interested in the
clustering problem that consists of identifying a partition of X into subsets C = C(h), h K=
{1,…,k} for specified values of k.
In this paper, we are specifically interested in clustering problems where the goal is to organize
clusters by a criterion based on the distances of points in each cluster from a common (suitably
defined) “central point,” but where the points lack coordinates that make it possible to draw on
the notion of a classical centroid as used in methods such as the K-Means algorithm. We
emphasize that we interpret “distance” in a broad sense to encompass any value assigned to a
pair of points that expresses their proximity or similarity. We may think of distance as a type of
repulsion, for example, where larger distances reflect a reduced desirability of placing the two
associated points in a common cluster. We also allow consideration of negative distances, which
may be viewed as a form of attraction.
To handle situations where distances are not restricted to constitute a metric based on spatial
coordinates of a vector, we introduce the notion of a pseudo-centroid (PC) to replace that of a
classical centroid. This enables us to describe K-PC algorithms by analogy with the K-Means
algorithm for applications in which points xr to be clustered include non-numerical coordinates
(or lack coordinates altogether). Such problems occur prominently, for example, in the social
media, marketing and economics areas, where distances, broadly conceived, may represent
affinity relationships, psychological preferences, advertising responses, polling data, market
interactions and so forth.
We establish the background and general ideas underlying pseudo-centroid clustering, and focus
particularly on two forms of this clustering we call K-MinMax and K-MinSum clustering. The
MinMax and MinSum-centroids make it possible to generate starting clusters for the K-MinMax
and K-MinSum clustering algorithms that embody a meaningful structure in relation to the goal
of producing a best collection of final clusters and which apply as well to other definitions of
pseudo-centroids discussed in Section 9. We demonstrate this by providing advanced starting
procedures for these clustering algorithms of two types, called diversity-based and intensitybased methods, which also offer ways to aid in determining good values of k and provide
variants that can be used with K-Means algorithms. We also including special adaptive versions
of the intensity-based methods designed to reduce the number of iterations subsequently
performed by K-PC algorithms and to increase the likelihood of yielding final clusters that are
the best achievable. The intensity-based methods are accompanied by theorems that identify the
quality of clusters they produce.
2. Background and related research
A variety of papers have undertaken to develop measures to replace Euclidean distances, or in
some cases to transform non-numerical coordinates (such as those for categorical variables) into
numerical coordinates to permit Euclidean distances to be calculated. For instance,
Ralambondrainy (1995) converts multiple category attributes into binary attributes (1 if a
2
category is present, 0 if absent) to treat binary attributes as numeric in the K-Means algorithm
(MacQueen, 1967). This approach needs to handle a large number of binary variables if used in
data mining, where categorical attributes often involve hundreds or thousands of categories.
Huang (1998) goes farther by developing ways to handle categorical variables utilizing a simple
dissimilarity measure for categorical objects proposed by (Kaufman and Rousseeuw, 1990) and
applies a variant of the K-Means algorithm using modes instead of means for clusters, together
with a frequency-based method to update modes to minimize the clustering cost function. The
foregoing approaches continue to rely on coordinates, however, rather than directly making use
of distance measures to characterize a notion of centrality without the need to refer to
coordinates.
Even in situations where Euclidean distances exist naturally and provide the possibility for
creating centroids, a reliance on such centroids is not always desirable. For example, as
illustrated in Cao and Glover (2010), centroids can create difficulties for problems where
communications or travel between points must take place within a specified network, since
centroids may lie in unreachable areas such as rivers or fields that cannot be traversed, or may
compel some points that lie in a centroid-based cluster to cross such areas to reach other cluster
points, thus linking points that are not desirable to be grouped within a cluster together.
On the other hand, there are many methods that rely on distances without making use of
centroids. Such methods avoid the problems that typically arise by centroid-based methods and
have a venerable tradition, although at the cost of abandoning an effort to group points with
reference to a point that satisfies a meaningful alternative definition of centrality. For example,
Fan (2009) describes a distance measure that modifies Euclidean distances to incorporate an
“obstructed distance” between two locations if their straight line link is intersected by geographic
obstacles such as rivers, mountains or highways. He then undertakes to use these distances by a
method that uses a simulated annealing approach to correct distortions introduced by an attempt
to use a K-Means algorithm. But no measure is offered to replace a centroid-based measure as
relied on by the K-Means procedure.
In a related setting, geographic information system (GIS) technology is often utilized due to its
ability to supply vital information such as geographic feature data, street network information,
speed limits on street segment and lengths of street segments. This technology is also able to
keep track of restrictions such as vehicle heights, weights and volumes that need to be considered
by optimization procedures. For example, Estivill-Castro and Lee (2001) combine data mining
and GIS as a means to consider geographic obstacles such as hills or rivers. The authors devise a
clustering algorithm utilizing a Voronoi diagram to set up a topological structure for a set of
points as a basis for retrieving spatial information related to various definitions of neighbors.
Once again, having established the topological structure, no way of representing centrality is
provided. Another use of a Voronoi diagram to yield a more appropriate space is given in Kwon
et al. (2007) as a basis for a tabu search method for solving capacitated vehicle routing problems.
However, relying on the Voronoi topology by itself without a new concept of central point was
not sufficient to allow the approach to beat the existing benchmark results.
3
A different type of application is addressed by Strehl and Ghosh (2002), who develop a
clustering algorithm for data mining problems found in the retail industry and some Web
applications, where data reside in a very high-dimensional space. This approach introduces a
similarity relationship defined on each pair of data samples and transforms the problem into one
over the similarity domain so that the original high-dimensional space is no longer needed. The
goal is to cluster data samples into k groups so that data samples for different clusters have
similar characteristics. Lacking a meaningful centrality measure, the authors formulated the
clustering problem as an NP-hard vertex-weighted graph partitioning problem and developed an
algorithm called Optimal Partitioning of Space Similarities Using Metis (OPOSSUM) using the
Metis approach proposed by Karypis and Kumar (1998) as the multi-objective graph partitioning
engine. Another recourse to a more complex model representation to cluster objects in the
absence of an ability to take advantage of central points occurs in Kochenberger et al. (2005),
who use clique partitioning to cluster microarray data.
A type of approach that is related to one of the types of pseudo-centroid methods described in
this paper is the K-Medoid algorithm embodied in the Partitioning Around Medoids (PAM)
algorithm of Kaufman and Rousseeuw (1990), and in its extended variant Clustering Large
Applications CLARA). Ng and Han (2002) further extend this approach with a method called
CLARANS (Clustering Large Applications based upon RANdomized Search). Specifically, our
definition of a MinMax-centroid corresponds to one of the definitions of a medoid, but the
medoid definition incorporated into the foregoing K-Medoid algorithms has a different
foundation that stems from a coordinate-centric perspective (seeking to minimize the maximum
deviation of cluster points from a classical centroid). Similarly, another treatment of medoids in
Park and Jun (2009) employs a definition corresponding to our definition of a MinSum pseudocentroid, but again resorts to a coordinate-centric perspective for its implementation and lacks
the generality that makes the pseudo-centroid framework applicable to other definitions as well.
Because of these differences, the K-Medoid algorithms cited above operate differently from the
K-MinMax and K-MinSum algorithms and fail to gain access to associated starting algorithms
and their associated theorems described in subsequent sections.
It is important to point out that not all forms of clustering can benefit from exploiting a measure
of centrality. A reliance on centrality is purposely missing in the case of the spanning tree
methods which provide an ability to generate clusters that may be embedded in others (see, e.g.
Xu et al., 2001; Paivinen, 2005; Shamsul, et al., 2011). Similarly, centrality measures are not
relevant to a variety of applications approached by the use of hierarchical clustering methods
(Anderberg, 1973, Jain and Dubes, 1988) or by the use of methods such as the cohesive
clustering approach of Cao et al. (2015). Many other examples of these types exist. At the same,
centrality plays a vital role in numerous clustering problems and hence constitutes the focus of
this paper.
The remainder of this paper is organized as follows. Section 3 introduces the basic concepts
underlying pseudo-centroids (PCs) and their embodiment in a K-PC algorithm. Sections 4 and 5,
respectively, introduce the two instances of a PC-centroid called the MinMax-centroid and the
MinSum-centroid, together with their associated K-MinMax and K-MinSum algorithms. Section
6 addresses the topic of advanced starting methods to generate initial sets of points for the K 4
MinMax and K-MinSum algorithms, and describes the class of diversity-based starting methods
that also can help to determine good values of k for the number of clusters to be generated.
Section 7 then introduces the class of intensity-based starting methods with a focus on the
primary intensity-based methods. These methods have an ability to successively generate clusters
that are locally optimal for the chosen number of elements to be included in a cluster, given the
set of elements available at the current stage of construction. The more advanced adaptive
intensity-based starting methods are introduced in Section 8, which allow the number of
elements in a cluster to be generated adaptively at each stage. Finally, a strategic variant of the
K-PC algorithm, called the regret-threshold PC algorithm, is introduced in Section 9, together
with a diversification method and a new criterion for evaluating a collection of clusters. The
paper concludes with observations concerning other types of pseudo-centroids and applications
involving metaheuristic optimization in Section 10.
3. Pseudo-Centroids and a K-PC Algorithm
Let d(i,j) denote a distance measure defined on the elements xi, xj X. We follow the
convention d(i,i) = 0, but do not require that d(i,j) satisfies the conditions to be a metric or that it
be non-negative. For convenience, as in the case of d(i,j), we allow points xi and xj to be
referenced by their indexes. Let C denote an arbitrary set of points (indexes) in N. In the
following we refer to a separation measure Separate(i: C) that represents the separation of i
from other points j C \ {i}. As developed here, Separate(i: C) is a function of the distances
d(i,j) for j C \ {i} and specific instances of Separate(i: C) are given in the following sections.
Definition PC: A Pseudo-Centroid of C is a point i* C that gives the smallest value of
Separate(i: C); i.e.:
Separate(i*, C) = Min(Separate(i: C): i C)
The value Separate(i*: C) will be called the span of C, and denoted Span(C). The span value
may be thought of as a “radius” about i* (defined by reference to the set separation measure
Separate(i: C)), conceiving i* as a “central point,” such that all points j C \ {i} lie within the
separation Span(C) from i*.
An important characteristic of pseudo-centroids, as will be evident from the two primary
instances we discuss later, is that unlike ordinary centroids they may not be unique. We account
for this by including a variation of our algorithms based on pseudo-centroids that refers to the set
C* of all points in C that qualify as a pseudo-centroid of C; i.e., C* = {i C: Separate(i: C) =
Span(C)). The point i* in Definition PC thus belongs to C*. We first describe the K-PC
algorithm by assuming we arbitrarily single out a particular element of C* to be identified as i*,
in the case where more than one such element exists, and then describe the variant that
references all of C*.
5
K-PC Algorithm
0. Begin with a collection of k points N(K) = {i(h), h K} N, and assign each of the n –
k points j N \ N(K) to the point i = i(h) that is closest to it by the measure d(i,j). (That
is, j N \ N(K) is assigned to a point i(h) = arg min (d(i(h),j): h K).) Identify an initial
collection of clusters C(h), h K, each consisting of i(h) and the points assigned to it.
1. Identify a pseudo-centroid i*(h) for each cluster C(h) in accordance with Definition PC;
i.e., Separate(i*: C) = Span(C) for i* = i(h) and C = C(h). Denote this set of pseudocentroids by N*(K) = {i*(h): h K}.
2. Reassign each of the n – k points j N \ N*(K), to the pseudo-centroid i = i*(h) that is
closest to it by the measure d(i,j), thus assigning each j to an associated new cluster C(h)
that contains i*(h).
3. Repeat Steps 1 and 2 (creating a new set of pseudo-centroids at Step 1) until the set
N*(K) does not change or until a chosen iteration limit is reached.
As the foregoing description shows, the KC-PC algorithm closely follows the format of the KMeans algorithm (considering both in their simplest forms) except for the reference to the
separation measure Separate(i: C) and the introduction of pseudo-centroids in place of classical
centroids.
The variant of the foregoing algorithm that incorporates the set C* of all pseudo-centroids
amends the definition of N*(K) so that it becomes the union of the sets C*(h), h K. Then Steps
1 and 2 of the K-PC algorithm become as follows.
Modified Steps 1 and 2 to incorporate all pseudo-centroids
1. Identify the set C*(h) of pseudo-centroids for each cluster C = C(h) by C*(h) = {i C:
Separate(i: C) = Span(C)}. Denote this set of pseudo-centroids by
N*(K) = (C*(h): h K).
2. Reassign each of the n – |N*(K)| points j N \ N*(K), to the pseudo-centroid i* N*(K)
that is closest to it by the measure d(i,j), thus assigning each j to the associated new
cluster C(h) such that i* C*(h) (retaining each i* C*(h) in the this new cluster).
An additional variant of the K-PC algorithm allows all n points j N to be reassigned in Step 2
instead of only the points j N \ N*(K). If the distances d(i,j) are non-negative, and we interpret
the convention d(i,i) = 0 to permit d(i,i) to be slightly negative, then this implies that a pseudocentroid will be re-assigned to itself, and hence the outcome of Step 2 will not change for this
variant. However, in the case where negative distances exist, the indicated variant could reassign a pseudo-centroid to a different pseudo-centroid, and thus reduce the number of clusters
6
generated. Later we discuss other variants of the K-PC algorithm, including a Regret-Threshold
PC algorithm, which in turn suggest new variants of the K-Means algorithm.
Finally, we note that the classical form of the K-Means algorithm terminates when an objective
function expressed as the sum of the distances, or squared distances, of each point from its
assigned centroid, summed over all centroids, reaches a local minimum. An analogous (though
clearly not equivalent) objective function for the K-PC algorithm is to minimize
∑(Span(C(h)): h K)
This function may likewise be replaced by one in which the Span(C(h)) values are squared. Thus
an alternative termination point for the K-PC algorithm is to terminate when this function attains
a local minimum, where the next iteration causes the function to increase. This objective may
also be modified to replace Span(C(h)) by Span(C(h))/|C(h)|, thus seeking to minimize the sum
of the “average” span values over the clusters. In Section 9 we present a somewhat different
function for evaluating the clusters which can also be used to identify good values of k.
4. The MinMax-Centroid and the K-MinMax Algorithm
The first principal instance of a pseudo-centroid we consider, the MinMax-centroid, arises by
defining Separate(i: C) = Max(d(i,j), j C \ {i}), which we also denote by MaxDist(i,C). Then
we obtain the instance of Definition PC given by
Definition MinMax: A MinMax-Centroid of C is a point i* C whose maximum distance from
all other points in C is smallest; i.e.:
MaxDist(i*: C) = Min(MaxDist(i: C): i C)
Then Span(C) = MaxDist(i*: C) and for this case will be denoted MMSpan(C).
The statement of the K-MinMax Algorithm then becomes essentially the same as the statement
of the K-PC algorithm, upon replacing “pseudo-centroid” with “MinMax-centroid,” “Separate(i:
C(h))” with “MaxDist(i: C(h))” and “Span(C) with MMSpan(C).”
K-MinMax Algorithm
0. Begin with a collection of k points N(K), and assign each of the n – k points j N \ N(K)
to the point i = i(h) N(K) that is closest to it by the measure d(i,j). Identify an initial
collection of clusters C(h), h K, each consisting of i(h) and the points assigned to it.
1. Identify a MinMax-centroid i*(h) for each cluster C(h); i.e., MaxDist(i*: C) =
MMSpan(C) for i* = i(h) and C = C(h). Denote the set of MinMax-centroids by N*(K) =
{i*(h): h K}.
7
2
Reassign each of the n – k points j N \ N*(K), to the MinMax-centroid i = i*(h) that is
closest to it by the measure d(i,j), ), thus assigning each j to an associated new cluster
C(h) that contains i*(h).
3
Repeat Steps 1 and 2 (creating a new set of MinMax-centroids at Step 1) until the set
N*(K) does not change or until a chosen iteration limit is reached.
The time complexity of the K-MinMax algorithm compared to that of the K-Means algorithm
depends on the dimension d of the points i N (vectors xi, i N), which are used to compute a
centroid in the K-Means algorithm, as determined by the execution of Step 1 (since Steps 0 and 2
are essentially the same in both algorithms). The complexity of identifying a MinMax-centroid
of a cluster C(h) in Step 1 of the K-MinMax algorithm is O(|C(h)|2 ), while the complexity of
identifying an ordinary centroid with the K-Means algorithm is O(|C(h)|d). If the distances d(i,j)
must be precomputed from coordinates, rather than being given directly from the problem data
base for the application of concern, then this will inflate the the complexity of the K -MinMax
algorithm. In this respect, it may be noted that even where distances rely on coordinate
calculations, commercial enterprises or government agencies may prefer not to disclose the
information contained in these coordinates, and therefore will preprocess the data to produce
distances before passing them along to a clustering algorithm. 1
Independently of such considerations, the computational effort of Step 1 in the K-MinMax
algorithm can potentially be improved as follows:
Accelerated update for Step 1. For each MinMax-centroid i* = i*(h) identified in Step 1, let
match(i*) be a point j* that determines the MMSpan(C(h)); that is, d(i*, j*) = MaxDist(i*:
C(h)). If j* belongs to the new cluster C(h) produced in Step 2, then the computation to identify
the MinMax-centroid of C(h) on the next execution of Step 1 can restrict attention to the set of
points C+(h) that were added to C(h) in Step 2. Specifically, for each i C+(h) compute the value
MaxDist(i: C(h)) = Max(d(i,j): j C(h) \ {i}) and from these values identify i'(h) = arg
min(MaxDist(i: C(h)): i C+(h)). Then the MinMax-centroid of the new C(h) passed from Step
2 to Step 1 is either i'(h) or i*(h), depending on which of MaxDist(i'(h): C(h)) and
MMSpan(C(h)) (= d(i*(h),j*)) is smaller.
The modification of the K-MinMax algorithm to refer to the full sets of MinMax-Centroids
C*(h), when these may contain more than one element, changes Steps 1 and 2 as indicated
earlier. In this case the changed form of Step 1 identifies C*(h) by C*(h) = {i C: MaxDist(i: C)
= MMSpan(C)} for C = C(h) (together with redefining N*(K) = (C*(h): h K)) and the
changed form of Step 2 reassigns each of the n – |N*(K)| points j N \ N*(K), to the MinMaxcentroid i* N*(K) that is closest to it, creating new clusters C(h) (which, respectively, inherit
the previous elements of C*(h)).
The Accelerated Update of Step 1 can be applied to the situation where we keep track of the sets
C*(h) as a result of the fact that MMSpan(C(h)) is the same value regardless of which element i*
1
Communication from B. Cao, regarding applications of the method of Cao et al. (2015) to commercial
applications.
8
C*(h) is used to identify it, and it is only necessary to retain a single associated element j* in
C(h) (for some i*) to imply the value MMSpan(C(h) (previously = d(i*,j*)) remains at least as
large in the new C(h) as it was before. Likewise we may refer to a single index i'(h) to compare
the two distances MaxDist(i'(h): C(h)) and MMSpan(C(h)). Then the new set of MinMaxcentroids will be determined according to which of these distances is smaller. Since all MinMaxcentroids of the old C(h) are required to be carried forward as elements of the new C(h), the new
set C*(h) will contain the old C*(h) plus any additional points of C +(h) that qualify to be
included.
5. MinSum-Centroids and a K-MinSum Algorithm
The second type of pseudo-centroid we focus on, called a MinSum-centroid, results by defining
defining Separate(i: C) = SumDist(i,C), where SumDist(i,C) = ∑ (d(i,j), j C \ {i}). Then we
obtain the instance of Definition PC given by
Definition MinSum: A MinSum-centroid of C is a point i* C whose sum of distances from all
other points in C is smallest; i.e.:
SumDist(i*: C) = Min(SumDist(i: C): i C)
Then SumDist(i*: C) identifies the value of Span(C), which in this case will be denoted
MSSpan(C).
A weighted version of a MinSum-centroid can be created without introducing weights separately
but simply by modifying the distances. Specifically, suppose it is desired to attach a weight w(D)
to each d(i,j) such that d(i,j) = D. Then it suffices to create the new value d'(i,j) := w(D)d(i,j) for
D = d(i,j) and then use d'(i,j) in place of d(i,j). To keep the ordering the same for d'(i,j) and d(i,j)
we would specify that w(D) is a monotonically increasing function of D. This occurs, for
example, for non-negative distances by setting w(D) = Dp where p 0. (E.g., setting p = 1 yields
d'(i,j) = d(i,j)2 .)
For weights w(D) that grow fast enough (as where w(D) = Dp for a very large power p), the
MinSum-centroid becomes a special instance of the MinMax-centroid – one where ties among
candidates for i* are implicitly broken by a lexicographic rule: if the largest d(i,j) values are the
same, then the second largest d(i,j) values are compared for the tied elements, and if ties remain
at this level, the third largest d(i,j) values are compared for the resulting tied elements, and so on.
However, a scheme employing such large weights is very likely to produce d(i,j) values too large
to be easily handled computationally, thus making the direct MinMax-centroid preferable to the
weighted MinSum-centroid for problems where maximum distances are more relevant than
(weighted) sums of distances.
On the other hand, when weights of reasonable size are employed, the structures of clusters
produced by a K-MinMax Algorithm and a K-MinSum Algorithm can differ appreciably, and
9
hence it becomes of interest to compare the two approaches to determine which may have
advantages in particular contexts.
While the statement of the K-MinSum Algorithm can be readily inferred from the statement of
the K-PC Algorithm and our preceding definitions, we give explicit instructions for the KMinSum Algorithm as follows. By this means, we are able to make observations about the KMinSum Algorithm that differ from those concerning the K-MinMax Algorithm, particularly in
the context of advanced starting methods of later sections.
K-MinSum Algorithm
0. Begin with a collection of k points N(K), and assign each of the n – k points j N \ N(K)
to the point i = i(h) N(K) that is closest to it by the measure d(i,j). Identify an initial
collection of clusters C(h), h K, each consisting of i(h) and the points assigned to it.
1. Identify a MinSum-centroid i*(h) for each cluster C(h); i.e., SumDist(i*: C) =
MSSpan(C) for i* = i(h) and C = C(h). (SumDist(i: C(h)): i C(h)). Denote the set of
MinSum-centroids by N*(K) = {i*(h): h K}.
2. Reassign each of the n – k points j N \ N*(K), to the MinSum-centroid i = i*(h) that is
closest to it by the measure d(i,j), ), thus assigning each j to an associated new cluster
C(h) that contains i*(h).
3. Repeat Steps 1 and 2 (creating a new set of MinSum-centroids at Step 1) until the set
N*(K) does not change or until a chosen iteration limit is reached.
The time complexity of the K-MinSum algorithm is the same as that of the K-MinMax
algorithm, and hence the comparison to the complexity of the K-Means algorithm again depends
on the dimension d of the points i N (vectors xi, i N) involved in computing a centroid, as
determined by the execution of Step 1. Specifically, the different complexities for executing this
step are O(|C(h)|2 ) for the K-MinSum algorithm and O(|C(h)|d) for the K-Means algorithm.
The computational effort of Step 1 in the K-MinSum algorithm can potentially be improved by a
somewhat different method than in the case of the K-MinMax algorithm.
Accelerated update for Step 1 of the MinSum Algorithm. Let C' = C'(h) denote the set of points
C(h) before executing Step 2 and C" = C"(h) denote the set of points after executing Step 2. Let
Co = C' C" (the points that remain in C(h) after the reassignment of Step 2), C + = C" \ C' (the
points added to C(h) in Step 2) and C - = C' \ C" (the points dropped from C(h) in Step 2).
Assume the values SumDist(i,C') have been saved before executing Step 2. Then for i Co , the
new SumDist(i: C(h)) value is given by SumDist(i: C") = SumDist(i,C') + ∑(d(i,j): j C+) –
∑(d(i,j): j C-). This update can be useful for saving computation on iterations where some of
the clusters C(h) do not greatly change their composition by Step 2 (i.e., where C o is relatively
large compared to C + and C-).
10
We now turn to one of the main features of the K-PC algorithm that motivates our study, which
is the ability to derive an initial collection of clusters and their pseudo-centroids by means of
special types of advanced starting algorithms.
6. Advanced Starting Methods
We introduce two types of advanced starting methods, called Diversity-Based Starting Methods
and Intensity-Based Starting Methods,2 to generate an initial set of points N(K) = {i(h), h K}
for the K-MinMax and the K-MinSum algorithms. The diversity-based methods are simpler and
require less computation than the intensity-based methods, and apply equally to the MinMax and
MinSum cases. (These methods can also be used to generate starting points for the K-Means
algorithm.)
The more advanced intensity-based methods take slightly different forms for the MinMax and
MinSum cases. Their greater computational complexity makes them less suited for large
clustering problems, though they afford an opportunity to generate a better set of initial points to
become N(k) in the K-MinMax and K-MinSum algorithms. They also generate MinMax and
MinSum-centroids to compose the set N*(k) for the first execution of Step 2, and simultaneously
produce associated clusters C(h).
The intensity-based methods are of two types, which we call Primary and Adaptive. Both
generate a best cluster of a given size from the elements available at each step. This does not
assure the final collection of clusters will be optimal since the cluster size may not be chosen
correctly and a best cluster at a given stage may not be a globally best choice over all stages.
Nevertheless, it increases the likelihood that the quality of the initial clusters will lead to
reducing the overall number of iterations of the K-PC algorithm and improve the quality of the
final clusters produced.
The remainder of this section is devoted to the diversity-based starting methods. The intensitybased starting methods are then described in the two subsequent sections, introducing the
Primary Intensity-Based Algorithms in Section 7 and the Adaptive Intensity-Based Algorithms
in Section 8.
Diversity-Based Starting Methods
The diversity-based starting methods constitute an elaboration of an approach suggested in
Glover (1994) which may be formulated in the present context by viewing N as the set from
The “intensity-based” terminology does not relate to “intensification methods” as used in metaheuristic
algorithms. However, the “diversity-based” terminology carries an association with “diversification
methods” used in such algorithms.
2
11
which to create a subset N(K) of k elements. In this setting, the approach may be described as
starting with an arbitrary seed element from N as the first element of N(K) and then sequentially
selecting each of the remaining k – 1 elements to maximize its minimum distance from the
elements chosen so far. We first depict a simple instance of this method based on the following
terminology. Let Ho = {i(h): h = 1, …, ko } denote the set of starting points selected as of iteration
ko , where ko ranges from 1 to k. As the points i(h) are successively selected from N and added to
Ho , we also refer to the set Io = N \ Ho which identifies the points remaining in N as candidates to
be selected for inclusion in Ho on the next iteration.
By the sequence of updating Ho , the value MinD(i) identified below is equivalent to the value
defined as MinD(i) = Min(d(i,i(1)), …, d(i,i(ko – 1)).
Simple Diversity-Based Starting Method
Initialization: Choose an arbitrary point i N, set i(1) = i, ko = 1, Ho = {i(1)} and Io = N \ {i(1)}.
Core Loop
For ko = 2 to k:
Identify MinD(i) = Min(d(i,j): j Ho ) for each i Io .
Choose i' = arg max(MinD(i): i Io ) and set i(ko ) = i', Ho := Ho {i'} and Io := Io \ {i'}3
EndFor
The final set Ho = {i(ko ): ko = 1 to k} becomes the set of starting points for the K-MinMax or the
K-MinSum algorithm. The computational complexity of this method is O(n∙k2 ), since we
examine O(n) points on each of the k – 1 iterations from ko = 2 to k, and examine the distances
from each point to ko – 1 other points.
The original proposal for the preceding method (in a different context) gave extended tiebreaking rules for choosing i(ko ) where multiple points qualified to be a point i' satisfying
MinD(i') = Max(MinD(i)). Instead of employing such rules, we now propose a refinement of the
Simple Diversity-Based Starting Method.
The underlying idea is to consider the maximum of the smallest MinD(i(ko )) value for the
elements i(ko ) in Ho = {i(ko ): ko = 1 to k} to be an indicator of the quality of Ho . Due to the
construction of Ho , this maximum MinD(i(ko )) value results when i(ko ) = i(k), the last element
added to Ho . Hence, we adopt the goal of maximizing MinD(i(k)), and re-run the algorithm by
allowing it to choose this last element i(k) as the new starting element i(1). This approach is
based on the supposition that if element i(k) (which is a limiting element in determining the
quality of the current Ho ) belongs to a better set Ho , then we increase the possibility of finding
such a better Ho by making i(k) the first element of this set, to increase the distance of other
points from it. At the conclusion of the execution that has specified the old i(k) to be the new
3
The MinD(i) values used in these computations need not be stored as a vector indexed by i, but can
generated and compared sequentially.
12
i(1), we therefore check whether the resulting final MaxMin distance MinD(i(k)) exceeds this
distance on the previous execution (denoted PreviousMaxMin). If so, the maximization objective
value is improving and we re-run the algorithm by again setting i(1) = i(k) for this new i(k). The
method stops when MinD(i(k)) < PreviousMaxMin, at which point the Ho that was generated at
the end of the previous execution (denoted PreviousHo ) is recovered, thus picking a local
optimum for maximizing MinD(i(k)). We refer below to the modified version of the previous
Core Loop that handles this objective as the Core Algorithm.
Refined Diversity-Based Starting Method
Initialization: Choose an arbitrary point i N, set i(1) = i, ko = 1, Ho = {i(1)} and Io = N \ {i(1)}.
Set PreviousMaxMin = – Large and CurrentMaxMin = 0.
Core Algorithm
While CurrentMaxMin > PreviousMaxMin
For ko = 2 to k:
Identify MinD(i) = Min(d(i,j): j Ho ) for each i Io
Choose i' = arg max(MinD(i): i Io ) and set i(ko ) = i', Ho := Ho {i'} and
Io := Io \ {i'}
EndFor
CurrentMaxMin = MinD(i')
If (CurrentMaxMin ≤ PreviousMaxMin) then
Break 4
Else
PreviousMaxMin = CurrentMaxMin
PreviousHo = Ho
i(1) = i'; Ho = {i(1)}; Io = N \ {i(1)}
Endif
EndWhile
The foregoing approach can also choose an element of Ho other than the last to become the new
i(1) (for example, selecting a “middle element” i(ko ) from the sequence, obtained by rounding ko
= (k – 1)/2 to an integer). However, the stopping criterion should continue to be defined in
reference to a local maximum for MinD(i(k)), the last element of Ho . Hence the only change in
the foregoing pseudocode is to select a different element than the last element i(k) in Ho to
become the new i(1). The computational complexity of the procedure remains O(n∙k2 ) provided a
constant limit (e.g., 5 or 10) is placed on the number of executions allowed in searching for a
locally maximum MinD(i(k)) value.
The value MinD(i(k)) for the last element of Ho can also serve another function, since it gives an
estimate of the least distance for separating the PC-centroids. If this estimated distance is small
we may conclude there is a significant chance that k has been chosen too large. Moreover, a
4
The instruction Break means to exit the “current loop” (the innermost loop the instruction is embedded in).
13
process of monitoring the successive MinD(i(ko )) values as ko ranges upward from 2 can disclose
ko values where MinD(i(ko )) abruptly decreases or falls below a minimum desired distance
separating the clusters generated.
The preceding Refined Diversity-Based method can be further refined by not waiting until ko = k
to apply it, but instead by selecting a value kCheck < k which is treated temporarily as if it were
the final ko . In this way, the objective of maximizing MinD(i(k)) can be additionally pursued at
an earlier stage of building the set Ho , thus potentially affording greater leeway for selecting
subsequent elements to add to Ho that will keep the MaxMin distance high.
This refinement may be executed for multiple (successively larger) values of kCheck, but we
indicate the form of the method for choosing only a single kCheck value. Thus we first iterate
from ko = 2 to kCheck, and then after finding a local maximum for the last MinD(i') (for i' =
i(kCheck)), we continue iterating from ko = kCheck + 1 to k to complete the method (concluding
by seeking a maximum value for MinD(i') for k o = k).5 We call this approach the Compound
Diversity-Based Starting Method which we organize to always iterate over ko from kFirst to
kCheck instead of over 2 to k.6
Compound Diversity-Based Starting Method
Initialization: Choose k' [2, k – 1]. Choose an arbitrary point i N, set ko = 1, i(1) = i, Ho =
{i(1)} and Io = N \ {i(1)}. Set PreviousMaxMin = – Large, CurrentMaxMin = 0 and kFirst = 2.
While kFirst ≤ k
Compound Core Algorithm
While CurrentMaxMin > PreviousMaxMin
For ko = kFirst to kCheck
Identify MinD(i) = Min(d(i,j): j Ho ) for each i Io
Choose i' = arg max(MinD(i): i Io ) and set i(ko ) = i', Ho := Ho {i'} and
Io := Io \ {i'}
EndFor
CurrentMaxMin = MinD(i')
If (CurrentMaxMin ≤ PreviousMaxMin) then
Break
Else
PreviousMaxMin = CurrentMaxMin
PreviousHo = Ho
i(1) = i'; Ho = {i(1)}; Io = N \ {i(1)}
kFirst = 2 (to reset kFirst after having assigned it the value
5
An interesting small value of k' to initiate such an approach is 3.
This permits starting with kFirst = 2 and later increasing kFirst to kCheck + 1 accompanied by resetting
kCheck = k for the final execution.
6
14
kCheck + 1 following the first execution of the Compound Core
Algorithm)
Endif
EndWhile
(End Compound Core Algorithm)
If kCheck < k then
kFirst = kCheck + 1
kCheck = k
PreviousMaxMin = – Large; CurrentMaxMin = 0
Elseif kCheck = k then
Break
Endif
EndWhile
We consider one last type of diversity-based starting method, called the Targeted DiversityBased Starting Method, which instead of seeking to maximize MinD(i) at each iteration selects a
target value T for MinD(i) and identifies Deviation(i) = |MinD(i) – T| for the goal of minimizing
this latter deviation. An appropriate value for T may be determined by first executing one of the
preceding diversity-based starting methods and then selecting T to be final the mean or the
median of the MinD(i(ko )) values generated as ko ranges from 2 to k (noting that no MinD(i(ko ))
value exists for ko = 1), This targeted diversity-based approach will typically yield the same
choice of i' = i(k) as by the MaxMin choice of the preceding methods on the final iteration for ko
= k.
We identify a simple instance of this method analogous to the Simple Diversity-Based Starting
Method described earlier.
Simple Targeted Diversity-Based Starting Method
Initialization: Choose an arbitrary point i N, set i(1) = i, ko = 1, Ho = {i(1)} and Io = N \ {i(1)}.
For ko = 2 to k:
Identify MinD(i) = Min(d(i,j): j Ho ) and Deviation(i) = |MinD(i) – T| for each i Io
Choose i' = arg min(Deviation(i): i Io ) and set i(ko ) = i', Ho := Ho {i'} and
Io := Io \ {i'}7
EndFor
As with the previous diversity-based starting methods, the Targeted Diversity-Based Starting
Method can be embodied in refined versions that identify different initial choices of i(1) for
restarting the method, continuing in this case as long as the T value computed at the end of each
execution is increasing, to yield a local maximum relative to the rule for selecting T. A special
7
As in the case of the MinD(i) values, the Deviation(i) values can be identified sequentially and do not
need to be stored as vectors indexed by i.
15
variant of the Targeted Diversity-Based Starting Method designed to create a more compact set
of diverse points is described in Section A of the Appendix.
7. Primary Intensity-Based Starting Methods
The intensity-based starting methods exploit the structure of the MinMax and MinSum-centroids
to generate a starting collection of clusters where the current cluster at each iteration is the best
one possible, subject to the elements available for generating a cluster and the chosen value for
the cluster size at that iteration. Rather than identifying starting points for the set N(K) in Step 0,
the intensity-based methods generate a starting collection of clusters C(h) and the associated set
N*(K) of MinMax-centroids or MinSum-centroids to be passed to Step 2, thereby bypassing the
first execution of Steps 0 and 1.
These intensity-based procedures require O(n2 ∙(k + log n)) effort in comparison with the O(n∙k2 )
effort of the diversity-based algorithms of Section 6, which should be taken into account in
considering their use for large problems. However, the ability of the intensity-based methods to
select the best cluster of a given size from the collection of elements available affords an
opportunity to reduce the number of iterations consumed by the K-MinMax and K-MinSum
algorithms and thereby compensate for the time invested in obtaining the initial clusters.
We employ the following notation and conventions which apply to the adaptive intensity-based
methods as well as to the primary methods.
Notation and Conventions
No = a subset of N consisting of all points at a particular stage of construction available to belong
to clusters C(h) not yet created. (Hence No begins equal to N, and then shrinks as elements are
removed to populate each cluster C(h) as it is created.) We let no = |N o | and to begin take N o = N.
δ(i) = a logical “indicator variable,” where δ(i) = True if i No and False otherwise. (The use of
this indicator variable allows the algorithms that follow to be expressed in a convenient form,
although more efficient versions of the algorithms may be produced by reference to doubly
linked lists.)
ko = the number of clusters (and hence the number of MinMax-centroids) that remain to be
generated, starting at ko = k and decreasing ko by 1 at each iteration until reaching 1. Each
iteration is identified by its ko value.
C(ko ) = the cluster generated at iteration ko .
MMCentroid(ko ) = the MinMax-centroid i* for C(ko ).
16
ClusterSize = the targeted number of elements |C(k o )| in the cluster C(ko ). In the simple (first)
version of the algorithm described below, ClusterSize receives a value that is (approximately) the
same for each ko , while in the succeeding more advanced version ClusterSize varies adaptively
for different ko values.
ScanSize = the number of distances d(i,j) scanned from element i at iteration k o . Since element i
itself will belong to the cluster C(ko ) to be generated, ScanSize = ClusterSize – 1.
We determine bounds MinSize and MaxSize on ClusterSize as follows. At least one of the ko
clusters remaining to be created must have a size that equals or exceeds the average number of
elements no /ko that can be assigned to these clusters. We adopt the goal of insuring the current
cluster C(ko ) will attain this size and hence set
MinSize = no /ko
(where is the roof function that rounds fractional values upward). In the version of the
algorithm that only allows a single value for the size of C(k o ), we choose ClusterSize = MinSize.
We observe that this choice is highly natural for applications such as Cao and Glover (2010)
where it is highly desirable to have clusters that are all approximately of the same size.
For the adaptive version of the algorithm we allow ClusterSize to vary between MinSize and an
upper bound MaxSize. Let GlobalMinSize be a constant denoting the fewest number of elements
permitted in any cluster (set externally, independent of the value MinSize). (For our following
purposes, it is useful to choose GlobalMinSize 2.) Then on the current iteration, with k o
clusters remaining to be constructed, the maximum number of elements permitted in the current
cluster C(ko ) is
MaxSize = no – GlobalMinSize∙(ko – 1)
Correspondingly, we set the lower and upper bounds on the number of elements scanned from
any given point i by MinScan = MinSize – 1 and MaxScan = MaxSize – 1.
Finally, all of the intensity-based methods are based on a preprocessing step that creates an
ordered set of indexes i(1), i(2), …, i(n-1) for each i N to sequence the distances d(i,j) for each
i N\{i} in ascending order so that
d(i,i(1)) ≤ d(i,i(2)) … ≤ d(i,i(n-1))
This preliminary ordering is potentially the most expensive part of the intensity-based methods,
with a time complexity of O(n2 ∙log n). Remaining operations of the intensity-based methods are
O(n2 ∙k) which makes the total complexity O(n2 ∙(k + log n)) which may reduce to the complexity
of creating the preliminary ordering if log n > k. An exception occurs if the distances d(i,j) are
integers and for each i, Max(d(i,j): j N) – Min(d(i,j): j N) has an O(n) range, in which case a
17
properly constructed bucket sort can sequence the distances for all i N in O(n2 ) time, and the
O(n2 ∙k) complexity dominates.
It is to be noted that the preliminary ordering needs to be done only once, and its information can
be re-used in refined versions of the algorithms subsequently described and also re-used to
generate K-MinMax and K-MinSum clusters for different values of k.
Since the intensity-based methods differ for the K-MinMax and the K-MinSum algorithms, we
begin by introducing these starting methods for the K-MinMax setting, starting with the simpler
“primary” version that selects the single value ScanSize = MinScan.
Primary MinMax Starting Method
Initialization: Begin with the ordered indexes i(1), i(2), …, i(n-1) for each i N and set N o = N,
no = n and δ(j) = True, j N.
For ko = k to 1,
ClusterSize = no /ko (ClusterSize = MinSize)
ScanSize = ClusterSize – 1
Execute the MinMax Distance Algorithm (described below)
(At the conclusion: i* = arg min(Max (d(i,,j): j C) (= MMSpan(C)) over all sets
C No such that |C| = ClusterSize.)
Generate C(ko ) and update the set No
(This step re-executes the relevant part of the MinMax Distance Algorithm for i = i* to
determine the set C = C(ko ) associated with i*.)
C(ko ) =
Scan = 0
For s = 1 to n – 1
j = i*(s)
If (δ(j) = True) then (j is an element of No \ {i}
Scan := Scan + 1
No := No \ {j}
C(ko ) = C(ko ) {j}
δ(j) = False
If (Scan = ScanSize) Break
Endif
Endfor
N o := No \ {i*}
C(ko ) = C(ko ) {i*}
δ(i*) = False
MMCentroid(ko ) = i*
no := no – ClusterSize
18
Endfor
On the final iteration of the Primary Algorithm when ko = 1, the step of generating C(ko ) can be
shortcut by simply setting C(1) = N o and MMCentroid(1) = i* (and there is no need to update No
because it will become empty). This ability to shortcut the final update of C(k o ) for ko = 1 also
holds for all intensity-based starting methods subsequently discussed.
The MinMax-centroids of MMCentroid(ko ), ko = 1, …, k, generated by the foregoing method
give the initial set of points N*(K) = (i(h), h K} that can be passed directly to Step 2 of the KMinMax Algorithm along with the associated clusters C(k o ). This likewise is true of all intensitybased methods subsequently discussed (referring to the array MSCentroid(k o ) in the case of the
K-MinSum algorithm).
The internal MinMax Distance algorithm executed within the Primary Algorithm, described
next, is understood to share its arrays and values with the Primary Algorithm.
MinMax Distance Algorithm
Conventions:
i* identifies a point i that yields a MinMax-centroid for a “best” C of size ClusterSize in N o .
BestDistance = the value Min((Max (d(i*,j): j C): C No ) that gives MMSpan(C) for this
“best” C.
Large = a large positive number
Initialization:
BestDistance = Large
For i N o
Scan = 0 (the number of elements j scanned from i)
For s = 1 to n – 1
j = i(s)
If (δ(j) = True) then (j is an element of No \ {i})
Scan := Scan + 1
If (Scan = ScanSize) Break
Endif
Endfor
If (d(i,j) < BestDistance) then
i* = i
BestDistance = d(i,j)
Endif
Endfor
The rationale underlying the foregoing method may be expressed as follows. Let v = ClusterSize
and no = |N o |, for ClusterSize and N o determined at iteration ko , and let CSet = {C N o : |C| = v}
19
denote the set of all clusters C in N o of size v. Assume at each iteration ko we seek to identify a
“best” cluster C best that minimizes MMSpan(C) over the clusters C in CSet. Evidently, one way
to do this, which we would hope to improve upon, is to examine the clusters in CSet (where
|CSet| = no !/v!(no – v)!) and then determine the MinMax-centroid i' of each C in CSet to compute
MMSpan(C) = Max(d(i',j): j C \ {i'}). The following result establishes the ability of the
Primary Starting Method to achieve the desired outcome with vastly less effort.
Theorem 1. The cluster C(ko ) and the associated point i* identified at the conclusion of
iteration ko of the Primary Starting Method respectively qualify to be Cbest and its MinMaxcentroid.
Proof: Let io (t), t = 1, …, no – 1 identify the subsequence of i(s), s = 1, …, n – 1 such that
δ(j) = True for j = i(s). Hence, defining jt = io (t) (with the identity of i implicit), the ordering of
distances d(i,j) for point i can be written d(i,j1 ) ≤ d(i,j2 ) ≤ … ≤ d(i,jno-1 ). Consider the cluster Ci
consisting of i and the points jt for t = 1 to v – 1. (Note that t corresponds to successive values of
Scan, and v – 1 = ScanSize.) It is possible that i is not a MinMax-centroid of C i and hence the
distance d(i,jv-1 ) may be larger than MMSpan(Ci). Neverthless, for a point i* that qualifies as a
MinMax-centroid of Cbest (at least one such i* must exist) there can be no better candidate for
Cbest than the set C i* , and the distance d(i*,jv-1 ) must equal MMSpan(Cbest ). Therefore, letting
BestDistance denote the minimum of the d(i,jv-1 ) values over all points i examined so far, the
Primary MinMax Starting Method goes through the successive indexes jt , t = 1 to v – 1, for each
i N o (where jt depends on i) and stops at t = v – 1 to check whether d(i,jv-1 ) < BestDistance. If
the inequality holds, then BestDistance is updated to be the current d(i,jv-1 ) and i* is recorded as
i, which verifies the validity of the Starting Method and establishes the claim of the theorem.
Note the foregoing proof also shows it would be possible to interrupt the sequence of t values for
a given i N o if d(i,jt ) BestDistance, and thus break from the “For loop” at this point. The
foregoing analysis also makes it clear that the Primary MinMax Starting Method succeeds in
finding a best C and its MinMax-centroid for a given N o in O(no ∙n) time. (This time could be
reduced to O(no ∙v) if the sequence io (t) were used directly instead of relying on the check δ(j) =
True for j = i(s)). For the iterations for ko = k to 1 the total effort is therefore bounded above by
O(n2 ∙k), as remarked earlier.
We are motivated to improve upon the outcome obtained by the Primary Starting Method by
accounting for the fact that an ideal collection of clusters may well have somewhat different
numbers of elements in different clusters. One way to pursue such an improvement is to look
beyond the end of the execution of the Primary Starting Method to identify values for
ClusterSize = ClusterSize(ko ), for ko = 1 to k which are then used to launch a second execution of
the Starting Method. We consider two such approaches next.
Refinement to Choose Different ClusterSize values.
Approach 1: After applying the Primary Starting Method (to bypass Steps 0 and 1 of the
K-MinMax algorithm), perform the first execution of Step 2 to determine a new collection of
clusters C(h), h K. Index these clusters so the values |C(h)| are in descending order for h = k to
20
1 and set ClusterSize(h) = |C(h)|. Finally, rerun the Primary Starting Method by setting ScanSize
= ClusterSize(ko ) – 1 in the “For loop” from ko = k to 1. We choose to put the ClusterSize(ko )
values in descending order as ko goes from k to 1 by noting that N o will be larger, and hence will
offer greater latitude for choosing C(ko ), for larger values of k. (This approach may be shortened
by interrupting the Primary Starting Method after executing it for any value k o * and reassigning
the elements of N \ No in Step 2 to obtain a subset of clusters C(h) for h = k to ko *. The indexes
of clusters are then arranged as indicated above to produce descending ClusterSize(ko ) for ko = k
to ko *. Remaining ClusterSize values for ko = ko *– 1 to 1 can be determined by ClusterSize = no
/ko exactly as before.)
Approach 2: This approach is the same as the first, except that it waits until a later
iteration of the K-MinMax Algorithm to select and order the collection of clusters C(h) at Step 2.
If the approach waits until the final iteration, the overall effort of generating the final clusters
may be expected to be less than twice the amount of the first execution of the K-MinMax
Algorithm, because effort is saved on the second execution by not having to re-do the
preprocessing step that determines the sequence (i(s): s = 1, …, n-1) for each i N. In addition,
fewer iterations may be required on the second execution due to the chance of generating a better
set of starting clusters.
The Adaptive Intensity-Based Starting Methods described in Section 8 give a different way to
account for the fact that the best clusters C(h) can vary in size.
Primary Intensity-Based Method for the K-MinSum Problem
We organize the Primary Intensity-Based Starting Method for the K-MinSum Algorithm in the
same manner as for the K-MinMax Algorithm. This method is quite similar to the Primary
MinMax Starting Method, the chief difference being to replace MaxDist(i) with SumDist(i), and
to replace the list MMCentroid(ko ) with a corresponding list MSCentroid(ko ) to identify the
MinSum-centroids generated. To avoid ambiguity concerning this correspondence, and to clarify
how the MinSum process differs from the MinMax process, we identify the Primary MinSum
Starting Method in Section B of the Appendix.
8. Adaptive MinMax and MinSum Starting Methods
As already noted, the adaptive starting methods for the K-MinMax and K-MinSum algorithms
add a layer of sophistication beyond that of the primary starting methods, in order to respond to
the challenge of determining how to select varying cluster sizes in the absence of prior
knowledge about the ideal sizes. We begin by considering an adaptive method for the KMinMax algorithm.
21
8.1 Foundation of the Adaptive MinMax Starting Method
The Adaptive MinMax method has two phases. Phase 1 resembles the Primary MinMax method
by identifying a BestDistance value over all i No for ClusterSize = MinSize. In addition, it
identifies a value DistanceLimit that limits the largest distance d(i,j) based on the successive
differences in the ascending d(i,j) values that are used to compute BestDistance.
Phase 2 then makes use of the information generated in Phase 1 by examining each i N o and
all elements j in the set J(i) = {j No \ {i}: d(i,j) ≤ DistanceLimit) to identify a set of “best
points” i as candidates for the MinMax-centroid i*. This set of points is defined by
BestSet = {i No : |J(i)| = MaxJ} where MaxJ = Max(|J(i)|: i No ). Finally, to choose among
elements in BestSet, the method conputes the value D(i) = Max(d(i,j): d(i,j) ≤ DistanceLimit) and
selects the element i* that gives.
i* = arg min (D(i): i BestSet)
This construction insures i* will be a MinMax-centroid over the set C = J(i*) {i*} and, as in
the case of the Primary MinMax method, will identify a best such set among those of the same
size. Since the adaptively determined cluster C can contain somewhat more than MinSize
elements, the upper limit MaxSize is imposed by the calculation indicated earlier as a safeguard
against the possibility that |C| may become too large.
If the procedure is modified to permit MaxSize to receive a value as large as |N o |, then the
current cluster could potentially absorb all of N o , and under such a condition the method will
terminate with fewer than k clusters.8 Such a variation can be of interest for enabling the
adaptive starting method to propose a new (smaller) value for k. An evaluation criterion to verify
when fewer clusters can be appropriate is suggested in Section 9.
In spite of the more advanced character of the adaptive starting method, the use of the ordering
(i(s): s = 1, …, n-1) makes it possible to identify the sets J(i) and values D(i) in Phase 2 of the
adaptive method implicitly, rather than explicitly, as part of the process of identifying i*.
Adaptive MinMax Starting Method
Initialization: Begin with the ordered indexes i(1), i(2), …, i(n-1) for each i N and set N o = N,
no = n and δ(j) = True, j N.
For ko = k to 1
MinSize = no /ko
MaxSize = no – MinSize∙(ko – 1)
MinScan = MinSize – 1
8
An exception occurs when ko = 1, at which point all of N o is absorbed in any case.
22
MaxScan = MaxSize – 1
Execute the Phase 1 MinMax Algorithm (to identify BestDistance)
Execute the Phase 2 MinMax Algorithm (to identify i*)
Generate C(ko ) and update the set No
(This portion of the Adaptive MinMax Method is identical to that of the Primary
MinMax Method, except that “If (Scan = BestScan) Break” replaces
“If (Scan = ScanSize) Break”.)
Endfor
Next we describe the Phase 1 and Phase 2 algorithms embedded in the Adaptive MinMax
method. We employ a bare minimum of descriptive comments, since the mnemonic names of the
variables and the instructions of the algorithm should make the interpretation clear.
Phase 1 MinMax Algorithm
Assumption: No contains at least 2 elements
BestDistance = Large
BestMinGap = 0
For i N o
MinGap = Large
Sum = 0
SumGap = 0
Scan = 0
For s = 1 to n – 1
j = i(s)
If (δ(j) = True) then
Scan := Scan + 1
Distance = d(i,j)
Sum := Sum + Distance
If (Scan > 1) then
Gap = Distance – PreviousDistance
SumGap = SumGap + Gap
MinGap = Min(Gap,MinGap)
Endif
If (Scan = MinScan) then
s(i) = s (initialize Phase 2 at s = s(i))
Break
Endif
PreviousDistance = Distance
Endif
Endfor
If ((Distance < BestDistance)
23
or ((Distance
BestDistance
BestMinGap
BestSumGap
= BestDistance) and (MinGap > BestMinGap)) ) then
= Distance
= MinGap
= SumGap
Endif
Endfor
BestMeanGap = BestSumGap/(MinScan – 1)
TargetGap = λ∙BestMeanGap + (1 – λ)∙BestMinGap) (suggested default, λ = .3)
(The following value is inherited in Phase 2 as a limit for i No )
DistanceLimit = BestDistance + TargetGap
Phase 2 Adaptive MinMax Algorithm
Conventions:
DistanceLimit = the largest distance allowed for d(i,j) when scanning from element i.
BestScan = the value of Scan that produces the best candidate for i*.
Inherited from Phase 1: s(i), i No , and DistanceLimit
Initialization:
BestDistance = Large
BestScan = MinScan – 1
For i N o
Scan = MinScan – 1 (Scan is immediately incremented to MinScan in the next loop)
For s = s(i) to n – 1
j = i(s)
If (δ(j) = True) then (automatically true for s = s(i))
Scan := Scan + 1 (reaches MinScan on first execution, when s = s(i))
If (d(i,j) > DistanceLimit) then
Scan := Scan – 1
Break
Endif)
Distance = d(i,j)
If (Scan = MaxScan) Break
Endif
Endfor
If (Scan MinScan) then (i is a candidate for i*)
If ((Scan > BestScan)
or ((Scan = BestScan) and (Distance < BestDistance))) then
i* = i
BestScan = Scan
BestDistance = Distance
Endif
Endif
Endfor
24
We observe that Approaches 1 and 2 of the “Refinement to Choose Different ClusterSize
Values” for the Primary MinMax Starting Method can also be used with the Adaptive MinMax
Starting Method, where upon restarting the K-MinMax algorithm for the second pass the primary
starting method is used together with setting ClusterSize = ClusterSize(ko ).
The justification of the Adaptive MinMax Starting Method is provided in Theorem 2, following.
As in the discussion that precedes Theorem 1, let v = ClusterSize , where in this case ClusterSize
is not specified in advance but is determined at iteration ko to be the value BestScan + 1 for
BestScan at the conclusion of this iteration. As before, let CSet = {C No : |C| = v}. Now at each
iteration ko we seek to identify a “best” cluster C best that minimizes MMSpan(C) over the clusters
C in CSet, where v is not fixed at the start of iteration ko . Subject to the restriction Scan ≤
MaxScan, which implies v ≤ MaxScan + 1, we undertake to make v as large as possible. Because
of this dependency on the variable value of v, we refer to C best as Cbest (v) and to the best value of
v as v*, hence identifying the best cluster C finally obtained as Cbest (v*).
Theorem 2. The cluster C(ko ) and the associated point i* identified at the conclusion of
iteration ko of the Adaptive MinMax Starting Method respectively qualify to be Cbest (v*) and its
MinMax-centroid.
Proof: As in the proof of Theorem 1, let io (t), t = 1, …, no – 1 identify the subsequence
of i(s), s = 1, …, n – 1 such that δ(j) = True for j = i(s). Hence, defining jt = io (t) (with the
identity of i implicit), the ordering of distances d(i,j) for point i can be written d(i,j1 ) ≤ d(i,j2 ) ≤
… ≤ d(i,jno-1 ). Now the cluster C i consisting of i and the points jt for t = 1 to v – 1 depends on v
as a variable rather than as a constant. The first priority is to make v as large as possible, subject
to requiring v ≤ MaxScan + 1, and more particularly subject to requiring d(i,jv-1 ) ≤ DistanceLimit
where DistanceLimit is determined by BestDistance and TargetGap computed in Phase 1
according to the choice of the parameter λ. The best value v* for v is the value BestScan + 1 for
BestScan as updated in Phase 2, whenever BestScan can be increased subject to BestScan ≤
MaxScan. The final “If … then” instruction in Phase 2 first increases BestScan (hence implicitly
increases v*) when possible, and otherwise decreases the value BestDistance when BestScan
remains unchanged. By this layer of priorities, the method always first increases v*, and for v*
unchanged selects the minimum value d(i,jv*-1 ) for BestDistance. Consequently, by the same
reasoning as in the proof of Theorem 1, it follows that ultimately C best (v*) identifies a cluster that
is the best for the largest admissible value v* for v, and thus qualifies to be treated as the cluster
denoted Cbest in Theorem 1 with i* constituting its associated MinMax-centroid.
The analysis of the complexity of the Primary MinMax Starting Method applies as well to the
Adaptive MinMax Starting Method in spite of its more elaborate form.
This result carries over with the replacement of “MinMax” by “MinSum” and “MMCentroid” by
“MSCentroid” to become applicable to the corresponding starting method for the K-MinSum
problem.
25
8.2 The Adaptive MinSum Starting Method
The Adaptive MinSum Starting Method follows a pattern similar to that of the Adaptive MinMax
Starting Method with some minor departures. As in the case of the Primary MinSum Starting
Method, we include the pseudo-code for the Adaptive MinSum method in Section B of the
Appendix to clearly disclose the details where the two methods differ.
9. A Regret-Threshold PC Algorithm and a Diversified Restarting
Method
We conclude by proposing a modification of the K-PC Algorithm called a Regret-Threshold PC
Algorithm that applies both to the K-MinMax and K-MinSum algorithms, and which can also be
used to modify the K-Means Algorithm by replacing references to the pseudo-centroid with
references to the ordinary centroid.
The strategy of the Regret-Threshold algorithm is to form a candidate list consisting of the
elements that are reassigned to new clusters in Step 2 of the K-PC Algorithm and to permit only
elements belonging to a subset of this candidate list to be reassigned. The elements eligible for
reassignment are determined by a threshold value T which is based on a regret measure, as
follows.
Referring to Step 2 of the K-PC Algorithm, let CandidateList = {j N: j qualifies to be
reassigned to a new cluster C(h)}, and for j CandidateList, let:
i' = Assign(j) denote the pseudo-centroid of the cluster C(h') that j is currently assigned to
i" = ReAssign(j) denote the pseudo-centroid of new cluster C(h") to which j would be
reassigned by Step 2.
ADist(j) = d(i',j) denote the assign distance for j
RDist(j) = d(i",j) denote the reassign distance for j.
Regret(j) = ADist(j) – RDist(j)
Hence, the larger the value of Regret(j), the greater is the regret for not reassigning j to the new
pseudo-centroid ReAssign(j) (and hence to the associated cluster C(h")) in Step 2. These
definitions equally apply in the situation where multiple points may be recorded as pseudocentroids of a cluster C(h), as represented by the set C*(h).
We then use the threshold T by identifying a selected number r of the “largest regret” elements in
CandidateList, as where r is determined to admit some fraction F of these elements for
consideration. As a simple example, F may be chosen from the interval [.05, .2] subject to
assuring that at least 1 element passes the threshold. When F = 1, this approach is the same as the
K-PC approach. (Instead of making F constant, a natural variant is to begin by selecting F larger
and then allowing it to become smaller as the number of iterations grows.)
26
Upon assigning T a value that will accept these specified elements, the eligible elements to be
reassigned in Step 2 are those belonging to
SelectList = {j CandidateList: Regret(j) T}
An upper bound may be additionally imposed on the size of SelectList when ties in the regret
values would cause SelectList to contain more elements than considered desirable.
The foregoing Regret-Threshold approach may be accompanied by a restarting algorithm to
create a new set of initial clusters that differ strategically from the set of clusters produced at the
termination of a current execution, as a basis for obtaining a new set of final clusters on an
ensuing execution. This algorithm, which we call the Diversified Restart Method, may be used
with the earlier K-PC algorithms as well as with the Regret-Threshold approach. Accordingly,
we use the term “Clustering Algorithm” to refer to any of these algorithms.
Diversified Restart Method
Terminology: Let C(h), h K, denote the final set of clusters produced by the Clustering
Algorithm, and let C*(h) denote the set of pseudo-centroids associated with C(h) and C* denote
the union of the sets C*(h), h K. (We allow the option of saving only one element in C*(h).)
For each j N, let h = h(j) identify the cluster C(h) containing j and denote the pseudo-centroid
to which j is assigned by i = PC 1 (j) C*(h). (Thus PC1 (j) is the pseudo-centroid in C* closest to
j.) Finally, let d1 (j) = d(j, PC1 (j)) and let d2 (j) = Min(d(j,i): i C* \ C*(h) for h = h(j)}. (Hence
d1 (j) is the distance from the pseudo-centroid “first closest” to j, while d2 (j) is the distance from
the pseudo-centroid “second closest” to j, restricting attention in the latter case to pseudocentroids outside of C*(h).)
1. Beginning from the final set of Clusters C(h), create a new set of clusters (to start another
execution of the clustering algorithm) by assigning each j N \ C* to the pseudo-centroid PC2 (j)
C* to j.
2. Restart the Clustering Algorithm with the new set of clusters and executing Step 1 of the
Clustering Algorithm to identify new pseudo-centroids for each C(h).
We observe that the first step above associates each j N \ C* with a cluster C(h') (containing
PC2 (j)) that is different from the cluster C(h). An alternative is to reassign only a portion of the
points j N \ C* to an alternative pseudo-centroid before restarting.
More pronounced forms of diversification can be achieved by redefining PC2 (j) to be the pseudocentroid that is “third closest” to j, or in the extreme to be the pseudo-centroid farthest from j.
However, according to the types of diversification often favored in metaheuristic methods, the
definition of PC 2 (j) above is likely to be preferred.
Alternative Measure for Evaluating the Collection of Clusters
The definitions underlying the Diversification Method give a natural way to create a function for
evaluating the clusters, which also can be used to identify good values for k.
27
Let do (j) = d2 (j) – d1 (j) ( 0). The magnitude of do (j) signals the relative appropriateness of
assigning j to a cluster with pseudo-centroid PC 1 (j), as opposed to assigning j to some other
clusters in the current collection. We can exclude elements of C* from having an impact by the
convention do (j) = 0 for j C*, or we can include them by defining d 1 (j) = 0 for j C* in the
definition of do (j).
Then an overall measure of the “quality” of C(h) is given by
Do (h) = ∑(do (j): j C(h)) or by Meano (h) = Do (h)/|C(h)|.
The denominator |C(h)| may be replaced by |C(h)\C*(h)| if pseudo-centroids are excluded from
consideration in defining do (j). Finally, by drawing on Do (h) we may evaluate the entire
collection of clusters C(h), h K by
Value = ∑ (Meano (h): h K)
Here larger values indicate higher quality. 9
The Value term may be considered an alternative to the Davies-Bouldin validity index (Davies
and Bouldin, 1979) which is frequently used to compare the quality of different collections of
clusters. Since Value can meaningfully compare cluster collections for different values of k, this
term can be used with other approaches such as those suggested in earlier sections for evaluating
collections to find a preferred k value.
The do (j) quantities used to define Do (h) and Meano (h) can also serve to create a stochastic
variant of the Diversified Restart approach. In this case, probabilities can be computed for
reassigning points that are inversely related to the magnitude of the do (j) values. Then, as an
extension of the variant that only permits a subset of the elements j N \ C* to be reassigned, a
cutoff value can be established to prevent reassignment of points with d o (j) values exceeding a
specified magnitude, and the probabilities for reassigning remaining elements can be used to
accept or reject such reassignments. Such a strategy reflects the heuristic notion that points with
the largest do (j) values in all likelihood should be maintained in a common cluster.
10. Conclusions
The proposals of the preceding sections provide a wide range of strategies for clustering with
pseudo-centroid methods, and lay a foundation for studies to compare these methods across
alternative classes of clustering problems. Other types of pseudo-centroids, such as a
9
A variation to accentuate the influence of points with larger D o (h) values is to define Meano (h) =
Do (h)2 /|C(h)|.
28
MinProduct-centroid, a MinMedian-centroid, which are defined in the obvious way, give rise to
associated K-MinProduct and K-MinMedian clustering algorithms whose forms can be identified
by analogy with the K-MinMax and K-MinSum algorithms.10 Similarly, we can identify K-PC
algorithms associated with a MinSpan-centroid for Span = Max – Min, and with various “ratiobased” pseudo-centroids, such as a MinSpanRatio-centroid based on a ratio such as Max/Min
(excluding Min = 0, conceiving that two points separated by a 0 distance represent the same
point) and a MinSumRatio-centroid based on a ratio such as Sum/Max ratio (or Sum/Median).
These variations employing more subtle types of pseudo-centroids also invite study.
Another potential area for exploitation arises by noting that the pseudo-centroid approaches may
be contrasted with fuzzy clustering approaches as in Sudha et al. (2012) and imprecise
knowledge clustering approaches as in Anwar et al. (1992). An interesting topic for research
would be to integrate elements of these latter approaches with the pseudo-centroid approaches.
In recent years a variety of metaheuristic algorithms for clustering have emerged to obtain better
outcomes than those produced by classical approaches, which in effect are heuristics for
obtaining local optima for associated evaluation functions (though these functions are not always
clearly defined). While the results and procedures of this paper are offered independent of the
metaheuristic context, we observe that metaheuristic algorithms can be employed to enhance the
effectiveness of these clustering approaches in the same way that metaheuristic algorithms are
used to enhance the K-Means and K-Medoid approaches. Conversely, there is a great potential to
enhance the operation of metaheuristic algorithms by making use of clustering. Proposals of this
type have been made, for example, to improve intensification and diversification strategies for
tabu search and to refine the rules for selecting solutions to be combined by scatter search and
path relinking (which could also apply to solutions combined by genetic algorithms).11 The
notions underlying pseudo-centroid clustering may find applications in these areas as well.
Compliance with Ethical Standards:
Funding: This study was not funded.
Conflict of Interest: The author declares that he has no conflict of interest.
Ethical approval: This article does not contain any studies with human or animal participants.
10
The K-MinProduct problem can also be approached via the K-MinSum algorithm by using logarithms.
The K-MinMedian problem is not to be confused with the K-median problem, which is based on a
different concept relying on coordinate vectors. We can also define a K-MedianMedian algorithm, among
other novel variants.
11
See, for example, Glover (1994, 1997)
29
References
Anderberg, M. R. (1973) Cluster Analysis for Applications, Academic Press.
Anwar, T. M., H. W. Beck and S. B. Navathe (1992) “Knowledge Mining by Imprecise
Querying: A Classification-based Approach,” Proceedings of the Eighth International
Conference on Data Engineering, IEEE Computer Society, Washington, D.C., pp. 622-630.
Cao, B. and F. Glover (2010) “Creating Balanced and Connected Clusters to Improve Service
Delivery Routes in Logistics Planning,” Journal of Systems Science and Systems Engineering,
Vol. 19, No. 4, pp. 453-480.
Cao, B., F. Glover and C. Rego (2015) “A tabu search algorithm for cohesive clustering
problems,” Journal of Heuristics, Vol. 21, pp. 457–477.
Davies, D. L. and D. W. Bouldin (1979) “A Cluster Separation Measure,” IEEE Transactions
Pattern Analysis and Machine Intelligence 1, pp. 224-227.
Estivill-Castro, V. and I. Lee (2001), Fast Spatial Clustering with Different Metrics and in the
Presence of Obstacles. GIS’01, November 9 – 10, pp. 142 – 147.
Fan, Bo (2009), A hybrid spatial data clustering method for site selection: The data driven
approach of GIS mining. Experts Systems with Applications, 36, pp. 3923 – 3936.
Glover, F. (1994) "Tabu Search for Nonlinear and Parametric Optimization (with Links to
Genetic Algorithms)," Discrete Applied Mathematics, 49, pp. 231-255.
Glover, F. (1997) “A Template for Scatter Search and Path Relinking,” in Artificial Evolution,
Lecture Notes in Computer Science, 1363, J.-K. Hao, E. Lutton, E. Ronald , M. Schoenauer and
D. Snyers, Eds. Springer, pp. 13-54.
Huang, Z. (1998) “Extensions to the k-Means Algorithm for Clustering Large Data Sets with
Categorical Values,“ Data Mining and Knowledge Discovery 2, pp. 283-304.
Jain, A. K. and R.C. Dubes (1988) Algorithms for Clustering Data, Prentice-Hall, Englewood
Cliffs, NJ.
Karypis, G., V. Kumar (1998), A Fast and High Quality Multilevel Scheme for Partitioning
Irregular Graphs. SIAM Journal of Scientific Computing 20, pp. 359 -- 392.
Kauffmann, L. and P.J. Rousseeuw (1990) Finding Groups in Data – an Introduction to Cluster
Analysis. Wiley.
30
Kochenberger, G., F. Glover, B. Alidaee and H. Wang (2005) “Clustering of Microarray data
via Clique Partitioning,” Journal of Combinatorial Optimization, 10, pp. 77-92.
Kwon, Y.-J., J. G. Kim, J. Seo, D. H. Lee, and D. S. Kim (2007), A Tabu Search Algorithm
using Voronoi Diagram for the Capacitated Vehicle Routing Problem, Proceeding of 5th
International Conference on Computational Science and Applications, IEEE Computer Society,
pp. 480 – 485.
MacQueen, J.B. (1967) “Some Methods for classification and Analysis of Multivariate
Observations,” Proceedings of 5-th Berkeley Symposium on Mathematical Statistics and
Probability", Berkeley, University of California Press, 1 pp. 281 -- 297.
Ng, R. and J. Han (2002) “CLARANS: a Method for Custering Objects for Satial Data Mining,”
IEEE Transactions Knowledge of Data Engineering 14(5), pp. 1003-1016.
Paivinen, N. (2005) “Clustering with a minimum spanning tree of scale-free-like structure,”.
Pattern Recognition Letters, 26(7) pp. 921–930.
Park, H-S. and C-H. Jun (2009) “A simple and fast algorithm for K-medoids clustering,” Expert
Systems with Applications 36, pp. 3336-3341.
Ralambondrainy, H. (1995) “A conceptual version of the k-means algorithm,” Pattern
Recognition Letters, 16:1147–1157.
Shamsul, A., M. Inostroza-Ponta, L. Mathieson,, R. Berretta, and P. Moscato (2011) “Clustering
Nodes in Large-Scale Biological Networks Using External Memory Algorithms,” In Xiang et al.
(Eds.): ICA3PP 2011 Workshops, Part II, LNCS 7017, pp. 375–386.
Strehl, A. and J. Ghosh (2002) “Relationship-Based Clustering and Visualization for HighDimensional Data Mining,” INFORMS Journal on Computing, pp. 1 – 23.
Sudha, K.R., Y.B. Raju and A.C. Sekhar (2012) “Fuzzy C-Means clustering for robust
decentralized load frequency control of interconnected power system with Generation Rate
Constraint,” Electrical Power and Energy Systems 37, pp. 58-66.
Xu, Y., V. Olman, and D. Xu (2001) “Minimum spanning trees for gene expression data
clustering,” Genome Informatics, 12 pp. 24–33.
31
Appendix: Additional Supporting Methods and Pseudocode
A. Diversity-Based Starting Methods
1. A Targeted Diversity-Based Starting Method to Produce More Compact
Sets of Points
A more elaborate variant of the Targeted Diversity-Based Starting Method stems from the
observation that for some number of iterations, there are likely to be numerous ties for the point
i' = arg min(Deviation(i): i Io ), that is, there may be many points i' whose minimum distance
from the points j Ho is close to T (yielding Deviation(i') close to 0), and we would prefer to
choose among them in a way that keeps the growing collection of points in H o compact. This
potential for ties is particularly evident if we choose T to be the value MinD(i(k)) which is the
final MaxMin value when the last element i(k o ) = i(k) is added to Ho by the Simple DiversityBased Starting Method. Since the point i(k) lies at a distance at least T = MinD(i(k)) from all
other points in Ho , selecting i(k) as the starting point i(1) yields numerous points i for which
Deviation(i) = |MinD(i) – T| is close to 0. As a result, a sequence of points chosen to populate Ho
could initially resemble a straight line (and subsequently a bent line with gradual fill- in), whose
endpoints get progressively farther away from the points added earlier. Such a “line
configuration” would be encouraged, for example, by breaking ties in the choice of i' to favor
points farthest from the most recent point i(k o ) added to Ho .
We introduce two refinements as a basis for generating a more compact collection of points in
Ho . First, we identify a target T and a starting point i(1) as follows. Relative to the set Ho =
{i(ko ): ko = 1 to k} generated from an execution of the Simple Diversity-Based method or one of
its refined variants, identify the min distance for each i = i(ko ) Ho from the other points in Ho ,
given by MinDo (i) = Min(d(i,j): j Ho \ {i}). (This is effectively equivalent to the previous
definition of MinD(i), since the latter is restricted to elements i not in Ho .) Then we identify the
mean of these MinDo (i) values given by
MeanMinD = ∑MinDo (i): i Ho )/k
Finally, we select the target T = MeanMinD and pick a point i# of Ho to become the new i(1) (to
generate a new Ho ) whose MinDo (i) value is closest to MeanMinD:
i# = arg min(|MinDo (i) – MeanMinD|: i Ho )
The choice of T = MeanMinD and i(1) = i# to launch the targeted Diversity-Based Starting
Method is then accompanied by introducing a tie-breaking rule for choosing i' that minimizes the
maximum distance from the points currently in Ho .
32
For greater latitude of choice, we choose a small value To and require all qualifying points to
belong to the set Io ' = {i Io : Deviation(i) ≤ MinDev + To }, where we define MinDev =
Min(Deviation(i): i Io ) (hence all points that satisfy Deviation(i) = MinDev qualify to be
chosen as i' in the Targeted Diversity-Based method). For example, To can be obtained by setting
To = f∙T for is a small non-negative fraction f. (When f = 0 and hence To = 0, the set Io ' consists
strictly of the points that qualify to be i' in the Simple Targeted Diversity-Based method.)
Finally we define MaxD(i) = Max(d(i,j): j Ho ) and choose i', the element to be added to Ho at
each iteration, by i' = arg min(MaxD(i): i Io '), thereby minimizing the maximum distance from
points in Ho , given that i' lies in Io '. This gives the following method.
Targeted Diversity-Based Starting Method with Flexible Tie-Breaking
Initialization: Choose the point i(1) = i# and the target T = MeanMinD, and set ko = 1, Ho =
{i(1)} and Io = N \ {i(1)}.
For ko = 2 to k:
Identify MinD(i) = Min(d(i,j): j Ho ), Deviation(i) = |MinD(i) – T| for each i Io ,
MinDev = Min(Deviation(i): i Io ), Io ' = {i Io : Deviation(i) ≤ MinDev + To }
and for i Io ', identify MaxD(i) = Max(d(i,j): j Ho ).
Choose i' = arg min(MaxD(i): i Io ') and set i(ko ) = i', Ho := Ho {i'} and
Io := Io \ {i'}
EndFor
This approach likewise can be iterated, by choosing a new the target T = MeanMinD and starting
point i(1) = i# from the Ho most recently generated. A natural variant of the method results by
replacing MaxD(i) with SumD(i) = ∑(d(i,j): j Ho ).
The earlier Diversity-Based Starting Methods of Section 6 can similarly be modified to generate
more compact sets of points by defining MaxMin = Max(MinD(i): i Io ) and Io ' = {i Io :
MinD(i) MaxMin – To }. Then as in the Targeted Tie-Breaking Approach above, for i Io ' we
identify MaxD(i) = Max(d(i,j): j Ho ) and choose i' = arg min(MaxD(i): i Io ') (or alternatively
replace MaxD(i) with SumD(i)). The starting point i(1) = i# can also be used, without reference to
the target T.
2. Successive Elimination Diversity-Based Starting Methods
Another type of diversity-based starting methods employs a successive elimination strategy that
shares an organization related to that of the Intensity-Based methods, except that it strictly aims
to create a diverse set of points, again denoted by Ho , instead of to generate clusters at the same
time. The basic strategy is as follows.
33
Initialization: Set No = N, no = n and Ho = .
For ko = 1 to k:
Apply an Elimination Choice Rule (as identified below) to select a point i' No. and let Ho
:= Ho {i'} and No := No \Proximity(i'), where Proximity(i') denotes the set of no /ko
points j No \{i'} that are closest to i. .
EndFor
A trivial possibility for the Elimination Choice Rule would be to simply choose i' randomly from
No . More interesting possibilities are as follows, where for each point i No , let Proximity(i) be
defined the same as the set Proximity(i'), i.e., the set of no /ko points j N o \{i} that are closest
to i.
Elimination Choice Rule. Let PC(i) = a PC-centroid of Proximity(i) according to the case in
question, and let V(i) = Span(Proximity(i)). (For example, in the MinMax case, PC(i) is a
point i* in Proximity(i) whose maximum distance from all other points in Proximity(i) is
smallest, and V(i) = MaxDist(i*: Proximity(i).) Then choose i' in N o to be a point that gives
a value of V(i) that is (a) largest, (b) smallest or (c) ‘closest to the mean’ of such values for
i No.
Using an instance (a), (b) or (c) of the preceding rule, an iterated version of the method can be
employed by successively restarting again from ko = 1, pre-selecting the value i' for ko = 1 to be
the “final i' value” for ko = k (chosen at the end of the preceding iteration). In this iterated
approach, a variation is to draw on ideas related to those of other Diversity-Based methods by redefining Proximity(i) after the first iteration (over ko = 1 to k) to consist not just of the no /ko
points j No \{i} that are closest to i, but to encompass all points j No such that d(i,j) ≤ U,
where U is the mean or maximum d(i',j) value for j Proximity(i') over i' Ho on the previous
iteration. After the first iteration of the method over ko = 1 to k, this approach requires less
computation than determining the no /ko points closest to i. Because of the increased speed of
using a definition of Proximity(i) based on U instead of no /ko , a value of U may be selected
earlier, after examining only one or a few ko values, and subsequently adjusted at each new
iteration according to the numbers of elements admitted to Proximity(i) on the preceding
iteration. Referring to U may exhaust all elements of N before Ho contains k elements, and if U is
estimated to be of reasonable size, this may signal that k is selected too large.
B. Intensity-Based Starting Methods
1. The Primary MinSum Starting Method
We now consider the intensity-based starting method associated with the MinSum-centroid. In
the following we omit explanatory comments for pseudocode that is made clear by the
observations for the corresponding Primary MinMax Starting Method.
34
Primary MinSum Method
Initialization: Begin with the ordered indexes i(1), i(2), …, i(n-1) for each i N and set N o = N,
no = n and δ(j) = True, j N.
For ko = k to 1
ClusterSize = no /ko
ScanSize = ClusterSize – 1
Execute the MinSum Distance Algorithm
(At the conclusion: i* = arg min(Max (d(i,,j): j C) (= MSSpan(C)) over all sets
C No such that |C| = ClusterSize.)
Generate C(ko ) and update the set N o
(This portion of the Primary MinSum Method is identical to that of the Primary MinMax
Method, except that MSCentroid(ko ) = i* replaces MMCentroid(ko ) = i*.)
Endfor
The internal MinSum Distance algorithm is also nearly a replica of the internal MinMax
Distance algorithm upon replacing BestDistance with a corresponding value BestSum. All arrays
and values of the MinSum Distance Algorithm are shared with the Primary MinSum Starting
Method.
MinSum Distance Algorithm
Conventions:
BestSum = the MinSum value ∑(d(i,j): = d(i*,j) associated with the best candidate for i*.
Initialization:
BestSum = Large
For i N o :
Sum = 0
Scan = 0
For s = 1 to n – 1
j = i(s)
If (δ(j) = True) then
Sum := Sum + d(i,j)
Scan := Scan + 1
If (Scan = ScanSize) Break
Endif
Endfor
If (Sum < BestSum) then
i* = i
BestSum = Sum
Endif
35
Endfor
An analog of Theorem 1 applies to this Primary MinSum Starting Method, which we do not
bother to state here.
2. Advanced Intensity-Based Starting: The Adaptive MinSum Starting
Method
The description of the “external” portion of the Adaptive MinSum method, given next, is nearly
identical to that of the Adaptive MinMax Starting Method. However, the internal Adaptive Phase
1 and Phase 2 algorithms differ more significantly from the MinMax case, as indicated below.
Adaptive MinSum Method
Initialization: Begin with the ordered indexes i(1), i(2), …, i(n-1) for each i N and set N o = N,
no = n and δ(j) = True, j N.
For ko = k to 1
MinSize = no /ko
MaxSize = no – MinSize∙(ko – 1)
MinScan = MinSize – 1
MaxScan = MaxSize – 1
Execute the Phase 1 MinSum Algorithm (to identify BestSum and associated values)
Execute the Phase 2 MinSum Algorithm (to identify i*)
Generate C(ko ) and update the set No
(This portion of the Adaptive MinSum Method is identical to that of the Primary
MinSum Method, except that “If (Scan = BestScan) Break” replaces
“If (Scan = ScanSize) Break”.)
Endfor
Phase 1 MinSum Algorithm
Assumption: No contains at least 2 elements
BestDistance = Large
BestMinGap = 0
BestSum = Large
For i N o
MinGap = Large
Sum = 0
SumGap = 0
Scan = 0
For s = 1 to n – 1
j = i(s)
36
If (δ(j) = True) then
Scan := Scan + 1
Distance = d(i,j)
Sum := Sum + Distance
If (Scan > 1) then
Gap = Distance – PreviousDistance
SumGap = SumGap + Gap
MinGap = Min(Gap,MinGap)
Endif
If (Scan = MinScan) then
Sum(i) = Sum (initialize Phase 2 at Sum = Sum(i))
s(i) = s (initialize Phase 2 at s = s(i))
Break
Endif
PreviousDistance = Distance
Endif
Endfor
If ((Sum < BestSum)
or ((Sum = BestSum) and (Distance < BestDistance))
or ((Sum = BestSum) and (Distance = BestDistance)
and (MinGap > BestMinGap)) ) then
BestSum = Sum
BestDistance = Distance
BestSumGap = SumGap
BestMinGap = MinGap
Endif
Endfor
BestMeanGap = BestSumGap/(MinScan – 1)
TargetGap = λ∙BestMeanGap + (1 – λ)∙BestMinGap) (suggested default, λ = .3)
(The following 3 values are inherited to apply for each i N o in Phase 2)
DistanceLimit = BestDistance + TargetGap (applies to MinMax version)
FirstSumLimit = BestSum + TargetGap
DeltaSum = DistanceLimit
Phase 2 Adaptive MinSum Algorithm
Conventions:
Sum = Sum d(i,j) using the ordered scanning of elements from i N o
BestSum = the Sum value for the current best candidate for i*.
DistanceLimit = the largest distance allowed for d(i,j) when scanning from element i.
SumLimit = the limit on Sum, increasing by DeltaSum at each Scan > MinScan
BestScan = the value of Scan that produces the best candidate for i*.
Inherited from Phase 1:
37
DistanceLimit, FirstSumLimit, DeltaSum and s(i), Sum(i), i No
Initialization:
BestDistance = Large
BestScan = Large
BestSum = Large
For i N o
Scan = MinScan – 1 (Scan will be immediately incremented to MinScan)
SumLimit = FirstSumLimit
Sum = Sum(i)
For s = s(i) to n – 1
j = i(s)
If (δ(j) = True) then (automatically true for s = s(i))
Scan := Scan + 1 (reaches MinScan on first execution when s = s(i))
If (Scan > MinScan) then
Sum := Sum + d(i,j)
SumLimit := SumLimit + DeltaSum
Endif
If (Sum > SumLimit) then
Scan := Scan – 1
Break (exit the “For loop”)
Endif
If (Scan = MaxScan) Break
Endif
Endfor
If (Scan MinScan) then (i is a candidate for i*)
If ((Scan > BestScan)
or ((Scan = BestScan) and (Sum < BestSum))) then
i* = i
BestSum = Sum
BestScan = Scan
Endif
Endif
Endfor
The analog of Theorem 2 and its proof that applies to the Adaptive MinSum Starting Method
gives BestSum the role taken by BestDisance, but otherwise establishes the same double layer of
priorities to assure the final C best (v*) (for v* = BestScan + 1) and its associated point i* give the
cluster and MinSum-centroid we seek.
38
| 8cs.DS
|
1
Automatic Pavement Crack Detection
Based on Structured Prediction
with the Convolutional Neural Network
arXiv:1802.02208v1 [cs.CV] 1 Feb 2018
Zhun Fan, Senior Member, IEEE, Yuming Wu, Jiewei Lu, and Wenji Li
Abstract—Automated pavement crack detection is a challenging task that has been researched for decades due to the
complicated pavement conditions in real world. In this paper, a
supervised method based on deep learning is proposed, which
has the capability of dealing with different pavement conditions.
Specifically, a convolutional neural network (CNN) is used to
learn the structure of the cracks from raw images, without any
preprocessing. Small patches are extracted from crack images as
inputs to generate a large training database, a CNN is trained
and crack detection is modeled as a multi-label classification
problem. Typically, crack pixels are much fewer than non-crack
pixels. To deal with the problem with severely imbalanced data, a
strategy with modifying the ratio of positive to negative samples
is proposed. The method is tested on two public databases and
compared with five existing methods. Experimental results show
that it outperforms the other methods.
Index Terms—Pavement crack detection, convolutional neural
network, deep learning, structured prediction, multi-label classification, imbalanced data.
I. I NTRODUCTION
P
AVEMENT distress detection is a significant step for
pavement management [1]. The goal is to obtain the
pavement condition information for maintenance. Cracking is
the most common type of pavement distress. Since automated
crack detection systems are safer, lower costing, more efficient,
and more objective, the research about them has attracted wide
attention from both the academy and the industry [2], [3].
Automated crack detection is a challenging task. In the past
few decades, image based algorithms of crack detection have
been widely discussed. In early studies, methods are mostly
based on combination or improvement of conventional digital
image processing techniques, like thresholding [4], mathematical morphology [5], and edge detection [6]. These methods
are generally based on photometric and geometric hypotheses
about properties of crack images [7]. The most distinguished
photometric property is that crack pixels are the darker pixels
in an image. Based on that, threshold value is determined
globally or locally to segment cracks and background [4], [8],
[9]. However, these approaches are very sensitive to noise as
they are performed on individual pixels. In order to overcome
this problem, other methods take geometric information into
account. For instance, the continuity property of cracks is
considered to reduce the false detection [5]. Local binary
Zhun Fan, Yuming Wu and Jiewei Lu are with the Guangdong Provincial
Key Laboratory of Digital Signal and Image Processing, College of Engineering, Shantou University, Shantou 515063, China (email: zfan, 16ymwu1,
[email protected]).
pattern operator is used to determine whether a pixel belongs
to cracks based on the local orientation [10]. Wavelet transform
is applied to separate crack regions and crack free regions
with multi-scale analysis [11]. These methods detect cracks
efficiently, but they are not precise enough for finding all the
cracks in an image. More recent works have been proposed
to improve the accuracy of the detection method, and can be
sorted into the following several branches.
Minimal Path Based Methods: Minimal path problem is
to find best paths between nodes of a graph. Several methods
based on the minimal path principle are proposed. Kaul et
al. propose an algorithm that can find an open curve without
knowing either the endpoints or topology of the curve [12].
Nguyen et al. take brightness and connectivity into account
simultaneously to provide features of anisotropic cracks along
free-form paths [13]. Amhaz et al. propose strategies to select
endpoints at a local scale and to select minimal paths at the
global scale [14]. These methods utilize distinguished features
of crack pixels in a global view. However, the drawback
is that they are too computationally intensive for practical
applications.
Machine Learning: With the development of machine
learning, several methods focusing on feature extraction and
pattern recognition have been proposed for crack detection
[15]–[18]. In CrackIT [16], Oliveira et al. use the mean and
the standard deviation for unsupervised learning to distinguish
blocks with crack from blocks without crack. Cord et al. use
AdaBoost to select textural descriptors that can describe crack
images [17]. In CrackForest [18], Shi et al. propose a new
descriptor based on random structured forests to characterize
cracks. The performance of these methods are great but
very dependent on the extracted features. However, due to
complicated pavement conditions, it is hard to find features
effective for all pavements.
Deep Learning: For image classification tasks, deep learning has been proven to work better than traditional machine
learning methods [19]. Recently, deep learning methods are
successfully applied on damage and distress detection [20]–
[23]. Cha et al. use a sliding window to divide the image into
blocks and CNN is used to predict whether the block contains
cracks or not [20]. But the method can only find patch level
cracks without considering the pixel level. In [23], Zhang et
al. use CNN to predict whether an individual pixel belongs
to crack based on the local patch information. However,
the method ignores the spatial relations between pixels and
overestimates crack width. In [21], Zhang et al. use CNN to
2
Fig. 1. Two examples of structured prediction based on CNN. (a) Two pixel examples of the original image. (x1 , y1 ), (x2 , y2 ) are the coordinate of the
pixel. (b) The extracted patches centered on the pixels. In the method the size of the patches is 27 × 27. (c) The structured prediction results of CNN with
the patches input. (c1 ) is obtained from (b1 ) and (c2 ) is obtained from (b2 ). In the method the size of the output structure is 5 × 5. (d) The probability map
obtained by applying structured prediction on all pixels.
predict class for each pixel of the image. However, it still
needs manually designed feature extractors for preprocessing,
and CNN is only used as a classifier. Besides, its network
architecture is strictly related to the input image size, which
prevents generalization of the method.
Since CNN can extract good features from raw data, this
paper propose a CNN based method that can learn the
crack structure of a small patch within an image to find the
whole crack on pixel level without preprocessing. Specifically,
patches are exacted on each individual pixel of a raw image
to build a training database. Then a CNN architecture is
proposed for structured prediction as a multi-label problem
[24]. Since crack pixels are much less than non-crack pixels
in a typical image, applying structured prediction directly
can hardly obtain a satisfied result. To solve the multi-label
classification problem with imbalanced samples, we propose
a strategy with modifying the ratio of positive to negative
training samples. After training, the CNN can predict the
centered structure of each input patch. Then all structure
results of an image are summed to obtain a probability output,
which reflects the condition of pavement generally, as shown
in Fig. 1. Finally a decision threshold is set to obtain a binary
output. The proposed method is compared with traditional
methods like Canny [25] and local thresholding [26], and the
state-of-art methods like Free-Form Anisotropy (FFA) [13],
CrackForest [18], and Minimal Path Selection (MPS) [14]) to
validate its effectiveness. The contributions of the paper are the
following: 1) apply CNN based structured prediction on crack
detection and obtain an outstanding performance; 2) propose
a strategy to solve the multi-label classification problem with
imbalanced samples and analyze its influence.
This paper is organized as the following. Section II describes the implementing details of the convolutional neural network, including data preparation, the architecture of
CNN and training details. Section III gives the results and
evaluations. Section IV describes the structured prediction in
detail. Section V describes the influence of the imbalanced
samples and the strategy to deal with it. The generalization
of the method is discussed in Section VI. Conclusions and
perspectives are given in Section VII.
Fig. 2. The extracted patch centered on pixel (x, y). h is the distance between
the edges and the center. Thus the size of the patch is (2h + 1) × (2h + 1).
II. T HE M ETHOD
A. Data Preparation
The proposed method is trained and tested on two databases:
CFD [18]1 with RGB images and AigleRN [7]2 with gray-level
images. The CFD database contains 118 RGB images with
a resolution of 320 × 480 pixels. All images are taken by
an iphone5 from pavements of Beijing, China. These images
contain noises such as shadows, oil spots and water stains, and
have non-uniform illumination. We use 72 images for training
and 46 images for testing. The AigleRN database contains
38 gray-level images, which have been collected on French
pavement and have been pre-processed to reduce the nonuniform illumination. Compared to the images in CFD, the
pavement images in AigleRN are with more complex texture.
We use 24 images for training and 14 images for testing.
The input of the network are patches extracted from images
based on each specific pixel, as shown in Fig. 2. In this work,
we set the patch size as 27 × 27 and thus h = 13. Typically,
RGB images with 3 channels and gray-level images with 1
channel are used as input of the network. So we propose a
network with 3 channels input and a network with 1 channel
input respectively.
In a typical crack image, non-crack pixels are a lot more
than crack pixels. For a classification task with imbalanced
samples, the proportion of positive and negative samples
has a great impact on the network performance. Specifically,
according to the manually labeled ground truth, if the center
pixel of the extracted patch is a crack pixel, the corresponding
1 https://github.com/cuilimeng/CrackForest-dataset
2 https://www.irit.fr/∼Sylvie.Chambon/Crack
Detection Database.html
3
Fig. 3. Example of negative (left) and positive (right) 27×27 training patches
extracted from the CFD database.
TABLE I
T RAINING DATABASE
CFD
Image dimensions
Images
Positive patches
Negative patches
Total patches
Positive : Negative
320×480
Training
72
185,483
556,449
741,932
1:3
Testing
46
107,112
6,958,488
7,065,600
1:65
we use s to denote the size of the output structure. Thus s = 5
as shown in Fig. 1(c). The s × s window is flattened into s2
neurons in the network, as the output layer of the CNN shown
in Fig. 4, and is fully connected to the second last layer. Thus
the structured prediction problem is modeled as an multi-label
problem and s2 labels are from the corresponding ground truth.
Details about the multi-label problem is discussed in Section
IV.
For a multi-label classification problem, sigmoid function is
the commonly used activation function for the output. Except
the output units, all hidden layers are equipped with Rectified
Linear Units (ReLU) [28] for dealing with nonlinearity.
C. Training and Testing
AigleRN
991×462/311×462
Training
24
39,208
117,624
156,832
1:3
Testing
14
42,317
4,168,351
4,210,668
1:98.5
patch is a positive sample and vice versa. To modify the
proportion and keep the samples uniform, for training, all
positive samples are extracted first. Then negative samples
are extracted randomly according to the setting proportion.
Generally, the ratio of positive to negative samples is set to 1:3,
which is discussed in Section V. Then the training database can
be obtained as shown in Fig. 3. In testing procedure, all pixels
from an image are used and symmetric padding is used to deal
with the boundary of the image. As a result, we can obtain
all samples from an image in testing. For instance, 153,600
samples can be obtained in a 320×480 image. Details of the
training and testing database are shown in Table I.
Data with zero mean and equal variance is preferred for
optimization in machine learning. For image input, a common
practice is scaling the raw data from [0, 255] to [-1, 1]. This
preprocessing is adopted our work.
B. Network Architecture
Typical CNN contains convolutional layers, max-pooling
layers and fully-connection (FC) layers. The proposed network
has 4 convolutional layers with 2 max-pooling layers and 3
FC layers. The architecture of the network is illustrated in
Fig. 4 and Table II. Extracted patches are used as input of the
network with 3 channels in CFD and 1 channel in AigleRN.
All convolutional layers are equipped with kernel of 3 × 3 and
stride of 1, whose effectiveness has been verified in VGG-Net
[27]. Zeros padding on the boundary of each convolutional
layer is assigned to preserve the spatial resolution of resulting
feature maps after convolution. Max pooling is performed with
stride 2 over a 2 × 2 window.
As shown in Fig. 1, the output of the network is a prediction
of the structure centered in the input patch. For generalization,
In a multi-label problem, all output units are not mutually
exclusive, which means that more than one positive output may
exist at the same time. In this case, cross entropy is commonly
used in the loss function, which is defined as:
2
L=−
s
X
(yi log yˆi + (1 − yi ) log(1 − yˆi ))
i=1
where yi and yˆi are the label and prediction of the ith output
unit, respectively; s2 is the number of labels. To prevent the
network from overfitting, weight decay is adopted in the paper
to penalize the large weights. The L2 penalty term is added
to the loss function and loss L0 used for optimization is:
1X 2
Wj
L0 = L + β ·
2 j
where L is the cross entropy function, β is the L2 penalty
factor and Wj is the jth weight in the network, including
weights in convolutional layers and FC layers. β is set to
0.0005 in the experiment.
Dropout [29] is also a simple and effective approach for
preventing overfitting. During training, value of hidden neurons is set to zero randomly. It is applied on the first two FC
layers and the dropout ratio is set to 0.5.
Weights in convolutional layers and FC layers are initialized
using Xavier method [30]. Adam [31] is adopted as the
optimizer with default setting in the original paper (learning
rate set to 0.001). The batch size for each iteration is set to
256, and 20,000 iterations (43 epochs) in AigleRN and 30,000
iterations (13 epochs) in CFD are conducted.
In testing, each pixel is used for generating an input patch,
so input patches overlap, leading to overlap of output windows.
Typically, s2 decisions can be obtained from a pixel except that
it is near the boundary. Then for a testing image, all outputs
from a pixel are summed and normalized to [0,1]. As a result,
we can obtain a probability map from the results of all the
pixels of an image.
III. R ESULTS
The network is programmed under TensorFlow. All experiments are conducted on the workstation with Intel Xeon E52690 2.9GHz CPU, 64GB RAM and Nvidia Quadro K5000
GPU.
4
Fig. 4. An illustration of the CNN architecture. The leftmost image is the input patch with 3 channels. Other cubes indicate the feature maps obtained from
convolution (Conv) or max pooling. All convolutional layers are with kernel of 3 × 3, stride 1 and zeros padding. Max pooling is performed with stride 2
over a 2 × 2 window.
TABLE II
A RCHITECTURE OF THE CNN. L AYER NAMES ARE F OLLOWED BY N UMBERS OF F EATURE M APS .
S QUARE B RACKETS S PECIFY K ERNEL S IZE , S TRIDE AND PADDING .
Input
→
conv16
[3×3,1,1]
→
conv16
[3×3,1,1]
→
maxpool
[2×2,2,0]
→
conv32
[3×3,1,1]
→
conv32
[3×3,1,1]
→
maxpool
[2×2,2,0]
→ FC64
→ FC64
→ FC25
A. Evaluation
We test the performance of the network on test images.
For evaluation, precision (P r), recall (Re) and F1 score (F 1)
are introduced, which are commonly used in classification
problems, and defined as:
TP
TP + FP
TP
Re =
TP + FN
2 × P r × Re
F1 =
P r + Re
where T P , F P , F N are the numbers of true positive, false
positive and false negative respectively. Considering that the
ground truth is subjective manually labels and there are
transitional areas between crack pixels and non-crack pixels
in real images, small distance between the detected result and
the reference, which is 2 pixels in [14] and 5 pixels in [18],
is accepted for evaluation. We accept 2 pixels in the paper.
Pr =
B. Results on CFD
Exemplar detections on CFD are shown in Fig. 5. The
output of the network is a probability map as shown on the
third row. Whiter color indicates that the pixel is more likely
to be a crack pixel. Then the decision probability is set to
0.5 to remove the pixels with low probability, and a binary
output is obtained. It can be observed that only few pixels
are removed from binarization. The reason is that the contrast
between the detected crack pixels and the detected non-crack
pixels in the probability map is high. According to the output
images in Fig. 5, our method can effectively deal with complex
background and complex crack topology.
Fig. 6 and Table III shows the comparison with some other
algorithms. It can be observed that the two traditional methods
Fig. 5. Part of crack detection results by the proposed method (from top to
bottom: original image, ground truth, probability map, binary output).
are not suitable for crack detection as they are sensitive to
noises. Since both CrackForest and the proposed method are
supervised learning methods, we generate the same training
and testing databases for them. Notice that crack width is
overestimated in CrackForest, leading to the high recall but
low precision as shown in Table III. Evaluation results shown
in Table III indicate that the proposed method outperforms
CrackForest.
5
Fig. 6. Results comparing on CFD (from left to right: original image, ground truth, Canny, local thresholding, CrackForest, the proposed method).
TABLE III
C RACK D ETECTION R ESULTS E VALUATION ON CFD
Canny
Local thresholding
CrackForest
The proposed method
Pr
Re
F1
0.4377
0.7727
0.7466
0.9119
0.7307
0.8274
0.9514
0.9481
0.4570
0.7418
0.8318
0.9244
TABLE IV
C RACK D ETECTION R ESULTS E VALUATION ON A IGLE RN
Canny
Local thresholding
FFA
MPS
The proposed method
Pr
Re
F1
0.1989
0.5329
0.7688
0.8263
0.9178
0.6753
0.9345
0.6812
0.8410
0.8812
0.2881
0.6670
0.6817
0.8195
0.8954
C. Results on AigleRN
Fig. 7 and Table IV present the results of different algorithms on AigleRN. It is again observed that the two traditional
methods are sensitive to noise and fail to find the continuous
cracks. FFA can find some cracks in a local scale, but can
hardly find all the continuous cracks. MPS is more effective
to find the continuous cracks on a global view. Compared to
MPS, our proposed method can find the cracks with better
precision and higher recall. From image results, it is clear that
the proposed method outperforms the others.
IV. O UTPUT S TRUCTURE
Generally, CNN is used for classification [19]. Typical
classification problems contain single-label classification and
multi-label classification. In a single-label classification problem, a sample can only be classified into one class. For
example, in [23], crack detection is modeled as a binary
classification problem and a pixel is classified as a crack pixel
or a non-crack pixel based on its neighbouring information.
However, in a multi-label classification problem, a sample
can be classified into more than one class. Structured prediction based on CNN, modeled as a multi-label classification
problem, is quite a new idea [24]. The method utilizes the
information of a patch to predict not only the centered pixel
but the centered structure. This operation takes the spacial
relations between pixels into account, i.e., a crack pixel is
more likely to be neighboured by another crack pixel and vice
versa.
In this experiment, we explore the performance of the structured prediction. The size of the output structure is denoted
by s. Results presented are obtained for s = 1, 3, 5, 7, i.e., the
number of output units is set to 1, 9, 25, 49, respectively. Other
configurations remain unchanged. Specifically, for s = 1,
the model is simplified to the pixel classification problem
similar to [23]. The performance with different sizes of output
structure is shown in Fig. 8.
From Fig. 8, it can be observed that all results for structured
prediction, particularly for s = 5, are better than that for pixel
classification. To identify the difference between the structured
prediction and the pixel classification, we present the probability maps for s = 1 and s = 5, as shown in Fig. 9. Both
of them can detect the cracks but pixel classification is much
more sensitive to noises. This is natural when considering the
procedure of patches extraction and classification. Classifying
a pixel based on its neighbouring information is the original
design. However, the network is essentially classifying the
patches in the procedure. Thus, limited by samples, it may
be trapped in overfitting and misunderstand the classification
problem. For example, if the mean of the extracted patches
distinguishes between positive samples and negative samples,
CNN may essentially turn to a classifier of patches based
on the mean. The mean is a good feature for patches but is
not enough to classify the pixel, since it is sensitive to the
interference like non-uniform illumination. It is very confusing
as we can not ensure what is CNN based on to classify the
patches. Structured prediction can solve this problem to some
extent, as it aims to predict several outputs at the same time.
The outputs between patch to patch differ slightly based on the
labels. Therefore, the network can not rely on single feature
6
Fig. 7. Results of different methods on AigleRN (from left to right: original image, ground truth, Canny, local thresholding, FFA, MPS, the proposed method).
to accurately predict all outputs of a patch. This prevents the
network from misunderstanding the problem.
Another explanation is for the testing procedure. In pixel
classification, only one decision is obtained from the network
for a pixel. However, in structured prediction, s2 decisions
can be obtained for a pixel by scanning all pixels through the
whole image. These decisions are summed to obtain a more
accurate result. Besides, normalizing the whole output image
to [0,1] gives a global view to smooth the detected pixels.
V. R ATIO OF P OSITIVE TO N EGATIVE S AMPLES
Applying structured prediction directly can hardly obtain
a satisfying result. As is shown in Table I, crack pixels are
much fewer than non-crack pixels in typical crack images. The
ratio of positive to negative samples is down to 1:65 in CFD,
and even down to 1:98.5 in AigleRN. In such a classification
task with severely imbalanced samples, the ratio of positive to
negative has a great impact on the training results. To clarify
the problem, assume that the natural ratio(i.e., 1:65) is used
in training, if the predicted results of CNN are all negative,
i.e., it predicts all pixels to be non-crack pixels, it can obtain a
high accuracy as 65/(65 + 1) = 0.985. Thus if no treatment is
conducted to deal with imbalanced data, the network prefers
predicting negative results in training, which would lead to
overestimation of the number of non-crack pixels in testing,
as shown in Fig. 11.
Typically, in a multi-label problem, one output unit represents one class. Specifically in our method, an output unit
represents a pixel on the particular position. For output units,
a significant characteristic is that if a crack pixel appears, it
is likely to find crack pixels on other units. Based on this
characteristic, we introduce the terms positive sample and
negative sample that used in binary classification, as we just
care about the probabilities of crack pixels and non-crack
7
Fig. 8. Precision, recall and F1 score value variations with different sizes of
output structure. Testing on CFD and AigleRN respectively.
Fig. 10. Precision, recall and F1 score value variations with different ratios
of positive to negative samples. Testing on CFD and AigleRN respectively.
Fig. 9. Example probability maps with s = 1 (middle) and s = 5 (right) for
the original image (left).
(a) Original image
(b) R = 1
pixels appearing. The particular position is not so important.
According to the patch extraction procedure, we define an
input extracted from a crack pixel as a positive sample, and an
input extracted from a non-crack pixel as a negative sample.
Experiments are conducted to make a thorough investigation
on the influence of the ratio.
In the experiment, the total number of positive samples and
negative samples are kept constant, and the ratio of positive
to negative samples is modified. The ratio R is defined as:
(c) R = 3
(d) R = 5
(e) R = 59
(f) Natural(R = 184)
R=
negative samples
positive samples
For example, if we keep the number of total samples to be
120, R = 1 means 60 positive samples and 60 negative
samples, while R = 2 means 40 positive samples and 80
negative samples. We set the number of total training samples
to be 360,000 for CFD and 60,000 for AigleRN. All training
samples are generated from training dataset randomly, and we
test the model with different training conditions on the whole
testing dataset. The result is shown in Fig. 10.
From Fig. 10, upward trend of precision and downward
trend of recall can be clearly seen. The reason is that increasing
negative samples in training would evidently increase the
probability of negative prediction of the model, as is discussed
above. This can be observed more clearly from Fig. 11. With
higher R in training, the model overestimate the number of the
non-crack pixels, which leads to higher precision and lower
Fig. 11. Results of the network with different ratios of positive to negative
samples in training. Natural means all samples are extracted randomly.
recall. F1 score varies accordingly. Besides, precision and
recall are computed according to manual labels, which are
different from database to database. Since cracks in ground
truth of CFD are wider than those of AigleRN generally, the
best ratios for different database are different. From Fig. 10,
the acceptable range of R is 2 6 R 6 5 for CFD and
2 6 R 6 3 for AigleRN.
VI. G ENERALIZATION S TUDY
From results shown in Section III, we can see that the model
works efficiently on CFD database and AigleRN database,
8
Fig. 12. Examples of cross database testing. From left to right on the top (training/testing): original image, CFD/CFD, AigleRN/CFD, Hybrid/CFD. From
left to right on the bottom: original image, AigleRN/AigleRN, CFD/AigleRN, Hybrid/AigleRN. All outputs are probability maps.
TABLE V
C ROSS DATABASE T ESTING
Test
AigleRN
CFD
AigleRN
P r = 0.9178
Re = 0.8812
F 1 = 0.8954
P r = 0.9651
Re = 0.3832
F 1 = 0.4812
CFD
P r = 0.6488
Re = 0.8819
F 1 = 0.7182
P r = 0.9119
Re = 0.9481
F 1 = 0.9244
Hybrid
P r = 0.9042
Re = 0.8448
F 1 = 0.8677
P r = 0.9018
Re = 0.9494
F 1 = 0.9210
Train
respectively. In this section, the generalization of the model is
discussed. In practical applications, different data acquisition
systems bring different types of data. Besides, conditions of
pavements differ from road to road. We conduct cross database
testing on the two databases to confirm the validation of the
model generalization.
We conduct three experiments: training on CFD and testing
on AigleRN, training on AigleRN and testing on CFD, and
training and testing on a hybrid database. In the first two experiments, the training and testing databases generated previously
are used. For the hybrid database, we take half of both training
databases to generate a new training database and test on the
other images. The examples of results are shown in Fig. 12
and the evaluations are shown in Table V.
It is observed that in cross testing, cracks detected by the
model trained on AigleRN are very thin, leading to low recall
and high precision, and the opposite is true for the model
trained on CFD. In addition, a hybrid data training can make
a trade-off. Moreover, results from hybrid data training suggest
that with more data from different pavement conditions, CNN
can learn better to predict cracks for different roads with the
same architecture.
VII. C ONCLUSION AND P ERSPECTIVES
In this paper, an effective method based on CNN is proposed
for pavement crack detection. Compared with other state-of-
art methods, the proposed method shows a better performance
in terms of dealing with different pavement texture. After
training, the model can predict cracks close to the manual
labels. This evidently shows that deep learning methods have
potential for crack structure prediction. The network can learn
from raw images without any preprocessing and output very
satisfactory results.
With training and testing on different databases, we observe
a good generalization of the model. With hybrid data training,
the potential of a general model for different conditions of
pavements may be brought out. Since CNN has a good
learning ability for images, the network architecture can also
be considered for other types of data, such as laser and 3D
camera images, which are commonly obtained from systems
of pavement distress detection.
Moreover, from cross testing we can see the influence
of the manual labels to the detection results. As shown in
experiments, thinner crack labels lead to thinner crack outputs.
This is basically because labelling still limited by judgements
of human experts. Our main perspective is to develop Semisupervised or unsupervised learning methods to break through
the limitations of human judgments.
R EFERENCES
[1] H. Zakeri, F. M. Nejad, and A. Fahimifar, “Image based techniques for
crack detection, classification and quantification in asphalt pavement: a
review,” Archives of Computational Methods in Engineering, vol. 24,
no. 4, pp. 935–977, 2017.
[2] Q. Zou, Y. Cao, Q. Li, Q. Mao, and S. Wang, “Cracktree: Automatic
crack detection from pavement images,” Pattern Recognition Letters,
vol. 33, no. 3, pp. 227–238, 2012.
[3] D. Zhang, Q. Li, Y. Chen, M. Cao, L. He, and B. Zhang, “An
efficient and reliable coarse-to-fine approach for asphalt pavement crack
detection,” Image and Vision Computing, vol. 57, pp. 130–146, 2017.
[4] H. Oliveira and P. L. Correia, “Automatic road crack segmentation using
entropy and image dynamic thresholding,” in Proc. Eur. Signal Process.
Conf., 2009, pp. 622–626.
[5] N. Tanaka and K. Uematsu, “A crack detection method in road surface
images using morphology.” in Proc. Workshop Mach. Vis. Appl., vol. 98,
1998, pp. 17–19.
[6] H. Zhao, G. Qin, and X. Wang, “Improvement of canny algorithm based
on pavement edge detection,” in Proc. Int. Conf. Image, Signal process.,
vol. 2. IEEE, 2010, pp. 964–967.
[7] S. Chambon and J.-M. Moliard, “Automatic road pavement assessment
with image processing: Review and comparison,” International Journal
of Geophysics, vol. 2011, 2011.
9
[8] J. Tang and Y. Gu, “Automatic crack detection and segmentation using
a hybrid algorithm for road distress analysis,” in Proc. Int. Conf. Syst.,
Man, Cybern. IEEE, 2013, pp. 3026–3030.
[9] Q. Li and X. Liu, “Novel approach to pavement image segmentation
based on neighboring difference histogram method,” in Proc. Int. Conf.
Image, Signal process., vol. 2. IEEE, 2008, pp. 792–796.
[10] Y. Hu and C.-x. Zhao, “A local binary pattern based methods for
pavement crack detection,” Journal of Pattern Recognition Research,
vol. 1, no. 20103, pp. 140–147, 2010.
[11] P. Subirats, J. Dumoulin, V. Legeay, and D. Barba, “Automation of pavement surface crack detection using the continuous wavelet transform,”
in Proc. Int. Conf. Image Process. IEEE, 2006, pp. 3037–3040.
[12] V. Kaul, A. Yezzi, and Y. Tsai, “Detecting curves with unknown endpoints and arbitrary topology using minimal paths,” IEEE Transactions
on Pattern Analysis and Machine Intelligence, vol. 34, no. 10, pp. 1952–
1965, 2012.
[13] T. S. Nguyen, S. Begot, F. Duculty, and M. Avila, “Free-form anisotropy:
A new method for crack detection on pavement surface images,” in Proc.
Int. Conf. Image Process. IEEE, 2011, pp. 1069–1072.
[14] R. Amhaz, S. Chambon, J. Idier, and V. Baltazart, “Automatic crack
detection on two-dimensional pavement images: An algorithm based on
minimal path selection,” IEEE Transactions on Intelligent Transportation Systems, vol. 17, no. 10, pp. 2718–2729, 2016.
[15] P. Delagnes and D. Barba, “A markov random field for rectilinear
structure extraction in pavement distress image analysis,” in Proc. Int.
Conf. Image Process., vol. 1. IEEE, 1995, pp. 446–449.
[16] H. Oliveira and P. L. Correia, “Automatic road crack detection and characterization,” IEEE Transactions on Intelligent Transportation Systems,
vol. 14, no. 1, pp. 155–168, 2013.
[17] A. Cord and S. Chambon, “Automatic road defect detection by textural
pattern recognition based on adaboost,” Computer-Aided Civil and
Infrastructure Engineering, vol. 27, no. 4, pp. 244–259, 2012.
[18] Y. Shi, L. Cui, Z. Qi, F. Meng, and Z. Chen, “Automatic road crack
detection using random structured forests,” IEEE Transactions on Intelligent Transportation Systems, vol. 17, no. 12, pp. 3434–3445, 2016.
[19] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Advances in Neural Information Processing Systems, 2012, pp. 1097–1105.
[20] Y.-J. Cha, W. Choi, and O. Büyüköztürk, “Deep learning-based crack
damage detection using convolutional neural networks,” Computer-Aided
Civil and Infrastructure Engineering, vol. 32, no. 5, pp. 361–378, 2017.
[21] A. Zhang, K. C. Wang, B. Li, E. Yang, X. Dai, Y. Peng, Y. Fei, Y. Liu,
J. Q. Li, and C. Chen, “Automated pixel-level pavement crack detection
on 3D asphalt surfaces using a deep-learning network,” Computer-Aided
Civil and Infrastructure Engineering, vol. 32, no. 10, pp. 805–819, 2017.
[22] K. Gopalakrishnan, S. K. Khaitan, A. Choudhary, and A. Agrawal,
“Deep convolutional neural networks with transfer learning for computer
vision-based data-driven pavement distress detection,” Construction and
Building Materials, vol. 157, pp. 322–330, 2017.
[23] L. Zhang, F. Yang, Y. D. Zhang, and Y. J. Zhu, “Road crack detection
using deep convolutional neural network,” in Proc. Int. Conf. Image
Process. IEEE, 2016, pp. 3708–3712.
[24] P. Liskowski and K. Krawiec, “Segmenting retinal blood vessels with
deep neural networks.” IEEE transactions on medical imaging, vol. 35,
no. 11, pp. 2369–2380, 2016.
[25] J. Canny, “A computational approach to edge detection,” IEEE Transactions on Pattern Analysis and Machine Intelligence, no. 6, pp. 679–698,
1986.
[26] W. Niblack, An introduction to digital image processing. Strandberg
Publishing Company, 1985.
[27] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
[28] V. Nair and G. E. Hinton, “Rectified linear units improve restricted
boltzmann machines,” in Proc. Int. Conf. Mach. Learn., 2010, pp. 807–
814.
[29] N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and
R. Salakhutdinov, “Dropout: a simple way to prevent neural networks
from overfitting.” Journal of Machine Learning Research, vol. 15, no. 1,
pp. 1929–1958, 2014.
[30] X. Glorot and Y. Bengio, “Understanding the difficulty of training deep
feedforward neural networks,” in Proc. Int. Confe. Art. Intell., Stat.,
2010, pp. 249–256.
[31] D. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
arXiv preprint arXiv:1412.6980, 2014.
| 1cs.CV
|
arXiv:1802.09904v3 [cs.AI] 25 Mar 2018
Ab initio Algorithmic Causal Deconvolution of
Intertwined Programs and Networks
by Generative Mechanism∗
Hector Zenil1,2,3,4 , Narsis A. Kiani1,2,3,4 , Jesper Tegnér2,3,5
1
Algorithmic Dynamics Lab, Centre for Molecular Medicine,
Karolinska Institute, Stockholm, Sweden
2
Unit of Computational Medicine, Department of Medicine,
Karolinska Institute, Stockholm, Sweden
3
Science for Life Laboratory, SciLifeLab, Stockholm, Sweden
4
Algorithmic Nature Group, LABORES for the Natural and
Digital Sciences, Paris, France
5
Biological and Environmental Sciences and Engineering Division,
Computer, Electrical and Mathematical Sciences and Engineering
Division, King Abdullah University of Science and
Technology (KAUST), Kingdom of Saudi Arabia
{hector.zenil, narsis.kiani, jesper.tegner}@ki.se
Abstract
To extract and learn representations leading to generative mechanisms from data, especially without making arbitrary decisions and
biased assumptions, is a central challenge in most areas of scientific
research particularly in connection to current major limitations of influential topics and methods of machine and deep learning as they
have often lost sight of the model component. Complex data is usually produced by interacting sources with different mechanisms. Here
we introduce a parameter-free model-based approach, based upon the
seminal concept of Algorithmic Probability, that decomposes an observation and signal into its most likely algorithmic generative mechanisms. Our methods use a causal calculus to infer model representations. We demonstrate the method ability to distinguish interacting
mechanisms and deconvolve them, regardless of whether the objects
produce strings, space-time evolution diagrams, images or networks.
We numerically test and evaluate our method and find that it can
disentangle observations from discrete dynamic systems, random and
complex networks. We think that these causal inference techniques can
contribute as key pieces of information for estimations of probability
distributions complementing other more statistical-oriented techniques
that otherwise lack model inference capabilities.
∗
An online implementation to estimations of graph complexity is available online at
http://www.complexitycalculator.com
1
Keywords: Model generation; signal decomposition; segmentation;
algorithmic renormalization; algorithmic image segmentation; graph
partitioning; algorithmic machine learning; model generation; feature
selection.
1
Introduction
Classical information theory has provided rigorous ways to capture our intuitive notions regarding uncertainty and information, and made an enormous
impact in doing so. One of the fundamental measures here is mutual information, which captures the average information contained in one variable
about another, and vice versa. If we have two source variables and a target, for example, we can measure the information held by one source about
the target, the information held by the other source about the target, and
the information held by those sources together about the target. Any other
notion about the directed information relationship between these variables,
which can be captured by classical information-theoretic measures (e.g., conditional mutual information terms) is linearly redundant with those three
quantities.
However, intuitively, there is strong desire to measure further notions
of how this directed information interaction may be decomposed, e.g., how
much information the two source variables hold redundantly about the target, how much each source variable holds uniquely, and how much information can only be discerned by synergistically examining the two sources
together. These notions go beyond the traditional information-theoretic
view of a channel serving the purpose of reliable communication, considering now the situation of multiple communication streams converging on a
single target. This is a common situation in biology, and in particular in
neuroscience, where, say, the ability of a target to synergistically fuse multiple information sources in a non-trivial fashion is likely to have its own
intrinsic value, independently of reliability of communication.
The absence of measures for such decompositions into redundant, unique
and synergistic information is arguably the most fundamental missing piece
in classical information theory. Triggered by the formulation of the Partial
Information Decomposition framework by Williams and Beer in 2010, the
past few years have witnessed a concentration of work by the community
in proposing, contrasting, and investigating new measures to capture these
notions of information decomposition.
Typically, models encode features of data in statistical form and in single
variables and merged models, even in cases where several data sources are
2
involved. Quantitative measures to disentangle complex signals and methods to tell apart causal mechanisms from noise and (ir)relevant interactions
are of broad interest in areas such as statistical mechanics, reverse engineering, network inference, data reconstruction and scientific discovery in
general, being germane to the challenge of causal analysis. For example, the
development of techniques for learning disentangled representations using
probabilistic machine and deep learning methods has recently gained significant momentum [11]. More broadly, model-based representations from
data is one of the main challenges in machine learning, artificial intelligence
and causal discovery and inference. Here we introduce a framework based
upon the theory of algorithmic probability, which in our formulation is capable of identifying different sources that may explain and provide different
models for each of the possible causes of convoluted or intertwined data.
Casual inference has been one of the most challenging problems in science.
The debate about causality has not prevented the development of successful
and mature mathematical and algorithmic frameworks, first in the form of
classical statistics and today in the form of computability and algorithmic information theories. Based on these latest mature mathematical frameworks
that are acknowledged to fully characterize the concept of randomness as
opposed to causal (deterministic), we introduced a suite of algorithms [20] to
study the algorithmic information dynamics of evolving systems, and also
methods to reduce the dimensions of data [21] based on the same principles. Algorithmic data dimension reduction and algorithmic deconvolution
are two challenges which can be viewed as opposite sides of the same coin.
On the one hand, data reduction is achieved by finding elements that are
considered redundant, using as a criterion their contribution to the algorithmic content of the description of the data. Such elements provide the basis
for compression and they can therefore safely be removed. On the other
hand, the algorithmic deconvolution by generative mechanisms operates by
identifying the elements that are likely to be part of the same causal path of
some production rule or generating mechanism. The deconvolution is thus
able to decompose causes and pinpoint the different sources generating the
observed data.
2
2.1
Notation and Background
Cellular automata
We use cellular automata as causal dynamical systems to illustrate the algorithm before demonstrating its capabilities on more sophisticated objects
3
of a convoluted nature and its applications to objects such as complex networks.
A cellular automaton is a computer program that applies in parallel a
global rule composed of local rules on a tape of cells with symbols (e.g.
binary). Thoroughly studied in [22], Elementary Cellular Automata (or
ECA) are one-dimensional cellular automata that take into consideration in
their local rules the cell next to the centre and the centre cell.
Definition 2.1. A cellular automaton (or CA) is a tuple hS, (L, +), T, f i
with a set S of states, a lattice L with a binary operation +, a neighbourhood
template T , and a local rule f .
The set
P of states S is a finite set with elements s taken from a finite
alphabet
with at least two elements.
Definition 2.2. The neighbourhood template T = hη1 , ..., ηm i is a sequence
of L. In particular, the neighbourhood of cell i is given by adding the cell i
to each element of the template T : T = hi + η1 , ..., i + ηm i. Each cell i of the
CA is in a particular state c[i] ∈ S. A configuration of the CA is a function
c : L → S. The set of all possible configurations of the CA is defined as SL .
As a discrete dynamic system, the evolution of the CA occurs in discrete
time steps t = 0, 1, 2, ..., n. The transition from a configuration ct at time
t to the configuration c(t+1) at time t + 1 is induced by applying the local
rule f . The local rule is to be taken as a function f : S |T | → S which maps
the states of the neighbourhood cells of time step t in the neighbourhood
template T to cell states of the configuration at time step t + 1:
ct+1 [i] = f (ct [i + η1 ], ..., ct [i + ηm ])
(1)
The general transition from configuration to configuration is called the global
map and is defined as: F : S L → S L .
In the case of 1-dimensional CA it is common to introduce the radius of
the neighbourhood template which can be written as h−r, −r + 1, ..., r − 1, ri
and has length 2r + 1 cells. With a given radius r the local rule is a function
|S|(2r+1)
|S|(2r+1)
f : Z|S|
→ Z|S| with Z|S|
rules. Elementary Cellular Automata
(ECA) have a radius r = 1 (closest neighbours), having the neighbourhood
template h−1, 0, 1i, meaning that the neighbourhood comprises a central
3
cell. From this it follows that the rule space for ECA contains 22 = 256
rules.
Enumeration of ECA rules: It is common to follow the lexicographic
ordering scheme introduced by Wolfram [22]. According to this encoding,
the 256 ECA rules can be encoded by 8-bits.
4
2.2
Causation and Algorithmic Probability
The concept of algorithmic complexity [5, 2] is at the core of the challenge of
complexity in discrete dynamic systems, as it involves finding the most statistically likely generating mechanism (computer program) that produces
a set of given data. Formally, the algorithmic complexity (also known as
Kolmogorov-Chaitin complexity) is the length of the shortest computer program that reproduces the data from its compressed form when running on
a universal Turing machine.
We apply the so-called Coding Theorem Method (CTM) and Block Decomposition Method (BDM) as introduced in [3, 9, 13, 19] based on the
seminal concept of Algorithmic Probability [10, 6] to estimate algorithmic
complexity [5, 2].
3
Methods and deconvolution algorithm
The aim of the deconvolution algorithm is to break a dataset into groups
that do not share certain features (essentially causal clustering and algorithmic partition by probable generative mechanism, completely different from
traditional clustering and partition in machine learning approaches). Usually these characteristics are a parameter to maximize, but ultimately the
purpose is to distinguish components that are generated similarly from those
that are generated differently. In information-theoretic terms the question
is therefore as follows: What are the elements (e.g. nodes or edges) that
can break a network into the components that maximize their algorithmic
information content, that is, those elements that preserve the information
about the underlying programs generating the data?
Formally, let E = E(G) be the set of all edges of a graph G, we seek the
edges {e1 , . . . , en }, ∈ Ej ⊂ G that break G into N disconnected components
such that the removal Ej minimizes the loss of information in G among all
subsets Ei ⊂ E, i.e. C(G\Ej∗ ) ≤ C(G\Ei∗ ) for all i.
The algorithm’s pseudo-code can thus be written as follows:
1. Find the subset of edges with index j such that C(G\Ej∗ ) ≤ C(G\Ei∗ )
for all i.
2. Remove the elements in Ej from G.
3. Repeat 1 with G := G\{Ej } until N number of subcomponents is
reached (or closest value to N ).
4. The number of components is the sum of the remaining subcomponents
that were deleted plus those that prevailed.
5
The only parameter that the algorithm requires is the number of components into which an object will be decomposed. However, there is a natural way to find the optimal terminating step and therefore the number
of maximum possible components that minimize the sum of the lengths of
the candidate generating mechanisms thereby truly making the algorithm
parameter-free as it is not required to have a preset number of desired components.
However, before introducing the terminating criterion (c.f. next section) for the number of components, let’s analyze what it might mean for
two components s1 and s2 to have same algorithmic information content
C(s1 ) = C(s2 ). Clearly that subcomponents s1 and s2 have the same algorithmic complexity (an integer—or a real value if using AP-based BDM—
indicating the size of the approximated minimal program) does not imply
that the two components are generated by exactly the same generating mechanism. However, because of the exponential decay of the algorithmic probability of an increasingly random object, we have it that the less random it
is, the exponentially more likely it is that the underlying mechanism will be
the same (see Fig. 4C). This is because there are exponentially fewer short
programs than long ones. For example, in the extreme case of connected
graphs, we have it that the complete graph denoted by Kn has the smallest possible algorithmic complexity ∼ log(n). If C(s1 ) = C(s2 ) ∼ log(n)
then s1 and s2 are, with extremely high probability, generated by the same
algorithm that generates either the complete graph or the empty graph
(with same lowest algorithmic complexity as it requires no description other
than either all nodes connected or all nodes disconnected). Conversely, if
C(s1 ) = C(s2 ) but C(s2 ) and C(s1 ) depart from log(n) (and approximate
algorithmic randomness) then the likelihood of being generated by the same
algorithm exponentially vanishes. So the information regarding both the
algorithmic complexity of the components and their relative size sheds light
on the candidate generating mechanisms and is less likely to coincide ‘by
chance’ for non-trivial cases.
3.0.1
Algorithm terminating criterion
The immediate question is where we should stop breaking down a system
into its causal components. The previous section suggests a terminating criterion. Let S be the object which has been produced by N mostly independent generative mechanisms. We decompose S into n parts s1 , . . . , sn in such
a way that each si , i ∈ {1 . . . n} has an underlying generating mechanism
found by running the algorithm iteratively for increasing n, but after each
6
iteration we calculate the minimum of the differences in algorithmic complexity among all subcomponents. The algorithm should then stop where
the number of subcomponents is exactly N when the sum of the lengths—
the estimated algorithmic complexity—of each of the programs will diverge
from the expected log(N ) because the length of the individual causal mechanisms producing each new component will be breaking a component that
could previously be explained by the causal mechanism at a previous iteration of the algorithm.
As a trivial example, let’s take the string 1n , where S n means that
the pattern S is repeated n times. After application of the algorithm, the
terminating criterion will suggest that 1n cannot be broken down into smaller
segments, each with a different causal generating mechanism, whose total
length sums will be shorter than the length of the generating mechanism
producingP
1n itself. This is because the sum of the length of the shortest
programs i |pi | running on a universal Turing machine generating segments
of 1n of length mi < n each, such that the concatenation ∪i=1 pi = 1n , will be
strictly greater than C(1n ), given that each pi halting criterion will require
i log mi bits more than C(1n ).
In the case of Fig. 2, the terminating criterion retrieves N = 3 components from the two interacting ECA (rule 60 and 110). This does not contradict the fact that we started from two generating mechanisms, because
there are three clear regimes that are actually likely to be reproducible by
three different generating mechanisms, as suggested by the deconvolution
algorithm itself, and as found in [7], where it has been shown that rule 110
can be emulated by the composition of two simpler ECA rules (rules 51
and 118). As seen in Fig. 2, among the possible causal partitions, N = 2
successfully deconvolves ECA rule 60 from rule 110 on the first run, with
a stronger difference than the difference found between N = 3 components
when breaking rule 110 into its two different regimes.
3.1
Graph complexity
The concept of algorithmic probability (also known as Levin’s semi-measure)
yields a method for approximating algorithmic complexity related to the frequency of patterns in the adjacency matrix of a network, including therefore
the number of subgraphs in a network. The algorithmic probability [10, 6, 2]
of a subgraph H ∈ G is a measure that describes the probability that a random computer program p will produce H when run on a 2-dimensional tape
7
1 ) Turing machine U also referred to as a Turmite. That
universal (prefix-free
P
is, m(G) = p:U (p)=H∈G 1/2|p| .
The probability semi-measure m(G) is related to algorithmic complexity C(G) in that m(G) is at least the maximum term in the summation
of programs m(G) ≥ 2−C(G) , given that the shortest program carries the
greatest weight in the sum. The Coding Theorem establishes the connection between m(G) and C(G) as ([6]): | − log2 m(G) − C(G)| < c (Eq. 1),
where c is some fixed constant, independent of s. The theorem implies that
one can estimate the algorithmic complexity of a graph from the frequency
of production from running random programs and applying the Coding theorem: C(G0 ) = − log2 m(G) + O(1). The Coding theorem establishes that
graphs produced with lower frequency by random computer programs have
higher algorithmic complexity, and vice versa.
The BDM of a graph thus consists in decomposing the adjacency matrix
of a graph into subgraphs of sizes for which complexity values have been estimated, then reconstructing an approximation of the algorithmic complexity
of the graph by adding the complexity of the individual pieces according to
the rules of information theory, as follows:
C(G) =
X
log2 (nu ) + C(ru )
(2)
(ru ,nu )∈Adj(G)d×d
where Adj(G)d×d represents the set with elements (ru , nu ), obtained
when decomposing the adjacency matrix of G into all subgraphs contained
in G of size d. In each (ru , nu ) pair, ru is one such submatrix of the adjacency matrix and nu its multiplicity (number of occurrences). As can be
seen from the formula, repeated subgraphs only contribute to the complexity value with the subgraph BDM complexity value once plus a logarithmic
term as a function of the number of occurrences. This is because the information content of subgraphs is only sub-additive, as one would expect from
the growth of their description lengths. Applications of m(G) and C(G)
have been explored in [3, 9, 8, 13], and include applications to graph theory
and complex networks [12] and [13] where the technique was first introduced.
The only parameters used for the application of BDM, as suggested
in [19], is to set the overlapping of the decomposition to the maximum 12
bits for strings and 4 square bits for arrays given the current best CTM approximations [9] from an empirical distribution based on all Turing machines
with up to 5 states, and no string/array overlapping in the decomposition
1
The group of valid programs forms a prefix-free set (no element is a prefix of any
other, a property necessary to keep 0 < m(G) < 1).
8
for maximum efficiency (as it runs in linear time) and for which the error
(due to boundary conditions) has been shown to be bounded [19].
However, the algorithm introduced here is independent of the method
used to approximate algorithmic complexity such as BDM. BDM assigns
an index associated with the size of the most likely generating mechanism
producing the data according to Algorithmic Probability [10]. BDM is capable of capturing features in data beyond statistical properties [19, 16] and
thus represents an improvement over classical information theory. Because
finding the program that reproduces a large object is computationally very
expensive– even to approximate–BDM finds short candidate programs using
another method [3, 9] that finds and reproduces fragments of the original
object and then puts them together as a candidate algorithmic model of the
whole object [19, 13]. These short computer programs are effectively model
candidates explaining each fragment, with the long finite sequence of short
models being itself a generating mechanistic model.
4
4.1
Numerical experiments
Decomposition of sequences and space-time diagrams
In this section, we will test the suggested algorithm on different types of
objects, in order to show its applicability and power. We start with the
simplest version of an object which conveys information, a string, and move
later to consider richer objects such as networks.
We will use different programs to produce different parts of a string, that
is a program p to generate segment s1 and program p0 to generate segment
s2 put next to each other. Clearly the string has been generated by two
generating mechanisms (p and p0 ). Now we use the algorithm to deconvolve
the string and find the number of generating mechanisms and most likely
model mechanisms (the program themselves) inducing a form of algorithmic
partition based on the likelihood of each segment to be produced by different
generating mechanisms.
Figs. 1A-E illustrate how strings that have short generating mechanisms
are significantly and consistently more sensitive to perturbations. The resulting string is 01010101010101010101010101010101010101010101010101
01110100101010100000001001100111100110000011100110 with the colours
corresponding to the parts suggested by the different regimes, according
to their algorithmic contribution and the segment’s resilience to perturbations (by deletion and replacement) to the original string. Behind every real
number approximating the algorithmic complexity of a string there is the
9
A
B
C
D
E
F
G
Figure 1: Proof of concept applied to a binary string composed of two
segments with different underlying generating mechanisms (computer programs). A: Log plot of complexity estimation of a regular segment (blue)
consisting of the repetition of ‘01’ 25 times followed by a random-looking
segment (red). B: Log plot reversing the order of A yet preserving the qualitative behaviour of the different segments. C: The code of the smallest
generating program (a non-terminating Turing machine) depicted visually
(states are arrows in different directions) producing the string of 01n for
any n (0 is white and 1 is orange) starting from a blank tape as shown in
the space-time diagram (E). D: the same computer program as a state diagram. F: Interacting programs with different generating mechanisms (ECA
rules 255 v 110) running for 60 steps. G: Algorithmic information footprint,
every pixel is deleted and its original contribution to the whole quantified
and coloured accordingly. If grey, then it makes the lowest contribution,
blue represents a low contribution and red the highest contribution (randomness).
10
discovery of a large set of generating programs when using the Algorithmic
Probability (AP)-based measure BDM producing the object.
We not only could find the number of mechanisms correctly (Figs. 1A and
B) but the candidate programs (which for this trivial example are exactly
the original) that generate each segment (Figs. 1C-E) by way of seeking for
the shortest computer programs in a bottom up approach (see [3, 9, 13,
19]). Finding the shortest programs is, however, secondary, because we only
care about the different explanatory power that different programs have to
explain the data in full or in part pinpointing the different causal nature of
the segments and helping in the deconvolution of the original observation.
Figs. 1C-E depict the computer program (a non-terminating Turing machine) that is found when calculating the BDM of the 01n string. The
BDM approximation to the algorithmic complexity of any 01n string is thus
the number of small computer programs that are found capable of generating the same string or, conversely (via the algorithmic Coding theorem,
see [3, 9, 13]), the length of the shortest program producing the string. For
example, the string 01n was trivially found to be generated by a large number
of small computer programs (in Fig. 1C,D depicted a non-terminating Turing machine with E its output) using our algorithmic methods (as opposed
to, e.g., using lossless compression, which would only obfuscate the possible
generating model) with only two rules out of 2×2 rules for the size of Turing
machine with only two states and two symbols and no more, thus of very
low algorithmic complexity compared to, e.g., generating a random-looking
string that would require a more complex (longer) computer program. The
computer program of a truly random string will grow in proportion to the
length of the random string, but for a low complexity string such as 01n ,
repeated any number of times n, the length of the computer program is
of (almost) fixed size, growing only by log(n) if the computer program is
required to stop after n iterations. In this case 01n is a trivial example with
a strong statistical regularity whose low complexity could be captured by
applying Shannon entropy alone on blocks of size 2.
Figs. 1 F and G illustrate how the algorithm can separate regions produced by generating mechanisms of different algorithmic information content
by observing their space-time dynamics, thereby contributing to the deconvolutin of regions that are produced by different generating mechanisms. In
this example both programs are sufficiently robust to not break down (see
Sup. Inf.) when they interact with each other, with rule 110 prevailing over
255. Yet, in the general case it is not always easy to tell these mechanisms
apart. In more sophisticated examples such as in Figs. 2 D to E, we see how
the algorithm can break down contiguous regions separating an object into
11
two major components corresponding to the different generating computer
programs that are intertwined and actively interacting with each other. The
experiment was repeated 20 times with programs with differing qualitative
(e.g. Wolfram class) behaviour.
Fig. 2 F demonstrates that perturbations to regions in red are considered
to have a more random effect and are thus by themselves less algorithmically
complex (random) versus simple. When regions are of the same algorithmic
complexity they are likely to be generated by similar algorithms, or more
precisely, algorithms that are of similar minimal length. Regions in blue
move the space-time evolution away from randomness and are themselves
more algorithmically random. Blue structures on the left hand side correspond to large triangles occurring in ECA rule 110 that are usually used to
compute and transfer information in the form of particles. However, triangular patterns transfer information in a limited way because their light cone
of influence reduces at the greatest possible speed of the automaton, and
they are assigned an absolute neutral information value. Absolute neutral
values are those closest to 0. Once separated, the two regions have clearly
different algorithmic characteristics given by their causal perturbation sensitivity, with the right hand side being more sensitive to both random and
non-random perturbations. Moreover, Fig. 2 F shows results compatible
with the theoretical expectation and findings in [20] where a measure of
reprogrammability associated with the number and magnitude of elements
that can move a dynamical system towards or away from randomness was introduced and shown to be related to fundamental properties of the attractor
space of the system.
4.2
Graph and network deconvolution
Classification can usually be viewed as solving a problem which has an underlying tree structure according to some measure of interest. One way to
think of optimal classification is to discover a tree structure at some level of
depth, with tree leaves closer to each other when such objects have a common or similar causal mechanism and for which no feature of interest has
been selected. Fig. 3 illustrates how the algorithm may data, in this case
starting from a trivial example that breaks complete K-ary trees. Traditionally, partitioning is induced by an arbitrary distance measure of interest
that determines the connections in a tree, with elements closer to a cluster
centre connected by edges. The algorithm breaks the trees (see Fig. 3) into
as many components as desired by iterating the algorithm until the number
of desired components is obtained or the terminating criterion is applied
12
A
B
C
D
E
F
Figure 2: A: The output of two different intertwined programs (ECA rules
60 and 110) with qualitatively complex output behaviour (11 to 60 steps
depicted here, from a random initial condition) interacting with each other
(one of which has been proven to be Turing-universal and the other has
also been conjectured to be universal [22]), each producing structures of a
similar type that, from an observer’s perspective, are difficult to distinguish
(see Subfigure C) as is artificially done in B (knowing which pixel is generated by which rule). C: What an observer of the last runtime would see
in the form of a stream of bits with no clear statistical distinction. D: The
algorithm pinpoints the regions of neutral, positive and negative, with the
contiguous largest blue component segmenting the image into two components. E: only negative vs positive causal contributions where both Shannon
entropy and a popular lossless compression algorithms fail (see SI). F: Sanity
check/validation: Statistically significant quantitative differences among the
parts after application of the algorithm as illustrated in E among apparently
weak qualitative differences as illustrated in Subfig. A.
13
(c.f. Subsection 3.0.1). Figs. 3A,B provide examples illustrating how to
maximize topological symmetry. The algorithm can be applied, without
loss of generalization, to any non-trivial graph, as in Figs. 3C,D or on any
dataset for that matter.
The same task using classical information theory (Shannon entropy) is
shown not to be sensitive enough (see Sup. Inf.), and a popular lossless
compression algorithm (Compress based on LZW) provided a noisy approximation (see Sup. Inf.) to the results obtained by using the Block Decomposition Method, as defined in [19], whose description is provided in the Sup.
Inf.
Figs. 3C-E illustrate how randomly connected graphs with different topologies can be broken into their respective generative mechanisms. Fig. 3C is
a complete graph of size 20 randomly connected by 3 edges to a scale-free
graph of size 100. The graphs are generated by different mechanisms, one
is a small program that, given a number N of nodes, produces a graph with
all nodes connected to all other N − 1 nodes and has a program of small
length that grows only by log N [15]. The scale-free network is generated
by the canonical preferential attachment algorithm with two edges per node
and requires a slightly longer algorithm that grows by log N + c [15] where c
is a small constant accounting for the pseudo-random choice of attachment
nodes. The algorithm breaks the graphs into two components, each of which
corresponds to the graphs with different degree distribution (depicted below each case) associated with its generating mechanism. This is because
|P (G1 )| + |P (G2 )| + . . . + |P (Gn )| + |P (eGi )| > |P (G1 G2 . . . Gn )| for any Gi ,
where eGi is the set of edges randomly connecting Gi to Gj for any i and j
for all G of low algorithmic complexity.
4.3
Robustness and limitations
Fig. 3 D illustrates a similar case to Fig. 3 C, but instead of a complete
graph an Erdös Rényi (E-R) graph with edge density 0.5 is produced and
connected by 3 random edges to a scale-free network produced in the same
fashion as in Fig. 3 C. Again, the algorithm was able to break it down into
the two corresponding subgraphs. Fig. 3 D represents a test case to evaluate
the effect of additive noise by connecting an E-R graph of increasing size
and with an increasingly greater number of random edges.
Next we ask how much structure, if any, can be recovered/extracted when
adding a random (E-R) graph to different types of structured networks. To
this end, we conducted a series of numerical experiments shedding light
on the limitations of the algorithm introduced here in the face of additive
14
A
B
C
D
E
Figure 3: A,B: Forced deconvolution of a tree by minimization of graph
algorithmic information loss thereby maximizing causal resemblance of the
resultant components (hence causal clustering). Depicted are the components of a K-ary trees of size 6 (A) and 10 (B) and their resulting graphs
after one iteration of the deconvolution algorithm. C: Deconvoluting a graph
composed by a complete graph and a scale-free (S-F) network generated by
preferential attachment and randomly connected between them. Negative
edges break down the graph into components corresponding to the different
underlying generating mechanisms. D: Deconvolution of a random graph
(E-R) and a scale-free (S-F) network. E: The algorithm first separates the
subcomponents with the largest algorithmic difference, followed by other
subcomponents.
15
A
B
C
Figure 4: A: Causal deconvolution of graphs, a S-F graph (2 new links per
added node) from an E-R (0.5 edge density) graph emulating noise (each
point represents the average of 10 replicates). When the number of links
increases as a function of the subgraph sizes, the separability is robust (red
square markers) compared to increasing the number of random links only for
graphs of fixed size (40 nodes), when successful separation is compromised,
noise (E-R) and structure (S-F) being indistinguishable. B: A small sample
of 6 graphs among the 10 × 20 = 200 graphs randomly connecting a S-F to
an E-R graph. Success at identifying randomly connecting links is shown
and was found to be very robust. C: Objects have exponentially greater
chances of being produced by the same generating mechanism if they are of
low algorithmic randomness and thus of high algorithmic probability.
16
noise. The same results were obtained for the simpler case of connecting any
complete graph of increasing size to any other graph such as E-R or S-F.
Fig. 4 shows the results from the experiments separating graphs, in this
case a scale-free graph (S-F) from an Erdös Rényi graph (E-R), the former generated by a Barábasi-Albert preferential attachment algorithm [1]
and the latter produced by a pseudo-random generator. Fig. 4A quantifies the error and optimal signal-to-noise ratio for optimal deconvolution,
testing the algorithm under additive noise both for fixed and growing size
subcomponents. Fig. 4B shows links coloured in red as identified by the
algorithm having the highest algorithmic information value when their removal sends the original composed system towards lower algorithmic information content, thereby telling apart the two subcomponents. Negative
links are mostly on the side of the E-R graph. In other words, the SF can be extracted with the greatest precision and a lower rate of false
positives from the mix, which is to be expected given the random nature of the added links that connect the graphs, making them more like
the E-R links than the S-F. If only the number of random links among
graphs increases for fixed size graphs (Fig. 4B blue circle marks) a maximum precision of about 0.9 is reached before degradation. That is, at around
32.5% of the links randomly connecting the components. In other words,
the algorithm is robust, telling apart noise from structure even after up to
0.325 (from the random links connecting the components) + 0.5 (from the
E-R component) = 0.825, i.e. 82.5% of all links are random. On the other
hand, the number of false positives is constant at about 5%, in the case
shown in Fig. 4B all of the false positives (red links not connecting the
two graphs with different topology) are inside the E-R graph and mostly
nonexistent on the side of the less random S-F graph.
5
Remarks and conclusions
For some more sophisticated yet successful examples on which these new algorithmic methods may outperform applications of entropy in general (without access to true probability distributions) and other computable measures
see [19], and [16] for a non-trivial example in which entropic measures fail
(by offering divergent descriptions of the same evolving system). The generation of these models is key in our approach because the integer (or real)
value assigned to a system as an estimation of its algorithmic complexity is
nothing but a guiding index of the number of specific models found by our
method that are capable of explaining and generating the data.
17
One possible objection is that any number of interacting rules can be
thought of as a single rule producing an intertwined output because we
known from Turing universality that any number of interacting programs can
also be rewritten as a single program in a larger rule space (defined by state
× symbol) incorporating all the behaviours together. That is, any computer
program can be decomposed into one or more computer programs producing
the same output. This means that separating programs and signals is, in
some way, not fundamental. Fig. 2C-F shows, for example, that by iterating
the deconvolution algorithm not only do the two main components of the
image correspond to the two generating ECA rules, but a second application
of the algorithm would produce a third or more components corresponding
to further resilient features generated by the rules, which can be considered
rules themselves within a smaller rule (state/symbol) space. However, in
the deconvolved observations the interacting rule determining how two or
more rules may interact effectively constitutes a third global rule to which
the algorithm has no direct access or an apparent region in the observed
window.
We have introduced and tested a parameter-free algorithm for causal
deconvolution of interlaced and interacting mechanisms using fundamental
concepts drawn from the theory of Algorithmic Probability and Algorithmic
Information Theory. Our approach enables a parameter-free analysis of the
deconvolution problem, since we have removed the need for pre-defined usercentric definitions of global properties or local rules (e.g. distance metrics)
to determine the algorithm. Instead, relating our algorithm to the algorithmic information theory provides us with a fundamental metric. While
the algorithm uses state-of-the-art algorithms to approximate algorithmic
complexity (CTM and BDM), the algorithm and methods introduced here
are independent of the approximating method chosen (e.g. lossless compression). However, the precision and accuracy is not. Yet, the algorithm
is sufficiently robust to disentangle sophisticated intertwined causal mechanisms. This opens the possibility of parlaying these algorithmic methods into
a more elaborate machine learning framework. For example, the extracted
causal information can be used as a prior distribution for machine and deep
learning techniques. This is useful because efficient techniques for pattern
recognition are as a rule weak on inferring models, and thus ill-equipped
to capture the underlying generative mechanisms and thereby produce predictive and prescriptive causal models that exceed the descriptive nature of
current classical and modern statistical approaches.
18
Acknowledgements
H.Z. was supported by the Swedish Research Council (Vetenskapsrådet)
grant No. 2015-05299.
References
[1] R. Albert, A.-L. Barabási, Statistical mechanics of complex networks,
Reviews of Modern Physics, 74 (1): 47–97, 2002.
[2] G.J. Chaitin. On the length of programs for computing finite binary
sequences Journal of the ACM, 13(4):547–569, 1966.
[3] J.-P. Delahaye and H. Zenil, Numerical Evaluation of the Complexity
of Short Strings: A Glance Into the Innermost Structure of Algorithmic Randomness, Applied Mathematics and Computation 219, 63–77,
2012.
[4] E. Hermo-Reyes and J.J. Joosten,
“Competing Cellular
Automata”
http://demonstrations.wolfram.com/
CompetingCellularAutomata/, Wolfram Demonstrations Project,
June 17, 2014.
[5] A.N. Kolmogorov. Three approaches to the quantitative definition
of information, Problems of Information and Transmission, 1(1):1–7,
1965.
[6] L.A. Levin. Laws of information conservation (non-growth) and aspects of the foundation of probability theory, Problems of Information
Transmission, 10(3):206–210, 1974.
[7] J. Riedel and H. Zenil, Rule Primality and Compositional Emergence
of Turing-universality from Elementary Cellular Automata, forthcoming, 2017. XXX
[8] F. Soler-Toscano, H. Zenil, J.-P. Delahaye and N. Gauvrit, Correspondence and Independence of Numerical Evaluations of Algorithmic
Information Measures, Computability, vol. 2, no. 2, pp. 125–140, 2013.
[9] F. Soler-Toscano, H. Zenil, J.-P. Delahaye and N. Gauvrit, Calculating
Kolmogorov Complexity from the Frequency Output Distributions of
Small Turing Machines, PLoS One 9(5), e96223, 2014.
19
[10] R.J. Solomonoff, A formal theory of inductive inference: Parts 1 and
2. Information and Control, 7:1–22 and 224–254, 1964.
[11] N. Siddharth, B. Paige, J-W van de Meent, A. Desmaison, N.D.
Goodman, P. Kohli, F. Wood, P.H.S. Torr, Learning Disentangled Representations with Semi-Supervised Deep Generative Models,
arXiv:1706.00400 [stat.ML], 2017.
[12] H. Zenil, F. Soler-Toscano, K. Dingle and A. Louis, Graph Automorphisms and Topological Characterization of Complex Networks by Algorithmic Information Content, Physica A: Statistical Mechanics and
its Applications, vol. 404, pp. 341–358, 2014.
[13] H. Zenil, F. Soler-Toscano, J.-P. Delahaye and N. Gauvrit, TwoDimensional Kolmogorov Complexity and Validation of the Coding
Theorem Method by Compressibility, 2013.
[14] H. Zenil, N.A. Kiani and J. Tegnér, Quantifying Loss of Information
in Network-based Dimensionality Reduction Techniques, Journal of
Complex Networks 4, 342–362, 2016.
[15] H. Zenil, N.A. Kiani and J. Tegnér, Methods of Information Theory
and Algorithmic Complexity for Network Biology, arXiv:1401.3604,
2014.
[16] H. Zenil, N.A. Kiani and J. Tegnér, Low-Algorithmic-Complexity
Entropy-deceiving Graphs, Physics Reviews E. 96, 012308, 2017.
[17] H. Zenil, Compression-based Investigation of the Dynamical Properties of Cellular Automata and Other Systems, Complex Systems, 19(1),
pages 1–28, 2010.
[18] H. Zenil and E. Villarreal-Zapata, Asymptotic Behaviour and Ratios of
Complexity in Cellular Automata Rule Spaces, International Journal
of Bifurcation and Chaos, vol. 13, no. 9, 2013.
[19] H. Zenil, S. Hernández-Orozco, N.A. Kiani, F. Soler-Toscano, A.
Rueda-Toicen, A Decomposition Method for Global Evaluation of
Shannon Entropy and Local Estimations of Algorithmic Complexity,
arXiv:1609.00110 [cs.IT], 2016.
[20] H. Zenil, N.A. Kiani, F. Marabita, Y. Deng, S. Elias, A. Schmidt,
G. Ball, J. Tegnér, An Algorithmic Information Calculus for
20
Causal Discovery and Reprogramming Systems, 2017. BioArXiv DOI:
https://doi.org/10.1101/185637
[21] H. Zenil, N.A. Kiani, J. Tegnér, Data Reduction and Network Sparsification by Minimal Algorithmic Information Loss, arXiv XXX
[22] S. Wolfram, A New Kind of Science, Wolfram Media, Champaign IL.,
2002.
21
Supplementary Information
5.1
Interacting programs
Figure 5: Interacting programs such as cellular automata require us to define
how the interaction happens by, e.g., deciding what set of rules apply at the
intersection. For instance, one of the two sets of local rules or a 3rd set
of rules effectively defining a super cellular automaton that most likely is
another cellular automaton in a larger rule-space (requiring more states to
define the two sub-cellular automata and the interaction). Depicted here is
the type of interaction we explored, and this case shows the richness of such
possible interactions where both rules ‘spill over’ each other.
The qualitative behaviour of each program can heuristically be identified
by what is known as its Wolfram class, which in turn has been formalized
using tools and methods from algorithmic complexity in [17, 18]. Informally,
Wolfram class 1 represents evolutions of programs that converge to a simple
fixed configuration, exemplars of Wolfram class two converge to repetitive
simple behaviour, those of Wolfram class 3 produce unbounded apparently
statistically random behaviour and exemplars of Wolfram class 4 reproduce
apparently open-ended persistent structures. None of what has been introduced here depends on this behavioural characterization based on different
heuristics, and it is thus in no way fundamental to the results reported.
The interaction rule is interaction rule number 1 according to the enumeration open-source program on the Wolfram Demonstrations website publicly available at http://demonstrations.wolfram.com/CompetingCellularAutomata/ [4]
determining the way in which the local rules from each global ECA rule will
22
be dictated and applied at their intersection.
5.2
Graph generation
The graphs used throughout this paper were generated using the Wolfram Language on the Mathematica platform using the function RandomGraph[] with uniform distribution (UniformGraphDistribution[]) for ErdösRényi graphs and a scale-free distribution (BarabasiAlbertGraphDistribution[]) for the scale-free networks constructed by starting from a cycle graph
of size 3 and a vertex of k edges added at each step according to the preferential attachment algorithm [1] following a distribution proportional to the
vertex degree. All experiments were replicated and averaged from a set of
at least 20 instances.
5.3
Comparison with entropy and lossless compression
Fig. 6 shows the results obtained by using classical information theory
(Shannon entropy) and one of the most popular lossless compression algorithms (Compress) based on LZW as an approximation to algorithmic
(Kolmogorov-Chaitin) complexity instead of BDM. Because all values collapse into a single value for entropy, the colours displayed are the result of
an artificial sorting of the pixels based on their indices, from top to bottom. Compression is a lower-quality approximation of what we reported in
Figs. 1C-F, where the reported algorithm based on the BDM is clearly an
improvement.
A
B
Figure 6: Shannon entropy (A) and lossless compression (Compress) underperform, not being sensitive enough in performing the same task reported
in Figs. 1C-F.
23
| 2cs.AI
|
Rise of the humanbot
Ricard Solé1, 2, 3
1
ICREA-Complex Systems Lab, Universitat Pompeu Fabra, Dr Aiguader 88, 08003 Barcelona, Spain
Institut de Biologia Evolutiva, CSIC-UPF, Pg Maritim de la Barceloneta 37, 08003 Barcelona, Spain
3
Santa Fe Institute, 1399 Hyde Park Road, Santa Fe NM 87501, USA
arXiv:1705.05935v1 [q-bio.NC] 16 May 2017
2
The accelerated path of technological development, particularly at the interface between hardware
and biology has been suggested as evidence for future major technological breakthroughs associated to our potential to overcome biological constraints. This includes the potential of becoming
immortal, having expanded cognitive capacities thanks to hardware implants or the creation of
intelligent machines. Here I argue that several relevant evolutionary and structural constraints
might prevent achieving most (if not all) these innovations. Instead, the coming future will bring
novelties that will challenge many other aspects of our life and that can be seen as other feasible
singularities. One particularly important one has to do with the evolving interactions between humans and non-intelligent robots capable of learning and communication. Here I argue that a long
term interaction can lead to a new class of “agent” (the humanbot). The way shared memories get
tangled over time will inevitably have important consequences for both sides of the pair, whose
identity as separated entities might become blurred and ultimately vanish. Understanding such
hybrid systems requires a second-order neuroscience approach while posing serious conceptual
challenges, including the definition of consciousness.
Keywords: Singularity, evolution,social interacting robots,major transitions,mind,memory,ageing
Can a machine think? Could it have pain?
Ludwing Wittgenstein
I. INTRODUCTION
The beginnings of the 21-st century have been marked
by a rapid increase in our understanding of brain organisation and a parallel improvement of robots as embodied cognitive agents (Steels 2003; Cangelosi 2010; Nolfi
and Mirolli 2009; Vershure et al 2014). This has taken
place along with the development of enormously powerful
connectionist systems, particularly within the domain of
convolutional neural networks (Lecun et al 2015; Kock
2015). Two hundred years after the rise of mechanical
automata, that became the technological marvels of the
Enlightnment (Woods 2003) new kinds of automata are
emerging, capable of interacting with humans in adaptive ways. The requirements for building an intelligent
or a conscious machine are likely to be still ahead in
the future, but some advances and new perceptions of
the problem are placing the possibility at the forefront
of ”what-if” questions (Vershure 2016). To a large extent, today’s discussion on what separates humans from
their artificial counterparts is deeply tied to the problem of how to properly define mind and consciousness
(Zarkadakis 2015).
In the 1950s, the development of cybernetics by Norbert Wiener and others along with the beginning of theoretical neuroscience and Turing’s proposal for an intelligence test (Turing 1950) were received with a similar
interest, triggering a philosophical debate on the limits
and potential of man-made imitations of life (Nourbakhsh
2013). The study of the first ”cybernetic machines” made
by a few pioneers such as Gray Walter generated a great
expectation. For the first time ”behaviour” emerged as
a word associated to mechanical machines, this time empowered by the rising technology that allowed to combine hardware and a new form of engineering inspired
-to some extent- by natural devices (Water 1950, 1951;
see also Braitenberg 1984). Those early experiments provided some interesting insights into the patterns of exploration of simple agents that where able to detect edges,
respond to light and modify their movements according
to some simple feedback mechanisms. Although their
simple behaviour was essentially predictable, it was not
completely free from surprise (Holland 1997). Later work
in the 1990s and afterwards incorporated artificial neural
systems as an explicit form of introducing learning and
behaviour largely inspired in biology (Vershure et al 1992;
Edelman 1992; Sporns and Alexander, 2002; Prescott et
al 2006).
Nowadays we are rapidly moving towards a new generation of domestic, human-friendly robots that will probably trigger a new technological revolution, similar in some
ways to the one proposed by the first personal computers.
There is a great promise in this revolution. The promise
of robots helping us, playing with us or simply having
some basic support functions is pushing forward software
and hardware companies towards designing cheap robotic
systems. For the general public the fascination remains
in the humanoids, those who really look like us. We
are actually witnessing a rehearsal of the mechanical automata mania of the 18th century, except that the new
automata will be much more autonomous and perhaps
closer to us than ever. Not surprisingly, the rise of the
new robots has come about with a parallel growth of new
fields associated to human-robot interactions (Breazeal
2003; Fong et al 2003; Dautenhahn 2007) and the exploration of learning, cognition and evolution potentials
of artificial agents (Clark and Grush 1999; Schaal 1999;
2
a
b
c
d
r
h
FIG. 1 Humans and robots have been interacting in increasingly more complex ways, most of them limited to simple tasks.
However, communicating robots (a) could in the future interact in deeper ways both among them and with humans (image
courtesy of Luc Steels -the human in the picture- and the Neurocybernetics group at Osnabruck). Fictional stories on future
human-robot interactions, including (b) the movie Robot and Frank or (c) TV series Humans have started to consider the
relevance of strong emotional ties established between elderly or impaired human beings and (possibly non-intelligent) robots
capable of learning and communicating through natural language. Evolving connectomes will inevitably result from HRI (d).
The long-term association between a human and a robotic agent, when the later is equipped with both communication skills
and learning capacities implies the creation of an association network Γ linking both ”brains” through a number of shared
memories. The relative importance of this shared connectome will be a function of the cognitive apparatus of each partner and
the depth of emotional engagement.
Floreano and Mattiusi 2008; Takeno 2013).
Along with the science and engineering of these humanlike agents, science fiction is also exploring some relevant
(and sometimes unexpected) consequences of these near
futures. Most early works on robots have been centered
on the possibility of intelligent, or even self-aware machines. Isaac Asimov published the Sci-Fi classic I robot
three years before Turing’s landmark paper on intelligence, picturing a would-be society where robots become
a central part of our lives (Asimov 1947). Perhaps the
most interesting reflection of Asimov’s tales is the unexpected, the unforeseen consequences of robotic actions
while interacting with their most complex part of the environment: us, the humans. The robots incorporate a set
of hardwired rules (the ”laws” of robotics) preventing the
machines from harming humans and harming themselves
but in special contexts, when a conflict among the laws
emerges, the unexpected should be expected.
The scientific progress made in the field of human-
robot interactions (HRI) and artificial intelligence (AI)
have renewed the interest in the potential consequences
of advanced cognitive robots (Wallach and Allen, 2009;
Bradshaw et al., 2004; Murphy and Woods, 2009) and
in particular the role played by embodiment in communicating agents (figure 1a) which is known to play a key
role in the development of complex behavioural traits
(Steels 2003, 2015). The plots in new Sci-Fi stories have
become more subtle, and much more interesting. In the
movie Robot and Frank, for example, we have a moving
story of a declining man (Frank) facing the initial phases
of Alzheimer’s and a robot companion which, despite a
lack of real intelligence, shares experiences and memories with his human partner (fig 1b). At some point,
the robot asks Frank to wipe out his memories. Frank
rejects the idea, while the robot tells him ”I am not a person, I am just an advanced simulation” but afterwards
he uses a number of very personal (Frank-like) sentences
to make his argument, resulting from previous learned
3
talks and experiences. A related situation also shows
up in an episode of the Sci-Fi series Humans where Dr
George Millican, an elderly man and one of the main
characters, owns a faulty robot named Odi (fig 1c). The
boy-like robot is an old model that stores memory chunks
of shared, unique experiences, some of them related to deceased Millican’s wife. Because of this, Millican does not
accept the mandatory replacement of old models, thus
hiding Odi inside a closet.
None of the previous examples is related to truly intelligent robots. Instead, the subtle ties and their consequences arise from shared cognition and emotions. Nowadays, despite our distance from the sci-fi proposals, robot
designs are already present in a wide range of situations
where they are social partners. This includes household pets, healthcare assistants or even educational companions. Even when the robotic agent has a limited
repertoire of interactions, its actions can be perceived
as part of some ”personality”, particularly in relation
with robotic pets (Min Lee et al 2006; Li and Chignell
2010; Miklosi and Gacsi 2012; Park et al 2012). One goal
of robotic research is the development of autonomous
robots, capable of perception but also of making decisions by themselves and more importantly to communicate (Breazeal 2002; 2003; Breazeal et al 2016). Here several levels of complexity are possible, from programmed
agents responding to a more or less simple and predictable environment to robots capable of complex interactions with humans, able to use natural language and
learn from experience. The later scenario implies sophisticated implementations grounded in neurorobotics (Arbib et al 2006; Oztop et al 2006; Arbib et al 2008). One
goal of this paper is actually provide a general space of
HRI complexity where different classes of interactions can
be located.
As will be argued in the next section, these shared
memories and other common relationships define a network of interactions describing the HRI that can contribute, but also replace, parts of the brain functions
(figure 1d). This diagram displays an idealised bipartite
system (Γ) including both the human cognitive network
(Γh , the brain) and the artificial neural network embedded in the robotic agent (Γr ). The whole network Γ will
change in time as changes in the two subnetworks occur
but also through the creation of all levels of correlations
emerging from the HRI. What happens when this network is the outcome of a long-term exchange? What are
the consequences for a HR pair when the human node is
affected by some kind of impairment? As will be argued
below, by considering the relative cognitive complexities
of each agent in this HRI and the influence of the degree of emotional engagement, a space of HRI can be
defined following the approach of second-person neuroscience (Schlibach et al 2013). Within this space of possibilities, one in particular can involve the emergence of
a new class of cognitive agent transcending both humans
and robots, perhaps defining a novel form of synthetic
evolutionary transition (Solé 2016). The consequences of
such novel hybrid system will be discussed.
II. THE SPACE OF HUMAN-ROBOT INTERACTIONS
Here we are interested in the problem of how a socially interacting robot can trigger emotional and even
cognitive changes in the human partner. In order to do
so we need to take into account several components of
this interaction, which is mediated by a complex network of exchanges and could be described in the context of distributed adaptive control theory (see Lallee and
Vershure 2015). Two major groups of HRI interactions
can be defined here. This first involves those scenarios
where interactions are short-lived (the robot companion
exchanges take place over a small time window) or predictable (the robot is programmed for simple tasks). The
second instead incorporates robotic agents that can learn
from experience and interact through long time spans. A
major difference between the two cases, in terms of the
HRI is the absence and presence of emotional engagement (EE). The choice of EE as a key axis in our approach is grounded in previous studies on second-person
neuroscience (Schlibach et al 2013). In this field, the key
assumption is that an active social engagement between
two (or more) individuals is fundamentally different from
a mere observation of subjects. Within our context, we
will distinguish between a situation of low EE where the
two partners in a HRI are essentially independent from
another one where each agent shapes others cognition.
The space of HRI proposed here is summarised in figure 2a, where three axes have been used to locate different classes of robotic agents. In a way, this can be seen as
the landscape of human-robot interactions. It provides a
tentative picture of the diverse array of qualitative classes
of HRI. Three axes are introduced, namely: the degree of
cognitive complexity of robots and humans and a third
axis linked to the emotional engagement, resulting from
both a long-term interaction and the potential for learning and adaptation displayed by the robot. As shown in
figure 2b and 2c, we can decompose this space, which we
have partitioned in eight arbitrary combinations, in two
main layers including low and high emotional ties. Below
we present and discuss these two layers separately.
A. Low emotional engagement
In this level of engagement, robots are typically associated to simple tasks with low (if any) of cognitive complexity and playing the role of decision-making systems
exhibiting low communication skills (figure 3b). Most
standard socially interacting robots (at least the first
generations of them) fall in this domain. The typical scenario is a programmed artificial agent that can perform
predictable tasks and is expected to operate in simple
environments.
Humans (from toddlers to adults) interacting with sim-
4
ple robots (such as toys or automata following simple
orders) would define a lower bound of low cognitive complexity for both agents. Cleaning robots such as Roomba
and receptionist robots answering requests from clients
(giving simple types of information) would be obvious
examples, although even in this case the use of personalization toolkits triggers emotional connections (Sung et
al 2009).
In this domain, robots with high cognitive skills interacting with cognitively impaired individuals over specific
tasks give the last option in our list. Robots can help in
providing support to the elderly or impaired in ways that
do not require learning potential from the artificial agent
side. Robots capable of identifying the needs of their
human partners can be helpful even if not emotionally
engaged, while robots with a large amount of available
data sets can be interesting as expert systems with userfriendly (humanoid) interfaces. Instead we weight cognition in terms of (pre-programmed) diverse repertoires of
responses.
Agents and operative systems exhibiting a rich repertoire of interactions, such as SIRI (which can assist blind
people) or Alexa would fit the high-cognitive complexity corner. Here of course ”cognition” does not have the
meaning that can be attributed to a neural system. For
some of these non-embodied systems, a very large repertoire of potential answers can be communicated with the
help of a natural language interface. The success of some
systems such as Watson, which was trained for a specific
goal (answering questions on the Jeopardy quiz show)
using hundreds of millions of webpages illustrate the potential for surpassing humans in searching and solving
questions.
This layer and the next layer in our space, now allowing higher emotional engagement, are not clear cut. This
is particularly relevant if we take into account the human
tendency to extract behavioural or emotional clues from
the interaction with even simple computer programs. It
is worth remembering that even the earliest attempts to
program machines (Weizenbaum 1966) capable for answering questions, such as ELIZA (which used pattern
matching to simulate a conversation) led to rather unexpected reactions (Weizenbaum 1976). ELIZA was supposed to imitate a speaking psychiatrist, essentially creating simple responses triggered by key words that were
then used to create simple questions. But in many cases,
people failed to believe they were talking with a computer
(which was far beyond in computer power from anything
existing today). Similarly, some robots can be simple
companions if their interactions span a short time scale
but create a strong bond (from human to machine) with
their owners if interactions occur over long, shared periods of time. To reach the second layer, we must allow
artificial agents to learn and adapt in flexible ways, as
well as be capable of exhibiting and detecting emotions.
B. High emotional engagement
Long-term relationships in HRI can lead to a rather
distinct set of patterns that strongly depart from the
previous layer. Here we consider the possibility that the
robotic agents have been interacting with a given individual in a flexible manner and over an extended period of
time. Such interaction might occur at different levels and
this too is strongly dependent on the relative cognitive
complexity or each partner (the human and the robot)
as well as the level and span of their emotional interaction. A healthy, adult human brain is a highly complex
system that can nevertheless engage in an emotionally
strong relationship with a pet. Humans and dogs (and
other pets) have been co-evolving over hundreds of thousands of years until today, where our animal companions
have limited cognition powers but a highly developed set
of skills associated to emotion recognition (Hare et al
2002). One side effect of this process is our tendency to
generate empathy for non-living objects resembling pets.
These evolutionary responses have left a detectable cognitive signal (Stoeckel et al 2014). What has been the
outcome of HRI between human and artificial pets? A
first glimpse of the implications of long-term exchanges
between humans and pet robots capable of learning was
provided by AIBO dogs, used as companions in a broad
range of conditions, from preschool children to elder patients (Kahn et al 2006, Melson et al 2009). In the future, these type of pets (as well as those emerging from
the Virtual reality world) are likely to become the rule
rather than the exception (Rault 2015).
Despite the lack of a complex communication, the domestication of dog’s cognition has lead to emotional ties
that can be strong. All studies on HRI involving robotic
pets reveal that human perception is markedly biased towards perceiving them as life-like entities and treated as
such, instead as artifacts (Kahn et al 2006). This creates
a number of interesting situations relevant for the problem addressed here, an in particular the blurring of lines
between ontological categories. As a consequence, emotional ties and their consequent effects are likely to be
shared. Such effects are enhanced by the development
of some ”personality” associated to robot learning capacities, which can strengthen emotional ties. These differential behaviours result from the historical sequence of
HRI events: by tapping AIBO’s head sensor after a given
behavioural display, it is possible to enforce or decrease
a given response, but each case and how it relates with
other responses will depend on each specific HRI. Among
other things, the loss of a pet can trigger a strong reaction in the human, but also in the animal end of the
tie. Perhaps not surprisingly a similar situation has been
emerging in relation with AIBO owners who, unable to
repair their old dog robots, have to accept their ”death”
and eventually organise funerals not different from those
made for the living counterparts. In this case, the human and the robot are separated by a huge cognitive
complexity gap, but the capacity of the robot for learn-
5
HAL9000
Expert AI
Artificial cognitive complexity
Watson
Humanbot
collaborative HRI
Siri
Alexa
recepcionist !
robots
Therapeutic!
robots
power exoskeleton
an c
ogn
GWT
Aibo (dog robot)
Roomba
Paro !
(seal robot)
Hum
Asimo
game-playing!
robots
itive
Nao
com
plexi
ty
KOG
ELIZA
Kismet
!
ot
Em
ion
a
nt
me
e
ag
ng
le
FIG. 2 The cognitive space of H-R interactions under the two-agent perspective taken here. Non-robotic systems (such as
ELIZA, SIRI or Watson) are also included. The axes represent (in a qualitative ordering) emotional engagement as well as
two dimensions associated to the complexity displayed by the human and the artificial agents, respectively. Here the relative
location of each system needs to be taken as indicative. Some of these robots appear in the high human cognitive complexity
domain, mainly because their full operational function is expected to take place here, although they could in principle be
useful in the other side of the cognition space. The domain where humanbot systems would emerge is indicated by a blurred
sphere. A broad array of possible HRI pairs could be expected here, involving strong cognitive dependences that could generate,
particularly when the human side is impaired, a new class of cognitive agent.
ing and somewhat adopt some personality makes a big
difference for their owners.
The previous example would correspond to the possible
interactions in the lower part of our space (figure 2c) associated with low robotic cognition. A relevant scenario
is the interaction between elderly patients requiring care
and therapy (the so called fourth age) and some robots,
such as Paro: an artificial, seal-shaped agent. Paro is
equipped with sensors detecting touch, light, sounds or
movement, with motors and actuators and other standard robotic features but also with some additional features such as responses to cuddling and a constant seeking of eye contact. Additionally, several emotions have
been programmed and it can respond to its own name
as well as learn other names. Paro has been used in
treatment of patients with dementia. The result of these
interactions was a reduction of anxiety and helping recover from chronic ailments as well as improving communication with other patients and often creating strong
attachments (Kidd et al 2006, Takayanagi et al 2014).
Similar patterns have been found using AIBO (Melson
et al 2009).
Most companion robots helping patients with dementia
are programmed to respond to specific tasks in a more or
less flexible way. On a basic level, the robot can be very
useful by reminding the human the name and/or location
of objects. This has been proposed as a memory prosthetics for elders (Ho et al 2013) where the robotic companion would be equipped with a visual episodic memory, but consider now a HRI capable of learning and using natural language. On another level, it can serve as a
medium to communicate with other humans and preventing isolation. Similarly, if capable of moving in outdoors
environments, it could greatly improve orientation. But
what if the artificial agent is equipped with a powerful
cognitive system beyond simple reactions and capable of
dealing with daily living environments using multimodal
integration learning? Here deep neural networks can play
a fundamental role, including the use of natural language
(Noda et al 2014).
Consider a robotic agent capable of maintaining a conversation, using natural language, and capable of gathering relevant information concerning past events related
with the life of the patient, detecting goals and wishes
as well as emotions and capable also of facial expression
to share emotional states. This would also be a very de-
6
a
Robot cognition
+
Low emotional engagement
Cognitive !
support!
robot
Robot cognition
f
g
h
i
j
k
Expert robots
Receptionist robot
Cleaning robot
-
+
e
Game-playing robots
-
b
d
Human cognition
+
High emotional engagement
HumanBot
Collaborative!
HR interaction
Therapeutic robots
Home support robots
-
Robotic pets
-
Human cognition
+
FIG. 3 The cognitive space of H-R interactions (a-b) here the two layers associated to different emotional engagement (related
to short/long term interactions) have been separated indicating potential examples within each discrete class. Several examples
of robotic agents engaged in HRI are shown in d-k, including programmed systems with little or no emotional responses nor
learning capabilities to robots capable of simulating emotions, seeking eye contact and other features that enhance human
emotional and social responses. Here we show: (d) Roomba cleaning robot, (e) REEM, from PAL robotics, (f) ASIMO, (g)
Pepper, (h) Aibo, (i) Paro the seal, (j) Nao and (k) Milo.
sirable form of interaction, since the loss of memories or
-more generally- the difficulties to access stored recollections can be a great source of stress for elder people with
mild cognitive impairment1 and Alzheimer’s patients in
particular, specially in early phases of the disease. An artificial agent capable of coping with memory decay and
disorientation by means of verbal communication would
leverage the anxieties associated to cognitive decline and
improve reasoning and judgement. If flexible enough, a
1
This refers usually to a transition stage between a normal process
of brain ageing and dementia, characterised by low performance
in memory tasks
neural agent can learn how to help the human in the
most useful and personalised way. And here comes the
problem.
The HRI involved here leads to a rather unique outcome. For example, if the artificial agent provides support to losses in episodic memory, a first paradoxical
situation might emerge: it can occur that some events
become absent from the impaired brain while they remain stored within the artificial agent. Such stored recollections can be easily over-interpreted by the artificial
agent, in particular their relative value and potential correlations among different memories. On a general level,
reminding the human subject where are the lost keys or a
misplaced wallet are simple and yet important tasks that
can easily counterbalance or alleviate early symptoms.
7
But it can also have a major impact on autobiographical
memory. This class of memory, associated to the left prefrontal cortex, provides the basis for putting together a
timeline of past events connected to visual and sensoryperceptual features. We should have in mind that a major limitation of robotic agents (and a crucial component
of the human mind, Suddendorf 2013) is their lack of
understanding or representation of time.
The agent, if properly equipped with visual recognition systems, can have seen pictures of family members
(alive or not) and learned about their stories from the patient. These stories can be true or not, but in both cases
the agent will contribute to store and recover them. A
great advantage of any neural-based system capable of
pattern recognition and generalisation (as deep learning
networks, see LeCun et al 2015, and references therein)
is that the robot can extract correlations required to
help in more complex tasks, particularly in relation to
episodic, semantic and working memory as well as language and executive functioning. But these correlations
rely on both the (possibly faulty) input provided by the
human and the emotional weight given to each memory
by both partners within the HRI. Because correlations
are likely to be generated from biased perceptions, the
resulting internal correlation matrix created by the robot
(and returned to the human through HRI events) can depart from the original correlations generated in the brain.
If we take into account that memories themselves are
not reliable (this in particular affects priming) the longterm HRI inevitably leads to considerable deviations
from the original memory web. In this respect, we have a
first glimpse of an anomalous pattern: potential memory
deficits are compensated by the reliable storage of information residing in the artificial neural network, which can
use false memories, create incorrect (but strong) correlations associated to emotional events and feed-back them
into the mind of the human. Can this process lead to a
runaway amplification phenomenon? What seems likely
to occur is that mismatches between the relevance and
emotional weight associated to different memory patterns
and their interactions might promote deep distortions of
the memory and behavioral landscapes. As the HRI proceeds, a new set of interactions will coalesce between the
human connectome Γh (figure 1d) and the network of
neural correlations created within the robot brain (Γr ).
The whole cognitive map must be found in the merged
structure Γ that includes both networks along with all the
human-robot correlations that have emerged and that we
also indicate as links between areas.
The loss of plasticity that results from ageing or damage reinforces dependencies among human and robotic
cognitive maps mediated by co-occurrence patterns. The
view of a given object or image elicits a response in both
sides that defines effectively an interaction between both,
since these responses will immediately lead to an information exchange. Consider for example a picture of someone
(or any other representation of it) that is identified by the
human and also stored by the robotic companion. Pre-
vious exchanges will have weighted the relevance of this
picture and the associated subnetwork of related objects,
people or actions.
As cognitive impairment grows in time, the relative
importance of the object representation within the artificial system might have been enhanced beyond its original relevance, while exciting and reinforcing other related
subnetworks. If decisions or actions derive from this perception, the loss of proper decision making by the human
might have been displaced towards the robot, thus shifting the deep correlates from one agent to the other and
helping to preserve episodic and semantic memory. Alternatively, different perception and decision layers might
become segregated between them. If the memory of this
specific image is gone from the human, it can nevertheless remain accessible in the artificial side, which would
now contain part of the autobiographical memory.
Since other related events connected to this memory
might indirectly interact with the brain network, it is possible that novel forms of hybrid memory might emerge.
However, despite the positive side of keeping otherwise
erased memories, the dynamics associated to the formation of the HR network can easily shift the importance of
events over time and even disrupt it. Unless under some
external supervision by close relatives of the patient, the
humanbot can lead to a shared mind that strongly departs from the organisation and coherence of the original subject. In other words, one outcome of this HRI is
the conscious experience of a different subject, like living
in the mind of someone else. The humanbot is a likely
outcome of future HRI and its potential to become real
is tied to the new generations of robots equipped with
powerful learning systems and high memory capacities.
These robots might be not yet here, but they are certainly
much closer than the coming of intelligent machines.
III. DISCUSSION
The increasing frequency of dementias, being
Alzheimer’s the most common one, will affect millions
of human beings in the next decades (Reitz et al 2011).
While prevention strategies are developed and new drugs
are been tested, the need to caregivers helping these patients is becoming a major issue. Beyond the staggering
economic costs, caregivers (often family members) are
also affected by strong physical and emotional stress.
In many cases, their health is also deteriorated. The
possibility of using a robotic agent providing help seems
a desirable option, although ethical issues need to be
considered (Ienca et al 2016). Because cognitive decay is
a central issue in most cases, the artificial agent should
be equipped with a flexible, adaptive system capable of
dealing with changing conditions and specific needs. But
such plasticity and learning potential can generate new
emergent phenomena. In this paper we have explored
the potential outcomes of long-term HRI with artificial
agents able to replace memory deficits and communicat-
8
ing through a natural language interface. As discussed
above, the increasing replacement of faulty cognitive
networks is likely to create a profound dependency of the
human companion that can eventually end in a blurred
boundary between the robot and the brain.
There are other implications derived from this class of
long-term HRI:
(a) The theory of attractor neural networks (Amari
and Maginu 1988; Amit 1992; Rojas 2013) has shown
that the qualitative responses of neural networks concerning their responses to noise, memory potential and other
properties can experience sharp changes as some parameters are tuned. One particular example is the rapid
decline of associative memory as the neural network is
damaged beyond a given threshold. Similarly, the qualitative changes associated to an mismatch between the
memory requirements and the available cognitive power
can lead to the emergence of spurious memory states that
is decoupled from the real repertoire of original memories. In all these cases, neural networks display different
phases separated by well defined phase transition points
(Amit 1992). If these results, grounded in simple neural
network models, can be extrapolated to our hybrid system, we would expect to observe tipping points in the
cognitive organisation of memories and other key properties.
(b) human beings display awareness, while machines
(so far) do not. A relevant problem concerning consciousness is how to define it and even how to measure
it (Tononi and Edelman 1998). Some efforts in this direction suggest that a single parameter Φ could be defined that can capture the degree of consciousness (Tononi
2012). Using this parameter, obtained from an information theoretic approach, it has been argued that we can
to the least order different case studies, from animals to
impaired human brains or machines (Tononi and Koch
2015). An interesting outcome of this approach is the
suggestion that machines (in particular those based on
von Neumann architecture) are not conscious (Tononi et
al. 2016). Without discussing this conclusion, it seems
clear that the humanbot, by inhabiting the boundaries
between human and machine will also incorporate some
level of consciousness (as measured by Φ). Under the
conditions described above, we need to ask the impact of
the cognitive replacement associated to the increasing interdependence and how is consciousness shared by both
parts.
(c) Once the human is gone, we should seriously evaluate what is left behind within the robot. As long term
interactions are likely to shape the robotic cognitive network, some key components of the human’s mind might
remain there for inspection or preservation. Once deceased, what is left can keep changing as other inputs
from the external world keep entering the system, thus
modifying or even erasing the previous stored memories.
What to do next? Should the capacity for inspecting the
environment be put on hold? Would the robot be ca-
pable of interacting with friends and close people of the
deceased in meaningful ways? Should the stored information be preserved as the only relevant remain of the
gone mind?
One of the most interesting and puzzling components
of Asimov’s vision of a future with robots playing a major role in our society was the existence of a novel research field: robopsychology. Experts in this area had to
deal with the sometimes unexpected behavior of robots,
emerging from the conflicts associated to the there rules
of robotics and their inevitable interaction with a complex external world. In the picture presented here, a
different (but related) class of robotic psychology might
emerge in the future. The interaction, merging and blurring of behavioral patterns resulting from the HRI described above defines an uncharted territory. Long before
machines might outsmart us or develop consciousness or
intelligent behaviour (Barrat 2013) we will need to either
face the rise of the humanbot or prevent it to happen.
Acknowledgments
The author would like to thank Luis Seoane, Marti
Sanchez-Fibla and Paul Vershure for a critical reading
of the manuscript and useful discussions on artificial intelligence, evolved agents and complexity. Also to Joan
Manel Solé, with whom I discussed and wrote as freshmen my first codes implementing the ELIZA program.
This work has been supported by the Botı́n Foundation
by Banco Santander through its Santander Universities
Global Division, a MINECO FIS2015-67616 fellowship
and by the Santa Fe Institute.
IV. REFERENCES
1. Amari, S. I., and Maginu, K. (1988). Statistical
neurodynamics of associative memory. Neural Networks 1, 63-73.
2. Amit, D. J. (1992). Modeling brain function: The
world of attractor neural networks. Cambridge University Press.
3. Arel, I., Rose, D. C., and Karnowski, T. P. 2010.
Deep machine learning-a new frontier in artificial
intelligence research [research frontier]. Comput.
Intell. Magazine IEEE 5, 13-18.
4. Barrat, J. (2013) Our final invention. Artificial intelligence and the end of the human era. Thomas
Dunne Books. New York.
5. Bradshaw, Jeffrey M., et al. 2004. Making agents
acceptable to people. In: Intelligent Technologies
for Information Analysis. Springer Berlin Heidelberg, pp. 361-406.
9
6. Breazeal, C. L. 2004. Designing sociable robots.
MIT press.
7. Breazeal, C. L. et al. 2016. Social Robotics.
In: Springer Handbook of Robotics. Springer. pp.
1935-1972.
8. Buttazzo, G. 2001. Artificial consciousness: Utopia
or real possibility?. Computer 34, 24-30.
9. Clark, A. and Grush, R. (1999). Towards a cognitive robotics. Adaptive Behavior, 7, 5-16.
10. Edelman, G. 1992. Bright air, brilliant fire. Basic
Books.
11. Fellous, J. M. and Arbib, M. A. (Eds.). 2005. Who
needs emotions?: The brain meets the robot. Oxford University Press.
12. Floreano, D., Ijspeert, A. J. and Schaal, S. 2014.
Robotics and Neuroscience. Curr. Biol. 24, R910R920.
13. Gorbenko, A., Popov, V. and Sheka, A. 2012.
Robot self-awareness: Exploration of internal
states. Applied Math. Sci. 6, 675-688.
14. Holland, O. 2003. Machine consciousness. Imprint
Academic.
22. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.
A. et al. 2015. Human-level control through deep
reinforcement learning. Nature 518, 529-533.
23. Noda, K., Arie, H., Suga, Y. and Ogata, T. (2014).
Multimodal integration learning of robot behavior using deep neural networks. Robotics and Autonomous Systems 62, 721-736.
24. Nourbakhsh, I. R. 2013. Robot futures. MIT Press.
25. Oztop, E., Kawato, M. and Arbib, M. 2006. Mirror
neurons and imitation: A computationally guided
review. Neural Networks 19, 254-271.
26. Prescott, T. J., Gonzalez, F. M. M., Gurney, K.,
Humphries, M. D. and Redgrave, P. (2006). A
robot model of the basal ganglia: behavior and intrinsic processing. Neural Networks, 19, 31-61.
27. Rault, J. L. (2015). Pets in the Digital Age: Live,
Robot, or Virtual?. Frontiers Vet. Science, 2.
28. Reitz, C., Brayne, C. and Mayeux, R. (2011). Epidemiology of Alzheimer disease. Nature Rev. Neurology, 7, 137-152.
29. Reggia, J. A. 2013. The rise of machine consciousness: Studying consciousness with computational
models. Neural Networks 44, 112-131.
15. Ienca, M., Jotterand, F., Vic?, C. and Elger, B.
(2016). Social and Assistive Robotics in Dementia
Care: Ethical Recommendations for Research and
Practice. International Journal of Social Robotics
8, 565-573.
30. Rojas, R. (2013). Neural networks: a systematic
introduction. Springer Science.
16. Kahn P. H., Friedman, B., Perez-Granados, D. R.
and Freier, N. G. (2004, April). Robotic pets in
the lives of preschool children. Interaction Studies
7, 405-436.
32. Sporns, O., and Alexander, W. H. (2002). Neuromodulation and plasticity in an autonomous robot.
Neural Networks, 15, 761-774.
17. Kidd, C. D., Taggart, W., and Turkle, S. (2006).
A sociable robot to encourage social interaction
among the elderly. In Proceedings 2006 IEEE International Conference on Robotics and Automation, 2006. ICRA 2006. pp. 3972-3976.
18. Koch, C. 2015. Do Androids Dream?. Sci. Am.
Mind 26, 24-27.
19. LeCun, Y., Bengio, Y. and Hinton, G. 2015. Deep
learning. Nature, 521, 436-444.
20. Marques, H. G. and Holland, O. 2009. Architectures for functional imagination. Neurocomputing
72, 743-759.
21. Melson, G. F., Kahn Jr, P. H., Beck, A. and Friedman, B. (2009). Robotic pets in human lives:
Implications for the human-animal bond and for
human relationships with personified technologies.
Journal Social Issues, 65, 545-567.
31. Solé, R. (2016). Synthetic transitions: towards a
new synthesis. Phil. Trans. R. Soc. B, 371,
20150438.
33. Sung, J., Grinter, R. E. and Christensen, H. I.
(2009, April). Pimp my roomba: Designing for personalization. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems
(pp. 193-196). ACM.
34. Suddendorf, T. 2013. The gap: The science of what
separates us from other animals. Basic Books.
35. Takayanagi, K., Kirita, T., and Shibata, T. (2014).
Comparison of Verbal and Emotional Responses of
Elderly People with Mild/Moderate Dementia and
Those with Severe Dementia in Responses to Seal
Robot, PARO. Frontiers Aging Neurosci. 6, 257.
36. Takeno, J. 2012. Creation of a Conscious Robot:
Mirror Image Cognition and Self-Awareness. CRC
Press.
37. Tononi, G. and Koch, C. 2015. Consciousness:
here, there and everywhere?. Phil. Trans Royal
Soc. London B 370, 20140167.
10
38. Tononi, G. 2012. Φ, Phi, a voyage from the brain
to the soul. Pantheon Books. New York.
39. Tononi, G. and Edelman, G. M. 1998. Consciousness and complexity. Science 282, 1846-1851.
40. Tononi, G., Boly, M., Massimini, M. and Koch,
C. (2016). Integrated information theory: from
consciousness to its physical substrate. Nature Reviews Neuroscience (in press).
41. Turing, A. M. 1950. Computing machinery and
intelligence. Mind 59, 433-460.
42. Verschure, P. F., Kröse, B. J. and Pfeifer, R. (1992).
Distributed adaptive control: The self-organization
of structured behavior. Robotics and Autonomous
Systems, 9, 181-196.
43. Verschure, P. F., Pennartz, C. M. and Pezzulo, G.
(2014). The why, what, where, when and how of
goal-directed choice: neuronal and computational
principles. Phil. Trans. R. Soc. B, 369, 20130483.
44. Verschure, P. F. (2016). Synthetic consciousness:
the distributed adaptive control perspective. Phil.
Trans. R. Soc. B, 371, 20150448.
45. Weizenbaum, J. (1966). ELIZA: a computer program for the study of natural language communication between man and machine. Comm. ACM.
36-45.
46. Zarkadakis, G. (2015). In Our Own Image: Will
artificial intelligence save or destroy us?. Random
House. New York.
| 2cs.AI
|
Sparse Sliced Inverse Regression for High Dimensional Data
Qian Lin
arXiv:1611.06655v1 [math.ST] 21 Nov 2016
Center of Mathematical Sciences and Applications
Harvard University
Cambridge, MA 02138
USA
E-mail: [email protected]
Zhigen Zhao
Department of Statistical Science
Fox School of Business
Temple University
Philadelphia, PA 19122
E-mail: [email protected]
Jun S. Liu
Department of Statistics
Harvard University
Cambridge, MA 02138-2901
USA
E-mail: [email protected]
Summary. For multiple index models, it has recently been shown that the sliced inverse regression
(SIR) is consistent for estimating the sufficient dimension reduction (SDR) subspace if and only if
the dimension p and sample size n satisfies that ρ = lim np = 0. Thus, when p is of the same or
a higher order of n, additional assumptions such as sparsity have to be imposed in order to ensure
consistency for SIR. By constructing artificial response variables made up from top eigenvectors of
the estimated conditional covariance matrix, vd
ar(E[x|y]), we introduce a simple Lasso regression
method to obtain an estimate of the SDR subspace. The resulting algorithm, Lasso-SIR, is shown
to be consistent and achieve the optimal convergence rate under certain sparsity conditions when p
is of order o(n2 λ2 ) where λ is the generalized signal noise ratio. We also demonstrate the superior
performance of Lasso-SIR compared with existing approaches via extensive numerical studies and
several real data examples.
1.
Introduction
Dimension reduction and variable selection have become indispensable steps for most data
analysts in this big data era, where thousands or even millions of features are easily obtained
for only hundreds or thousands of samples (n). With these ultra high-dimensional data, an
effective modeling strategy is to assume that only a few features and/or a few linear combinations of these features carry the information that researchers are interested in. Namely,
2
Lin, Zhao, and Liu
one can consider the following multiple index model [Li, 1991]:
y = f (βτ1 x, β τ2 x, ..., βτd x, ǫ),
(1)
where x follows a p dimensional elliptical distribution with covariance matrix Σ, the βi ’s are
unknown projection vectors, d is unknown but is assumed to be much smaller than p, and the
error ǫ is independent of x and has mean 0. When p is very large, it is perhaps reasonable
to further restrict each βi to be a sparse vector.
Since the introduction of the sliced inverse regression (SIR) method by Li [1991], many
sufficient dimension reduction (SDR, a term coined by Cook [1998]) methods have been
proposed to estimate the space spanned by (β 1 , · · · , βd ) with few assumptions on the link
function f (·). With the multiple index model (1), the objective of all SDR methods is to find
the minimal subspace S ⊆ Rp such that y ⊥
⊥ x | PS x, where PS stands for the projection
operator to the subspace S. When the dimension of x is moderate, all SDR methods including
SIR, which is most popular due to its simplicity and computational efficiency, have shown
successes in various application fields [Xia et al., 2002, Ni et al., 2005, Li and Nachtsheim,
2006, Li, 2007, Zhu et al., 2006]. However, these methods were previously known to work
only when the sample size n grows much faster than the dimension p, an assumption that
becomes inappropriate for many modern datasets, such as those from biomedical researches.
It is thus of great interest to have a thorough investigation of “the behavior of these SDR
estimators when n is not large relative to p”, as raised by Cook et al. [2012].
Lin et al. [2015] made an attempt to address the aforementioned challenge for SIR. They
showed that, under mild conditions, the SIR estimate of the central space is consistent if and
only if ρn = lim np goes to zero as n grows. Additionally, they showed that the convergence
rate of the SIR estimate of the central space (without any sparsity assumption) is ρn . When
p is much greater than n, however, certain constraints must be imposed in order for SIR to be
consistent. The sparsity assumption, i.e., the number of active variables (s) must be an order
of magnitude smaller than n and p, appears to be a reasonable one. In a follow-up work,
Neykov et al. [2016] studied the sign support recovery problem of the single index model
(d = 1), suggesting that the correct optimal convergence rate for estimating the central space
, a speculation that is partially confirmed in Lin et al. [2016]. It is shown that,
might be s log(p)
n
for multiple index models with bounded dimension d and the identity covariance matrix, the
log(p/s)
optimal rate for estimating the central space is ds+s nλ
, where λ is the smallest non-zero
eigenvalue of var(E[x|y]). They further showed that the Diagonal-Thresholding algorithm
proposed in Lin et al. [2015] achieves the optimal rate for the single index model with the
identity covariance matrix.
The main idea. In this article, we introduce an efficient Lasso variant of SIR for the multiple
index model (1) with a general covariance matrix Σ. Consider the single index model:
y = f (βτ x, ǫ). Let η be the eigenvector associated with the largest eigenvalue of var(E[x|y]).
Since β ∝ Σ−1 η, there are two immediate ways to estimate the space spanned by β. The
first approach, as discussed in Lin et al. [2015], estimates the inverse of Σ and η separately
(see Algorithm 1). The second one avoids estimating the inverse of the covariance matrix Σ
directly by solving the following penalized least square problem k n1 XX τ β − ηk22 + µkβk1 ,
where X is the p × n covariate matrix formed by the n samples (see Algorithm 2). However,
Sparse Sliced Inverse Regression for High Dimensional Data
3
similar to most L1 penalized approaches for nonlinear models, theoretical underpinning of this
approach has not been well understood. Since these two approaches provide good estimates
compared with other existing approaches (e.g.,Li [1991], Li and Nachtsheim [2006], Li [2007]),
we set them as benchmarks for further comparisons.
b of vd
We note that an eigenvector η
ar(E[x|y]), where vd
ar(E[x|y]) is an estimate of the
conditional covariance matrix var(E[x|y]) using SIR [Li, 1991], must be a linear combination
e ∈ Rn
of the column vectors of X. Thus, we can construct an artificial response vector y
1
e , and estimate β by solving another penalized least square problem:
b = nXy
such that η
τ
1 e
2
ky − X βk2 + µkβk1 (see Algorithm 3). We call this algorithm “Lasso-SIR”, which is
2n
computationally very efficient. In Section 3, we further show that the convergence rate of
, which is optimal if s = O(p1−δ ) for some
the estimator resulting from Lasso-SIR is s log(p)
nλ
positive constant δ. Note that Lasso-SIR can be easily extended to other regularization and
SDR methods, such as SCAD (Fan and Li [2001]), Group Lasso (Yuan and Lin [2006]), sparse
Group Lasso (Simon et al. [2013]), SAVE (Cook [2000]), etc.
Connection to Other work. Estimating the central space is widely considered as a generalized eigenvector problem in the literature [Li, 1991, Li and Nachtsheim, 2006, Li, 2007,
Chen and Li, 1998]. Lin et al. [2016] explicitly described the similarities and differences between SIR and PCA (as first studied by Jung and Marron [2009]) under the “high dimension,
low sample size (HDLSS)” scenario. However, after comparing their results with those for
Lasso regression, Lin et al. [2016] advocated that a more appropriate prototype of SIR (at
least for the single index model) should be the linear regression. In the past three decades,
tremendous efforts have been put into the study of linear regression models y = xτ β + ǫ for
HDLSS data. By imposing the L1 penalty on the regression coefficients, the Lasso approach
[Tibshirani, 1996] is rate optimal in estimating β in the sense that it achieves the maximum convergence rate [Raskutti et al., 2011]. Because of apparent limitations of linear models, there are many attempts to build flexible and computationally friendly semi-parametric
models, such as the projection pursuit regression [Friedman and Stuetzle, 1981, Chen, 1991],
sliced inverse regression [Li, 1991], and MAVE [Xia et al., 2002]. However, none of these
methods work under the HDLSS setting. Existing theoretical results for HDLSS data mainly
focus on linear regressions [Raskutti et al., 2011] and submatrix detections [Butucea et al.,
2013], and are not applicable to index models. In this paper, we provide a new framework
for the theoretical investigation of regularized SDR methods for HDLSS data.
The rest of the paper is organized as follows. After briefly reviewing SIR, we present
the Lasso-SIR algorithm in Section 2. The consistency of the Lasso-SIR estimate and its
connection to the Lasso regression are presented in Section 3. Numerical simulations and
real data applications are reported in Sections 4 and 5. Some potential extensions are briefly
discussed in Section 6. To improve the readability, we defer all the proofs and brief reviews
of some existing results to the appendix.
4
2.
Lin, Zhao, and Liu
Sparse SIR for High Dimensional Data
2.1. Notation
We adopt the following notations throughout this paper. For a matrix V , we call the space
generated by its column vectors the column space and denote it by col(V ). The i-th row
and j-th column of the matrix are denoted by V i,∗ and V ∗,j , respectively. For (column)
vectors x and β ∈ Rp , we denote their inner product hx, βi by x(β), and the k-th entry
of x by x(k). For two positive numbers a, b, we use a ∨ b and a ∧ b to denote max{a, b}
and min{a, b} respectively; We use C, C ′ , C1 and C2 to denote generic absolute constants,
though the actual value may vary from case to case. For two sequences {an } and {bn }, we
denote an ≻ bn and an ≺ bn if there exist positive constants C and C ′ such that an ≥ Cbn
and an ≤ C ′ bn , respectively. We denote an ≍ bn if both an ≻ bn and an ≺Pbn hold. The
(1, ∞) norm and (∞, ∞) norm of matrix A are defined by kAk1,∞ = max1≤j≤p pi=1 |Ai,j | and
is sufficiently
max1≤i,j≤n kAi,j k respectively. To simplify discussions, we assume that s log(p)
nλ
small. We emphasize again that our covariate data X is a p × n instead of the traditional
n × p matrix.
2.2. A brief review of Sliced Inverse Regression (SIR)
In the multiple index model (1), the matrix B formed by the vectors β1 , ..., βd is not identifiable. However, col(B), the space spanned by the columns of B is uniquely defined. Given
n i.i.d. samples (yi , xi ), i = 1, · · · , n, as proposed in Li [1991], SIR first divides the data into
H equal-sized slices according to the order statistics y(i) , i = 1, . . . , n. To ease notations and
arguments, we assume that n = cH and E[x] = 0, and re-express the data as yh,j and xh,j ,
where h refers to the slice number and j refers to the order number of a sample in the h-th
slice, i.e., yh,j = y(c(h−1)+j) , xh,j = x(c(h−1)+j) . Here x(k) is the concomitant of y(k) . Let the
sample mean in the h-th slice be denoted by xh,· , then Λ , var(E[x|y]) can be estimated by:
b
Λ
H
H
1 X
1
=
x̄h,· x̄τh,· = X τH X H
H h=1
H
(2)
where X H is a p×H matrix formed by the H sample means, i.e., X H = (x1,· , . . . , xH,· ). Thus,
cH ), where V
cH is the matrix formed by the top d eigenvectors
col(Λ) is estimated by col(V
b
c
of ΛH . The col(V H ) was shown to be a consistent estimator of col(Λ) under a few technical
conditions when p is fixed [Duan and Li, 1991, Hsing and Carroll, 1992, Zhu et al., 2006, Li,
1991, Lin et al., 2015], which are summarized in the online supplementary file. Recently,
cH ) is consistent for col(Λ) if and only if ρn = p → 0
Lin et al. [2015, 2016] showed that col(V
n
as n → ∞, and that the number of slices H can be chosen as a fixed integer independent
of n and p when the dimension d of the central space is bounded. When x’s distribution is
elliptically symmetric, Li [1991] showed that
Σcol(B) = col(Λ),
(3)
and thus our goal is to recover col(B) by solving the above equation, which can be easily
b −1 col(V
cH ) when ρn → 0 [Lin et al., 2015], where
\ =Σ
and consistently achieved by col(B)
5
Sparse Sliced Inverse Regression for High Dimensional Data
b = 1 XX τ , the sample covariance matrix of x. However, this simple approach breaks
Σ
n
down when ρn 6→ 0, especially when p ≫ n. Although stepwise methods [Zhong et al., 2012,
Jiang and Liu, 2014] can work under HDLSS settings, the sparse SDR algorithms proposed
in Li [2007] and Li and Nachtsheim [2006] appeared to be ineffective. Below we describe two
intuitive non-stepwise methods for HDLSS scenarios, which will be used as benchmarks in
our simulation studies to measure the performance of newly proposed SDR algorithms.
Diagonal Thresholding-SIR. When p ≫ n, the Diagonal Thresholding (DT) screening
method [Lin et al., 2015] proceeds by marginally screening all the variables via the diagonal
b and then applying SIR to those retained variables to obtain an estimate of
elements of Λ
H
col(B). The procedure is shown to be consistent if the number of nonzero entries in each row
of Σ is bounded even when p ≫ n.
Algorithm 1 DT-SIR
b to select the set of important predic1: Use the magnitudes of the diagonal elements of Λ
H
tors I, with |I| = o(n)
2: Apply SIR to the data (y, xI ) to estimate a subspace SbI .
3: Extend SbI to a subspace in Rp by filling in 0′ s for unimportant predictors.
Matrix Lasso. We can bypass the estimation and inversion of Σ by solving an L1
penalization problem. Since (3) holds at the population level, we speculate that a reasonable
estimate of col(B) can be obtained by solving a sample-version of the equation with an
b 1, · · · , η
b d be the
appropriate regularization term to cope with the high dimensionality. Let η
b
eigenvectors associated with the largest d eigenvalues of ΛH . Replacing Σ by its sample
version n1 XX τ and imposing an L1 penalty, we obtain a penalized sample version of (3):
1
b i k22 + µi kβk1
k XX τ β − η
n
(4)
for an appropriate choice of parameter µi .
Algorithm 2 Matrix Lasso
b ;
b 1 , ..., η
b d be the eienvectors associated with the largest d eigenvalues of Λ
1: Let η
H
b
2: For 1 ≤ i ≤ d, let β i be the minimizer of equation (4) for an appropriate choice of µi ;
b , ..., β
b ).
3: Estimate the central space col(B) by col(β
1
d
This simple procedure can be easily implemented to produce sparse estimates of βi ’s.
Empirically it works reasonably well, so we set it as another benchmark to compare with.
Since we later observed that its numerical performance was consistently worse than that of
our main algorithm, Lasso-SIR, we did not further investigate its theoretical properties.
6
Lin, Zhao, and Liu
2.3. Lasso-SIR
Let us first consider the single index model
y = f (xτ β0 , ǫ).
(5)
Without loss of generality, we assume that (xi , yi ), i = 1, . . . , n, are sorted i.i.d. samples
such that y1 ≤ y2 ≤ · · · ≤ yn . Construct an n × H matrix M = I H ⊗ 1c , where 1c is
the c × 1 vector with all entries being 1. Then, according to the definition of X H , we can
b = 1 X X τ and let η
b be the largest eigenvalue of Λ
b be the
write X H = XM /c. Let λ
H
H
H
H
corresponding eigenvector of length 1. That is,
b=
λ̂η
Thus, by defining
1 τ
1 τ
b=
b.
XH XH η
X M M τ Xη
H
nc
e=
y
1
b
cλ
b
M Mτ Xη
(6)
e . Note that a key in estimating the central space col(β) of SIR is the
we have η̂ = n1 X y
b and n1 XX τ respectively, this equation
equation η ∝ Σβ. If approximating η and Σ by η
τ
1
1
b ∝ β, one can consider the
e ∝ n XX β. To recover a sparse vector β
can be written as n X y
following optimization problem
min ||β||1 ,
e − X τ β)||∞ ≤ µ,
subject to ||X(y
which is known as the Dantzig selector [Candes and Tao, 2007]. A related formulation is the
Lasso regression, where β is estimated by the minimizer of
Lβ =
1
e − X τ βk22 + µkβk1 .
ky
2n
(7)
As shown by Bickel et al. [2009], the Dantzig Selector is asymptotically equivalent to the
Lasso for linear regressions. We thus propose and study the Lasso-SIR algorithm in this
paper and leave the discussion of the Dantzig selector in a future work.
Algorithm 3 Lasso-SIR (for single index models)
b , respectively;
b and η
b be the first eigenvalue and eigenvector of Λ
1: Let λ
H
e=
2: Let y
1
b
M M τ Xη
cb
λ
and solve the Lasso optimization problem
b
β(µ)
= arg min Lβ , where Lβ =
3: Estimate Pβ by Pβ
b(µ) .
1
e − X τ βk22 + µkβk1 ;
ky
2n
Clearly, we do not need to estimate the inverse of Σ in Lasso-SIR. Moreover, since
the optimization problem (7) is well studied for linear regression models [Tibshirani, 1996,
Sparse Sliced Inverse Regression for High Dimensional Data
7
Efron et al., 2004, Friedman et al., 2010], we may formally “transplant” their results to the
index models. Specifically, we use the R package glmnet to solve the optimization problem
where the tuning parameter µ is chosen based on cross-validation.
Last but not least, Lasso-SIR can be easily generalized to the multiple index model (1).
b and η
b , 1 ≤ i ≤ d, be the d-top eigenvalues of Λ
b = (η
b 1, · · · , η
b d ) be the corresponding
Let λ
i
H
eigenvectors. Similar to the definition of the “pseudo response variable” for the single index
f as
model, we define a multivariate pseudo response Y
1
1
1
f
b diag( , · · · ,
).
Y = M M τ Xη
b
b
c
λ1
λd
(8)
We then apply the Lasso on each column of the pseudo response matrix to produce the
b ′ s.
corresponding β
i
Algorithm 4 Lasso-SIR (for multiple index model)
b respectively.
b and η
b i , i = 1, · · · , d be the top d eigenvalues and eigenvectors of Λ
1: Let λ
i
H
f = 1 M M τ Xη
b diag(
2: Let Y
c
tion problem
1
b
λ1
, · · · , b1 ) and for each 1 ≤ i ≤ d solve the Lasso optimizaλd
b = arg min L
β
β,i where Lβ,i =
i
1 f
kY ∗,i − X τ βk22 + µkβk1 .
2n
c be the matrix formed by β
b ,··· ,β
b . The estimate of PB is given by P .
3: Let B
1
d
b
B
The number of directions d plays an important role when implementing Algorithm 4. A
common practice is to locate the maximum gap among the ordered eigenvalues of the matrix
b , which does not work well under HDLSS settings. In Section 3, we show that there exists
Λ
H
a gap among the adjusted eigenvalues λ̂ai = λ̂i ||β̂i ||2 where β̂i is the i-th output of Algorithm
4. Motivated by this, we estimate d according to the following algorithm:
Algorithm 5 Estimation of the number of directions d
1: Apply Algorithm 4 by setting d = H;
2: For each i, calculate λ̂ai = λ̂i ||β̂i ||2 ;
3: Apply the k-means method on λ̂ai with k being 2 and the total number of points in the
cluster with larger λ̂a is the estimated value of d.
3.
Consistency of Lasso-SIR
For simplicity, we assume that x ∼ N (0, Σ). The normality assumption can be relaxed to
elliptically symmetric distributions with sub-Gaussian tail; however, this will make technical
arguments unnecessarily tedious and is not the main focus of this paper. From now on, we
assume that d, the dimension of the central space, is bounded; thus we can assume that H,
8
Lin, Zhao, and Liu
the number of slices, is also finite [Lin et al., 2016, 2015]. In order to prove the consistency,
we need the following technical conditions:
A1) There exist two positive constants Cmin < Cmax , such that Cmin < λmin (Σ) ≤ λmax (Σ) <
Cmax ;
A2) There exists a constant κ ≥ 1, such that 0 < λ = λd (var(E[x|y]) ≤ ... ≤ λ1 (var(E[x|y]) ≤
κλ ≤ λmax (Σ);
A3) The central curve m(y) = E[x|y] satisfies the sliced stability condition.
Condition A1 is quite mild and standard. Condition A2 is merely a refinement of the coverage condition, rank(var(E[x|y])) = d, a common assumption in the SIR literature. A3
is a condition on the central curve, or equivalently, a regularity condition on the link funcb γ converges to var(γ τ E[x|y]) for any
tion f (·) and the noise ǫ. It guarantees that γ τ Λ
H
γ ∈ col(var(E[x|y])). These conditions play indispensable roles in establishing the phase
transition phenomenon of SIR [Lin et al., 2015] and the minimax rate of SIR over a large
class of models [Lin et al., 2016]. As discussed in Lin et al. [2016] and Neykov et al. [2016],
Condition A3 can be derived from a slightly stronger version of the commonly used condition introduced by Hsing and Carroll [1992]. To improve the readability and avoid duplicating existing results, we leave further discussion of Condition A3 in the online supplementary file. For single index model, there is a more intuitive explanation of the condition
A2. Since rank(var(E[x|y])) = 1, condition A2 is simplified to 0 < λ = λ1 ≤ λmax (Σ)
which is a direct corollary of the total variance decomposition identity ( i.e., var(x) =
var(E[x|y]) + E[var(x|y)]). We may treat λ as a generalized SN R and A2 simply requires
that the generalized SNR is non-zero.
Remark 1 (Generalized SN R and eigenvalue bound). Recall that the the signalto-noise ratio (SN R) for the linear model y = βτ x + ǫ, where x ∼ N (0, Σ) and ǫ ∼ N (0, 1),
is defined as
SN R =
kβk22 βτ0 Σβ0
E[(βτ x)2 ]
.
=
E[y 2 ]
1 + kβk22 βτ0 Σβ 0
where β0 = β/kβk2 . A simple calculation shows that
var(E[x|y]) =
Σββτ Σ
β τ0 Σβ0 kβk22
+1
,
and λ(var(E[x|y])) =
β τ0 ΣΣβ0 kβk22
,
βτ0 Σβ0 kβk22 + 1
where λ(var(E[x|y])) is the unique non-zero eigenvalue of var(E[x|y]). This leads to the
identity
βτ ΣΣβ 0
λ(var(E[x|y])) = 0 τ
SN R.
β0 Σβ 0
Thus, we call λ, the smallest non-zero eigenvalue of var(E[x|y]), the generalized SN R for
multiple index models.
Sparse Sliced Inverse Regression for High Dimensional Data
9
b be the
Theorem 1 (Consistency of Lasso-SIR for Single Index Models). Let β
α
output of Algorithm 3. Assume that nλ = p for some α > 1/2 and that conditions A1-A3
hold for the single index model, y = f (βτ0 x, ǫ), where β 0 is a unit vector, then
kPβb − Pβ0 kF ≤ C1
s
s log(p)
nλ
holds with probability at least 1 − C2 exp(−C3 log(p)) for some constants C2 and C3 .
The technical assumption nλ = pα for some α > 1/2 could be replaced by some sparsity
conditions on the covariance matrix Σ. Since our focus here is to introduce the Lasso-SIR
algorithm, we leave this extension in our future study. Next, we state the theoretical result
regarding the multiple index model (1).
b be the output of Algorithm 4 where
Theorem 2 (Consistency of Lasso-SIR). Let B
d, the dimension of central subspace, is known. Assume that that nλ = pα for some α > 1/2
(where λ is the smallest nonzero eigenvalue of var(E[x|y]) ), and that conditions A1-A3 hold
for the multiple index model (1), then
kPB
b − PB kF ≤ C1
s
s log(p)
.
nλ
holds with probability at least 1 − C2 exp(−C3 log(p)) for some constants C2 and C3 .
s log(p/s)
2
Lin et al. [2016] have speculated that the lower bound of the risk EkPB
b − PB kF is
nλ
when the dimension of the central space is bounded. This implies that if s = O(p1−δ ) for
some positive constant δ, the Lasso-SIR algorithm achieves the optimal rate, i.e., we have
the following corollary.
Corollary 1. Assume that conditions A1-A3 hold for the multiple index model (1) and
d, the dimension of central subspace, is bounded. If nλ = pα for some α > 1/2 and s =
O(p1−δ ), the Lasso-SIR estimate PB
b achieves the minimax rate.
Remark 2 (Comparison with Lasso regression). Consider the linear regression y =
β τ x + ǫ, where x ∼ N (0, Σ) and ǫ ∼ N (0, 1). It is shown in Raskutti et al. [2011] that the
lower bound of the minimax rate of the l2 distance between any estimator and the true β
s log(p)
b
is s log(p/s)
and the convergence rate of Lasso estimator β
. Namely, the Lasso
Lasso is
n
n
1−δ
estimator is rate optimal for linear regression when s = O(p ) for some positive constant
δ. A simple calculation shows that λ(var(E[x|y])) ∼ kβk22 , if kβk is bounded away from ∞.
Consequently,
b
kβ
− βk2
≤C
kPβb
− Pβ kF ≤ 4 Lasso
Lasso
kβk2
s
s log(p)
nλ(var(E[x|y]))
(9)
holds with high probability. In other words, the projection matrix Pβb
based on Lasso is
Lasso
rate optimal. From this view point, the Lasso-SIR has extended the Lasso to the non-linear
10
Lin, Zhao, and Liu
multiple index models. This justifies a suggestion in Chen and Li [1998], stating that SIR
should be viewed as an alternative or generalization of the multiple linear regression. The
connection also justifies a speculation in Lin et al. [2016] that a more appropriate prototype
of the high dimensional SIR problem should be the sparse linear regression rather than the
sparse PCA and the generalized eigenvector problem.
Determining the dimension d of the central space is a challenge problem for SDR, especially
for HDLSS cases. If we want to discern signals (i.e., the true directions) from noises (i.e.,
b of Λ
b , i = 1, ..., H, we face the problem
the other directions) simply via the eigenvalues λ
i
H
b
that all these λi ’s are of order p/n, but the gap between the signals and noises is of order λ
(≤ Cmax ). With the Lasso-SIR, we can bypass this difficulty by using the adjusted eigenvalues
b kβ
b k , i = 1, ..., H. To this end, we have the following theorem.
λ̂ai = λ
i
i 2
b be the output of Algorithm 4 for i = 1, ..., H. Assume that nλ = pα
Theorem 3. Let β
i
for some α > 1/2, s log(p) = o(nλ), and H > d, then, for some constants C1 , C2 C3 and C4 ,
√
λ̂ai ≥ C1 λ − C2
λ̂ai ≤ C3
p
s
p log(p) √
λ + C4
nλ
s log(p)
, for 1 ≤ i ≤ d, and
n
s
s log(p)
, for d + 1 ≤ i ≤ H,
n
hold with probability at least 1 − C5 exp(−C6 log(p)) for some constants C5 and C6 .
In the sparse linear regression literature, the region s2 = o(p) is often referred to as the “highly
sparse” region [Ingster et al., 2010]. Theorem 3 shows that there is a clear gap between the
signal (i.e., the first d adjusted eigenvalues) and noise (i.e., the other adjusted eigenvalues)
in the highly sparse region if p1/2 = o(nλ).
4.
Simulation Stuides
4.1. Single Index Models
Let β be the vector of coefficients and let S be the active set; namely, βi = 0, ∀i ∈ S c .
Furthermore, for each i ∈ S, we simulated independently βi ∼ N (0, 1). Let x be the design
matrix with each row following N (0, Σ). We consider two types of covariance matrices: (i)
Σ = (σij ) where σii = 1 and σij = ρ|i−j| ; and (ii) σii = 1, σi,j = ρ when i, j ∈ S or i, j ∈ S c ,
and σi,j = 0.1 when i ∈ S, j ∈ S c or vice versa. The first one represents a covariance
matrix which is essentially sparse and we choose ρ among 0, 0.3, 0.5, and 0.8. The second
one represents a dense covariance matrix with ρ chosen as 0.2. In all the simulations, we set
n = 1, 000 and let p vary among 100, 1,000, 2,000, and 4,000. For all the settings, the random
error ǫ follows N (0, I n ). For single index models, we consider the following model settings:
I. y = xβ + ǫ where S = {1, 2, · · · , 10};
II. y = (xβ)3 /2 + ǫ where S = {1, 2, · · · , 20};
III. y = sin(xβ) ∗ exp(xβ) + ǫ where S = {1, 2, · · · , 10};
Sparse Sliced Inverse Regression for High Dimensional Data
Table 1. Estimation error for the first type covariance matrix with ρ = 0.5.
p
Lasso-SIR DT-SIR Lasso M-Lasso Lasso-SIR(Known d)
100
0.12
0.47
0.11
0.19
0.12
1000
0.18
0.65
0.15
0.26
0.18
I
2000
0.20
0.74
0.16
0.30
0.20
4000
0.23
0.90
0.18
0.39
0.23
100
0.07
0.60
0.23
0.27
0.07
1000
0.12
0.78
0.31
0.17
0.12
II
2000
0.15
0.86
0.34
0.20
0.15
4000
0.20
0.99
0.37
0.28
0.19
100
0.21
0.55
1.25
0.26
0.21
1000
0.28
0.74
1.32
0.51
0.27
III
2000
0.35
0.87
1.34
0.66
0.31
4000
0.46
1.00
1.33
0.83
0.39
100
0.46
0.92
0.78
0.58
0.45
1000
0.62
1.07
0.87
0.78
0.58
IV
2000
0.71
1.22
0.89
0.94
0.59
4000
0.71
1.30
0.91
1.00
0.63
100
0.12
0.37
0.42
0.15
0.12
1000
0.20
0.55
0.55
0.41
0.20
V
2000
0.38
0.80
0.60
0.67
0.29
4000
0.78
1.22
0.77
1.06
0.48
11
dˆ
1
1
1
1
1
1
1
1
1
1
1.1
1.1
1
1.1
1.3
1.2
1
1
1.2
1.5
IV. y = exp(xβ/10) + ǫ where S = {1, 2, · · · , 50};
V. y = exp(xβ + ǫ) where S = {1, 2, · · · , 7}.
The goal is to estimate col(β), the space spanned by β. As in Lin et al. [2015], the
\ col(β)), where D(M, N ), the distance between two
estimation error is defined as D(col(β),
p
subspaces M, N ⊂ R , is defined as the Frobenius norm of PM − PN where PM and PN are
the projection matrices associated with these two spaces. The methods we compared with
are DT-SIR, matrix Lasso (M-Lasso), and Lasso. The number of slices H is chosen as 20 in
all simulation studies. The number of directions d is chosen according to Algorithm 5. Note
that both benchmarks (i.e., DT-SIR and M-Lasso) require the knowledge of d as well. To
be fair, we use the d estimated based on Algorithm 5 for both benchmarks. For comparison,
we have also included the estimation error of Lasso-SIR assuming d is known. For each p, n,
and ρ, we replicate the above steps 100 times to calculate the average estimation error for
each setting. We tabulated the results for the first type of covariance matrix with ρ = 0.5 in
Table 1 and put the results for other settings in Tables C1-C4 in the online supplementary
file. The average of estimated directions dˆ is reported in the last column of these tables.
The simulations show that Lasso-SIR outperformed both DT-SIR and M-Lasso under all
settings. The performance of DT-SIR has become worse when the dependence is stronger
and denser. The reason is that this method is based on the diagonal threshold and is only
supposed to work well for the diagonal covariance matrix. Overall, Algorithm 5 provided a
reasonable good estimate of d especially for moderate covariance matrix. When assuming d
is known, the performances of both DT-SIR and M-Lasso are inferior to Lasso-SIR, and are
12
Lin, Zhao, and Liu
Table 2. Estimation error for the first type covariance matrix with ρ = 0.5.
p
Lasso-SIR DT-SIR M-Lasso Lasso-SIR(Known d)
100
0.26
0.57
0.31
0.26
1000
0.33
0.74
0.62
0.33
VI
2000
0.36
0.92
0.73
0.38
4000
0.44
1.12
0.87
0.42
100
0.32
0.67
0.42
0.32
1000
0.60
0.93
1.02
0.66
VII
2000
0.95
1.18
1.35
0.83
4000
1.17
1.43
1.47
1.08
100
0.29
0.61
0.34
0.25
1000
0.37
0.82
0.69
0.35
VIII
2000
0.54
1.00
0.92
0.47
4000
0.88
1.37
1.27
0.71
100
0.43
0.74
0.48
0.43
1000
0.47
0.91
0.91
0.48
IX
2000
0.58
1.11
1.12
0.50
4000
0.57
1.25
1.23
0.56
dˆ
2
2
2
2
2
2.1
2.3
2.1
2
2
2.2
2.5
2
2
2.1
2
thus not reported.
Under Setting I when the true model is linear, Lasso performed the best among all the
methods, as expected. However, the difference between Lasso and Lasso-SIR is not significant,
implying that Lasso-SIR does not sacrifice much efficiency without the knowledge of the
underlying linearity. On the other hand, when the models are not linear (Case II-VI), LassoSIR worked much better than Lasso. It is seen that Lasso works better than Lasso-SIR
for Setting V when ρ is large or the covariance matrix is dense. One explanation is that
Lasso-SIR tends to overestimate d under these conditions while Lasso used the actual d. If
assuming d = 1 in Lasso-SIR, its estimation error is smaller than that of Lasso.
4.2. Multiple Index Models
Let β be the p × 2 matrix of coefficients and S be the active set. Let x be simulated similarly
as in Section 4.1, and denote z = xβ. Consider the following settings:
VI. yi = |zi2 /4 + 2|3 ∗ sgn(zi1 ) + ǫi where S = {1, 2, · · · , 7} and β1:4,1 = 1, β5:7,2 = 1,
βi,j = 0 otherwise;
VII. yi = zi1 ∗ exp(zi2 ) + ǫi where S = {1, 2, · · · , 12} and β1:7,1 , β8:12,2 ∼ N (0, 1), and βi,j
otherwise;
VIII. yi = zi1 ∗ exp(zi2 + ǫi ) where S = {1, 2, · · · , 12} and β1:7,1 , β8:12,2 ∼ N (0, 1), and βi,j
otherwise;
IX. yi = zi1 ∗(2+zi2 /3)2 +ǫi where S = {1, 2, · · · , 12} and β1:8,1 = 1, β9:12,2 = 1 and βi,j
otherwise.
and
=0
=0
=0
For the multiple index models, we compared both benchmarks (DT-SIR and M-Lasso)
with Lasso-SIR. Lasso is not applicable for these cases and is thus not included. Similar to
Sparse Sliced Inverse Regression for High Dimensional Data
13
Section 4.1, we tabulated the results for the first type covariance matrix with ρ = 0.5 in Table
2 and put the results for others in Tables C5-C8 in the online supplementary file.
For the identity covariance matrix (ρ = 0), there was little difference between performances
of Lasso-SIR and DT-SIR. However, Lasso-SIR was substantially better than DT-SIR in
other cases. Under all settings, Lasso-SIR worked much better than the matrix Lasso. For
the dense covariance matrix Σ2 , Algorithm 5 tended to underestimate d, which is worthy of
further investigation.
4.3. Discrete Response
We consider the following simulation settings where for the response variable Y is discrete.
X.
XI.
XII.
XIII.
y = 1(xβ + ǫ > 0) where S = {1, 2, · · · , 10};
y = 1(exp(xβ) + ǫ > 0) where S = {1, 2, · · · , 7};
y = 1((xβ)3 /2 + ǫ) where S = {1, 2, · · · , 20};
Let z = xβ where S = {1, 2, · · · , 12}, β is a p by 2 matrix with β1:7,1 , β8:12,2 ∼ N (0, 1)
and βi,j = 0 otherwise. The response yi is
yi =
where ǫij ∼ N (0, 1).
1,
2,
3,
if zi1 + ǫi1 < 0,
if zi1 + ǫi1 > 0 and zi2 + ǫi2 < 0,
if zi1 + ǫi1 > 0 and zi2 + ǫi2 > 0,
In settings X, XI, and XII, the response variable is dichotomous, and βi ∼ N (0, 1) when
i ∈ S and βi = 0 otherwise. Thus the number of slices H can only be 2. For Setting XIII
where the response variable is trichotomous, the number of slices H is chosen as 3. The
number of direction d is chosen as H − 1 in all these simulations.
Similar to the previous two sections, we calculated the average estimation errors for LassoSIR (Algorithm 4), DT-SIR, M-Lasso, and generalized-Lasso based on 100 replications and
reported the result in Table 3 for the first type covariance matrix with ρ = 0.5 and the
results for other cases in Tables C9-C12 in online supplementary file. It is clearly seen that
Lasso-SIR performed much better than DT-SIR and M-Lasso under all settings and the
improvements were very significant. The generalized Lasso performed as good as Lasso-SIR
for the dichotomous response; however, it performed substantially worse for Setting XIII.
5.
Applications to Real Data
5.1. Arcene Data Set
We first apply the methods to a two-class classification problem, which aims to distinguish
between cancer patients and normal subjects from using their mass-spectrometric measurements. The data were obtained by the National Cancer Institute (NCI) and the Eastern
Virginia Medical School (EVMS) using the SELDI technique, including samples from 44 patients with ovarian and prostate cancers and 56 normal controls. The dataset was downloaded
from the UCI machine learning repository (Lichman [2013]), where a detailed description can
be found. It has also been used in the NIPS 2003 feature selection challenge (Guyon et al.
14
Lin, Zhao, and Liu
Table 3. Estimation error for the first type covariance matrix with ρ = 0.5.
p
Lasso-SIR DT-SIR M-Lasso Lasso
100
0.21
0.58
0.26
0.21
1000
0.26
1.21
0.51
0.28
X
2000
0.27
1.33
0.59
0.29
4000
0.28
1.39
0.65
0.31
100
0.36
0.85
0.57
0.34
1000
0.44
1.32
1.07
0.45
XI
2000
0.46
1.38
1.16
0.48
4000
0.52
1.41
1.20
0.52
100
0.24
0.63
0.49
0.22
1000
0.34
1.18
0.53
0.32
XII
2000
0.36
1.29
0.61
0.35
4000
0.40
1.37
0.68
0.39
100
0.38
1.08
0.60
1.06
1000
0.38
1.79
1.13
1.08
XIII
2000
0.40
1.92
1.24
1.09
4000
0.41
1.98
1.31
1.10
[2004]). For each subject, there are 10,000 features where 7,000 of them are real variables
and 3,000 of them are random probes. There are 100 subjects in the validation set.
After standardizing X, we estimated the number of directions d as 1 using Algorithm 5.
We then applied Algorithm 3 and the sparse PCA to calculate the direction of β and the
corresponding components, followed by a logistic regression model. We applied the fitted
model to the validation set and calculated the probability of each subject being a cancer
patient. We also fitted a Lasso logistic regression model to the training set and applied it to
the validation set to calculate the corresponding probabilities.
In Figure 1, we plot the Receiver Operating Characteristic (ROC) curves for various methods. Lasso-SIR, represented by the red curve, was slightly better than Lasso (insignificant)
and the sparse PCA, represented by the green and blue curves respectively. The areas under
these three curves are 0.754, 0.742, and 0.671, respectively.
5.2. HapMap
In this section, we analyzed a data set with a continuous response. We consider the gene
expression data from 45 Japanese and 45 Chinese from the international “HapMap” project
(Thorisson et al. [2005], Thorgeirsson et al. [2010]). The total number of probes is 47,293.
According to Thorgeirsson et al. [2010], the gene CHRNA6 is the subject of many nicotine
addiction studies. Similar to Fan et al. [2015], we treat the mRNA expression of CHRNA6 as
the response Y and expressions of other genes as the covariates. Consequently, the number
of dimension p is 47,292, much greater than the number of subjects n=90.
We first applied Lasso-SIR to the data set with d being chosen as 1 according to Algorithm
5. The number of selected variables was 13. Based on the estimated coefficients β and X,
we calculated the first component and the scatter plot between the response Y and this
component, showing a moderate linear relationship between them. We then fitted a linear
15
0.4
Sensitivity
0.6
0.8
1.0
1.2
Sparse Sliced Inverse Regression for High Dimensional Data
−0.2
0.0
0.2
LASSO−SIR
Lasso
Sparse PCA
1.0
0.8
0.6
0.4
0.2
0.0
Specificity
Fig. 1. ROC curve of various methods for Arcene Data set.
regression between them. The R-sq of this model is 0.5596 and the mean squared error of
the fitted model 0.045.
We also applied Lasso to estimate the direction β. The tuning parameter λ is chosen
as 0.1215 such that the number of selected variables is also 13. When fitting a regression
model between Y and the component based on the estimated β, the R-sq is 0.5782 and the
mean squared error is 0.044. There is no significant difference between these two approaches.
This confirms the message that Lasso-SIR performs as good as Lasso when the linearity
assumption is appropriate.
We have also calculated a direction and the corresponding components based on the sparse
PCA (Zou et al. [2006]). We then fitted a regression model. The R-sq is only 0.1013 and the
mean squared error is 0.093, significantly worse than the above two approaches.
5.3. Classify Wine Cultivars
We finally investigate the popular wine data set that people have used to compare various
classification methods. This is a three-class classification problem. The data, available from
the UCI machine learning repository (Lichman [2013]), consists of 178 wines grown in the
same region in Italy under three different cultivars. For each wine, the chemical analysis
was conducted and the quantities of 13 constituents are obtained, which are Alcohol, Malic
acid, Ash, Alkalinity of ash, Magnesium, Total Phenols, Flavanoids, Nonflavanoid Phenols,
Proanthocyanins, Color intensity, Hue, OD280/OD315 of diluted wines, and Proline. One of
the goals is to use these 13 features to classify the cultivar.
The number of directions d is chosen as 2 according to Algorithm 5. We tested PCA,
DT-SIR, M-Lasso, and Lasso-SIR, for obtaining these two directions. In Figure 2, we plotted
the projection of the data onto the space spanned by two components. The colors of the
points correspond to three different cultivars. It is clearly seen that Lasso-SIR provided the
best separation of the three cultivars. When using one vertical and one horizontal line to
classify three groups, only one subject would be wrongly classified.
16
Lin, Zhao, and Liu
PCA
Bench I
2
2
1
C2
C2
0
0
−2
−1
−2
−4
−2.5
0.0
2.5
−1
0
C1
C1
Bench II
LASSO−SIR
1
2
2
C2
C2
0
0
−2
−2
−3
−2
−1
0
C1
1
2
−2
−1
0
1
2
3
C1
Fig. 2. We plotted the second component versus the first component for all the wines, which are labeled
with different colors, representing different cultivars (1–red, 2–green, 3–blue). The four methods for
calculating the directions are PCA, DT-SIR, M-Lasso, and Lasso-SIR from top-left to bottom-right. It is
clearly seen that Lasso-SIR offered the best separation among these three groups.
Sparse Sliced Inverse Regression for High Dimensional Data
6.
17
Discussion
In this paper, we have proposed Lasso-SIR, an efficient high-dimensional variant of the sliced
inverse regression [Li, 1991], to obtain a sparse solution to the estimation of the sufficient
dimensional reduction space for multiple index models, and showed that it is rate optimal if
nλ = pα for some α > 1/2, where λ is the smallest nonzero eigenvalue of var(E[x|y]). This
technical assumption on n, λ and p is slightly disappointing from the ultra-high dimensional
perspective. We believe that this technical assumption arises from an intrinsic limitation
in estimating the central subspace, i.e., some further sparsity assumptions on either Σ or
var(E[x|y]) or both are needed to show the consistency of any estimation method. We will
address such extensions in our future researches.
Cautious reader may find that the concept of “pseudo-response variable” is not essential
for developing the theory of the Lasso-SIR algorithm. However, by re-formulating the SIR
method as a linear regression problem using the pseudo-response variable, we can formally
consider the model selection consistency, regularization path and many others for the LassoSIR. In other words, the Lasso-SIR does not only provide an efficient high dimensional variant
of SIR, but also extends the rich theory developed for Lasso linear regression in the past
decades to the multiple index models.
Acknowledgments
This research was partly supported by NSF Grant IIS-1633283, NSF Grant DMS-1120368,
and NIH Grant R01 GM113242-01.
Appendix: Sktech of Proofs of Theorems 1 and 2
We need two propositions to prove Theorems 1 and 2. The first one reveals some properties
b .
of the eigenvectors of Λ
H
b = 1 X Xτ
Proposition 1. Assume that Conditions A1), A2) and A3) hold. Let Λ
H
H
H
H
b j be the eigenvector of unit length
be the SIR estimate of Λ = var(E[x|y]) using (2). Let η
b j = 1, ..., H. If nλ = pα for some α > 1/2, there
b of Λ,
associated with the j-th eigenvalue λ
j
exist positive constants C1 and C2 such that
b j k2 ≥ C1
kPΛ η
s
λ
bj
λ
b j k2 ≤ C2
, 1 ≤ j ≤ d and kPΛ η
p
p log(p)
nλ
s
λ
bj
λ
,d + 1 ≤ j ≤ H
(10)
hold with high probability. Here we say that an event Ω happens with high probability if
P(Ω) ≥ 1 − C3 exp(−C4 log(p)) for some constants C3 and C4 .
Remark: This result might be of independent interest. In order to justify that the sparsity
assumption for the high dimensional setting is necessary, Lin et al. [2015] have shown
q that
p
b 1 )] = 0 if and only if lim nλ
= 0. Proposition 1 states that the projection of
E[∠(η1 , η
j = 1, ..., d, onto the true direction is non-zero if nλ > p1/2 .
b η
λ
j bj,
18
Lin, Zhao, and Liu
The proof of Proposition 1 is technical. To improve the readability, we sketch the key
steps here and put the detailed proof in the online supplementary file. Let x = z + w be the
orthogonal decomposition with respect to col(var(E[x|y])) and its orthogonal complement.
We define two p×H matrices Z H = (z 1,· , ..., z H,· ) and W H = (w1,· , ..., wH,· ) whose definition
are similar to the definition of X H . We then have the following decomposition
XH = ZH + W H .
(11)
1/2
By definition, we know that Z H ⊥
⊥ W H , Z τH W H = 0, and W H = √1c Σ1 E H , where Σ1 is
the covariance matrix of w and E H is a p × H matrix with i.i.d. standard normal entries.
We have the following lemma:
Lemma 1. Assume that ν1 is a large enough constant only depends on the constant κ.
For sufficiently large constant a, the event Ω = Ω1 ∩ Ω2 holds with high probability where
n
o
κ
1
1
1
)λ ≤ λmin ( Z τH Z H ) ≤ λmax ( Z τH Z H ) ≤ (1 +
)κλ
2ν1
H
H
2ν1
p
o
tr(Σ
)
p
log(p)
1
1
Wτ WH −
IH
≤a
.
H H
n
n
F
Ω1 = ω | (1 −
n
Ω2 = ω |
(12)
(13)
Proof. The proof is presented in the online supplementary file.
✷
For any ω ∈ Ω, we can choose a p × p orthogonal matrix T and an H × H orthogonal
matrix S such that
B1 0
0
0
1
1
0 0
B2 B3
T Z H (ω)S =
and T W H (ω)S =
H
H
0 B4
0 0
(14)
where B1 is a d × d matrix, B2 is a d × d matrix, B3 is a d × (H − d) matrix and B4 is a
(p − 2d) × (H − d) matrix. By definition of the event ω, we have
1
κ
)λ ≤ λmin (B1τ B1 ) ≤ λmax (B1τ B1 ) ≤ (1 +
)κλ
2ν1
2ν1
p
τ
tr(Σ1 )
p log(p)
B2τ B3
B2 B2
−
IH
≤ a′
τ
τ
τ
B3 B2 B3 B3 + B4 B4
n
n
F
(1 −
for some constant a′ . Thus, we only need to prove the following lemma:
(15)
B1 0
α
B2 B3 be a p × H
Lemma 2. Assume that nλ = p for some α > 1/2. Let M =
0 B4
matrix, where B1 is a d × d matrix, B2 is a d × d matrix, B3 is a d × (H − d) matrix and
b j be the eigenvector associated with
B4 is a (p − 2d) × (H − d) matrix satisfying (15). Let η
τ
b
b onto
the j-th eigenvalue λj of M M , j = 1, ..., H. Then the length of the projection of η
√
r
rj
p log(p)
λ
its first d-coordinates is at least C bλ for j = 1, ..., d and is at most C nλ
for
b
λj
j = d + 1, ..., H.
λj
Sparse Sliced Inverse Regression for High Dimensional Data
19
Proof. See the online supplementary file.
Next, we briefly review the restricted eigenvalue (RE) property, which was first introduced
in Raskutti et al. [2010], before stating the second proposition. Given a set S ⊂ [p] =
{1, ..., p}, define the set C(S, α) as
C(S, α) = {θ ∈ Rp | kθS c k1 ≤ αkθS k1 }.
We say that a sample matrix XX τ /n satisfies the restricted eigenvalue condition over S with
parameter (α, κ) ∈ [1, ∞) × (0, ∞) if
1 τ
θ XX τ θ ≥ κ2 kθk22 ,
n
∀θ ∈ C(S, α).
(16)
If (16) holds uniformly for all the subsets S with cardinality s, we say that XX τ /n satisfies
the restricted eigenvalue condition of order s with parameter (α, κ). Similarly, we say that the
covariance matrix Σ satisfies the RE condition over S with parameter (α, κ) if kΣ1/2 θk2 ≥
κkθk for all θ ∈ C(S, α). Additionally, if this condition holds uniformly for all the subsets S
with cardinality s, we say that Σ satisfies the restricted eigenvalue condition of order s with
parameter (α, κ). The following Corollary is borrowed from Raskutti et al. [2010].
Corollary 2. Suppose that Σ satisfies the RE condition of order s with parameter
(α, κ). Let X be the p × n matrix formed by n i.i.d samples from N (0, Σ). For some
universal positive constants a1 , a2 and a3 , if the sample size satisfies
n > a3
(1 + α)2 maxi∈[p] Σii
s log(p),
κ2
then the matrix n1 XX τ satisfies the RE condition of order s with parameter (α, κ8 ) with
probability at least 1 − a1 exp (−a2 n).
It is clear that λ√min (Σ) ≥ Cmin implies that Σ satisfies the RE condition of any order s
with parameter (3, Cmin ). Thus, we have the following proposition.
Proposition 2. Assume that Condition A1) holds. For some universal constants a1 ,
a2 and a3 , if the sample size satisfies that n > a1 s log(p), then the matrix n1 XX τ satisfies
√
the RE condition for any order s with parameter (3, Cmin /8) with probability at least 1 −
a2 exp (−a3 n).
Next, we sketch key points and intermediate results of the proofs of the two theorems and
leave details in the online supplementary files.
Sketch of the proof of Theorem 1. Recall that for single index model y = f (βτ0 x, ǫ)
b associated with the
b the eigenvector of Λ
where β0 is a unit vector, we have denoted by η
H
b Let β
b be the minimizer of
largest eigenvalue λ.
Lβ =
1
e − X τ βk + µkβk1 ,
ky
2n
20
Lin, Zhao, and Liu
e = Σ−1 η
e . Let η 0 = Σβ 0 , η
e = Pη0 η
b and β
e ∝ β0 .
e ∈ Rn such that η
b = n1 X y
where y
b
Since we are interested in the distance between the directions of β and β0 , we consider the
b − β.
e A slight modification of the argument in Bickel et al. [2009] implies
difference δ = β
r
r
that, if we choose µ = C
log(p)
nb
λ
for sufficiently large constant C, we have kδk2 ≤ C1
s log(p)
nb
λ
with high probability. The detailed arguments are put in the online
supplementary file. The
q
−1
λ
e
e
e , imply that C1
Proposition 1, Condition A1) and β = Σ η
≤ kβk
2 ≤ C2 holds with
b
λ
high probability for some constants C1 and C2 . Thus, we have
kPβb − Pβ0 kF = kPβb − Pβe kF ≤ 4
b − βk
e
kβ
2
holds with high probability.
e 2
kβk
e ≤C
= 4kδk2 /kβk
2
s
s log(p)
nλ
(17)
✷
b j be the (unit) eigenvectors associated with the j-th eigenProof of Theorem 2. Let η
b
e = Σ−1 η
e /kβ
e k . Applying the
e j = PΛ η
bj, β
e j and γ j = β
values of ΛH , j = 1, ...., d. Let η
j
j
j 2
argument in Theorem 1 to other eigenvectors gives us that
v
s
u
u s log(p)
s log(p)
b −β
e k ≤ Ct
kβ
and kPβb − Pβe kF ≤ C
j
j 2
b
j
j
nλ
nλj
(18)
e , j = 1, ..., d, are bounded below by
hold with high probability. Thus, if i) the lengths of β
j
C
r
λ
e , j = 1, ..., d, are
with high probability and, ii) the angles between any two vectors of β
j
b
λj
bounded below by some q
constant with high probability, then the Gram-Schmit process implies
that kPB
b − PB kF ≤ C
that d is bounded.
s log(p)
nλ
holds with high probability from (18) and the assumption
We only need to verify i) and ii). i) follows from the Proposition 1 and the Condition A1),
e = Σ−1 η
e j , j = 1, ..., d. ii) can be easily derived from the following two intuitively
since β
j
straightforward statements.
e j ′ s , j = 1, ..., d are nearly π/2. Since
I. The angles between any two vectors in η
α
nλ = p for some α > 1/2, we only need to prove that
ej, η
ej )
cos ∠(η
≤C
p
p log(p)
nλ
(19)
holds with high probability for any i 6= j. Recall that We have the following
√ decomposition
X H = Z H + W H . It is easy to see that col(Z H ) = col(var(E[x|y])) and c cov(w)−1/2 W H
is identically distributed to a matrix, E1 , with all the entries are i.i.d. standard
normal
A
1
and
random variables. Let us choose an orthogonal matrix T such that √H T Z H =
0
0
√1 T W H =
bj
where A is a d × H matrix and B is a (p − d)× H matrix. Thus, T η
H
B
Sparse Sliced Inverse Regression for High Dimensional Data
21
b , j = 1, ..., d. If we
associated with the j-th eigenvalue λ
j
√
r
p log(p)
τ
τ
λ
b
have a) λmin (AA ) ≥ λ, b) kPcol(T Z H ) (T η j )k2 ≥ C b and c) kB B − µI H kF ≤ C
n
is the eigenvector of
τ
1
τ
H T XHXHT
λj
for some scalar µ > 0, then the statement I is reduced to the following linear algebra lemma.
Lemma 3. Let A be a d×H matrix√(d < H) with λmin (AAτ ) = λ. Let B be a (p−d)×H
p log(p)
. Let ξbj be the j-th (unit) eigenvector of CC τ
matrix such that kB τ B − µI H k2F ≤ C
n
b where C τ = (Aτ , B τ ) and ξ
e be the projection of ξ
b
associated with the j-th eigenvalue λ
j
j
j
onto its first d-coordinates. If kξej k2 ≥ C
r
λ
b
λj
cos ∠(ξei , ξej )
, then for any i 6= j,
≤C
p
p log(p)
.
nλ
(20)
Thus, ξei ′ s are nearly orthogonal if nλ = pα for some α > 1/2.
b , then ξ
b = 1 Cα and ξ
e = 1 Aα . It is easy to see that
b j = Cτ ξ
Proof. Let α
j
j
j
j
j
b
b
λj
λj
q
q
b j and kC α
b j . Since α
b j is also the (unit) eigenvector of
b j k2 ≥ λ
b j/ λ
b j k2 = λ
kα
C τ C = Aτ A + µI + (B τ B − µI),
for any i 6= j, we have
bi
bi + α
b τj (B τ B − µI)α
b j + µα
b τj α
bi = α
b τj Aτ Aα
b τj C τ C α
0=α
τ
bj λ
bi ξ
e ξ
e
b i.
b τj (B τ B − µI)α
=λ
j i+α
√
r
p log(p)
b k2 ≥ C λ , ∀i 6= j, we have
and
k
ξ
Since kB τ B − tr(Σ)I H kF ≤ C
j
n
b
λj
p
b 1/2 λ
b 1/2
b τj
λ
ξτj ξi
bi
τ
α
1 α
p log(p)
j
i
τ
b
b
(B B − µI) 1/2 ≤ C
≤ C ξj ξi
.
=C
1/2
b
b
kξi k2 kξj k2
λ
λλ
nλ
λ
j
i
✷
Note that a) follows from the Lemma 3, b) follows from Proposition 1 and c) follows from
the Lemma 8. Thus statement I holds.
e ′ s are bounded away from 0.
II. The angles between any two vectors in β
Since
j
−1 e
e
β j = Σ η j , we only need to prove that there exists a positive constant ζ < 1 such that
ei
e τi Σ−1 Σ−1 η
η
≤ ζ.
−1 e
−1 e
kΣ η i k2 kΣ η j k2
(21)
e 1 /kη
e 1 k2 , ..., η
e d /kη
e d k2 ) = T M , where T is a p × d orthogonal matrix. Since η
e j /kη
e j k′2 s
Let (η
τ
are nearly mutually orthogonal, we know that M M is nearly an identity matrix. Thus, by
some continuity argument, the statement is reduced to the following linear algebra lemma.
22
Lin, Zhao, and Liu
Lemma 4. Let A be a p×p positive definite matrix such that Cmin ≤ λmin (A) ≤ λmax (A) ≤
Cmax for some positive constants Cmin and Cmax . There exists constant 0 < ζ < 1 such that
for any p × d orthogonal matrix B, we have
B τ∗,i Aτ AB ∗,j
≤ζ
kAB ∗,i k2 kAB ∗,j k2
∀i 6= j.
(22)
Proof. When d is finite, without loss of generality, we can assume that B is a p × 2
matrix. Note that the expression on the left side is invariant under orthogonal transformation
of B. We can simply assume that B is a matrix with the last p − 2-rows consisting of all
zeros. The result follows immediately based on basic calculation.
✷
References
P. J. Bickel, Y. Ritov, and A. B. Tsybakov. Simultaneous analysis of Lasso and Dantzig
selector. The Annals of Statistics, 37(4):1705–1732, 2009.
Cristina Butucea, Yuri I Ingster, et al. Detection of a sparse submatrix of a high-dimensional
noisy matrix. Bernoulli, 19(5B):2652–2688, 2013.
E. Candes and T. Tao. The Dantzig selector: Statistical estimation when p is much larger
than n. The Annals of Statistics, 35(6):2313–2351, 2007.
C. H. Chen and K. C. Li. Can SIR be as popular as multiple linear regression? Statistica
Sinica, 8(2):289–316, 1998.
H. Chen. Estimation of a projection-pursuit type regression model. The Annals of Statistics,
19(1):142–157, 1991.
D. R. Cook. Regression graphics. Wiley Series in Probability and Statistics: Probability and
Statistics. John Wiley & Sons, Inc., New York, 1998.
D. R. Cook. SAVE: a method for dimension reduction and graphics in regression. Communications in statistics-Theory and methods, 29(9-10):2109–2121, 2000.
D. R. Cook, L. Forzani, and A. J. Rothman. Estimating sufficient reductions of the predictors
in abundant high-dimensional regressions. The Annals of Statistics, 40(1):353–384, 2012.
N. Duan and K. C. Li. Slicing regression: a link-free regression method. The Annals of
Statistics, 19(2):505–530, 1991.
B. Efron, T. Hastie, I. Johnstone, and R. Tibshirani. Least angle regression. The Annals of
statistics, 32(2):407–499, 2004.
J. Fan and R. Li. Variable selection via nonconcave penalized likelihood and its oracle
properties. Journal of the American statistical Association, 96(456):1348–1360, 2001.
J. Fan, Q. Shao, and W. Zhou. Are discoveries spurious? distributions of maximum spurious
correlations and their applications. arXiv preprint arXiv:1502.04237, 2015.
Sparse Sliced Inverse Regression for High Dimensional Data
23
J. Friedman, T. Hastie, and R. Tibshirani. Regularization paths for generalized linear models
via coordinate descent. Journal of statistical software, 33(1):1, 2010.
J. H. Friedman and W. Stuetzle. Projection pursuit regression. Journal of the American
statistical Association, 76(376):817–823, 1981.
I. Guyon, S. Gunn, A. Ben-Hur, and G. Dror. Result analysis of the NIPS 2003 feature
selection challenge. In Advances in neural information processing systems, pages 545–552,
2004.
T. Hsing and R. J. Carroll. An asymptotic theory for sliced inverse regression. The Annals
of Statistics, 20(2):1040–1061, 1992.
Yuri I Ingster, Alexandre B Tsybakov, Nicolas Verzelen, et al. Detection boundary in sparse
regression. Electronic Journal of Statistics, 4:1476–1526, 2010.
B. Jiang and J. S. Liu. Variable selection for general index models via sliced inverse regression.
The Annals of Statistics, 42(5):1751–1786, 2014.
S. Jung and J. S. Marron. PCA consistency in high dimension, low sample size context. The
Annals of Statistics, 37(6B):4104–4130, 2009.
K. C. Li. Sliced inverse regression for dimension reduction. Journal of the American Statistical
Association, 86(414):316–327, 1991.
L. Li. Sparse sufficient dimension reduction. Biometrika, 94(3):603–613, 2007.
L. Li and C. J. Nachtsheim. Sparse sliced inverse regression. Technometrics, 48(4):503–510,
2006.
M.
Lichman.
UCI
machine
http://archive.ics.uci.edu/ml.
learning
repository,
2013.
URL
Q. Lin, Z. Zhao, and J. S. Liu. On consistency and sparsity for sliced inverse regression in
high dimensions. arXiv preprint arXiv:1507.03895, 2015.
Q. Lin, X. Li, H. Dong, and J. S. Liu. On optimality of sliced inverse regression in high
dimensions. 2016.
M. Neykov, Q. Lin, and J. S. Liu. Signed support recovery for single index models in highdimensions. Annals of Mathematical Sciences and Applications, 1(2):379–426, 2016.
L. Ni, D. R. Cook, and C. L. Tsai. A note on shrinkage sliced inverse regression. Biometrika,
92(1):242–247, 2005.
G. Raskutti, M. J. Wainwright, and B. Yu. Restricted eigenvalue properties for correlated
gaussian designs. The Journal of Machine Learning Research, 11:2241–2259, 2010.
G. Raskutti, M. J. Wainwright, and B. Yu. Minimax rates of estimation for high-dimensional
linear regression over-balls. Information Theory, IEEE Transactions on, 57(10):6976–6994,
2011.
24
Lin, Zhao, and Liu
N. Simon, J. Friedman, T. Hastie, and R. Tibshirani. A sparse-group lasso. Journal of
Computational and Graphical Statistics, 22(2):231–245, 2013.
T. E. Thorgeirsson, D. F. Gudbjartsson, and many others. Sequence variants at CHRNB3CHRNA6 and CYP2A6 affect smoking behavior. Nature genetics, 42(5):448–453, 2010.
G. A. Thorisson, A. V. Smith, L. Krishnan, and L. D. Stein. The international HapMap
project web site. Genome research, 15(11):1592–1593, 2005.
R. Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical
Society. Series B, pages 267–288, 1996.
Y. Xia, H. Tong, W. K. Li, and L. X. Zhu. An adaptive estimation of dimension reduction
space. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 64(3):
363–410, 2002.
M. Yuan and Y. Lin. Model selection and estimation in regression with grouped variables.
Journal of the Royal Statistical Society: Series B (Statistical Methodology), 68(1):49–67,
2006.
W. Zhong, T. Zhang, Y. Zhu, and J. S. Liu. Correlation pursuit: forward stepwise variable
selection for index models. Journal of the Royal Statistical Society: Series B, 74(5):849–870,
2012.
L. Zhu, B. Miao, and H. Peng. On sliced inverse regression with high-dimensional covariates.
Journal of the American Statistical Association, 101(474):640–643, 2006.
Hui Zou, Trevor Hastie, and Robert Tibshirani. Sparse principal component analysis. Journal
of computational and graphical statistics, 15(2):265–286, 2006.
1
Sparse Sliced Inverse Regression for High Dimensional Data
Web-based supporting materials for
“Sparse Sliced Inverse Regression for High Dimensional Data”
by “Lin, Q., Zhao, Z., and Liu, S. J.”
A.
Proofs of Theorems
A.1. Proof of Theorem 1
Let T = supp(β0 ), then |T | ≤ s. For a vector γ ∈ Rp , let γ T and γ T c be the sub-vector
consists of the components of γ in T and T c respectively. Let us introduce the event E =
E1 ∩ E2 ∩ E3 ∩ E4 where
E1 =
n
n
o
ω | kδ T c k1 ≤ 3kδT k1 ,
1
E2 = ω | XX τ satisfies the RE condition of order s with parameter(3,
n
s
n
log(p) o
1
E3 = ω | k XX τ δk∞ ≤ µ where µ = b1
b
n
nλ
n
√
Cmin o
)
8
s
λ o
.
b
λ
Here b1 and b2 are sufficiently large constants to be specified later. Because Proposition 1
implies that E4 happens with high probability( i.e., P (E4 ) ≥ 1 − C1 exp (−C2 log(p)) for some
constants C1 and C2 ), Proposition 2 implies that E2 happens with high probability, Lemma 5
(see below) implies that E3 happens with high probability, and Lemma 7 (see below) implies
that E1 happens with high probability, we conclude that E happens with high probability.
Conditioning on E, we have
√
1
1
1
Cmin kδk22 ≤ kX τ δk2 ≤ kXX τ δk∞ kδk1 ≤ 4µkδ T k1 ≤ 4 sµkδ T k2 .
64
n
n
r
q
√
s log(p)
λ
.
Since
we
have
C
Thus, conditioning on E, we have kδk2 ≤ C256
sµ
=
C
1 b ≤
min
b
E4 =
b k 2 ≥ b2
ω | kPΛ η
nλ
e ≤ C conditioning on E, we know that
kβk
2
2
kPβb − Pβ0 kF = kPβb − Pβe kF ≤ 4
b − βk
e
kβ
2
e
kβk
2
≤C
λ
s
s log(p)
.
nλ
r
Lemma 5. Assume that condition A1), A2) and A3) hold. Let µ = A
ciently large A, we have that
1
k XX τ δk∞ ≤ µ
n
✷
log(p)
.
nb
λ
For suffi-
(A.1)
2
Lin, Zhao, and Liu
holds with high probability.
e =η
b − β,
e η
e and Σβ
e , we have
b = n1 X y
Proof. Since δ = β
1
1
1
1
b − Xy
e
e k∞ + kX ỹ − nη
e k∞ + k( XX τ − Σ)βk
kXX τ δk∞ ≤ kXX τ β
∞.
n
n
n
n
{z
} |
{z
} |
{z
}
|
I
II
b we have 0 ∈
For I. By the definition of β,
e k∞ ≤ µ.
Xy
1
n
(A.2)
III
b , i.e.,
b − Xy
e + µsgn β
XX τ β
1
b
kXX τ β
n
−
For II. Let x = z +w be the orthogonal decomposition with respect to col(var(E[x|y])) and its
orthogonal complement. Recall that we have introduced
the decomposition: X H = Z H +W H .
√
It is easy to see that col(Z H ) = col(η 0 ) and c cov(w)−1/2 W H is identically distributed to
a matrix, E1 , with all the entries are i.i.d. standard normal random variables. Let
Since
1
1
1
b , α2 = √ W τH η
b and α = √ X τH η
b = α1 + α2 .
α1 = √ Z τH η
H
H
H
τ
1
b
H XH XH η
bη
b and η
b , we know that kαk22 = λ
b=
=λ
e = Pη0 η
b=
we know η
1
√
ZH α
Hb
λ
b − Pη0 η
b k∞
we know that kη
A2 (> 1), we have
P k
cov(w)
√
c
1/2
√
1
√
W H α.
Hb
λ
1
Z H α + √ 1 b W H α.
Hb
λ
Hλ
From this,
p
√
b
Since kαk1 ≤ Hkαk2 = H λ,
b −Pη0 η
b=
and η
≤ √1 kW H k∞,∞ . It is easy to see that for positive constant
b
λ
E1 k∞,∞ > λ1/2
max (Σ)
s
A2 H log(pH)
≤ 2 exp−(A2 −1) log(pH) ,
n
r
√
1/2
b − Pη0 η
b k∞ ≤ A
i.e., by letting A > 2λmax (Σ) A2 H, we have that kη
probability.
log(p)
nb
λ
holds with high
For III. Let x = Σ1/2 ǫ where ǫ ∼ N (0, I p ), then X = Σ1/2 E where E is a p × n matrix with
i.i.d standard normal entries. Let
n
o
1 τ
F(t) = ω | kΣ1/2
EE − I p Σ−1/2 η 0 k∞ > tλmax (Σ)1/2 λmin (Σ)−1/2 .
n
We have the following elementary bound on this event.
Lemma 6.
P(F(t)) ≤ 4 exp−
3nt2
16
+log(p)
.
(A.3)
Proof. For any two deterministic vectors α1 and α2 , let
e 1 , α2 ) =
E(α
ω | |α1
1 τ
EE − I p α2 | > tkα1 k2 kα2 k2 .
n
(A.4)
3
Sparse Sliced Inverse Regression for High Dimensional Data
We know that for 0 < t < 1/2,
3
nt2
e 1 , α2 ) ≤ 4 exp− 16
.
P E(α
1/2
(A.5)
Combining with the facts kΣi,∗ k ≤ λmax (Σ)1/2 and kΣ−1/2 η 0 k2 ≤ λmin (Σ)−1/2 , we know
that
1/2
kΣ
1 τ
EE − I p Σ−1/2 η 0 k∞ > tλmax (Σ)1/2 λmin (Σ)−1/2
n
3
happens with probability at most 4 exp− 16 nt
2 +log(p)
(A.6)
.
✷
e = Σ−1 η
e, η
e = kη
e kη 0 and kη
e k ≤ 1, we know that
Since β
1 τ
1
e
e kkΣ1/2
k (XX τ − nΣ) βk
EE − I p Σ−1/2 η 0 k∞ .
∞ = kη
n
n
Conditioning on F(
q
16A3 log(p)
)
3n
(A.7)
∩ E4 , we have
1
e
k (XX τ − nΣ) βk
∞ >C
n
s
λ 16A3 log(p)
λmax (Σ)1/2 λmin (Σ)−1/2 .
b
3n
λ
(A.8)
Combining the Lemma 6 and Proposition 1, we know that (A.8) happens with high probability.
To summarize, we know that, for sufficiently large constant A,
holds with high probability.
r
Lemma 7. Let µ = A
holds with high probability.
log(p)
.
nb
λ
k n1 XX τ δk∞
r
≤A
log(p)
nb
λ
✷
For sufficiently large constant A, we have kδT c k1 ≤ 3kδT k1
Proof. Since Lβb ≤ Lβe , we have
1
1
e ∞ ≤ µ(kβk
e 1 − kβk
b 1 ).
−kδk1 k X ye − XX τ βk
n
n
(A.9)
e − kβk
b ≤ kδ k − kδ c k and kδk = kδ k + kδ c k , if k 1 X ye − 1 XX τ βk
e
Since kβk
1
1
T 1
T 1
1
T 1
T 1
∞ ≤
n
n
A1
r
log(p)
nb
λ
and A > 2A1 , we have kδT c k1 ≤
A+A1
A−A1 kδT k1
≤ 3kδT k1 .
e
e
e − XX τ βk
b−η
e k∞ + k( n1 XX τ − Σ)βk
Since n1 kX y
∞ = II + III where II and
∞ ≤ kη
III are introduced in (A.2).
r From the bounds of II and III in Lemma 5, we know that
e ∞ ≤ A1
k n1 X ye − n1 XX τ βk
log(p)
nb
λ
holds with high probability for some A1 (does not depend
on choice of µ) and some constants C1 and C2 .
✷
4
Lin, Zhao, and Liu
A.2. Proof of Theorem 3
b , η
b i of Λ
b i be the eigenvector associated with the i-th eigenvalue λ
b i and
Let η
H e i = PΛ η
e = Ση
e
β
,
i
=
1,
.
.
.
,
H.
The
argument
in
Theorem
1
implies
that,
for
any
1
≤
i
≤
H,
i
i
The Proposition 1 implies that
e k ≥C
kβ
1
i 2
s
λ
b
λ
i
b −β
e k ≤C
kβ
i
i 2
s
e k ≤C
, 1 ≤ i ≤ d and kβ
2
i 2
s log(p)
s
b
nλ
λ
b
λ
i
.
(A.10)
p
p log(p)
, d + 1 ≤ i ≤ H.
nλ
(A.11)
The above two statements give us the desried result in Theorem 3.
✷
A.3. Proof of Proposition 1
Proof of Lemma 1 By Lemma 10, one of main contributions of Lin et al. [2016], we
know that Ω1 happens with high probability follows. By the following lemma, we know that
Ω2 happens with high probability.
Lemma 8. Let d1 , . . . , dp be positive constants. We have the following statements:
i) For p i.i.d standard normal random variables x1 , . . . , xp , there exist constants C1 and
C2 such that for any sufficiently small a , we have
!
p2 a2
− P 2
1X
d
C
2
di (xi − 1)| > a ≤ C1 exp 2 j j .
P |
p i
(A.12)
ii) For 2p i.i.d standard normal random variables x1 , . . . , xp y1 , · · · , yp , there exist constants
C1 and C2 such that for any sufficiently small a , we have
!
p2 a2
− P 2
1X
d
C
di xi yi | > a ≤ C1 exp 2 j j .
P |
p i
(A.13)
Proof. ii) is a direct corollary of i). And i) followsQfrom a typical deviation argument,
i.e., all we need to do is choosing proper t such that e−pt pj=1 √ 1
is bounded by the RHS
1−2tdj
of (A.12). This could be done by simple Taylor expansion argument.
✷
✷
Proof of Lemma 2 The proof needs the Sin-Theta theroem, which we reviewed below:
Lemma 9. Let A and A + E be symmetric matrices satisfying
A = [F 0 , F 1 ]
A0 0
0 A1
F τ0
F τ1
,
A + E = [G0 , G1 ]
Λ0 0
0 Λ1
Gτ0
Gτ1
Sparse Sliced Inverse Regression for High Dimensional Data
5
where [F 0 , F 1 ] and [G0 , G1 ] are orthogonal matrices. If the eigenvalues of A0 are contained
in an interval (a,b) , and the eigenvalues of Λ1 are excluded from the interval (a − δ, b + δ)
for some δ > 0, then
min(kF τ1 EG0 kF , kF τ0 EG1 kF )
1
√ kF 0 F τ0 − G0 Gτ0 kF ≤
.
δ
2
E1 E2
D1 0
E τ1 E τ3
Let us consider the eigen-decompositions: Q1 , M M =
E3 E4
0 D2
E τ2 E τ4
τ
B1 B1 + pµ
0
F1 0
D ′1 0
F τ1 0
n Id
Q2 ,
=
whereD1 , D ′1
pµ
′
0
0 F2
0 D2
0 F τ2
n I H−d
( resp. D 2 , D′2 ) are d × d (resp. (H − d) × (H − d) ) diagonal matrices. From (15), we know
that
τ
p
p
p log(p)
pµ
p log(p)
pµ
−C
≤ λmin (D1 ) ≤ λmax (D1 ) ≤ κλ +
+C
λ+
n
np
n
p n
pµ
p log(p)
p log(p)
pµ
and
−C
≤ λmin (D2 ) ≤ λmax (D2 ) ≤
+C
.
n
n
n
n
√
p log(p)
Thus the eigengap is of order λ −
( which is of order λ, since nλ = pα for some
n
α > 1/2). √
Let us apply the Sin-Theta theorem on them. From (15) , we know that kQ1 −
Q2 k F ≤ C
p log(p)
.
n
Thus, we have
p
p log(p)
,
nλ
F
√
√
p log(p)
p log(p)
τ
τ
i.e., kE 3 E 3 kF ≤ C nλ . Similar argument gives us kE 2 E 2 kF ≤ C nλ .
E1
E3
E τ1
E τ3
−
Id 0
0 0
≤C
(A.14)
b of M M τ . Let
Let η be the (unit) eigenvector associated with the non-zero eigenvalue λ
η1
α1
us write η = η 2 where η 1 , η 2 ∈ Rd and η 3 ∈ Rp−2d . Let α =
= M τ η, where
α2
η3
α1 = B1τ η 1 + B2τ η 2 ∈ Rd and α2 = B3τ η 2 + B4τ η 3 ∈ RH−d . It is easy to verify that √α is the
b of M τ M and
(unit) eigenvector associted with the eigenvalue λ
B1 α 1
η1 = p p ,
b λ
b
λ
B2 α 1
B3 α 2
B4 α 2
η 2 = p p + p p , and η 3 = p p .
b λ
b
b λ
b
b λ
b
λ
λ
λ
p
b
λ
b is among the first d eigenvalues of M τ M , then kα / λk
b is bounded below by some
If λ
1
2
q
λ
b is among the last H − d eigenvalues of M τ M ,
positive constant. Thus kη 1 k2 ≥ C b . If λ
λ
√
q √
p
p log(p)
p log(p)
λ
b
then kα1 / λk1 = O
. Thus kη 1 k2 ≤ O κ b nλ
.
✷
nλ
λ
6
Lin, Zhao, and Liu
B.
Conditions for SIR to be Consistent
All the results in this appendix are borrowed from Lin et al. [2016] and Lin et al. [2015] with
some minor modifications. We state them here mainly for the purpose of self-content. In order
that SIR gives a consistent estimate of the central space, researchers (for more discussion, see
Li [1991], Hsing and Carroll [1992], Zhu et al. [2006] and Lin et al. [2015]) require the several
technical assumptions,
A′ ) Linearity Condition and Coverage Condition.
n
span E[x|y]
o
n
= span V ∗,1 , ..., V ∗,d
o
(B.1)
where V ∗,i is the i-th columns of the orthogonal matrix V and a smoothness and tail condition
on the central curve E[x|y], i.e.,
B′ ) Smoothness and Tail Condition.
Smoothness condition: For B > 0 and n ≥ 1, let Πn (B) be the collection of all the n-point
partitions −B ≤ y(1) ≤ · · · ≤ y(n) ≤ B of [−B, B]. The central curve m(y) satisfies the
following:
lim
sup n−1/4
n→∞ y∈Π (B)
n
Tail condition:
such that
n
X
i=2
km(yi ) − m(yi−1 )k2 = 0, ∀B > 0.
e
for some B0 > 0, there exists a non-decreasing function m(y)
on (B0 , ∞),
e 4 (y)P (|Y | > y) → 0 as y → ∞
m
e
e ′ )| for y, y ′ ∈ (−∞, −B0 ) ∪ (B0 , ∞)
km(y) − m(y ′ )k2 ≤ |m(y)
− m(y
(B.2)
Recent results in Lin et al. [2015], Lin et al. [2016] and Neykov et al. [2016] suggest that the
following sliced stability condition (see Definition 1) is more convenient to build the high
dimensional theory of SIR.
Definition 1. Let Y be a univariate random variable. For 0 < γ 1 < 1 < γ 2 , let
AH (γ 1 , γ 2 ) denote all partitions {−∞ = a1 ≤ a2 ≤ . . . ≤ aH+1 = +∞} of R, such that
γ
γ1
≤ P(ah ≤ Y ≤ ah+1 ) ≤ 2 .
H
H
A curve m(y) is ϑ-sliced stable with respect to y, if there exist positive constants γ 1 , γ 2 , γ 3
such that for any partition ∈ AH (γ 1 , γ 2 ) and any β ∈ Rp , one has:
X
γ
1 H+1
var β τ m(Y ) ah−1 ≤ Y < ah ≤ 3ϑ var (β τ m(Y )) .
H h=1
H
A curve is sliced stable if it is ϑ-sliced stable for some positive constant ϑ.
(B.3)
Sparse Sliced Inverse Regression for High Dimensional Data
7
τ
Note that the LHS of (B.3) ≈ var(E[β m(Y )|Y ]) = 0, so the sliced stability condition is
merely requiring the convergence rate to be a power of the slice number and is believed to be
a general condition. In fact, Neykov et al. [2016] shows that if we replace the tail condition
(B.2) by a slightly stronger condition
e )4 ] < ∞,
E[m(Y
(B.4)
then B′ ) implies the sliced stability condition with ϑ = 12 .
One of the main implication of the sliced stability condition is the following lemma (see
Lin et al. [2015]), which plays the key role in their sequential works.
Lemma 10. Let x ∈ Rp be a sub-Gaussian random variable. For any unit vector β ∈ Rp ,
let x(β) = hx, βi and m(β) = hm, βi = E[x(β) | y], we have the following:
i) If var(m(β)) = 0 , there exists positive constants C1 , C2 and C3 such that for any
b = O(1) and sufficiently large H, we have
P(varH (x(β)) > b) ≤ C1 exp −C2
nb
+ C3 log(H) .
H2
ii) If var(m(β)) 6= 0 , there exists positive constants C1 , C2 and C3 such that, for any
ν > 1, we have
1
|varH (x(β)) − var(m(β))| ≥
var(m(β))
2ν
with probability at most
n var(m(β))
+ C3 log(H) .
C1 exp −C2
H 2ν 2
where we choose H such that H ϑ > C4 ν for some sufficiently large constant C4 .
We summarize serveral simple implications of this lemma into the following Proposition,
the proof of which is simple and thus omited. For interest readers, we refer to Lin et al.
[2016].
Proposition 3. There exist positive constants C1 ,C2 and C3 , such that
kβτ ΛH β − βτ var(E[x|y])βk2 ≥
1 τ
β var(E[x|y])β
2ν
(B.5)
with probability at most C1 exp −C2 Hnλ
2 ν 2 + C3 log(H) .
Proof. It follows from Lemma 10 and the fact that for any β ∈ col(var(E[x|y])),
var(m(β))) ≥ λ.
✷
C.
Results of simulations
8
Lin, Zhao, and Liu
Table C1. Estimation error: Σ = Σ1 , and ρ = 0.
p
Lasso-SIR DT-SIR Lasso M-Lasso
100
0.09
0.21
0.08
0.09
1000
0.12
0.21
0.10
0.22
I
2000
0.14
0.22
0.10
0.29
4000
0.18
0.22
0.11
0.39
100
0.05
0.29
0.23
0.05
1000
0.09
0.35
0.30
0.12
II
2000
0.12
0.38
0.31
0.18
4000
0.15
0.41
0.33
0.27
100
0.17
0.23
1.14
0.20
1000
0.27
0.30
1.25
0.63
III
2000
0.35
0.34
1.31
0.77
4000
0.45
0.42
1.29
0.93
100
0.35
0.79
0.41
0.98
1000
0.59
0.96
0.61
0.84
IV
2000
0.72
1.02
0.67
1.01
4000
0.95
1.14
0.71
1.23
100
0.10
0.18
0.55
0.11
1000
0.12
0.19
0.69
0.30
V
2000
0.15
0.19
0.72
0.37
4000
0.18
0.20
0.74
0.47
Lasso-SIR(Known d)
0.09
0.12
0.14
0.18
0.05
0.09
0.11
0.15
0.18
0.23
0.26
0.34
0.35
0.56
0.64
0.82
0.09
0.13
0.15
0.18
dˆ
1
1
1
1
1
1
1
1
1
1.1
1.1
1.3
1
1.1
1.2
1.4
1
1
1
1
Table C2. Estimation error: Σ = Σ1 and ρ = 0.3.
p
Lasso-SIR DT-SIR Lasso M-Lasso
100
0.10
0.435
0.09
0.12
1000
0.15
0.50
0.12
0.24
I
2000
0.18
0.50
0.13
0.31
4000
0.21
0.49
0.14
0.42
100
0.06
0.46
0.22
0.08
1000
0.11
0.55
0.28
0.14
II
2000
0.14
0.55
0.30
0.20
4000
0.19
0.58
0.32
0.32
100
0.19
0.50
1.18
0.24
1000
0.29
0.60
1.30
0.58
III
2000
0.35
0.63
1.32
0.73
4000
0.57
0.75
1.33
0.98
100
0.38
0.85
0.56
0.54
1000
0.56
0.97
0.70
0.77
IV
2000
0.65
1.03
0.76
0.92
4000
0.79
1.12
0.79
1.09
100
0.10
0.47
0.48
0.14
1000
0.14
0.55
0.60
0.35
V
2000
0.17
0.56
0.66
0.44
4000
0.30
0.60
0.72
0.66
Lasso-SIR(Known d)
0.1
0.15
0.17
0.22
0.06
0.11
0.14
0.19
0.19
0.25
0.29
0.4
0.37
0.54
0.58
0.65
0.1
0.15
0.18
0.25
dˆ
1
1
1
1
1
1
1
1
1
1
1.1
1.4
1
1
1.1
1.3
1
1
1
1
Sparse Sliced Inverse Regression for High Dimensional Data
Table C3. Estimation error: Σ = Σ1 , and ρ = 0.8.
p
Lasso-SIR DT-SIR Lasso M-Lasso
100
0.18
1.34
0.16
1.01
1000
0.24
1.38
0.22
0.79
I
2000
0.27
1.39
0.23
0.53
4000
0.32
1.39
0.25
0.45
100
0.10
1.34
0.33
1.17
1000
0.16
1.39
0.55
1.08
II
2000
0.19
1.39
0.71
0.92
4000
0.23
1.40
0.92
0.54
100
0.28
1.34
1.26
1.00
1000
0.45
1.38
1.29
0.92
III
2000
0.54
1.39
1.30
0.84
4000
0.76
1.43
1.29
0.89
100
0.74
1.40
1.21
0.91
1000
0.75
1.41
1.23
0.88
IV
2000
0.79
1.44
1.26
0.94
4000
0.93
1.52
1.27
1.09
100
0.19
1.31
0.36
1.10
1000
0.31
1.38
0.56
0.55
V
2000
0.50
1.42
0.74
0.71
4000
1.15
1.66
0.82
1.25
Lasso-SIR(Known d)
0.18
0.24
0.27
0.32
0.11
0.16
0.19
0.23
0.28
0.44
0.54
0.68
0.72
0.76
0.75
0.76
0.19
0.32
0.47
0.80
dˆ
1
1
1
1
1
1
1
1
1
1
1
1.1
1
1
1.1
1.4
1
1
1.1
2.1
Table C4. Estimation error: Σ = Σ2 and ρ = 0.2.
p
Lasso-SIR DT-SIR Lasso M-Lasso
100
0.13
1.22
0.09
0.16
1000
0.33
1.37
0.11
0.65
I
2000
0.30
1.37
0.12
0.74
4000
0.36
1.38
0.13
0.81
100
0.10
1.26
0.24
0.11
1000
0.25
1.37
0.31
0.47
II
2000
0.30
1.38
0.33
0.59
4000
0.31
1.40
0.35
0.65
100
0.24
1.24
1.19
0.33
1000
0.55
1.33
1.30
0.98
III
2000
0.59
1.35
1.28
1.08
4000
0.58
1.36
1.28
1.14
100
0.54
1.37
1.19
0.60
1000
0.63
1.41
1.25
0.87
IV
2000
0.64
1.41
1.27
0.99
4000
0.65
1.41
1.26
1.07
100
0.23
1.23
0.49
0.29
1000
1.03
1.10
0.61
1.15
V
2000
1.09
0.96
0.67
1.20
4000
1.07
0.99
0.71
1.21
Lasso-SIR(Known d)
0.13
0.26
0.30
0.30
0.10
0.25
0.29
0.32
0.23
0.40
0.45
0.47
0.54
0.63
0.65
0.66
0.13
0.79
0.86
0.87
dˆ
1
1.1
1
1.1
1
1
1
1
1
1.3
1.3
1.3
1
1
1
1
1.1
1.6
1.6
1.7
9
10
Lin, Zhao, and Liu
Table C5. Estimation error: Σ = Σ1 and ρ = 0.
p
Lasso-SIR DT-SIR M-Lasso Lasso-SIR(Known d)
100
0.15
0.18
0.23
0.14
1000
0.18
0.17
0.61
0.17
VI
2000
0.22
0.20
0.72
0.20
4000
0.28
0.20
0.86
0.27
100
0.27
0.35
0.32
0.27
1000
0.37
0.40
0.93
0.37
VII
2000
0.44
0.41
1.09
0.45
4000
0.75
0.64
1.40
0.60
100
0.87
0.88
0.90
0.23
1000
0.45
0.44
0.91
0.31
VIII
2000
0.34
0.35
0.80
0.34
4000
0.57
0.53
1.04
0.41
100
0.87
0.89
0.91
0.26
1000
0.60
0.54
1.10
0.39
IX
2000
0.78
0.71
1.18
0.59
4000
0.96
0.83
1.25
0.83
dˆ
2
2
2
2
2
2
2
2.4
1.2
1.8
2
1.8
1.2
1.7
1.6
1.5
Table C6. Estimation error: Σ = Σ1 and ρ = 0.3.
p
Lasso-SIR DT-SIR M-Lasso Lasso-SIR(Known d)
100
0.20
0.34
0.25
0.20
1000
0.24
0.30
0.61
0.23
VI
2000
0.26
0.36
0.71
0.26
4000
0.31
0.41
0.84
0.29
100
0.28
0.63
0.42
0.28
1000
0.41
0.71
0.95
0.40
VII
2000
0.58
0.78
1.17
0.54
4000
0.97
0.92
1.46
0.78
100
0.25
0.55
0.35
0.22
1000
0.32
0.59
0.77
0.29
VIII
2000
0.34
0.69
0.81
0.34
4000
0.57
0.77
1.11
0.47
100
0.31
0.50
0.43
0.31
1000
0.35
0.47
0.99
0.35
IX
2000
0.42
0.55
1.17
0.40
4000
0.51
0.56
1.28
0.44
dˆ
2
2
2
2
2
2
2.1
2.3
2
2
2
2.2
2
2
2.1
2
Sparse Sliced Inverse Regression for High Dimensional Data
Table C7. Estimation error: Σ = Σ1 and ρ = 0.8.
p
Lasso-SIR DT-SIR M-Lasso Lasso-SIR(Known d)
100
0.52
1.86
1.01
0.51
1000
0.79
1.92
0.93
0.79
VI
2000
0.96
1.94
1.05
0.94
4000
1.14
2.01
1.26
1.06
100
0.80
1.77
1.07
0.72
1000
1.09
1.78
1.23
1.34
VII
2000
1.09
1.76
1.23
1.39
4000
1.12
1.76
1.27
1.42
100
0.42
1.81
0.79
0.34
1000
1.00
1.97
1.22
0.86
VIII
2000
1.12
1.93
1.27
1.17
4000
1.16
1.89
1.27
1.28
100
0.78
1.90
0.95
0.79
1000
0.92
1.95
1.08
0.90
IX
2000
0.97
1.97
1.17
0.95
4000
1.12
2.03
1.37
1.01
dˆ
2
2
2
2.2
1.6
1.3
1.3
1.2
2
2.2
2.1
1.8
2
2
2
2.3
Table C8. Estimation error: Σ = Σ2 and ρ = 0.2.
p
Lasso-SIR DT-SIR M-Lasso Lasso-SIR(Known d)
100
0.27
1.73
0.43
0.21
1000
1.01
1.73
1.11
0.26
VI
2000
1.01
1.73
1.14
0.29
4000
1.02
1.73
1.18
0.38
100
0.39
1.70
0.55
0.31
1000
1.03
1.73
1.25
0.47
VII
2000
1.04
1.73
1.30
0.55
4000
1.04
1.73
1.34
0.69
100
0.24
1.69
0.34
0.24
1000
0.97
1.73
1.15
0.33
VIII
2000
1.03
1.74
1.24
0.35
4000
1.03
1.74
1.26
0.41
100
1.00
1.69
1.04
0.40
1000
1.03
1.73
1.22
0.67
IX
2000
1.03
1.73
1.27
0.73
4000
1.04
1.73
1.30
0.89
dˆ
1.9
1
1
1
1.9
1
1
1
2
1.1
1
1
1
1
1
1
11
12
Lin, Zhao, and Liu
Table C9. Estimation error: Σ = Σ1 and ρ = 0.
p
Lasso-SIR DT-SIR M-Lasso
100
0.19
0.23
0.22
1000
0.22
1.13
0.60
X
2000
0.23
1.24
0.67
4000
0.24
1.29
0.71
100
0.36
0.82
0.41
1000
0.44
1.29
1.16
XI
2000
0.41
1.35
1.22
4000
0.45
1.37
1.23
100
0.23
0.53
0.25
1000
0.30
1.12
0.62
XII
2000
0.32
1.23
0.71
4000
0.34
1.29
0.75
100
0.36
0.92
0.43
1000
0.44
1.69
1.22
XIII
2000
0.44
1.80
1.30
4000
0.43
1.85
1.33
Lasso
0.18
0.25
0.27
0.29
0.37
0.43
0.48
0.47
0.21
0.30
0.34
0.36
1.05
1.08
1.09
1.10
Table C10. Estimation error: Σ = Σ1 and ρ = 0.3.
p
Lasso-SIR DT-SIR M-Lasso Lasso
100
0.20
0.50
0.26
0.19
1000
0.24
1.15
0.56
0.26
X
2000
0.25
1.25
0.63
0.28
4000
0.26
1.31
0.68
0.30
100
0.34
0.86
0.57
0.34
1000
0.43
1.31
1.13
0.43
XI
2000
0.42
1.36
1.20
0.45
4000
0.45
1.38
1.22
0.48
100
0.22
0.54
0.28
0.21
1000
0.29
1.10
0.57
0.30
XII
2000
0.31
1.22
0.66
0.34
4000
0.33
1.29
0.71
0.36
100
0.38
0.99
0.61
1.06
1000
0.39
1.73
1.19
1.08
XIII
2000
0.41
1.84
1.29
1.09
4000
0.42
1.88
1.33
1.10
Sparse Sliced Inverse Regression for High Dimensional Data
Table C11. Estimation error: Σ = Σ1 and ρ = 0.8.
p
Lasso-SIR DT-SIR M-Lasso Lasso
100
0.28
1.36
0.99
0.28
1000
0.45
1.41
0.90
0.39
X
2000
0.55
1.41
0.82
0.46
4000
0.65
1.41
0.77
0.54
100
0.40
1.38
0.84
0.39
1000
0.71
1.41
1.04
0.69
XI
2000
0.89
1.41
1.11
0.88
4000
1.02
1.4
1.17
1.02
100
0.36
1.37
1.17
0.30
1000
0.63
1.41
1.16
0.52
XII
2000
0.83
1.41
1.14
0.71
4000
1.03
1.41
1.09
0.91
100
0.48
1.92
0.89
1.10
1000
0.56
1.99
1.14
1.12
XIII
2000
0.59
2.00
1.21
1.14
4000
0.61
2.00
1.31
1.15
Table C12. Estimation error: Σ = Σ2 and ρ = 0.2.
p
Lasso-SIR DT-SIR M-Lasso Lasso
100
0.19
1.24
0.28
0.19
1000
0.23
1.41
0.61
0.26
X
2000
0.24
1.41
0.68
0.28
4000
0.26
1.41
0.73
0.31
100
0.36
1.29
0.58
0.35
1000
0.42
1.41
1.16
0.43
XI
2000
0.46
1.41
1.22
0.48
4000
0.46
1.41
1.24
0.49
100
0.24
1.25
0.31
0.22
1000
0.32
1.40
0.62
0.32
XII
2000
0.34
1.41
0.71
0.34
4000
0.35
1.41
0.77
0.38
100
0.40
1.72
0.67
1.06
1000
0.44
1.99
1.22
1.09
XIII
2000
0.46
2.00
1.31
1.10
4000
0.45
2.00
1.37
1.11
13
| 10math.ST
|
Multi-Pointer Co-Attention Networks for Recommendation
Yi Tay
Nanyang Technological University
Singapore
[email protected]
Luu Anh Tuan
Institute for Infocomm Research
Singapore
[email protected]
arXiv:1801.09251v1 [cs.CL] 28 Jan 2018
ABSTRACT
Many recent state-of-the-art recommender systems such as D-ATT,
TransNet and DeepCoNN exploit reviews for representation learning. This paper proposes a new neural architecture for recommendation with reviews. Our model operates on a multi-hierarchical paradigm and is based on the intuition that not all reviews are created
equal, i.e., only a select few are important. The importance, however,
should be dynamically inferred depending on the current target.
To this end, we propose a review-by-review pointer-based learning
scheme that to extract important reviews from user and item reviews, subsequently matches them in a word-by-word fashion. This
enables not only the most informative reviews to be utilized for prediction but also a deeper word-level interaction. Our pointer-based
method operates with a novel gumbel-softmax based pointer mechanism that enables incorporation discrete vectors within differentiable neural architectures. Our pointer mechanism is co-attentive
in nature, i.e., learning pointers which are co-dependent on useritem relationships. Finally, we propose a multi-pointer learning
scheme that learns to combine multiple views of interactions between user and item. Overall, we demonstrate the effectiveness
of our proposed model via extensive experiments on 24 benchmark datasets from Amazon and Yelp. Empirical results show that
our approach significantly outperforms existing state-of-the-art,
with up to 19% and 71% relative improvement when compared to
TransNet and DeepCoNN respectively. We study the behavior of
our multi-pointer learning mechanism, shedding light on evidence
aggregation patterns in review-based recommender systems.
KEYWORDS
Deep Learning; Recommendation; Collaborative Filtering
ACM Reference Format:
Yi Tay, Luu Anh Tuan, and Siu Cheung Hui. 2018. Multi-Pointer Co-Attention
Networks for Recommendation. In Proceedings of ACM Conference (Conference’17). ACM, New York, NY, USA, 9 pages. https://doi.org/10.1145/
nnnnnnn.nnnnnnn
1
INTRODUCTION
On most e-commerce platforms today, the ability to write and share
reviews is not only a central feature but is also a strongly encouraged act. Reviews are typically informative, pooling an extensive
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from [email protected].
Conference’17, July 2017, Washington, DC, USA
© 2018 Association for Computing Machinery.
ACM ISBN 978-x-xxxx-xxxx-x/YY/MM. . . $15.00
https://doi.org/10.1145/nnnnnnn.nnnnnnn
Siu Cheung Hui
Nanyang Technological University
Singapore
[email protected]
wealth of knowledge for prospective customers. However, the extensive utility of reviews do not only end at this point. Reviews are
also powerful in capturing preferences of authors, given the rich
semantic textual information that cannot be conveyed via implicit
interaction data or purchase logs. As such, there have been immense interest in collaborative filtering systems that exploit review
information for making better recommendations [4, 17, 37].
Recent advances in deep learning has spurred on various innovative models that exploit reviews for recommendation [2, 24, 37].
The intuition is simple yet powerful, i.e., each user is represented
as all reviews he (she) has written and an item is represented by
all reviews that was written for it. All reviews are concatenated to
form a single user (item) document. Subsequently, a convolutional
encoder is employed to learn a single latent representation for the
user (item). User and item embeddings are then matched using a
parameterized function such as Factorization Machines [21]. This
has shown to be highly performant [24], outperforming a wide
range of traditionally strong baselines such as Matrix Factorization
(MF). Models such as DeepCoNN [37], TransNets [2] and D-ATT
[24] are recent state-of-the-art models that are heavily grounded
in this paradigm.
Intuitively, this modeling paradigm leaves a lot to be desired.
Firstly, the naive concatenation of reviews into a single document is
unnatural, ad-hoc and noisy. In this formulation, reviews are treated
indiscriminatingly irregardless of whether they are important or
not. A user’s bad review about a coffee shop should be mostly
irrelevant when deciding if a Spa is a good match. Secondly, user
and item representations are static irregardless of the target match.
For example, when deciding if a coffee shop is a good match for a
user, the user’s past reviews about other coffee shops (and eateries)
should be highly relevant. Conversely, reviews about spas and gyms
should not count. Hence, a certain level of dynamism is necessary.
Finally, the only accessible interaction between user and item is
through a fixed dimensional representation, which is learned via
excessive compression of large user-item review banks into lowdimensional vector representations. For richer modeling of user and
item reviews, deeper and highly accessible interaction interfaces
between user-item pairs should be mandatory.
Recall that reviews were fundamentally written independently,
at different times, and for different products (or by different people).
There should be no reason to squash everything into one long
document if they can be modeled independently and then combined
later. More importantly, a user may write hundreds of reviews over
an extended period of time while an item may effortlessly receive
a thousand of reviews. As such, existing modeling paradigms will
eventually hit a dead-end. Overall, this work proposes four major
overhauls have to be made to existing models, i.e., (1) reviews
should be modeled independently, (2) not all reviews are equally
important and should be weighted differently, (3) the importance
of each review should be dynamic and dependent on the target
match and finally (4), user and item reviews should interact not
only through compressed vector representations but at a deeper
granularity, i.e., word-level.
To this end, we propose a Multi-Pointer Co-Attention Network
(MPCN), a novel deep learning architecture that elegantly satisfies our key desiderata. Our model is multi-hierarchical in nature,
i.e., each user is represented by n reviews of ℓ words each. Subsequently, all user and item reviews (for this particular instance pair)
is matched to determine the most informative reviews. In order to
do so, we design a novel pointer-based co-attention mechanism. The
pointer mechanism extracts the named reviews for direct reviewto-review matching. At this stage, another co-attention mechanism
learns a fixed dimensional representation, by modeling the wordlevel interaction between these matched reviews. This forms the
crux of our review-by-review modeling paradigm. Finally, we introduce a multi-pointer learning scheme that can be executed an
arbitrary k times, extracting multiple multi-hierarchical interactions between user and item reviews.
1.1
Our Contributions
In summary, the prime contributions of this paper are as follows:
• We propose a state-of-the-art neural model for recommendation with reviews. Our proposed model exploits a novel
pointer-based learning scheme. This enables not only noisefree but deep word-level interaction between user and item.
• We conduct experiments on 24 benchmark datasets. Our
proposed MPCN model outperforms all state-of-the-art baselines by a significant margin across all datasets. Our compared baselines are highly competitive, encompassing not
only review-based models but also state-of-the-art interactiononly models. We outperform models such as Neural Matrix
Factorization (NeuMF) [8], DeepCoNN [37], D-ATT [24] and
TransNet [2]. Performance improvement over DeepCoNN,
TransNets and D-ATT are up to 71%, 19% and 5% respectively.
• We investigate the inner workings of our proposed model
and provide insight about MPCN works under the hood. Additionally, analyzing the behavior of our pointer mechanism
allows us to better understand the nature of the problem.
Through analysis of our pointer mechanism, we show that
different problem domains have different patterns of ‘evidence aggregation’, which might require different treatment
and special care.
2
RELATED WORK
In this section, we identify and review several key areas which are
highly relevant to our work.
2.1
Review-based Recommendation
The utility of exploiting reviews for recommendations have been
extensively discussed and justified in many works [2, 11, 17, 24, 37].
This not only enables a mitigation of cold-start issues but also provide a richer semantic modeling of user and item characteristics.
While relatively earlier works have mainly concentrated efforts
on topic modeling and language modeling approaches [4, 17, 30],
the recent shift towards deep learning models is prominent. The
advantages of neural architectures are clear, i.e., not only do these
models dispense with laborious feature engineering altogether, they
are often highly competitive. In many recent works, Convolutional
neural networks (CNN) act as automatic feature extractors, encoding a user (item) into a low-dimensional vector representation. User
and item embeddings are then compared with a matching function.
An earlier neural model, the Deep Co-operative Neural Networks (DeepCoNN) [37] represents a user as all the reviews that
he (she) has written. Likewise, an item is represented as all the
reviews ever written (by other users) for it. User and item documents are then encoded with CNNs and passed into a Factorization
Machine (FM) [21] for matching. It was later argued that DeepCoNN’s competitive performance exploits the fact that test reviews
were leaked (into the training set) [2]. As such, this reduces the
recommendation problem to resemble a noisy adaption of standard document-level sentiment analysis. To this end, [2] proposed
TransNets, augmenting a DeepCoNN-like neural network with an
additional multi-task learning scheme. More specifically, it learns to
transform the penultimate hidden layer of DeepCoNN into a CNNencoded representation of the test review. This signal was found to
be useful, improving the performance on multiple benchmarks.
DeepCoNN and TransNet are relatively simple model architectures. Another recently proposed model, the Dual Attention CNN
model (D-ATT) [24] proposed augmenting CNNs with neural attention. The key idea of neural attention [1] is to emphasize important
segments of documents by weighting each word by a learned attention vector. The final representation comprises a weighted linear
combination of all input embeddings. Two variants of attention
mechanism are proposed, i.e., local and global, both modeling different views of user-item review documents. However, these models
are not without flaws. As mentioned, these models follow the same
paradigm of representing user and items as a giant concatenated
document of all their reviews which suffers inherent drawbacks
such as (1) noise, (2) lack of dynamic target-dependent and (3) lack
of interaction interfaces.
2.2
Text Matching and Co-Attentional Models
Our work is closely related to the problems domains of sequence
pair modeling. A wide spectrum of models have been proposed
for modeling relationships between two sequences, e.g., questionanswer [5, 26], premise-hypothesis [20, 22, 27] which are very similar to the user-item modeling problem at hand. In these fields,
learning representations without fine-grained interaction modeling, i.e., absence of interaction interfaces ala DeepCoNN, is known
to be far outperformed by new models which utilize co-attentional
mechanisms [5, 36]. Co-attentions learn pairwise attention between
two sequences [35] (or modalities [15]), enabling pair-aware attention weights to be learned.
2.3
Recent Advances in Deep Learning
Notably, our network solely relies on attention mechanisms, and
showcases the potential neural architectures that do not use convolutional and recurrent layers. This is reminiscent of (and inspired
by) the Transformer [28] architecture which uses multi-headed
attention, concatenating outputs of each attention call.
Additionally, our work is characterized by the usage of pointers,
which have been popularized by both Pointer Networks [29] which
learn to predict an output token which exists in the sequence itself,
i.e., pointing to a token. The usage of pointers are primarily motivated for discrete problems and have also been widely adopted for
answer span prediction in question answering [32]. In this models,
pointers are commonly applied at the last layer and have no issues
since the model optimizes a loss function such as the cross entropy
loss. However, our model requires the usage of pointers within
the network (between layers). As such, a form of hard attention is
required. Due to the non-differentiability of hard attention, it has
been much less popular than the standard soft attention. In order
to generate discrete pointers, we utilize the recent gumbel-softmax
[10] trick which is capable of learning one-hot encoded vectors.
Notably, the gumbel-softmax was recently adapted for automatic
compositional learning of Gumbel TreeLSTMs [3], in which the
gumbel-softmax is exploited to adaptively learn to make merging
decisions.
2.4
OUR PROPOSED MODEL
In this section, we present a layer-by-layer description of our proposed model. The overall model architecture is illustrated in Figure
2.
3.1
Sum
Input Encoding
Our model accepts two input sequences, a (user) and b (item). Each
input sequence is a list of reviews {r 1 , r 2 · · · r ℓd } where ℓd is the
maximum number of reviews.
3.1.1 Embedding Layer. Each review is a sequence of ℓw words
which are represented as one-hot encoded vectors. For a and b, we
pass all words into an embedding matrix Wd × |V | where V is the
vocabulary, retrieving a d dimensional vector for each word. Our
network is hierarchical in nature, i.e., instead of representing all
user (or item) reviews as one long document, we represent each user
(or item) as a sequence of reviews. Each review is then hierarchically
constructed from a sequence of words.
3.1.2 Review Gating Mechanism. Each review is represented as
a sum of its constituent word embeddings to form the vector ∈ Rd .
Intuitively, not all reviews that a user writes and not all reviews
Sum
Word-level
Co-Attention
Expand
Review To
Word-level
Expand
Review To
Word-level
Review
Pointers
0
1
0
0
0
0
0
1
Row/Col
Max Pooling
Review
Pointers
Review-level
Co-Attention
Deep Learning for Recommendation
Factorization-based models [9, 19, 21] were popular standard machine learning baselines for interaction-based recommendation.
Today, we see a shift into deep learning, in which neural models
are claiming state-of-the-art performance. He et al. [8] proposed
a neural framework that combines a generalized matrix factorization formulation with multi-layered perceptrons. He and Chua
[7] proposed a neural adaptation of factorization machines. [34]
proposed recurrent models for sequential modeling of interaction
data. Tay et al. [25] proposed a translation-based framework that
exploits neural attention for modeling user-item relations. Li et al.
[13] proposed a encoder-decoder based framework for both rating
prediction and generating tips. A separate class of deep models
based on auto-encoders [14, 23, 31] have also been proposed for
recommendation tasks.
3
FM
Review
Embeddings
User Reviews
Item Reviews
Figure 1: Illustration of proposed model architecture for
Pointer-based Learning (Best viewed in color). This example
illustrates a one pointer example. Review gating mechanism
and multi-pointer learning is omitted for clarity.
written for a product is important. We design a first-stage filter, i.e.,
a review gating mechanism that accepts each review as an input
and controls how much information passes through to the next
level. Given the input x ∈ Rℓr ×d , which represents either a or b.
x¯i = σ (Wд x i ) + bд ⊙ tanh(Wu x i + bu )
(1)
where ⊙ is the Hadamard product and σ is the sigmoid activation
function. x i is the i-th review of sequence x. Wд , Wu ∈ Rd ×d and
bд , bu ∈ Rn are parameters of this layer. While the purpose of
the co-attention layer is to extract important reviews, we hypothesize that applying a pre-filter (gating mechanism) helps improve
performance on certain datasets.
3.2
Review-level Co-Attention
In this layer, the aim is to select the most informative review from
the review bank of each user and item respectively.
3.2.1 Affinity Matrix. Given a list of review embeddings from
each user (a ∈ Rℓr ×d ) and item (b ∈ Rℓr ×d ) banks, we calculate an
affinity matrix between them. This is described as follows:
si j = F (ai )⊤ M F (b j )
(2)
where Md ×d and S ∈ Rℓr ×ℓr . F (.) is a feed-forward neural network
function with l layers. In practice, l is tuned amongst [0, 2] where
l = 0 reverts Equation (2) to the bilinear form.
3.2.2 Pooling Function. By taking the row and column wise
maximum of the matrix s and using them to weight the original list
of reviews a and b, we are able to derive the standard co-attention
mechanism. This is described as follows:
a ′ = (G(max(s)))⊤a
col
b ′ = (G(max(s)))⊤b
r ow
(3)
There are different choices for the pooling operation. The choice
of the max pooling is because, intuitively max pooling selects the
review which has the maximum influence (or affinity) with all
reviews from its partner. This type of co-attention is known to be
extractive, characterized by its usage of max-pooling.
Note that we apply the function G(.) to maxcol (s) and maxr ow (s).
In most applications, G(.) would correspond to the standard softmax function, which converts the input vector into a probability
distribution. Vectors a ′, b ′ would then be the co-attentional vector
representations). However, in our case, we desire further operations
on the selected reviews and therefore do not make use of these representations. Instead, G(.) has to return a one-hot encoded vector,
pointing to the selected reviews, forming the real objective behind
this co-attentional layer. However, the Softmax function returns a
continuous vector, which is unsuitable for our use-case. The usage
of discrete vectors in neural architectures is known to be difficult,
as the arg max operation is non-differentiable. Hence, we leverage
a recent advance, the Gumbel-Max trick, for learning to point. The
next section describes this mechanism.
3.3
Review Pointers
We leverage a recent advance, the Gumbel-Softmax [10] for incorporating discrete random variables in the network.
3.3.1 Gumbel-Max. In this section, we describes Gumbel-Max
[16], which facilitates the key mechanism of our MPCN model.
Gumbel-Max enables discrete random variables (e.g., one hot vectors) to be utilized within an end-to-end neural network architecture. Consider a k-dimensional categorical distribution where class
probabilities p1 , · · · pk are defined in terms of unnormalized log
probabilities π1 , · · · πk :
exp(log(πi ))
pi = Í
k exp(log(π ))
j
j=1
(4)
Rk
A one hot sample z = (z 1 , · · · zk ) ∈
from the distribution can be
drawn by the following:
(
1 , i = arg maxj (log(π j ) + дj )
zi =
(5)
0 , otherwise
дi = − log(− log(ui ))
ui ∼ U ni f orm(0, 1)
(6)
where дi is the Gumbel noise which perturbs each log(πi ) term
such that the arg max operation is equivalent to drawing a sample
weighted by pi , · · · pk .
3.3.2 Straight-Through Gumbel-Softmax. In the Gumbel-Softmax,
the key difference is that the arg max function is replaced by the
differentiable softmax function as described as follows:
log(π )+д
i
i
exp(
)
τ
yi = Í
log(π
)+д
j
i
k exp(
)
τ
j=1
(7)
where τ is a temperature parameter. As τ approaches zero, the output of the Gumbel-Softmax distribution becomes cold, i.e., becomes
closer to a one-hot vector. In the straight-through (ST) adaptation,
the forward-pass is discretized by converting the vector output to
a one-hot vector via arg max:
(
1 , i = arg maxj (y j )
yi =
(8)
0 , otherwise
However, the backward pass maintains the flow of continuous
gradients which allows the model to be trained end-to-end. This
is useful as we only want important reviews to be selected (i.e.,
hard selected) to be considered in computation of the loss function.
Notably, alternatives such as REINFORCE [33] exists. However, it
is known to suffer from high variance and slow convergence [10].
3.3.3 Learning to Point. In order to compute a review pointer
(for user and item), we set G(.) in Equation 3 to use the Gumbel
Softmax. However, since we are interested only in the pointer (to
be used in subsequent layers), the pointer is then calculated as:
pa = (Gumbel(max(s)))
col
pb = (Gumbel(max(s)))
r ow
(9)
By applying pa to a, we select the pa − th review of user a and
pb − th review of item b. The selected reviews are then passed into
the next layer where rich interactions are extracted between these
reviews.
3.4
Word-level Co-Attention
The review-level co-attention smooths over word information as
it compresses each review into a single embedding. However, the
design of model allows the most informative reviews to be extracted
by the use of pointers. These reviews can then be compared and
modeled at word-level. This allows a user-item comparison of finer
granularity which facilitates richer interactions as compared to
simply composing the two review embeddings. Let ā, b̄ be the selected reviews using the pointer learning scheme. Similar to the
review-level co-attention, we compute a similarity matrix between
ā and b̄. The key difference is that the affinity matrix is computed
word-by-word and not review-by-review.
w i j = F (āi )⊤ Mw F (b̄ j )
(10)
where Mw ∈ Rd ×d and w ∈ Rℓw ×ℓw . Next, to compute the coattentional representation of reviews ā, b̄, we take the mean pooling.
ā ′ = (S(avдcol (s)))⊤ā
b̄ ′ = (S(avдr ow (s)))⊤b̄
(11)
where S(.) is the standard Softmax function and F (.) is a standard
feed-forward neural network with l layers. The rationale for using
the average pooling operator here is also intuitive. At the reviewlevel, a large maximum affinity score of a review with respect to all
opposite reviews (even when a low average affinty socre) warrants it
being extracted, i.e., a strong signal needs to be further investigated.
However, at a word-level, max-pooling may be biased towards
identical words and are have a great inclination to act as a word
matching operator. Hence, we want to maintain a stable co-attentive
extractor. Our early empirical experiments also justify this design.
Finally ā ′ and b̄ ′ are the output representations.
3.5
i=1
Multi-Pointer Learning
While our objective is to eliminate noisy reviews by the usage
of hard pointers, pointing to only a single pair of review might
be insufficient. Hence, we a devise a multi-pointer composition
mechanism. The key idea is to use multiple pointers where the
number of pointers np is a user-defined hyperparameter. Our model
runs the Review-level Co-Attention np times, each generating a
unique pointer. Each of the np review pairs are then modeled with
the Word-level Co-Attention mechanism. The overall output is a
list of vectors {ā 1′ , · · · ān′ p } and {b̄1′ , · · · b̄n′ p }. Additionally, we also
found it useful include the sum embedding of all words belonging to
the user (item). This mainly helps in robustness, in-case where user
and item does not have any matching signals that was found by
our pointer mechanism. We explore three ways to compose these
embeddings.
• Concatenation - All pointer outputs are concatenated, e.g.,
[ā 1′ ; · · · ān′ p ]. This has implications to the subsequent layers,
especially if np is large. Hence we consider the last alternative.
• Additive Composition - All pointer outputs are summed,
e.g., sum(ā 1′ , · · · ān′ p ).
• Neural Network - All pointer outputs are concatenated
and passed through a single layered non-linear layer. e.g.,
ReLU (W ([ā 1′ ; · · · ān′ p ]) +b which maps the concatenated vector into a d dimension vector.
Note that this is applied to b̄ ′ as well but omitted for brevity. Let
the output of this layer be a f and bf . In our experiments, we tune
amongst the three above-mentioned schemes. More details are
provided in the ablation study.
FM
FM
Neural Network
Multi-Pointer
Learning
p1
Additive
Multi-Pointer
Sum
p1
p3
Model
User Reviews Item Reviews
Multi-Pointer
Learning
p3
Model
User Reviews
Item Reviews
Figure 2: Illustration of Neural Network (left) and Additive
(right) based Multi-Pointer Learning.
3.6
real-valued feature vector and models the pairwise interactions
between features using factorized parameters. The FM function is
defined as follows:
n
n Õ
n
Õ
Õ
F (x) = w 0 +
wi xi +
⟨vi , v j ⟩ x i x j
(12)
Prediction Layer
This layer accepts a f , bf as an input. The concatenation of [a f ; bf ]
is passed into a factorization machine (FM) [21]. FMs accepts a
i=1 j=i+1
where x ∈ Rk is a real-valued input feature vector. ⟨., ⟩ is the
dot product. The parameters {v 1 . . . vn } are factorized parameters
(vectors of v ∈ Rk ) used to model pairwise interactions ((x i , x j )).
Í
w 0 is the global bias and ni=1 w i x i represents a linear regression
component. The output of F (x) is a scalar, representing the strength
of the user-item interaction. The network is trained end-to-end by
minimizing the standard mean squared error loss following [24].
4
EMPIRICAL EVALUATION
In this section, we present our experimental setup and empirical
evaluation. Our experiments are designed to answer the following
research questions (RQs):
(1) RQ1 - Does our proposed approach outperform state-of-theart models such as D-ATT and DeepCoNN? How much is
the relative improvement?
(2) RQ2 - What are the impacts of some of the design / architectural choices of MPCN?
(3) RQ3 - What are the effects of the key hyperparameters of our
model (e.g., number of pointers etc.) on model performance?
(4) RQ4 - Are we able to derive any insight about how MPCN
works by analyzing the behavior of the pointer layer?
4.1
Datasets
We utilize datasets from two different sources which are described
as follows:
(1) Yelp Dataset Challenge - Yelp is an online review platform
for businesses such as restaurants, bars and spas etc. We use
the dataset from the latest challenge1 .
(2) Amazon Product Reviews - Amazon is a well-known Ecommerce platform. Users are able to write reviews for products they have purchased. We use 23 datasets from the Amazon Product Review corpus2 [6, 18].
In total, we provide model comparisons over 24 benchmark datasets.
For all datasets, we split interactions into training, development
and testing sets. We utilize a time-based split, i.e., the last item of
each user is added to the test set while the penultimate is used
for development. For Amazon, the datasets are preprocessed in a
5-core fashion (i.e., each user and item have at least 5 reviews to be
included). Since the datasets can be found in the official webpage,
we do not restate their statistics to save space. For Yelp, we use a
20-core setting, providing a comparison on a denser dataset. We
tokenize the reviews using NLTK and retain words that appear at
least 10 times in the vocabulary. We would like to emphasize that,
when building user and item representations using their respective
reviews, all reviews belonging to interactions from the test and development sets were not included. This is to prevent the problem
from reverting to a sentiment analysis task, albeit noisier [2].
1 https://www.yelp.com/dataset/challenge
2 http://jmcauley.ucsd.edu/data/amazon/
Dataset
Yelp17
Instant Video
Instruments
Digital Music
Baby
Patio / Lawn
Gourmet Food
Automotive
Pet Supplies
Office Products
Android Apps
Beauty
Tools / Home
Video Games
Toys / Games
Health
CellPhone
Sports / Outdoors
Kindle Store
Home / Care
Clothing
CDs / Vinyl
Movies / TV
Electronics
MF
1.735
2.769
6.720
1.956
1.755
2.813
1.537
5.080
1.736
1.143
1.922
1.950
1.569
1.627
1.802
1.882
1.972
1.388
1.533
1.667
2.396
1.368
1.690
1.962
Interaction
FM
MLP
1.726 1.727
1.331 1.532
1.166 1.081
1.382 1.361
1.614 1.585
1.311 1.279
1.348 1.442
1.599 1.071
1.618 1.649
0.998 1.122
1.871 1.805
1.711 1.631
1.310 1.356
1.665 1.576
1.195 1.286
1.506 1.522
1.668 1.622
1.195 1.120
1.217 1.231
1.547 1.584
1.492 1.462
1.555 1.432
1.852 1.518
2.120 1.950
NEUMF
1.691
1.451
1.189
1.332
1.598
1.251
1.464
1.013
1.646
1.069
1.842
1.552
1.314
1.568
1.222
1.415
1.606
1.299
1.398
1.654
1.535
1.368
1.775
1.651
D-CON
1.385
1.285
1.483
1.202
1.440
1.534
1.199
1.130
1.447
0.909
1.517
1.453
1.208
1.307
1.057
1.299
1.524
1.039
0.823
1.259
1.322
1.045
1.960
1.372
Reviews
TNET D-ATT
1.363
1.428
1.007
1.004
1.100
0.964
1.004
1.000
1.338
1.325
1.123
1.037
1.129
1.143
0.946
0.881
1.346
1.337
0.840
0.805
1.515
1.509
1.404
1.409
1.122
1.101
1.276
1.269
0.974
0.982
1.249
1.269
1.431
1.452
0.994
0.990
0.797
0.813
1.235
1.237
1.197
1.207
1.010
1.018
1.176
1.187
1.365
1.368
MPCN
1.349
0.997
0.923
0.970
1.304
1.011
1.125
0.861
1.328
0.779
1.494
1.387
1.096
1.257
0.973
1.238
1.413
0.980
0.775
1.220
1.187
1.005
1.144
1.350
Improvement (%)
∆DC ∆T N ∆DA
2.7
1.0
5.9
28.9
1.0
0.7
60.7
19.2
4.4
23.9
3.5
3.1
10.4
2.6
1.6
51.7
11.1
2.6
6.6
0.4
1.6
31.2
9.9
2.3
9.0
1.4
0.7
16.7
7.8
3.3
1.5
1.4
1.0
4.8
1.2
1.6
10.2
2.4
0.5
4.0
1.5
1.0
8.6
0.1
0.9
4.9
0.9
2.5
7.9
1.3
2.8
6.0
1.4
1.0
6.2
2.8
4.9
3.2
1.2
1.4
11.4
0.8
1.7
4.0
0.5
1.3
71.3
2.8
3.8
1.6
1.1
1.3
Table 1: Performance comparison (mean squared error) on 24 benchmark datasets. The best performance is in boldface.
∆DC , ∆T N , ∆DA are the relative improvements (%) of MPCN over DeepCoNN (D-CON), TransNet (T-NET) and D-ATT respectively. MPCN achieves state-of-the-art performance, outperforming all existing methods on 24 benchmark datasets.
4.2
• Dual Attention CNN Model (D-ATT) [24] is a recently
proposed state-of-the-art CNN-based model that uses reviews for recommendation. This model is characterized by
its usage of two forms of attention (local and global). A final
user (item) representation is learned by concatenating representations learned from both local and global attention. The
dot product between user and item representations is then
used to estimate the rating score.
Compared Methods
We compare against a series of competitive baselines.
• Matrix Factorization (MF) is a standard and well-known
baseline for CF. It represents the user and item rating with
the inner product, i.e., p ⊤q.
• Factorization Machines (FM) [21] are general purpose machine learning algorithms that use factorized parameters to
model pairwise interactions a real-valued feature vector. We
concatenate the user-item latent embedding together and
pass it through the FM model.
• Multi-layered Perceptrons (MLP) are strong neural baselines for CF and was used as a baseline in [8]. We use the
same pyramidal scheme of 3 layers.
• Neural Matrix Factorization (NeuMF) [8] is the state-ofthe-art model for interaction-only CF. It casts the MF model
within a neural framework and combines the output with
Multi-layered perceptrons.
• Deep Co-Operative Neural Networks (DeepCoNN) [37]
is a review-based convolutional recommendation model. It
trains convolutional representations of user and items and
passes the concatenated embedding into a FM model.
• TransNet [2] is an improved adaptation of DeepCoNN which
incorporates transform layers and an additional training step
that enforces the transformed representation to be similar
to the embedding of the actual target review. user and item
review documents and passes the learned embeddings into
a FM model.
Given our already extensive comparisons against the state-of-theart, we omit comparisons with models such as HFT [17], Collaborative Topic Regression [30], Collaborative Deep Learning (CDL)
[31] and ConvMF [11] since they have been outperformed by the
recently proposed DeepCoNN or D-ATT model.
4.3
Experimental Setup
The evaluation metric is the well-known (and standard) meansquared error which measures the square error between the rating
prediction and ground truth. We implement all models ourselves in
Tensorflow. All models are trained with Adam [12] with an initial
learning rate of 10−3 . We train all models for a maximum of 20
epochs with early stopping (i.e., if model performance does not
improve for 5 epochs) and report the test result from the best performing model on the development set. We found that models tend
to converge before 20 epochs. However, an exception is that the
MF baseline requires many more epochs to converge. As such, we
train the MF model till convergence. For interaction only models,
the embedding size is set to 50. For TransNet and DeepCoNN, the
number of filters is set to 50 and the filter size is 3. For D-ATT, the
global attention layer uses filter sizes of [2, 3, 4]. The word embedding layer is also set to 50 dimensions. We regularize models with
a dropout rate of 0.2 and a fixed L2 regularization of 10−6 . Dropout
is applied after all fully-connected and convolutional. We use two
transform layers in the TransNet model. All word embeddings are
learned from scratch as we found that using pretrained embeddings
consistently degrades performance across all datasets (and models).
The maximum document length is set to 600 words (20 reviews
of 30 tokens each) which we empirically found to be a reasonable
length-specific performance bound. We assign a special delimiter
token to separate reviews within a user (item) document for DeepCoNN, TransNet and D-ATT. If FM is used, the number of factors is
set to 10. For our proposed model, the number of pointers p is tuned
amongst {1, 3, 5, 8, 10}. On most datasets, the optimal performance
is reached with 2 − 3 pointers.
4.4
Experimental Results
Table 1 reports the results of our experiments. Firstly, we observe
that our proposed MPCN is the top performing model on all 24
benchmark datasets. This ascertains the effectiveness of our proposed model and clearly answers RQ1. MPCN consistently and
significantly outperforms DeepCoNN, TransNet and D-ATT, which
are all recent competitive review-based methods for recommendation. The relative improvement is also encouraging with gains
of up to 71% (DeepCoNN), 19% (TransNet) and 5% (D-ATT). On
majority of the datasets, performance gains are modest, seeing an
improvement of 1% − 3% for most models. Notably, the average
percentage improvement of MPCN over DeepCoNN is 16%. The
average performance gain over TransNet and D-ATT is a modest
3.2% and 2.2% respectively.
Pertaining to the relative ranking of the review-based models,
our empirical evaluation reaffirms the claim of [2], showing that
TransNet always outperforms DeepCoNN. However, the relative
ranking of D-ATT and TransNet switches positions frequently.
Notably, TransNet uses the test review(s) as an additional data
source (albeit as a training target) while D-ATT does not make
use of this information. The additional training step of TransNet is
actually quite effective and hypothetically could be used to enhance
D-ATT or MPCN as well. However, we leave that for future work.
Next, the performance of interaction-only models (MF, FM etc.)
is consistently lower than review-based models (e.g., DeepCoNN).
The relative performance of all interaction models is generally quite
inconsistent over various datasets. However, one consistent fact
is that MF performs worse to models most of the time. The top
scoring interaction model often switches between FM and MLP.
Finally, we give readers a sense of computational runtime. We
provide a estimate that we found generally quite universal across
multiple datasets. Let t be the runtime of DeepCoNN, the runtime of
MPCN p = 1 is approximately ≈ 0.4t. MPCN with 2 and 3 pointers
are 0.8t and 1.2t respectively. TransNet and D-ATT runs at ≈ 2t. On
medium size datasets (e.g., Amazon Beauty), t ≈ 40s on a GTX1060
GPU (batch size is 128). We found that if popt ≤ 2, then MPCN is
faster than DeepCoNN. While popt ≤ 5 is the threshold for being
equal with D-ATT and TransNet in terms of runtime.
5
HYPERPARAMETER & ABLATION
ANALYSIS
In this section, we study the impact of key hyperparameter settings
and various architectural choices on model performance.
5.1
Ablation Analysis
We study the impacts of various architectural decisions on model
performance (RQ2). Table 2 reports an ablation analysis conducted
on the development sets of four benchmark datasets (Beauty, Office,
Musical Instruments (M-Instr) and Amazon Instant Video (Inst-Vid)).
We report the results of several different model variations. We first
begin describing the default setting in which ablation reports are
deviated from. In the default setting, we use the standard model
with all components (review gates, word-level co-attention, FM
prediction layer). The Multi-Pointer aggregation (aggr) is set to use
a neural network (single layer nonlinear transform). The number
of layers in the co-attention layer is set to l = 0.
We report the validation results from 7 different variations, with
the aims of clearly showcasing the importance of each component.
In (1) we remove the review gating mechanism. In (2), we replace
the FM with the simple inner product. In (2-3), we investigate the
effects of different pointer aggregation (aggr) functions. They are
the concatenate and additive operators respectively. In (4-5), we
set l = 0 (remove layer) and l = 2. In (6) , we remove the word
level co-attention layer. In this case, the representation for user and
item is simply the pointed review embedding. In (7), we remove
the review-level co-attention (RLCA). This variation is no longer
‘hierarchical’, and simply applies word-level co-attention to user
and item reviews.
Architecture
(0) Default
(1) Remove Gates
(2) Remove FM
(2) Agrr (Concat)
(3) Aggr (Additive)
(4) set l = 0
(5) set l = 2
(6) Remove WLCA
(7) Remove RLCA
Beauty
1.290
1.299
1.286
1.279
1.290
1.293
1.295
1.296
1.304
Office
0.770
0.760
0.808
0.788
0.767
0.776
0.775
0.778
0.789
M-Instr
0.827
0.837
0.924
0.832
0.829
0.830
0.829
0.831
0.839
Inst-Vid
0.975
0.979
0.997
0.971
0.971
0.976
0.974
0.999
0.1003
Table 2: Ablation analysis (validation MSE) on four datasets.
Firstly, we observe the default setting is not universally the best
across four datasets. As mentioned, the review gating mechanism
helps in 3 out of 4 presented datasets. In the office dataset, removing the review gating layer improves performance. We found this
to be true across most datasets, i.e., the review gating mechanism
helps more often than not, but not always. The impacts of removing
FM is quite easily noticed, leading to huge performance degradation on the M-Instr dataset. Deprovement on Inst-Vid and Office
is also significant. On the other hand, removing FM marginally
improved performance on Beauty. We also discovered that there is
no straightforward choice of aggr functions. Notably, the relative
ranking of all three variants (concat, additive and neural network)
are always interchanging across different datasets. As such, they
have to be tuned. We also noticed that the choice of l = 1 is safe
across most datasets, as increasing or decreasing would often lead to
performance degradation. Finally, removing the WLCA and RLCA
consistently lowered performance on all datasets, which ascertains
the effectiveness of the WLCA layer. Notably, removing RLCA
seems to hurt performance more, which signifies that modeling at
a review-level is essential.
5.2
Effect of Number of Pointers
Table 3 reports the effect of varying pointers on performance (RQ3).
We use 4 datasets of varying sizes as an illustrative example (Patio,
Automotive, Sports and Video Games). The datasets shown are sorted
from smallest to largest in terms of number of interactions. Clearly,
we observe that the optimal number of pointers varies across all
datasets. We found this to hold true for the remainder datasets that
are not shown. This seems to more of a domain-dependent setting
since we were not able to find any correlation with dataset size. For
most datasets, the optimal pointers falls in the range of 1 − 3. In
exceptional cases (Video Games), the optimal number of pointers
was 5.
Ptr
1
2
3
4
5
Patio
0.992
0.980
0.975
0.991
0.991
Automotive
0.842
0.855
0.843
0.844
0.844
Sports
0.926
0.933
0.938
0.938
0.935
Video Games
1.885
1.178
1.194
1.189
1.121
Table 3: Validation MSE on various datasets when varying
the number of pointers. The best result is in boldface. The
optimal number of pointers is domain-dependent.
we believe that this is trying to solicit the user’s preferences about
a type of product. We provide some qualitative examples in Table
4. Consider the context of video games, it finds that the user has
written a review about rpg (roleplaying games). At the same time,
it finds that the item review consists of a (positive) review of the
item being a good rpg game. As a result, it surfaces both reviews
and concurrently points to both of them. This follows suit for the
other examples, e.g., it finds a matching clue of puzzle games (turnbased) in the second example. The last example is taken from the
Gourmet Food dataset. In this case, it discovers that the user likes
cocoa, and concurrently finds out that the product in question has
some relevant information about chocolate.
6.2
Behavior of Multi-Pointer Learning
In this section, we study the behavior of our multi-pointer mechanism. First and foremost, this serves as another sanity check and
to observe if multi-pointers are really necessary, i.e., if pointers
are not pointing all to the same reviews. Hence, this section aims
to provide better insight into the inner workings of our proposed
model. We trained a MPCN model with four pointers. A quick observation is that all four pointers point to different reviews (given
the same user item pair). This is automatic, and does not require
any special optimization constraint (such as explicitly enforcing the
model to choose different reviews through an extra optimization
term). Moreover, we analyze the affinity matrix from the reviewlevel co-attention. Figure 3 shows the affinity matrix for pointers
one to four.
1.0
1
2
0.8
5
6
IN-DEPTH MODEL ANALYSIS
0.6
9
6.1
What are the pointers pointing to?
User Review
game is really beautiful! coolest rpg
ever...
.. game is completely turned based, so
you have time to ponder you actions..
just love this little guy ... phone is reasonably easy to put in and take out
is a nice charger..but after a few
momths it wasn’t charging...
cocoa is a wonderful, rich tasting, not
overly sweet product ..
Item Review
..a gift for a friend who really loves rpg
games. he really loved it!
..is a great and engaging puzzler game
but wasn’t too challenging
..case really fits the 5s like a glove..
it is clearly a used or refurbished battery..
used to eat the dark and milk chocolate,
and then i tried this and can’t explain
how good they are
Table 4: Excerpts from top matching User and Item reviews
extracted by MPCN’s pointer mechanism.
In this section, we list some observations by analyzing the behavior
of the MPCN model. Firstly, we observed that pointers react to
aspects and product sub-categories. In many cases, we observed
that the pointer mechanism tries to find the most relevant review
written by the user for a given product. If the target item is a phone
case, then our model tries to find a review written by the user
which is directed towards another phone case product. Intuitively,
4
7
0.6
9
8
10
10
10
11
11
11
12
12
12
0.4
0.4
13
14
15
18
0.2
17
0.0
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19 20
(a) p = 1
0.2
17
0.0
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19 20
(b) p = 2
16
0.2
17
18
19
20
0.4
15
16
18
19
20
13
14
15
16
18
19
0.4
13
14
15
0.2
17
0.6
9
12
14
0.8
5
6
8
11
13
1.0
3
0.8
5
7
0.6
9
1
2
4
6
8
10
16
In this section, we present several insights pertaining to the inner
workings of our proposed model. This aims to answer RQ4.
0.8
5
7
8
1.0
3
4
6
7
1
2
3
4
6
1.0
1
2
3
19
20
0.0
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19 20
(c) p = 3
20
0.0
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19 20
(d) p = 4
Figure 3: Visualisation of Review-level Co-Attention. Matching patterns differ significantly across multiple calls, hence
generating different pointers
Secondly, it is also intuitive that it is not absolutely necessary
for MPCN to always point to unique reviews given the same useritem pair. We observed a relatively significant one-to-many pointer
pattern on top of the usual one-to-one pattern. In this case, the same
review for user (item) is being matched with n (many) different
reviews from the item (user). This is also observed to be dataset /
domain dependent. In a small minority of cases, all pointers pointed
to same reviews constantly (all repeated condition). However, this
is understandable as there is just insufficient information in the user
and item review bank. Additionally, we analyzed a small sample
from the test set, determining if any of the following conditions
hold for each test case.
Table 5 reports the percentages of test samples which falls into
each category. We report results on five datasets Digital Music
(D-Music), Android Apps (Apps), Video Games (V-Games), Gourmet
Food (Food) and Yelp17. Here, we observe that pointer behavior is
largely influenced by domain. The first three are concerned with
electronic domains while the last two are more focused on food
Condition
(1) All unique
(2) 1 Repeated
(3) All Repeated
(4) One-to-Many
D-Music
85.2%
13.2%
1.6%
64.8%
Apps
87.5%
12.5%
0.0%
57.8%
V-Games
82.0%
17.2%
1.6%
57.8%
Food
99.2%
0.7%
0.0%
23.4%
Yelp17
99.2%
0.7%
0.0%
43.7%
Table 5: Analysis of Multi-Pointer Behavior of MPCN on five
datasets using np = 3.
(and restaurants). We clearly observe that Food and Yelp have very
similar pointer behavior. In general, electronic problem domains
usually make an inference using a select few subset of reviews.
This is made evident by the high one-to-many ratio which signifies
that there is often one important review written by the user (or
item) that contributes more (and needs to be matched with multiple
opposing reviews). Conversely, food problem domains require more
evidences across multiple reviews. We believe this is one of the
biggest insights that our work offers, i.e., shedding light on how
evidence aggregation works (and differs across domains) in reviewbased recommendation.
7
CONCLUSION
We proposed a new state-of-the-art neural model for recommendation with reviews. Our proposed Multi-Pointer Co-Attention
Networks outperforms many strong competitors on 24 benchmark
datasets from Amazon and Yelp. We conduct extensive analysis
on the inner workings of our proposed multi-pointer learning
mechanism. By analyzing the pointer behavior across multiple
domains, we conclude that different domains (such as food-related
and electronics-related) have different ‘evidence aggregation’ patterns. While our model dynamically handles this aspect, we believe
this warrants further investigation.
REFERENCES
[1] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly learning to align and translate. arXiv preprint
arXiv:1409.0473 (2014).
[2] Rose Catherine and William Cohen. 2017. TransNets: Learning to Transform for
Recommendation. arXiv preprint arXiv:1704.02298 (2017).
[3] Jihun Choi, Kang Min Yoo, and Sang-goo Lee. 2017. Unsupervised learning of
task-specific tree structures with tree-lstms. arXiv preprint arXiv:1707.02786
(2017).
[4] Qiming Diao, Minghui Qiu, Chao-Yuan Wu, Alexander J. Smola, Jing Jiang, and
Chong Wang. 2014. Jointly modeling aspects, ratings and sentiments for movie
recommendation (JMARS). In The 20th ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining, KDD ’14.
[5] Cícero Nogueira dos Santos, Ming Tan, Bing Xiang, and Bowen Zhou. 2016.
Attentive Pooling Networks. CoRR abs/1602.03609 (2016).
[6] Ruining He and Julian McAuley. 2016. Ups and downs: Modeling the visual
evolution of fashion trends with one-class collaborative filtering. In Proceedings
of the 25th International Conference on World Wide Web. International World Wide
Web Conferences Steering Committee, 507–517.
[7] Xiangnan He and Tat-Seng Chua. 2017. Neural Factorization Machines for Sparse
Predictive Analytics. (2017).
[8] Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng
Chua. 2017. Neural Collaborative Filtering. In Proceedings of the 26th International
Conference on World Wide Web (WWW ’17). 173–182.
[9] Xiangnan He, Hanwang Zhang, Min-Yen Kan, and Tat-Seng Chua. [n. d.]. Fast
Matrix Factorization for Online Recommendation with Implicit Feedback. In
Proceedings of the 39th International ACM SIGIR conference on Research and Development in Information Retrieval, SIGIR 2016, Pisa, Italy, July 17-21, 2016.
[10] Eric Jang, Shixiang Gu, and Ben Poole. 2016. Categorical reparameterization
with gumbel-softmax. arXiv preprint arXiv:1611.01144 (2016).
[11] Donghyun Kim, Chanyoung Park, Jinoh Oh, Sungyoung Lee, and Hwanjo Yu.
2016. Convolutional Matrix Factorization for Document Context-Aware Recommendation. In Proceedings of the 10th ACM Conference on Recommender Systems
(RecSys ’16). ACM, New York, NY, USA.
[12] Diederik P. Kingma and Jimmy Ba. 2014. Adam: A Method for Stochastic Optimization. CoRR abs/1412.6980 (2014).
[13] Piji Li, Zihao Wang, Zhaochun Ren, Lidong Bing, and Wai Lam. 2017. Neural
Rating Regression with Abstractive Tips Generation for Recommendation. (2017).
[14] Xiaopeng Li and James She. 2017. Collaborative Variational Autoencoder for
Recommender Systems. In Proceedings of the 23rd ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining (KDD ’17).
[15] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. 2016. Hierarchical
question-image co-attention for visual question answering. In Advances In Neural
Information Processing Systems. 289–297.
[16] Chris J Maddison, Daniel Tarlow, and Tom Minka. 2014. A* sampling. In Advances
in Neural Information Processing Systems. 3086–3094.
[17] Julian McAuley and Jure Leskovec. 2013. Hidden factors and hidden topics:
understanding rating dimensions with review text. In Proceedings of the 7th ACM
conference on Recommender systems. ACM, 165–172.
[18] Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel.
2015. Image-based recommendations on styles and substitutes. In Proceedings
of the 38th International ACM SIGIR Conference on Research and Development in
Information Retrieval. ACM, 43–52.
[19] Andriy Mnih and Ruslan R Salakhutdinov. 2008. Probabilistic matrix factorization.
In Advances in neural information processing systems. 1257–1264.
[20] Ankur P. Parikh, Oscar Täckström, Dipanjan Das, and Jakob Uszkoreit. 2016. A
Decomposable Attention Model for Natural Language Inference. In Proceedings of
the 2016 Conference on Empirical Methods in Natural Language Processing, EMNLP
2016, Austin, Texas, USA, November 1-4, 2016. 2249–2255.
[21] Steffen Rendle. 2010. Factorization Machines. In ICDM 2010, The 10th IEEE
International Conference on Data Mining, Sydney, Australia, 14-17 December 2010.
995–1000.
[22] Tim Rocktäschel, Edward Grefenstette, Karl Moritz Hermann, Tomáš Kočiskỳ,
and Phil Blunsom. 2015. Reasoning about entailment with neural attention. arXiv
preprint arXiv:1509.06664 (2015).
[23] Suvash Sedhain, Aditya Krishna Menon, Scott Sanner, and Lexing Xie. 2015.
Autorec: Autoencoders meet collaborative filtering. In Proceedings of the 24th
International Conference on World Wide Web. ACM, 111–112.
[24] Sungyong Seo, Jing Huang, Hao Yang, and Yan Liu. 2017. Interpretable Convolutional Neural Networks with Dual Local and Global Attention for Review
Rating Prediction. In Proceedings of the Eleventh ACM Conference on Recommender
Systems (RecSys ’17).
[25] Yi Tay, Anh Tuan Luu, and Siu Cheung Hui. 2017. Translational recommender
networks. arXiv preprint arXiv:1707.05176 (2017).
[26] Yi Tay, Minh C. Phan, Anh Tuan Luu, and Siu Cheung Hui. [n. d.]. Learning
to Rank Question Answer Pairs with Holographic Dual LSTM Architecture.
In Proceedings of the 40th International ACM SIGIR Conference on Research and
Development in Information Retrieval, Shinjuku, Tokyo, Japan, August 7-11, 2017.
[27] Yi Tay, Luu Anh Tuan, and Siu Cheung Hui. 2017. A Compare-Propagate Architecture with Alignment Factorization for Natural Language Inference. (2017).
arXiv:arXiv:1801.00102
[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones,
Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention Is All
You Need. arXiv preprint arXiv:1706.03762 (2017).
[29] Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. 2015. Pointer networks. In
Advances in Neural Information Processing Systems. 2692–2700.
[30] Chong Wang and David M Blei. 2011. Collaborative topic modeling for recommending scientific articles. In Proceedings of the 17th ACM SIGKDD international
conference on Knowledge discovery and data mining. ACM, 448–456.
[31] Hao Wang, Naiyan Wang, and Dit-Yan Yeung. 2015. Collaborative deep learning
for recommender systems. In Proceedings of the 21th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining. ACM, 1235–1244.
[32] Shuohang Wang and Jing Jiang. 2016. Machine comprehension using match-lstm
and answer pointer. arXiv preprint arXiv:1608.07905 (2016).
[33] Ronald J Williams. 1992. Simple statistical gradient-following algorithms for
connectionist reinforcement learning. Machine learning 8, 3-4 (1992), 229–256.
[34] Chao-Yuan Wu, Amr Ahmed, Alex Beutel, Alexander J. Smola, and How Jing.
2017. Recurrent Recommender Networks. In Proceedings of the Tenth ACM
International Conference on Web Search and Data Mining, WSDM 2017, Cambridge,
United Kingdom, February 6-10, 2017. 495–503.
[35] Caiming Xiong, Victor Zhong, and Richard Socher. 2016. Dynamic Coattention
Networks For Question Answering. CoRR abs/1611.01604 (2016).
[36] Xiaodong Zhang, Sujian Li, Lei Sha, and Houfeng Wang. 2017. Attentive Interactive Neural Networks for Answer Selection in Community Question Answering.
In Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, February 4-9, 2017, San Francisco, California, USA. 3525–3531.
[37] Lei Zheng, Vahid Noroozi, and Philip S. Yu. 2017. Joint Deep Modeling of Users
and Items Using Reviews for Recommendation. In Proceedings of the Tenth ACM
International Conference on Web Search and Data Mining, WSDM 2017. 425–434.
| 2cs.AI
|
Visual7W: Grounded Question Answering in Images
arXiv:1511.03416v4 [cs.CV] 9 Apr 2016
Yuke Zhu†
Oliver Groth‡
Michael Bernstein†
Li Fei-Fei†
†
Computer Science Department, Stanford University
‡
Computer Science Department, Dresden University of Technology
Abstract
Where
does this
scene take
place?
We have seen great progress in basic perceptual tasks
such as object recognition and detection. However, AI models still fail to match humans in high-level vision tasks due
to the lack of capacities for deeper reasoning. Recently the
new task of visual question answering (QA) has been proposed to evaluate a model’s capacity for deep image understanding. Previous works have established a loose, global
association between QA sentences and images. However,
many questions and answers, in practice, relate to local
regions in the images. We establish a semantic link between textual descriptions and image regions by object-level
grounding. It enables a new type of QA with visual answers,
in addition to textual answers used in previous work. We
study the visual QA tasks in a grounded setting with a large
collection of 7W multiple-choice QA pairs. Furthermore,
we evaluate human performance and several baseline models on the QA tasks. Finally, we propose a novel LSTM
model with spatial attention to tackle the 7W QA tasks.
Why is there
foam?
A) Because of a wave. ✔
B) Because of a boat.
C) Because of a fire.
D) Because of a leak.
A) In the sea. ✔
B) In the desert.
C) In the forest.
D) On a lawn.
What is
the dog
doing?
What is the dog
standing on?
A) Surfing. ✔
B) Sleeping.
C) Running.
D) Eating.
A) On a surfboard. ✔
B) On a table.
C) On a garage.
D) On a ball.
Which paw is lifted?
Figure 1: Deep image understanding relies on detailed knowledge about different image parts. We employ diverse questions to
acquire detailed information on images, ground objects mentioned
in text with their visual appearances, and provide a multiple-choice
setting for evaluating a visual question answering task with both
textual and visual answers.
suggest, a tighter semantic link between textual descriptions and corresponding visual regions is a key ingredient
for better models. As Fig. 1 shows, the localization of objects can be a critical step to understand images better and
solve image-related questions. Providing these image-text
correspondences is called grounding. Inspired by Geman
et al.’s prototype of a visual Turing test based on image
regions [8] and the comprehensive data collection of QA
pairs on COCO images [25] such as VQA [1] and Baidu [7],
we fuse visual QA and grounding in order to create a new
QA dataset with dense annotations and a more flexible
evaluation environment. Object-level grounding provides
a stronger link between QA pairs and images than global
image-level associations. Furthermore, it allows us to resolve coreference ambiguity [19, 35] and to understand object distributions in QA, and enables visually grounded answers that consist of object bounding boxes.
Motivated by the goal of developing a model for visual QA based on grounded regions, our paper introduces
a dataset that extends previous approaches [1, 7, 36] and
proposes an attention-based model to perform this task. We
collected 327,939 QA pairs on 47,300 COCO images [25],
together with 1,311,756 human-generated multiple-choices
and 561,459 object groundings from 36,579 categories. Our
data collection was inspired by the age-old idea of the W
1. Introduction
The recent development of deep learning technologies
has achieved successes in many perceptual visual tasks such
as object recognition, image classification and pose estimation [15, 21, 25, 38, 39, 42, 43]. Yet the status quo of
computer vision is still far from matching human capabilities, especially when it comes to understanding an image
in all its details. Recently, visual question answering (QA)
has been proposed as a proxy task for evaluating a vision
system’s capacity for deeper image understanding. Several
QA datasets [1, 7, 27, 36, 49] have been released since last
year. They contributed valuable data for training visual QA
systems and introduced various tasks, from picking correct
multiple-choice answers [1] to filling in blanks [49].
Pioneer work in image captioning [4, 5, 13, 45, 47],
sentence-based image retrieval [14, 40] and visual QA [1, 7,
36] shows promising results. These works aimed at establishing a global association between sentences and images.
However, as Flickr30K [34, 48] and Visual Madlibs [49]
1
questions in journalism to describe a complete story [22].
The 7W questions roughly correspond to an array of standard vision tasks: what [9, 15, 39], where [24, 50],
when [30, 32], who [35, 42], why [33], how [23, 31] and
which [16, 17]. The Visual7W dataset features richer questions and longer answers than VQA [1]. In addition, we
provide complete grounding annotations that link the object
mentions in the QA sentences to their bounding boxes in the
images and therefore introduce a new QA type with image
regions as the visually grounded answers. We refer to questions with textual answers as telling questions (what, where,
when, who, why and how) and to such with visual answers
as pointing questions (which). We provide a detailed comparison and data analysis in Sec. 4.
A salient property of our dataset is the notable gap
between human performance (96.6%) and state-of-the-art
LSTM models [28] (52.1%) on the visual QA tasks. We add
a new spatial attention mechanism to an LSTM architecture
for tackling the visually grounded QA tasks with both textual and visual answers (see Sec. 5). The model aims to
capture the intuition that answers to image-related questions
usually correspond with specific image regions. It learns to
attend to the pertinent regions as it reads the question tokens
in a sequence. We achieve state-of-the-art performance with
55.6%, and find correlations between the model’s attention
heat maps and the object groundings (see Sec. 6). Due to
the large performance gap between human and machine, we
envision our dataset and visually grounded QA tasks to contribute to a long-term joint effort from several communities
such as vision, natural language processing and knowledge
to close the gap together.
The Visual7W dataset constitutes a part of the Visual
Genome project [20]. Visual Genome contains 1.7 million
QA pairs of the 7W question types, which offers the largest
visual QA collection to date for training models. The QA
pairs in Visual7W are a subset of the 1.7 million QA pairs
from Visual Genome. Moreover, Visual7W includes extra
annotations such as object groundings, multiple choices and
human experiments, making it a clean and complete benchmark for evaluation and analysis.
2. Related Work
Vision + Language. There have been years of effort in connecting the visual and textual information for joint learning [2, 19, 33, 35, 40, 52]. Image and video captioning has
become a popular task in the past year [4, 5, 13, 37, 45, 47].
The goal is to generate text snippets to describe the images
and regions instead of just predicting a few labels. Visual
question answering is a natural extension to the captioning
tasks, but is more interactive and has a stronger connection
to real-world applications [3].
Text-based question answering. Question answering in
NLP has been a well-established problem. Successful applications can be seen in voice assistants in mobile devices, search engines and game shows (e.g., IBM Waston).
Traditional question answering system relies on an elaborate pipeline of models involving natural language parsing,
knowledge base querying, and answer generation [6]. Recent neural network models attempt to learn end-to-end directly from questions and answers [12, 46].
Visual question answering. Geman et al. [8] proposed a
restricted visual Turing test to evaluate visual understanding. The DAQUAR dataset is the first toy-sized QA benchmark built upon indoor scene RGB-D images. Most of the
other datasets [1, 7, 36, 49] collected QA pairs on Microsoft
COCO images [25], either generated automatically by NLP
tools [36] or written by human workers [1, 7, 49]. Following these datasets, an array of models has been proposed
to tackle the visual QA tasks. The proposed models range
from probabilistic inference [27, 44, 51] and recurrent neural networks [1, 7, 28, 36] to convolutional networks [26].
Previous visual QA datasets evaluate textual answers on images while omitting the links between the object mentions
and their visual appearances. Inspired by Geman et al. [8],
we establish the link by grounding objects in the images and
perform experiments in the grounded QA setting.
3. Creating the Visual7W Dataset
We elaborate on the details of the data collection we
conducted upon 47,300 images from COCO [25] (a subset of images from Visual Genome [20]). We leverage the
six W questions (what, where, when, who, why, and how)
to systematically examine a model’s capability for visual
understanding, and append a 7th which question category.
This extends existing visual QA setups [1, 7, 36] to accommodate visual answers. We standardize the visual QA
tasks with multi-modal answers in a multiple-choice format.
Each question comes with four answer candidates, with one
being the correct answer. In addition, we ground all the
objects mentioned in the QA pairs to their corresponding
bounding boxes in the images. The object-level groundings
enable examining the object distributions and resolve the
coreference ambiguity [19, 35].
3.1. Collecting the 7W Questions
The data collection tasks are conducted on Amazon Mechanical Turk (AMT), an online crowdsourcing platform.
The online workers are asked to write pairs of question and
answer based on image content. We instruct the workers to
be concise and unambiguous to avoid wordy or speculative
questions. To obtain a clean set of high-quality QA pairs,
we ask three AMT workers to label each pair as good or bad
independently. The workers judge each pair by whether an
average person is able to tell the answer when seeing the
Q: What endangered animal
is featured on the truck?
Q: Where will the driver go
if turning right?
Q: When was the picture
taken?
Q: Who is under the
umbrella?
A:
A:
A:
A:
A:
A:
A:
A:
A:
A:
A:
A:
A:
A:
A:
A:
A bald eagle.
A sparrow.
A humming bird.
A raven.
Q: Which pillow is farther
from the window?
Onto 24 ¾ Rd.
Onto 25 ¾ Rd.
Onto 23 ¾ Rd.
Onto Main Street.
Q: Which step leads to the
tub?
During a wedding.
During a bar mitzvah.
During a funeral.
During a Sunday church
service.
Q: Which is the small
computer in the corner?
Two women.
A child.
An old man.
A husband and a wife.
Q: Which item is used to
cut items?
Q: Why was the hand of the
woman over the left
shoulder of the man?
Q: How many magnets are
on the bottom of the
fridge?
A: They were together and
engaging in affection.
A: The woman was trying to
get the man’s attention.
A: The woman was trying to
scare the man.
A: The woman was holding
on to the man for balance.
A:
A:
A:
A:
Q: Which doughnut has
multicolored sprinkles?
Q: Which man is wearing the
red tie?
5.
2.
3.
4.
Figure 2: Examples of multiple-choice QA from the 7W question categories. The first row shows telling questions where the green answer
is the ground-truth, and the red ones are human-generated wrong answers. The what, who and how questions often pertain to recognition
tasks with spatial reasoning. The where, when and why questions usually involve high-level common sense reasoning. The second row
depicts pointing (which) questions where the yellow box is the correct answer and the red boxes are human-generated wrong answers.
These four answers form a multiple-choice test for each question.
image. We accept the QA pairs with at least two positive
votes. We notice varying acceptance rates between categories, ranging from 92% for what to 63% for why. The
overall acceptance rate is 85.8%.
VQA [1] relied on both human workers and automatic
methods to generate a pool of candidate answers. We find
that human-generated answers produce the best quality; on
the contrary, automatic methods are prone to introducing
candidate answers paraphrasing the ground-truth answers.
For the telling questions, the human workers write three
plausible answers to each question without seeing the image. To ensure the uniqueness of correct answers, we provide the ground-truth answers to the workers, and instruct
them to write answers of different meanings. For the pointing questions, the workers draw three bounding boxes of
other objects in the image, ensuring that these boxes cannot
be taken as the correct answer. We provide examples from
the 7W categories in Fig. 2.
3.2. Collecting Object-level Groundings
We collect object-level groundings by linking the object
mentions in the QA pairs to their bounding boxes in the images. We ask the AMT workers to extract the object mentions from the QA pairs and draw boxes on the images. We
collect additional groundings for the multiple choices of the
pointing questions. Duplicate boxes are removed based on
the object names with an Intersection-over-Union threshold
question ambiguity
What is the man wearing?
A black shirt.
answer ambiguity
What is the vehicle with ads on it?
A red bus.
Figure 3: Coreference ambiguity arises when an object mention
has multiple correspondences in an image, and the textual context
is insufficient to tell it apart. The answer to the left question can
be either gray, yellow or black, depending on which man is meant.
In the right example, the generic phrase red bus can refer to both
buses in the image. Thus an algorithm might answer correctly even
if referring to the wrong bus.
of 0.5. In total, we have collected 561,459 object bounding
boxes, on average 12 boxes per image.
The benefits of object-level groundings are three-fold: 1)
it resolves the coreference ambiguity problem between QA
sentences and images; 2) it extends the existing visual QA
setups to accommodate visual answers; and 3) it offers a
means to understand the distribution of objects, shedding
light on the essential knowledge to be acquired for tackling
the QA tasks (see Sec. 4).
We illustrate examples of coreference ambiguity in
Fig. 3. Ambiguity might cause a question to have more than
Table 1: Comparisons on Existing Visual Question Answering Datasets
DAQUAR [27]
Visual Madlibs [49]
COCO-QA [36]
Baidu [7]
VQA [1]
Visual7W (Ours)
# QA
12,468
56,468
117,684
316,193
614,163
327,939
# Images
1,447
9,688
69,172
316,193
204,721
47,300
AvgQLen
11.5 ± 2.4
4.9 ± 2.4
8.7 ± 2.7
6.2 ± 2.0
6.9 ± 2.4
AvgALen
1.2 ± 0.5
2.8 ± 2.0
1.0 ± 0
1.1 ± 0.4
2.0 ± 1.4
LongAns
3.4%
47.4%
0.0%
3.8%
27.6%
TopAns
96.4%
57.9%
100%
82.7%
63.5%
HumanPerf
COCO
X
X
X
X
X
X
MC
X
X
Human-Machine Performance Gap We expect that a
good QA benchmark should exhibit a sufficient performance gap between humans and state-of-the-art models,
leaving room for future research to explore. Additionally
a nearly perfect human performance is desired to certify
the quality of its questions. On Visual7W, we conducted
two experiments to measure human performance (cf. Table 1, HumanPerf ), as well as examining the percentage of
questions that can be answered without images. Our results
show both strong human performance and a strong interdependency between images and QA pairs. We provide the
1 We report the statistics of VQA dataset [1] with its real images and
Visual Madlibs [49] with its filtered hard tasks. The fill-in-the-blank tasks
in Visual Madlibs [49], where the answers are sentence fragments, differ
from other QA tasks, resulting in distinct statistics. We omit some statistics
for Baidu [7] due to its partial release.
X
X
Top 20 objects
man
trees
shirt
wall
grass
tree
woman
people
sky
building
4. Comparison and Analysis
Advantages of Grounding The unique feature of our Visual7W dataset is the grounding annotations of all textually mentioned objects (cf. Table 1, Grounding). In total
we have collected 561,459 object groundings, which enables the new type of visual answers in the form of bounding boxes (cf. Table 1, VisualAns). Examining the object
distribution in the QA pairs sheds light on the focus of the
questions and the essential knowledge to be acquired for answering them. Our object groundings spread across 36,579
categories (distinct object names), thereby exhibiting a long
tail pattern where 85% of the categories have fewer than 5
instances (see Fig. 4). The open-vocabulary annotations of
objects, in contrast with traditional image datasets focusing
on predefined categories and salient objects [25, 38], provide a broad coverage of objects in the images.
VisualAns
X
one plausible answers at test time, thus complicating evaluation. Our online study shows that, such ambiguity occurs
in 1% of the accepted questions and 7% of the accepted answers. This illustrates a drawback of existing visual QA
setups [1, 7, 27, 36, 49], where in the absence of objectlevel groundings the textual questions and answers are only
loosely coupled to the images.
In this section, we analyze our Visual7W dataset collected on COCO images (cf. Table 1, COCO), present its
key features, and provide comparisons of our dataset with
previous work. We summarize important metrics of existing
visual QA datasets in Table 1.1
Grounding
person
window
table
head
water
snow
hair
sign
ground
wood
Figure 4: Object distribution in telling and pointing QA. The rank
of an object category is based on its frequency with rank #1 referring to the most frequent one. The pointing QA pairs cover an
order of magnitude more objects than the telling QA pairs. The
top 20 object categories indicate that the object distribution’s bias
towards persons, daily-life objects and natural entities.
detailed analysis and comparisons with the state-of-the-art
automatic models in Sec. 6.
Table 2: Model and Human Performances on QA Datasets
VQA (open-ended) [1]
VQA (multiple-choice) [1]
Facebook bAbI [46]
Ours (telling QA)
Ours (pointing QA)
Model
0.54
0.57
0.92
0.54
0.56
Human
0.83
0.92
∼1.0
0.96
0.97
∆
0.29
0.35
0.08
0.42
0.41
Table 2 compares Visual7W with VQA [1] and Facebook
bAbI [46], which have reported model and human performances. Facebook bAbI [46] is a textual QA dataset claiming that humans can potentially achieve 100% accuracy yet
without explicit experimental proof. For VQA [1], numbers
are reported for both multiple-choice and open-ended evaluation setups. Visual7W features the largest performance gap
(∆), a desirable property for a challenging and long-lasting
evaluation task. At the same time, the nearly perfect human
performance proves high quality of the 7W questions.
QA Diversity The diversity of QA pairs is an important
feature of a good QA dataset as it reflects a broad coverage
of image details, introduces complexity and potentially requires a broad range of skills for solving the questions. To
obtain diverse QA pairs, we decided to rule out binary questions, contrasting Geman et al.’s proposal [8] and VQA’s
approach [1]. We hypothesize that this encourages workers
to write more complex questions and also prevents inflating
answer baselines with simple yes/no answers.
softmax
5.1. Recurrent QA Models with Spatial Attention
LSTM
CNN
CNN
which
is
the
brown
bread
?
ht-1
convolutional
feature maps C(I)
attention terms at
Q: Which is the brown bread?
A1:
A2:
A3:
A4:
Figure 5: Diagram of the recurrent neural network model for
pointing QA. At the encoding stage, the model reads the image
and the question tokens word by word. At each word, it computes
attention terms based on the previous hidden state and the convolutional feature map, deciding which regions to focus on. At the
decoding stage, it computes the log-likelihood of an answer by a
dot product between its transformed visual feature (fc7) and the
last LSTM hidden state.
When examining the richness of QA pairs, the length of
questions and answers (cf. Table 1, AvgQLen, AvgALen) is
a rough indicator for the amount of information and complexity they contain. The overall average question and
answer lengths are 6.9 and 2.0 words respectively. The
pointing questions have the longest average question length.
The telling questions exhibit a long-tail distribution where
51.2%, 21.2%, and 16.6% of their answers have one, two
or three words respectively. Many answers to where and
why questions are phrases and sentences, with an average
of 3 words. In general, our dataset features long answers
where 27.6% of the questions have answers of more than
two words (cf. Table 1, LongAns). In contrast, 89% of answers in VQA [1], 90% of answers in DAQUAR [27] and
all answers in COCO-QA [36] are a single word. We also
capture more long-tail answers as our 1,000 most frequent
answers only account for 63.5% of all our answers (cf. Table 1, TopAns). Finally we provide human created multiplechoices for evaluation (cf. Table 1, MC).
5. Attention-based Model for Grounded QA
The visual QA tasks are visually grounded, as local image regions are pertinent to answering questions in many
cases. For instance, in the first pointing QA example of
Fig. 2 the regions of the window and the pillows reveal the
answer, while other regions are irrelevant to the question.
We capture this intuition by introducing a spatial attention
mechanism similar to the model for image captioning [47].
LSTM models [11] have achieved state-of-the-art results
in several sequence processing tasks [5, 13, 41]. They have
also been used to tackle visual QA tasks [1, 7, 28]. These
models represent images by their global features, lacking a
mechanism to understand local image regions. We add spatial attention [10, 47] to the standard LSTM model for visual
QA, illustrated in Fig. 5. We consider QA as a two-stage
process [7, 28]. At the encoding stage, the model memorizes the image and the question into a hidden state vector
(the gray box in Fig. 5). At the decoding stage, the model
selects an answer from the multiple choices based on its
memory (the softmax layer in Fig. 5). We use the same encoder structure for all visual QA tasks but different decoders
for the telling and pointing QA tasks. Given an image I and
a question Q = (q1 , q2 , . . . , qm ), we learn the embeddings
of the image and the word tokens as follow:
v0
= Wi [F (I)] + bi
(1)
vi
= Ww [OH(ti )], i = 1, . . . , m
(2)
where F (·) transforms an image I from pixel space to a
4096-dimensional feature representation. We extract the
activations from the last fully connected layer (fc7) of a
pre-trained CNN model VGG-16 [39]. OH(·) transforms
a word token to its one-hot representation, an indicator column vector where there is a single one at the index of the
token in the word vocabulary. The Wi matrix transforms the
4096-dimensional image features into the di -dimensional
embedding space v0 , and the Ww transforms the one-hot
vectors into the dw -dimensional embedding space vi . We
set di and dw to the same value of 512. Thus, we take the
image as the first input token. These embedding vectors
v0,1,...,m are fed into the LSTM model one by one. The
update rules of our LSTM model can be defined as follow:
it
=
σ(Wvi vt + Whi ht−1 + Wri rt + bi )
(3)
ft
=
σ(Wvf vt + Whf ht−1 + Wrf rt + bf )
(4)
ot
=
σ(Wvo vt + Who ht−1 + Wro rt + bo )
(5)
gt
=
φ(Wvg vt + Whg ht−1 + Wrg rt + bg )
(6)
ct
=
ft
ct−1 + it
(7)
ht
=
ot
φ(ct )
gt
(8)
where σ(·) is the sigmoid function, φ(·) is the tanh function,
and is the element-wise multiplication operator. The attention mechanism is introduced by the term rt , which is
a weighted average of convolutional features that depends
upon the previous hidden state and the convolutional features. The exact formulation is as follows:
et
=
waT tanh(Whe ht−1 + Wce C(I)) + ba
(9)
at
=
softmax(et )
(10)
rt
=
aTt C(I)
(11)
Figure 6: Response time of human subjects on the telling QA
tasks. The boxes go from the first quartile to the third quartile
of the response time values. The bars in the centers of the boxes
indicate the median response time of each category.
where C(I) returns the 14 × 14 512-dimensional convolutional feature maps of image I from the fourth convolutional layer from the same VGG-16 model [39]. The attention term at is a 196-dimensional unit vector, deciding the
contribution of each convolutional feature at the t-th step.
The standard LSTM model can be considered as a special
case with each element in at set uniformly. Wi , bi , Ww and
all the W s and bs in the LSTM model and attention terms
are learnable parameters.
5.2. Learning and Inference
The model first reads the image v0 and all the question
tokens vq1 , vq2 , . . . , vqm until reaching the question mark
(i.e., end token of the question sequence). When training
for telling QA, we continue to feed the ground-truth answer
tokens va1 , va2 , . . . , van into the model. For pointing QA,
we compute the log-likelihood of an candidate region by
a dot product between its transformed visual feature (fc7)
and the last LSTM hidden state (see Fig. 5). We use crossentropy loss to train the model parameters with backpropagation. During testing, we select the candidate answer with
the largest log-likelihood. We set the hyperparameters using the validation set. The dimensions of the LSTM gates
and memory cells are 512 in all the experiments. The model
is trained with Adam update rule [18], mini-batch size 128,
and a global learning rate of 10−4 .
6. Experiments
We evaluate the human and model performances on the
QA tasks. We report a reasonably challenging performance
delta leaving sufficient room for future research to explore.
6.1. Experiment Setups
As the 7W QA tasks have been formulated in a multiplechoice format, we use the same procedure to evaluate hu-
man and model performances. At test time, the input is an
image and a natural language question, followed by four
multiple choices. In telling QA, the multiple choices are
written in natural language; whereas, in pointing QA, each
multiple choice corresponds to an image region. We say
the model is correct on a question if it picks the correct answer among the candidates. Accuracy is used to measure
the performance. An alternative method to evaluate telling
QA is to let the model predict open-ended text outputs [1].
This approach works well on short answers; however, it performs poorly on long answers, where there are many ways
of paraphrasing the same meaning. We make the training,
validation and test splits, each with 50%, 20%, 30% of the
pairs respectively. The numbers are reported on the test set.
6.2. 7W QA Experiments
6.2.1
Human Experiments on 7W QA
We evaluate human performances on the multiple-choice
7W QA. We want to measure in these experiments 1) how
well humans can perform in the visual QA task and 2)
whether humans can use common sense to answer questions
without seeing the images.
We conduct two sets of human experiments. In the first
experiment (Question), a group of five AMT workers are
asked to guess the best possible answers from the multiple choices without seeing the images. In the second experiment (Question + Image), we have a different group of
five workers to answer the same questions given the images.
The first block in Table 3 reports the human performances
on these experiments. We measure the mean accuracy over
the QA pairs where we take the majority votes among the
five human responses. Even without the images, humans
manage to guess the most plausible answers in some cases.
Human subjects achieve 35.3% accuracy, 10% higher than
chance. The human performance without images is remarkably high (43.9%) for the why questions, indicating that
many why questions encode a fair amount of common sense
that humans are able to infer without visual cue. However,
images are important in the majority of the questions. Human performance is significantly improved when the images are provided. Overall, humans achieve a high accuracy
of 96.6% on the 7W QA tasks.
Fig. 6 shows the box plots of response time of the human subjects for telling QA. Human subjects spend double
the time to respond when the images are displayed. In addition, why questions take a longer average response time
compared to the other five question types. Human subjects spend an average of 9.3 seconds on pointing questions.
However, that experiment was conducted in a different user
interface, where workers click on the answer boxes in the
image. Thus, the response time is not comparable with the
telling QA tasks. Interestingly, longer response time does
not imply higher performance. Human subjects spend more
Overall
0.353
0.966
0.383
0.305
0.352
0.462
0.359
0.521
0.556
w/ Image w/o Image Multiple Choices
Image
Table 3: Human and model performances in the multiple-choice 7W QA tasks (in accuracy)
Telling
Pointing
Method
What Where When Who
Why
How
Which
Human (Question)
0.356
0.322
0.393 0.342 0.439 0.337
Human (Question + Image)
0.965
0.957
0.944 0.965 0.927 0.942
0.973
Logistic Regression (Question)
0.420
0.375
0.666 0.510 0.354 0.458
0.354
Logistic Regression (Image)
0.408
0.426
0.438 0.415 0.337 0.303
0.256
Logistic Regression (Question + Image) 0.429
0.454
0.621 0.501 0.343 0.356
0.307
LSTM (Question)
0.430
0.414
0.693 0.538 0.491 0.484
LSTM (Image)
0.422
0.497
0.660 0.523 0.475 0.468
0.299
LSTM (Question + Image) [28]
0.489
0.544
0.713 0.581 0.513 0.503
0.521
Ours, LSTM-Att (Question + Image)
0.515
0.570
0.750 0.595 0.555 0.498
0.561
Q: Who is behind the
batter?
Q: What adorns the
tops of the post?
Q: How many cameras
are in the photo?
Q: Why is there rope?
Q: What kind of stuffed Q: What animal is being
animal is shown?
petted?
A: Catcher.
A: Umpire.
A: Fans.
A: Ball girl.
A: Gulls.
A: An eagle.
A: A crown.
A: A pretty sign.
A: One.
A: Two.
A: Three.
A: Four.
A: To tie up the boats.
A: To tie up horses.
A: To hang people.
A: To hit tether balls.
A: Teddy Bear.
A: Monkey.
A: Tiger.
A: Bunny rabbit.
A: A sheep.
A: Goat.
A: Alpaca.
A: Pig.
H: Catcher. ✓
H: Gulls.
H: To hit tether balls. ✗
H: Monkey. ✗
M: Gulls.
✓
✓
H: Three. ✗
✗
M: One. ✓
M: To hang people. ✗
M: Teddy Bear.
✓
M: A sheep. ✓
H: Catcher. ✓
H: Gulls.
✓
H: One. ✓
H: To tie up the boats. ✓
H: Teddy Bear.
M: A crown. ✗
M: One. ✓
M: To hang people. ✗
M: Teddy Bear.
✓
✓
H: Goat. ✗
M: Catcher. ✓
M: Umpire.
H: A sheep. ✓
M: A sheep. ✓
Figure 7: Qualitative results of human subjects and the state-of-the-art model (LSTM-Att) on multiple-choice QAs. We illustrate the
prediction results of six multiple-choice QAs, with and without images. The green answer corresponds to the correct answer to each
question, and the rest three are wrong answer candidates. We take the majority votes of five human subjects as the human predictions (H)
and the top predictions from the model (M). The correct predictions are indicated by check marks.
time on questions with lower accuracy. The Pearson correlation coefficient between the average response time and
the average accuracy is −0.135, indicating a weak negative
correlation between the response time and human accuracy.
class labels. At test time, we select the top-scoring answer
candidate. For pointing QA, we perform k-means to cluster
training set regions by fc7 features into 5000 clusters, used
as class labels. At test time, we select the answer candidate
closest to the centroid of the predicted cluster.
6.2.2
LSTM The LSTM model in Malinowski and Fritz [28]
for visual QA with no attention modeling, which can be
considered as a simplified version of our full model with
the attention terms set to be uniform.
Model Experiments on 7W QA
Having examined human performance, our next question is
how well the state-of-the-art models can perform in the 7W
QA task. We evaluate automatic models on the 7W QA
tasks in three sets of experiments: without images (Question), without questions (Image) and with images (Question
+ Image). In the experiments without images (questions),
we zero out the image (questions) features. We briefly describe the three models we used in the experiments:
Logistic Regression A logistic regression model that predicts the answer from a concatenation of image fc7 feature
and question feature. The questions are represented by 200dimensional averaged word embeddings from a pre-trained
model [29]. For telling QA, we take the top-5000 most frequent answers (79.2% of the training set answers) as the
LSTM-Att Our LSTM model with spatial attention introduced in Sec. 5, where the attention terms in Eq. (10) determines which region to focus on at each step.
We report the results in Table 3. All the baseline models surpass the chance performance (25%). The logistic regression baseline yields the best performance when only
the question features are provided. Having the global image features hurts its performance, indicating the importance of understanding local image regions rather than a
holistic representation. Interestingly, the LSTM performance (46.2%) significantly outperforms human perfor-
A
B
cat
cake
What kind of animal is in the photo?
A cat.
C
Where are the carrots?
At the top.
Why is the person holding a knife?
To cut the cake with.
D
How many people are there?
Three.
Figure 8: Object groundings and attention heat maps. We visualize the attention heat maps (with Gaussian blur) on the images.
The brighter regions indicate larger attention terms, i.e., where the
model focuses. The bounding boxes show the object-level groundings of the objects mentioned in the answers.
mance (35.3%) when the images are not present. Human
subjects are not trained before answering the questions;
however, the LSTM model manages to learn the priors of
answers from the training set. In addition, both the questions and image content contribute to better results. The
Question + Image baseline shows large improvement on
overall accuracy (52.1%) than the ones when either the
question or the image is absent. Finally, our attention-based
LSTM model (LSTM-Att) outperforms other baselines on
all question types, except the how category, achieving the
best model performance of 55.6%.
We show qualitative results of human experiments and
the LSTM models on the telling QA task in Fig. 7. Human subjects fail to tell a sheep apart from a goat in the last
example, whereas the LSTM model gives the correct answer. Yet humans successfully answer the fourth why question when seeing the image, where the LSTM model fails in
both cases.
The object groundings help us analyzing the behavior
of the attention-based model. First, we examine where the
model focuses by visualizing the attention terms of Eq. (10).
The attention terms vary as the model reads the QA words
one by one. We perform max pooling along time to find
the maximum attention weight on each of the 14×14 image
grid, producing an attention heat map. We see if the model
attends to the mentioned objects. The answer object boxes
occupy an average of 12% of image area; while the peak of
the attention heat map resides in answer object boxes 24%
of the time. That indicates a tendency for the model to attend to the answer-related regions. We visualize the atten-
Figure 9: Impact of object category frequency on the model accuracy in the pointing QA task. The x-axis shows the upper bound
object category frequency of each bin. The y-axis shows the mean
accuracy within each bin. The accuracy increases gradually as the
model sees more instances from the same category. Meanwhile,
the model manages to handle infrequent categories by transferring
knowledge from larger categories.
tion heat maps on some example QA pairs in Fig. 8. The
top two examples show QA pairs with answers containing
an object. The peaks of the attention heat maps reside in the
bounding boxes of the target objects. The bottom two examples show QA pairs with answers containing no object.
The attention heat maps are scattered around the image grid.
For instance, the model attends to the four corners and the
borders of the image to look for the carrots in Fig. 8(c).
Furthermore, we use object groundings to examine the
model’s behavior on the pointing QA. Fig. 9 shows the impact of object category frequency on the QA accuracy. We
divide the object categories into different bins based on their
frequencies (by power of 2) in the training set. We compute the mean accuracy over the test set QA pairs within
each bin. We observe increased accuracy for categories
with more object instances. However, the model is able to
transfer knowledge from common categories to rare ones,
generating an adequate performance (over 50%) on object
categories with only a few instances.
7. Conclusions
In this paper, we propose to leverage the visually
grounded 7W questions to facilitate a deeper understanding of images beyond recognizing objects. Previous visual
QA works lack a tight semantic link between textual descriptions and image regions. We link the object mentions
to their bounding boxes in the images. Object grounding
allows us to resolve coreference ambiguity, understand object distributions, and evaluate on a new type of visually
grounded QA. We propose an attention-based LSTM model
to achieve the state-of-the-art performance on the QA tasks.
Future research directions include exploring ways of utilizing common sense knowledge to improve the model’s performance on QA tasks that require complex reasoning.
Acknowledgements We would like to thank Carsten Rother
from Dresden University of Technology for establishing the
collaboration between the Computer Vision Lab Dresden
and the Stanford Vision Lab which enabled Oliver Groth
to visit Stanford to contribute to this work. We would also
like to thank Olga Russakovsky, Lamberto Ballan, Justin
Johnson and anonymous reviewers for useful comments.
This research is partially supported by a Yahoo Labs Macro
award, and an ONR MURI award.
References
[1] S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. L.
Zitnick, and D. Parikh. VQA: Visual question answering.
ICCV, 2015. 1, 2, 3, 4, 5, 6
[2] K. Barnard, P. Duygulu, D. Forsyth, N. De Freitas, D. M.
Blei, and M. I. Jordan. Matching words and pictures. The
Journal of Machine Learning Research, 3:1107–1135, 2003.
2
[3] J. P. Bigham, C. Jayant, H. Ji, G. Little, A. Miller, R. C.
Miller, R. Miller, A. Tatarowicz, B. White, S. White, et al.
Vizwiz: nearly real-time answers to visual questions. In
Proceedings of the 23nd annual ACM symposium on User
Interface Software and Technology, 2010. 2
[4] X. Chen and C. L. Zitnick. Mind’s eye: A recurrent visual
representation for image caption generation. In CVPR, 2015.
1, 2
[5] J. Donahue, L. Anne Hendricks, S. Guadarrama,
M. Rohrbach, S. Venugopalan, K. Saenko, and T. Darrell. Long-term recurrent convolutional networks for visual
recognition and description. In CVPR, 2015. 1, 2, 5
[6] D. Ferrucci et al. Building Watson: An overview of the
DeepQA project. AI Magazine, 2010. 2
[7] H. Gao, J. Mao, J. Zhou, Z. Huang, L. Wang, and W. Xu.
Are you talking to a machine? dataset and methods for multilingual image question answering. NIPS, 2015. 1, 2, 4,
5
[8] D. Geman, S. Geman, N. Hallonquist, and L. Younes. Visual
turing test for computer vision systems. Proceedings of the
National Academy of Sciences, 112(12):3618–3623, 2015. 1,
2, 4
[9] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic
segmentation. In CVPR, 2014. 2
[10] K. Gregor, I. Danihelka, A. Graves, and D. Wierstra. Draw:
A recurrent neural network for image generation. arXiv
preprint arXiv:1502.04623, 2015. 5
[11] S. Hochreiter and J. Schmidhuber. Long short-term memory.
Neural computation, 9(8):1735–1780, 1997. 5
[12] M. Iyyer, J. Boyd-Graber, L. Claudino, R. Socher, and
H. Daumé III. A neural network for factoid question answering over paragraphs. In Empirical Methods in Natural
Language Processing, 2014. 2
[13] A. Karpathy and L. Fei-Fei. Deep visual-semantic alignments for generating image descriptions. CVPR, 2015. 1,
2, 5
[14] A. Karpathy, A. Joulin, and L. Fei-Fei. Deep fragment embeddings for bidirectional image sentence mapping. In NIPS,
pages 1889–1897, 2014. 1
[15] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar,
and L. Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, 2014. 1, 2
[16] S. Kazemzadeh, V. Ordonez, M. Matten, and T. L. Berg.
Referit game: Referring to objects in photographs of natural scenes. In EMNLP, 2014. 2
[17] M. H. Kiapour, X. Han, S. Lazebnik, A. C. Berg, and T. L.
Berg. Where to buy it: Matching street clothing photos in
online shops. ICCV, 2015. 2
[18] D. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6
[19] C. Kong, D. Lin, M. Bansal, R. Urtasun, and S. Fidler. What
are you talking about? text-to-image coreference. In CVPR,
2014. 1, 2
[20] R. Krishna, Y. Zhu, O. Groth, J. Johnson, K. Hata, J. Kravitz,
S. Chen, Y. Kalantidis, L.-J. Li, D. A. Shamma, M. Bernstein, and L. Fei-Fei. Visual genome: Connecting language
and vision using crowdsourced dense image annotations. In
arXiv preprint arxiv:1602.07332, 2016. 2
[21] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet
classification with deep convolutional neural networks. In
NIPS, pages 1097–1105, 2012. 1
[22] R. Kuhn and E. Neveu. Political journalism: New challenges, new practices. Routledge, 2013. 2
[23] C. H. Lampert, H. Nickisch, and S. Harmeling. Learning to
detect unseen object classes by between-class attribute transfer. In CVPR, 2009. 2
[24] T.-Y. Lin, Y. Cui, S. Belongie, and J. Hays. Learning
deep representations for ground-to-aerial geolocalization. In
CVPR, 2015. 2
[25] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. Microsoft coco: Common objects in context. In ECCV. 2014. 1, 2, 4
[26] L. Ma, Z. Lu, and H. Li. Learning to answer questions from
image using convolutional neural network. arXiv preprint
arXiv:1506.00333, 2015. 2
[27] M. Malinowski and M. Fritz. A multi-world approach to
question answering about real-world scenes based on uncertain input. In NIPS, pages 1682–1690, 2014. 1, 2, 4, 5
[28] M. Malinowski, M. Rohrbach, and M. Fritz. Ask your neurons: A neural-based approach to answering questions about
images. ICCV, 2015. 2, 5, 7
[29] T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and
J. Dean. Distributed representations of words and phrases
and their compositionality. In NIPS, 2013. 7
[30] F. Palermo, J. Hays, and A. A. Efros. Dating historical color
images. In ECCV. 2012. 2
[31] G. Patterson and J. Hays. Sun attribute database: Discovering, annotating, and recognizing scene attributes. In CVPR,
2012. 2
[32] L. C. Pickup, Z. Pan, D. Wei, Y. Shih, C. Zhang, A. Zisserman, B. Scholkopf, and W. T. Freeman. Seeing the arrow of
time. In CVPR, 2014. 2
[33] H. Pirsiavash, C. Vondrick, and A. Torralba. Inferring the
why in images. arXiv preprint arXiv:1406.5472, 2014. 2
[34] B. Plummer, L. Wang, C. Cervantes, J. Caicedo, J. Hockenmaier, and S. Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-tosentence models. ICCV, 2015. 1
[35] V. Ramanathan, A. Joulin, P. Liang, and L. Fei-Fei. Linking
people with ”their” names using coreference resolution. In
ECCV, 2014. 1, 2
[36] M. Ren, R. Kiros, and R. Zemel. Exploring models and data
for image question answering. NIPS, 2015. 1, 2, 4, 5
[37] M. Rohrbach, W. Qiu, I. Titov, S. Thater, M. Pinkal, and
B. Schiele. Translating video content to natural language
descriptions. In ICCV, 2013. 2
[38] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh,
S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein,
A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual
Recognition Challenge. IJCV, pages 1–42, April 2015. 1, 4
[39] K. Simonyan and A. Zisserman. Very deep convolutional
networks for large-scale image recognition. ICLR, 2014. 1,
2, 5, 6
[40] R. Socher, A. Karpathy, Q. V. Le, C. D. Manning, and A. Y.
Ng. Grounded compositional semantics for finding and describing images with sentences. Transactions of the Association for Computational Linguistics, 2:207–218, 2014. 1,
2
[41] I. Sutskever, O. Vinyals, and Q. V. Le. Sequence to sequence
learning with neural networks. In NIPS, pages 3104–3112,
2014. 5
[42] Y. Taigman, M. Yang, M. Ranzato, and L. Wolf. Deepface:
Closing the gap to human-level performance in face verification. In CVPR, 2014. 1, 2
[43] A. Toshev and C. Szegedy. Deeppose: Human pose estimation via deep neural networks. In CVPR, 2014. 1
[44] K. Tu, M. Meng, M. W. Lee, T. E. Choe, and S.-C. Zhu. Joint
video and text parsing for understanding events and answering queries. In IEEE MultiMedia, 2014. 2
[45] O. Vinyals, A. Toshev, S. Bengio, and D. Erhan. Show and
tell: A neural image caption generator. In CVPR, 2015. 1, 2
[46] J. Weston, A. Bordes, S. Chopra, and T. Mikolov. Towards
ai-complete question answering: a set of prerequisite toy
tasks. arXiv preprint arXiv:1502.05698, 2015. 2, 4
[47] K. Xu, J. Ba, R. Kiros, K. Cho, A. Courville, R. Salakhutdinov, R. Zemel, and Y. Bengio. Show, attend and tell: Neural
image caption generation with visual attention. In ICML,
2015. 1, 2, 5
[48] P. Young, A. Lai, M. Hodosh, and J. Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67–
78, 2014. 1
[49] L. Yu, E. Park, A. C. Berg, and T. L. Berg. Visual Madlibs:
Fill in the blank Image Generation and Question Answering.
ICCV, 2015. 1, 2, 4
[50] B. Zhou, A. Lapedriza, J. Xiao, A. Torralba, and A. Oliva.
Learning deep features for scene recognition using places
database. In NIPS, 2014. 2
[51] Y. Zhu, A. Fathi, and L. Fei-Fei. Reasoning about object affordances in a knowledge base representation. ECCV, 2014.
2
[52] C. L. Zitnick, D. Parikh, and L. Vanderwende. Learning the
visual interpretation of sentences. In ICCV, 2013. 2
| 9cs.NE
|
Autonomous Take-Off and Flight of a Tethered Aircraft for Airborne
Wind Energy
Lorenzo Fagiano, Eric Nguyen-Van, Felix Rager, Stephan Schnez, and Christian Ohler∗†
arXiv:1608.01889v2 [cs.SY] 16 Feb 2017
Abstract
A control design approach to achieve fully autonomous take-off and flight maneuvers with a tethered aircraft is
presented and demonstrated in real-world flight tests with a small-scale prototype. A ground station equipped with a
controlled winch and a linear motion system accelerates the aircraft to take-off speed and controls the tether reeling
in order to limit the pulling force. This setup corresponds to airborne wind energy systems with ground-based energy
generation and rigid aircrafts. A simple model of the aircraft’s dynamics is introduced and its parameters are identified
from experimental data. A model-based, hierarchical feedback controller is then designed, whose aim is to manipulate
the elevator, aileron and propeller inputs in order to stabilize the aircraft during the take-off and to achieve figure-of-eight
flight patterns parallel to the ground. The controller operates in a fully decoupled mode with respect to the ground station.
Parameter tuning and stability/robustness aspect are discussed, too. The experimental results indicate that the controller
is able to achieve satisfactory performance and robustness, notwithstanding its simplicity, and confirm that the considered
take-off approach is technically viable and solves the issue of launching this kind of airborne wind energy systems in a
compact space and at low additional cost.
1
Introduction
Airborne Wind Energy (AWE) systems aim at extracting energy from the wind by using autonomous tethered aircrafts,
kept aloft by either aerodynamic or aerostatic lift [1, 2]. By using light and flexible tethers instead of a massive structure,
like the tower of wind turbines, AWE technologies feature relatively low capital costs and can operate above 200 m from
the ground, where the wind is generally faster and more consistent. The price paid for these advantages is a higher system
complexity, which makes the development of AWE difficult. At the current stage, power production has been studied
theoretically and assessed experimentally by several companies and academic research groups worldwide, for different
AWE concepts [3, 4, 5, 6, 7, 8, 9, 10, 11].
However, system reliability and robustness during power generation has not been demonstrated yet, not even on
a small scale, at least judging from the publicly available literature. Due to the complex dynamics of AWE systems,
involving the interaction of ground-based equipments with long, flexible tether(s) and with the aircraft immersed in an
uncertain and variable wind field, successful simulations and small experiments do not represent a significant proof of
long-term reliability and performance: only an experimental validation carried over several months can provide enough
confidence in any AWE system to really foster its industrial development, and these long-term tests have not been carried
out or documented yet. In turn, a prerequisite for such a long-term experimental validation is the development of a viable
(technically and economically) approach for autonomous take-off and landing in compact space. For AWE systems that
employ the so-called “pumping” or “yoyo” operation, i.e. where the tether is repeatedly reeled-out under high load from
a winch installed on the ground and reeled-in under low load, this important feature appears to be still missing, again
judging from the scientific literature. In fact, while for the power production phase there is a consensus on how the system
shall operate, the related modeling and control aspects are relatively well-understood and experimental results have been
reported [12, 13, 10, 14, 11], for take-off and landing it is not even clear what approach to pursue. Consequently, there
appears to be a lack of understanding of what are the most relevant problems involved, and there are no documented
experimental results in the literature.
There is therefore a need for research and development efforts targeted at studying and demonstrating take-off and
landing strategies for pumping AWE systems, from first-principle, conceptual analyses down to the specific engineering
problems that need to be solved during implementation.
In this paper, we present a contribution in this direction, focusing our attention on the take-off phase of systems
that employ rigid aircrafts. In this context, outside the scientific literature there is evidence of autonomous take-off
∗ This is the pre-print of a paper accepted for publication on the IEEE Transactions on Control systems Technology and is subject to IEEE Copyright.
doi: 10.1109/TCST.2017.2661825. L. Fagiano is with Politecnico di Milano, Dip. di Elettronica, Informazione e Bioingegneria, 20133 Milano, Italy.
E. Nguyen-Van, F. Rager, S. Schnez and C. Ohler are with ABB Switzerland Ltd., Corporate Research, 5405 Baden-Dättwil - Switzerland. E-mail
addresses: [email protected], {eric-nguyen-van | felix.rager | stephan.schnez | christian.ohler}@ch.abb.com.
† Corresponding author: Lorenzo Fagiano.
1
operation via a winch launch [15], which however requires a significant space in all directions in order to adapt to the
wind conditions. Another take-off approach that has been practically demonstrated employs vertical-axis propellers to
slowly lift the aircraft to a safe altitude before shifting to the power production phase [16]. In this case, the required
land occupation is rather small, however the additional onboard mass might compromise the system’s performance during
power generation. Within the scientific literature, the take-off of pumping AWE generators has been addressed only to a
limited extent, mainly considering a rotational take-off approach [17, 18]. In recent contributions [19, 20] we compared
different take-off approaches on the basis of qualitative and quantitative criteria, and concluded that a linear approach
is among the most promising solutions, with a good tradeoff between ground-based equipments, land occupation and
additional onboard mass. In such an approach, the aircraft is accelerated to take-off speed over a short distance by a motor
installed on the ground and then it climbs to a safe altitude using relatively small onboard propellers. At a recent AWE
conference, the company Ampyx Power also announced the development of a similar approach for their system [21].
The main contribution of this paper is to prove experimentally such a linear take-off concept on a small scale, with
fully autonomous operation. Additional contributions are represented by the modeling and control design steps that we
carried out in order to achieve this result. The coordination between the ground station and the aircraft is essential to
achieve a successful take-off. Here, we adopt a completely decoupled strategy, where the controller of the ground station
and the one on the glider do not exchange any information. Rather, the ground station controller exploits a local, indirect
measurement of the tether force to adjust the reeling speed. For the sake of brevity, this control algorithm is only briefly
described here; the full details are presented in [22]. Instead, the focus of the present paper is on the design of the onboard
control system and on the results of the autonomous take-off tests. We first introduce a simple model of the aircraft and
identify its parameters using experimental data. Then, we use the model to design a hierarchical controller able to achieve
autonomous take-off, transition to figure-of-eight patterns and flight along such patterns at roughly constant elevation.
We finally present and discuss experimental results. Albeit the tests have been performed on a small-scale prototype, the
control approach and most of the findings can be easily carried over to larger system sizes. We conclude with some final
remarks and indicate the potential improvements that can be subject of future research. It is important to point out that
our whole research activity towards autonomous take-off and flight has been conceived as a proof-of-concept study rather
than a final, optimized solution. To the best of our knowledge, this is nevertheless the first contribution in the scientific
literature that provides the full details and the experimental validation of a control approach for the take-off of a tethered
aircraft in compact space, to be used for AWE systems.
The paper is organized as follows. Section 2 presents the system layout, the features of the employed experimental
setup and the considered control objectives. Section 3 describes the modeling, identification and control design aspects.
The experimental results are presented in section 4, and conclusions are drawn in section 5.
2
System description and control objectives
The considered system is composed of a ground station connected to a rigid aircraft by a single tether, see Fig. 1 for a
sketch and Fig. 2 for a picture of the employed small-scale prototype. We first describe the two main subsystems, i.e. the
ground station and the aircraft, followed by a brief overview of the ground station control strategy. We finally present the
desired system operation and the consequent control objectives for the aircraft’s controller.
2.1
Ground station
Referring to Figs. 1-2, the ground station is equipped with a first motor/generator (“winch motor”) connected to a winch
that stores the tether attached to the aircraft. In a power-generating system, the winch motor also acts as generator during
the power production phase. Before leaving the ground station, the tether passes through a series of pulleys, one of which
is attached to a mass-spring mechanism used to reduce the stiffness of the link between the aircraft and the ground station.
The last pulley is installed on a slide able to move on linear rails. The slide is used to accelerate the aircraft up to take-off
speed. The slide position is controlled by a second motor (“slide motor”) installed on the ground station. In this way,
most of the power injected during the take-off maneuver is provided by the slide motor and the resulting required space
can be quite compact: in our prototype, only about 3.5 m of travel distance are needed to accelerate to about 9 m/s and
then brake the slide. Finally, the ground station also comprises a power supply, the drives to control the two motors, a
real-time machine for control and data-acquisition, a human-machine interface and the mechanical frame that supports all
the mentioned components. In our prototype, the power supply is composed of two 12V batteries connected in parallel
and a DC-AC inverter providing a 220 VAC, 3 kW single-phase line to the drives and to all the measurement and control
hardware. The main features of our small-scale prototype are reported in Table 1, while the full details are given in [22].
2.2
Aircraft
We consider a rigid aircraft with conventional design, equipped with an horizontal axis propeller and with four control
surfaces - elevator, ailerons, rudder and flaps. As a matter of fact, in our control approach we use only the motor, ailerons
2
Z
𝑊𝑊
𝒑𝒑̇
7
𝑣𝑣𝑎𝑎
x
𝜑𝜑
𝜃𝜃
y
pZ
z
pX
2
5
pY
1
3
X
6
𝛾𝛾ref,to
Y
4
Figure 1: Sketch of the considered system, showing: 1. the winch motor, 2. the winch, 3. the tether connecting the winch
to the aircraft, 4. the tether tensioning system, 5. the slide, 6. the rails, 7. the aircraft. The inertial coordinate system
(X, Y, Z) and the local one (x, y, z) are also shown, together with the aircraft position p and velocity ṗ, the roll and pitch
rotations ϕ, θ, the wind velocity vector relative to ground, W , the resulting apparent wind velocity va , and the angle
γref,to between the take-off direction (dotted arrow) and the X-axis.
7
5
6
4
3
2
1
9
11
10
8
12
13
Figure 2: Picture of the small-scale prototype built at ABB Corporate Research. The numbers 1. to 7. correspond to the
components also depicted in Fig. 1. In addition, the picture shows: 8. the slide motor and drum, 9. the box containing
power supply, inverter, and drives, 10. the real-time machine for data-logging and high-level control of the ground station,
11. the laptop used to interface with the drives and the real-time machine.
and elevator to control the aircraft, since these proved to be sufficient to achieve good performance. Further improvements
can be obtained by exploiting also the remaining control surfaces. A mechanism to attach and detach the tether is also
installed on the aircraft. The following onboard measurements are available to be used for feedback control: position and
velocity with respect to the ground; roll, pitch and yaw angles and their rates; 3D accelerations with respect to a local (i.e.
fixed with the aircraft) reference frame; airspeed along the local longitudinal axis.
In our prototype, we use a commercial model glider made of styrofoam (visible in Fig. 2), which we reinforced
with fiberglass and modified to add the tether attachment mechanism. In the glider we installed an inertial measurement
unit (IMU) (an SBG Systems “Ellipse-N” unit [23]) and a pitot probe, which together provide the above-mentioned
measurements at a rate of 50 Hz, large enough for the considered control problem since the aircraft dynamics exhibit
dominant poles lower than 5 Hz (see sections 3-4 for more details on the time constants of the glider dynamics). We
also installed an embedded platform (an Arduino MEGA 2560 board [24]) to fuse all the measurements, log the data,
and compute and actuate the control inputs. A commercial remote controller and the corresponding onboard receiver are
3
Table 1: Main technical features of the employed prototype.
Ground station
Winch and slide motors’ maximum continuous torque
13
Winch and slide motors’ peak torque
26
Winch and slide motors’ rated speed
208
Tether length
150
Rails’ length
4.5
Slide mass
9
Spring maximum compression
0.32
Spring stiffness
60
Battery bank capacity
260
Battery bank voltage
12
Inverter continuous power
3000
Inverter peak power
6000
Glider
Wingspan
1.68
Mean Aerodynamic Chord
0.194
Aspect Ratio
8.89
Wing Surface Area
0.3174
Mass
1.2
Peak Motor Power
360
Rated Motor Power
290
Battery voltage
14.4
Propeller diameter
0.22
Nm
Nm
rad/s
m
m
kg
m
N/m
Ah
V
W
W
m
m
m2
kg
W
W
V
m
used to pilot the aircraft manually if needed, and to switch between manual and automatic flight modes. Fig. 3 shows the
inside of the glider’s fuselage, with the IMU and the control board. The control surfaces and the tether release mechanism
are actuated by standard servomotors widely employed in the model glider business, and the front propeller is driven by
a standard brushless DC motor. For the sake of brevity, only the main characteristics of the glider are shown in Table 1,
while the full details on the employed hardware, sensors, actuators etc. are reported in [22].
4
2
3
1
7
6
5
Figure 3: Onboard measurement and control hardware for the glider. 1. DC-DC power converter; 2. motor driver; 3.
inertial measurement unit; 4. GPS antenna; 5. Arduinor Mega 2560 embedded platform; 6. extension boards; 7. RC
receiver.
2.3
Ground station control
As anticipated in the introduction, the details of the ground station design and control are outside the scope of this paper
and can be found in [22]. However, we recall here the principle of operation of the winch control system, since this is
instrumental to understand one of the main assumptions that we consider when modeling the aircraft and designing its
onboard controller.
A first goal for the winch control strategy is to limit the force exerted by the tether on the aircraft to relatively small
values. This aspect is fundamentally different from what is usually considered in AWE during the power generation
4
phase, where the tether force is very large. The main reason for this difference is that, during power generation, crosswind
trajectories are flown at large speed and the aircraft’s velocity vector is roughly perpendicular to the tether, thus the tether
force does not affect the aircraft forward velocity. On the contrary, during take-off the aircraft speed is much lower and
its velocity vector is almost aligned with the tether, such that even a small tether force can slow-down the aircraft and
provoke a stall situation.
At the same time, reeling out with no load must be also avoided, in order to prevent entanglement of the tether
and to limit the tether sag, which increases the weight and drag forces acting on the aircraft and might as well lead to
entanglement with obstacles on the ground. Such conflicting control objectives (low tether force and low tether sag)
require some sort of coordination between the aircraft and the ground station. A natural approach to this problem would
be to exploit communication between the two sub-systems to try to coordinate their motion. However, such an exchange
of information between the controllers might be still not enough to achieve the required coordination, since one would
need to gather not only the aircraft velocity vector and the tether length and speed, but also the relative angle between
the aircraft velocity and the tether. The latter might be very difficult to estimate or measure with good enough accuracy.
Furthermore, even with perfect knowledge of the position and velocity of both the aircraft and tether/winch, the torque
limitation and the inertia of the winch might still prevent the control system from acting fast enough to avoid a stall
situation when the tether force becomes too large.
In our study, a key design choice we made is to have no exchange of information between the ground station controller
and the aircraft. The main advantage of this choice is that communication reliability issues and delays do not have to be
dealt with, while the main disadvantage is that the achievable system performance, for example in terms of flight pattern
repeatability and tolerance to wind disturbances, is inferior to the case when communication between the aircraft and
the ground is exploited. Since our activity was conceived as a proof of concept and performance optimization was not
a priority, we decided not to exploit communication. In order to solve the winch/aircraft coordination problem without
using an active exchange of information, we designed the above-mentioned mass-spring system. The employed spring is
relatively soft, such that a small pulling force on the tether produces a rather large displacement. The spring compression
helps to reduce the harshness with which the tether pulls on the aircraft, and gives more time to the winch control system
to react. The spring displacement is measured by a potentiometer and employed as a feedback variable to compute the
reference reeling speed for the winch. A standard proportional controller in an inner control loop is then in charge of
tracking such a reference speed by adjusting the motor current, which is subject to saturation (hence imposing a saturation
on the motor torque). The full details of the ground station controller and more in general of the ground station design
are described in [22]. However, for the sake of completeness we include here a short description of the feedback control
strategy that computes the reference winch speed.
More specifically, let us denote with xs (k) the measured spring compression, where k ∈ Z indicates the discrete time
instants employed by the ground station controller. We set two threshold values, xIs , xII
s , which divide the available spring
travel in three zones:
• Zone a (0 ≤ xs (k) ≤ xIs ): the spring is practically uncompressed, the winch shall decrease speed and eventually
reel-in;
• Zone b (xIs < xs (k) < xII
s ): the spring is subject to low force, the winch speed shall be held (constant tether
speed);
• Zone c: (xII
s ≤ xs (k) ≤ xs ): the spring is subject to relatively large force, the winch shall increase its speed and
reel-out to release the tether.
Then, the reference winch speed, θ̇ref,w (k), is computed according to the following strategy (see Fig. 4):
If 0 ≤ xs (k) < xIs (Zone a)
xs (k) − xIs
I
xI,a
s −xs
fbck
a
θ̇ref,w (k) = min 0, max θ̇ref,w , θ̇ref,w (k − 1) + Ts θ̈ref,w
x̄s (k)
x̄s (k) =
Else if xIs ≤ xs (k) < xII
s (Zone b)
θ̇ref,w (k) = θ̇ref,w (k − 1)
Else if xII
s ≤ xs (k) ≤ xs (Zone c)
x̄s (k) =
xs (k) − xII
s
xII,c
−xII
s
s
fbck
c
θ̇ref,w (k) = max 0, min θ̇ref,w , θ̇ref,w (k − 1) + Ts θ̈ref,w
x̄s (k)
5
fbck
fbck
a
c
where θ̇ref,w , θ̇ref,w are the desired minimum and maximum reference speed values that can be issued, and θ̈ref,w
, θ̈ref,w
are the desired angular accelerations for the reference speed. Such values are scaled according to the position of the
II,c
I
potentiometer relative to the values xI,a
> xII
s < xs and xs
s , respectively for zones a and c, which are design parameters
as well. Equation (1) represents an integral controller, where the integrated quantity is the distance of the spring position
xs (k) from the interval (xIs , xII
s ) (i.e. zone b) and the gain is piecewise constant, since it is different in zones a and c.
Moreover, a saturation of the integrated variable to negative (resp. positive) values is operated whenever the spring enters
zone a (resp. c), in order to quickly start to reel-in (resp. reel-out) when the tether is released (resp. pulled). As shown
fbck
fbck
a
c
< 0, θ̇ref,w , θ̈ref,w
> 0, xI,a
≈ xIs /2 and
in [22], a sensible choice for the involved design parameters is θ̇ref,w , θ̈ref,w
s
II,c
xs ≈ (xs + xII
s )/2.
𝑥𝑥𝑠𝑠
a b
c
𝑥𝑥𝑠𝑠
a
𝜃𝜃̈ ref,w
c
𝜃𝜃̈ ref,w
×
Integrator saturation in zones a and c
�
𝜃𝜃̇ref,w
Figure 4: Block-diagram of the feedback contribution to the reference winch speed.
The actual winch control strategy is slightly more complex, since it involves also a feed-forward term during the
take-off transient, in order to latch the winch speed to the slide speed, see [22] for details.
An example of the typical behavior obtained with the described winch control system is shown in Fig. 5, which
presents the aircraft airspeed, winch speed, reference winch speed, spring compression and an estimate of the tether force
during autonomous tethered flight, while the aircraft is pulling on the tether and then starting a turn maneuver. The
difference between the aircraft airspeed and the tether speed is due to the relative orientation between the two. Referring
to Fig. 5, the initial spring position is such that, according to the winch control strategy, the reeling speed is held constant
(equal to zero in this specific case). Then, at t ' 15 s the tether becomes taut and the spring compresses, entering a region
where the reference winch speed is increased (according to the winch control law). When the tether length becomes larger
than the distance between the aircraft and the ground station, the tether becomes slack again and the spring is released.
Then, the reference winch speed is first held in place (as the spring travels through the "hold" region) and then quickly
ramped down toward negative values. However, as the winch slows down, the aircraft might pull again on the tether, hence
giving place to the oscillating behavior of the winch speed, until finally after t ' 18.5 s the tether force settles again on a
roughly constant value, corresponding to a fixed spring position. This kind of transient is very common with the adopted
strategy, where the winch controller and the aircraft controller are fully decoupled and coordination has to be done via
the spring compression. An oscillating behavior of the spring while the tether is taut indicates that the (average) reel-out
speed is matching well with the aircraft speed projected along the tether: the alternative would be either to reel-out too
fast, with consequent full decompression of the spring and tether entanglement on the ground, or to reel-out too slowly,
with consequent stall of the aircraft.
It can be noted that the typical perturbation of the glider speed due to the tether force is less than 1 m/s, which is
an acceptable performance for the sake of our study. A rough estimate of the the tether force, neglecting the inertia and
viscous friction of the spring mechanism, can be obtained by simply scaling the spring compression by its stiffness, of
about 60 N/m, and considering that, since the tether runs 180 deg around the pulley attached to the spring (see Fig. 1),
the force acting on the spring equals two times the tether force. Such an estimate is plotted in Fig. 5, too, showing that
impulses of about 3-8 N peak and 0.3-0.7 s of duration are experienced by the aircraft. These values are acceptable for
the glider used in our tests, since they don’t compromise the flight dynamics.
The described winch control strategy and its performance represent the main motivation for which, in the remainder
of this paper, we will refrain from including a model of the tether or of the ground station to design the aircraft controller.
In fact, we will consider a rather simple model of the aircraft dynamics, where the force exerted by the tether is embedded,
together with other neglected aspects and high-order dynamics, in an unknown-but-bounded additive disturbance term.
2.4
System operation and control objectives
The overall aim of the onboard controller is to reach a repetitive flight pattern above the ground station, starting from a
standstill position on the slide. We divided such a control task in two subsequent phases, whose control objectives are
6
Speed (m/s), compression (10-1 m), Est. tether force (N)
12
10
8
6
4
2
0
-2
14
15
16
17
Time (s)
18
19
20
Figure 5: Experimental results: example of transient of spring compression and related loss of speed of the aircraft when
the tether is taut. Solid line: airspeed (m/s), dashed: winch speed (m/s), dotted: reference winch speed (m/s), dash-dot:
spring compression (10−1 m). Thin dash-dotted line: tether force estimated from the spring compression.
described next.
Take-Off and climbing: The glider is initially attached to the slide. The take-off maneuver is initiated by the ground
station, which accelerates the slide. The onboard controller has to detect this situation (we recall that there is no exchange
of information between the ground station controller and the onboard one) and stabilize the aircraft’s attitude until it
detaches from the slide, while at the same time powering up the onboard motor to sustain the climb. After detaching from
the slide, the first goal of the controller is to reach a safe altitude, before performing any other maneuver. A relatively
large climb angle is thus targeted to rapidly reach such a safe altitude. A large climbing angle is also beneficial when the
tether is taut, since a smaller component of the tether force adds to the aerodynamic drag.
Transition phase and figure-of-eight patterns: Once a safe altitude has been reached, the controller shall carry out
a transition maneuver to engage a repetitive figure-of-eight pattern parallel to the ground at a target altitude, withstanding
the presence of wind gusts and of the tether. Assuming that the wind flow is mainly parallel to the ground as well, such
a pattern gives rise to low tether forces. In a power generating system, this phase can be then exploited to check the
system and estimate the wind conditions before carrying out a further transition to crosswind figure-of-eight paths. Such
a transition can be achieved by smoothly adjusting the eight-pattern from horizontal to crosswind. At the same time, the
winch control system has to be changed from a “low-tether-force” strategy, like the one considered in this paper (which
allows the aircraft to take-off notwithstanding the tether), to a “high-tether-force” strategy, e.g. where the reel-in torque is
set proportional to the square of the reel-out speed, see e.g. [11]. Since in our study the goal is to demonstrate only the
autonomous take-off and the transition to low-force flight, we stop the experiments once the horizontal eight-pattern has
been reached.
3
System modeling and control design
We will now describe in detail the design of a controller able to achieve the objectives described in section 2.4. The
design is model based: we first derive a relatively simple model of the aircraft dynamics and identify its parameters using
experimental data collected in preliminary flights, and then we design the final controller on the basis of this model.
3.1
Aircraft modeling and parameter identification
3.1.1
Model equations
There exist a vast literature and many textbooks concerned with the flight mechanics of a rigid aircraft like the one
considered in this work, albeit without the tether. A frequent approach is to employ a six-degrees-of-freedom (dof) model
with lumped parameters accounting for the geometrical, inertial and aerodynamic characteristics of the aircraft. In a
preliminary work we used such a model (taken from [25]), together with models of the ground station and of the tether,
in order to obtain a simulation tool for our control algorithms [26]. This model is quite realistic as it includes the most
7
important system nonlinearities, e.g. the stall behavior, and considers the coupling among the different states. However,
it is rather complex, since it comprises 12 states (three-dimensional position and rotation and their time derivatives) and
5 inputs (motor, rudder, ailerons, elevator, and flaps). If the presence of the ground station (slide, winch and massspring system) and of the tether were also considered (the latter with a simple purely elastic model as done in [26]), the
number of states and inputs would raise to 18 and 7, respectively. Such a model is very useful to simulate the system
and carry out a preliminary test of control algorithms, however its use for control design is not trivial. In fact, even the
parameter identification phase raises difficulties due to the presence of nonlinearities, open-loop unstable behavior and
poorly controllable and expensive experimental tests, where the wind vector, which would be required for an accurate
identification of the aerodynamic characteristics, is anyways not measured.
Therefore, for the sake of control design we decided to take a radically opposite approach and adopt a very simple
model of the system dynamics to attain our goals, with the same spirit of previous contributions on AWE control design,
[27, 13, 10]. Namely, we consider two main dynamical modes: the first one to describe the turning behavior, and the
second one to describe the vertical motion. A third dynamical equation describes the link between the airspeed and the
propeller thrust. The couplings among all these modes are neglected and embedded into additive disturbance/uncertainty
terms. The core building blocks of the resulting model are either linear time-invariant, second-order dynamical equations
whose parameters can be easily identified from the available measurements, or kinematic relationships that feature no parameter at all. Thus, the parameter identification phase is rather straightforward and effective, as shown below. Moreover,
the model is readily usable for control design using a hierarchical approach, as detailed in section 3.2.
We start by introducing an inertial coordinate system (X, Y, Z) fixed to the ground with the Z−axis pointing up, and a
local one (x, y, z) fixed to the aircraft (body frame), with the z−axis pointing down when the aircraft is in horizontal flight,
see Fig. 1. The rails’ orientation (take-off direction) forms an angle γref,to with the inertial X-axis. The position vector of
.
the aircraft is denoted by p (bold letters indicate vectors in a three-dimensional space), its velocity by ṗ = dp/dt, where
t is the continuous time variable. When considering the components of a vector, we add as a subscript the corresponding
axis, e.g. pX is the component of vector p along the inertial X−axis. We further denote by ϕ, θ, ψ the standard roll,
pitch and yaw angles (Euler angles), and with va the magnitude of the airspeed toward the body x−axis direction. The
full airspeed vector is in principle computed as W − ṗ, where W is the wind velocity vector, which however is not
measured in practice. On the other hand, the value of va is measured thanks to the onboard pitot probe aligned with the
body x−axis. Finally, we denote with uϕ , uθ and um the control inputs for the ailerons, elevator, and propeller thrust,
respectively.
The equations for the first mode (turning behavior) are derived starting from the roll angle dynamics:
ϕ̈(t) = aϕ ϕ̇(t) + bϕ uϕ (t) + dϕ (t)
(1)
where aϕ and bϕ are parameters to be identified, and dϕ is a disturbance term accounting for neglected dynamics, wind,
and the presence of the tether. The parameter aϕ is expected to be negative, so to have a stable first-order model linking
the aileron input uϕ to the roll rate ϕ̇. Equation (1) is motivated by physical insight, since a constant aileron input gives
rise to a constant roll rate in steady-state(due to the equilibrium of the aerodynamic roll moment induced by the aileron
with the opposite one generated by the roll rate), by experimental evidence (see section 3.1.2 below) and by the structure
of the equation obtained by linearizing a more complex model, like the one derived in [26], in a steady-state, straight-flight
configuration.
Considering a turning maneuver at constant tangential speed and with constant radius along a circular trajectory
parallel to the ground (i.e. the (X, Y )-plane), the roll angle ϕ can be then linked to the lateral acceleration of the aircraft
by considering the equilibrium of aerodynamic lift, gravitational force, and centrifugal force during the turn:
L sin (ϕ) = malat
(2a)
L cos (ϕ) = mg
(2b)
where L is the aerodynamic lift force, m the mass of the aircraft, and alat is the centripetal acceleration. Fig. 6 presents
an intuitive representation of equations (2). As mentioned, such equations hold with good approximation when the path
is parallel to the ground, such that when ϕ = 0 the lift force points along the inertial Z direction. Moreover, under the
considered assumptions the tangential speed equals |ṗ(t)| and from the kinematics of rigid bodies we have:
alat = |ṗ(t)|ψ̇
(3)
where ψ̇ is the yaw rate. Now, inserting (3) into (2a), dividing (2a) by (2b) and assuming small roll angles, such that we
can use the linearization of the trigonometric functions at ϕ = 0 (in our experiments we seldom exceeded 30◦ of roll), we
obtain:
g
ϕ(t)
(4)
ψ̇(t) =
|ṗ(t)|
8
𝜑𝜑
L
alat
malat
z
mg
y
Z
Figure 6: Sketch of the considered turning model for the aircraft
Finally, it is useful to also introduce the link between the trajectory curvature 1/R(t) (where R(t) is the turning radius)
|ṗ(t)|
and the roll angle in the model. Considering that ψ̇(t) '
, we have:
R(t)
1
g
ϕ(t)
=
R(t)
|ṗ(t)|2
(5)
Together, equations (1) and (4) form our model for the turning behavior, since they link the aileron input to the roll angle,
and the latter to the flown trajectory on a plane parallel to the ground. Such equations have the advantage of having just
two parameters to be identified within a second-order, linear time invariant (LTI) system (eq. (1)), while all the other
involved quantities are measured by the available sensors, in particular magnetometers for ϕ, gyroscopes for ψ̇, GPS and
accelerometers for |ṗ(t)|.
About the vertical motion, we start from the pitch dynamics:
θ̈(t) = aθ θ̇(t) + bθ uθ (t) + dθ (t)
(6)
where, in a way similar to equation (1), aθ and bθ are unknown parameters to be identified and dθ accounts for neglected
effects. Again, the parameter aθ is expected to be negative, i.e. the dynamical behavior from the elevator input uθ to the
pitch rate is asymptotically stable. The model (6) is also based on physical insight and prior knowledge, it is confirmed by
the experimental data and it holds within certain limits, in particular for small values of the pitch angle and of the angle of
attack. Such conditions are met in our experiments. Assuming now a straight-line flight parallel to the ground at constant
forward speed, a constant pitch value θ0 exists, such that this motion is at steady state, i.e. the lift force equals the weight
of the aircraft and the propeller’s thrust counteracts the drag. For simplicity and without loss of generality, we assume
θ0 = 0. Exploiting again the assumption of small θ values, we can then approximate the link between the pitch angle and
the vertical velocity along the inertial Z−axis as:
ṗZ (t) = |ṗ(t)|θ(t)
(7)
Fig. 7 provides a graphical interpretation of (7). Equations (6) and (7) link the pitch dynamics to the vertical inertial
𝑥𝑥
𝑍𝑍
𝒑𝒑̇
(𝑝𝑝̇𝑋𝑋 , 𝑝𝑝̇𝑌𝑌 )
𝑝𝑝̇𝑍𝑍
𝑧𝑧
Figure 7: Sketch of the considered pitch/vertical motion model for the aircraft
component of the aircraft trajectory, and form our model of the vertical dynamics. Also in this case there are only few
parameters, which can be easily identified from experimental data.
9
Last, we employ the following equations to model the effect of the propeller thrust on the airspeed va :
um (t) =
1
ρACD va (t)2 + dva (t)
2
(8)
where ρ is the air density, CD the aerodynamic drag coefficient, A the effective area of the aircraft, and dva (t) a term
accounting for neglected effects. Equation (8) is derived by assuming equilibrium between the motor thrust and the aerodynamic drag, and a small angle of attack (i.e. assuming that the apparent wind is mainly aligned with the body x−axis).
Similarly to the previous equations of motion, we include the approximation errors arising from such an assumption in
the term dva (t). The simple equation (8) is good enough for the considered application.
Equations (1), (4), (6), (7) and (8) form the model of the aircraft that we will use for control design. The only unknown
parameters are aϕ , bϕ , aθ , bθ , which can be identified from data as shown in the following. Regarding the value of CD
in (8), a good estimate can be obtained from the airfoil shape and the aircraft design; moreover such a parameter is not
critical for control design as we comment later on in section 3.2.4.
3.1.2
Parameter identification
For the sake of brevity, we present here the identification approach and results only for the roll angle dynamics (1). For
the pitch angle, we employed the same approach and obtained similar results. Equation (1) involves two parameters to
identify, aϕ and bϕ . We first estimated a value for both parameters by linearizing the 6-dof model of [26] and designed a
preliminary proportional controller Kid to track a reference roll angle ϕref , and we tested such a controller in fly-by-wire,
i.e. with a human pilot issuing the desired roll angle from the ground. Using a batch of the collected data, we then
identified the model parameters in closed-loop with the same preliminary controller by solving the following optimization
problem:
N
min {ϕ̃}N
1 − {ϕ}1
aϕ ,bϕ
2
N
˜ N
+ {ϕ̇}
1 − {ϕ̇}1
2
(9a)
subject to
ϕ̇(τ + 1) = ϕ̇(τ ) + Ts (aϕ ϕ̇(τ ) + bϕ Kid (ϕ̃ref (τ ) − ϕ(τ )))
(9b)
ϕ(0) = ϕ̃(0)
˜
ϕ̇(0) = ϕ̇(0)
(9c)
(9d)
where Ts is the employed sampling time, τ = 0, 1, . . . , N − 1 denotes the discrete-time sampling instants, N is the
.
T
N
number of employed samples, {ϕ̃}N
1 = [ϕ̃(1), . . . , ϕ̃(N )] is the sequence of measured roll angle values and {ϕ}1 of
N
N
˜ 1 , {ϕ̇}1 are the measured and estimated roll rate values). The notation˜in (9) denotes
estimated ones (and similarly {ϕ̇}
the collected noise-corrupted experimental data. Namely, problem (9) aims to find values of the parameters that minimize
the 2-norm of the error between the measured and simulated roll rate, in the presence of the same preliminary controller
and applying the measured input reference roll angle. The resulting problem is a nonlinear program (NLP) which we
solved for a local minimum by using a sequential quadratic programming approach (the function fmincon in Matlabr ).
An example of the obtained results is illustrated in Fig 8, where the roll rate obtained using the parameters computed by
solving (9) is compared with the measured one, using a second batch of data not employed for the parameter identification.
The obtained results further confirm that the considered second order model (1) approximates the roll motion well and
that we can avoid the complexity of the 6-dof model for the controller design. In section 4.1, we present the obtained
numerical values of the parameters.
3.2
Control Design
3.2.1
Controller structure
We employ a hierarchical control structure (see Fig 9), with low-level controllers designed to track reference values for
the roll angle, ϕref , pitch angle, θref , and front airspeed, va,ref , and high-level controllers that compute such references in
order to achieve the goals outlined in section 2.4 for each operational phase. Following our modeling approach, we design
decoupled low- and high-level controllers for the roll, pitch and forward motion modes. The employed control inputs are
those appearing in the model of section 3.1, i.e. the aileron and elevator positions, uϕ and uθ , and the propeller thrust, um .
In the next sub-sections, we describe the structure of the low- and high-level controllers, followed by short discussions
about the controllers’ tuning and closed-loop stability guarantees.
3.2.2
Low-level controllers
Since ϕ, ϕ̇, θ, θ̇, and va are directly measured and their dynamics are approximately linear for the sake of our goals (see
section 3.1), we can use static, linear state-feedback control laws to stabilize these quantities and track their reference
values.
10
Roll rate (rad/s)
Roll angle (rad)
1
0.5
0
-0.5
-1
30
40
50
60
70
Time (s)
80
90
100
2
1.5
1
0.5
0
-0.5
-1
-1.5
30
40
50
60
70
Time (s)
80
90
100
Figure 8: Experimental results. Solid lines: measured roll angle (upper plot) and roll rate (lower plot) obtained in fly-bywire mode with a preliminary controller. Dashed lines: roll angle and rate obtained with the model (1), whose parameters
have been identified by solving problem (9). The data shown in the plot are different from those used for the identification.
For the roll dynamics, we define the angular tracking error as
.
eϕ (t) = ϕref (t) − ϕ(t),
(10)
and employ the state-feedback control law
uϕ (t) = Keϕ eϕ (t) + Kėϕ ėϕ (t).
(11)
From (1) and (10)-(11) we have:
ėϕ (t)
ëϕ (t)
=
|
0
−bϕ Keϕ
eϕ (t)
1
ėϕ (t)
aϕ − bϕ Kėϕ
{z
}
Aϕ,CL
0
+
wϕ (t)
(12)
.
where wϕ (t) = (ϕ̈ref (t) − dϕ (t) − aϕ ϕ̇ref ) can be seen as an external, bounded disturbance term accounting for neglected dynamics and the time variation of the reference roll angle and rate issued by the high-level controller. We use
a classical pole-placement technique [28] to design the gains Keϕ , Kėϕ in order to assign the eigenvalues of the matrix
Aϕ,CL describing the closed-loop dynamics of the tracking error. In particular, in order to easily tune the controller also
during flight, we implemented the explicit equations that link Keϕ , Kėϕ to the desired closed-loop eigenvalues of the roll
dynamics, denoted as λϕ,1 , λϕ,2 :
λϕ,1 λϕ,2
Keϕ =
bϕ
(13)
λϕ,1 + λϕ,2 − aϕ
Kėϕ =
−bϕ
In (13), the only design parameters are λϕ,1 , λϕ,2 , which can be intuitively tuned on the basis of the desired closed-loop
bandwidth (we provide the values employed in our experimental tests in section 4.1).
Following the same approach, we design the low-level controller for the pitch as:
.
eθ (t) = θref (t) − θ(t)
(14a)
uθ (t) = Keθ eθ (t) + Kėθ ėθ (t)
λθ,1 + λθ,2 − aθ
Kėθ = −
−bθ
λθ,1 λθ,2
K eθ =
bθ
λθ,1 + λθ,2 − aθ
Kėθ =
−bθ
11
(14b)
(14c)
(14d)
(14e)
Turning
controller
𝜑𝜑ref
Altitude
controller
Roll
controller
𝜃𝜃ref
𝑢𝑢𝜑𝜑
Pitch
controller
High-level 𝑣𝑣
𝑎𝑎,ref
airspeed
controller
Aircraft
Airspeed 𝑢𝑢m
controller
𝜑𝜑, 𝜑𝜑̇
High-level controller
𝑢𝑢𝜃𝜃
𝜃𝜃, 𝜃𝜃̇
𝑣𝑣𝑎𝑎
Low-level controller
𝒑𝒑(𝑋𝑋,𝑌𝑌,𝑍𝑍) , 𝒑𝒑̇ (𝑋𝑋,𝑌𝑌,𝑍𝑍) , 𝑝𝑝̈𝑥𝑥
Figure 9: Controller architecture showing the feedback variables used by the three low-level controllers and by the corresponding three high-level controllers.
Finally, for the airspeed controller we employ a static law as well, however considering the square of the airspeed and
of its reference:
2
um (t) = Km va,ref
(t) − va2 (t) ,
(15)
where Km > 0 is a design parameter. From (8) and (15) we obtain the closed-loop relationship:
s
Km
1
2
va,ref
(t) +
dv (t)
va (t) =
1
Km + 2 ρACD
Km + 12 ρACD a
(16)
which for large enough values of Km (as compared with the terms 21 ρACD and dva (t)) gives:
va (t) ≈ va,ref (t).
As a final remark before proceeding to the description of the high-level controllers, we note that all the control inputs
have hard saturations that need to be enforced, i.e. uϕ ∈ [uϕ , uϕ ], uθ ∈ [uθ , uθ ] and um ∈ [um , um ]. For the sake of
simplicity, we did not include such saturations in the control laws (11), (14b) and (15), however in practice we trim the
control inputs in order not to violate these bounds. Since we have no integrators in our controllers, the presence of the
saturations does not give rise to windup issues. Moreover, as a matter of fact in our tests the saturation values were hit
rarely and always for very short time, indicating that the available control authority was large enough to accomplish the
targeted task. We provide the actual saturation values of our experimental setup in section 4.1.
3.2.3
High-level controllers
We adopt three high-level controllers for the steering, altitude and airspeed dynamics, which respectively compute the
references ϕref (t), θref (t), and va,ref (t). While the low-level controllers never change during operation, we design different
high-level control strategies according to the operational phases outlined in section 2.4, and we switch among them when
a transition from the first phase to the second one is detected. Therefore, in the following we present the high-level
controllers organized by operational phase, as well as the conditions that trigger the phase transition.
Take-Off and climbing. In this phase, the aircraft is initially at rest on the slide, waiting for the acceleration provided
by the latter. A minimum forward acceleration threshold p̈x,to is set by the control designer. When the following condition
is detected:
p̈x ≥ p̈x,to
(17)
then a constant airspeed reference is issued:
va,ref (t) = va,ref,to ,
12
(18)
where the design parameter va,ref,to is chosen as a much larger value than the cruise airspeed of the aircraft, in order to
provide full propeller thrust during the initial climb. Regarding the altitude controller, a constant, relatively large reference
pitch angle θref,to (e.g. 40◦ ) is used:
θref (t) = θref,to
(19)
which gives place to a large vertical speed (see equation (7)). Finally, the roll reference is computed in order to keep a
straight trajectory in the inertial (X, Y ) plane. To this end, we consider the course angle:
ṗY (t)
γ(t) = arctan
(20)
ṗX (t)
and we employ the following feedback controller:
ϕref (t) = Kϕ
|ṗ(t)|
(γref,to − γ(t)) ,
g
(21)
where Kϕ > 0 is a design parameter and γref,to corresponds to the orientation of the ground station’s rails in the inertial
reference frame. The latter can be easily and accurately estimated onboard by using the yaw angle provided by the IMU
in a time interval before the initial acceleration, hence once again avoiding active communication between the ground
station and the aircraft. The rationale behind equation (21) is the following: from (10) we can write
ϕ(t) = ϕref (t) − eϕ (t),
(22)
where the tracking error eϕ (t) can be considered to be small thanks to the low-level controller of the roll dynamics (11),
particularly if we assume that the dominant poles of (12) have been assigned such that the closed-loop roll motion is faster
than the yaw motion (please see section 3.2.4 for more details on the tuning and section 4.1 for the actual values used in
our experiments). We further assume that
ψ̇(t) ' γ̇(t),
(23)
i.e. that the yaw rate and the rate of the course angle are similar, which is a reasonable assumption when the sideslip angle
of the aircraft is small like in our application. Then, using (4) and (21)-(23), we have:
γ̇(t) ' Kϕ (γto − γ(t)) ,
(24)
which is (since Kϕ > 0) a stable first-order system with time constant τγ = 1/Kϕ . Thus, under the proportional control
law (21), the roll angle of the aircraft is controlled in order to track the desired course angle. Finally, we limit the turning
radius of the aircraft to a minimum value Rmin by setting the following bounds on the reference yaw rate (see equation
(5)):
|ṗ(t)|2
|ṗ(t)|2
≤ ϕref (t) ≤
.
(25)
−
g Rmin
g Rmin
Similarly to the other design parameters, in section 4.1 we provide the numerical value chosen for Rmin .
Transition phase and figure-of-eight patterns. When the “safe-altitude” condition:
pZ (t) ≥ Z
(26)
has been reached, the transition phase begins. The goal is to steer the aircraft back towards the ground station and engage
a repetitive figure-of-eight pattern roughly above it, while at the same time continuing to ascend to the target altitude
Zref > Z. To achieve this result, in the airspeed controller we set a constant reference airspeed va,ref (t) = va,ref,flight equal
to the cruise speed of the aircraft.
About the turning controller, we employ the same control law (21), but with a time-varying reference course angle
γref (t) (which was set to the constant value γref,to during the climb) in order to steer the aircraft towards two switching
target points, which are fixed w.r.t. the ground and suitably chosen to achieve the desired flight patterns. In particular,
let us consider two target points p I and p II defined in the inertial (X, Y, Z) plane. On the (X, Y ) plane, these points
are computed in order to be symmetrical w.r.t. to the location of the ground station along the take-off direction γto , and
slightly shifted to one side along the direction perpendicular to γto . The altitude of the points is set equal to Zref . As an
example and without loss of generality, if we assume that the ground station is located at the origin of the inertial frame
and that γto = 0 (i.e. the take-off is carried out along the X−direction), then a suitable choice of the target points is:
−∆X
∆X
ref
I
ref
2
p = Yref , p II =
Zref
2
Yref
,
(27)
Zref
where ∆Xref > 0 is the chosen distance between target points along the take-off direction. A graphical example of this
type of parametrization of the target points (but for γto ' 15◦ ) is shown in the experimental results of Fig. 10.
13
At the start of the transition phase (i.e. when condition (26) is detected), the target point farthest away from the aircraft
is chosen as the active one, denoted as p a . Then, the reference course angle is computed as:
a
pY (t) − pY (t)
γref (t) = arctan
,
(28)
pXa (t) − pX (t)
i.e. the course angle corresponding to a straight line connecting the current (X, Y ) position of the aircraft with that of
the target. The switching of the active target point happens when the aircraft’s (X, Y ) position surpasses the position of
the current target point, after projecting both positions on a direction corresponding to the take-off course γref,to . In the
previous example (27) with γref,to = 0, the corresponding switching rule would be:
I
ref
+ δX
p if pX (t) < − ∆X
2
a
∆Xref
II
p (t) =
(29)
p if pX (t) > 2 − δX
a −
p (t ) else
where δX > 0 is a small tolerance (e.g. 0.5 m) to avoid possible numerical issues when computing the arctan function
in (28), and p a (t− ) is the previous active target point (in the discrete-time implementation of the controller –see section
4.1– this corresponds to the target point at the previous sampling instant). The value of γref (t) computed by (28) is finally
plugged into (21) in place of γref,to to close the high-level turning control loop.
About the altitude controller, in this phase the aim is to regulate the aircraft Z position close to the reference Zref .
To achieve this goal, differently from the initial take-off and climbing phase we employ a static proportional feedback
controller to compute θref (t):
Kθ
(Zref − pZ (t))
(30)
θref (t) =
|ṗ(t)|
where Kθ > 0 is a design parameter. Similarly to what shown in (24) for the turning dynamics, from (14) we have:
θ(t) = θref (t) − eθ (t),
with small tracking error eθ (t) thanks to the low-level controller. Then, considering also (7) and (30) we obtain:
ṗZ (t) ' Kθ (Zref − pZ (t)) ,
(31)
which is a stable first-order system with time constant τθ = 1/Kθ .
3.2.4
Controller tuning and stability/robustness aspects
The described controller features several tuning parameters, which can be initially set on the basis of standard control
design guidelines and physical insight, and then fine-tuned directly during experimental tests. We now provide for completeness some suggestions about possible tuning choices and procedures. The final numerical values employed in our
tests are shown in section 4.1
The closed-loop eigenvalues λϕ,1 , λϕ,2 of the low-level roll controller and the gain Kϕ of the high-level turning
controller have to be chosen such that the bandwidth of the inner closed-loop dynamics (12) is sufficiently larger than the
one of the outer loop (24), i.e.:
|λϕ,1 |, |λϕ,2 | > Kϕ ,
(32)
and similarly for the pitch/altitude controllers:
|λθ,1 |, |λθ,2 | > Kθ .
(33)
The low-level controllers can be fine-tuned in real-world tests by using a fly-by-wire mode, where a pilot on the ground
issues reference roll and pitch values to the aircraft and adjusts the values of the closed-loop eigenvalues. When the latter
have been fixed, the gains of the outer loop can be also fine-tuned during fully autonomous flight tests.
The airspeed controller gain Km can be initially set in order to provide full motor power (resp. stop the motor) when
the measured airspeed is larger (resp. smaller) than the reference by a certain amount, e.g. 30 %, and then fine-tuned
during experiments.
The acceleration threshold p̈x,to should be chosen in order to be sure to detect the slide movement, while avoiding
false-positives e.g. due to vibrations induced by wind gusts. The actual acceleration experienced by the glider depends
mainly on the rails’ length and on the take-off speed (assuming that all of the available length is exploited in order to limit
the peak power consumed during take-off). As an example, in our tests the initial acceleration peak was about 40 N/m2 ,
and a threshold of 20 N/m2 was large enough to avoid false-positives.
As already mentioned, the values of va,ref,to and θref,to shall be chosen large enough to provide full motor thrust (taking
into account the value of Km ) and a large climb angle during the take-off, in order to quickly gain altitude and increase
14
the relative angle between the tether and the aircraft’s velocity vector, hence reducing the risk of stall due to the tether’s
pulling force. The “safe-altitude” value Z shall be chosen to have enough safety margin in terms of distance from the
ground, while avoiding to climb for too long on a straight line, with consequent large values of unreeled tether length.
Finally, the minimum radius Rmin , the cruise airspeed va,ref,flight , and the position of the target points p I and p II shall be
chosen in order to keep the resulting flight paths compact (to reduce the occupied air volume and the length of the tether)
but at the same time avoiding too large roll angles during the turns, taking into account the aerodynamic/manoeuvrability
features of the employed aircraft.
Before proceeding to the experimental results, we finally comment on the closed-loop stability and robustness of the
proposed control approach. Regarding nominal stability (i.e. assuming that the model equations introduced in section 3.1
hold with no uncertainty or disturbances), the choice of closed-loop eigenvalues with negative real-part for the low-level
controllers and of positive gains Km , Kϕ and Kθ is sufficient to guarantee that the closed-loop system is asymptotically
stable, which implies that bounded exogenous perturbations (e.g. due to the wind or the tether) give rise to bounded
tracking errors. Moreover, since all of the feedback variables are directly measured, i.e. there is no observer involved,
the use of state-feedback controllers provides good robustness margins against neglected dynamics, couplings between
the different modes (which are neglected in our model) and parametric uncertainty [28]. Thus, the control system can
tolerate quite large inaccuracies in the estimate of all of the unknown parameters in the model (i.e. aϕ , bϕ , aθ , bθ , CD ,
see section 3.1) without compromising stability and performance.
Albeit these few comments do not represent a rigorous robustness analysis, which lies beyond the scope of this paper,
they provide some insight on why a relatively simple approach like the one presented here provides good results in realworld experiments, as shown in the next section.
4
4.1
Experimental results
Controller implementation, system and design parameters
We implemented the described controller on the Arduino MEGA 2560 board installed on the glider. We employed a
sampling frequency of 50 Hz, which is about two orders of magnitude larger than the fastest closed-loop dynamics of the
system. Since the controller features only static feedback functions, there is no modification needed due to the discretization. In addition to the main control functions presented in the previous section, we programmed several consistency
checks on the measured variables, in order to detect a possible sensor malfunctioning, as well as safety measures, like an
automatic tether detach procedure if the distance of the aircraft from the ground station is too close to the maximum tether
length (150 m). Since these additional functionalities do not change the system behavior in normal conditions, we omit
further details here.
The model parameters (resulting either from the identification procedure described in section 3.1 or from prior knowledge of the aircraft) are shown in Table 2, together with the employed values of the controller’s parameters. We followed
the tuning guidelines described in the previous section, fine-tuned the parameters during initial flight tests and then fixed
them to the values of Table 2 for all the subsequent experiments.
4.2
Results
We carried out several outdoor tests using our experimental setup and the described control system. All the tests were
carried out close to Leibstadt, Switzerland, at a location with little/moderate wind on average, with prevalent direction
from East/North-East. We present here the experimental data of 14 tests carried out with the same take-off course angle
γto . During the tests, the wind conditions changed from little/zero wind to front wind of about 4-5 m/s, to side wind of
about 3-4 m/s. Thus, we are able to present results with both undisturbed conditions (i.e. either no wind or little front
wind) and in presence of moderate wind disturbance coming from different directions. All these considerations hold for
the wind at ground level, while at the target altitude (50 m) a moderate wind and wind gusts were always present, of the
order of 3-4 m/s as estimated by experienced model glider pilots. Since we had no equipment to measure the wind at
altitudes higher than a few meters, we cannot provide more precise estimates of the wind conditions experienced by the
glider after the very first take-off instants. A movie is available online [29], showing an autonomous take-off and flight
test with the initial take-off, the transition phase and finally the figure-of-eight flights.
We first present the results of a test with little wind at ground level. The flight trajectory projected on the inertial
(X, Y ) plane centered at the ground station is shown in Fig. 10(a), and the same trajectory overlayed on a satellite image
of the corresponding GPS coordinates is presented in Fig. 10(b). The initial take-off phase on a straight line is clearly
visible, as well as the transition to the figure-of-eight paths.
Figs. 11(a)-(c) show the behavior of the turning dynamics (roll and course angle controllers) during the test. The
discontinuities in γref (t) (Fig. 11(a), solid line) correspond to the switching instants between the target points. It can be
clearly seen that the actual course angle γ(t) (Fig. 11(a), dashed line) changes at a constant rate during these transitions,
which corresponds to the turning rate at cruise speed and with the minimum turning radius of 20 m set in the controller.
15
Table 2: System and controller parameters for the experimental tests
System parameters
aϕ
−2.3 s−1
bϕ
12.6 s−2
aθ
−4.65 s−1
bθ
30 s−2
CD
0.05 ρ
1.2 kg m−3
A
0.3 m2
Controller parameters - low level
λϕ,1
−2.7 s−1
λϕ,2
−3.1 s−1
λθ,1
−2.7 s−1
λθ,2
−3.1 s−1
Km
0.5 kg m−1
[uϕ uϕ ]
[−0.34 0.34] rad
[uθ uθ ]
[−0.34 0.34] rad
[um um ]
[0 20] N
Controller parameters - high level
Kϕ
1 s−1
Kθ
0.1 s−1
p̈x,to
20 m s−1
va,ref,to
16 m s−1
θref,to
0.69 rad
Z
20 m
Rmin
20 m
va,ref,flight
13 m s−1
I
I
I
[pX pY pZ ]
[30 55 50] m
[pXII pYII pZII ]
[−30 40 50] m
The corresponding roll angle ϕ(t) (Fig. 11(b)) is approximately constant and equal to about 35◦ during these turns.
Finally, the employed roll input values uϕ (Fig. 11(c)) are usually smaller than 10◦ with few peaks at 15◦ , quite far
from the saturation values of ±20◦ . Overall, the inner and outer loops for the roll and turning dynamics show very good
tracking performance and provided a very consistent and robust behavior.
The closed-loop altitude and pitch motions are shown in Fig 12(a)-(b). The aircraft altitude quickly approaches the
target of 50 m during the initial take-off and transition. The vertical rate is constant because, similarly to what happens
for the rate of the course angle, the aircraft sets to constant pitch and airspeed, which according to our model (7) yields a
constant vertical position rate. During flight, the high-level controller is able to track with good accuracy (about 3-4 m of
error) the target altitude, see Fig. 12(a) e.g. between 45 s and 60 s, except for periodic, quick drops in the altitude which
are limited to less than 10 m of tracking error. A comparison between the distance of the aircraft from the ground station
and the tether length, shown in Fig. 12(c), reveals that such disturbances are due to the tether pull: during each figureof-eight pattern the aircraft pulls on the tether for a fraction of the time and some length of slack line (also shown in Fig.
12(c)) is created. After the initial transient, the behavior of such a slack line becomes periodic as well, since the overall
system (aircraft and ground station) converges to a stable periodic motion. The length of slack line depends on how fast
the winch reels-in the line compared to how fast the glider changes its directions and flies back towards the ground station
during the figure-of-eight patterns. The compression of the spring installed on the ground station is depicted in 12(c): we
recall that this signal is used by the winch controller to regulate the winch reeling speed and decide whether to reel-out or
reel-in (as briefly described in section 2.3 and thoroughly presented in [22]).
Overall, the presented results show that the altitude controller behaves as expected and as discussed in section 3.2.4:
in nominal/unperturbed conditions the tracking performance is very good, and in the presence of external disturbances the
system reacts with bounded tracking errors. Moreover, it is shown that there is quite some room for improvement, either
by better tuning or more advanced control strategies of the ground station, or by allowing the aircraft and ground station
to communicate and optimize their behavior in a coordinated way.
Regarding the propeller controller, Fig. 13 shows the courses of the reference airspeed and of the actual one, as well
as the corresponding propeller thrust um . Similarly to the pitch controller, with no tether perturbations a small tracking
16
(a)
100
75
Y (m)
50
25
0
-25
-50
-100
-75
-50
-25
0
X (m)
25
50
75
100
(b)
47.589
Latitude (deg)
47.5888
47.5886
47.5884
47.5882
47.588
8.197
8.1975
8.198
Longitude (deg)
8.1985
8.199
Figure 10: Experimental results of a test with little wind at ground level. (a) Typical flight pattern in the inertial (X, Y )
plane centered at the ground station (marked with ’◦’). The target points p I and p II are marked with ’4’. (b) Flight pattern
in GPS coordinates, overlaid to a satellite map of the area. The white dots along the trajectory are position measurements
down-sampled at 2 Hz.
error of about 0.5 m s−1 is achieved, while when the tether pulls on the aircraft a drop of airspeed is noticed, to which the
controller reacts with an increase of thrust up to the maximum value. The loss of airspeed has consequences on all the
other motions (pitch and roll), since it changes suddenly the angle of attack and the incoming flow speed. However, as
already commented, the proposed controller is able to effectively cope with all such nonlinear effects.
Finally, we present the results pertaining to all the 14 tests carried out with the same starting course angle, to show
the repeatability and robustness of the controller in conditions with little to moderate wind disturbance. Fig. 14 shows
the course of aircraft altitude for all the tests: it can be seen that the trajectories are quite consistent, except for one
occurrence. The aircraft position on the (X, Y ) plane (in GPS coordinates) for all the tests is shown in Fig. 15: there is
some variability in the initial transient and in the first few figure-of-eight patterns, until the aircraft and the whole system
settle to a periodic motion. After such a transient, the patterns are more consistent but still exhibit some variability, which
is expected: since the high-level controller relies on a two-point strategy only (instead for example on the tracking of a
fully prescribed orbit), the actual obtained patterns can change e.g. depending on the wind conditions, similarly to what
happens for the two-point strategies used for crosswind kite control in e.g. [13] and [10]. However, as shown in Fig. 16
where the (X, Y ) trajectories are presented without the initial transient, in all the performed tests the control system was
able to reach and consistently follow the figure-of-eight paths, which was the ultimate intended goal of our feasibility
study.
17
5
Conclusions and future work
We presented a feedback controller for a tethered aircraft, able to consistently carry-out tethered take-off and low-force
flight maneuvers. The controller works in combination (yet without any exchange of information) with a ground station
that manages the initial acceleration on ground and controls the tether reeling, in order to limit the pulling force. The
presented approach includes both modeling and control design aspects and has been extensively tested in experiments,
whose results have been reported here as well.
Our experimental results lead to essentially two concluding remarks. The first one is that a rather simple control
approach, based on proportional controllers only and neglecting system nonlinearities, couplings among the aircraft’s
modes and the presence of the tether, is able to achieve satisfactory results in terms of performance and robustness. This
indicates that the considered take-off approach appears to be technically viable and, on the basis of our previous studies
[20], it has the potential to also be economically viable, since it requires relatively small additional power both on ground
and onboard and is able to carry out the take-off in very compact space. In our small-scale prototype, less than 2 m of rails’
length were enough to take-off. According to [20], with a larger wing loading the required take-off distance would still
remain of the order of the aircraft’s wingspan, similarly to other AWE technologies which employ e.g. vertical take-off
approaches, like Makani Power [7].
The second consideration is that there is quite some room for improvement of the controller, along the following (and
possibly more) lines of further research: advanced multi-variable controllers taking into account the couplings among the
aircraft’s modes; nonlinear control approaches; adaptive approaches based on the estimation of the wind conditions and
on the aircraft’s angle of attack and sideslip; coordinated control of the aircraft and of the ground station; iterative learning
approaches; improved/specialized aircraft design also considering the presence of the tether.
As a final remark, we believe that the next major technical bottleneck for the considered AWE technology, on which
future research efforts should be targeted, is the feasibility proof (theoretical and experimental) of an autonomous landing
strategy with the same qualitative features as the take-off approach that we demonstrated in this paper, i.e. small required
land area, low cost, repeatability and robustness. It may well be that the removal of such a bottleneck entails not only
control technology, but also the development of new aircraft concepts specifically designed for this purpose.
Acknowledgment
The authors would like to thank Alessandro Lauriola and Stefan Schmidt for their helpful contributions during the project.
References
[1] U. Ahrens, M. Diehl, and R. Schmehl, Eds., Airborne Wind Energy.
Berlin Heidelberg: Springer, 2013.
[2] L. Fagiano and M. Milanese, “Airborne wind energy : an overview,” in Proceedings of the 2012 American Control
Conference. Montréal, Canada: IEEE, June 2012, pp. 3132–3143.
[3] M. L. Loyd, “Crosswind kite power,” Journal of Energy, vol. 4, no. 3, pp. 106–111, 1980.
[4] A. Ilzhöfer, B. Houska, and M. Diehl, “Nonlinear MPC of kites under varying wind conditions for a new class of
large-scale wind power generators,” International Journal of Robust and Nonlinear Control, vol. 17, no. 17, pp.
1590–1599, 2007.
[5] L. Fagiano, M. Milanese, and D. Piga, “High-altitude wind power generation,” IEEE Transactions on Energy
Conversion, vol. 25, no. 1, pp. 168–180, 2010. [Online]. Available: http://ieeexplore.ieee.org/xpls/abs_all.jsp?
arnumber=5350676
[6] M. Canale, L. Fagiano, and M. Milanese, “High altitude wind energy generation using controlled power kites,” IEEE
Transactions on Control Systems Technology, vol. 18, no. 2, pp. 279 –293, mar. 2010.
[7] D. Vander Lind, “Analysis and flight test validation of high performance airborne wind turbines,” in Airborne Wind
Energy. Berlin Heidelberg: Springer, 2013, ch. 28, pp. 473–490.
[8] R. van der Vlugt, J. Peschel, and R. Schmehl, “Design and experimental characterization of a pumping kite power
system,” in Airborne Wind Energy, U. Ahrens, M. Diehl, and R. Schmehl, Eds. Berlin Heidelberg: Springer, 2013,
ch. 23, pp. 403–425.
[9] C. Vermillion, T. Grunnagle, R. Lim, and I. Kolmanovsky, “Modeling, control design, and experimental validation
of a prototype lighter-than-air wind energy system,” IEEE Transactions on Control Systems Technology, vol. 22,
no. 2, pp. 531–542, 2014.
18
[10] M. Erhard and H. Strauch, “Flight control of tethered kites in autonomous pumping cycles for airborne wind energy,”
Control Engineering Practice, vol. 40, pp. 13–26, 2015.
[11] A. Zgraggen, L. Fagiano, and M. Morari, “Automatic retraction and full-cycle operation for a class of airborne wind
energy generators,” IEEE Transactions on Control Systems Technology, vol. 24, no. 2, pp. 594–698, 2016.
[12] R. Ruiterkamp and S. Sieberling, “Description and preliminary test results of a six degrees of freedom rigid wing
pumping system,” in Airborne Wind Energy, U. Ahrens, M. Diehl, and R. Schmehl, Eds. Berlin Heidelberg:
Springer, 2013, ch. 26, pp. 443–458.
[13] L. Fagiano, A. U. Zgraggen, M. Morari, and M. Khammash, “Automatic crosswind flight of tethered wings for
airborne wind energy: modeling, control design and experimental results,” IEEE Transactions on Control System
Technology, vol. 22, no. 4, pp. 1433–1447, 2014.
[14] U. Fechner, R. van der Vlugt, E. Schreuder, and R. Schmehl, “Dynamic model of a pumping kite power system,”
Renewable Energy, 2015.
[15] Ampyx Power website, http://www.ampyxpower.com/.
[16] TwingTec website, http://www.twingtec.ch.
[17] S. Gros, M. Zanon, and M. Diehl, “Control of airborne wind energy systems based on nonlinear model predictive
control & moving horizon estimation,” in European Control Conference (ECC) 2013, Zuerich, Switzerland, July
2013, 2013, pp. 1017–1022.
[18] E. Bontekoe, “Up! - how to launch and retrieve a tethered aircraft,” Master’s thesis, TU Delft, August 2010, accessed
in August 2015 at http://repository.tudelft.nl/.
[19] L. Fagiano, E. Nguyen Van, and S. Schnez, Linear Take-Off and Landing of a Rigid Aircraft for Airborne Wind
Energy Extraction, ser. Green Energy and Technology. Springer Singapore, 2016, vol. Airborne Wind Energy,
Advances in Technology Development and Research, iSBN 978-981-10-1947-0.
[20] L. Fagiano and S. Schnez, “On the take-off of airborne wind energy systems based on rigid wings,” Submitted,
available on arXiv:1510.06701v1.
[21] M. Kruijff, “Preparing for a commercially viable awe system in the utility sector,” oral presentation at The International Airborne Wind Energy Conference 2015,
available online:
https://collegerama.tudelft.nl/Mediasite/Play/2e1f967767d541b1b1f2c912e8eff7df1d (last accessed 26 October 2105).
[22] L. Fagiano, E. Nguyen-Van, F. Rager, S. Schnez, and C. Ohler, “A small-scale prototype to study the take-off
of tethered rigid aircrafts for airborne wind energy,” Submitted. Preliminary version available on ArXiv e-prints:
1608.01846, vol. http://arxiv.org/abs/1608.01846, 2016, submitted.
[23] SBG Systems. (2016) Ellipse-N navigation unit. [Online]. Available: http://www.sbg-systems.com/
[24] Arduino. (2016) Arduino MEGA 2560 board. [Online]. Available:
ArduinoBoardMega2560
[25] B. Etkin, Dynamics of Atmospheric Flight.
https://www.arduino.cc/en/Main/
Dover, 2005.
[26] E. Nguyen-Van, L. Fagiano, and S. Schnez, “On the autonomous take-off and landing of tethered wings for airborne
wind energy,” in American Control Conference 2016, Boston, MA, July 2016, pp. 4077–4082
[27] M. Erhard and H. Strauch, “Control of towing kites for seagoing vessels,” IEEE Transactions on Control Systems
Technology, vol. 21, no. 5, pp. 1629–1640, 2012.
[28] G. C. Goodwin, S. F. Graebe, and M. E. Salgado, Control System Design.
2001.
Upper Saddle River, NJ: Prentice-Hall,
[29] L. Fagiano, E. Nguyen-Van, F. Rager, S. Schnez, and C. Ohler, “Autonomous tethered take-off and flight for airborne
wind energy - movie,” YouTube: https://www.youtube.com/watch?v=UPiTiHPXciE, March 2016.
19
(a)
500
450
400
Course angle (°)
350
300
250
200
150
100
50
0
0
20
40
60
80
100
Time (s)
120
140
160
180
120
140
160
180
120
140
160
180
(b)
60
40
Roll angle (°)
20
0
-20
-40
-60
0
20
40
60
80
100
Time (s)
(c)
15
10
Aileron input (°)
5
0
-5
-10
-15
-20
0
20
40
60
80
100
Time (s)
Figure 11: Experimental results corresponding to Fig. 10, turning dynamics. (a) Course angle γ (dashed) and its reference
γref (solid) during tethered launch, climb and flight. (b) Roll angle ϕ (dashed) and its reference ϕref (solid) in the same
test. (c) Aileron input uϕ .
20
(a)
60
Altitude from ground (m)
50
40
30
20
10
0
0
20
40
60
80
100
120
140
160
180
120
140
160
180
Time (s)
(b)
50
40
Pitch angle (deg)
30
20
10
0
-10
-20
-30
0
20
40
60
80
100
Time (s)
(c)
Distance (m), spring compression (10 -2 m)
120
100
80
60
40
20
0
0
30
60
90
Time (s)
120
150
180
Figure 12: Experimental results corresponding to Fig. 10, pitch and altitude dynamics. (a) Courses of the reference
altitude pZ,ref (dotted line) and the actual one, pZ (dashed). (b) Courses of the reference pitch angle θref (solid), and the
actual one, θ (dash-dot). (c) Courses of the aircraft’s distance from the ground station, kpk2 (solid line), of the tether
21 slack tether length (dash-dotted). All plotted variables are
length (dashed), of the spring compression (dotted) and of the
−2
in m, except for the spring compression in 10 m.
Speed (m/s)
16
12
8
4
0
0
20
40
60
80
100
Time (s)
120
140
160
180
20
40
60
80
100
Time (s)
120
140
160
180
Motor thrust (N)
20
16
12
8
4
0
0
Figure 13: Experimental results corresponding to Fig. 10, airspeed and motor thrust. Upper plot: courses of the reference
airspeed va,ref (solid) and of the actual one, va (t) (dashed). Lower plot: propeller thrust um .
60
50
Altitude (m)
40
30
20
10
0
0
5
10
15
20
25
30
Time (s)
35
40
45
50
Figure 14: Experimental results of 14 overlayed tests. Course of the aircraft altitude.
22
47.589
Latitude (deg)
47.5888
47.5886
47.5884
47.5882
47.588
47.5878
8.197
8.1975
8.198
Longitude (deg)
8.1985
8.199
Figure 15: Experimental results of 14 overlayed tests. Aircraft position in GPS coordinates. The white dots correspond to
position measurements down-sampled to 1 Hz. The ground station position is marked with ’◦’. The target points p I and
p II are marked with ’4’.
47.5892
47.589
Latitude (deg)
47.5888
47.5886
47.5884
47.5882
47.588
8.197
8.1975
8.198
Longitude (deg)
8.1985
8.199
Figure 16: Experimental results of 14 overlayed tests. Aircraft position in GPS coordinates without the initial take-off
and transition phases. The white dots correspond to position measurements down-sampled to 1 Hz. The ground station
position is marked with ’◦’. The target points p I and p II are marked with ’4’.
23
| 3cs.SY
|
arXiv:1608.03137v1 [math.GR] 10 Aug 2016
On the structure of virtually nilpotent compact
p-adic analytic groups
Billy Woods
August 11, 2016
Abstract
Let G be a compact p-adic analytic group. We recall the well-understood finite radical ∆+ and FC-centre ∆, and introduce a p-adic analogue
of Roseblade’s subgroup nio(G), the unique largest orbitally sound open
normal subgroup of G. Further, when G is nilpotent-by-finite, we introduce the finite-by-(nilpotent p-valuable) radical FNp (G), an open characteristic subgroup of G contained in nio(G). By relating the already
well-known theory of isolators with Lazard’s notion of p-saturations, we
introduce the isolated lower central (resp. isolated derived) series of a
nilpotent (resp. soluble) p-valuable group of finite rank, and use this to
study the conjugation action of nio(G) on FNp (G). We emerge with a
structure theorem for G,
1 ≤ ∆+ ≤ ∆ ≤ FNp (G) ≤ nio(G) ≤ G,
in which the various quotients of this series of groups are well understood.
This sheds light on the ideal structure of the Iwasawa algebras (i.e. the
completed group rings kG) of such groups, and will be used in future work
to study the prime ideals of these rings.
1
Contents
1 Preliminaries
7
2 The Roseblade subgroup nio(G)
12
3 p-saturations
16
4 Conjugation action of G
21
5 The finite-by-(nilpotent p-valuable) radical
24
2
Introduction
We aim to study the structure of certain compact p-adic analytic groups G. This
will crucially underpin later work in which we will explore the ring-theoretic
properties of completed group rings kG, where k is a finite field of characteristic
p and G is a nilpotent-by-finite compact p-adic analytic group.
There is much in common between the theory of polycyclic-by-finite groups
(and their group rings) and the theory of polyprocyclic-by-finite compact p-adic
analytic groups (and their completed group rings). See [4] and [5], or the more
recent survey paper by Ardakov and Brown [3], for an overview of the latter.
It is known that
)
(
compact p-adic
uniform groups
p-valuable groups
analytic groups
,
⊆
⊆
[4, Definition 4.1]
[5, III, 2.1.2]
[4, Definition 8.14]
with the first inclusion coming from [4, Definition 1.15; notes at end of chapter
4], and the second from [4, Corollary 8.34].
Note also that compact p-adic analytic groups G are profinite groups satisfying
Max: every nonempty set of closed subgroups of G contains a maximal element.
Indeed, in the case when G is p-valuable, this follows from [5, III, 3.1.7.5]; in the
general case, G contains a uniform open normal subgroup U by [4, 8.34]. Indeed,
[4, 8.34] implies that compact p-adic analytic groups are precisely extensions of
uniform (or p-valuable) groups by finite groups.
We aim eventually to extend some of the work of Roseblade [13] and Letzter
and Lorenz [6] to the domain of compact p-adic analytic groups, building on
work by Ardakov [2]. Our main results are as follows.
Let G be a compact p-adic analytic group, and H a closed subgroup. Following
Roseblade [13], we will say that H is orbital (or G-orbital ) if it only has finitely
many G-conjugates, or equivalently if its normaliser NG (H) is open in G; and
H is isolated orbital (or G-isolated orbital ) if H is orbital, and given any other
closed orbital subgroup H ′ of G with H H ′ , we have [H ′ : H] = ∞. G is then
said to be orbitally sound if all its isolated orbital closed subgroups are in fact
normal.
We define the Roseblade subgroup
nio(G) =
\
NG (H),
H
where this intersection is taken over all isolated orbital closed subgroups H of
G. In section 2, we prove:
Theorem A. Let G be a compact p-adic analytic group. Then nio(G) is an
orbitally sound, open, characteristic subgroup of G, and contains all finite-bynilpotent closed normal subgroups of G.
3
Now let G be a p-valuable group. Ardakov shows, in [2, Lemma 8.4(a)], that
the centre Z(G) is isolated orbital in G; we may then deduce from Lemma 1.5(i)
that the usual upper central series
1 ≤ Z(G) ≤ Z2 (G) ≤ . . .
of G as an abstract group consists of isolated orbital subgroups of G, so that any
p-valuation on G naturally induces p-valuations on each Zi (G) and G/Zi (G).
Unfortunately, in general, the (abstract) lower central series {γi } (or derived
series {Di }) of G will not necessarily consist of isolated orbital subgroups, so
G/γi (or G/Di ) will not necessarily remain p-valuable.
In section 3, we introduce appropriate “isolated” analogues of these series for
p-valuable groups:
Theorem B. Let G be a p-valuable group. Then there exists a unique fastest
descending series of isolated orbital closed normal subgroups of G, the isolated
lower central series,
G = G1 ✄ G2 ✄ . . . ,
with the properties that each Gi is characteristic in G, Gi /Gi+1 is abelian for
each i, [Gi , Gj ] ≤ Gi+j for all i and j, and there exists some r with Gr = 1 if
and only if G is nilpotent.
There exists also a unique fastest descending series of isolated orbital closed
normal subgroups of G, the isolated derived series,
G = G(0) ✄ G(1) ✄ . . . ,
with the properties that each G(i) is characteristic in G, G(i) /G(i+1) is abelian
for each i, and there exists some r with G(r) = 1 if and only if G is soluble.
Now let G be a fixed compact p-adic analytic group. Define the two closed
subgroups
∆ = x ∈ G [G : CG (x)] < ∞ ,
∆+ = x ∈ ∆ o(x) < ∞ ,
where o(x) denotes the order of x. We will show that we always have an inclusion
of subgroups
1 ≤ ∆+ ≤ ∆ ≤ nio(G) ≤ G.
Suppose further that G is a nilpotent-by-finite compact p-adic analytic group.
Consider the set of finite-by-(nilpotent p-valuable) open normal subgroups H of
G – that is, the set of open normal subgroups H that contain a finite normal
subgroup F ✁ H, such that H/F is nilpotent and p-valuable. The main result
of section 5 is:
4
Theorem C. This set contains a unique maximal element H, which is characteristic in G. There is an inclusion of subgroups
∆ ≤ H ≤ nio(G),
and the quotient group nio(G)/H is isomorphic to a subgroup of the group of
torsion units of Zp .
We will denote this H by FNp (G), the finite-by-(nilpotent p-valuable) radical
of G.
Putting this together with the technical results of section 4, we can now understand the conjugation action of nio(G) on FNp (G):
Theorem D. Let G be a nilpotent-by-finite compact p-adic analytic group, and
H = FNp (G) its finite-by-(nilpotent p-valuable) radical. Write
N = N1 ✄ N2 ✄ · · · ✄ Nr = 1
for the isolated lower central series of N = H/∆+ , and let Hi be the full
preimage in G of Ni . Then nio(G)/H is isomorphic to a subgroup of t(Z×
p ), the
group of torsion units of Zp , and so is cyclic; let a be a preimage in nio(G) of
a generator of nio(G)/H. Then conjugation by a acts on each Hi , and hence
induces an action on the (free, finite-rank) Zp -modules Hi /Hi+1 = Ai . In
a
ζi
multiplicative notation, there is some scalar ζ ∈ t(Z×
for
p ) such that x = x
all x ∈ Ai .
Putting these ingredients together allows us to understand the structure of a
nilpotent-by-finite compact p-adic analytic group G, which, for convenience, we
display in the following diagram.
5
G
nio(G)
.t(Z×
p )
H1
a: ζ
H
1∼
Zd
p =A1
H2
..
.
nilpotent, p-valuable
Hr−1
a: ζ r−1
d
Zpr−1 ∼
=Ar−1
∆+
Hr
1
6
1
Preliminaries
Definition 1.1. Let G be a profinite group. A closed subgroup H of G is Gorbital (or just orbital, when the group G is clear from context) if H has only
finitely many G-conjugates, i.e. if NG (H) is open in G. Similarly, an element
x ∈ G is orbital if [G : CG (x)] < ∞.
Remark. Note that, if G is compact p-adic analytic, it is profinite (by [4, 8.34]).
Definition 1.2. The FC-centre ∆(G) of an arbitrary group G is the subgroup
of all orbital elements of G. The finite radical ∆+ (G) of G is the subgroup of
all torsion orbital elements of G.
Remark. Throughout this paper, we will write as shorthand ∆+ = ∆+ (G)
and ∆ = ∆(G). Also throughout this paper, all subgroups will be closed, all
homomorphisms continuous, etc. unless otherwise specified.
Lemma 1.3. Let G be a compact p-adic analytic group. For convenience, we
record a few basic properties of ∆ and ∆+ .
(i) ∆+ is finite.
(ii) If H is an open subgroup of G, then ∆+ (H) ≤ ∆+ (G) and ∆(H) ≤ ∆(G).
(iii) When G is compact p-adic analytic, ∆+ and ∆ are closed in G.
(iv) ∆+ and ∆ are characteristic subgroups of G.
(v) ∆/∆+ is a torsion-free abelian group.
Proof.
(i) ∆+ is generated by the finite normal subgroups of G [11, 5.1(iii)]. It is
obvious that the compositum of two finite normal subgroups is again finite
and normal. Now suppose that ∆+ is infinite, and take an open uniform
subgroup H of G [4, 4.3]: then ∆+ ∩ H is non-trivial, and so we must have
some finite normal subgroup F with F ∩ H non-trivial. But F is torsion,
so this contradicts the fact that H is torsion-free [4, 4.5].
(ii) If an element x ∈ H has finitely many H-conjugates, and H has finite
index in G, then x has finitely many G-conjugates.
(iii) ∆+ is closed because it is finite.
For the case of ∆, suppose first that G is p-valued [5, III, 2.1.2]. Now,
any orbital x ∈ G has CG (x) open in G, and so, for any g ∈ G, there
n
n
n
exists some n with g p ∈ CG (x), i.e. g p x = xg p . This implies that
x pn
pn
x
(g ) = g , and so by [5, III, 2.1.4], we get g = g. Hence CG (x) = G.
In other words, ∆ = Z(G), which is closed in G.
When G is not p-valued, it still has an open p-valued subgroup N [4, 4.3].
Clearly ∆(N ) = ∆(G) ∩ N , and so [∆(G) : ∆(N )] ≤ [G : N ] < ∞. So
7
∆(G) is a finite union of translates of Z(N ), which is closed in N and
hence closed in G.
(iv) See [10, discussion after lemma 4.1.2 and lemma 4.1.6].
(v) See [10, lemma 4.1.6].
Throughout the remainder of this subsection, G is a profinite group
unless stated otherwise.
Definition 1.4. An orbital closed subgroup H of G is isolated if, for all orbital
closed subgroups H ′ of G with H
H ′ ≤ G, we have [H ′ : H] = ∞. (We
will sometimes say that a closed subgroup is G-isolated orbital as shorthand for
isolated as an orbital closed subgroup of G.) Following Passman [11, definition
19.1], if all isolated orbital closed subgroups of G are in fact normal, we shall
say that G is orbitally sound.
We record a few basic properties, before showing that this definition is the same
as the one given in [13, 1.3] and [2, 5.8] (in Lemma 1.10 below).
Lemma 1.5. Let N be a closed normal subgroup of G.
(i) Suppose H is a closed subgroup of G containing N . Then H/N is (G/N )orbital if and only if H is G-orbital; and H/N is (G/N )-isolated orbital if
and only if H is G-isolated orbital.
(ii) Suppose G is orbitally sound. Then G/N is orbitally sound.
(iii) Suppose N is finite and G/N is orbitally sound. Then G is orbitally sound.
Proof.
(i) It is easily checked that NG/N (H/N ) = NG (H)/N , and so
[G : NG (H)] = [G/N : NG (H)/N ] = [G/N : NG/N (H/N )].
So H is orbital if and only if H/N is orbital. Suppose these two groups
are both orbital, and let H ′ be an orbital closed subgroup of G with
H H ′ ≤ G: then [H ′ : H] = [H ′ /N : H/N ], so H is isolated if and only
if H ′ is isolated.
(ii) Let H/N be an isolated orbital closed subgroup of G/N . Then, by (i), H
is an isolated orbital subgroup of G, so H ✁ G, and so H/N ✁ G/N .
(iii) Let H be an isolated orbital closed subgroup of G, and H ′ an orbital
closed subgroup of G with H
H ′ ≤ G. If H contains N , then we
′
may apply (i) to show that [H : H] = [H ′ /N : H/N ] = ∞ as G/N is
orbitally sound. But H must contain N : indeed, as H is G-orbital, clearly
HN/N is G/N -orbital, and so, by (i), HN is G-orbital. But N is finite,
so [HN : H] < ∞, and H is isolated, so H = HN .
8
From now on, we assume that G is a profinite group satisfying the
maximum condition on closed subgroups: every nonempty set of closed
subgroups of G has a maximal element.
Remark. Note that, if G is compact p-adic analytic, it satisfies the maximum
condition on closed subgroups. Indeed, this is true for p-valuable G by [5, III,
3.1.7.5], and hence true for any compact p-adic analytic group G, as G contains
a uniform (hence p-valuable) subgroup of finite index [4, 8.34].
Definition 1.6. If H is an orbital closed subgroup of G, we define its isolator
iG (H) in G to be the closed subgroup of G generated by all orbital closed
subgroups L of G containing H as an open subgroup, i.e. with [L : H] < ∞.
Once we have proved that iG (H) is indeed an isolated orbital closed subgroup
of G containing H as an open subgroup, it will be clear from the definition that
it is the unique such closed subgroup.
We now prove some basic properties of iG (H), following [11].
Proposition 1.7. Suppose H is an orbital closed subgroup of G. Then H is
open in iG (H).
Proof. We first show that, if L1 and L2 are orbital subgroups of G T
containing
H as an open subgroup, then [hL1 , L2 i : H] < ∞. Write (−)◦ for g∈G (−)g .
Suppose without loss of generality that G = hL1 , L2 i, and that H ◦ = 1 (by
passing to G/H ◦ ).
For i = 1, 2, as [Li : H] < ∞ and as H, Li are all orbital, we may take an open
normal subgroup N of G such that [N, Li ] ⊆ H. Indeed, NG (Li ) is a subgroup
of finite index in G, and permutes the (finitely many) left cosets of H in Li by
left multiplication; take Ni to be the kernel of this action, and set N = N1 ∩ N2 .
Hence [N ∩ H, Li ] ⊆ N ∩ H, i.e. both L1 and L2 normalise N ∩ H, so G
normalises N ∩ H. So N ∩ H is a normal subgroup of G contained in H, and
by assumption must be trivial. But N was an open subgroup of G, so H must
have been finite, and so L1 and L2 must be finite orbital subgroups of G. This
implies that Li ≤ ∆+ , and hence G = ∆+ , so that G is finite, as required.
Now, in the general case, G satisfies the maximal condition on closed subgroups,
so we can choose L maximal subject to L being orbital and [L : H] < ∞. This
L is iG (H) and contains H as an open subgroup.
Lemma 1.8.
(i) Suppose H is an orbital closed subgroup of G. Then iG (H) is an isolated
orbital closed subgroup of G. Furthermore, if H is normal in G, then so
is iG (H).
(ii) Suppose G is orbitally sound and H is a closed subgroup of finite index.
Then H is orbitally sound.
9
Proof.
(i) If iG (H) is orbital, then by Proposition 1.7, it is isolated (by construction). But H has finite index in iG (H), so iG (H) must be generated by
a finite number of subgroups L1 , . . . , Ln containing H as a subgroup of
finite index. So
n
\
NG (Li ) ≤ NG (iG (H)),
i=1
and as each NG (Li ) is open in G, so is NG (iG (H)).
Now suppose that H is normal in G. To see that iG (H) is normal in G,
fix g ∈ G, and note that conjugation by g fixes H and therefore simply
permutes the set of orbital closed subgroups L of G containing H as an
open subgroup, i.e. permutes the set of subgroups of G that generate
iG (H) (see Definition 1.6).
(ii) Let K be an isolated H-orbital closed subgroup of H. Then K is G-orbital,
so iG (K) is an isolated orbital subgroup of G, and so is normal in G. Hence
iG (K)∩H is normal in H. But [iG (K) : K] < ∞, so [iG (K)∩H : K] < ∞,
and hence iG (K) ∩ H = K, as K was assumed to be isolated in H.
Lemma 1.9. Let H be an open normal subgroup of G. Then there is a oneto-one correspondence
o
n
o
n
isolated orbital
isolated orbital
oo
//
closed subgroups of G
closed subgroups of H
′ ✤
// iH (G′ ∩ H),
G
✤ ′
iG (H ′ ) oo
H.
Proof. Suppose first that H ′ is an arbitrary orbital closed subgroup of H. That
is, NH (H ′ ) is open in H, hence also in G, and so NG (H ′ ) must be open in G.
Therefore H ′ is also G-orbital, and so, by Lemma 1.8(i), iG (H ′ ) is an isolated
orbital closed subgroup of G.
Conversely, take a G-isolated orbital closed subgroup G′ : then NG (G′ ) is open
in G, as G′ is G-orbital, which implies that NH (G′ ∩ H) is open in H, i.e. that
G′ ∩ H is H-orbital. Now iH (G′ ∩ H) is H-isolated orbital by Lemma 1.8(i).
Now we will show that these correspondences are mutually inverse.
In one direction, we must take H ′ as above, assume further that it is H-isolated
orbital, and show that iH (iG (H ′ ) ∩ H) = H ′ . To do this, note that iG (H ′ )
contains H ′ as an open subgroup by Proposition 1.7, so iG (H ′ ) ∩ H is an Horbital closed subgroup (by the correspondence above) containing the H-isolated
orbital H ′ as an open subgroup, and so by definition the two must be equal.
For the converse direction, we must take G′ as above and show that iG (iH (G′ ∩H))
is equal to G′ . But clearly iG iH = iG , and both G′ and iG (G′ ∩H) are G-isolated
10
orbital (G′ by assumption, iG (G′ ∩ H) by definition) and contain G′ ∩ H as an
open subgroup, so by uniqueness (see Definition 1.6), they are equal.
Lemma 1.10. The following are equivalent:
(i) Any isolated orbital closed subgroup H of G is normal.
(ii) Any orbital closed subgroup K of G contains a subgroup N of finite index
in K which is normal in G.
Proof.
(i) ⇒ (ii) Let K be an orbital closed subgroup of G. By Lemma 1.8(i), iG (K)
is an isolated orbital closed subgroup of G, and so (by assumption) is normal
in G. Therefore, as it contains K as a subgroup of finite index (by Proposition
1.7), it must contain each conjugate K g (for any g ∈ G) as a subgroup of finite
index. But as K is G-orbital, it only has finitely many G-conjugates, and so
their intersection K ◦ still has finite index in iG (K) and hence also in K, and
K ◦ is normal in G by construction.
(ii) ⇒ (i) Let H be an isolated orbital closed subgroup of G, and write H ◦
for the largest normal subgroup of G contained in H, which by (ii) must have
finite index in H. Now clearly H ≤ iG (H ◦ ) by definition of iG (H ◦ ), but also
iG (H ◦ ) ≤ H as H is isolated and contains H ◦ . So H is the G-isolator of a
normal subgroup, and so by Lemma 1.8(i), H is also normal in G.
11
2
The Roseblade subgroup nio(G)
We begin this section by remarking that “orbitally sound” is not too restrictive
a condition. Recall:
Lemma 2.1. Let G be a p-valuable group. Then G is orbitally sound.
Proof. This is [2, Proposition 5.9], after remarking that the definitions of “orbitally sound” given in Definition 1.4 and in [2, 5.8] are equivalent by Lemma
1.10.
The following two lemmas will allow us to find a large class of orbitally sound
groups.
For the next lemma, fix the following notation. Let G be a compact
h i p-adic
analytic group, and consider its Qp -Iwasawa algebra Qp G := (Zp G) p1 . Write
I for its augmentation ideal
I = ker(Qp G → Qp ).
Recall that I k is generated over Qp G by {(x1 − 1) . . . (xk − 1) xi ∈ G}. Now it
is clear that G acts unipotently on the series
Qp G > I > I 2 > I 3 > . . . ,
i.e. for all g ∈ G, we have (g − 1)Qp G ⊆ I and (g − 1)I k ⊆ I k+1 .
Write also Un for the subgroup of GLn (Qp ) consisting of upper triangular unipotent matrices.
Lemma 2.2. Write
Dk = ker(G → (Qp G/I k )× ),
the k-th rational dimension subgroup of G, for all k ≥ 1. Then the Dk are a
descending chain of isolated orbital closed normal subgroups of G. This chain
eventually stabilises: that is, there exists some t such that Dn = Dt for all
n ≥ t.
Furthermore, if G is torsion-free and nilpotent, Dt = 1, and G is isomorphic to
a closed subgroup of Um for some m.
Proof. By definition, it is clear that the Dk are closed normal (hence orbital)
subgroups of G; to show that they are isolated orbital, we will show that each
G/Dk is torsion-free.
Fix k. Consider the series of finite-dimensional Qp -vector spaces
Qp G/I k > I/I k > I 2 /I k > I 3 /I k > · · · > I k /I k ,
12
and choose a basis for Qp G/I k which is filtered relative to this series: i.e. by
repeatedly extending a basis for I r /I k to a basis for I r−1 /I k , we get a basis
B = {e1 , . . . , er }
and integers
0 = nk < nk−1 < nk−2 < · · · < n1 < n0 = r
with the property that {e1 , . . . , enr } is a basis for I r /I k for each 0 ≤ r ≤ k
(where we write I 0 = Qp G for convenience).
As G acts unipotently (by left multiplication) on Qp G/I k , and by definition of
the basis B, we see that with respect to B, each g ∈ G acts by a unipotent uppertriangular matrix, i.e. we get a continuous group homomorphism G → Ur . Now
Dk is just the kernel of this map; but Ur is torsion-free, so Dk must be isolated.
Recall the dimension dim H of a pro-p group H of finite rank from [4, 4.7]. As
G has finite rank, it also has finite dimension [4, 3.11, 3.12], and we must have
dim Di ≥ dim Di+1 for all i by [4, 4.8]. But if dim Di = dim Di+1 , then Di /Di+1
is a p-valued group (as Di+1 is isolated) of dimension 0 (again by [4, 4.8]), and so
must be trivial. Hence the sequence (Di ) stabilises after at most t := 1 + dim G
terms, and so
\
Dt =
Dn .
n≥1
Now suppose that G is nilpotent. Then, by [1, Theorem A], it follows that I
is localisable. Let R = (Qp G)I be its localisation, and J(R) = m its unique
maximal ideal: then the ideal
\
A=
mn
n≥1
satisfies A = mA, so by Nakayama’s lemma [7, 0.3.10], we must have A = 0.
This implies that
\
I n ⊆ ker(Qp G → R).
n≥1
Assuming further that G is torsion-free, we see that Qp G is a domain [9, Theorem 1], and so the localisation map Qp G → R is injective. Hence ∩n≥1 I n = 0,
and so
\
\
\
Dt =
Dn =
(I n + 1) ∩ G ⊆
I n + 1 = 1.
n≥1
n≥1
n≥1
Now the representation G → Aut(Qp G/I t ) ∼
= GLm (Qp ) is faithful and has
image in Um .
Lemma 2.3. Let G be a (topologically) finitely generated nilpotent pro-p
group. Then G is p-valuable if and only if it is torsion-free.
13
Proof. Lemma 2.2 gives an injective map G → Um . Now, as G is topologically
finitely generated, its image in Um must lie inside the set p1t Mm (Zp ) ∩ Um for
some t. Hence, by conjugating by the diagonal element
diag(p, p2 , . . . , pm )t+ε ∈ GLm (Qp ),
where
ε=
(
1
2
p > 2,
p = 2,
we see that G is isomorphic to a subgroup of
Γε = γ ∈ GLm (Zp ) γ ≡ 1(mod pε ) ,
the εth congruence subgroup of GLm (Zp ), which is uniform (and hence pvaluable) by [4, Theorem 5.2].
The reverse implication is clear from the definition of a p-valuation [5, III,
2.1.2].
Now we have found a large class of orbitally sound compact p-adic analytic
groups.
Corollary 2.4. If G is a finite-by-nilpotent compact p-adic analytic group, then
it is orbitally sound.
Proof. G := G/∆+ must be a nilpotent compact p-adic analytic group with
∆+ (G) = 1, and so G is torsion-free by [12, 5.2.7]. Now Lemma 2.3 shows that
G is p-valuable, and from Lemma 2.1 we may deduce that G is orbitally sound.
But now Lemma 1.5(iii) implies that G is orbitally sound.
Remark. It is well known that finite-by-nilpotent implies nilpotent-by-finite, but
not conversely. Not all nilpotent-by-finite compact p-adic analytic groups are
orbitally sound: indeed, the wreath product
G = Zp ≀ C2 = (Zp × Zp ) ⋊ C2
is abelian-by-finite, and the infinite procyclic subgroup H = Zp × {0} is orbital,
but the largest G-normal subgroup contained in H is the trivial subgroup.
We can now define the Roseblade subgroup.
Definition 2.5. As in Roseblade [13, 1.3], write nio(G) for the closed subgroup
of G defined by
\
nio(G) =
NG (H),
H
where the intersection ranges over the isolated orbital closed subgroups H of G.
Theorem 2.6. Let G be a compact p-adic analytic group.
14
(i) An orbitally sound open normal subgroup N ✁ G normalises every closed
isolated orbital subgroup H ≤ G.
(ii) nio(G) is an orbitally sound open characteristic subgroup of G.
Proof.
(i) Since H is isolated orbital in G, we must have that H ∩ N is isolated
orbital (and hence normal) in N . However, it follows from Proposition
1.7 that H = iG (H ∩ N ). Hence H is generated by all the (finitely many)
closed orbital subgroups L1 , . . . , Lk of G containing H ∩ N as a subgroup
of finite index. Conjugation by n ∈ N permutes these Li , and so fixes H.
(ii) Let N be a complete p-valued open normal subgroup of G (e.g. by [4,
8.34]). Then [2, Proposition 5.9] shows that N is orbitally sound, and
hence by (i) N normalises all closed isolated orbital subgroups of G. So,
by definition, N ≤ nio(G), and so [G : nio(G)] ≤ [G : N ] < ∞. Therefore
nio(G) is open in G as required. But by definition, nio(G) is the largest
subgroup that normalises all isolated orbital subgroups of G, so by the
correspondence of Lemma 1.9 and Lemma 1.8(i), it normalises all isolated
orbital subgroups of nio(G), i.e. it is orbitally sound.
Proof of Theorem A. It is clear from Definition 2.5 and Theorem 2.6(i), (ii) that
nio(G) is the unique maximal orbitally sound closed normal subgroup of G, and
is hence characteristic in G. Corollary 2.4, together with Fitting’s theorem [14,
1B, Proposition 15], implies that it contains all finite-by-nilpotent closed normal
subgroups.
15
3
p-saturations
Recall [5, III, 2.1.2] that a p-valuation of a group G is a function
ω : G → R ∪ {∞}
satisfying the following properties:
• ω(x) = ∞ if and only if x = 1,
• ω(x) > (p − 1)−1 ,
• ω(x−1 y) ≥ inf{ω(x), ω(y)},
• ω([x, y]) ≥ ω(x) + ω(y),
• ω(xp ) = ω(x) + 1,
for all x, y ∈ G. The group G, when endowed with the p-valuation ω, is called
p-valued. On the other hand, a group G is called p-valuable [5, III, 3.1.6] if there
exists a p-valuation ω of G with respect to which G is complete of finite rank.
Let G be a p-valuable group, and fix a p-valuation ω on G, so that G is complete
p-valued of finite rank. Recall the definition of the p-saturation Sat G of G (with
respect to ω) from [5, IV, 3.3.1.1]: this is again a complete p-valued group of
finite rank, and there is a natural isometry identifying G with an open subgroup
of Sat G [5, IV, 3.3.2.1]. We will prove a few basic facts about p-saturations.
Firstly, we will prove a basic relationship between isolators and p-saturations.
Lemma 3.1. Let G be a complete p-valued group of finite rank, and let H be
a closed normal (and hence orbital) subgroup of G. Then iG (H) = Sat H ∩ G
(considered as subgroups of Sat G).
Proof. [Sat H : H] < ∞ by [5, IV, 3.4.1], and Sat H is a closed normal subgroup
of Sat G by [5, IV, 3.3.3], so S := Sat H ∩ G is a normal (and hence orbital) subgroup of G, and contains H as a subgroup of finite index. Hence, by Definition
1.6, S is contained in iG (H).
To show the reverse inclusion, we will consider the group iG (H)/S, which is
a finite subgroup of G/S (as it is a quotient of iG (H)/H, which is finite by
Proposition 1.7). But G/S is isomorphic to GSat H/Sat H, a subgroup of the
torsion-free group Sat G/Sat H (see [5, IV, 3.4.2] or [5, III, 3.3.2.4]). In particular, G/S has no non-trivial finite subgroups, so we must have iG (H) = S.
Remark. Of course, iG (H) is independent of the choice of ω.
Lemma 3.2. Let G be a complete p-valued group of finite rank, which we
again identify with an open subgroup of its p-saturation S. Suppose S ′ is a
p-saturated closed normal subgroup of S, and set G′ = S ′ ∩ G. Then there is a
natural isometry S/S ′ ∼
= Sat (G/G′ ).
16
Proof. We will show that S/S ′ satisfies the universal property for Sat (G/G′ ) [5,
IV, 3.3.2.4]. Clearly we may regard G/G′ ∼
= GS ′ /S ′ as a subgroup of S/S ′ . Note
′
also that S/S is p-saturated, by [5, III, 3.3.2.4]. Also, as G′ is open in S ′ and
S ′ is p-saturated, we have that S ′ = Sat G′ .
Let H be an arbitrary p-saturated group and ϕ : G/G′ → H a homomorphism
of p-valued groups. We must first construct a map ψ : S/S ′ → H. To do this, we
first compose ϕ with the natural surjection G → G/G′ to get a map α : G → H,
which we may then extend uniquely to a map β : S → H using the universal
property of S = Sat G, so that α = β|G and the following diagram commutes.
α
G _
S
// // G/G′
ϕ
//&& H
;;
β
Now we wish to show that β descends to a map S/S ′ → H. To do this, we
first study the restriction of α to G′ and of β to S ′ . The following diagram
commutes:
α|G′
// H
G′ _
>>
⑥⑥
⑥
⑥
⑥⑥
⑥⑥ β|S′
S′
and so, since S ′ = Sat G′ , β|S ′ must be the unique extension of α|G′ to a map
S ′ → H, as S ′ = Sat G′ . But α factors through G/G′ , i.e. α|G′ is the trivial
homomorphism G′ → H, so it extends to the trivial homomorphism S ′ → H.
By uniqueness, we must have S ′ ⊆ ker β. This shows that β induces a map
ψ : S/S ′ → H.
Finally, suppose ϕ : G/G′ → H has two distinct extensions ψ1 , ψ2 : S/S ′ → H.
Then we may compose them with the natural surjection S → S/S ′ to get two
distinct maps β1 , β2 : S → H. Their restrictions α1 , α2 : G → H to G must
therefore also be distinct, for if not, then the map α1 = α2 : G → H has
(at least) two distinct extensions to maps S → H, contradicting the universal
property of S = Sat G. Finally, if α1 and α2 are distinct, then they descend
to distinct maps ϕ1 , ϕ2 : G/G′ → H, contradicting our assumption. So the
extension of ϕ to ψ is unique.
Remark. Lemma 3.2 holds even if G does not have finite rank, and hence is only
closed (not necessarily open) in its p-saturation S.
Definition 3.3. Let G be an arbitrary group. A central series for G is a
sequence of subgroups
G = G1 ✄ G2 ✄ · · · ✄ Gn = 1
17
with the property that [G, Gi ] ≤ Gi+1 for each i. (For the purposes of this
definition, Gj is understood to mean 1 if j > n, and G if j < 1.)
We will say that a central series is strongly central if also [Gi , Gj ] ≤ Gi+j for
all i and j.
An abelian series for G a sequence of subgroups
G = G1 ✄ G2 ✄ · · · ✄ Gn = 1
with the property that [Gi , Gi ] ≤ Gi+1 for each i.
When G is a topological group, we will insist further that all of the Gi should
be closed subgroups of G.
Remark. We will be working with nilpotent p-valuable groups G. It will be
useful for us to define the isolated lower central series of G, which will turn out
to be the fastest descending central series of closed subgroups
G = G1 ✄ G2 ✄ · · · ✄ Gr = 1
with the property that the successive quotients Gi /Gi+1 are torsion-free (and
hence p-valuable, by [5, IV, 3.4.2]). We will also prove that the isolated lower
central series is a strongly central series. (We demonstrate an isolated derived
series for soluble p-valued groups at the same time.)
Lemma 3.4. Let G be a complete p-valued group of finite rank, and G1 ≥ G2
closed normal subgroups of G with G1 /G2 an abelian pro-p group (which is
not necessarily p-valued). Let Si = Sat Gi for i = 1, 2. Then S1 /S2 is abelian
and torsion-free (and hence p-valued), and has the same rank as G1 /G2 as a
Zp -module.
Proof. As S2 is p-saturated, S1 /S2 is torsion-free, and so
G1 /(S2 ∩ G1 ) ∼
= G1 S2 /S2 ≤ S1 /S2
is torsion-free. G1 /G2 maps onto G1 /(S2 ∩G1 ) with finite kernel (by Lemma 3.1
and Proposition 1.7, and the assumption that G has finite rank), so G1 /(S2 ∩G1 )
is abelian of the same Zp -rank as G1 /G2 . By Lemma 3.2, S1 /S2 is the psaturation of G1 /(S2 ∩ G1 ), so is still abelian of the same Zp -rank.
Before proving the main result of this section, we first need a technical lemma.
Lemma 3.5. Let G be a complete p-valued group of finite rank, and let H and
N be two closed normal subgroups. Then
[iG (H), iG (N )] ≤ iG ([H, N ]).
Proof. Write L := iG ([H, N ]). This is normal in G, as G is orbitally sound [2,
5.9], and the quotient G/L is still p-valued as it is torsion-free [5, IV, 3.4.2].
18
Suppose first that L = 1, so that [H, N ] = 1. Then, for any h ∈ iG (H), there is
n
n
some integer n such that hp ∈ H, so that [g, hp ] = 1 for all g ∈ N . But this
pn
g pn
is the same as saying that h = (h ) ; and, as G is p-valued, [5, III, 2.1.4]
implies that h = hg . As g and h were arbitrary, we see that [iG (H), N ] = 1.
Repeat this argument for N to show that [iG (H), iG (N )] = 1.
If L 6= 1, we may pass to G/L. Write π : G → G/L for the natural surjection,
so that
π [iG (H), iG (N )] = [π(iG (H)), π(iG (N ))].
(1)
Now, π(H) is a closed orbital subgroup of π(G), and π(iG (H)) is a closed orbital
subgroup of π(G) containing π(H) as an open subgroup, so that
iπ(G) (π(H)) ≥ π(iG (H)),
and similarly for N . Together with (1), this implies that
π [iG (H), iG (N )] ≤ [iπ(G) (π(H)), iπ(G) (π(N ))].
But the right-hand side is now equal to π(1), by the previous case, which shows
that [iG (H), iG (N )] ≤ L as required.
Corollary 3.6. Let G be a p-valuable
( group. Define two series:
γ1 = G,
Gi = iG (γi ), where
γi+1 = [γi , G] for i ≥ 1;
(i)
G
= iG (Di ), where
(
D0 = G,
Di+1 = [Di , Di ] for i ≥ 0,
where the bars denote topological closure inside G. If G is nilpotent, then (Gi )
is a strongly central series for G, i.e. a central series in which [Gi , Gj ] ≤ Gi+j .
If G is soluble, then (G(i) ) is an abelian series for G. The quotients Gi /Gi+1
and G(i) /G(i+1) are torsion-free, and hence p-valuable.
Remark. We prove this using p-saturations, but the resulting closed subgroups
Gi and G(i) are independent of the choice of p-valuation ω on G.
The series (G(i) ) above is a generalisation of the series studied in [8, proof of
lemma 2.2.1], there called (Gi ).
Proof. Fix a p-valuation ω on G throughout.
Firstly, we will show that (Gi ) is an abelian series. The claim that (G(i) ) is an
abelian series will follow by an identical argument.
The (abstract) lower central series (γi ) is an abelian series for G as an abstract
group (i.e. the subgroups γi are not necessarily closed in G), and so the series
(γi ) is a series of closed normal subgroups of G, which is still an abelian series
by continuity. Now, applying Lemma 3.4 shows that (Sat γi ) is also an abelian
19
series; and by Lemma 3.1, we see that Gi = Sat γi ∩ G for each i, so that (Gi )
is an abelian series.
Secondly, we address the claim that the quotients Gi /Gi+1 are torsion-free and
hence p-valuable: this follows from [5, III, 3.1.7.6 / IV, 3.4.2], as the Gi+1 are
isolated in G. The case of the quotients G(i) /G(i+1) is again identical.
Thirdly, we must show that Gi−1 /Gi is central in G/Gi . Certainly γi−1 Gi /Gi
is central in G/Gi , because γi ≤ Gi , and so
γi−1 Gi /Gi ≤ Z(G/Gi )
by continuity. However, [2, lemma 8.4(a)] says that Z(G/Gi ) is isolated in G/Gi ,
so by taking (G/Gi )-isolators of both sides, we must have
iG/Gi γi−1 Gi /Gi ≤ Z(G/Gi );
and the left-hand side is clearly equal to Gi−1 /Gi by Lemma 1.5(i) and Definition 1.6.
Finally, note that
[γi , γj ] ≤ γi+j
by [12, 5.1.11(i)], and so by taking closures,
[γi , γj ] ≤ γi+j .
But [γi , γj ] ≤ [γi , γj ], as the function G × G → G given by (a, b) 7→ [a, b] is
continuous. Hence
[γi , γj ] ≤ γi+j ,
which implies
[γi , γj ] ≤ γi+j ,
and so, by Lemma 3.5, we may take isolators to show that
[iG (γi ), iG (γj )] ≤ iG [γi , γj ] ≤ iG (γi+j ),
i.e. [Gi , Gj ] ≤ Gi+j .
Definition 3.7. When G is a nilpotent (resp. soluble) p-valued group of finite
rank, the series (Gi ) (resp. (G(i) )) defined in Corollary 3.6 is the isolated lower
central series (resp. isolated derived series) of G.
Proof of Theorem B. This is the content of Corollary 3.6.
20
4
Conjugation action of G
In this subsection, we will study how nilpotent-by-finite compact p-adic analytic groups G act by conjugation on certain torsion-free abelian and nilpotent
subquotients. First, we slightly extend the term “orbitally sound”.
Definition 4.1. Let G and H be profinite groups, and suppose G acts (continuously) on H. Then G permutes the closed subgroups of H. We say that the
action of G on H is orbitally sound if, for any closed subgroup K of H with
finite G-orbit, there exists an open subgroup K ′ of K which is normalised by
G.
Recall the group of torsion units of Zp :
(
{±1} p = 2
×
t(Zp ) =
F×
p > 2.
p
Lemma 4.2. Let A be a free abelian pro-p group of finite rank. Let G be a
profinite group acting orbitally soundly and by automorphisms of finite order
on M . Then, for each g ∈ G, there exists
ζ = ζg ∈ t(Z×
p)
such that g · x = ζx for all x ∈ A. This is multiplicative in G, in the sense that
ζg ζh = ζgh for all g, h ∈ G.
Proof. Write ϕ for the automorphism of A given by conjugation by g. We may
view ϕ as an automorphism of the Qp -vector space AQp := A ⊗ Qp .
Zp
As the action of G on A is orbitally sound, in particular, we have
hxi ∩ hϕ(x)i =
6 {0}
(as Qp -vector subspaces) for every x ∈ AQp . But this just means that x is an
eigenvector of the linear map ϕ. If all elements of AQp are eigenvectors of ϕ,
then they must have a common eigenvalue, say ζ. The statement that G acts
on A by automorphisms of finite order means that the eigenvalue ζ for x is of
finite order, ζ ∈ t(Z×
p ).
Multiplicativity is clear from the fact that (gh)·x = g ·(h·x) for all g, h ∈ G.
Remark. Assume that G is a nilpotent-by-finite, orbitally sound compact p-adic
analytic group. In the case when H is an open subgroup of G containing ∆+ ,
with the property that N := H/∆+ is nilpotent p-valuable, we may consider
the isolated lower central series of Corollary 3.6 for N :
N = N1 ✄ N2 ✄ · · · ✄ Nr = 1,
21
and take their preimages in G to get a series of characteristic subgroups of H:
H = H1 ✄ H2 ✄ · · · ✄ Hr = ∆+ ,
with the property that each Ai := Hi /Hi+1 is a free abelian pro-p group of finite
rank.
G clearly acts orbitally soundly on each Ai , as G is itself orbitally sound. Furthermore, as [H, Hi ] ≤ Hi+1 for each i, we see that the action G → Aut(Ai )
contains the open subgroup H in its kernel, and so G acts by automorphisms of
finite order. Thus we may apply Lemma 4.2 to see that G acts on each Ai via
a homomorphism ξi : G → t(Z×
p ).
That is, given any g ∈ G and h ∈ Hi , and writing ζ = ξi (g) and a = nNi+1 ∈ Ai ,
we have
(ng )n−ζ ∈ Ni+1 ,
or equivalently (still in multiplicative notation)
ag = aζ .
We now show that the action of an automorphism of G on the quotients Ai is
strongly controlled by its action on A1 . This is an important property that the
isolated lower central series shares with the usual lower central series of abstract
nilpotent groups; cf. [12, 5.2.5] and the surrounding discussion.
Lemma 4.3. Let H be a finite-by-(nilpotent p-valuable) group, and continue
to write Ai := Hi /Hi+1 as in the remark above. Let α be an automorphism of
i
H inducing multiplication by ζi ∈ t(Z×
p ) on each Ai . Then ζi = ζ1 for each i.
Proof. The map
A1 ⊗ Ai → Ai+1
Zp
xH2 ⊗ yHi+1 7→ [x, y]Hi+2
is a Zp hαi-module homomorphism, and its image is open in Ai+1 (by definition
of the isolated lower central series). Write ζ1 = ζ, and proceed by induction on
i: suppose that ζi = ζ i . Now, for any positive integers a and b, we have
[xa , y b ]Hi+2 = [x, y]ab Hi+2
by [4, 0.2(i), (ii)] and by using the fact that [x, y]Hi+2 is central in H/Hi+2 .
Hence, by continuity, this is true for any a, b ∈ Zp , and so
i
α([x, y]Hi+2 ) = [xζ , y ζ ]Hi+2
= [x, y]ζ
We deduce:
22
i+1
Hi+2 .
Corollary 4.4. Let G be a nilpotent-by-finite, orbitally sound compact p-adic
analytic group, and H an open normal subgroup of G containing ∆+ such that
H/∆+ is nilpotent p-valuable. Then the conjugation action of G on H induces
an action of G on H/H2 given by the map ξ1 : G → t(Z×
p ) ≤ Aut(H/H2 ) defined
above. Moreover, H ≤ ker ξ1 .
Remark. If N = H/∆+ is p-saturable, we may take its corresponding Lie algebra
L by Lazard’s isomorphism of categories [5, IV, 3.2.6]. As in [2, proof of lemma
8.5]: using [2, lemma 4.2] and the fact that the N/Ni are torsion-free, we can
pick an ordered basis [5, III, 2.2.4] B for N which is filtered relative to the
filtration on N : that is,
B = {n1 , n2 , . . . , ne },
and there exists a filtration of sets
B = B1 ⊃ B2 ⊃ · · · ⊃ Br−1 6= ∅
such that Bi is an ordered basis for Ni for each 1 ≤ i ≤ r − 1. We may order
the elements so that, for some integers 1 = k1 < k2 < · · · < kr−1 < e, we have
Bi = {nki +1 , . . . , ne } for each 1 ≤ i ≤ r − 1. Taking logarithms of these basis
elements gives us a basis for L, and then Lemma 4.3 implies that, with respect
to this basis, the automorphism of L induced by α has the special block lower
triangular form
0
0
...
0
ζId1
∗
0
...
0
ζ 2 Id2
3
∗
0
∗
ζ Id3 . . .
,
..
..
.
..
..
..
.
.
.
.
∗
∗
∗
...
ζ r−1 Idr−1
where di = rk(Li /Li+1 ) = rk(Mi ) and I denotes the identity matrix.
23
5
The finite-by-(nilpotent p-valuable) radical
Let G be a nilpotent-by-finite compact p-adic analytic group. Consider the set
S(G) = H ✁ G H/∆+ (H) is nilpotent and p-valuable ,
O
where “H ✁ G” means “H is an open normal subgroup of G”. S(G) is nonempty,
O
as we can pick an open normal nilpotent uniform subgroup of G by [4, 4.1], and
hence contains a maximal element. We will show that this maximal element is
unique, and we will call this element the finite-by-(nilpotent p-valuable) radical
of G, and once we have shown its uniqueness we will denote it by FNp (G).
Remark. Once we have shown the existence and uniqueness of FNp (G), it will
be clear that it is a characteristic open subgroup of G (as automorphisms of G
leave S(G) invariant), and contained in nio(G) (by Corollary 2.4 and Theorem
A).
The quotient group
nio(G)/FNp (G)
×
is isomorphic to a subgroup of t(Z×
p ) by Corollary 4.4. When p > 2, t(Zp ) is a
′
+
p -group, and so FNp (G)/∆ is the unique Sylow pro-p subgroup of nio(G)/∆+ .
(This fails for p = 2: the “2-adic dihedral group” G = Z2 ⋊ C2 has ∆+ (G) = 1,
nio(G) = G, and is its own Sylow 2-subgroup, but FNp (G) = Z2 .)
In looking for maximal elements H of S(G), we may make an immediate simplification. By maximality, any such H must have ∆+ (H) = ∆+ , i.e. maximal
elements of S(G) are in one-to-one correspondence with maximal elements of
′
+
+
S (G) = H ✁ G ∆ ≤ H, H/∆ is nilpotent and p-valuable ,
O
and this set is clearly in order-preserving one-to-one correspondence with the
set S(G/∆+ ). Hence we may immediately assume without loss of generality
that ∆+ = 1.
Lemma 5.1. Let G be a nilpotent-by-finite compact p-adic analytic group with
∆+ = 1. Then
(i) there exists a nilpotent uniform open normal subgroup H of G which
contains ∆,
(ii) any such H satisfies the property that Z(H) = ∆.
Proof. First, suppose we are given a nilpotent uniform open normal subgroup
H. Take x ∈ ∆: CG (x) is open in G by definition, and so
CH (x) = CG (x) ∩ H
24
is open in H. Therefore, for any h ∈ H, we can find some integer k such that
k
k
k
hp ∈ CH (x). This means that (x−1 hx)p = hp , and so by [5, III, 2.1.4],
we may take (pk )th roots inside H to see that x−1 hx = h. In other words,
x ∈ C∆ (H).
Now suppose further that H contains ∆. Then x ∈ Z(H). In fact, as we have
Z(H) ≤ ∆(H) by definition and ∆(H) ⊆ ∆ by Lemma 1.3(ii), we see that ∆ is
all of the centre of H. This establishes (ii).
To prove (i), let N be an open normal nilpotent uniform subgroup [4, 4.1] of
G. Form H = N ∆, again an open normal subgroup of G. The first paragraph
above shows that [N, ∆] = 1; we also know from Lemma 1.3(v) that ∆ is abelian
and N is nilpotent. This forces H to be nilpotent and open in G, and to contain
∆ in its centre.
It remains only to show that H is uniform. As H is nilpotent, its set t(H) of
torsion elements forms a normal subgroup [12, 5.2.7], and if t(H) is non-trivial
then t(H) ∩ Z(H) must be non-trivial [12, 5.2.1]; but Z(H) = ∆ is torsion-free
by Lemma 1.3(v), so H must be torsion-free. Now it is easy to check that H is
powerful as in [4, 3.1], so that H is uniform by [4, 4.5].
Lemma 5.2. Let G be a nilpotent-by-finite compact p-adic analytic group.
Then S(G) is closed under finite joins, and hence contains a unique maximal
element H, which is characteristic as a subgroup of G.
Proof. First, observe that, for an open normal subgroup K of G, we have
K ∈ S(G) if and only if K ∈ S(G) (where bars denote quotient by ∆+ ). So we
continue to assume without loss of generality that ∆+ = 1.
Suppose we are given K, L ∈ S(G): then it remains to show that KL ∈ S(G).
As K and L are open and normal, it is obvious that KL is too; and since K
and L are also nilpotent, Fitting’s theorem [14, 1B, Proposition 15] implies that
KL is nilpotent. But now, again by [12, 5.2.7], t(KL) = ∆+ (KL) ≤ ∆+ = 1 –
that is, KL is torsion-free, and hence p-valuable by Lemma 2.3.
Now let H be a maximal element of S(G). Assume for contradiction that H
does not contain every other element of S(G) as a subgroup. Then we may pick
some L ∈ S(G) not contained in H, and form HL ∈ S(G); but now H HL,
a contradiction to the maximality of H. So H must be the unique maximal
element of S(G).
As the set S(G) is invariant under automorphisms of G, this maximal element
H is characteristic in G.
Definition 5.3. Let G be a nilpotent-by-finite compact p-adic analytic group.
Its finite-by-(nilpotent p-valuable) radical FNp (G) is the open characteristic
subgroup defined in Lemma 5.2.
Proof of Theorem C. The uniqueness of H = FNp (G) is the content of Lemma
5.2. As H is a finite-by-nilpotent closed subgroup of G, Corollary 2.4 and
25
Theorem B show that H is contained in nio(G). The inclusion ∆ ≤ H follows
from Lemma 5.1(i). Finally, the isomorphism from nio(G)/H to a subgroup of
t(Z×
p ) is induced by the map
ξ1 : nio(G) → t(Z×
p)
of Corollary 4.4, whose kernel contains H.
Proof of Theorem D. This now follows from Theorem C together with Lemma
4.3.
References
[1] K. Ardakov. Localisation at augmentation ideals in Iwasawa algebras. Glasgow Mathematical Journal, 48(2):251–267, 2006.
[2] K. Ardakov. Prime ideals in nilpotent Iwasawa algebras. Inventiones mathematicae, 190(2):439–503, 2012.
[3] K. Ardakov, K.A. Brown. Primeness, semiprimeness and localisation in
Iwasawa algebras. Trans. Amer. Math. Soc., 359:1499–1515, 2007.
[4] J.D. Dixon, M.P.F. du Sautoy, A. Mann and D. Segal. Analytic Pro-p
Groups. Cambridge University Press, 1999.
[5] Michel Lazard. Groupes analytiques p-adiques. Publications Mathématiques
de l’IHÉS, 26:5–219, 1965.
[6] M. Lorenz, E. Letzter. Polycyclic-by-finite group algebras are catenary.
Math. Res. Lett., 6:183–194, 1999.
[7] J.C. McConnell and J.C. Robson. Noncommutative Noetherian Rings.
American Mathematical Society, 2001.
[8] J. Nelson. Localisation and Euler Characteristics of Soluble Iwasawa Algebras. PhD thesis, University of Cambridge, 2013.
[9] A. Neumann. Completed group algebras without zero divisors. Arch. Math.,
51:496–499, 1988.
[10] Donald S. Passman. Algebraic Structure of Group Rings. John Wiley &
Sons, 1977.
[11] Donald S. Passman. Infinite Crossed Products. Academic Press Inc., 1989.
[12] Derek J.S. Robinson. A course in the theory of groups. Springer-Verlag
New York, 1996.
[13] J.E. Roseblade. Prime ideals in group rings of polycyclic groups. Proc.
London Math. Soc., 36:385–447, 1978.
[14] D. Segal. Polycyclic groups. Cambridge University Press, 2005.
26
| 4math.GR
|
Fisher Kernel for Deep Neural Activations
Donggeun Yoo, Sunggyun Park, Joon-Young Lee, and In So Kweon
arXiv:1412.1628v2 [cs.CV] 19 Dec 2014
KAIST
Daejeon, 305-701, Korea.
[email protected], [email protected], [email protected], [email protected]
Abstract
In recent computer vision researches, drastic advances
of visual recognition are achieved by deep convolutional
neural networks (CNNs) [5], which jointly learn the whole
feature hierarchies starting from image pixels to the final
class posterior with stacked non-linear processing layers.
A deep representation is quite efficient since its intermediate templates are reused. However, the deep CNN is nonlinear and have millions of parameters to be estimated. It
requires strong computing power for the optimization and
large training data to be generalized well. The recent presence of large scale ImageNet [6] database and the raise of
parallel computing contribute to the breakthrough in visual
recognition. Krizhevsky et al. [17] achieved an impressive
result using a CNN in large-scale image classification.
Compared to image representation based on low-level
local descriptors, deep neural activations of Convolutional
Neural Networks (CNNs) are richer in mid-level representation, but poorer in geometric invariance properties. In this
paper, we present a straightforward framework for better
image representation by combining the two approaches. To
take advantages of both representations, we propose an efficient method to extract a fair amount of multi-scale dense
local activations from a pre-trained CNN. We then aggregate the activations by Fisher kernel framework, which has
been modified with a simple scale-wise normalization essential to make it suitable for CNN activations. Replacing
the direct use of a single activation vector with our representation demonstrates significant performance improvements: +17.76 (Acc.) on MIT Indoor 67 and +7.18 (mAP)
on PASCAL VOC 2007. The results suggest that our proposal can be used as a primary image representation for
better performances in visual recognition tasks.
Instead of training a CNN for a specific task, intermediate activations extracted from a CNN pre-trained on independent large data have been successfully applied as a
generic image representation. Combining the CNN activations with a classifier has shown impressive performance
in wide visual recognition tasks such as object classification [26, 8, 22, 13, 4], object detection [11, 13], scene classification [26, 12, 36], fine-grained classification [26, 34],
attribute recognition [35], image retrieval [2], and domain
transfer [8].
1. Introduction
Image representation is one of the most important factors that affect performance on visual recognition tasks.
Barbu et al. [3] introduced an interesting experiment that
a simple classifier along with human brain-scan data substantially outperforms the state-of-the-art methods in recognizing action from video clips.
With a success of local descriptors [19], many researches
devoted deep study to global image representation based on
a Bag-of-Word (BOW) model [29] that aggregates abundant local statistics captured by hand-designed local descriptors. The BOW representation is further improved with
VLAD [14] and Fisher kernel [24, 23] by adding higher order statistics. One major benefit of these global representations based on local descriptors is their invariance property
to scale changes, location changes, occlusions and background clutters.
For utilizing CNN activations as a generic image representation, a straightforward way is to extract the responses
from the first or second fully connected layer of a pretrained CNN by feeding an image and to represent the image with the responses [8, 2, 13, 11]. However, this representation is vulnerable to geometric variations. There are
techniques to address the problem. A common practice
is exploiting multiple jitterred images (random crops and
flips) for data augmentation. Though the data augmentation
has been used to prevent over-fitting [17], recent researches
show that average pooling, augmenting data and averaging
the multiple activation vectors in a test stage, also helps to
achieve better geometric invariance of CNNs while improving classification performance by +2.92% in [4] and +3.3%
1
Linear SVM
-normalization
Norm. …
Power normalization
Norm.
FV 3
Average pooling
Norm.
FV 2
…
FV 1
Dictionary (GMM)
…
Replaced with
convolution
…
Scale pyramid
Multi-scale
dense activations
Conv. 7
Conv. 6
Conv. 5
Conv. 4
Conv. 3
Conv. 2
Conv. 1
Target
task …
Activation
pyramid
Dimensionality Reduction (PCA)
Transfer
parameters
Prob.
FC 8
FC 7
FC 6
Conv. 5
Conv. 4
Conv. 3
Conv. 2
Conv. 1
Pre-trained
CNN
Multi-scale Pyramid Pooling (MPP) layer
Figure 1. A pipeline of the proposed method. Given a pre-trained CNN, we replace the first two fully connected layers with the two
equivalent convolutional layers to efficiently obtain large amount of multi-scale dense activations. The activations are followed by the
Multi-scale Pyramid Pooling (MPP) layer we suggest. The consequent image representation is combined with the linear SVM for the
target classification task.
in [26] on PASCAL VOC 2007.
A different experiment for enhancing the geometric invariance on CNN activations was also presented. Gong et
al. [12] proposed a method to exploit multi-scale CNN activations in order to achieve geometric invariance characteristic while improving recognition accuracy. They extracted
dense local patches at three different scales and fed each local patch into a pre-trained CNN. The CNN activations are
aggregated at finer scales via VLAD encoding which was
introduced in [14], and then the encoded activations are
concatenated as a single vector to obtain the final representation.
In this paper, we introduce a multi-scale pyramid pooling to improve the discriminative power of CNN activations
robust to geometric variations. A pipeline of the proposed
method is illustrated in Figure 1. Similar to [12], we also
utilize multi-scale CNN activations, but present a different
pooling method that shows better performance in our experiments. Specifically, we suggest an efficient way to obtain abundant amount of multi-scale local activations from
a CNN, and aggregate them using the state-of-the-art Fisher
kernel [24, 23] with a simple but important scale-wise normalization, so called multi-scale pyramid pooling. Our proposal demonstrates substantial improvements on both scene
and object classification tasks compared to the previous representations including a single activation, the average pooling [26, 4], and the VLAD of activations [12]. Also, we
demonstrate object confidence maps which is useful for object detection/localization though only category-level labels
without specific object bounding boxes are used in training.
According to our empirical observations, replacing a
VLAD kernel with a Fisher kernel does not present significant impact, however it shows meaningful performance
improvements when our pooling mechanism that takes an
average pooling after scale-wise normalization is applied.
It implies that the performance improvement of our representation does not come just from the superiority of Fisher
kernel but from the careful consideration of neural activation’s property dependent on scales.
2. Multi-scale Pyramid Pooling
In this section, we first review the Fisher kernel framework and then introduce a multi-scale pyramid pooling
which adds a Fisher kernel based pooling layer on top of
a pre-trained CNN.
2.1. Fisher Kernel Review
The Fisher kernel framework on a visual vocabulary is
proposed by Perronnin et al. in [23]. It extends the conventional Bag-of-Words model to a probabilistic generative
model. It models the distribution of low-level descriptors
using a Gaussian Mixture Model (GMM) and represents
an image by considering the gradient with respect to the
model parameters. Although the number of local descriptors varies across images, the consequent Fisher vector has
a fixed-length, therefore it is possible to use discriminative
classifiers such as a linear SVM.
Let x denote a d-dimensional local descriptor and Gλ =
{gk , k = 1...K} denote a pre-trained GMM with K Gaussians where λ = {ωk , µk , σk , k = 1...K}. For each visual
word gk , two gradient vectors, Gµk ∈ <d and Gσk ∈ <d ,
are computed by aggregating the gradients of the local descriptors extracted from an image with respect to the mean
and the standard deviation of the k th Gaussian. Then, the
final image representation, Fisher vector, is obtained by
concatenating all the gradient vectors. Accordingly, the
Fisher kernel framework represents an image with a 2Kddimensional Fisher vector G ∈ <2Kd .
Intuitively, a Fisher vector includes the information
2
Scale
pyramid
Conv. 1
(11ൈ11ൈ3ൈ96)
FC 6
(6ൈ6ൈ256ൈ4,096)
…
3
227
55
55
227
Scale 2
…
322
4,096
6
3
322
Reform
Multi-scale
dense activations
256
6
96
Scale 1
Activation
pyramid
4,096
1
1
1
1
256
8
96
78
78
4,096
4,096
3
8
1
1
3
Figure 2. Obtaining multi-scale local activations densely from a pre-trained CNN. In this figure, the target layer is the first fully connected
layer (FC6). Because FC6 can be equally implemented by a convolutional layer containing 4,096 filters of 6×6×256 size, we can obtain
an activation map where spatial ordering of local descriptors is conserved. A single pre-trained CNN is shared for all scales.
Image scales
1∼4
1∼5
1∼6
1∼7
Number of activations
270
754
1,910 4,410
Naive extraction (sec)
1.702 4.941 11.41 27.64
Proposed extraction (sec) 0.0769 0.1265 0.2420 0.4635
2.3. Multi-scale Pyramid Pooling (MPP)
For representing an image, we first generate a scale pyramid for the input image where the minimum scale image
has a fixed size of a CNN and each scale image has two
times larger resolution than the previous scale image. We
feed all the scaled images into a pre-trained CNN and extract dense CNN activation vectors. Then, all the activation
vectors are merged into a single vector by our multi-scale
pyramid pooling.
Table 1. Average time for extracting multi-scale dense activations
per image. With Caffe reference model [15], FC7 activations are
extracted from 100 random images of PASCAL VOC 2007. All
timings are based on a server with a CPU of 2.6GHz Intel Xeon
and a GPU of GTX TITAN Black.
If we consider each activation vector as a local descriptor, it is straightforward to aggregate all the local activations into a Fisher vector as explained in Sec. 2.1. However,
CNN activations have different scale properties compared
to SIFT-like local descriptors, as will be explained in Sec. 3.
To adopt the Fisher kernel suitable to CNN activation characteristics, we introduce adding a multi-scale pyramid pooling layer on top of the modified CNN as follows.
about directions of model parameters to best fit the local descriptors of an image to the GMM. The fisher kernel framework is further improved in [24] by the additional two-stage
normalizations: power-normalization with the factor of 0.5
followed by `2 -normalization. Refer to [24] for the theoretical proofs and details.
2.2. Dense CNN Activations
Given a scale pyramid S containing N scaled image and
local activation vectors xs extracted from each scale s ∈ S,
we first apply PCA to reduce the dimension of activation
vectors and obtain x0s . Then, we aggregate the local activation vectors x0s of each scale s to each Fisher vector G s .
After Fisher encoding, we have N Fisher vectors and they
are merged into one global vector by average pooling after
`2 -normalization as
To obtain multi-scale activations from a CNN without
modification, previous approach cropped local patches and
fed the patches into a network after resizing the patches to
the fixed size of CNN input. However, when we extract
multi-scale local activations densely, the approach is quite
inefficient since many redundant operations are performed
in convolutional layers for overlapped regions.
To extract dense CNN activations without redundant operations, we simply replace the fully connected layers of
an existing CNN with equivalent multiple convolution filters along spatial axises. When an image larger than the
fixed size is fed, the modified network outputs multiple activation vectors where each vector is CNN activations from
the corresponding local patch. The procedure is illustrated
in Fig. 2. With this method, thousands of dense local activations (4,410 per image) from multiple scale levels are
extracted in a reasonable extraction time (0.46 seconds per
image) as shown in Table 1.
GS =
1 X Gs
N
kG s k2
s∈S
s.t.
Gs =
1 X
∇λ log Gλ (x),
|x0s |
0
x∈xs
(1)
where |·| denotes the cardinality of a set. We use an average
pooling since it is a natural pooling scheme for Fisher kernel
rather than vector concatenation. Following the Improved
Fisher Kernel framework [24], we finally apply power normalization and `2 -normalization to the Fisher vector G S .
The overall pipeline of MPP is illustrated in Figure 1.
3
50
SIFT−Fisher
80
CNN−Fisher
MIT Indoor 67
PASCAL VOC 2007
Oxford 102 Flowers
94
Image Scale (Num. of Desc.)
Image scales
90
88
86
7
6
1~
1~
4
5
Naive Fisher
Proposed MPP
84
1~
7
6
1~
1~
4
5
1~
65
1~
7
6
1~
5
1~
Image scales
Image Scale (Num. of Desc.)
Naive Fisher
Proposed MPP
92
1~
Accuracy (%)
75
70
4
60
Image scales
Figure 4. Classification performance of our multi-scale pyramid
pooling in Eq. (1) and the naive Fisher pooling in Eq. (2). The tick
labels of the horizontal axis scale levels in a scale pyramid.
Figure 3. Classification performance of SIFT-Fisher and CNNFisher according to image scale on PASCAL VOC 2007. The tick
labels of the horizontal axis denote image scales and their average
number of local descriptors.
empirical analysis with scale-wise classification scores on
PASCAL VOC 2007 [9]. For the analysis, we first diversify
dataset into seven different scale levels from the smallest
scale of 227×227 resolution to the biggest scale of 1, 816×
1, 816 resolution and extract both dense SIFT descriptors
and local activation vectors in the seventh layer (FC7) of our
CNN. Then, we follow the standard framework to encode
Fisher vectors and to train an independent linear SVM for
each scale, respectively.
In Fig. 3, we show the results of classification performances using SIFT-Fisher and CNN-Fisher according to
scale. The figure demonstrates clear contrast between SIFTFisher and CNN-Fisher. CNN-Fisher performs worst at
the largest image scale since local activations come from
small image regions in an original image, while SIFTFisher performs best at the same scale since SIFT properly captures low-level contents within such small regions.
If we aggregate the CNN activations of all scales into one
Fisher vector by Eq. (2), the poorly performing 2,500 activations will have dominant influence with the large weight
of 2,500/4,410 in the image representation.
One possible strategy for aggregating multi-scale CNN
activations is to choose activations of a set of scales relatively performing well. However, the selection of good
scales is dependent on dataset and the activations from the
large image scale can also contribute to geometric invariance property if we balance the influence of each scale. We
empirically examined various combinations of pooling as
will be shown in Sec. 4 and we found that scale-wise Fisher
vector normalization followed by an simple average pooling
is effective to balance the influence.
We perform an experiment to compare our pooling
method in Eq. (1) to the naive Fisher pooling in Eq. (2).
In the experiment, we apply both of two pooling methods
with five different numbers of scales and perform classification on PASCAL VOC 2007. Despite the simplicity of
our multi-scale pyramid pooling, it demonstrates superior
3. Analysis of Multi-scale CNN Activations
We compare scale characteristics between traditional local features and CNN activations. It tells us that it is
not suitable to directly adopt a Fisher kernel framework to
multi-scale local CNN activations for representing an image. To investigate the best way for aggregating the CNN
activations into a global representation, we perform empirical studies and conclude that applying scale-wise normalization of Fisher vectors is very important.
S
A naive way to obtain a Fisher vector G 0 given multiscale local activations X = {x ∈ xs , s ∈ S} is to aggregate
them as
1 XX
S
G0 =
∇λ log Gλ (x).
(2)
|X|
x∈x
s∈S
65
1~
60
70
Naive Fisher
Proposed MPP
1(1
)
2(9
3(6 )
4(1 4)
9
5( 6)
6(1 484)
,
7(2 156)
,50
0)
20
1(1
2
2(6 )
3(1 5)
7
4( 3)
5(1 456)
,
6(2 036)
,35
7(4 4)
,95
0)
65
80
mAP (%)
70
75
1~
30
Accuracy (%)
mAP(%)
mAP(%)
75
40
s
Here, every multi-scale local activation vector is pooled to
one Fisher vector with an equal weight of 1/|X|.
To better combine a Fisher kernel with mid-level neural activations, the property of CNN activations according
to patch scale should be took in consideration. In the traditional use of Fisher kernel on visual classification tasks,
the hand-designed local descriptors such as SIFT [19] have
been often densely computed in multi-scale. This local descriptor encodes low-level gradient information within an
local region and captures detailed textures or shapes within
a small region rather than the global structure within a larger
region. In contrast, a mid-level neural activation extracted
from a higher layer of CNNs (e.g. FC6 or FC7 of [17]) represents higher level structure information which is closer to
class posteriors. As shown in the CNN visualization proposed by Zeiler and Fergus in [33], image regions strongly
activated by a certain CNN filter of the fifth layer usually
capture a category-level entire object.
To figure out the different scale properties between the
Fisher vector of traditional SIFT (SIFT-Fisher) and that of
neural activation from FC7 (CNN-Fisher), we conduct an
4
performances as depicted in Fig. 4. The performance of the
naive Fisher kernel pooling in Eq. (2) deteriorates rapidly
when finer scale levels are involved. This is because indistinctive neural activations from finer scale levels become
dominant in forming a Fisher vector. Our representation,
however, exhibits stable performance that the accuracy is
constantly increasing and finally being saturated. It verify
that our pooling method aggregates multi-scale CNN activations effectively.
the OverFeat [27], is also composed of five convolutional
layers (three in [27]) and three fully connected layers. It
shows 15.5% top-5 error on the ILSVRC’12 dataset with the
same center-crop. Compared to Alex, it uses 7×7 smaller
filters but dense stride of 2 in the first convolutional layer.
Our experiments are conducted mostly with the Alex
by default. The CNNS is used only for the PASCAL
VOC 2007 dataset to compare our method with [4], which
demonstrates excellent performance with the CNNS. Both
of the two pre-trained models are available online [31].
4. Experiments
4.3. Implementation Details
4.1. Datasets
We use an image pyramid of seven scales by default
since the seven scales can cover large enough scale variations and performance in all datasets as shown in Fig. 4.
The overall procedure of our image representation is as
follows. Given an image, we make an image pyramid containing seven scaled images. Each image in the pyramid has
twice resolution than the previous scale starting from the
standard size defined in each CNN (e.g. 227×227 for Alex).
We then feed each scale image to the CNN and obtain 4,410
vectors of 4,096 dimensional dense CNN activations from
the seventh layer. The dimensionality of each activation
vector is reduced to 128 by PCA where a projection is
trained with 256,000 activation vectors sampled from training images. A visual vocabulary (GMM of 256 Gaussian
distributions) is also trained with the same samples. Consequently, one 65,536 dimensional Fisher vector is computed
by Eq. (1), and further power- and `2 -normalization follow.
One-versus-rest linear SVMs with a quadratic regularizer
and a hinge loss are trained finally.
Our system is mostly implemented using open source libraries including VLFeat [30] for a Fisher kernel framework
and MatConvNet [31] for CNNs.
To evaluate our proposal as a generic image representation, we conduct three different visual recognition tasks
with following datasets.
MIT Indoor 67 [25] is used for a scene classification
task. The dataset contains 15,620 images with 67 indoor
scene classes in total. It is a challenging dataset because
many indoor classes are characterized by the objects they
contain (e.g. different type of stores) rather than their spatial properties. The performance is measured with top-1 accuracy.
PASCAL VOC 2007 [9] is used for an object classification task. It consists of 9,963 images of 20 object classes
in total. The task is quite difficult since the scales of the
objects fluctuate and multiple objects of different classes
are often contained in the same image. The performance is
measured with (11-points interpolated) mean average precision.
Oxford 102 Flowers [21] is used for a fine-grained object classification task, which distinguishes the sub-classes
of the same object class. This dataset consists of 8,189 images with 102 flower classes. Each class consists of various
numbers of images from 20 to 258. The performance is
measured with top-1 accuracy.
4.4. Results and Analysis
We perform comprehensive experiments to compare various methods on the three recognition tasks. We first
show the performance of our method and baseline methods.
Then, we compare our result with state-of-the-art methods
for each dataset. For simplicity, we use a notation protocol
“A(B)” where A denotes a pooling method and B denotes
descriptors to be pooled by A. The notations are summarized in Table 2.
We compare our method with several baseline methods. The baseline methods include intermediate CNN activations from a pre-trained CNN with a standard input, an
average pooling with multiple jittered images, and modified
versions of our method. The comparison results for each
dataset are summarized in Table 3(a), 4(a), 6(a). As expected, the most basic representation, Alex-FC7, performs
the worst for all datasets. The average pooling in AP10
and AP50 improves the performance +1.39%∼+3%, how-
4.2. Pre-trained CNNs
We use two CNNs pre-trained on the ILSVRC’12
dataset [6] to extract multi-scale local activations. One is the
Caffe reference model [15] composed of five convolutional
layers and three fully connected layers. This model performed 19.6% top-5 error when a single center-crop of each
validation image are used for evaluation on the ILSVRC’12
dataset. Henceforth, we denote this model by “Alex” since
it is nearly the same architecture of Krizhevsky et al.’s
CNN [17].
The other one is Chatfield et al.’s CNN-S model [4]
(“CNNS”, henceforth). This model, a simplified version of
5
ever the improvement is bounded regardless of the number of data augmentation. The other two baseline methods
(MPP w/o SN and CSF) exploit multi-scale CNN activations and they show better results than single-scale representations. Compared to the AP10, the performance gains
from multi-scale activations exceed +10%, +1%, and +5%
for each dataset. It shows that image representation based
on CNN activations can be enriched by utilizing multi-scale
local activations.
Even though baseline methods exploiting multi-scale
CNN activations show substantial improvements compared
to the single-scale baselines, we can also verify that handling multi-scale activations is important for further improvement. Compared to the naive Fisher kernel pooling
(NFK) in Eq. (2), our MPP achieves an extra but significant
performance gain of +4.18%, +4.58%, and 2.84% for each
dataset. Instead of pooling multi-scale activations as our
MPP, concatenating encoded Fisher vectors can be another
option as done in Gong et al.’s method [12]. The concatenation (CSF) also improves the performance, however the
CSF without an additional dimension reduction raises the
dimensionality proportional to the number of scales and the
MPP still outperforms the CSF for all datasets. The comprehensive test with various pooling strategies so far shows
that the proposed image representation can be used as a primary image representation in wide visual recognition tasks.
We also apply the spatial pyramid (SP) kernel [18] to
our representation. We construct a spatial pyramid into four
sub-regions (whole, top, middle, bottom) and it increases
the dimensionality of our representation four times. The results are unequable but the differences are marginal for all
datasets. This result is not surprising because the rich activations from smaller image scales already cover the global
layout. It makes the SP kernel redundant.
In Table 3(b), we compare our result with various stateof-the-art methods on Indoor 67. Similar to ours, Gong et
al. [12] proposed a pooling method for multi-scale CNN activations. They performed VLAD pooling at each scale and
concatenated them. Compared to [12], our representation
largely outperforms the method with a gain of +7.07%. The
performance gap possibly comes from 1) the large number
of scales, 2) the superiority of the Fisher kernel, and 3) the
details of pooling strategy. While they use only three scales,
we extract seven-scale activations with a quite efficient way
(Fig. 2). Though adding local activations from very finer
scales such as 6 or 7 in a naive way may harm the performance, it actually contribute to a better invariance property by the proposed MPP. In addition, as our experiment
of the “CSF” was shown, the MPP is more suitable for aggregating multi-scale activations than the concatenation. It
implies that our better performance does not just come from
the superior Fisher kernel, but from the better handling of
multi-scale neural activations.
The record holder in the Indoor 67 dataset has been
Zuo et al. [37] who combined the Alex-FC6 and their complementary features so called DSFL. The DSFL learns discriminative and shareable filters with a target dataset. When
we stack an additional MPP at the Pool5 layer, we already
achieve a state-of-the-art performance (77.76%) with a pretrained Alex only. We also stack the DSFL feature1 over
our representation and the result shows the performance of
80.78%. It shows that our representation is also improved
by combining complementary features.
The results on VOC 2007 is summarized in Table 4(b).
There are two methods ([22] and [26]) that use the same
Alex network. Razavian et al. [26] performed target data
augmentation and Oquab et al. [22] used a multi-layer perceptron (MLP) instead of a linear SVM with ground truth
bounding boxes. Our representation outperforms the two
methods using the pre-trained Alex without data augmentation or the use of bounding box annotations. The gains are
+1.84% and +2.34% respectively.
There are recent methods outperforming our method.
All of them are adopting better CNNs for the source task
(i.e. ImageNet classification) or the target task, such as
Spatial Pyramid Pooling (SPP) network [13], Multi-label
CNN [32] and the CNNS [4]. Our basic MPP(Alex-FC7)
demonstrates slightly lower precisions (79.54%) compared
to them, however we use the basic Alex CNN without finetuning on VOC 2007. When our representation is equipped
with the superior CNNS [4], which is not fine-tuned on
VOC 2007, our representation (81.40%) reaches nearly statof-the-art performance and our method is further improved
to 82.13% by stacking MPP(CNNS-FC8). The performance
is still lower than [4], who conduct target data augmentation
and fine-tuning. We believe our method can be further improved by additional techniques such as fine-tuning, target
data augmentation, or use of ground truth bounding boxes,
we leave the issue as future work because our major focus
is a generic image representation with a pre-trained CNN.
Table 6(b) shows the classification performances on 102
Flowers. Our method (91.28%) outperforms the previous
state-of-the-art method [26] (86.80%). Without the use of
a powerful CNN representation, various previous methods
show much lower performances.
Table 5 shows per-class performances on VOC 2007.
Compared to state-of-the-art methods, our method performs
best in 6 classes among 20 classes. It is interesting that the
6 classes include “bottle”, “pottedplant”, and “tvmonitor”,
which are the representative small objects in the VOC 2007
dataset. The results clearly demonstrates the benefit of our
MPP that aggregates activations from very finer-scales as
well, which are prone to harm the performance if it is handled inappropriately.
1 Pre-computed DSFL vectors for the MIT Indoor 67 dataset are provided by the authors.
6
Method
CNN-FC7
AP10(CNN-FC7)
AP50(CNN-FC7)
NFK(CNN-FC7)
CSF(CNN-FC7)
MPP(CNN-FC7)
Description
A standard activation vector from FC7 of a CNN with a center-crop of a 256 × 256 size input image.
Average pooling of a 5 crops and their flips, given a 256 × 256 size input image.
Average pooling of a 25 crops and their flips, given a 256 × 256 size input image.
Naive Fisher kernel pooling without scale-wise vector normalization, given a multi-scale image pyramid.
Concatenation of scale-wise normalized Fisher vectors, given a multi-scale image pyramid.
The proposed representation, given a multi-scale image pyramid.
Table 2. Summary of our notation protocol. Consequent image representations by the listed methods are finally `2 -normalized.
Method
Alex-FC7
AP10(Alex-FC7)
MPP(Alex-FC7)
MPP(CNNS-FC7)
Perronnin et al. [24] 10’
Razavian et al. [26] ’14
Oquab et al. [22] ’14
Wei et al. [32] ’14
Chatfieldet al. [4] ’14
FT
No.
No.
No.
No.
No.
No.
No.
Yes.
Yes.
plane
85.0
85.7
90.2
90.2
75.7
90.1
88.5
95.1
95.3
bike
79.7
80.8
86.9
88.6
64.8
84.4
81.5
90.1
90.4
bird
82.8
83.3
86.6
89.0
52.8
86.5
87.9
92.8
92.5
boat
80.4
80.7
84.4
84.7
70.6
84.1
82.0
89.9
89.6
bottle
39.7
40.4
54.0
58.2
30.0
48.4
47.5
51.5
54.4
bus
69.3
71.5
80.0
82.8
64.1
73.4
75.5
80.0
81.9
car
82.9
83.8
87.9
88.1
77.5
86.7
90.1
91.7
91.5
cat
81.7
82.7
86.0
89.0
55.5
85.4
87.2
91.6
91.9
chair
58.7
60.7
63.4
64.9
55.6
61.3
61.6
57.7
64.1
cow
57.8
60.5
72.2
77.0
41.8
67.6
75.7
77.8
76.3
table
68.5
70.6
75.7
78.4
56.3
69.6
67.3
70.9
74.9
dog
75.9
79.0
83.1
86.9
41.7
84.0
85.5
89.3
89.7
horse
83.0
84.5
87.8
89.2
76.3
85.4
83.5
89.3
92.2
motor
72.5
75.0
83.9
86.7
64.4
80.0
80.0
85.2
86.9
person
90.6
91.3
93.0
92.8
82.7
92.0
95.6
93.0
95.2
plant
51.7
53.4
64.8
61.2
28.3
56.9
60.8
64.0
60.7
sheep
71.1
70.1
75.8
81.3
39.7
76.7
76.8
85.7
82.9
sofa
60.8
62.6
69.6
70.0
56.6
67.3
58.0
62.7
68.0
train
85.0
86.5
89.9
89.8
79.7
89.1
90.4
94.4
95.5
tv
70.5
72.1
75.9
79.3
51.5
74.9
77.9
78.3
74.4
mAP
72.4
73.7
79.5
81.4
58.3
77.2
77.7
81.5
82.4
Table 5. Per-class classification performances on PASCAL VOC 2007. “FT” represents fine-tuning of a pre-trained CNN on VOC2007.
Method
Baseline
Baseline
Baseline
Baseline
Baseline
Ours
Ours
Ours
Ours
Description
CNN Acc.
Alex-FC7
Yes. 57.91
AP10(Alex-FC7)
Yes. 60.90
AP50(Alex-FC7)
Yes. 60.37
NFK(Alex-FC7)
Yes. 71.49
CSF(Alex-FC7)
Yes. 72.24
MPP(Alex-FC7)
Yes. 75.67
MPP(Alex-FC7)+SP
Yes. 75.97
MPP(Alex-FC7,Pool5)
Yes. 77.56
MPP(Alex-FC7)+DSFL[37] Yes. 80.78
(a) baselines and our methods.
Method
Description
CNN
Singh et al. [28] ’12
Part+GIST+DPM+SP
No.
Juneja et al. [16] ’13
IFK+Bag-of-Parts
No.
Doersch et al. [7] ’13 IFK+MidlevelRepresent. No.
Zuo et al. [37] ’14
DSFL
No.
Zuo et al. [37] ’14
DSFL+Alex-FC6
Yes.
Zhou et al. [36] ’14
Alex-FC7
Yes.
Zhou et al. [36] ’14
Alex-FC7
Yes.
Razavian et al. [26] ’14 AP(Alex)+PT+TargetAug. Yes.
Gong et al. [12] ’14
VLAD Concat.(Alex-FC7) Yes.
(b) state-of-the-art methods on MIT Indoor 67.
Method Description
FT BB CNN mAP
Baseline Alex-FC7
No. No. Yes. 72.36
Baseline AP10(Alex-FC7)
No. No. Yes. 73.75
Baseline AP50(Alex-FC7)
No. No. Yes. 73.60
Baseline NFK(Alex-FC7)
No. No. Yes. 74.96
Baseline CSF(Alex-FC7)
No. No. Yes. 78.46
Ours
MPP(Alex-FC7)
No. No. Yes. 79.54
Ours
MPP(Alex-FC7)+SP No. No. Yes. 79.29
Ours
MPP(CNNS-FC7)
No. No. Yes. 81.40
Ours
MPP(CNNS-FC7,FC8) No. No. Yes. 82.13
(a) Baselines and our methods.
Method
Description
FT BB CNN mAP
Perronnin et al. [24] 10’ IFK(SIFT+color)
No. No. No. 60.3%
He et al. [13] ’14
SPPNET-FC7
No. No. Yes. 80.10%
Wei et al. [32] ’14
Multi-label CNN
Yes. No. Yes. 81.50%
Razavian et al. [26] ’14 AP(Alex)+PT+TA No. No. Yes. 77.20%
Oquab et al. [22] ’14 Alex-FC7+MLP
No. Yes. Yes. 77.70%
Chatfield et al. [4] ’14 AP(CNNS-FC7)+TA No. No. Yes. 79.74%
Chatfield et al. [4] ’14 AP(CNNS-FC7)+TA Yes. No. Yes. 82.42%
(b) state-of-the-art methods on
PASCAL VOC 2007 classification.
Table 4. Classification performances on PASCAL VOC 2007 classification. “FT” represents fine-tuning of a pre-trained CNN on
VOC2007 and “BB” denotes the use of ground truth object bounding boxes in training. (SP: Spatial Pyramid, IFK: Improved Fisher
Kernel, SPPNET: Spatial Pyramid Pooling Network, PT: Power
Transform, TA: Target data Augmentation in training, MLP: Multilayer Perceptron.)
Acc.
49.40
63.18
66.87
52.24
76.23
68.24
70.80
69.00
68.90
Table 3. Classification performances on MIT Indoor 67. (SP: Spatial Pyramid, DPM: Deformable Part-based Model, PT: Power
Transform, IFK: Improved Fisher Kernel, DSFL: Discriminative
and Shareable Feature Learning.)
7
Method Description
Seg. CNN Acc.
Baseline Alex-FC7
No. Yes. 81.43
Baseline AP10(Alex-FC7)
No. Yes. 83.40
Baseline AP50(Alex-FC7)
No. Yes. 83.56
Baseline NFK(Alex-FC7)
No. Yes. 88.44
Baseline CSF(Alex-FC7)
No. Yes. 89.35
Ours
MPP(Alex-FC7)
No. Yes. 91.28
Ours
MPP(Alex-FC7)+SP No. Yes. 90.05
(a) baselines and our methods.
Method
Description
Seg. CNN Acc.
Nilsback and Zisserman [21] ’08 Multple kernel learning
Yes. No. 77.70
Angelova and Zhu [1] ’13
Seg+DenseHoG+LLC+MaxPooling Yes. No. 80.70
Murray and Perronnin [20] ’14 GMP of IFK(SIFT+color)
No. No. 81.50
Fernando et al. [10] ’14
Bag-of-FLH
Yes. No. 72.70
Razavian et al. [26] ’14
AP(Alex)+PT+TA
No. Yes. 86.8
(b) state-of-the-art methods on Oxford 102 Flowers.
Table 6. Classification performances on Oxford 102 Flowers. “Seg.” denotes the use of ground truth segmentations in training. (SP:
spatial pyramid, LLC: Locality-constrained Linear Coding, GMP: generalized max pooling, FLH: Frequent Local Histograms, PT: power
transform, TA: target data augmentation in training.)
Bottle
Bottle
Dining Table
Dining Table
Car
Car
Sofa
Sofa
TV monitor
TV monitor
Train
Train
Potted plant
Potted plant
Bus
Bus
Bicycle
Bicycle
Aeroplane
Aeroplane
Person
Person
Sheep
Sheep
Figure 5. Examples of object confidence maps obtained by our image representation on the PASCAL VOC 2007. All examples are test
images, not training images.
8
4.5. Weakly-Supervised Object Confidence Map
One interesting feature of our method is that we can
present object confidence maps for object classification
tasks, though we train the SVM classifiers without bounding box annotation but only with class-level labels. To recover confidence maps, we trace how much weight is given
to each local patch and accumulate all the weights of local activations. Tracing the weight of local activations is
possible because our final representation can be formed regardless of the number of scales and the number of local
activation vectors. To trace the weight of each patch, we
compute our final representation per patch using the corresponding single activation vector only and compute the
score from the pre-trained SVM classifiers we used for object classification.
Fig. 5 and Fig. 6 show several examples of object confidence map on the VOC 2007 test images. In the figures,
we can verify our image representation encodes the discriminative image patches well, despite large within-class
variations as well as substantial geometric changes. As we
discussed in Sec. 4.4, the images containing small-size objects also present the accurate confidence maps. These maps
may further be utilized as an considerable cue for object detection/localization and also be useful for analyzing image
representation.
Person
Boat
Car
Person
Aeroplane
Car
Bicycle
Person
Dog
Person
Bottle
Person
Cat
Dog
Car
Person
Person
Sheep
5. Discussion
We have proposed the multi-scale pyramid pooling for
better use of neural activations from a pre-trained CNN.
There are several conclusions we can derive through our
study. First, we should take the scale characteristic of neural activations into consideration for the successful combination of a Fisher kernel and a CNN. The activations become uninformative as a patch size becomes smaller, however they can contribute to better scale invariance when they
meet a simple scale-wise normalization. Second, dense
deep neural activations from multiple scale levels are extracted with reasonable computations by replacing the fully
connection with equivalent multiple convolution filters. It
enables us to pool the truly multi-scale activations and to
achieve significant performance improvements on the visual recognition tasks. Third, reasonable object-level confidence maps can be obtained from our image representation even though only class-level labels are given for supervision, which can be further applied to object detection
or localization tasks. In the comprehensive experiments on
three different recognition tasks, the results suggest that our
proposal can be used as a primary image representation for
better performances in various visual recognition tasks.
Figure 6. Examples of multi-object confidence map obtained by
our image representation on the PASCAL VOC 2007. All examples are test images, not training images.
9
References
[15] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. arXiv preprint
arXiv:1408.5093, 2014.
[16] M. Juneja, A. Vedaldi, C. V. Jawahar, and A. Zisserman.
Blocks that shout: Distinctive parts for scene classification.
In Proceedings of IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), 2013.
[17] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet
classification with deep convolutional neural networks. In
Advances in Neural Information Processing Systems (NIPS),
2013.
[18] S. Lazebnik, C. Schmid, and J. Ponce. Beyond bags of
features: Spatial pyramid matching for recognizing natural
scene categories. In Proceedings of IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), pages II:
2169–2178, 2006.
[19] D. G. Lowe. Distinctive image features from scale-invariant
keypoints. International Journal on Computer Vision (IJCV),
60(2):91–110, 2004.
[20] N. Murray and F. Perronnin. Generalized max pooling. Computing Research Repository (CoRR), abs/1406.0312, 2014.
[21] M.-E. Nilsback and A. Zisserman. Automated flower classification over a large number of classes. In Proceedings of the
Indian Conference on Computer Vision, Graphics and Image
Processing, 2008.
[22] M. Oquab, L. Bottou, I. Laptev, and J. Sivic. Learning and
transferring mid-level image representations using convolutional neural networks. In Proceedings of IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), 2014.
[23] F. Perronnin and C. R. Dance. Fisher kernels on visual
vocabularies for image categorization. In Proceedings of
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2007.
[24] F. Perronnin, J. Sánchez, and T. Mensink. Improving the
fisher kernel for large-scale image classification. In Proceedings of European Conference on Computer Vision (ECCV),
2010.
[25] A. Quattoni and A.Torralba. Recognizing indoor scenes. In
Proceedings of IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), 2009.
[26] A. S. Razavian, H. Azizpour, J. Sullivan, and S. Carlsson. CNN features off-the-shelf: an astounding baseline
for recognition. In Computing Research Repository (CoRR),
2014.
[27] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus,
and Y. LeCun. Overfeat: Integrated recognition, localization and detection using convolutional networks. In Proceedings of International Conference on Learning Representations (ICLR), 2014.
[28] S. Singh, A. Gupta, and A. A. Efros. Unsupervised discovery of mid-level discriminative patches. In Proceedings of
European Conference on Computer Vision (ECCV), 2012.
[29] J. Sivic and A. Zisserman. Video google: A text retrieval
approach to object matching in videos. In Proceedings of
IEEE International Conference on Computer Vision (ICCV),
2003.
[1] A. Angelova and S. Zhu. Efficient object detection and segmentation for fine-grained recognition. In Proceedings of
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 811–818. IEEE, 2013.
[2] A. Babenko, A. Slesarev, A. Chigorin, and V. Lempitsky.
Neural codes for image retrieval. In Proceedings of European Conference on Computer Vision (ECCV), 2014.
[3] A. Barbu, D. P. Barrett, W. Chen, S. Narayanaswamy,
C. Xiong, J. J. Corso, C. D. Fellbaum, C. Hanson, S. J. Hanson, S. Hélie, E. Malaia, B. A. Pearlmutter, J. M. Siskind,
T. M. Talavage, and R. B. Wilbur. Seeing is worse than believing: Reading people’s minds better than computer-vision
methods recognize actions. In Proceedings of European
Conference on Computer Vision (ECCV), 2014.
[4] K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman.
Return of the devil in the details: Delving deep into convolutional nets. In Proceedings of British Machine Vision
Conference (BMVC), 2014.
[5] Y. L. Cun, B. Boser, J. S. Denker, D. Henderson, R. Howard,
W. Hubbard, and L. Jackel. Backpropagation applied to
handwritten zip code recognition. Neural Computation,
1:541–551, 1989.
[6] J. Deng, W. Dong, R. Socher, L. J. Li, K. Li, and L. F. Fei.
Imagenet: A large-scale hierarchical image database. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009.
[7] C. Doersch, A. Gupta, and A. A. Efros. Mid-level visual element discovery as discriminative mode seeking. In Advances
in Neural Information Processing Systems (NIPS), 2013.
[8] J. Donahue, Y. Jia, O. Vinyals, J. Hoffman, N. Zhang,
E. Tzeng, and T. Darrell. DeCAF: A deep convolutional activation feature for generic visual recognition. In Computing
Research Repository (CoRR), 2013.
[9] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn,
and A. Zisserman. The PASCAL Visual Object Classes
Challenge 2007 (VOC2007) Results. http://www.pascalnetwork.org/challenges/VOC/voc2007/workshop/index.html.
[10] B. Fernando, É. Fromont, and T. Tuytelaars. Mining midlevel features for image classification. International Journal
on Computer Vision (IJCV), 108(3):186–203, 2014.
[11] R. B. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich
feature hierarchies for accurate object detection and semantic
segmentation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014.
[12] Y. Gong, L. Wang, R. Guo, and S. Lazebnik. Multi-scale
orderless pooling of deep convolutional activations features.
In Proceedings of European Conference on Computer Vision
(ECCV), 2014.
[13] K. He, X. Zhang, S. Ren, and J. Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition.
In Proceedings of European Conference on Computer Vision
(ECCV), 2014.
[14] H. Jegou, M. Douze, C. Schmid, and P. Pérez. Aggregating
local descriptors into a compact image representation. In
Proceedings of IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), 2010.
10
[30] A. Vedaldi and B. Fulkerson. VLFeat: An open and portable
library of computer vision algorithms. http://www.
vlfeat.org/, 2008.
[31] A. Vedaldi and K. Lenc. MatConvNet: Convolutional neural networks for matlab. http://www.vlfeat.org/
matconvnet/, 2014.
[32] Y. Wei, W. Xia, J. Huang, B. Ni, J. Dong, Y. Zhao, and
S. Yan. Cnn: Single-label to multi-label. In Computing Research Repository (CoRR), 2014.
[33] M. D. Zeiler and R. Fergus. Visualizing and understanding
convolutional networks. In Proceedings of European Conference on Computer Vision (ECCV), 2014.
[34] N. Zhang, J. Donahue, R. B. Girshick, and T. Darrell.
Part-based R-CNNs for fine-grained category detection. In
Proceedings of European Conference on Computer Vision
(ECCV), 2014.
[35] N. Zhang, M. Paluri, M. Ranzato, T. Darrell, and L. D. Bourdev. PANDA: Pose aligned networks for deep attribute modeling. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014.
[36] B. Zhou, A. Lapedriza, J. Xiao, A. Torralba, and A. Oliva.
Learning deep features for scene recognition using places
database. In Advances in Neural Information Processing Systems (NIPS), 2014.
[37] Z. Zuo, G. Wang, B. Shuai, L. Zhao, Q. Yang, and X. Jiang.
Learning discriminative and shareable features for scene
classification. In Proceedings of European Conference on
Computer Vision (ECCV), 2014.
11
| 9cs.NE
|
T HREE FACTORS INFLUENCING MINIMA IN SGD
Stanisław Jastrz˛ebski∗†, Zachary Kenton∗‡, Devansh Arpit‡, Nicolas Ballas§,
Asja Fischer¶ , Yoshua Bengio‡k, Amos Storkey∗∗
arXiv:1711.04623v1 [cs.LG] 13 Nov 2017
A BSTRACT
We study the properties of the endpoint of stochastic gradient descent (SGD). By
approximating SGD as a stochastic differential equation (SDE) we consider the
Boltzmann-Gibbs equilibrium distribution of that SDE under the assumption of
isotropic variance in loss gradients. Through this analysis, we find that three factors – learning rate, batch size and the variance of the loss gradients – control the
trade-off between the depth and width of the minima found by SGD, with wider
minima favoured by a higher ratio of learning rate to batch size. We have direct
control over the learning rate and batch size, while the variance is determined
by the choice of model architecture, model parameterization and dataset. In the
equilibrium distribution only the ratio of learning rate to batch size appears, implying that the equilibrium distribution is invariant under a simultaneous rescaling
of learning rate and batch size by the same amount.
We then explore experimentally how learning rate and batch size affect SGD from
two perspectives: the endpoint of SGD and the dynamics that lead up to it. For the
endpoint, the experiments suggest the endpoint of SGD is invariant under simultaneous rescaling of batch size and learning rate, and also that a higher ratio leads
to flatter minima, both findings are consistent with our theoretical analysis. We
note experimentally that the dynamics also seem to be invariant under the same
rescaling of learning rate and batch size, which we explore showing that one can
exchange batch size and learning rate for cyclical learning rate schedule. Next,
we illustrate how noise affects memorization, showing that high noise levels lead
to better generalization. Finally, we find experimentally that the invariance under
simultaneous rescaling of learning rate and batch size breaks down if the learning
rate gets too large or the batch size gets too small.
1
I NTRODUCTION
Despite being massively over-parameterized (Zhang et al., 2016), deep neural networks (DNNs)
have demonstrated good generalization ability and achieved state-of-the-art performances in many
application domains such as image (He et al., 2016) and speech recognition (Amodei et al., 2016).
The reason for this success has been a focus of research recently but still remains an open question.
Our work provides new theoretical insights and useful suggestions for deep learning practitioners.
The standard way of training DNNs involves minimizing a loss function using SGD and its variants (Bottou, 1998). In SGD, parameters are updated by taking a small discrete step depending on
the learning rate in the direction of the negative loss gradient, which is approximated based on a
small subset of training examples (called a mini-batch). Since the loss functions of DNNs are highly
non-convex functions of the parameters, with complex structure and potentially multiple minima
and saddle points, SGD generally converges to different regions of parameter space depending on
optimization hyper-parameters and initialization.
∗
First two authors contributed equally
Jagiellonian University
‡
MILA, Université de Montréal
§
Facebook AI Research
¶
University of Bonn
k
CIFAR Senior Fellow
∗∗
University of Edinburgh
†
1
Recently, several works (Arpit et al., 2017; Advani & Saxe, 2017; Shirish Keskar et al., 2016) have
investigated how SGD impacts generalization in DNNs. It has been argued that wide minima tend to
generalize better than sharp minima (Hochreiter & Schmidhuber, 1997; Shirish Keskar et al., 2016).
This is entirely compatible with a Bayesian viewpoint that emphasizes targeting the probability mass
associated with a solution, rather than the density value at a solution (MacKay, 1992). Specifically,
(Shirish Keskar et al., 2016) find that larger batch sizes correlate with sharper minima. In contrast,
we find that it is the ratio of learning rate to batch size which is correlated with sharpness of minima,
not just batch size alone. In this vein, while (Dinh et al., 2017) discuss the existence of sharp
minima which behave similarly in terms of predictions compared with wide minima, we argue that
SGD naturally tends to find wider minima at higher noise levels in gradients, and such wider minima
seem to correlate with better generalization.
In order to achieve our goal, we approximate SGD as a continuous stochastic differential equation
(Bottou, 1991; Mandt et al., 2017; Li et al., 2017). Assuming isotropic gradient noise, we derive the
Boltzmann-Gibbs equilibrium distribution of this stochastic process, and further derive the relative
probability of landing in one local minima as compared to another in terms of their depth and width.
Our main finding is that the ratio of learning rate to batch-size along with the gradient’s covariances
influence the trade-off between the depth and sharpness of the final minima found by SGD, with a
high ratio of learning rate to batch size favouring flatter minima. In addition, our analysis provides
a theoretical justification for the empirical observation that scaling the learning rate linearly with
batch size (up to a limit) leads to identical performance in DNNs (Krizhevsky, 2014; Goyal et al.,
2017).
We verify our theoretical insights experimentally on different models and datasets. In particular, we
demonstrate that high learning rate to batch size ratio (due to either high learning rate or low batchsize) leads to wider minima and correlates well with better validation performance. We also show
that a high learning rate to batch size ratio helps prevent memorization. Furthermore, we observe that
multiplying each of the learning rate and the batch size by the same scaling factor results in similar
training dynamics. Extending this observation, we validate experimentally that one can exchange
learning rate and batch size for the recently proposed cyclic learning rate (CLR) schedule (Smith,
2015), where the learning rate oscillates between two levels. Finally, we discuss the limitations of
our theory in practice.
2
R ELATED WORK
The relationship between SGD and sampling a posterior distribution via stochastic Langevin methods has been the subject of discussion in a number of papers (Chaudhari et al., 2017; Chen et al.,
2014; Ding et al., 2014; Vollmer et al., 2015; Welling & Teh, 2011; Shang et al., 2015; Sato &
Nakagawa, 2014). In particular, (Mandt et al., 2017) describe the dynamics of stochastic gradient
descent (SGD) as a stochastic process that can be divided into three distinct phases. In the first
phase, weights diffuse and move away from the initialization. In the second phase the gradient
magnitude dominates the noise in the gradient estimate. In the final phase, the weights are near the
optimum. (Shwartz-Ziv & Tishby, 2017) make related observations from an information theoretic
point of view and suggest the diffusion behaviour of the parameters in the last phase leads to the
minimization of mutual information between the input and hidden representation. We also relate the
SGD dynamics to the stationary distribution of the stochastic differential equation. Our derivation
bears similarity with (Mandt et al., 2017). However, while (Mandt et al., 2017) study SGD as an
approximate Bayesian inference method in the final phase of optimization in a locally convex setting, our end goal is to analyze the stationary distribution over the entire parameter space reached by
SGD. Further, our analysis allows us to compare the probability of SGD ending up in one minima
over another (in terms of width and depth), which is novel in our case.
Our work is also closely related to the ongoing discussion about the role of large batch size and the
sharpness of minima found in terms of generalization (Shirish Keskar et al., 2016). (Shirish Keskar
et al., 2016) showed that SGD ends up in sharp minimum when using large batch size. (Goyal et al.,
2017; Hoffer et al., 2017) empirically observed that scaling up the learning rate, and training for
more epochs, leads to good generalization when using large batch size. Our novelty is in explaining
the importance of the ratio of learning rate to batch size. In particular, our theoretical and empirical
2
results show that simultaneously rescaling the batch size and learning rate by the same amount leads
SGD to minima having similar width despite using different batch sizes.
Concurrent with this work, (Smith & Le, 2017; Chaudhari & Soatto, 2017) have both analyzed SGD
approximated as a continuous time stochastic process and stressed the importance of the learning
rate to batch size ratio. (Smith & Le, 2017) focused on the training dynamics while (Chaudhari &
Soatto, 2017) explored the stationary non-equilibrium solution for the stochastic differential equation for non-isotropic gradient noise, but assuming other conditions on the covariance and loss to
enforce the stationary distribution to be path-independent. Their solution does not have an explicit
solution in terms of the loss in this case. In contrast to other work, we strictly focus on the explicitly solvable case of the Boltzmann-Gibbs equilibrium distribution with isotropic noise. This focus
allows us to relate the noise in SGD, controlled by the learning rate to batch size ratio, with the
width of its endpoint. We empirically verify that the width and height of minima correlates with the
learning rate to batch size ratio in practice.
Our work continues the line of research on the importance of noise in SGD (Bottou, 1998; Roux
et al., 2008; Neelakantan et al., 2015; Mandt et al., 2017). Our novelty is in formalizing the impact
of batch size and learning rate (i.e. noise level) on the width and depth of the final minima, and
empirical verifications of this.
3
T HEORETICAL R ESULTS
Our focus in this section is on finding the relative probability with which we end optimization in
a certain minimum, as compared to another minimum. We will find that the relative probability
depends on the local geometry of the loss function at each minimum along with batch size, learning
rate and the covariance of the loss gradients. To reach this result, we first derive the equilibrium
distribution of SGD over the parameter space under a stochastic differential equation treatment. We
make the assumption of isotropic covariance of the loss gradients, which allows us to write down
an explicit expression for the equilibrium distribution, which turns out to be a Boltzmann-Gibbs
distribution.
3.1
S ETUP
We follow a theoretical setup similar to (Mandt et al., 2017), approximating SGD with a continuoustime stochastic process, which we now outline.
Let us consider a model parameterized by θ = {θ1 , . . . , θq }. For N training examples xi , i ∈
{1, ..., N }, the loss function, L(θ), and the corresponding gradient g(θ), are defined based on the
sum over the loss values for all training examples. Stochastic gradients g(S) (θ) arise when we
consider a batch B of size S < N of random indices drawn uniformly from {1, ..., N } and form an
(unbiased) estimate of loss and gradient based on the corresponding subset of training examples
1 X
∂ (S)
L(S) (θ) =
l(θ, xn ) ,
g(S) (θ) =
L (θ) .
S
∂θ
n∈B
We consider stochastic gradient descent (SGD) with learning rate η, as defined by the update rule
θ(t + 1) = θ(t) − ηg (S) (θ) .
We now make the following assumptions:
(1) By the central limit theorem (CLT), we assume the noise in the stochastic gradient is Gaussian with covariance matrix S1 C(θ)
1
g(S) (θ) = g(θ) + √ ∆g(θ), where ∆g(θ) ∼ N (0, C(θ)) .
S
We note that the covariance is symmetric positive-semidefinite, and so can be decomposed
into the product of two matrices C(θ) = B(θ)B> (θ) .
(2) We assume the discrete process of SGD can be approximated by the continuous time limit
of the following stochastic differential equation (known as a Langevin equation)
dθ
η
= −ηg(θ) + √ B(θ)f (t)
(1)
dt
S
3
where f (t) is a normalized Gaussian time-dependent stochastic term.
Note that the continuous time approximation of SGD as a stochastic differential equation has been
shown to hold in a weak approximation on the condition that the learning rate is small (Li et al.,
2017).
3.2
M AIN R ESULTS
The Langevin equation is a stochastic differential equation, and we are interested in its equilibrium distribution which gives insights into the behavior of SGD and the properties of the points
it converges to. Assuming isotropic noise, the Langevin equation is well known to have a GibbsBoltzmann distribution as its equilibrium distribution. This equilibrium distribution can be derived
by finding the stationary solution of the Fokker-Planck equation, with detailed balance, which governs the evolution of the probability density of the value of the parameters with time. The FokkerPlanck equation and its derivation can be found in Appendix A in equation (5). In Appendix C
we restate the standard proofs of the stationary distribution of a Langevin system, and provide the
resulting Gibbs-Boltzmann equilbirium distribution here, using the notation of this paper:
Theorem 1 (Equilibrium Distribution). Assume1 that the gradient covariance is isotropic,
i.e. C(θ) = σ 2 I, where σ 2 is a constant. Then the equilibrium distribution of the stochastic differential equation 1 is given by
2L(θ)
P (θ) = P0 exp −
,
(2)
nσ 2
where n ≡ Sη and P0 is a normalization constant, which is well defined for loss functions with L2
regularization.
Discussion: Here P (θ) defines the density over the parameter space. The above result says that if
we run SGD for long enough (under the assumptions made regarding the SGD sufficiently matching
the infinitesimal limit), then the probability of the parameters being in a particular state asymptotically follows the above density. Note, that n ≡ Sη is a measure of the noise in the system set by the
choice of learning rate η and batch size S. The fact that the loss is divided by n emphasizes that the
higher the noise n, the less granular the loss surface appears to SGD. The gradient variance C(θ)
on the other hand is determined by the dataset and model priors (e.g. architecture, model parameterization, batch normalization etc.). This reveals an important area of investigation, i.e., how different
architectures and model parameterizations affect the gradient’s covariance structure. We note that in
the analysis above, the assumption of the gradient covariance C(θ) being fixed and isotropic in the
parameter space is unrealistic. However it is a simplification that enables straightforward insights
regarding the relationship of the noise, batch size and learning rate in the Gibbs-Boltzmann equilibrium. We empirically show that various predictions based on this relationship hold in practice.
Returning to SGD as an optimization method, we can ask, given the probability density P (θ) can
we derive the probability of ending at a given minimum, θ A , which we will denote by lowercase
pA = p̃A C, where C is a normalization constant which is the same for all minima (the unnormalized
probability p̃A is all we are interested in when estimating the relative probability of finishing in a
given minimum compared to another one). This probability is derived in Appendix D, and given in
the following theorem, which is the core result of our theory.
Theorem 2 (Probability of ending in region near minima θ A ). Assume the loss is locally strictly
convex with Hessian HA and loss LA at a minimum θ A . Then the unnormalized probability of
ending in minima θ A is
1
2LA
p̃A = √
exp − 2
(3)
nσ
det HA
where n =
η
S
is the noise used in the SGD process to reach θA .
Discussion: For this analysis, we qualitatively categorize a minima θ A by its loss LA (depth) and
the determinant of the Hessian det HA (a larger determinant implies a sharper minima). The above
1
Here we also assume a weak regularity condition that the loss L(θ) includes the regularization term τ kθk22
for some τ > 0.
4
result shows that the probability of landing in a specific minimum depends on three factors - learning
rate, batch-size and covariance of the gradients. The two factors that we directly control only appear
in the ratio given by the noise n = η/S.
To study which kind of minima are more likely if we were to reach equilibrium, it is instructive to
consider the ratio of probabilities pA and pB at minima θ A and θ B respectively given by
r
pA
2
det HB
(LB − LA ) .
=
exp
pB
det HA
nσ 2
To highlight that towards the equilibrium solution SGD favors wider rather than sharper minima,
let’s consider the special case when LA = LB , i.e., both minima have the same loss value. Then,
r
pA
det HB
=
.
pB
det HA
This case highlights that in equilibrium, SGD favors the minimum with lower determinant of the
Hessian (i.e. the flatter minima) when all other factors are identical. On the flip side, it can be seen
that if two minima have the same curvature (det HA = det HB ), then SGD will favor the minima
with lower loss. Finally in the general case when LA ≥ LB , it holds that pA ≥ pB if and only if
q
det HB
2
2σ
log
det HA
1
≤
≡Y .
n
(LA − LB )
That is, there is an upper bound on the inverse of the noise for θ A to be favored in the case that its
loss is higher than at θ B , and this upper bound depends on the difference in the heights compared
to the ratio of the widths. In particular we can see that if det HB < det HA , then Y < 0, and so
no amount of noise will result in θ A being more probable than θ B . In words, if the minimum at θ A
is both higher and sharper than the minimum at θ B , it is never reached with higher probability than
θ B , regardless of the amount of noise. However, if det HB > det HA then Y > 0, and there is a
lower bound on the noise
(LA − LB )
q
n>
(4)
det HB
2σ 2 log
det HA
to make θ A more probable than θ B . In words, if the minimum at θ A is higher but flatter than the
minimum at θ B , it is favored over θ B , as long as the noise is large enough, as defined by eq. (4).
To summarize, the presented theory shows that the noise level in SGD (which is defined by the ratio
of learning rate to batch size) controls the extent to which optimization favors wider over deeper
minima. Increasing the noise by increasing the ratio of learning rate to batch size increases the
probability of wider compared to deeper minima. For a discussion on the relative probabilities of
critical points that are not strictly minima, see appendix D.
4
4.1
E XPERIMENTS
I MPACT OF
η
S ON THE MINIMA
SGD FINDS
In this section, we empirically study the impact of learning rate η and batch size S on the local
minimum that SGD finds. We first focus on a 4-layer Batch Normalized ReLU MLP trained on
Fashion-MNIST (Xiao et al., 2017). We study how the noise ratio n = Sη leads to minima with
different curvatures and validation accuracy. To measure the curvature at a minimum, we compute
the norm of its Hessian (a higher norm implies higher sharpness of the minimum) using the finite
difference method (Wu et al., 2017). In Figure 1a, we report the norm of the Hessian for local
minima obtained by SGD for different n = Sη , where η ∈ [5e − 3, 1e − 1] and S ∈ [25, 1000].
Each experiment is run for 200 epochs; most models reach approximately 100% accuracy on train
set. As n grows, we observe that the norm of the Hessian at the minima also decreases, suggesting
that higher Sη pushes the optimization towards flatter minima. This agrees with Theorem 2, Eq. (3),
that higher Sη favors flatter over sharper minima.
Figure 1b shows the results from exploring the impact of n = Sη on the final validation performance,
which confirms that better generalization correlates with higher n. Taken together, Fig. 1a and
5
Validation performance
Log Hessian norm
90%
90%
89%
89%
88%
88%
87%
87%0E+00 1E-03 2E-03 3E-03 4E-03
4E-01
2E-01
0E+00
0E+00
2E-03
4E-03
/S
η
S
(a) Correlation of
norm of Hessian.
/S
with logarithm of (b) Correlation of
accuracy.
η
S
with validation
Figure 1: Impact on SGD with ratio of learning rate η and batch size S for 4 layer ReLU MLP on
FashionMNIST.
train accuracy
val accuracy
train loss
val loss
40
20
0
1.0
0.5
(a) left
0.0
0.5
η=0.1
,
S=128
1.0
right
1.5
2.0
η=0.1
S=1024
100
train accuracy
val accuracy
train loss
val loss
80
60
40
20
0
1.0
(b) left
0.5
CIFAR10 (Resnet56): /S = 0.1/1024, /S = 0.01/128
Accuracy (and scaled cross-entropy)
80
60
CIFAR10 (Resnet56): /S = 0.1/128, /S = 0.01/128
Accuracy (and scaled cross-entropy)
Accuracy (and scaled cross-entropy)
CIFAR10 (Resnet56): /S = 0.1/128, /S = 0.1/1024
100
0.0
0.5
η=0.1
,
S=128
1.0
right
1.5
η=0.01
S=128
2.0
100
80
60
40
20
0
train accuracy
val accuracy
train loss
val loss
1.0
(c) left
0.5
0.0
η=0.1
,
S=1024
0.5
1.0
right
1.5
2.0
η=0.01
S=128
Figure 2: Interpolation of Resnet56 networks trained with different learning rate to batch size ratio,
η
η
S . α (x-axis) corresponds to the interpolation coefficient. As predicted by our theory, lower S ratio
leads to sharper minima (as shown by the left and middle plot).
Fig. 1b imply wider minima correlate well with better generalization. As n = Sη increases, SGD
finds local minima that generalize better. In Appendix F.1, we report similar results for Resnet56
applied on CIFAR10 in Figure 9, for a 20 layer ReLU network with good initialization schemes in
Figures 10a and 10c, and with bad initilization in Figure 10b.
100
80
60
40
20
0
CIFAR10 (VGG11): = 1, = 4
Accuracy (and scaled cross-entropy)
Accuracy (and scaled cross-entropy)
To further illustrate the behavior of SGD with different noise levels, we train three Resnet56 models
on CIFAR10 using SGD (with a learning rate of 0.1 and without momentum) with different Sη .
η=0.1
η=0.1
Our baseline model uses S=128
. In comparision, we investigate a large batch model with S=1024
train accuracy
val accuracy
train loss
val loss
1.0 0.5 0.0 0.5 1.0 1.5 2.0
CIFAR10 (VGG11): = 1, = 0.25
100
80
train accuracy
60
val accuracy
train loss
40
val loss
20
0
1.0 0.5 0.0 0.5 1.0 1.5 2.0
(a) β = 1 corresponds to model at α = 0 and (b) β = 1 corresponds to model at α = 0 and
β = 4 corresponds to model at α = 1
β = 0.25 corresponds to model at α = 1
Figure 3: Interpolation between parameters of models trained with the same learning rate (η) to
batch-size (S) ratio: η=0.1×β
S=50×β , but different η and S values determined by β. As predicted by our
theory, the minima for models with identical noise levels should be qualitatively similar as can be
seen by these plots.
6
100
80
80
Accuracy
Accuracy
100
60
40
20
0
100
Epoch
Cyclic BS Train
Cyclic LR Train
Cyclic BS Test
Cyclic LR Test
200
300
60
40
bs=128, lr=0.001 train
bs=640, lr0.005 train
bs=128, lr=0.001 test
bs=640, lr=0.005 test
100
200
300
Epoch
20
0
Figure 4: Learning rate schedule can be replaced by an equivalent batch size schedule. The ratio of
learning rate to batch size is equal at all times for both red and blue curves in each plot. Above plots
show train and test accuracy for experiments involving VGG-11 architecture on CIFAR10 dataset.
Left: cyclic batch size schedule (blue) in range 128 to 640, compared to cyclic learning rate schedule
(red) in range 0.001 to 0.005. Right: constant batch size 128 and constant learning rate 0.001 (blue),
compared to constant batch size 640 and constant learning rate 0.005 (red).
η
and a small learning rate model with η=0.01
S=128 , which have approximately the same S ratio. We
follow (Goodfellow et al., 2014) by investigating the loss on the line interpolating between the
parameters of two models. More specifically, let θ 1 and θ 2 be the final parameters found by SGD
using different Sη , we report the loss values L((1 − α)θ 1 + αθ 2 ) for α ∈ [−1, 2]. Results indicate
that models with large batch size (Fig. 2-left) or low learning rate (Fig. 2-middle; both having a
lower Sη than the baseline) end up in a sharper minimum relative to the baseline model. These plots
are consistent with our theoretical analysis that higher n = η/S gives preference to wider minima
over sharper minima. On the other hand, figure 2 (right) shows that models trained with roughly
the same level of noise end up in minima of similar quality. The following experiment explores this
aspect further.
We train VGG-11 models (Simonyan & Zisserman, 2014) on CIFAR-10, such that all the models are trained with the same noise level but with different values of learning rate and batch size.
Specifically, we use η=0.1×β
S=50×β , where we set β = 0.25, 1, 4. We then interpolate between the model
parameters found when training with β = 1 and β = 4 (Fig. 3-left), and β = 1 and β = 0.25
(Fig. 3-right). The interpolation results indicate that all the minima have similar width and depth,
qualitatively supporting our theoretical observation that for the same noise ratio SGD ends up in
minima of similar quality.
4.2
η
S RATIO INFLUENCES LEARNING DYNAMICS OF
SGD
In this section we look at two experimental phenomena: firstly, the equilibrium endpoint of SGD
and secondly the dynamical evolution of SGD. The former, was theoretically analysed in the theory
section, while the latter is not directly addressed in the theory section, but we note that the two are
related - the endpoint is the result of the intermediate dynamics.
We experimentally study both phenomena in the following four experiments involving the VGG11 architecture on the CIFAR10 dataset, shown in Fig 4. The left plot compares two experiments:
cyclic batch size schedule (blue) in range 128 to 640, compared to cyclic learning rate schedule (red)
in range 0.001 to 0.005. The right plot compares two other experiments: constant learning rate to
η
0.005
batch-size ratio of Sη = 0.001
128 (blue) and S = 640 (red).
Regarding the first phenomena, of the endpoint of SGD, the test accuracy when training with a
cyclic batch size and cyclic learning rate is 89.39% and 89.24%, respectively, and we emphasize
that these are similar scores. For a constant learning rate to batch-size ratio of Sη = 0.001
128 and
η
0.005
=
,
the
test
accuracy
is
87.25%
and
86.92%,
respectively,
and
we
again
emphasize
that
these
S
640
two scores are similar to each other. That in each of these experiments the endpoint test accuracies
7
Random label accuracy
25.0% random labels 55% 50.0% random labels
76%
74%
50%
72%
45%
70%
68%
40%
65%
63%
35%
62%
60%0% 20% 40% 60% 80% 100% 30%0% 20% 40% 60% 80% 100%
Validation accuracy
Validation accuracy
25.0% random labels 55% 50.0% random labels
76%
74%
50%
72%
45%
70%
68%
40%
65%
63%
35%
62%
60%0% 20% 40% 60% 80% 100% 30%0% 20% 40% 60% 80% 100%
Random label accuracy
Random label accuracy
Random label accuracy
Figure 5: Impact of Sη on memorization of MNIST when 25% and 50% of labels in the training set
are replaced with random labels, using no momentum (on the right) or a momentum with parameter 0.9 (on the left). We observe that for a specific level of memorization, high Sη leads to better
generalization. Red has higher value of Sη than blue.
are similar shows exchangability of learning rate for batch size for the endpoint, and is consistent
with our theoretical calculation which says characteristics of the minima found at the endpoint are
determined by the ratio of learning rate to batch-size, but not individually on learning rate or batch
size.
Regarding the second phenomena of the dynamical evolution, we note the similarity of the training and test accuracy curves for each pair of same-noise curves in each experiment. Our theoretical analysis does not explain this phenomena, as it does not determine the dynamical distribution.
Nonetheless, we report it here as an interesting observation, and point to Appendix B for some intuition on why this may occur from the Fokker-Planck equation. In Appendix F.2, Fig. 11 we show
in more detail the loss curves. While the epoch-averaged loss curves match well when exchanging
batch size for learning rate, the per-iteration loss is not invariant to switching batch size for learning
rate. In particular, we note that each run with smaller batch-size has higher variance in per-iteration
loss than it’s same-noise pair. This is expected, since from one iteration to the next, the examples
will have higher variance for a smaller batch-size.
The take-away message from this section is that the endpoint and dynamics of SGD are approximately invariant if the batch size and learning rate are simultaneously rescaled by the same amount.
This is in contrast to a commonly used heuristic consisting
√ of scaling the learning rate with the
square root of the batch size, i.e. of keeping the ratio η/ S constant. This is used for example by
(Hoffer et al., 2017) as a way of keeping the covariance matrix of the parameter update step the
same for any batch size. However, our theory and experiments suggest changing learning rate and
batch size in a way that keeps the ratio n = η/S constant instead, since this results in the same
equilibrium distribution.
4.3
I MPACT OF SGD ON M EMORIZATION
To generalize well, a model must identify the underlying pattern in the data instead of simply perfectly memorizing each training example. An empirical approach to test for memorization is to
analyze how good a DNN can fit a training set when the true labels are partly replaced by random
labels (Zhang et al., 2016; Arpit et al., 2017). The experiments described in this section highlight
that SGD with a sufficient amount of noise improves generalization at a given level of memorization.
Experiments are performed on the MNIST dataset with an MLP similar to the one used by (Arpit
et al., 2017), but with 256 hidden units. We train the MLP with different amounts of random labels
in the training set. For each level of label noise, we evaluate the impact of Sη on the generalization
performance. Specifically, we run experiments with Sη taking values in a grid with batch size in
{50, 100, 200, 500, 1000}, learning rate in {0.005, 0.01, 0.02, 0.05, 0.07, 0.1}, and momentum in
{0.0, 0.9}. Models are trained for 300 epochs. Fig. 5 reports the MLPs performances on both the
noisy training set and the validation set. The results show that larger noise in SGD (regardless if
induced by using a smaller batch size or a larger learning rate) leads to solutions which generalize
better for the same amount of random labels memorized on the training set. Thus, our analysis
highlights that SGD with low noise n = Sη steers the endpoint of optimization towards a minimum
with low generalization ability.
8
Discrete S
Hessian norm
Loss
1.6E+05
Discrete
1.6E+05
Triangle
= 0.001
1.6E+05
1.4E+05
1.4E+05
1.4E+05
1.2E+05
1.2E+05
1.2E+05
1.2E+05
1.2E+05
1.0E+05
1.0E+05
1.0E+05
1.0E+05
1.0E+05
8.0E+04
8.0E+04
8.0E+04
8.0E+04
8.0E+04
6.0E+04
6.0E+04
6.0E+04
6.0E+04
6.0E+04
4.0E+04
4.0E+04
4.0E+04
4.0E+04
4.0E+04
2.0E+04
2.0E+04
2.0E+04
2.0E+04
2.0E+04
0 50 100 150 200 250 300
Epoch
0 50 100 150 200 250 300
Epoch
0 50 100 150 200 250 300
Epoch
= 0.005
1.6E+05
1.4E+05
Hessian norm
1.4E+05
0 50 100 150 200 250 300
Epoch
2.0
1.5
Train loss
1.6E+05
1.0
0.5
0 50 100 150 200 250 300
Epoch
0.0
Figure 6: Cyclical schemes oscillate between sharp and wide regions. Additionally, cyclical schemes
find wider minima than baseline run for same level of loss, which might explain their better generalization. All cyclical schedules use base η = 0.005 and cycle length 15 epochs, which approximate
convergence at the end of each cycle. Plots from left to right: discrete S, discrete η, triangle η,
constant learning rate η = 0.001, constant learning rate η = 0.005. On vertical axis we report loss
(red) and approximated norm of Hessian (blue).
Discrete η
Discrete S
Triangle η
Baseline
Test acc
Valid acc
Loss
Hessian norm.
90.04% ± 0.18%
90.07% ± 0.32%
90.03% ± 0.10%
87.70% ± 0.56%
90.30% ± 0.07%
90.25% ± 0.06%
90.04% ± 0.23%
88.36% ± 0.13%
0.048 ± 0.001
0.050 ± 0.002
0.068 ± 0.002
0.033 ± 0.001
36470
13918
35310
57838
Table 1: Comparison between different cyclical training schedules (cycle length and learning rate
are optimized using a grid search). Discrete schedules perform similarly, or slightly better than
triangular. Additionally, discrete S schedule leads to much wider minima for similar loss. Hessian
norm is approximated on 1 out of 6 repetitions and measured at minimum value (usually endpoint
of training).
While Fig 5 reports the generalization at the endpoint, we observe that SGD with larger noise continuously steers away from sharp solutions throughout the dynamics. We also reproduce the observation reported by (Arpit et al., 2017): that memorization roughly starts after reaching maximum
generalization. For runs with momentum we exclude learning rates higher than 0.02 as they lead to
divergence. Full learning curves are reported in Fig. 12 included in Appendix F.3.
4.4
C YCLICAL BATCH AND L EARNING R ATE S CHEDULES
It has been observed that a cyclic learning rate (CLR) schedule leads to better generalization (Smith,
2015). In Sec. 4.2 we demonstrated that one can exchange cyclic learning rate schedule (CLR) with
cyclic batch size (CBS) and approximately preserve the practical benefit of CLR. This exchangeability shows that the generalization benefit of CLR must come from the varying noise level, rather
than just from cycling the learning rate. To explore why this helps generalization, we run VGG11 on CIFAR10 using 4 training schedules: we compared two discrete schedules (where either η
or S switches discretely from one value to another between epochs) and two baseline schedules,
one constant (η is constant) and one triangle (η is interpolated linearly between its maximum and
minimum value). We track the norm of the Hessian and the training loss throughout training. Each
experiment is repeated six times. For each schedule we optimize η in [1e − 3, 5e − 2] and cycle
length in {5, 10, 15} on a validation set. In all cyclical schedules the maximum value (of η or S) is
5× larger than the minimum value.
First, we observe that cyclical schemes oscillate between sharp and wide regions of the parameter
space, see Fig. 6. Next, we empirically demonstrate that a discrete schedule varying either S or η
performs similarly, or slightly better than triangular CLR schedule, see Tab. 1. Finally, we observe
that cyclical schemes reach wider minima at the same loss value, see Fig. 6 and Tab. 1. All of the
above suggest that by changing noise levels cyclical schemes reach different endpoints than constant
learning rate schemes with same final noise level. We leave the exploration of the implications and
a more thorough comparison with other learning schedules for future work.
9
90
70
= 1 (83.0±0.15)
= 3 (82.7±0.07)
= 5 (82.5±0.29)
= 7 (82.1±0.06)
= 9 (81.4±0.27)
= 11 (81.0±0.3)
= 13 (69.3±7.83)
= 15 (77.5±1.05)*
60
50
40
30
20
10
0
25
50
75
100
epoch
125
150
175
validation accuracy
validation accuracy
80
80
70
50
40
30
20
10
200
= 1 (87.1±0.13)
= 3 (87.0±0.2)
= 5 (86.6±0.15)
= 7 (86.4±0.14)
= 9 (86.3±0.17)
= 11 (77.8±13.9)
= 13 (85.8±0.01)*
= 15 (80.4±3.09)*
60
0
(a) Train dataset size 12000
50
75
100
epoch
125
150
175
200
(b) Train dataset size 22500
90
80
70
= 1 (90.4±0.12)
= 3 (90.3±0.12)
= 5 (90.3±0.14)
= 7 (90.2±0.09)
= 9 (90.0±0.11)
= 11 (89.7±0.22)
= 13 (89.1±0.42)
= 15 (88.6±0.0)*
60
50
40
30
20
10
0
25
50
75
100
epoch
125
150
175
validation accuracy
90
validation accuracy
25
80
70
50
40
30
20
10
200
= 1 (89.8±0.09)
= 5 (89.6±0.14)
= 9 (89.5±0.11)
= 13 (89.3±0.11)
= 17 (89.2±0.1)
= 21 (88.9±0.16)
= 25 (85.4±5.43)
= 29 (85.1±1.01)*
60
(c) Train dataset size 45000
0
25
50
75
100
epoch
125
150
175
200
(d) Half the noise
Figure 7: Breaking point analysis: Our theory suggests the final performance should be similar when
β×η
the SGD noise level β×S
is kept the same. Here we study its breaking point in terms of too large
a learning rate or too small a batch size. (a,b,c)- Validation accuracy for different dataset sizes and
different β values for a VGG-11 architecture trained on CIFAR10. In each experiment, we multiply
the learning rate (η) and batch size (S) with β such that the ratio β×(η=0.1)
β×(S=50) is fixed. We observe
that for the same ratio, increasing the learning rate and batch size yields similar performance up
to a certain β value, for which the performance drops significantly. (d)- Breaking point analysis
when half the noise level β×(η=0.05)
β×(S=50) is used. The breaking point happens for much larger β when
using a smaller noise. All experiments are repeated 5 times with different random seeds. The graphs
denote the mean validation accuracies and the numbers in the brackets denote the mean and standard
deviation of the maximum validation accuracy across different runs. The * denotes at least one seed
lead to divergence.
4.5
B REAKING P OINT OF THE T HEORY IN P RACTICE
Our analysis relies on the assumption that the gradient step is sufficiently small to guarantee that
the first order approximation of a Taylor’s expansion is a good estimate of the loss function. In
the case where the learning rate becomes too high, this approximation is no longer suitable, and
the continuous limit of the discrete SGD update equation will no longer be valid. In this case, the
stochastic differential equation doesn’t hold, and hence neither does the Fokker-Planck equation,
and so we don’t expect our theory to be valid. In particular, we don’t expect to arrive at the same
stationary distribution as indicated by a fixed ratio η/S, if the learning rate gets too high.
This is exemplified by the empirical results reported in Fig. 7, where similar learning dynamics
and final performance can be observed when simultaneously multiplying the learning rate and batch
size by a factor β up to a certain limit. This is done for different training set sizes to investigate if
the breaking point depends on this factor (Fig. 7 a,b,c). The plots suggest that the breaking point
happens for smaller β values if the dataset size is smaller. We also investigate the influence of β
10
when half the noise level is used, due to halving the learning rate, in (figure 7 d). These experiments
strongly suggest that the reason behind breaking point is the use of a high learning rate because the
performance drops at much higher β when the base learning rate is halved. A similar experiment is
performed on Resnets (for results see Fig 8 in the appendix). We highlight other limitations of our
theory in appendix E.
5
D ISCUSSION
In the theoretical section of this work we treat the learning rate as fixed throughout training. However, in practical applications, the learning rate is annealed to a lower value, either gradually or in
discrete jumps. When viewed within our framework, at the beginning with high noise, SGD favors
width over depth of a region, then as the noise decreases, SGD prioritizes the depth more strongly this can be seen from Theorem 3 and the comments that follow.
In the theoretical section we made the additional assumption that the covariance of the gradients
is isotropic, in order to be able to derive a closed form solution for the equilibrium distribution.
We do not expect this assumption to hold in practice, but speculate that there may be mechanisms
which drive the covariance towards isotropy, for example one may be able to tune learning rates
on a per-parameter basis in such a way that the combination of learning rate and covariance matrix
is approximately isotropic – this may lead to improvements in optimization. Perhaps some existing mechanisms such as batch normalization or careful initialization give rise to more equalized
covariance - we leave study of this for future work.
We note further that our theoretical analysis considered an equilibrium distribution, which was independent of the intermediate dynamics. However, this may not be the case in practice. Without the
isotropic covariance, the system of partial differential equations in the late time limit will in general have a solution which will depend on the path through which optimization occurs, unless other
restrictive assumptions are made to force this path dependence to disappear (Chaudhari & Soatto,
2017). Despite this simplifying assumption, our empirical results are consistent with the developed
theory. We leave study of path dependence and dynamics to future work.
In experiments investigating memorization we explored how the noise level changes the preference
of wide minima over sharp ones. (Arpit et al., 2017) argues that SGD first learns true labels, before
focusing on random labels. Our insight is that in the second phase the high level of noise maintains generalization. This illustrates the trade-off between width of minima and depth in practice.
When the noise level is lower, DNNs are more likely to fit random labels better, at the expense of
generalizing less well on true ones.
6
C ONCLUSIONS
We shed light on the role of noise in SGD optimization of DNNs and argue that three factors (batch
size, learning rate and gradient variance) strongly influence the properties (loss and width) of the
final minima at which SGD converges. The learning rate and batch size of SGD can be viewed as
one effective hyper-parameter acting as a noise factor n = η/S. This, together with the gradient
covariance influences the trade-off between the loss and width of the final minima. Specifically,
higher noise favors wider minima, which in turn correlates with better generalization.
Further, we experimentally verify that the noise n = η/S determines the width and height of the
minima towards which SGD converges. We also show the impact of this noise on the memorization
phenomenon. We discuss the limitations of the theory in practice, exemplified by when the learning
rate gets too large. We also experimentally verify that η and S can be simultaneously rescaled as
long as the noise η/S remains the same.
ACKNOWLEDGMENTS
We thank Jason Jo, and Nicolas Le Roux, Mike Rabbat, Leon Bottou, and James Griffin for helpful
discussions. SJ was supported by Grant No. DI 2014/016644 from Ministry of Science and Higher
Education, Poland and ETIUDA stipend No. 2017/24/T/ST6/00487 from National Science Centre,
Poland. ZK was supported in part by an EA Grant from the Centre for Effective Altruism. We
11
acknowledge the computing resources provided by ComputeCanada and CalculQuebec. We made
use of the following frameworks TensorFlow (Abadi et al., 2015), Theano (Theano Development
Team, 2016) and PyTorch.
R EFERENCES
Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S.
Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew
Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath
Kudlur, Josh Levenberg, Dan Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah,
Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning
on heterogeneous systems, 2015. URL https://www.tensorflow.org/. Software available from tensorflow.org.
Madhu S Advani and Andrew M Saxe. High-dimensional dynamics of generalization error in neural
networks. arXiv preprint arXiv:1710.03667, 2017.
Dario Amodei, Sundaram Ananthanarayanan, Rishita Anubhai, Jingliang Bai, Eric Battenberg, Carl
Case, Jared Casper, Bryan Catanzaro, Qiang Cheng, Guoliang Chen, et al. Deep speech 2: Endto-end speech recognition in english and mandarin. In International Conference on Machine
Learning, pp. 173–182, 2016.
Devansh Arpit, Stanisław Jastrz˛ebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S. Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, and Simon
Lacoste-Julien. A closer look at memorization in deep networks. In Doina Precup and Yee Whye
Teh (eds.), Proceedings of the 34th International Conference on Machine Learning, volume 70
of Proceedings of Machine Learning Research, pp. 233–242, International Convention Centre,
Sydney, Australia, 06–11 Aug 2017. PMLR. URL http://proceedings.mlr.press/
v70/arpit17a.html.
Léon Bottou. Stochastic gradient learning in neural networks. Proceedings of Neuro-Nımes, 91(8),
1991.
Léon Bottou. Online learning and stochastic approximations. On-line learning in neural networks,
17(9):142, 1998.
Pratik Chaudhari and Stefano Soatto. Stochastic gradient descent performs variational inference,
converges to limit cycles for deep networks. arXiv preprint arXiv:1710.11029, 2017.
Pratik Chaudhari, Adam Oberman, Stanley Osher, Stefano Soatto, and Guillame Carlier. Deep
relaxation: partial differential equations for optimizing deep neural networks. arXiv preprint
arXiv:1704.04932, 2017.
T. Chen, E. B. Fox, and C. Guestrin. Stochastic gradient Hamiltonian Monte Carlo. In Proceedings
of the 31st International Conference on Machine Learning, pp. 1683–1691, 2014.
N. Ding, Y. Fang, R. Babbush, C. Chen, R. D. Skeel, and H. Neven. Bayesian sampling using
stochastic gradient thermostats. In Advances in Neural Information Processing Systems 27, pp.
3203–3211, 2014.
L. Dinh, R. Pascanu, S. Bengio, and Y. Bengio. Sharp Minima Can Generalize For Deep Nets. ArXiv
e-prints, March 2017.
C. Gardiner. Stochastic Methods: A Handbook for the Natural and Social Sciences. Springer
Series in Synergetics. Springer Berlin Heidelberg, 2010. ISBN 9783642089626. URL https:
//books.google.ca/books?id=321EuQAACAAJ.
Ian J Goodfellow, Oriol Vinyals, and Andrew M Saxe. Qualitatively characterizing neural network
optimization problems. arXiv preprint arXiv:1412.6544, 2014.
12
P. Goyal, P. Dollár, R. Girshick, P. Noordhuis, L. Wesolowski, A. Kyrola, A. Tulloch, Y. Jia, and
K. He. Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour. ArXiv e-prints, June 2017.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.
770–778, 2016.
Sepp Hochreiter and Jürgen Schmidhuber. Flat minima. Neural Computation, 9(1):1–42, 1997.
E. Hoffer, I. Hubara, and D. Soudry. Train longer, generalize better: closing the generalization gap
in large batch training of neural networks. ArXiv e-prints, May 2017.
Alex Krizhevsky. One weird trick for parallelizing convolutional neural networks. arXiv preprint
arXiv:1404.5997, 2014.
Qianxiao Li, Cheng Tai, and Weinan E. Stochastic modified equations and adaptive stochastic gradient algorithms. In Doina Precup and Yee Whye Teh (eds.), Proceedings of the 34th International
Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pp.
2101–2110, International Convention Centre, Sydney, Australia, 06–11 Aug 2017. PMLR. URL
http://proceedings.mlr.press/v70/li17f.html.
David JC MacKay. Bayesian methods for adaptive models. PhD thesis, California Institute of
Technology, 1992.
S. Mandt, M. D. Hoffman, and D. M. Blei. Stochastic Gradient Descent as Approximate Bayesian
Inference. ArXiv e-prints, April 2017.
A. Neelakantan, L. Vilnis, Q. V. Le, I. Sutskever, L. Kaiser, K. Kurach, and J. Martens. Adding
Gradient Noise Improves Learning for Very Deep Networks. ArXiv e-prints, November 2015.
Nicolas L. Roux, Pierre antoine Manzagol, and Yoshua Bengio.
Topmoumoute online natural gradient algorithm.
In J. C. Platt, D. Koller, Y. Singer, and S. T.
Roweis (eds.), Advances in Neural Information Processing Systems 20, pp. 849–
856. Curran Associates, Inc., 2008.
URL http://papers.nips.cc/paper/
3234-topmoumoute-online-natural-gradient-algorithm.pdf.
Issei Sato and Hiroshi Nakagawa. Approximation analysis of stochastic gradient langevin dynamics by using fokker-planck equation and ito process. In Eric P. Xing and Tony Jebara (eds.),
Proceedings of the 31st International Conference on Machine Learning, volume 32 of Proceedings of Machine Learning Research, pp. 982–990, Bejing, China, 22–24 Jun 2014. PMLR. URL
http://proceedings.mlr.press/v32/satoa14.html.
Xiaocheng Shang, Zhanxing Zhu, Benedict Leimkuhler, and Amos J Storkey. Covariance-controlled
adaptive Langevin thermostat for large-scale Bayesian sampling. In C. Cortes, N. D. Lawrence,
D. D. Lee, M. Sugiyama, and R. Garnett (eds.), Advances in Neural Information Processing
Systems 28, pp. 37–45, 2015.
N. Shirish Keskar, D. Mudigere, J. Nocedal, M. Smelyanskiy, and P. T. P. Tang. On Large-Batch
Training for Deep Learning: Generalization Gap and Sharp Minima. ArXiv e-prints, September
2016.
Ravid Shwartz-Ziv and Naftali Tishby. Opening the black box of deep neural networks via information. CoRR, abs/1703.00810, 2017. URL http://arxiv.org/abs/1703.00810.
Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image
recognition. arXiv preprint arXiv:1409.1556, 2014.
L. N. Smith. Cyclical Learning Rates for Training Neural Networks. ArXiv e-prints, June 2015.
Samuel L Smith and Quoc V Le. Understanding generalization and stochastic gradient descent.
arXiv preprint arXiv:1710.06451, 2017.
Theano Development Team. Theano: A Python framework for fast computation of mathematical
expressions. arXiv e-prints, abs/1605.02688, May 2016. URL http://arxiv.org/abs/
1605.02688.
13
S. J. Vollmer, K. C. Zygalakis, and Y. W. Teh. (Non-) asymptotic properties of stochastic gradient
Langevin dynamics. arXiv preprint arXiv:1501.00438, 2015.
M. Welling and Y. W. Teh. Bayesian learning via stochastic gradient Langevin dynamics. In Proceedings of the 28th International Conference on Machine Learning, pp. 681–688, 2011.
Lei Wu, Zhanxing Zhu, et al. Towards understanding generalization of deep learning: Perspective
of loss landscapes. arXiv preprint arXiv:1706.10239, 2017.
H. Xiao, K. Rasul, and R. Vollgraf. Fashion-MNIST: a Novel Image Dataset for Benchmarking
Machine Learning Algorithms. ArXiv e-prints, August 2017.
Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding
deep learning requires rethinking generalization. arXiv preprint arXiv:1611.03530, 2016.
A PPENDIX
A
D ERIVATION OF THE F OKKER -P LANCK E QUATION
In this appendix we derive the Fokker-Planck equation. For any stochastic differential equation,
since the evolution is noisy we can’t say exactly where in parameter space the parameter values
will be at any given time. But we can talk about the probability P (θ, t|θ 0 , t0 ) that a parameter
takes a certain value θ at a certain time t, given that it started at θ 0 , t0 . That is captured by the
Fokker-Planck equation, which reads
∂P (θ, t)
η2
= ∇θ · ηg(θ)P (θ, t) +
∇θ · [C(θ)P (θ, t)] .
(5)
∂t
2S
In this appendix we will derive the above equation from the stochastic differential equation (6).
We will not be interested in pure mathematical rigour, but intend the proof to add intuition for the
machine learning audience.
For brevity we will sometimes write the probability just as P (θ, t). We will sometimes make use
of tensor index notation, where a tensor is denoted by its components, (for example θi are the
components of the vector θ), and we will use the index summation convention, where a repeated
index is to be summed over.
We start with the stochastic differential equation
dθ
η
= −ηg(θ) + √ B(θ)f (t).
dt
S
(6)
A formal expression for the probability for a given noise function f is given by P (θ, t) = δ(θ − θ f ),
but since we don’t know the noise, we instead average over all possible noise functions to get
P (θ, t) = E[δ(θ − θ f )].
(7)
While this is just a formal solution we will make use of it later in the derivation.
We now consider a small step in time δt, working at first order in δt, and ask how far the parameters
move, denoted δθ, which is given by integrating (6)
Z t+δt
η
δθ = −ηgδt + √ B
f (t0 )dt0
(8)
S
t
where we’ve assumed that δθ is small enough that we can evaluate g at the original θ. We now look
at expectations of δθ. Using the fact that the noise is normalized Gaussian, E(f (t)) = 0, we get
(switching to index notation for clarity)
E(δθi ) = −ηgi δt
(9)
14
and using that the noise is normalized Gaussian again, we have that E(f (t)f (t0 )) = Iδ(t − t0 ),
leading to
η2
Cij δt + O(δt2 ).
(10)
S
If at time t we are at position θ 0 and end up at position θ = θ 0 + δθ at time t + δt, then we can take
(7) with θ f = θ 0 and Taylor expand it in δθi
!
∂
1
∂2
0
2
P (θ, t + δt|θ , t) = 1 + E(δθi ) 0 + E(δθi δθj ) 0 0 + O(δt ) δ(θ − θ 0 ). (11)
∂θi
2
∂θi ∂θj
E(δθi δθj ) =
We deal with the derivatives of the delta functions in the following way. We will use the following
identity, called the Chapman-Kolmogorov equation, which reads
Z +∞
dθ 0 P (θ, t + δt|θ 0 , t0 )P (θ 0 , t0 |θ 0 , t0 )
(12)
P (θ, t + δt|θ 0 , t0 ) =
−∞
0
0
for any t , such that t0 ≤ t ≤ t + δt. This identity is an integral version of the chain rule of
probability, stating there are multiple paths of getting from an initial position θ 0 to θ at time t + δt
and one should sum all of these paths.
We will now substitute (11) into the first term on the right hand side of (12) with t0 set to be t,
and apply integration by parts (assuming vanishing boundary conditions at infinity) to move the
derivatives off of the delta functions and onto the other terms. We end up with
∂
(E(δθi )P (θ, t|θ 0 , t0 ))
∂θi
1 ∂2
(E(δθi δθj )P (θ, t|θ 0 , t0 )) .
+
2 ∂θi0 ∂θj0
P (θ, t + δt|θ 0 , t) = P (θ, t|θ 0 , t0 ) −
(13)
We can then take the first term on the right hand side of (13) to the other side, insert (9) and (10),
divide by δt and take the limit δt → 0, getting a partial derivative with respect to time on the left
hand side, leading directly to the Fokker-Planck equation quoted in the text (5) (where we have
reverted back to vector notation from index notation for conciseness)
η2
∂P (θ, t)
= ∇θ · ηg(θ)P (θ, t) +
∇θ · [C(θ)P (θ, t)] .
(14)
∂t
2S
B
I NTUITIONS FROM F OKKER -P LANCK
In this appendix we will give some supplementary comments about the intuition we can gain from
the Fokker-Planck equation (5). If the learning rate, batch size are constant and the covariance is
proportional to the identity C = σ 2 I and σ 2 is constant, then we can rewrite the Fokker-Planck
equation in the following form
∂P (θ, tη )
ησ 2 2
= ∇θ · [g(θ)P (θ, tη )] +
∇ P (θ, tη ) ,
∂tη
2S θ
(15)
where we have rescaled the time coordinate to be tη ≡ tη. One can now see that in terms of this
rescaled time coordinate, the ratio between the drift and diffusion terms is governed by the following
ratio
ησ 2
.
(16)
2S
In terms of the balance between drift and diffusion, we see that a higher value of this ratio gives
rise to a more diffusive evolution, while a lower value allows the potential drift term to dominate.
In the next section we will see how this ratio controls the stationary distribution at which SGD
converges to. For now we highlight that in terms of this rescaled coordinate, only this ratio controls
the evolution towards this stationary distribution (not just its endpoint), in terms of the rescaled time
tη . That is, learning rate and batch size are interchangable in the sense that the ratio is invariant
15
under transformations S → aS, η → aη, for a > 0. But note that the time it takes to reach the
stationary distribution depends on η as well, because of the rescaled time variable. For example,
for a higher learning rate, but constant ratio, one arrives at the same stationary distribution, but in a
quicker time by a factor of 1/η.
However, a caution here is necessary. The first order SGD update equation only holds for small
enough η that a first order approximation to a Taylor expansion is valid, and hence we expect the
first order approximation of SGD as a continuous stochastic differential equation to break down for
high η. Thus, we expect learning rate and batch size to be interchangable up to a maximum value of
η at which the approximation breaks.
C
S TATIONARY D ISTRIBUTION
In this appendix we will prove equation (2) quoted in the main text, which claims that for an isotropic
covariance, C(θ) = σ 2 I, with σ constant, the equilibrium solution of the Fokker-Planck equation
has the following form
2L(θ)
P (θ) = P0 exp −
,
(17)
nσ 2
where n ≡ Sη and P0 is a normalization constant, which is well defined for loss functions with L2
regularization.
In order to prove this is the equilibrium distribution, we need to solve the Fokker-Planck equation
(5) with the left hand side set equal to zero (which would give a stationary distribution) and further
we require for equilibrium that detailed balance holds. To do this, we begin by writing the FokkerPlanck equation in a slightly different form, making use of a probability current J , defined
η2
J ≡ ηg(θ)P (θ, t) +
∇θ · (C(θ)P (θ, t))
(18)
2S
in which the Fokker-Planck equation becomes
∂P (θ, t)
= ∇θ · J .
(19)
∂t
2
At this point we use the assumption that C(θ) = σ I to get
η2 σ2
J = ηg(θ)P (θ, t) +
∇θ P (θ, t).
(20)
2S
(θ,t)
The stationary solution has ∂P∂t
= 0 and hence ∇θ · J = 0. But we require the equilibrium
solution, which is a stronger demand than just the stationary solution. The equuilibrium solution
is a particular stationary solution in which detailed balance occurs. Detailed balance means that in
the stationary solution, each individual transition balances precisely with its time reverse, resulting
in zero probability currents, see §5.3.5 of (Gardiner, 2010), i.e. that J = 0. Detailed balance is a
sufficient condition for having entropy increasing with time. Non-zero J with ∇θ · J = 0 would
correspond to a non-equilibrium stationary distribution, which we don’t consider here.
For the equilibrium solution, J = 0, we get the result for the probability distribution
2S
(21)
P (θ) = P0 exp − 2 L(θ) .
ησ
which is the desired stationary solution we intended to find. Finally, to ensure that P0 is a finite
value, we note that the loss function L(θ) can be decomposed as,
L(θ) = L0 (θ) + τ kθk2
(22)
where τ is some non-negative constant controlling L2 regularization. Then we see that,
Z
2S
P0 =
exp − 2 L(θ)
(23)
ησ
θ
Z
2S
(24)
=
exp − 2 L0 (θ) − τ kθk2
ησ
θ
Z
2S
(25)
=
exp − 2 L0 (θ) exp −τ kθk2
ησ
θ
16
2S
2S
Since L0 (θ) ≥ 0 and ησ
− ησ
≤ 1, thus,
2 > 0, we have that exp
2 L0 (θ)
Z
Z
2S
exp −τ kθk2
(26)
exp − 2 L0 (θ) exp −τ kθk2 ≤
ησ
θ
θ
and we now note that the right hand side of (26) is finite because it is just a multidimensional
Gaussian integral. Thus, P0 has a finite value and hence the stationary distribution P (θ) is well
defined.
D
D ERIVATION OF P ROBABILITY OF E NDING IN A GIVEN M INIMA
In this appendix we derive the discrete set of probabilities of ending at each minima, as given in (3).
Essentially, we will use Laplace’s method to approximate the integral of the probability density in
the region near a minimum. We work locally near θ A and take the following approximation of the
loss function, since it is near a minimum,
1
L(θ) ≈ LA + (θ − θ A )> HA (θ − θ A ).
(27)
2
where HA is the Hessian, and is positive definite for a minimum.
The distribution PA (θ) is a probability density, while we are interested in the discrete set of probabilities of ending at a given minimum, which we will denote by lowercase pA . To calculate this
discrete set of probabilities, for each minimum we need to integrate this stationary distribution over
an interval containing the minimum.
Integrating (17) in some region RA around θ A and using (27)
Z
2S
p A ≈ P0
(28)
exp − 2 L(θ)
ησ
RA
Z
2S
1
≈ P0
(29)
exp − 2 LA + (θ − θ A )> HA (θ − θ A )
ησ
2
RA
r
2SLA
1
≈ P̃0 exp −
(30)
ησ 2
det HA
where in the last line we assume the region is large enough that an approximation to the full Gaussian integral can be used (i.e. that the tails don’t contribute, which is fine as long as the region is
sufficiently larger than det HA ) – note that the region can’t be too large, otherwise we would invalidate our local assumption. The picture is that the minima are sufficiently far apart that the region
can be taken sufficiently large for this approximation to be valid. Note that P̃0 is different to P0 and
includes the normalization factors from performing the Gaussian integral, which are the same for all
minima.
Since we are interested in relative probabilities between different minima, so we can consider the
unnormalized probability, dropping factors that are common amongst all minima we get
r
1
2SLA
p̃A = exp −
(31)
ησ 2
det HA
This is the required expression given in the main text.
We note that the derivation used above talks about strict minima, i.e., minima with a positive definite
Hessian. In practice however, for deep neural networks with a large number of parameters, it is
unrealistic to expect the endpoint of training to be in a strict minimum. Instead, it is more likely
to be a point at which the Hessian has positive eigenvalues in a few directions, while the other
eigenvalues are (approximately) zero. In such cases, to understand which minima SGD favors, we
can consider the fact that at equilibrium, the iterate of SGD follows the distribution,
2L(θ)
P (θ) = P0 exp −
(32)
nσ 2
By definition this means that at any time during equilibrium, the iterate is more likely to be found in
a region of higher probability. In the restrictive case of strict minimum, we model it as a Gaussian
and characterize the probability of landing in a minimum depending on the curvature and depth of
the loss around that minimum. In the general case of minima with degenerate (flat) directions, we
can say that a minimum with more volume is a more probable one.
17
Valid Acc
80
60
baseline
times 2
times 3
times 4
times 5
times 6
times 7
times 8
times 9
40
20
0
50
100
150
Epoch
200
250
300
Figure 8: Experiments involving Resnet56 architecture on CIFAR10 dataset. In each curve, we
multiply the Sη ratio by a given factor (increasing both batch size and learning rate). We observe that
multiplying the ratio by a factor up to 5 results in similar performances. However, the performances
degrades for factor superior to 5.
E
L IMITATIONS OF OUR A NALYSIS
Due to our assumptions, we expect the theory to become unreliable when the discrete to continuous
approximation fails, when the covariance in gradients is non-isotropic, when the batch size becomes
comparable to the finite size of the training set and when momentum is considered.
We discussed the limits of the discrete to continous approximation in 4.5 (further illustrated in 8).
When the covariance in gradients is highly non-isotropic, the equilibrium solution to the FokkerPlanck equation is the solution to a complicated partial differential equation, and one can’t easily
spot the solution by inspection as we do in Appendix C. We expect this approximation to break
down especially in the case of complicated architectures where different gradient directions will be
have very different gradient covariances.
Our theory does not involve the finite size of the training set, which is a drawback of the theory. This
may be especially apparent when the batch size becomes large, compared to the training set size,
and we expect the theory to break down at this point.
Finally, we mention that momentum is used in practical deep learning optimization algorithms. Our
theory does not consider momentum, which is a drawback of the theory and we expect it to break
down in models in which momentum is important. We can write a Langevin equation for the case
of momentum, with momentum damping coefficient µ,
dv
ηB
= −µv − ηg + √ f (t)
dt
S
where the velocity is v = dθ/dt.
The Fokker-Planck equation corresponding to this Langevin equation with momentum is
∂
µ
η2 σ2 2
+ v · ∇θ P = η∇θ ·
vP + gP +
∇ P
∂t
η
2S θ
where P is now a function of velocity v as well as θ and t, and we’ve again assumed the gradient
covariance varies slowly compared to the loss.
From this more complicated Fokker-Planck equation, it is hard to spot the equilibrium distribution.
We leave the study of this for further work. For now, we just note that a factor of η can be taken
from the right hand side, and then the ratio between the diffusion and drift terms will again be the
ratio ησ 2 /(2S).
18
92
91
Validation Accuracy
90
89
88
87
86
85
84
10
4
LR/BS
10
3
Figure 9: Validation accuracy of Resnet56 networks against different ratios learning rate to batch
size on CIFAR10. Trained with SGD.
89%
8E-02
88%
87%
4E-02
86%
85%
0E+00
0E+00
4E-04
(a) Correlation of
Hessian Norm.
/S
η
S
8E-04
89%
89%
88%
88%
87%
87%
86%
86%
85%
Validation performance
Validation performance
Log. Hessian norm
1E-01
0E+001E-04 2E-04 3E-04 4E-04 5E-04 6E-04
/S
η
S
0E+00 2E-04 4E-04 6E-04 8E-04 1E-03
with Log (b) Correlation of with valida- (c) Correlation of
tion accuracy for bad initializa- tion accuracy.
tion.
/S
η
S
with valida-
Figure 10: Impact on SGD with ratio of learning rate η and batch size S for 20 layer ReLU MLP
without Batch Normalization on FashionMNIST.
F
F.1
OTHER E MPIRICAL R ESULTS
F URTHER I MPACTS OF SGD ON M INIMA
In this appendix we look at other experiments exploring the correlation between learning rate to
batch-size ratio and sharpness of minima, and to validation performance. In Figure 9, we report the
validation accuracy for Resnet56 models trained on CIFAR10 with different learning rate to batchsize ratio. We notice there is a peak validation accuracy at a learning rate to batch-size ratio of
around 2 × 10−3 . In particular, this example emphasizes that higher learning rate to batch-size ratio
doesn’t necessarily lead to higher validation accuracy. Instead, it acts as a control on the validation
accuracy, and there is an optimal learning rate to batch-size ratio.
Figures 10a and 10c show the results of a variant of 4 layer ReLU experiment, where we use 20
layers and remove Batch Normalization. The inspiration is to test predictions of theory in the more
challenging setup. We again observe a correlation of Hessian norm with learning rate to batch-size
ratio (though weaker), and similarly between validation performance and learning rate to batch-size
ratio.
F.2
M ORE L EARNING R ATE & BATCH S IZE E XCHANGEABILITY
In this appendix we show in more detail the experiments of Section 4.2 which show the exchangeability of learning rate and batch size. In Fig. 11 we show the log cross entropy loss, the epochaveraged log cross entropy loss, train, validation and test accuracy as well as the batch size schedule,
learning rate schedule. We see in the learning rate to batch-size ratio plot that the orange and blue
lines (cyclic schedules) have the same ratio of learning rate to batch-size as each other throughout,
and that their dynamics are very similar to each other. The same holds for the green and red lines, of
constant batch size and learning rate. This supports the theoretical result of this paper, that the ratio
S/η governs the stationary distribution of SGD.
19
Figure 11: We show in more detail exchangeability of batch size and learning rate in a one-to-one
ratio. In blue, cyclic batch size schedule between size 128 and 640 and fixed learning rate 0.005,
is exchangeable with orange cyclic learning rate schedule between learning rates 0.001 and 0.005
with fixed batch size 128. In green, constant batch size 640 and constant learning rate 0.005 is
exchangeable with, in red, constant batch size 128 and constant learning rate 0.005.
We note that it is not just in the stationary distribution that the exchangeability holds in these plots - it
appears throughout training, highlighted especially in the cyclic schedules. We postulate that due to
the scaling relation in the Fokker-Planck equation, the exchangeability holds throughout learning as
well as just at the end, as long as the learning rate does not get so high as to ruin the approximations
under which the Fokker-Planck equation holds. We note similar behaviour occurs also for standard
learning rate annealing schedules, which we omit here for brevity.
20
25.0% random labels
Accuracy
100%
100%
80%
80%
60%
60%
40%
40%
20%
20%
0 50 100 150 200 250 300
50.0% random labels
100%
Accuracy
25.0% random labels
0 50 100 150 200 250 300
100%
80%
80%
60%
60%
40%
40%
20%
20%
0 50 100 150 200 250 300
Epoch
50.0% random labels
0 50 100 150 200 250 300
Epoch
Figure 12: Learning curves for memorization experiment with momentum 0.0. Solid lines represent
training accuracy, dotted validation accuracy. Warm color indicates higher Sη ratio.
6.8E+00
Log. Hessian norm
6.6E+00
6.4E+00
6.2E+00
6.0E+00
5.8E+00
5.6E+00
0.0E+00
1.0E-03
/S
0.0E+00
2.0E-03
1.0E-03
/S
2.0E-03
Figure 13: Correlation between (approximate) norm of Hessian of best validation minima and learning rate to batch-size ratio for 0.0 (left) and 0.9 (right) momentum.
F.3
D ETAILS ON MEMORIZATION EXPERIMENT
We report learning curves from memorization experiment with 0.0 momentum, see Fig. 12. We
additionally confirm similar results to the previous experiments and show correlation between batch
size and learning rate ratio and norm of Hessian, see Fig. 13 without momentum (left) and with
momentum 0.9 (right).
21
| 1cs.CV
|
arXiv:1705.06057v1 [cs.CV] 17 May 2017
Joint Learning from Earth Observation and OpenStreetMap Data
to Get Faster Better Semantic Maps
Nicolas Audebert, Bertrand Le Saux
ONERA, The French Aerospace Lab
F-91761 Palaiseau, France
Sébastien Lefèvre
Univ. Bretagne-Sud, UMR 6074, IRISA
F-56000 Vannes, France
(nicolas.audebert,bertrand.le_saux)@onera.fr
[email protected]
Abstract
In this work, we investigate the use of OpenStreetMap
data for semantic labeling of Earth Observation images.
Deep neural networks have been used in the past for remote sensing data classification from various sensors, including multispectral, hyperspectral, SAR and LiDAR data.
While OpenStreetMap has already been used as ground
truth data for training such networks, this abundant data
source remains rarely exploited as an input information
layer. In this paper, we study different use cases and deep
network architectures to leverage OpenStreetMap data for
semantic labeling of aerial and satellite images. Especially, we look into fusion based architectures and coarseto-fine segmentation to include the OpenStreetMap layer
into multispectral-based deep fully convolutional networks.
We illustrate how these methods can be successfully used
on two public datasets: ISPRS Potsdam and DFC2017. We
show that OpenStreetMap data can efficiently be integrated
into the vision-based deep learning models and that it significantly improves both the accuracy performance and the
convergence speed of the networks.
1. Introduction
Dense labeling of remote sensing data is a common task
to generate semantic maps of large areas automatically, either to perform cartography of urban areas or to determine
land use covers at a large scale. Many techniques that originated in the computer vision community for semantic segmentation can be applied on Red-Green-Blue (RGB) remote sensing data to this end and has been successfully used
with state-of-the-art results, e.g. Convolutional Neural Networks (CNN) for land use classification [4] or Fully Convolutional Networks (FCN) for semantic labeling of urban
areas [1]. Yet, remote sensing data is rarely limited to RGB
channels. Often, multiple sensors are used to obtain complementary physical information about the observed area,
with the final products being co-registered multispectral,
LiDAR and SAR data. Several works successfully investigated semantic labeling of remote sensing data using fusion of multiple heterogeneous sensors [1, 26, 16] to leverage complementary informations sources. However, there
is still a huge untapped source of information: the public
geo-information system (GIS) databases that rely on voluntary crowdsourced annotations. These databases, such as
the very popular OpenStreetMap (OSM), are an extremely
valuable source of information, that is mostly used as a
training ground truth [24] as it contains data layers such as
roads, buildings footprints or forest coverage. But this data
can also be seen as an alternative input layer that will provide another information channel compared to the classical
sensors. Consequently, we investigate here the following
question: how can we use OSM data to improve sensorbased semantic labeling of remote sensing data ?
2. Related work
Semantic labeling of aerial and satellite images using
deep neural networks has been investigated several times
in the literature. Since the early work on road extraction
using CNN [24], many studies investigated deep neural networks for automatic processing of aerial and satellite data.
The most recent works using deep learning mostly focus
on the use of Fully Convolutional Networks [20], an architecture initially designed for semantic segmentation of
multimedia images, and later successfully applied to remote sensing data at various resolutions. This type of
deep models obtained excellent results on very high resolution datasets [28], often combined with multiscale analysis [1], graphical model post-processing [33] and boundary prediction [22]. The same architectures were also successful for building and roads extraction from satellite images [21, 31, 34] at a significantly lower resolution.
Although most of the works are limited to optical data,
e.g. RGB or multispectral images, the fusion of heterogeneous data layers has also been investigated in several
works. Notably, multimodal processing of very high resolution images has been successfully applied to a combination
of deep features and superpixel classification in [3], while
a deep framework for the fusion of video and multispectral
data was proposed in [25]. Fusion of multitemporal data for
joint registration and change detection was investigated in
[32]. Heterogeneous data fusion was also explored through
deep features mixed with hand-crafted features for random
forest classification [26], and later using end-to-end deep
networks in [1] for LiDAR and RGB data, and [16] for hyperspectral and SAR data.
While all these works investigated data fusion of various sensors, they did not study the inclusion of highly
processed, semantically richer data such as OpenStreetMap
layers. Indeed, since the launch of OpenStreetMap (OSM)
and Google Maps in 2004, map data became widely available and have been largely used within remote sensing
applications. First, they can be used as targets for deep
learning algorithms, such as in the seminal work of Mnih
[24, 23], since those layers already provide accurate information about the buildings and roads footprints. More
recent works in simultaneous registration and classification [32] and large-scale classification [21, 6] also rely on
OSM data to perform supervised learning. The generation
of OpenStreetMap rasters from optical images with Generative Adversarial Networks has also been investigated [17],
but the authors did not evaluate their method with classification metrics as they were only interested by perceptually
coherent image-to-image translation. Second, the map layers can be used as inputs of a processing flow to produce
new geo-spatial data. Although the coverage and the quality of the annotations from open GIS vary a lot depending
on the users’ knowledge and number of contributors, this
data may contain relevant information for mapping specific
areas and classes. A deterministic framework to create land
use and land cover maps from crowdsourced maps such as
OSM data was proposed in [10]. Machine learning tools (a
random forest variant) also allow coupling remote sensing
and volunteered geographic information (VGI) to predict
natural hazard exposure [11] and local climate zones [8],
while active deep learning helps finding unlabeled objects
in OSM [5]. However, to the best of our knowledge, no
VGI has been used as an input (opposite of a target) in deep
neural networks yet.
3. Method
In this work, we investigate two use cases for highly
structured semantic data as input to deep networks. The first
one occurs when the semantic data is close to the ground
truth, in our case where OpenStreetMap already contains a
part of the information we want to label, e.g. finding roads
in airborne images. The second one occurs when what we
want to label can be inferred from OpenStreetMap but only
indirectly, for example when we want to derive the type of
settlement based on building footprints.
3.1. Coarse-to-fine segmentation
In some cases, OpenStreetMap data can be used as a
coarse approximation from the ground truth, e.g. for objects
such as buildings and roads. Therefore, we would need only
to learn how to refine this coarse map to obtain the final very
high resolution pixel-wise segmentation. Especially, learning only this difference can be seen as a form of residual
learning [15] which should help the training process. This
was also suggested in [19] to help the classification process
by first learning coarse approximations of the ground truth
before iteratively refining the predictions.
In our case, we can use a simple FCN with only two layers to convert the raw rasterized OSM data into a semantic
map that approximates the ground truth, trained with the
usual backpropagation technique. This FCN will be denoted OSMNet in the following sections. OSMNet only
manipulates the original OSM map to project it in a representation space compatible with the ground truth classes.
The optical data will then be used as input to a FCN model
to infer a semantic segmentation that will complete the approximate map derived from OSM. In this work, we will
use the SegNet architecture [2] that we will train following
the guidelines from [1]. SegNet is an encoder-decoder deep
neural network that projects an input image into a semantic map with the same resolution. Using these two models,
we can compute the average prediction computed using the
two data sources. In this framework, if we denote I the
input image, O the input OSM raster, Popt the prediction
function for SegNet and Posm the prediction function for
OSMNet, the final prediction P is:
P (I, O) =
1
(Popt (I) + Posm (O)) .
2
(1)
If Posm (O) is already an adequate approximation of the
ground truth GT , then the training process will try to minimize:
Popt ∝ GT − Posm (O) GT ,
(2)
which should have a similar effect on the training process
than the residual learning from [15].
Moreover, to refine even further this average map, we use
a residual correction network [1]. This module consists in a
residual three-layer FCN that learns how to correct the average prediction using the complementary information from
the OSM and optical data sources. If we denote C the prediction function of the residual correction network, we finally have:
P (I, O) =
1
(Popt (I)+Posm (O))+C(Zopt (I), Zosm (O)) ,
2
(3)
(a) Optical and OSM data fusion using residual correction [1].
(b) FuseNet [13] architecture applied to optical and OSM data.
Figure 1: Deep learning architectures for joint processing of optical and OpenStreetMap data.
where Zopt and Zosm are the last feature maps from SegNet
and OSMNet, respectively.
The residual learning using this module can be seen as
learning an error correction to refine and correct occasional
errors in the prediction. The full pipeline is illustrated in
the Fig. 1a. The whole architecture is trained jointly.
3.2. Dual-stream convolutional neural networks
FCN with several sources have been investigated several
times in the past, notably for processing RGB-D (or 2.5D)
images in the computer vision community [9]. In this work,
we use the FuseNet architecture [13] to combine optical and
OSM data. It is based on the popular SegNet [2] model.
FuseNet has two encoders, one for each source. After each
convolutional block, the activations maps from the second
encoder are summed into the activations maps from the first
encoder. This allows the two encoders to learn a joint representation of the two modalities. Then, a single decoder performs both the spatial upsampling and the pixel-wise classification. As detailed in Fig. 1b, one main branch learns
this joint summed representation while the ancillary branch
learns only OSM-dependent activations. If we denote P the
prediction function from FuseNet, I the input image, O the
{opt,osm}
input OSM rasters, Ei
the encoded feature maps
{opt,osm}
after the ith block, Bi
the encoding functions for
the ith block and D the decoding function:
P (I, O) = D(E5opt (I, O))
(4)
and
opt
Ei+1
(I, O) = Biopt (Eiopt (I, O))+Biosm (Eiosm (O)) . (5)
This architecture allows us to fuse both data streams in the
internal representation learnt by the network. This model is
illustrated in Fig. 1b.
4. Experiments
4.1. Datasets
ISPRS Potsdam The ISPRS Potsdam Semantic Labeling
dataset [27]1 is comprised of 38 ortho-rectified aerial IRRGB images (6000 × 6000 px) at 5cm spatial resolution,
taken over the city of Potsdam (Germany). A comprehensive pixel-level ground truth is provided for 24 tiles with
annotations for the following classes: “impervious surface”,
“building”, “low vegetation”, “tree”, “car” and “clutter”, as
illustrated in Fig. 2a. As the tiles are geo-referenced, we
download and generate the corresponding OpenStreetMap
rasters with the footprints for roads, buildings, vegetation
and water bodies using Maperitive2 . Results on this dataset
are cross-validated on a 3-fold train/test split (18 tiles for
training, 6 tiles for testing).
Data Fusion Contest 2017 The Data Fusion Contest
(DFC) 2017 dataset3 or “grss_dfc_2017” [12] is comprised
of various acquisitions over 8 different cities: Berlin, Hong
Kong, Paris, Rome and Sao Paulo for training, and Amsterdam, Chicago, Madrid an Xi’an for testing. It includes
multispectral Sentinel-2 and Landsat data, OpenStreetMap
rasters for the roads, vegetation, land use types, buildings
and water bodies and a sparse ground truth containing an1 http://www2.isprs.org/commissions/comm3/wg4/
semantic-labeling.html
2 http://maperitive.net/
3 http://www.grss-ieee.org/community/
technical-committees/data-fusion/
data-fusion-contest/
Potsdam (RGB)
Potsdam (GT)
(a) ISPRS Potsdam dataset
DFC 2017 (false RGB)
DFC 2017 (GT)
(b) DFC 2017 dataset (extract from São Paulo)
Figure 2: Extract of the ISPRS Potsdam and DFC2017 datasets
notations for several Local Climate Zones (LCZ), as illustrated in Fig. 2b. The LCZ define urban or rural areas such
as “sparsely built urban area”, “water body”, “dense trees”
and so on, using the taxonomy from [30]. The annotations
cover only a part of the cities and are provided at 100m/pixel
resolution. The goal is to generalize those annotations to
the testing cities. In this work, we use the multispectral
20m/pixel resolution Sentinel-2 data and the OSM raster for
roads, buildings, vegetation and water bodies. We preprocess the multispectral data by clipping the 2% highest values. All 13 bands are kept and stacked as input to the neural
network. As Sentinel-2 multispectral data includes bands
at 10m/pixel, 20m/pixel and 60m/pixel resolutions, bands
that have a resolution lower or higher than 20m/pixel are
rescaled using bilinear interpolation. Results on this dataset
are computed on the held-out testing set from the benchmark.
4.2. Experimental setup
We train our models on the ISPRS Potsdam dataset in an
end-to-end fashion, following the guidelines from [1]. We
randomly extract 128×128 patches from the RGB and OSM
training tiles on which we apply random flipping or mirroring as data augmentation. The optimization is performed
with a batch size of 10 on the RGB tiles using a Stochastic Gradient Descent (SGD) with a learning rate of 0.01
divided by 10 every 2 epochs (' 30 000 iterations). SegNet’s encoder for the RGB data is initialized using VGG16 [29] weights trained on ImageNet, while the decoder
is randomly initialized using the MSRA [14] scheme. The
learning rate for the encoder is set to half the learning rate
for the decoder. During testing, each tile is processed by
sliding a 128 × 128 window with a stride of 64 (i.e. 50%
overlap). Multiple predictions for overlapping regions are
averaged to smooth the map and reduce visible stitching on
the patch borders. Training until convergence (' 150,000
iterations) takes around 20 hours on a NVIDIA K20c, while
(a) Binary representation.
(b) Signed distance transform.
Figure 3: Representations of the OSM layer for roads.
evaluating on the validation set takes less than 30 minutes.
On the DFC2017, we re-train SegNet from scratch and
the weights are initialized using the MSRA scheme. As
the input data has a resolution of 20m/pixel and the output LCZ are expected to be at 100m/pixel resolution, we
use a smaller decoder by removing the last three convolutional blocks and the associated pooling layers. The resulting predictions have a resolution of 1:4 the input data
and are interpolated to the 100m/pixel resolution. We train
the network on randomly flipped 160 × 160 patches with a
50% overlap. The patches are randomly selected but have
to contain at least 5% of annotated pixels. To avoid learning
on non-labeled pixels from the sparse LCZ annotations, we
ignore the undefined pixels in the ground truth during loss
computation. The network is trained using the Adam [18]
optimizer with a learning rate of 0.01 with a batch size of
10. Training until convergence (' 60,000 iterations) takes
around 6 hours on a NVIDIA Titan Pascal, while evaluating
on the test set takes less than 5 minutes.
OSM data modelization has to be carefully investigated.
Most sensor data is continuous both in the numerical meaning but also in the spatial repartition. In many cases, if the
original data is not continuous but sparse, well-chosen representations are used to get the continuity back, e.g. the Digital Surface Model which is a continuous topology extracted
from the sparse LiDAR point cloud. In our case, the OSM
data is sparse and discrete like the labels. Therefore, it is
dubious if the deep network will be able to handle all the
information using such a representation. We compare two
representations, illustrated in Fig. 3:
• Sparse tensor representation, which is discrete. For
each raster, we have an associated channel in the tensor which is a binary map denoting the presence of the
raster class in the specified pixel (cf. Fig. 3a).
• Signed distance transform (SDT) representation,
which is continuous. We generate for each raster the
associated channel corresponding to the distance transform d, with d > 0 if the pixel is inside the class and
d < 0 if not (cf. Fig. 3b, with a color range from blue
to red).
The signed distance transform was used in [34] for building extraction in remote sensing data. The continuous representation helped densifying the sparse building footprints
that were extracted from a public GIS database and successfully improved the classification accuracy.
4.3. Results
OSM data representation As can be seen in Table 1,
learning and testing on the signed distance transform
(SDT) representation for the OpenStreetMap layers performs slightly worse than its binary counterpart. This might
seem counter-intuitive, as the distance transform contains a
denser information. However, we suggest that this information might be too diffuse and that the model loses the sharp
boundaries and clear transitions of the binary raster on some
parts of the dataset. Yet, the difference between the two representations does not impact strongly the final accuracy.
ISPRS Potsdam We report in Table 1 the results of our
methods on our validation set of the ISPRS Potsdam dataset.
In accordance with the dataset guidelines, we compare our
predictions with the ground-truth where the borders have
been eroded by a disk of radius 3 pixels. We report the
overall pixel-wise accuracy and the F1 score for each class,
computed as follows:
F 1i = 2
precisioni × recalli
,
precisioni + recalli
(6)
tpi
tpi
, precisioni =
,
(7)
Ci
Pi
where tpi the number of true positives for class i, Ci the
number of pixels belonging to class i, and Pi the number of
pixels attributed to class i by the model.
recalli =
As could be expected, the inclusion of OSM data improves the semantic labeling performance, with significant
improvements on “road” and “building” classes. This is not
surprising considering that those classes already have a representation in the OSM layers which can help disambiguating predictions coming from the optical source. This is illustrated qualitatively in Fig. 4 and quantitatively in Fig. 5.
Moreover, OSM data accelerates the training process as it
allows the main network to focus on the harder parts of the
image. Indeed, OSM data already covers the majority of the
roads and the buildings, therefore simplifying the inference
of the “impervious surface” and “building” classes. OSM
data also helps discriminating between buildings and roads
that have similar textures.
An interesting side effect of the integration of the OSM
data into the learning process is the significant speedup in
convergence time that can be observed. Indeed, on the
same dataset, the coarse-to-fine model converges approximately 25% faster to the same overall accuracy compared
to the classical RGB SegNet, i.e. the network requires 25%
less iterations to reach the same classification performance.
Moreover, this accuracy is reached with a mean loss of 0.45
for the latter, while the former has a mean loss of only 0.39,
which indicates a better generalization capability. This is
similar to the findings from [15] on residual learning. Finally, the inclusion of the OSM data helps regularizing spatially the predictions when the network is still in early training. Fig. 6 illustrates how the same patch, classified at several stages in the training, is visually better represented using both OSM and RGB sources compared to the RGB image only.
DFC2017 We report in Table 2 the detailed results of our
methods on the DFC2017 dataset. The evaluations metrics
are the overall accuracy (OA) and the pixel-wise accuracies for each class. The integration of the OSM layers is
mostly useful for two types of classes: the high density urban areas and the dense vegetation. Indeed, classes 1, 2
and 3 denote different sorts of compact urban areas. The
buildings footprints from OSM help detecting those classes
and estimating the local building density. For the vegetation, the relevant OSM annotations are in the natural terrain
layer. These annotations are mostly concentrated on forests
and large meadows, which belong to the classes 11 (“dense
trees”) and 14 (“low plants”). Consequently, those classes
obtain a significant increase in classification accuracy when
fusing OSM and multispectral data. A qualitative example
on the city of Chicago is shown in Fig. 7.
5. Discussion
The methods presented in this work allow us to improve
semantic labeling of aerial and satellite multispectral images thanks to the integration of several OSM rasters, no-
Table 1: Test results on the ISPRS Potsdam dataset (pixel-wise overall accuracy and F1 score per class).
OSM
Method
imp. surfaces
buildings
low veg.
trees
cars
Overall
Binary
∅
OSMNet
SegNet RGB
54.8
93.0
90.0
92.9
51.5
85.0
0.0
85.1
0.0
95.1
60.3
89.7
Binary
Residual Correction RGB+OSM
FuseNet RGB+OSM
93.9
95.3
92.8
95.9
85.1
86.3
85.2
85.1
95.8
96.8
90.6
92.3
SDT
Residual Correction RGB+OSM
FuseNet RGB+OSM
93.8
95.2
92.7
95.9
85.2
86.4
84.8
85.0
95.9
96.5
90.5
92.3
Table 2: Test results on the DFC2017 dataset (pixel-wise accuracies)
Urban
LCZ
Compact
1
2
Open
3
4
5
6
Rural
Misc.
Trees Vegetation Soils and water
buildings
7 8 9 10 11 12 13 14 15 16 17 OA
SegNet multispectral
34.7 25.4 8.6 19.7 14.6 17.5 0.0 62.3 0.0 1.0 66.9 4.3 13.1 62.5 0.0 0.0 89.2 41.7
FuseNet multispectral + OSM 34.3 39.1 26.0 16.7 6.2 37.1 0.0 45.2 9.2 0.0 83.4 1.8 0.0 80.2 1.4 0.0 87.3 46.5
tably the roads, buildings and vegetation land use. However,
OpenStreetMap data is much more exhaustive than such
layers and also contains specific information (e.g. swamps,
agriculture fields, industrial areas, different categories of
roads. . . ). However, if all information are stacked by using one map per layer of interest, the OSM memory footprint would become huge very quickly, especially considering that OSM provides vector information that can be rasterized at any spatial resolution. In our case, we rasterize
the OSM layers to the same resolution as our input image,
which can be very high for airborne acquisitions. Moreover,
we have not addressed here the question of the subclassification, while this is definitely a source of future improvement. Indeed, thanks to OSM data, we can know that some
specific buildings have a particular type, e.g. a building can
be a church, a grocery store or a house. Point annotations,
such as parking lots, are also dismissed but could provide
meaningful insights about the semantics of the area. Furthermore, we underline that even though the OSM layers
that we used were more recent (2 years) than to the optical
data, there were few enough disagreements so that the models were robust to those conflicts. Yet, data fusion should
be done carefully if the sources do not represent the same
underlying reality. In the case of the OSM data, this could
be worked around by extracting the layers from the OSM
archives if the optical data is not recent enough. Finally,
mapping style and coverage can vary a lot based on the
observed regions. For example, urban areas in developed
countries are thoroughly mapped, whereas annotations are
very scarce in rural areas in developing countries. This enforces the need for the model to be robust to errors and miss-
ing OSM input data for very large scale mapping.
6. Conclusion
In this work, we showed how to integrate ancillary GIS
data in deep learning-based semantic labeling. We presented two methods: one for coarse-to-fine segmentation,
using deep learning on RGB data to refine OpenStreetMap
semantic maps, and one for data fusion to merge multispectral data and OSM rasters to predict local climate zones. We
validated our methods on two public datasets: the ISPRS
Potsdam 2D Semantic Labeling Challenge and the Data Fusion Contest 2017. We increase our semantic labeling overall accuracy by 2.5% on the former and by nearly 5% on
the latter by integrating OpenStreetMap data in the learning
process. Moreover, on the ISPRS Potsdam dataset, using
OSM layers in a residual correction fashion accelerates the
model convergence by 25%. Our findings show that GIS
sparse data can be leveraged successfully for semantic labeling on those two use cases, as it improves significantly
the classification accuracy of the models. We think that using crowdsourced and open GIS data is an exciting topic
of research, and this work provides new insights on how to
use this data to improve and accelerate learning based on
traditional sensors.
Acknowledgements
The Vaihingen dataset was provided by the German Society for Photogrammetry, Remote Sensing and
Geoinformation (DGPF) [7]:
http://www.ifp.
uni-stuttgart.de/dgpf/DKEP-Allg.html.
(c) Target ground truth
(b) OSM input
(a) RGB input
(d) SegNet (RGB)
(e) FuseNet (OSM+RGB)
Figure 4: Excerpt from the classification results on Potsdam
The authors thank the ISPRS for making the Vaihingen and
Potsdam datasets available and organizing the semantic
labeling challenge. The authors would like to thank
the WUDAPT (http://www.wudapt.org/) and
GeoWIKI (http://geo-wiki.org/) initiatives and
the IEEE GRSS Image Analysis and Data Fusion Technical
Committee. Nicolas Audebert’s work is supported by the
Total-ONERA research project NAOMI.
References
[1] N. Audebert, B. Le Saux, and S. Lefèvre. Semantic Segmentation of Earth Observation Data Using Multimodal and
Multi-scale Deep Networks. In Computer Vision – ACCV
2016, pages 180–196. Springer, Cham, Nov. 2016.
[2] V. Badrinarayanan, A. Kendall, and R. Cipolla. SegNet:
A Deep Convolutional Encoder-Decoder Architecture for
Scene Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, PP(99):1–1, 2017.
[3] M. Campos-Taberner, A. Romero-Soriano, C. Gatta,
G. Camps-Valls, A. Lagrange, B. Le Saux, A. Beaupère,
A. Boulch, A. Chan-Hon-Tong, S. Herbin, H. Randrianarivo,
M. Ferecatu, M. Shimoni, G. Moser, and D. Tuia. Processing
of Extremely High-Resolution LiDAR and RGB Data: Outcome of the 2015 IEEE GRSS Data Fusion Contest Part A: 2D Contest. IEEE Journal of Selected Topics in Applied Earth
Observations and Remote Sensing, PP(99):1–13, 2016.
1e7
1126605
1878448
725054
21838
buildings
1664120
51520875
2967059
111546
76285
low vegetation
1108857
558643
35856853
2083046
352
521341
57615
3122028
19490186
44788
trees
49485
14444
impervious surfaces buildings
1754
35150
3010013
low vegetation
trees
cars
6.0
6.0
62840673
906387
1553010
697701
28881
buildings
1316018
52340387
2541862
112340
90877
low vegetation
1485836
485991
35536974
2206183
658
561322
66035
3004559
19578633
39992
impervious surfaces
4.5
3.0
trees
1.5
cars
1e7
1e7
6.0
62297705
impervious surfaces
1084647
1848230
860691
30171
4.5
buildings
2549169
50056828
3264962
133645
77819
3.0
low vegetation
1218568
583730
35235082
2715956
213
trees
474568
49960
2658349
20038376
44156
cars
48348
25658
2871
37085
3001551
57425
18174
impervious surfaces buildings
665
39127
2999795
low vegetation
trees
cars
low vegetation
trees
cars
4.5
3.0
1.5
1.5
cars
(a) SegNet RGB
62461202
impervious surfaces
impervious surfaces buildings
(b) Residual Correction OSM+RGB
(c) FuseNet OSM+RGB
Figure 5: Confusion matrices on Potsdam using the different methods.
RGB only
RGB
RGB + OSM
GT
iteration
10,000
20,000
50,000
80,000
120,000
Figure 6: Evolution of the predictions coming from SegNet using RGB only vs. RGB + OSM. Integrating OSM data makes
the output more visually coherent, even in the early learning stages.
Legend: white: impervious surfaces, blue: buildings, cyan: low vegetation, green: trees, yellow: vehicles, red: clutter, black: undefined.
(a) RGB (composite)
(b) Prediction (FuseNet)
Figure 7: Partial results on the city of Chicago (DFC2017)
Legend: cyan: compact high rise, blue: compact mid-rise, yellow:
open high rise, brown: dense trees, grey: water
[4] M. Castelluccio, G. Poggi, C. Sansone, and L. Verdoliva.
Land use classification in remote sensing images by convolutional neural networks. arXiv:1508.00092 [cs], Aug. 2015.
[5] J. Chen and A. Zipf. DeepVGI: Deep Learning with Volunteered Geographic Information. In 26th International World
Wide Web Conference (Poster). ACM, 2017.
[6] D. Costea and M. Leordeanu. Aerial image geolocalization
from recognition and matching of roads and intersections.
arXiv:1605.08323 [cs], May 2016. arXiv: 1605.08323.
[7] M. Cramer. The DGPF test on digital aerial camera evaluation – overview and test design. Photogrammetrie – Fernerkundung – Geoinformation, 2:73–82, 2010.
[8] O. Danylo, L. See, B. Bechtel, D. Schepaschenko, and
S. Fritz. Contributing to WUDAPT: A Local Climate Zone
Classification of Two Cities in Ukraine. IEEE Journal of
Selected Topics in Applied Earth Observations and Remote
Sensing, 9(5):1841–1853, May 2016.
[9] A. Eitel, J. T. Springenberg, L. Spinello, M. Riedmiller, and
W. Burgard. Multimodal deep learning for robust RGB-D
object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 681–
687, Sept. 2015.
[10] C. C. Fonte, J. A. Patriarca, M. Minghini, V. Antoniou,
L. See, and M. A. Brovelli. Using OpenStreetMap to Create Land Use and Land Cover Maps. In C. E. C. Campelo,
M. Bertolotto, and P. Corcoran, editors, Volunteered Geographic Information and the Future of Geospatial Data,
pages 113–137. IGI Global, IGI Global, 2017.
[11] C. Geiß, A. Schauß, T. Riedlinger, S. Dech, C. Zelaya,
N. Guzmán, M. A. Hube, J. J. Arsanjani, and H. Taubenböck. Joint use of remote sensing data and volunteered geographic information for exposure estimation: evidence from
Valparaíso, Chile. Natural Hazards, 86(1):81–105, Mar.
2017.
[12] GRSS. 2017 IEEE GRSS Data Fusion Contest, 2017.
[13] C. Hazirbas, L. Ma, C. Domokos, and D. Cremers. FuseNet:
Incorporating Depth into Semantic Segmentation via FusionBased CNN Architecture. In Computer Vision – ACCV 2016,
pages 213–228. Springer, Cham, Nov. 2016.
[14] K. He, X. Zhang, S. Ren, and J. Sun. Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet
Classification. In 2015 IEEE International Conference on
Computer Vision (ICCV), pages 1026–1034, 2015.
[15] K. He, X. Zhang, S. Ren, and J. Sun. Deep Residual Learning
for Image Recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–
778, June 2016.
[16] J. Hu, L. Mou, A. Schmitt, and X. X. Zhu. FusioNet: A
Two-Stream Convolutional Neural Network for Urban Scene
Classification using PolSAR and Hyperspectral Data. In
2017 Joint Urban Remote Sensing Event (JURSE), Mar.
2017.
[17] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros. Image-toImage Translation with Conditional Adversarial Networks.
arXiv:1611.07004 [cs], Nov. 2016. arXiv: 1611.07004.
[18] D. Kingma and J. Ba. Adam: A Method for Stochastic
Optimization. arXiv:1412.6980 [cs], Dec. 2014. arXiv:
1412.6980.
[19] G. Lin, A. Milan, C. Shen, and I. Reid. RefineNet:
Multi-Path Refinement Networks with Identity Mappings
for High-Resolution Semantic Segmentation. arXiv preprint
arXiv:1611.06612, 2016.
[20] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional
networks for semantic segmentation. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
pages 3431–3440, June 2015.
[21] E. Maggiori, Y. Tarabalka, G. Charpiat, and P. Alliez. Fully
convolutional neural networks for remote sensing image
classification. In 2016 IEEE International Geoscience and
Remote Sensing Symposium (IGARSS), pages 5071–5074,
July 2016.
[22] D. Marmanis, K. Schindler, J. D. Wegner, S. Galliani,
M. Datcu, and U. Stilla. Classification With an Edge:
Improving Semantic Image Segmentation with Boundary
Detection. arXiv:1612.01337 [cs], Dec. 2016. arXiv:
1612.01337.
[23] V. Mnih. Machine Learning for Aerial Image Labeling. PhD
thesis, University of Toronto, 2013.
[24] V. Mnih and G. E. Hinton. Learning to Detect Roads in HighResolution Aerial Images. In K. Daniilidis, P. Maragos, and
N. Paragios, editors, Computer Vision – ECCV 2010, number
6316 in Lecture Notes in Computer Science, pages 210–223.
Springer Berlin Heidelberg, Sept. 2010.
[25] L. Mou and X. X. Zhu. Spatiotemporal scene interpretation of space videos via deep neural network and tracklet analysis. In Geoscience and Remote Sensing Symposium (IGARSS), 2016 IEEE International, pages 1823–1826.
IEEE, 2016.
[26] S. Paisitkriangkrai, J. Sherrah, P. Janney, and A. Van
Den Hengel. Effective semantic pixel labelling with convolutional networks and Conditional Random Fields. In 2015
IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 36–43, June 2015.
[27] F. Rottensteiner, G. Sohn, J. Jung, M. Gerke, C. Baillard,
S. Benitez, and U. Breitkopf. The ISPRS benchmark on
urban object classification and 3d building reconstruction.
ISPRS Annals of the Photogrammetry, Remote Sensing and
Spatial Information Sciences, 1:3, 2012.
[28] J. Sherrah.
Fully Convolutional Networks for Dense
Semantic Labelling of High-Resolution Aerial Imagery.
arXiv:1606.02585 [cs], June 2016.
[29] K. Simonyan and A. Zisserman.
Very Deep Convolutional Networks for Large-Scale Image Recognition.
arXiv:1409.1556 [cs], Sept. 2014.
[30] I. D. Stewart and T. R. Oke. Local Climate Zones for Urban
Temperature Studies. Bulletin of the American Meteorological Society, 93(12):1879–1900, 2012.
[31] M. Vakalopoulou, K. Karantzalos, N. Komodakis, and
N. Paragios. Building detection in very high resolution multispectral data with deep learning features. In Geoscience
and Remote Sensing Symposium (IGARSS), 2015 IEEE International, pages 1873–1876. IEEE, 2015.
[32] M. Vakalopoulou, C. Platias, M. Papadomanolaki, N. Paragios, and K. Karantzalos. Simultaneous registration, segmentation and change detection from multisensor, multitemporal satellite image pairs. July 2016.
[33] M. Volpi and D. Tuia. Dense Semantic Labeling of Subdecimeter Resolution Images With Convolutional Neural
Networks. IEEE Transactions on Geoscience and Remote
Sensing, 55(2):881–893, 2017.
[34] J. Yuan.
Automatic Building Extraction in Aerial
Scenes Using Convolutional Networks. arXiv preprint
arXiv:1602.06564, 2016.
| 9cs.NE
|
Error Reporting in Parsing Expression Grammars
André Murbach Maidl
Polytechnic School – PUCPR – Curitiba – Brazil
Fabio Mascarenhas
Department of Computer Science – UFRJ – Rio de Janeiro – Brazil
arXiv:1405.6646v3 [cs.PL] 13 Jul 2016
Sérgio Medeiros
School of Science and Technology – UFRN – Natal – Brazil
Roberto Ierusalimschy
Department of Computer Science – PUC-Rio – Rio de Janeiro – Brazil
Abstract
Parsing Expression Grammars (PEGs) describe top-down parsers. Unfortunately, the errorreporting techniques used in conventional top-down parsers do not directly apply to parsers
based on Parsing Expression Grammars (PEGs), so they have to be somehow simulated.
While the PEG formalism has no account of semantic actions, actual PEG implementations
add them, and we show how to simulate an error-reporting heuristic through these semantic
actions.
We also propose a complementary error reporting strategy that may lead to better error
messages: labeled failures. This approach is inspired by exception handling of programming
languages, and lets a PEG define different kinds of failure, with each ordered choice operator
specifying which kinds it catches. Labeled failures give a way to annotate grammars for
better error reporting, to express some of the error reporting strategies used by deterministic
parser combinators, and to encode predictive top-down parsing in a PEG.
Keywords: parsing, error reporting, parsing expression grammars, packrat parsing, parser
combinators
1. Introduction
When a parser receives an erroneous input, it should indicate the existence of syntax
errors. Errors can be handled in various ways. The easiest is just to report that an error
Email addresses: [email protected] (André Murbach Maidl), [email protected] (Fabio
Mascarenhas), [email protected] (Sérgio Medeiros), [email protected] (Roberto
Ierusalimschy)
Preprint submitted to Science of Computer Programming
July 14, 2016
was found, where is was found and what was expected at that point and then abort. At the
other end of the spectrum we find mechanisms that attempt to parse the complete input,
and report as many errors as best as possible.
The LL(k) and LR(k) methods detect syntax errors very efficiently because they have
the viable prefix property, that is, these methods detect a syntax error as soon as k tokens
are read and cannot be used to extend the thus far accepted part of the input into a viable
prefix of the language [1]. LL(k) and LR(k) parsers can use this property to produce suitable,
though generic, error messages.
Parsing Expression Grammars (PEGs) [2] are a formalism for describing the syntax
of programming languages. We can view a PEG as a formal description of a top-down
parser for the language it describes. PEGs have a concrete syntax based on the syntax of
regexes, or extended regular expressions. Unlike Context-Free Grammars (CFGs), PEGs
avoid ambiguities in the definition of the grammar’s language due to the use of an ordered
choice operator.
More specifically, a PEG can be interpreted as a the specification of a recursive descent
parser with restricted (or local) backtracking. This means that the alternatives of a choice
are tried in order; as soon as an alternative recognizes an input prefix, no other alternative
of this choice will be tried, but when an alternative fails to recognize an input prefix, the
parser backtracks to try the next alternative.
On the one hand, PEGs can be interpreted as a formalization of a specific class of topdown parsers [2]; on the other hand, PEGs cannot use error handling techniques that are
often applied to predictive top-down parsers, because these techniques assume the parser
reads the input without backtracking [3]. In top-down parsers without backtracking, it is
possible to signal a syntax error as soon as the next input symbol cannot be accepted. In
PEGs, it is more complicated to identify the cause of an error and the position where it
occurs, because failures during parsing are not necessarily errors, but just an indication that
the parser cannot proceed and a different choice should be made elsewhere.
Ford [3] has already identified this limitation of error reporting in PEGs, and, in his
parser generators for PEGs, included an heuristic for better error reporting. This heuristic
simulates the error reporting technique that is implemented in top-down parsers without
backtracking. The idea is to track the position in the input where the farthest failure
occurred, as well as what the parser was expecting at that point, and report this to the user
in case of errors.
Tracking the farthest failure position and context gives us PEGs that produce error
messages similar to the automatically produced error messages of other top-down parsers;
they tell the user the position where the error was encountered, what was found in the input
at that position, and what the parser was expecting to find.
In this paper, we show how grammar writers can use this error reporting technique even
in PEG implementations that do not implement it, by making use of semantic actions that
expose the current position in the input and the possibility to access some form of mutable
state associated with the parsing process.
We also propose a complementary approach for error reporting in PEGs, based on the
concept of labeled failures, inspired by the standard exception handling mechanisms as found
2
Tiny
CmdSeq
Cmd
IfCmd
RepeatCmd
AssignCmd
ReadCmd
WriteCmd
Exp
SimpleExp
Term
Factor
← CmdSeq
← (Cmd SEMICOLON) (Cmd SEMICOLON)∗
← IfCmd / RepeatCmd / AssignCmd / ReadCmd / WriteCmd
← IF Exp THEN CmdSeq (ELSE CmdSeq / ε) END
← REPEAT CmdSeq UNTIL Exp
← NAME ASSIGNMENT Exp
← READ NAME
← WRITE Exp
← SimpleExp ((LESS / EQUAL) SimpleExp / ε)
← Term ((ADD / SUB) Term)∗
← Factor ((MUL / DIV) Factor )∗
← OPENPAR Exp CLOSEPAR / NUMBER / NAME
Figure 1: A PEG for the Tiny language
in programming languages. Instead of just failing, a labeled PEG can produce different
kinds of failure labels using a throw operator. Each label can be tied to a more specific
error message. PEGs can also catch such labeled failures, via a change to the ordered choice
operator. We formalize labeled failures as an extension of the semantics of regular PEGs.
With labeled PEGs we can express some alternative error reporting techniques for topdown parsers with local backtracking. We can also encode predictive parsing in a PEG, and
we show how to do that for LL(∗) parsing, a powerful predictive parsing strategy.
The rest of this paper is organized as follows: Section 2 contextualizes the problem of
error handling in PEGs, explains in detail the failure tracking heuristic, and shows how it
can be realized in PEG implementations that do not support it directly; Section 3 discusses
related work on error reporting for top-down parsers with backtracking; Section 4 introduces
and formalizes the concept of labeled failures, and shows how to use it for error reporting;
Section 5 compares the error messages generated by a parser based on the failure tracking
heuristic with the ones generated by a parser based on labeled failures; Section 6 shows how
labeled failures can encode some of the techniques of Section 3, as well as predictive parsing;
finally, Section 7 gives some concluding remarks.
2. Handling Syntax Errors with PEGs
In this section, we use examples to present in more detail how a PEG behaves badly in
the presence of syntax errors. After that, we present a heuristic proposed by Ford [3] to
implement error reporting in PEGs. Rather than using the original notation and semantics
of PEGs given by Ford [2], our examples use the equivalent and more concise notation and
3
semantics proposed by Medeiros et al. [4, 5, 6]. We will extend both the notation and the
semantics in Section 4 to present PEGs with labeled failures.
A PEG G is a tuple (V, T, P, pS ) where V is a finite set of non-terminals, T is a finite
set of terminals, P is a total function from non-terminals to parsing expressions and pS
is the initial parsing expression. We describe the function P as a set of rules of the form
A ← p, where A ∈ V and p is a parsing expression. A parsing expression, when applied
to an input string, either fails or consumes a prefix of the input and returns the remaining
suffix. The abstract syntax of parsing expressions is given as follows, where a is a terminal,
A is a non-terminal, and p, p1 and p2 are parsing expressions:
p = ε | a | A | p1 p2 | p1 /p2 | p ∗ | !p
Intuitively, ε successfully matches the empty string, not changing the input; a matches
and consumes itself or fails otherwise; A tries to match the expression P (A); p1 p2 tries to
match p1 followed by p2 ; p1 /p2 tries to match p1 ; if p1 fails, then it tries to match p2 ; p∗
repeatedly matches p until p fails, that is, it consumes as much as it can from the input; the
matching of !p succeeds if the input does not match p and fails when the the input matches
p, not consuming any input in either case; we call it the negative predicate or the lookahead
predicate.
Figure 1 presents a PEG for the Tiny language [7]. Tiny is a simple programming
language with a syntax that resembles Pascal’s. We will use this PEG, which can be seen
as the equivalent of an LL(1) CFG, to show how error reporting differs between top-down
parsers without backtracking and PEGs.
PEGs usually express the language syntax at the character level, without the need of a
separate lexer. For instance, we can write the lexical rule IF as follows, assuming we have
non-terminals Skip, which consumes whitespace, and IDRest, which consumes any character
that may be present on a proper suffix of an identifier1 :
IF ← if !IDRest Skip
Now, we present an example of erroneous Tiny code so we can compare approaches for
error reporting. The program in Figure 2 is missing a semicolon (;) in the assignment in
line 5. A predictive top-down parser that aborts on the first error presents an error message
like:
factorial.tiny:6:1: syntax error, unexpected ’until’, expecting ’;’
The error is reported in line 6 because the parser cannot complete a valid prefix of the
language, since it unexpectedly finds the token until when it was expecting a command
terminator (;).
In PEGs, we can try to report errors using the remaining suffix, but this approach usually
does not help the PEG produce an error message like the one shown above. In general, when
1
In the presented PEG, we omitted the lexical rules for brevity.
4
01
02
03
04
05
06
07
n := 5;
f := 1;
repeat
f := f * n;
n := n - 1
until (n < 1);
write f;
Figure 2: Program for the Tiny Language with a Syntax Error
a PEG finishes parsing the input, a remaining non-empty suffix means that parsing did not
reach the end of file due to a syntax error. However, this suffix usually does not indicate
the actual place of the error, as the error will have caused the PEG to backtrack to another
place in the input.
In our example, the problem happens when the PEG tries to recognize the sequence of
commands inside the repeat command. Even though the program has a missing semicolon
(;) in the assignment in line 5, making the PEG fail to recognize the sequence of commands
inside the repeat command, this failure is not treated as an error. Instead, this failure makes
the recognition of the repeat command also fail. For this reason, the PEG backtracks the
input to line 3 to try to parse other alternatives for CmdSeq, and since these do not exist, its
ancestor Cmd. Since it is not possible to recognize a command other than repeat at line 3,
the parsing finishes without consuming all the input. Hence, if the PEG uses the remaining
suffix to produce an error message, the PEG reports line 3 instead of line 6 as the location
where no further progress can be made.
There is no perfect method to identify which information is the most relevant to report
an error. In our example it is easy for the parser to correctly report what the error is, but
it is easy to construct examples where this is not the case. If we add the semicolon in the
end of line 6 and remove line 3, a predictive top-down parser would complain about finding
an until where it expected another statement, while the actual error is a missing repeat.
According to Ford [3], using the information of the farthest position that the PEG reached
in the input is a heuristic that provides good results. PEGs define top-down parsers and try
to recognize the input from left to right, so the position farthest to the right in the input
that a PEG reaches during parsing usually is close to the real error [3]. The same idea for
error reporting in top-down parsings with backtracking was also mentioned in Section 16.2
of [8].
Ford used this heuristic to add error reporting to his PEG implementation using packrat
parsers [3]. A packrat parser generated by Pappy [9], Ford’s PEG parser generator, tracks
the farthest position and uses this position to report an error. In other words, this heuristic helps packrat parsers to simulate the error reporting technique that is implemented in
deterministic parsers.
Although Ford only has discussed his heuristic in relation to packrat parsers, we can use
the farthest position heuristic to add error reporting to any implementation of PEGs that
5
provides semantic actions. The idea is to annotate the grammar with semantic actions that
track this position. While this seems onerous, we just need to add annotations to all the
lexical rules to implement error reporting.
For instance, in Leg [10], a PEG parser generator with Yacc-style semantic actions, we
can annotate the rule for SEMICOLON as follows, where | is Leg’s ordered choice operator,
and following it is a semantic action (in the notation used by Leg):
SEMICOLON = ";" Skip | &{ updateffp() }
The function updateffp that the semantic action calls updates the farthest failure position in a global variable if the current parsing position is greater than the position that is
stored in this global, then makes the whole action fail so parsing continues as if the original
failure had occurred.
However, storing just the farthest failure position does not give the parser all the information it needs to produce an informative error message. That is, the parser has the information
about the position where the error happened, but it lacks the information about what terminals failed at that position. Thus, we extend our approach by including the terminals
in the annotations so the parser can also track these names in order to compute the set of
expected terminals at a certain position:
SEMICOLON = ";" Skip | &{ updateffp(";") }
The extended implementation of updateffp keeps, for a given failure position, the names
of all the symbols expected there. If the current position is greater than the farthest failure,
updateffp initializes this set with just the given name. If the current position equals the
farthest failure, updateffp adds this name to the set.
Parsers generated by Pappy also track the set of expected terminals, but with limitations.
The error messages include only symbols and keywords that were defined in the grammar
as literal strings. That is, the error messages do not include terminals that were defined
through character classes.
Our approach of naming terminals in the semantic actions avoids the kind of limitation
found in Pappy, though it increases the annotation burden because the implementor of the
PEG is also responsible for adding one semantic action for each terminal and its respective
name.
The annotation burden can be lessened in implementations of PEGs that treat parsing
expressions as first-class objects, because this makes it possible to define functions that
annotate the lexical parts of the grammar to track errors, record information about the
expected terminals to produce good error messages, and enforce lexical conventions such as
the presence of surrounding whitespace. For instance, in LPEG [11, 12], a PEG implementation for Lua that defines patterns as first-class objects, we can annotate the rule CmdSeq
as follows, where the patterns V"A", p1 * p2, and p^0 are respectively equivalent to parsing
expressions A, p1 p2 , and p∗ (in the notation used by LPEG):
CmdSeq = V"Cmd" * symb(";") * (V"Cmd" * symb(";"))^0;
6
The function symb receives a string as its only argument and returns a parser that is
equivalent to the parsing expression that we used in the Leg example. That is, symb(";")
is equivalent to ";" Skip | &{ updateffp(";") }.
We implemented error tracking and reporting using semantic actions as a set of parsing
combinators on top of LPeg and used these combinators to implement a PEG parser for
Tiny. It produces the following error message for the example we have been using in this
section:
factorial.tiny:6:1: syntax error, unexpected ’until’,
expecting ’;’, ’=’, ’<’, ’-’, ’+’, ’/’, ’*’
We tested this PEG parser for Tiny with other erroneous inputs and in all cases the parser
identified an error in the same place as a top-down parser without backtracking. In addition,
the parser for Tiny produced error messages that are similar to the error messages produced
by packrat parsers generated by Pappy. We annotated other grammars and successfully
obtained similar results. However, the error messages are still generic; they are not as
specific as the error messages of a hand-written top-down parser.
3. Error Reporting in Top-Down Parsers with Backtracking
In this section, we discuss alternative approaches for error reporting in top-down parsers
with backtracking.
Mizushima et al. [13] proposed a cut operator (↑) to reduce the space consumption of
packrat parsers; the authors claimed that the cut operator can also be used to implement
error reporting in packrat parsers, but the authors did not give any details on how the
cut operator could be used for this purpose. The cut operator is borrowed from Prolog
to annotate pieces of a PEG where backtracking should be avoided. PEGs’ ordered choice
works in a similar way to Prolog’s green cuts, that is, they limit backtracking to discard
unnecessary solutions. The cut proposed for PEGs is a way to implement Prolog’s white
cuts, that is, they prevent backtracking to rules that will certainly fail.
The semantics of cut is similar to the semantics of an if-then-else control structure
and can be simulated through predicates. For instance, the PEG (with cut) A ← B ↑ C/D
is functionally equivalent to the PEG (without cut) A ← BC/!BD that is also functionally
equivalent to the rule A ← B[C, D] on Generalized Top-Down Parsing Language (GTDPL),
one of the parsing techniques that influenced the creation of PEGs [3, 9, 2]. On the three
cases, the expression D is tried only if the expression B fails. Nevertheless, this translated
PEG still backtracks whenever B successfully matches and C fails. Thus, it is not trivial to
use this translation to implement error reporting in PEGs.
Rats! [14] is a popular packrat parser that implements error reporting with a strategy
similar to Ford’s, with the change that it always reports error positions at the start of
productions, and pretty-prints non-terminal names in the error message. For example, an
error in a ReturnStatement non-terminal becomes return statement expected.
Even though error handling is an important task for parsers, we did not find any other
research results about error handling in PEGs, beyond the heuristic proposed by Ford and
7
the cut operator proposed by Mizushima et al. However, parser combinators [15] present
some similarities with PEGs so we will briefly discuss them for the rest of this section.
In functional programming it is common to implement recursive descent parsers using
parser combinators [15]. A parser is a function that we use to model symbols of the grammar. A parser combinator is a higher-order function that we use to implement grammar
constructions such as sequencing and choice. One kind of parser combinator implements
parsers that return a list of all possible results of a parse, effectively implementing a recursive descent parser with full backtracking. Despite being actually deterministic in behavior
(parsing the same input always yields the same list of results), these combinators are called
non-deterministic parser combinators due to their use of a non-deterministic choice operator. We get parser combinators that have the same semantics as PEGs by changing the
return type from list of results to Maybe. That is, we use deterministic parser combinators
that return Maybe to implement recursive descent parsers with limited backtracking [16]. In
the rest of this paper, whenever we refer to parser combinators we intend to refer to these
parser combinators with limited backtracking.
Like PEGs, most deterministic parser combinator libraries also use ordered choice, and
thus suffer from the same problems as PEGs with erroneous inputs, where the point that
the parser reached in the input is ususally far away from the point of the error.
Hutton [15] introduced the nofail combinator to implement error reporting in a quite
simple way: we just need to distinguish between failure and error during parsing. More
specifically, we can use the nofail combinator to annotate the grammar’s terminals and
non-terminals that should not fail; when they fail, the failure should be transformed into an
error. The difference between an error and a failure is that an ordered choice just propagates
an error in its first alternative instead of backtracking and trying its second alternative, so
any error aborts the whole parser. This technique is also called the three-values technique [17]
because the parser finishes with one of the following values: OK, Fail or Error.
Röjemo [18] presented a cut combinator that we can also use to annotate the grammar
pieces where parsing should be aborted on failure, on behalf of efficiency and error reporting.
The cut combinator is different from the cut operator2 (↑) for PEGs because the combinator
is abortive and unary while the operator is not abortive and nullary. The cut combinator
introduced by Röjemo has the same semantics as the nofail combinator introduced by
Hutton.
Partridge and Wright [17] showed that error detection can be automated in parser combinators when we assume that the grammar is LL(1). Their main idea is: if one alternative
successfully consumes at least one symbol, no other alternative can successfully consume
any symbols. Their technique is also known as the four-values technique because the parser
finishes with one of the following values: Epsn, when the parser finishes with success without consuming any input; OK, when the parser finishes with success consuming some input;
Fail, when the parser fails without consuming any input; and Error, when the parser fails
consuming some input. Three values were inspired by Hutton’s work [15], but with new
2
Throughout this paper, we refer to combinators of parser combinators and to operators of PEGs, but
these terms are effectively interchangeable.
8
p1
Error
Error
Error
Error
Fail
Fail
Fail
Fail
Epsn
Epsn
Epsn
Epsn
OK (x)
OK (x)
OK (x)
OK (x)
p2
Error
Fail
Epsn
OK (x)
Error
Fail
Epsn
OK (x)
Error
Fail
Epsn
OK (x)
Error
Fail
Epsn
OK (y)
p1 p2
Error
Error
Error
Error
Fail
Fail
Fail
Fail
Error
Fail
Epsn
OK (x)
Error
Error
OK (x)
OK (y)
p1 | p2
Error
Error
Error
Error
Error
Fail
Epsn
OK (x)
Error
Epsn
Epsn
OK (x)
OK (x)
OK (x)
OK (x)
OK (x)
Table 1: Behavior of sequence and choice in the four-values technique
meanings.
In the four-values technique, we do not need to annotate the grammar because the
authors changed the semantics of the sequence and choice combinators to automatically
generate the Error value according to the Table 1. In summary, the sequence combinator
propagates an error when the second parse fails after consuming some input while the choice
combinator does not try further alternatives if the current one consumed at least one symbol
from the input. In case of error, the four-values technique detects the first symbol following
the longest parse of the input and uses this symbol to report an error.
The four-values technique assumes that the input is composed by tokens which are provided by a separate lexer. However, being restricted to LL(1) grammars can be a limitation
because parser combinators, like PEGs, usually operate on strings of characters to implement both lexer and parser together. For instance, a parser for Tiny that is implemented
with Parsec [19] does not parse the following program: read x;. That is, the matching of
read against repeat generates an error. Such behavior is confirmed in Table 1 by the third
line from the bottom.
Parsec is a parser combinator library for Haskell that employs a technique equivalent
to the four-values technique for implementing LL(1) predictive parsers that automatically
report errors [19]. To overcome the LL(1) limitation, Parsec introduced the try combinator,
a dual of Hutton’s nofail combinator. The effect of try is to translate an error into a
backtrackeable failure. The idea is to use try to annotate the parts of the grammar where
arbitrary lookahead is needed.
Parsec’s restriction to LL(1) grammars made it possible to implement an error reporting
9
technique similar to the one used in top-down parsers. Parsec produces error messages that
include the error position, the character at this position and the FIRST and FOLLOW sets
of the productions that were expected at this position. Parsec also implements the error
injection combinator (<?>) for naming productions. This combinator gets two arguments:
a parser p and a string exp. The string exp replaces the FIRST set of a parser p when all
the alternatives of p failed. This combinator is useful to name terminals and non-terminals
to get better information about the context of a syntax error.
Swierstra and Duponcheel [20] showed an implementation of parser combinators for error
recovery, although most libraries and parser generators that are based on parser combinators
implement only error reporting. Their work relies on the fact that the grammar is LL(1) and
shows an implementation of parser combinators that repairs erroneous inputs, produces an
appropriated message, and continues parsing the rest of the input. This approach was later
extended to also deal with grammars that are not LL(1), including ambiguous grammars [21].
The extended approach relies heavily on some features that the implementation language
should have, such as lazy evaluation.
4. Labeled Failures
Exceptions are a common mechanism for signaling and handling errors in programming
languages. Exceptions let programmers classify the different errors their programs may
signal by using distinct types for distinct errors, and decouple error handling from regular
program logic.
In this section we add labeled failures to PEGs, a mechanism akin to exceptions and
exception handling, with the goal of improving error reporting while preserving the composability of PEGs. In the next section we discuss how to use PEGs with labeled failures to
implement some of the techniques that we have discussed in Section 3: the nofail combinator [15], the cut combinator [18], the four-values technique [17] and the try combinator [19].
A labeled PEG G is a tuple (V, T, P, L, pS ) where L is a finite set of labels that must
include the fail label, and the expressions in P have been extended with the throw operator,
explained below. The other parts use the same definitions from Section 2.
The abstract syntax of labeled parsing expressions adds the throw operator ⇑l , which
generates a failure with label l, and adds an extra argument S to the ordered choice operator,
which is the set of labels that the ordered choice should catch. S must be a subset of L.
p = ε | a | A | p1 p2 | p1 /S p2 | p ∗ | !p | ⇑l
Figure 3 presents the semantics of PEGs with labels as a set of inference rules. The
PEG
semantics of PEGs with labels is defined by the relation
among a parsing expression, an
PEG
input string and a result. The result is either a string or a label. The notation G[p] xy
y
means that the expression p matches the input xy, consumes the prefix x and leaves the
PEG
suffix y as the output. The notation G[p] xy
l indicates that the matching of p fails
with label l on the input xy.
Now a matches and consumes itself and fails with label fail otherwise; p1 p2 tries to
match p1 , if p1 matches an input prefix, then it tries to match p2 with the suffix left by p1 ,
10
Empty
G[ε] x
Terminal
PEG
(empty.1)
x
PEG
G[a] ax
(char.1)
x
G[P (A)] x
Non-terminal
G[A] x
PEG
G[p1 ] xy
Concatenation
PEG
PEG
PEG
PEG
G[p] x
PEG
G[p∗] x
PEG
PEG
G[p] x
G[p∗] x
PEG
Negative Predicate
PEG
fail
l
l
l
PEG
y
PEG
l
/S
G[p1
G[⇑l ] x
G[p2 ] y
G[p1 /S p2 ] xy
G[p1 ] x
Throw
y
G[p1 ] xy
y
, b 6= a (char.2)
G[a] ε
fail
p2 ] x
PEG
X
X
(ord.1)
G[p2 ] x
PEG
(rep.1)
x
PEG
PEG
X
X
PEG
G[p1 p2 ] x
G[p1 ] x
PEG
G[p1 /S p2 ] x
l
l
PEG
l
PEG
G[!p] x
(char.3)
fail
fail
PEG
y
G[p∗] xy
G[p] x
PEG
l
G[!p] x
PEG
l
x
(not.1)
, l 6∈ S (ord.2)
PEG
l
G[p∗] y
PEG
PEG
X
X
G[p] xy
G[!p] xy
PEG
PEG
y
fail
, l 6= fail (not.3)
(throw.1)
Figure 3: Semantics of PEGs with labels
11
(con.2)
, l ∈ S (ord.3)
PEG
G[p] xy
G[p1 ] x
(con.1)
(rep.2)
, l 6= fail (rep.3)
G[p] x
PEG
(var.1)
X
G[p1 p2 ] xy
Ordered Choice
Repetition
X
G[b] ax
(not.2)
the label l is propagated otherwise; p1 /S p2 tries to match p1 in the input and tries to match
p2 in the same input only if p1 fails with a label l ∈ S, the label l is propagated otherwise; p∗
repeatedly matches p until the matching of p silently fails with label fail, and propagates a
label l when p fails with this label; !p successfully matches if the input does not match p with
the label fail, fails producing the label fail when the input matches p, and propagates a
label l when p fails with this label, not consuming the input in all cases; ⇑l produces the
label l.
We faced some design decisions in our formulation that are worth discussing. First, we
use a set of labels in the ordered choice as a convenience. We could have each ordered choice
handling a single label, and it would just lead to duplication: an expression p1 /{l1 ,l2 ,...,ln} p2
would become ( ... ((p1 /l1 p2 ) /l2 p2 ) ... /ln p2 ).
Second, we require the presence of a fail label to maintain compatibility with the
original semantics of PEGs, where we only have fail to signal both error and failure. For
the same reason, we define the expression p1 /p2 as syntactic sugar for p1 /{fail} p2 .
Another choice was how to handle labels in a repetition. We chose to have a repetition
stop silently only on the fail label to maintain the following identity, which holds for
unlabeled PEGs: an expression p∗ is equivalent to a fresh non-terminal A with the rule
A ← p A / ε.
Finally, the negative predicate succeeds only on the fail label to allow the implementation of the positive predicate: the expression &p that implements the positive predicate in
the original semantics of PEGs [3, 9, 2] is equivalent to the expression !!p. Both expressions
successfully match if the input matches p, fail producing the label fail when the input does
not match p, and propagate a label l when p fails with this label, not consuming the input
in all cases.
Figure 4 presents a PEG with labels for the Tiny language from Section 2. The expression
l
[p] is syntactic sugar for (p / ⇑l ).
The strategy we used to annotate the grammar was the following: first, annotate every
terminal that should not fail, that is, making the PEG backtrack on failure of that terminal
would be useless, as the whole parse would either fail or not consume the whole input in
that case. For an LL(1) grammar like the one in the example, that means all terminals in a
production except the one in the very beginning of the production.
After annotating the terminals, we do the same for whole productions. We annotate
productions where failing the whole production always implies an error in the input, adding
a new alternative that throws an error label specific to that production.
For Tiny, we end up annotating just two productions, Factor and CmdSeq. Productions
Exp, SimpleExp, and Term also should not fail, but after annotating Factor they always
either succeed or throw the label exp. The Cmd production can fail, because it controls
whether the repetition inside CmdSeq stops or continue.
Notice that this is just an example of how a grammar can be annotated. More thorough
analyses are possible: for example, we can deduce that Cmd is not allowed to fail unless the
next token is one of ELSE, END, UNTIL, or the end of the input (the FOLLOW set of Cmd),
and instead of ⇑cmd add !(ELSE / END / UNTIL / !.) ⇑cmd as a new alternative. This would
remove the need for the ⇑cmd annotation of CmdSeq.
12
Tiny ← CmdSeq
CmdSeq
Cmd
IfCmd
RepeatCmd
AssignCmd
ReadCmd
WriteCmd
Exp
SimpleExp
Term
Factor
← (Cmd [SEMICOLON]sc ) (Cmd [SEMICOLON]sc )∗ / ⇑cmd
← IfCmd / RepeatCmd / AssignCmd / ReadCmd / WriteCmd
← IF Exp [THEN]then CmdSeq (ELSE CmdSeq / ε) [END]end
← REPEAT CmdSeq [UNTIL]until Exp
← NAME [ASSIGNMENT]bind Exp
← READ [NAME]read
← WRITE Exp
← SimpleExp ((LESS / EQUAL) SimpleExp / ε)
← Term ((ADD / SUB) Term)∗
← Factor ((MUL / DIV) Factor )∗
← OPENPAR Exp [CLOSEPAR]cp / NUMBER / NAME / ⇑exp
Figure 4: A PEG with labels for the Tiny language
The PEG reports an error when parsing finishes with an uncaught label. Each label is
associated with a meaningful error message. For instance, if we use this PEG for Tiny to
parse the code example from Section 2, parsing finishes with the sc label and the PEG can
use it to produce the following error message:
factorial.tiny:6:1: syntax error, there is a missing ’;’
Note how the semantics of the repetition works with the rule CmdSeq. Inside the repetition, the fail label means that there are no more commands to be matched and the
repetition should stop while the sc label means that a semicolon (;) failed to match. It
would not be possible to write the rule CmdSeq using repetition if we had chosen to stop
the repetition with any label, instead of stopping only with the fail label, because the repetition would accept the sc label as the end of the repetition whereas it should propagate
this label.
Although the semantics of PEGs with labels presented in Figure 3 allows us to generate
specific error messages, it does not give us information about the location where the failure
probably is, so it is necessary to use some extra mechanism (e.g., semantic actions) to get
this information. To avoid this, we can adapt the semantics of PEGs with labels to give us
a tuple (l, y) in case of a failure, where y the suffix of the input that PEG was trying to
match when label l was thrown. Updating the semantics of Figure 3 to reflect this change
is straightforward.
In the next section, we try to establish a comparison between the farthest failure position
heuristic and the labeled failure mechanism by contrasting two different implementations of
a parser for a dialect of the Lua language.
13
5. Labeled Failures versus Farthest Failure Position
In this section we will compare two parser implementations for the Typed Lua language,
one that uses the farthest failure position heuristic for error reporting, which was implemented first, and one based on labeled failures.
Typed Lua [22] is an optionally-typed extension of the Lua programming language [23].
The Typed Lua parser recognizes plain Lua programs, and also Lua programs with type
annotations. The first version of the parser was implemented using Ford’s heuristic and the
LPeg library 3 .
As LPeg does not have a native error reporting mechanism based on Ford’s strategy, the
failure tracking heuristic was implemented following the approach described in Section 2,
which uses semantic actions.
Below we have the example of a Lua statement with a syntax error:
a = function (a,b,) end
In this case, the parser gives us the following error message, which is quite precise:
test.lua:1:19: syntax error, unexpected ’)’, expecting ’...’, ’Name’
In the previous case, the list of expected tokens had only two candidates, but this is not
always the case. For example, let us consider the following Lua program, where there is no
expression after the elseif in line 5:
01
02
03
04
05
06
07
if a then
return x
elseif b then
return y
elseif
end
The corresponding error message has a lengthy list of tokens, which does not help much
to fix the error:
test.lua:7:1: syntax error, unexpected ’end’, expecting ’(’, ’Name’, ’{’,
’function’, ’...’, ’true’, [9 more tokens]
When using the Typed Lua parser based on Ford’s heuristic it is not uncommon to get
a message like this. An analysis of the test cases available in the parser package shows us
that around half of the expected error messages have a list of at least eight expected tokens
(there were messages with a list of 39 expected tokens).
3
The first version of Typed Lua parser is available at https://github.com/andremm/typedlua/blob/master/typedlua
14
The second implementation of the Typed Lua parser was based on labeled failures and
used the LPegLabel library [24], which is an extension of the LPeg library that supports
labeled failures 4 .
The use of labeled failures implies in an annotation burden to specify when each label
should be thrown. In the case of Typed Lua grammar, we defined almost 50 different labels,
using the same basic strategy that we used to annotate the Tiny grammar of Section 4.
Given the previous Lua program, the error message presented now is:
test.lua:7:1: expecting <exp> after ’elseif’
This error message is more informative than the previous one, which was generated
automatically. We analyzed the error messages generated by the two parsers in 53 examples,
and considered that in more than half of these examples the parser based on labeled failures
produced a better error message. In about 20% of the cases we considered the error messages
of both approaches similar, and in other 20% of the cases the parser based on Ford’s heuristic
generated better error messages.
The error location indicated by the two parsers in the examples analyzed was essentially
the same. This seems to indicate that the main difference in practice between both approaches is related to the length of the error message generated. By using labeled failures
we can probably get a simple error message at the cost of annotating the grammar, while by
using the farthest failure tracking heuristic we can automatically generate error messages,
which sometimes may contain a long list of expected tokens.
A point that is worth mentioning about the labeled failure approach is that it is not
mandatory to annotate the entire grammar. The grammar can be annotated incrementally,
at the points where the current error message is not good enough, and when no specific
label is thrown, i.e., when the label fail is thrown, an error message can be generated
automatically by using the position where the failure occurred. This means that combining
labeled failures with the farthest failure position reduces the annotation burden, and helps
to identify the places in the parser where the a label is desirable.
In the next section, we discuss some applications of labeled failures: we can use labeled
PEGs to express the error reporting techniques that we have discussed in Section 3 [15, 18,
17, 19], and also to efficiently parse context-free grammars that can use the LL(∗) parsing
strategy [25].
6. Applications of Labeled Failures
This section shows that PEGs with labeled failures can express several error reporting
techniques used in the realm of parsing combinators. They can also efficiently parse contextfree grammars that are parseable by the LL(∗) top-down parsing strategy.
In Hutton’s deterministic parser combinators [15], the nofail combinator is used to
distinguish between failure and error. We can express the nofail combinators using PEGs
4
The second version of Typed Lua parser is available at https://github.com/sqmedeiros/lpeglabel/tree/master/ex
15
with labels as follows:
nofail p ≡ p / ⇑error
That is, nofail is an expression that transforms the failure of p into an error to abort
backtracking. Note that the error label should not be caught by any ordered choice.
Instead, the ordered choice propagates this label and catches solely the fail label. The
idea is that parsing should finish with one of the following values: success, fail or error.
The annotation of the Tiny grammar to use nofail is similar to the annotation we have
done using labeled failures. Basically, we just need to change the grammar to use nofail
instead of [p]l . For instance, we can write the rule CmdSeq as follows:
CmdSeq ← (Cmd (nofail SEMICOLON)) (Cmd (nofail SEMICOLON))∗
If we are writing a grammar from scratch, there is no advantage to use nofail instead
of more specific labels, as the annotation burden is the same and with nofail we lose more
specific error messages.
The cut combinator [18] was introduced to reduce the space inefficiency of backtracking
parsers, where the possibility of backtracking implies that any input that has already been
processed must be kept in memory until the end of parsing. Semantically it is identical to
nofail, differing only in the way the combinators are implemented: to implement cut the
parser combinators use continuation-passing style, so cut can drop the failure continuation
and consequently any pending backtrack frames. Hutton’s nofail is implemented in direct
style, and is not able to drop pending backtrack frames. Expressing a cut operator with the
same properties is not possible in our semantics of PEGs.
The four-values technique changed the semantics of parser combinators to implement
predictive parsers for LL(1) grammars that automatically identify the longest input prefix
in case of error, without needing annotations in the grammar. We can express this technique
using labeled failures by transforming the original PEG with the following rules:
JεK ≡ ⇑epsn
JaK ≡ a
JAK ≡ A
Jp1 p2 K ≡ Jp1 K (Jp2 K / ⇑error /{epsn} ε) /{epsn} Jp2 K
Jp1 /p2 K ≡ Jp1 K /{epsn} (Jp2 K / ⇑epsn ) / Jp2 K
(1)
(2)
(3)
(4)
(5)
This translation is based on three labels: epsn means that the expression successfully
finished without consuming any input, fail means that the expression failed without consuming any input, and error means that the expression failed after consuming some input.
In our translation we do not have an ok label because a resulting suffix means that the
expression successfully finished after consuming some input. It is straightforward to check
that the translated expressions behave according to the Table 1 from Section 3.
Parsec introduced the try combinator to annotate parts of the grammar where arbitrary
lookahead is needed. We need arbitrary lookahead because PEGs and parser combinators
16
usually operate at the character level. The authors of Parsec also showed a correspondence
between the semantics of Parsec as implemented in their library and Partridge and Wright’s
four-valued combinators, so we can emulate the behavior of Parsec using labeled failures by
building on the five rules above and adding the following rule for try:
Jtry pK ≡ JpK /{error} ⇑fail
(6)
If we take the Tiny grammar of Figure 1 from Section 2, insert try in the necessary
places, and pass this new grammar through the transformation J K, then we get a PEG that
automatically identifies errors in the input with the error label. For instance, we can write
the rule RepeatCmd as follows:
RepeatCmd ← (try REPEAT) CmdSeq UNTIL Exp
LL(∗) [25] is a parsing strategy used by the popular parsing tool ANTLR [26, 27] 5 .
An LL(∗) parser is a top-down parser with arbitrary lookahead. The main idea of LL(∗)
parsing is to build a deterministic finite automata for each rule in the grammar, and use
this automata to predict which alternative of the rule the parser should follow, based on the
rest of the input. Each final state of the DFA should correspond to a single alternative, or
we have an LL(∗) parsing conflict.
Mascarenhas et al. [6] shows how CFG classes that correspond to top-down predictive
parsing strategies can be encoded with PEGs by using predicates to encode the lookahead
necessary for each alternative. As translating a Deterministic Finite Automata (DFA) to a
PEG is straightforward [12, 6], this gives us one way of encoding an LL(∗) parsing strategy
in a PEG, at the cost of encoding a different copy of the lookahead DFA for each alternative.
Labeled PEGs provide a more straightforward encoding, where instead of a predicate
for each alternative, we use a single encoding of the lookahead DFA, where each final state
ends with a label corresponding to one of the alternatives. Each alternative is preceded by
a choice operator that catches its label.
To make the translation clearer, let us consider the following example, from Parr and
Fisher [25], where non-terminal S uses non-terminal Exp (omitted) to match arithmetic
expressions:
S → ID | ID ‘=’ Exp | ‘unsigned’ ‘∗’ ‘int’ ID | ‘unsigned’ ‘∗’ ID ID
After analyzing this grammar, ANTLR produces the DFA of Figure 5. When trying to
match S, ANTLR runs this DFA on the input until it reaches a final state that indicates
which alternative of the choice of rule S should be tried. For example, ANTLR chooses the
second alternative if the DFA reaches state s4 .
Figure 6 gives a labeled PEG that encodes the LL(∗) parsing strategy for rule S. Rules
S0 , S1 , and S2 encode the lookahead DFA of Figure 5, and correspond to states s0 , s1 ,
and s2 , respectively. The throw expressions correspond to the final states. As the throw
expressions make the input backtrack to where it was prior to parsing S0 , we do not need
to use a predicate. We can also turn any uncaught failures into errors.
5
The recently released version 4 of ANTLR uses adaptive LL(∗) as its parsing strategy.
17
s4 → 2
=
s1
EOF
ID
start
s0
s6 → 1
ID
unsigned
s2
ID
s5 → 4
unsigned int
int
s3 → 3
Figure 5: LL(∗) lookahead DFA for rule S
7. Conclusions
In this paper, we discussed error reporting strategies for Parsing Expression Grammars.
PEGs behave badly on the presence of syntax errors, because backtracking often makes the
PEG lose track of the position where the error happened. This limitation was already known
by Ford, and he tried to fix it in his PEG implementation by having the implementation
track the farthest position in the input where a failure has happened [3].
We took Ford’s failure tracking heuristic and showed that it is not necessary to modify a
PEG implementation to track failure positions as long as the implementation has mechanisms
to execute semantic actions, and the current parsing position is exposed to these actions.
In addition, we also showed how it is easy to extend the semantics of PEGs to incorporate
failure tracking, including information that can indicate what the PEG was expecting when
the failure happened.
Tracking the farthest failure position, either by changing the PEG implementation, using
semantic actions, or redefining the semantics of PEGs, helps PEG parsers produce error messages that are close to error messages that predictive top-down parsers are able to produce,
but these are generic error messages, sometimes with a long list of expected tokens.
As a way of generating more specific error messages, we introduced a mechanism of
labeled failures to PEGs. This mechanism closely resembles standard exception handling in
programming languages. Instead of a single kind of failure, we introduced a throw operator
⇑l that can throw different kinds of failures, identified by their labels, and extended the
18
S
S0
S1
S2
← S0 /1 ID /2 ID ‘=’ Exp /3 ‘unsigned’ ‘∗’ ‘int’ ID /4 ‘unsigned’ ‘∗’ ID ID / ⇑error
← ID S1 / ‘unsigned’ S2 / ‘int’ ⇑3
← ‘=’ ⇑2 / !. ⇑1 / ID ⇑4
← ‘unsigned’ S2 / ID ⇑4 / ‘int’ ⇑3
Figure 6: PEG with Labels that Simulates the LL(∗) Algorithm
ordered choice operator to specify the set of labels that it catches. The implementation of
these extensions in parser generator tools based on PEGs is straightforward.
We showed how labeled failures can be used as a way to annotate error points in a grammar, and tie them to more meaningful error messages. Labeled failures are orthogonal to the
failure tracking approach we discussed earlier, so grammars can be annotated incrementally,
at the points where better error messages are judged necessary.
We also showed that the labeled failures approach can express several techniques for error
reporting used in parsers based on deterministic parser combinators, as presented in related
work [15, 18, 17, 19]. Labeled failures can also be used as a way of encoding the decisions
made by a predictive top-down parser, as long as the decision procedure can be encoded as
a PEG, and showed an example of how to encode an LL(∗) grammar in this way.
Annotating a grammar with labeled failures demands care: if we mistakenly annotate
expressions that should be able to fail, this modifies the behavior of the parser beyond
error reporting. In any case, the use of labeled PEGs for error reporting introduces an
annotation burden that is lesser than the annotation burden introduced by error productions
in LR parsers, which also demand care, as their introduction usually lead to reduce-reduce
conflicts [28].
We showed the error reporting strategies in the context of a small grammar for a toy
language, and we also discussed the implementation of parsers for the Typed Lua language,
an extension of the Lua language, based on these strategies. Moreover, we also implemented
parsers for other languages, such as Céu [29], based on theses approaches, improving the
quality of error reporting either with generic error messages or with more specific error
messages.
References
[1] A. V. Aho, M. S. Lam, R. Sethi, J. D. Ullman, Compilers: Principles, Techniques, and Tools, 2nd
Edition, Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA, 2006.
[2] B. Ford, Parsing expression grammars: A recognition-based syntactic foundation, in: Proceedings of
the 31st ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages, POPL ’04,
ACM, New York, NY, USA, 2004, pp. 111–122.
[3] B. Ford, Packrat parsing: a practical linear-time algorithm with backtracking, Master’s thesis, Massachusetts Institute of Technology (September 2002).
[4] S. Medeiros, F. Mascarenhas, R. Ierusalimschy, From regular expressions to parsing expression grammars, in: Brazilian Symposium on Programming Languages, 2011.
19
[5] S. Medeiros, F. Mascarenhas, R. Ierusalimschy, Programming Languages: 16th Brazilian Symposium, SBLP 2012, Nat
Springer Berlin Heidelberg, Berlin, Heidelberg, 2012, Ch. Left Recursion in Parsing Expression Grammars, pp. 27–41. doi:10.1007/978-3-642-33182-4\_4.
URL http://dx.doi.org/10.1007/978-3-642-33182-4_4
[6] F.
Mascarenhas,
S.
Medeiros,
R.
Ierusalimschy,
On
the
relation
between
context-free
grammars
and
parsing
expression
grammars,
Science
of
Computer
Programmingdoi:http://dx.doi.org/10.1016/j.scico.2014.01.012.
[7] K. C. Louden, Compiler Construction: Principles and Practice, PWS Publishing Co., Boston, MA,
USA, 1997.
[8] D. Grune, C. J. Jacobs, Parsing Techniques: A Practical Guide, 2nd Edition, Springer Publishing
Company, Incorporated, 2010.
[9] B. Ford, Packrat parsing: Simple, powerful, lazy, linear time, in: Proceedings of the Seventh ACM
SIGPLAN International Conference on Functional Programming, ICFP ’02, ACM, New York, NY,
USA, 2002, pp. 36–47.
[10] I.
Piumarta,
peg/leg
—
recursive-descent
parser
generators
for
C,
http://piumarta.com/software/peg/, [Visited on March 2013] (2007).
[11] R.
Ierusalimschy,
LPeg
parsing
expression
grammars
for
Lua,
http://www.inf.puc-rio.br/~roberto/lpeg/lpeg.html, [Visited on March 2013] (2008).
[12] R. Ierusalimschy, A text pattern-matching tool based on parsing expression grammars, Software: Practice & Experience 39 (3) (2009) 221–258.
[13] K. Mizushima, A. Maeda, Y. Yamaguchi, Packrat parsers can handle practical grammars in mostly
constant space, in: Proceedings of the 9th ACM SIGPLAN-SIGSOFT Workshop on Program Analysis
for Software Tools and Engineering, PASTE ’10, ACM, New York, NY, USA, 2010, pp. 29–36.
[14] R. Grimm, Better extensibility through modular syntax, in: Proceedings of the 2006 ACM SIGPLAN
Conference on Programming Language Design and Implementation, PLDI ’06, ACM, New York, NY,
USA, 2006, pp. 38–51. doi:10.1145/1133981.1133987.
URL http://doi.acm.org/10.1145/1133981.1133987
[15] G. Hutton, Higher-order functions for parsing, Journal of Functional Programming 2 (3) (1992) 323–
343.
[16] M. Spivey, When maybe is not good enough, Journal of Functional Programming 22 (2012) 747–756.
doi:10.1017/S0956796812000329.
URL http://journals.cambridge.org/article_S0956796812000329
[17] A. Partridge, D. Wright, Predictive parser combinators need four values to report errors, Journal of
Functional Programming 6 (2) (1996) 355–364.
[18] N. Röjemo, Efficient parsing combinators, Tech. rep., Department of Computer Science, Chalmers
University of Technology (1995).
[19] D. Leijen, E. Meijer, Parsec: Direct style monadic parser combinators for the real world, Tech. Rep.
UU-CS-2001-35, Department of Computer Science, Utrecht University (2001).
[20] S. D. Swierstra, L. Duponcheel, Deterministic, error-correcting combinator parsers, in: Advanced Functional Programming, Vol. 1129 of Lecture Notes in Computer Science, Springer, 1996, pp. 184–207.
[21] S. D. Swierstra, Combinator parsers: a short tutorial, in: A. Bove, L. Barbosa, A. Pardo, J. Sousa Pinto
(Eds.), Language Engineering and Rigorous Software Development, Vol. 5520 of LNCS, Springer-Verlag,
2009, pp. 252–300.
[22] A. M. Maidl, F. Mascarenhas, R. Ierusalimschy, Typed Lua: An optional type system for Lua, in: Proceedings of the Workshop on Dynamic Languages and Applications, Dyla’14, ACM, New York, NY,
USA, 2014, pp. 3:1–3:10. doi:10.1145/2617548.2617553.
URL http://doi.acm.org/10.1145/2617548.2617553
[23] R. Ierusalimschy, Programming in Lua, 3rd Edition, Lua.Org, 2013.
[24] S. Medeiros,
LPegLabel - an extension of LPeg that supports labeled failures,
https://github.com/sqmedeiros/lpeglabel, [Visited on October 2015] (2014).
[25] T. Parr, K. Fisher, Ll(*): The foundation of the antlr parser generator, in: Proceedings of the 32Nd
20
[26]
[27]
[28]
[29]
ACM SIGPLAN Conference on Programming Language Design and Implementation, PLDI ’11, ACM,
New York, NY, USA, 2011, pp. 425–436. doi:10.1145/1993498.1993548.
URL http://doi.acm.org/10.1145/1993498.1993548
T. Par, The Definitive ANTLR 4 Reference, 2nd Edition, Pragmatic Bookshelf, 2013.
T. Parr, ANTLR, http://www.antlr.org, [Visited on March 2014].
C. L. Jeffery, Generating LR syntax error messages from examples, ACM Transactions on Programming
Languages and Systems (TOPLAS) 25 (5) (2003) 631–640.
F. Sant’Anna, Safe system-level concurrency on resource-constrained nodes with céu, Ph.D. thesis,
PUC–Rio (2013).
21
Appendix A. Formalization of Farthest Failure Tracking
In Section 2, we saw that a common approach to report errors in PEGs is to inform
the position of the farthest failure, and how we can use semantic actions to implement
this approach in common PEG implementations. In this appendix, we show a conservative
extension of the PEG formalism that expresses this error reporting approach directly, making
it a part of the semantics of PEGs instead of an ad-hoc extension to PEG implementations.
The result of a PEG on a given input is either fail or a suffix of the input. Returning a
suffix of the input means that the PEG succeeded and consumed a prefix of the input. PEGs
with farthest failure tracking return the product of the original result and either another
suffix of the input, to denote the position of the farthest failure during parsing, or nil to
denote that there were no failures.
Figure A.7 presents a formal semantics of PEGs with farthest failure tracking as a set
PEG
of inference rules. The notation G[p] xy
(y, v?) represents a successful match of the
parsing expression p in the context of a PEG G against the subject xy, consuming x and
PEG
leaving the suffix y. In a similar way, G[p] xy
(fail, v?) represents an unsuccessful
match. The term v? denotes a value that is either a suffix v of the original input or nil.
We cannot use an empty string to denote that there were no failures, as an empty string
already means that a failure has occurred after the last position of the input. The term X
denotes a value that is either a suffix of the current input or fail.
The auxiliary function smallest that appears on Figure A.7 compares two possible
error positions, denoted by a suffix of the input string, or nil if no failure has occurred,
and returns the furthest: any suffix of the input is a further possible error position than nil
and a shorter suffix is a further possible error position than a longer suffix.
Rule empty.1 deals with the empty expression. This expression never fails, so the failure
position is always nil. Rule var.1 deals with non-terminals, so it just propagates the result
of matching the right-hand side of the non-terminal. Rules char.1, char.2, and char.3 deal
with terminals. The latter two rules denote failures, so they return the subject as the failure
position.
Rules con.1 and con.2 deal with concatenation. The second rule just propagates the
failure position, but rule con.1 needs to take the farthest position between the two parts of
the concatenation. The rules for ordered choice (ord.1 and ord.2) and repetition (rep.1
and rep.2) work in a similar way: whenever there are two possible farthest failure positions,
we use smallest to take the farthest of them.
Finally, rules not.1 and not.2 deal with the syntactic predicate. Fitting this predicate
in this error reporting framework is subtle. The rules not only need to make sense for
the predicate in isolation, but also have to make sense for !!p, which works as a “positive
lookahead” predicate that succeeds if p succeeds but fails if p fails while never consuming
any input.
For not.1, the case where !p succeeds, we have two choices: either propagate p’s farthest
failure position, or ignore it (using nil). The first choice can lead to an odd case, where the
failure that made !p succeed can get the blame for the overall failure of the PEG, so not.1
takes the second choice.
22
Empty
G[ε] x
PEG
Terminal
(x, nil)
G[a] ax
PEG
G[b] ax
Concatenation
Ordered Choice
Repetition
PEG
(empty.1)
(x, nil)
(fail, ax)
G[p1 ] xy
PEG
(y, v?)
G[p1 p2 ] xy
G[p1 ] xy
PEG
G[p1 / p2 ] xy
G[p] x
PEG
G[p∗] x
Negative Predicate
(y, v?)
(fail, v?)
PEG
(x, v?)
G[p] x
PEG
G[!p] x
G[p2 ] y
(y, v?)
(rep.1)
(x, nil)
PEG
(X, w?)
(ord.1)
(fail, v?)
PEG
PEG
G[a] ε
(X, smallest(v?, w?))
PEG
G[A] x
PEG
PEG
(X, v?)
(X, v?)
(var.1)
(char.1)
, b 6= a (char.2)
PEG
G[P (A)] x
Non-terminal
G[p] xyz
(con.1)
G[p1 ] x
PEG
PEG
(yz, v?)
PEG
G[p1 ] x
PEG
G[p1 p2 ] x
PEG
(fail, v?)
PEG
(fail, v?)
G[p2 ] x
G[p∗] yz
PEG
PEG
PEG
(con.2)
(X, w?)
(X, smallest(v?, w?))
PEG
(z, w?)
(z, smallest(v?, w?))
G[p] xy
G[!p] xy
(char.3)
(fail, v?)
G[p1 / p2 ] x
G[p∗] xyz
(not.1)
(fail, ε)
(y, v?)
(fail, xy)
(rep.2)
(not.2)
Figure A.7: Semantics of PEGs with farthest failure tracking
We also have two choices for the case where !p fails: either propagate the failure position
from p or just use the current position. The first choice can also lead to an odd cases where
the overall failure of the PEG might be blamed on something in p that, if corrected, still
makes p succeed and !p fail, so not.2 also takes the second choice.
The end result is that what happens inside a predicate simply does not take part in error
reporting at all, which is the simplest approach, and also gives a consistent result for !!p.
As an example that shows the interaction of these rules, let us consider again the Tiny
program in Figure 1, reproduced below:
01
02
03
04
05
06
07
n := 5;
f := 1;
repeat
f := f * n;
n := n - 1
until (n < 1);
write f;
23
(ord.2)
The missing “;” at the end of line 5 makes the repetition (Cmd SEMICOLON)* inside
CmdSeq succeed through rule rep.1 and propagate the failure position to the concatenation
inside this same non-terminal. Rule con.2 then propagates the same failure position to the
concatenation inside RepeatCmd.
Another failure occurs in RepeatCmd when until does not match n, but rule con.2 again
propagates the position of the missing “;”, which is farthest. This leads to a failure inside
the repetition on CmdSeq again, which propagates the position of the missing “;” through
rule rep.2. Finally, rules con.2 and var.1 propagate this failure position to the top of the
grammar so the missing semicolon gets the blame for the PEG not consuming the whole
program.
We can translate the failure suffix into a line and column number inside the original
subject, and extract the first token from the beginning of this suffix to produce an error
message similar to the error messages in Section 2:
factorial.tiny:6:1: syntax error, unexpected ’until’
While this message correctly pinpoints the location of the error, it could be more informative. Appendix B shows how we can extend the semantics of Figure A.7 to gather
more information than just the farthest failure position, thus making us able to generate
error messages as informative as the ones produced by Pappy, which we discussed in the
end of Section 2. In the next section, we will introduce another approach for error reporting
in PEGs, which can produce more precise error messages, at the cost of annotating the
grammar.
Appendix B. Formalization of Farthest Failure Tracking and Error Lists
In this appendix we show how the formalization of the farthest failure tracking from
Appendix A can easily be adapted to build a more elaborate error object, instead of just
returning a position.
The new semantics formalizes a strategy similar to the one used by Ford [3] in his PEG
implementation. The basic idea is to keep a list of the simple expressions that the PEG
was expecting to match when a failure occurred, and use this list to build an error message.
PEG
Figure B.8 gives inference rules for the
relation of this new semantics.
The result of the matching of a PEG G against an input x is still a pair, but the second
component is now another pair, the error pair. The first component of the error pair is the
farthest error position, same as in the previous semantics. The second component is a list
of parsing expressions that were expected at this error position. If the grammar does not
have syntactic predicates, the expressions in this list are just terminals and non-terminals.
Rules empty.1 and char.1 do not change, and rules char.2 and char.3 just return a
list with the terminal that they tried to match.
24
Empty
G[ε] x
PEG
Terminal
G[A] x
(X, joinVar((v?, L), x, A))
G[a] ax
PEG
(x, (nil, {}))
PEG
PEG
G[p1 p2 ] x
(fail, (ax, {b}))
G[p1 ] xy
(fail, (v?, L))
G[p1 / p2 ] xy
(fail, (v?, L1 ))
G[p2 ] y
(y, (v?, L))
PEG
(y, (v?, L))
G[p2 ] x
PEG
G[p] x
PEG
(fail, (v?, L))
G[p] xyz
PEG
(x, (v?, L))
(yz, (v?, L1 ))
G[p∗] xyz
PEG
Negative Predicate
G[p∗] yz
PEG
G[!p] x
(char.3)
(con.1)
PEG
(z, (w?, L2 ))
(fail, (v?, L))
PEG
(ord.2)
(rep.1)
(z, join((v?, L1 ), (w?, L2 )))
G[p] x
(X, (w?, L2 ))
(X, (w?, L2 ))
(X, join((v?, L1 ), (w?, L2 )))
G[p∗] x
(fail, (ε, {a}))
(ord.1)
G[p1 / p2 ] x
PEG
PEG
PEG
(X, join((v?, L1 ), (w?, L2 )))
PEG
Repetition
G[a] ε
(con.2)
PEG
G[p1 ] xy
, b 6= a (char.2)
PEG
(var.1)
(char.1)
(y, (v?, L1 ))
(fail, (v?, L))
PEG
PEG
PEG
G[p1 p2 ] xy
Ordered Choice
G[p1 ] x
(X, (v?, L))
PEG
Concatenation
G[p1 ] x
PEG
G[P (A)] x
Non-terminal
G[b] ax
(empty.1)
(x, (nil, {}))
(x, (nil, {}))
(not.1)
(rep.2)
G[p] xy
G[!p] xy
PEG
PEG
(y, (v?, L))
(fail, (xy, {!p}))
Figure B.8: Semantics of PEGs with farthest failure tracking and error lists
25
(not.2)
Rule var.1 uses an auxiliary function joinVar, defined as follows:
joinVar((x, L), x, A) = (x, {A})
joinVar((nil, {}), x, A) = (nil, {})
joinVar((v, L), x, A) = (v, L)
where x 6= v
The idea behind joinVar is simple: if the right-hand side of terminal A has a farthest
error position that is the same as the current position in the input, it means that we can
treat this possible error as an error while expecting A itself. This is expressed in the first
case of joinVar by replacing the list of expected expressions L with just {A}. The other
cases just propagate the failure information returned by P (A).
To further understand how joinVar works, let us consider an example, adapted from
the grammar in Figure 1:
Factor ← “(” Exp “)” / Digit Digit∗
Digit ← “0” / “1” / “2” / “3” / “4” / “5” / “6” / “7” / “8” / “9”
When we try to match Factor against an input that does not match either “(” or Digit,
such as id, both alternatives of the ordered choice fail and the farthest failure position of
both matches is the same input that we were trying to match with Factor. So, instead of
keeping an error list that would later give us an error message like the following one:
Unexpected ’id’, expecting ’(’ or ’0’ or ’1’ or ’2’ or ’3’ or ’4’ or ’5’
or ’6’ or ’7’ or ’8’ or ’9’
Or even an error message like this one:
Unexpected ’id’, expecting ’(’ or Digit
We replace the error list built during the matching of the right-hand side of Factor with
just {Factor}. From this new list we can get the following higher-level error message:
Unexpected ’id’, expecting Factor
If the failure occurred in the middle of the first alternative of Factor (for example, because
of a missing “)’, we would keep the original error list instead of replacing it with {Factor}.
Using the names of the non-terminals in the error message instead of a list of terminals
that might have started this non-terminal might not be better if the names of the nonterminals give no indication to the user of what was expected. It is easy to replace joinVar
with just (v?, L), propagating the failure information from the expansion of the non-terminal
and keeping the error message just with symbols that the user can add to the input. Another
possibility is to have two kinds of non-terminals, and use joinVar for informative nonterminals and simple propagation for the others.
In the case of concatenation, a failure in the first expression of the concatenation means
we just need to propagate the error list, so case con.2 is straightforward. If both expressions
26
succeed, we might need to merge both error lists, if the two parts of the concatenation have
the same failure position. Rule con.1 uses the auxiliary join function, defined below:
join((x, L), (nil, {})) = (x, L)
join((nil, {}), (x, L)) = (x, L)
join((xy, L1 ), (y, L2)) = (y, L2 )
join((y, L1 ), (xy, L2)) = (y, L1 )
join((x, L1 ), (x, L2 )) = (x, L1
where x 6= ε
where x =
6 ε
L2 )
The first four cases of join keep the furthest and associated set of expected expressions.
In the first two cases, nilmeans that no error has occurred, so any error automatically
becomes the farthest error position. In the next two cases, one of the furthest error positions
is a strict suffix of the other, so is the furthest of the two and the other is discarded. The
only remaining possibility, expressed in the last case, is that the two positions are identical,
and we merge their expected sets.
The rules for ordered choice and repetition use the same join function, where applicable.
Finally, the rules for the syntactic predicate !p also ignore the error information inside p.
Rule not.2 blames the failure on the predicate itself.
Going back to our running example (Figure 2), our error tracking semantics will give an
error list that lets us generate the following error message:
factorial.tiny:6:1: syntax error, unexpected ’until’,
expecting ’;’, ’=’, ’<’, ’-’, ’+’, ’/’, ’*’
The operators also end up in the error list because their lexical rules all fail in the same
position as the semicolon. This error message is similar to the error message we get using
the error tracking combinators that we implemented on LPeg, at the end of Section 2.
We might try to tweak the error tracking heuristics of repetition and ordered choice to
ignore errors that happen in the first symbol of the input, which would let us take out
the operators from the error list in our previous example, and give a more succinct error
message:
factorial.tiny:6:1: syntax error, unexpected ’until’, expecting ’;’
This heuristic is not sound in the general case, though. Suppose we replace line 6 of
Figure 2 with the following line:
6 n ; until (n < 1);
The tweaked heuristic would still produce an error list with just the semicolon, which is
clearly wrong, as the problem is now a missing operator.
It is common in PEGs to mix scanning and parsing in the same grammar, as syntactic
predicates and repetition make lexical patterns convenient to express. But this can lead to
problems in the automatic generation of error messages because failures are expected while
27
recognizing a token, and these failures related to scanning can pollute the error list that we
use to generate the error message.
As an example, let us consider the lexical rule THEN from the PEG for the Tiny language
of Section 2:
THEN ← then !IDRest Skip
The pattern in the right-hand side fails if any alphanumeric character follows then in
the input, putting the predicate in the error list. The error will be reported to the user as
an unexpected character after then instead of a missing then keyword.
One solution is to split the set of non-terminals into lexical and non-lexical non-terminals.
Non-lexical non-terminals follow rule var.1, but lexical terminals follow a pair of new rules:
G[P (A)] xy
G[A] xy
PEG
G[P (A)] x
G[A] x
PEG
(y, (nil, {}))
PEG
PEG
(y, (v?, L))
(lvar.1)
(fail, (v?, L))
(fail, (x, {A}))
(lvar.2)
Intuitively, a lexical non-terminal reports errors just like a terminal. Any failures that
happen in a successful match are ignored, as they are considered to be expected, and a
failed match of the whole non-terminal gets blamed on the non-terminal itself, at its starting
position in the input.
All the extensions to the semantics of PEGs that we discussed in this section are conservative: if a PEG fails with some subject in the original semantics, it will also fail in all
of our extensions, and if a PEG matches some subject and leaves a suffix, it will also match
and leave the same suffix in all of our extensions. This is stated by the following lemma,
PEG
where we use the symbol
to represent the regular semantics of PEGs [6], and the symbol
PEGF
to represent the extended semantics of PEGs presented in Figure A.7:
Lemma Appendix B.1 (Conservativeness of farthest failure tracking). Given a PEG G, a
PEG
PEGF
parsing expression p and a subject xy, we have that G[p] xy
y iff G[p] xy
(y, v?),
PEG
PEGF
and that G[p] xy
fail iff G[p] xy
(fail, v?).
Proof. By induction on the height of the respective proof trees. The proof is straightforward,
PEG
PEGF
since the only difference between the rules of
and
is presence of the farthest failure
position, but this position has no influence on whether the expression successfully consumes
part of the input or fails.
A similar Lemma for the semantics with expected expression lists of Figure B.8 is also
PEGL
straightforward to prove, where we use the symbol
to represent those extended semantics:
Lemma Appendix B.2 (Conservativeness of farthest failure tracking with expected lists).
PEG
Given a PEG G, a parsing expression p and a subject xy, we have that G[p] xy
y iff
PEGL
PEG
PEGL
G[p] xy
(y, (v?, L)), and that G[p] xy
fail iff G[p] xy
(fail, (v?, L)).
28
Proof. By induction on the height of the respective proof trees. The proof is also straightPEG
PEGL
forward, since the only difference between the rules of
and
is presence of the farthest
failure position and list of expected expressions, but this extra information has no influence
on whether the expression successfully consumes part of the input or fails.
29
| 6cs.PL
|
Automatic estimation of the regularization parameter in 2-D
focusing gravity inversion: an application to the Safo manganese
mine in northwest of Iran
Saeed Vatankhaha , Vahid E Ardestania , Rosemary A Renautb,∗
a
arXiv:1310.0068v2 [cs.CE] 10 Jan 2014
b
Institute of Geophysics, University of Tehran,Tehran, Iran
School of Mathematical and Statistical Sciences, Arizona State University, Tempe, AZ 85287-1804, USA
Abstract
We investigate the use of Tikhonov regularization with the minimum support stabilizer for
underdetermined 2-D inversion of gravity data. This stabilizer produces models with nonsmooth properties which is useful for identifying geologic structures with sharp boundaries.
A very important aspect of using Tikhonov regularization is the choice of the regularization
parameter that controls the trade off between the data fidelity and the stabilizing functional.
The L-curve and generalized cross validation techniques, which only require the relative
sizes of the uncertainties in the observations are considered. Both criteria are applied in an
iterative process for which at each iteration a value for regularization parameter is estimated.
Suitable values for the regularization parameter are successfully determined in both cases
for synthetic but practically relevant examples. Whenever the geologic situation permits,
it is easier and more efficient to model the subsurface with a 2-D algorithm, rather than
to apply a full 3-D approach. Then, because the problem is not large it is appropriate to
use the generalized singular value decomposition for solving the problem efficiently. The
method is applied on a profile of gravity data acquired over the Safo mining camp in MakuIran, which is well known for manganese ores. The presented results demonstrate success in
reconstructing the geometry and density distribution of the subsurface source.
Keywords: Gravity inversion, manganese exploration, regularization parameter, L-curve
criterion, generalized cross validation, generalized singular value decomposition
PACS: 93.30.Bz,, 93.85.Hj
1. Introduction
Gravity inversion reconstructs models of subsurface density distribution using measured
data on the surface. There are two important ambiguities in the inversion of gravity data.
Corresponding Author: Rosemary Renaut, 480 965 3795
Email addresses: [email protected] (Saeed Vatankhah), [email protected] (Vahid E Ardestani),
[email protected] ( Rosemary A Renaut)
∗
Preprint submitted to Journal of Geophysics and Engineering
January 13, 2014
Theoretical ambiguity is caused by the nature of gravity; many equivalent sources in the subsurface can produce the same data at the surface. Algebraic ambiguity occurs when parameterization of the problem creates an underdetermined situation with more unknowns than
observations. Then, there is no unique density distribution which satisfies the observed data.
Further, the measurement process at the Earth’s surface is necessarily error-contaminated
and such errors can introduce arbitrarily large changes in the reconstructed solutions; namely
the solutions are sensitive to errors in the measurements. Thus, the inversion of gravity data
with under sampling is a typical example of an ill-posed problem that requires the inclusion
of a priori information in order to find a feasible reconstruction.
Tikhonov regularization is a well-known and well-studied method for stabilizing the solutions of ill-posed problems, (Hansen, 1998; Vogel, 2002, e.g.). The objective function of
the Tikhonov formulation includes a data fidelity (misfit term), and a stabilizing term that
controls the growth of the solution with respect to a chosen weighted norm. The choice of the
weighting for the regularization term impacts the properties of the solution. For example, a
smoothing stabilizer which employs the first or second derivative of the model parameters,
such as used in (Li and Oldenburg, 1996, 1998, e.g.) produces smooth images of the subsurface density distribution. There are, however, situations in which the potential field sources
are localized and have material properties that vary over relatively short distances. Then, a
regularization that does not not penalize sharp boundaries should be used. Last and Kubik
(1983) presented a compactness criteria for gravity inversion that seeks to minimize the volume of the causative body. This concept was developed by introducing minimum support
(MS) and minimum gradient support (MGS) stabilizers Portniaguine and Zhdanov (1999),
Zhdanov ( 2002), which are applied iteratively, generating repeatedly updated weighted
quadratic stabilizers.
In any regularization method, the trade off between the data fit and the regularization
term is controlled by a regularization parameter. Methods to find this regularization parameter, called parameter-choice methods, can be divided into two classes Hansen (1998): (i)
those that are based on knowledge of, or a good estimate of, the error in the observations,
such as Morozov’s discrepancy principle (MDP), and (ii) those that, in contrast, seek to
extract such information from the observations, such as the L-curve or generalized crossvalidation (GCV) methods. The use of the MDP is dominant in papers related to potential
field inversion,(Li and Oldenburg, 1996, 1998, e.g.), and the original paper for focusing inversion Portniaguine and Zhdanov (1999). In many practical applications, little knowledge
about the noise or error in the data measurements is available. The MDP then reduces to a
trial and error procedure for finding the optimal regularization parameter Li and Oldenburg
(1999). Here we discuss the use of the L-curve and GCV methods for use in focusing inversion
of gravity data in situations in which there is information about the relative magnitudes of
the standard deviations across the measured data Farquharson and Oldenburg (2004). Due
to the iterative nature of the algorithm, the regularization parameter is determined each
iteration.
Depending on the type of problem to be tackled, gravity inversion can be carried out
either in two or three dimensions (2-D or 3-D). 2-D methods are suitable for the recovery
2
of geologic structures such as faults, dikes and rift zones for which the length of the source
body in one direction is much longer than its extension in other directions. Then, it may be
possible to consider the gravitational sources as completely invariant in the direction parallel
to the length direction. Additionally, 2-D sources are both easier to conceptualize and model
than their 3-D counterparts, Blakely (1996).
The outline of this paper is as follows. In Section 2 we review the derivation of the
analytic calculation of the gravity anomaly derived from a 2-D cell model and then present
an overview of numerical methods for focusing inversion. The use, and rationale for the use,
of the generalized singular value decomposition, Paige and Saunders ( 1981), for the solution
is given in Section 2.2.1. The L-curve and GCV parameter-choice methods are discussed in
Section 2.3. Results for synthetic data are illustrated in Section 3. The approach is applied
on a profile of gravity data acquired from the Safo mining camp in Maku-Iran in Section 4.
Future directions and conclusions follow in Section 5.
2. Gravity modeling
2.1. The theoretical model
A simple 2-D model is obtained by dividing the subsurface under the survey area in
to a large number of infinitely long horizontal prisms in the invariant y−direction, with
variations in densities only assumed for the x and z directions. The cross-section of the
subsurface under the gravity profile for the model is shown in Figure 1 in which the cells
have square cross section and unknown densities. The dimensions of the cells are equal to
the distances between two observation points and the unknown density is considered to be
constant for each block. This type of parameterization, originally used by Last and Kubik
(1983), could indeed be improved. Yet, increasing the resolution of the models, and hence
the number of parameters, by dividing the subsurface into smaller cells, makes the problem
more ill-posed. Here the unknown density is considered to be constant for each block and the
data and model parameters are linearly related. The vertical component of the gravitational
attraction gi of a two-dimensional body at the origin using Cartesian coordinates is given
by, Blakely (1996),
Z Z ′ ′ ′
z dx dz
gi = 2Γρ
.
(1)
x′ 2 + z ′ 2
Here Γ is the universal gravitational constant and the density ρ is assumed to be constant
within the body. A solution of this integral for an ℓ-sided polygon is given by, Blakely (1996),
ℓ
X
gi
rp+1
νp
log
= 2Γ
− ψp (θp+1 − θp ) ,
ρj
1 + ψp2
rp
p=1
(2)
where ψp = (xp+1 − xp )/(zp+1 − zp ) and νp = xp − ψp zp . The variables rp , rp+1 , θp , and θp+1
are as displayed for the upper side of a square block in Figure 1. The term on the right-hand
side of (2), which quantifies the contribution to the ith datum of a unit density in the jth
3
Figure 1: The cross-section of the subsurface under the gravity profile. Gravity stations are located at the
centers of the blocks at the ground surface, indicated by the ∆ symbols. The cells are square and their
dimensions are equal to the distances between two observation points. Each cell extends as an infinitely long
prism in the invariant y−direction.
cell, is denoted by the kernel weight Gij , and is valid only at station i for cell j. The total
response for station i is obtained by summing over all cells giving
gi =
n
X
Gij ρj ,
i = 1, . . . , m,
j=1
m ≤ n,
(3)
which leads to the matrix equation
d = Gm + e,
(4)
where we have used the standard notation that vector d is the set of measurements given by
the gi , m is the vector of unknown model parameters, here the densities ρj , and e represents
the error in the measurements. The purpose of the gravity inverse problem is to find a
geophysically plausible density model that reproduces d.
2.2. Numerical approaches for focusing inversion
The conventional method for solving ill-posed inverse problems as described by (4) is
based on minimization of the parametric functional
P α (m) = φ(d) + α2 S(m).
(5)
Here φ(d) measures the data fidelity, which is usually measured by the weighted discrepancy
φ(d) = kWd (d − dobs )k22 ,
(6)
where d = Gm is the vector of predicted data, dobs contains the observations and Wd is a
data weighting matrix. Under the assumption that the noise e is Gaussian and uncorrelated,
4
Wd = diag(1/σ1 , . . . , 1/σm ) where σi is the standard deviation of the noise in the ith datum. Following Farquharson and Oldenburg (2004), rather than always assuming that the
absolute magnitudes of the error are known, we assume that the relative magnitudes of the
error can be estimated. Then for an unknown σ0 each σi is reexpressed as σi = σ0 σ̃i and
Wd = diag(1/σ̃1 , . . . , 1/σ̃m ). In (5) S(m) is a stabilizing regularization functional and α is a
regularization parameter.
Different choices are possible for S(m). Here we use the minimum support (MS) stabilizer
introduced in Last and Kubik (1983) and used in Portniaguine and Zhdanov (1999). This
stabilizer generates a compact image of a geophysical model with sharp boundaries and,
following Zhdanov ( 2002), is of the form:
SMS (m) = (m − mapr )T (m̂ − m̂apr )2 + ǫ2 I)
= (m − mapr )T We2 (m − mapr ),
−1
(m − mapr )
We = (m̂ − m̂apr )2 + ǫ2 I
(7)
−1/2
,
(8)
where m̂ and m̂apr are diagonal matrices of the current model parameters m and an estimate
of the model parameters mapr . If good prior knowledge of the properties of the subsurface
distribution exists, a full model of the expected physical properties can be used for mapr ,
otherwise it is often set to 0. In We , ǫ ≥ 0 is a focusing parameter that is introduced to
provide stability as m → mapr component wise. Small values for ǫ lead to compact models
but also increase the instability in the solution. For large ǫ the image will not be focused.
In general we are interested in the case where ǫ → 0. A trade-off curve method can be used
to select ǫ, Zhdanov and Tolstaya (2004); Ajo et al (2007); Vatankhah et al (2013).
It is well known that in potential data inversion the reconstructed models tend to concentrate near the surface. The depth weighting matrix, Wdepth = 1/(zj + ζ)β introduced in
Li and Oldenburg (1996, 1998); Pilkington (1997) can then be incorporated into the stabilizer term. Here zj is the mean depth of cell j and ζ > 0 is a small number imposed to
avoid singularity at the surface. The choice for β is important. Small β provide a shallow
reconstruction for the solution and large values concentrate the solution at depth. For all
inversions considered here β = 0.6 is selected. The hard constraint matrix Whard is also
important for the inversion process. If field operation geological and geophysical information
are able to provide the value of the density of cell j this information should be included
in mapr . Then, Whard is the diagonal identity matrix but with (Whard )jj = 100 for those
diagonal entries j for which information on the density is known.
Combining the diagonal weighting matrices, Whard , We and Wdepth , (5) is replaced by
P α (m) = kWd (Gm − dobs )k2 + α2 kD(m − mapr )k2 ,
D = We Whard Wdepth .
(9)
To obtain solution m := arg minm P α (m) linear transformation of the original model parameters is introduced via m(α) = m − mapr . Then m(α) solves the normal equations
(GT Wd2 G + α2 D T D)m(α) = GT Wd2 (dobs − Gmapr ) providing
m = mapr + m(α).
5
(10)
(11)
When (9) is solved iteratively due to the dependence of We on k, we use mapr = m(k−1) ,
(k)
m(0) = 0 and D (k) = We Whard Wdepth with
We(k) = (m̂(k−1) − m̂(k−2) )2 + ǫ2 I
−1/2
k > 1,
We(1) = I.
(12)
Then m(k) is obtained from the iteratively regularized equation, with updated regularization
parameter α(k) ,
(GT Wd2 G + (α(k) )2 (D (k) )T D (k) )m(α(k) ) = GT Wd2 (dobs − Gm(k−1) ) yielding
m(k) = m(k−1) + m(α(k) ).
(13)
(14)
This technique in which the weighting in We is frozen at each iteration, creating the possibility
to solve using the standard Tikhonov update, was introduced in the context of focusing
inversion in Zhdanov ( 2002). Because the MS stabilizer tends to produce the smallest
possible anomalous domain we follow the approach of Portniaguine and Zhdanov (1999);
Boulanger and Chouteau (2001) to produce a reliable image of the subsurface when using
focusing inversion. Based on geologic information, upper and lower bounds, mmin ≤ mj ≤
mmax , can be determined for the model parameters. If during the iterative process a given
density value falls outside the bounds, projection is employed to force the value back to the
exceeded value, and a hard constraint is imposed at that cell via (Whard )jj = 100.
2.2.1. Numerical Solution by the Generalized Singular Value Decomposition
We now discuss the numerical procedure for finding the solution to (13). For large
scale problems, iterative methods such as conjugate gradients, or other Krylov methods
should be employed to find m(α(k) ), (Hansen, 1998, e.g.). For small scale problems it is
feasible to use the singular value decomposition (SVD) for the matrix G̃ = Wd G, when
matrix D is the identity. Otherwise the generalized singular value decomposition (GSVD),
Paige and Saunders ( 1981), is needed. But again is effective to use for this problem because
it facilitates efficient determination of the regularization parameter. We assume G̃ ∈ Rm×n ,
m < n, D ∈ Rn×n and N (G̃) ∩ N (αD) = 0, where N (G̃) is the null space of matrix G̃.
Then there exist orthogonal matrices U ∈ Rm×m , V ∈ Rn×n and a nonsingular matrix
X ∈ Rn×n such that G̃ = UΛX T , D = V MX T where Λ of size m × n is zero except for
entries 0 < Λ1,q+1 ≤ . . . Λm,n < 1 with q = n − m, M is diagonal of size n × n with entries
1 = M1,1 = . . . Mq,q > Mq+1,q+1 ≥ M2,2 ≥ . . . Mn,n > 0. The generalized singular values
of the matrix pair G̃, D are γi = λi /µi , where γ1 = · · · = γq = 0 < γq+1 ≤ · · · ≤ γn , and
ΛT Λ = diag(0, . . . 0, λ2q+1 , . . . , λ2n ), M T M = diag(1, . . . , 1, µ2q+1, . . . , µ2n ), and λ2i + µ2i = 1,
∀i = 1 : n, i.e. M T M + ΛT Λ = In .
Using the GSVD, introducing ui as the ith column of matrix U and r̃(k) = Wd (dobs −
6
Gm(k−1) ), we may immediately write the solution of (13) as
(k)
m(α ) =
(k)
m
n
X
uTi−q r̃(k) T −1
γi2
(X )i ,
γ 2 + (α(k) )2 λi
i=q+1 i
(k−1)
=m
+
n
X
i=q+1
fi
(15)
uTi−q r̃(k) T −1
γi2
(X )i , fi = 2
, q < i ≤ n, fi = 0, 1 ≤ i ≤ q,
λi
γi + (α(k) )2
(16)
is the ith column of the inverse of the matrix X T and fi are the filter factors.
where (X T )−1
i
Therefore the algorithm proceeds by first updating the matrix We at step k using (12),
calculating the GSVD for the matrix pair [G̃, D (k)] and then updating m(k) using (16) which
depends on α(k) .
Three criteria are used to terminate the iterative procedure. Following Farquharson and Oldenburg
(2004) the iteration is seen to have converged and is thus terminated when either (i) a suffi(k−1)
(k)
(k)
cient decrease in the functional is observed,√P α
− P α < τ (1 + P α ), or (ii) the change
in the density satisfies km(k−1 − m(k) k < τ (1 + km(k) k). If neither of these conditions is
satisfied by an upper limit on the number of iterations, the procedure is terminated without convergence as measured in this manner. The parameter τ is taken as τ = .01 for the
inversions considered here.
The remaining issue is the determination of the regularization parameter at each step of
the iteration. As noted, when a priori information in the form of the standard deviations
on the noise in the data is available, the MDP can be used to find α. Here, in the absence
of the exact information on the error in the data we investigate the use of the L-curve and
GCV methods to find α(k) for which the formulation using the GSVD is advantageous.
2.3. Regularization Parameter Estimation
2.3.1. The L-curve
The L-curve approach developed by Hansen (1992, 1998) for linear inverse problems is
a robust criterion for determining the regularization parameter. It is based on the tradeoff between the norm of the regularized solution and the norm of the corresponding fidelity
term residual as the regularization parameter varies. According to Hansen (1992, 1998) when
these two norms are plotted on a log-log scale, the curve has an L shape with an obvious
corner. This corner separates the flat and vertical parts of the curve where the solution is
dominated by regularization errors and perturbation errors, respectively. Picking αopt as the
α responsible for the corner point gives the optimal trade off between the two terms, and
the corresponding model is selected as the optimal solution. For α > αopt the regularized
solution does not change dramatically, while the residual does. In contrast, for α < αopt
the regularized solution increases rapidly with little decrease in the residual. Because of the
relation of αopt with the shape of the curve, Hansen (1998) recommends estimating αopt by
finding the maximum of the local curvature in the neighborhood of the dominant corner
of the plot. Although the L-curve technique can be robust for problems generating welldefined corners, it may not work so well in other cases. For an underdetermined problem the
7
recovered model can change more slowly with the degree of regularization, Li and Oldenburg
(1999), and the L-curve is thus smoother. This makes it difficult to find maximum point of
curvature of the curve. On the other hand, given the solution of the regularized problem in
terms of the GSVD, as in (15), finding the L-curve is relatively efficient and is thus one of
the conventional ways to estimate αopt .
It was shown by Farquharson and Oldenburg (2004) that the L-curve choice for λ at
early iterations may be too small which may lead to inclusion of excessive structure in the
model that needs to be eventually removed, hence requiring more iterations for the inversion.
Hence here we follow the approach suggested by Farquharson and Oldenburg (2004)
and
(k)
(k)
(k−1)
∗
impose a so-called cooling process in which α is given by α = max c α
, α where
∗
0.01 ≤ c ≤ 0.5 and α is the point of maximum curvature of the L-curve. Moreover,
choosing a relatively large value for α(1) improves the performance of the algorithm. We use
α(1) = max(γi )/mean(γi ) and c = 0.4 which work well for the presented inversion examples.
2.3.2. Generalized Cross Validation
The major motivation of using the GCV to find an optimal value for α is that a good
value should predict missing data values. Specifically, if an arbitrary measurement is removed
from the data set, then the corresponding regularized solution should be able to predict the
missing observation. The choice of α should be independent of an orthogonal transformation
of the data, Hansen (1998). The GCV functional is given by
P
k ni=q+1 (1 − fi )uTi−q r̃k2
kG̃m(α) − r̃k2
kG̃m(k) − d̃obs k2
P
P
=
, (17)
GCV (α) =
=
(m − ni=q+1 fi )2
(m − ni=q+1 fi )2
trace(Im − G̃G(α))2
where the final expressions follow immediately from the GSVD, with G(α) = (G̃T G̃ +
α2 D T D)−1 G̃T . Here the numerator is the squared residual norm and the denominator is
effectively the square of the number of degrees of freedom, Hansen (1998). It can be shown
that the value of α which minimizes the expected value of the GCV function is near the minimizer of the expected value of the predictive mean-square error, kGm(α) − dexact k2 , Hansen
(1998). Hence, finding the minimum for GCV (α) should lead to a reasonable estimate for α.
We have seen in our experiments that when the GCV does not fail, in which case it produces
a very small α, it usually leads to α which is slightly larger than the optimal value. Failure
occurs when GCV (α) is almost flat near the optimal alpha, leading to numerical difficulties
in computing its minimum. The cooling procedure, as described for the L-curve, is also applied to the GCV estimation of α(k) but with α∗ now chosen as the current minimum of the
GCV function. For clarity we summarize the steps of the inversion that are applied every
iteration in Algorithm 1. Here we state this for the L-curve method. The GCV solutions are
found equivalently but at all steps using the L-curve instead the GCV method is applied.
Algorithm 1. The steps taken for the inversion at every iteration assuming the use of the
L-curve to find the regularization parameter
1. Calculate the GSVD for matrix pair [G̃, D].
8
2. Calculate the solutions, and the associated L-curve function, for a range of α , the
optimal α is found using the L-curve.
3. The cooling process is implemented for deciding whether the obtained α from step 2
should be used or not.
4. With α from step 3, the model parameters are computed using equation (16).
5. Density limits are implemented on model parameters, from step 4 , and then We and
Whard are updated.
6. Data misfit, S(m) and P α (m) are computed for model parameters obtained from step
5.
7. If the termination criteria are satisfied the iteration terminates. Otherwise, the a priori
density model is set equal to the density model from step 5 and the iteration returns
to step 1.
3. Numerical Results: Simulated Model
We evaluated the use of the L-curve and GCV in focusing inversion, as described in
Sections 2.3.1 and 2.3.2, for several synthetic data examples. In these simulations data are
calculated at 50 stations with 10m spacing on the surface and the inversion is required at the
subsurface on a rectangular grid of 50 × 10 with cell size 10m, hence in this case m = 50 and
n = 500. The iterations are initialized with m(0) = 0, Whard = We = I and bound constraints
on the density are set such that 0gr/cm3 ≤ mj ≤ 1gr/cm3 . The focusing parameter is fixed
as ǫ = 0.02 for the inversions. It should be noted that the regularization parameter depends
on the choice for ǫ, a large value requires a larger value for α and generates a smoother
model. In these simulations the maximum number of iterations is set to 20.
The synthetic gravity data are generated in each case for the rectangular body that has
density contrast equal to 1gr/cm3 with an homogeneous background, Figure 2(a). In generating noise-contaminated data we use zero mean Gaussian noise with a standard deviation
σ̃i = (η1 (dexact )i + η2 kdexact k), as indicated in Figure 2(b) for η1 = .03 and η2 = .001. In
each case we calculated the χ2 measure of the actual noise in the noisy data and report the
data fidelity values, relative error k(mexact − m(K) )k2 /kmexact k2 , and final values α(K) for
each inversion in Table 1.
The density models obtained for the first data set, Figure 2(b) are presented in Figures 5(a) and 5(b) for the L-curve and GCV inversions, respectively. In each case the geometry and density of the reconstructed models are close to those of the original model, although
the inversion using the L-curve criterion is more focused. This feature was present in all the
examples we have analyzed; inversion using GCV always provides a smoother reconstruction
than that obtained using the L-curve. Figures 4(a)-4(b) demonstrate the progression of the
solutions with iteration k for the data fidelity φ(d(k) ), the stabilizer S(m(k) ), the parametric
(k)
functional P α (m(k) ), and regularization parameter, α(k) , again for the L-curve and GCV
respectively for this first example. In our experience the behavior indicated is consistent
when using the GCV to find the regularization parameter α; it generally decreases initially,
but then increases to converge to a fixed value by the maximum number of iterations. On
9
(a) Simulated Model
(b) Gravity anomaly contaminated by uncorrelated noise
Figure 2: In 2(a) the synthetic model of a body set in a grid of square cells each of size 10m, the density
contrast of the body is 1gr/cm3 . In 2(b) the gravity anomaly due to the synthetic model contaminated by
uncorrelated noise with η1 = 0.03 and η2 = .001. The exact anomaly indicated by the solid line and the
contaminated data by the symbols.
(a) L-curve
(b) GCV
Figure 3: Density model obtained from inverting the data of Figure 2(a) with MS stabilizer and noise level
with η1 = 0.03 and η2 = .001 using bounds on density 0gr/cm3 ≤ mj ≤ 1gr/cm3 . The regularization
parameter was found using in (a) the L-curve and in (b) the GCV.
the other hand, for the L-curve the progression of α(k) is more erratic, generally oscillating
in final iterations toward a converged value as shown in Figure 4(a). Manual intervention
may then be needed to force the overall convergence of the algorithm Ajo et al (2007). The
main problem for the L-curve, as mentioned in section section 2.3.1, is its smooth shape, that
makes it difficult to find the corner, namely the point of maximum curvature. To illustrate
we plot the L-curve for all iterations in Figure 5(a), showing that overall the approach is
successful, although at a given middle iteration the apparent corner is missed, Figure 5(c).
Still the starting and final iterations in Figure 5(b) and Figure 5(d) do find useful corners.
The same formulation was used to generate two further synthetic data sets, with quantitative results shown also in Table 1. Noise generated using η1 = 0.01 and η = 0.05 was
considered, with in both cases η2 = .001. The results of the inversions are illustrated in
Figures 6 and 7, respectively. We note from Table 1 that the fidelity values at convergence
(K)
(K)
are less than the initial χ2 measure of the noise, that always αL−curve < αGCV , but that
there is no fixed conclusion about the relation between the final relative errors and fidelity
estimates by the L-curve and GCV inversions. We conclude that both the GCV and the
10
(a) L-curve
(b) GCV
Figure 4: The data fidelity φ(d), the stabilizer S(m), the parametric functional P α (m), and the regularization parameter, α ,all plotted against iteration k. The regularization parameter was found using in (a) the
L-curve and in (b) the GCV.
L-curve are successful in providing reasonable solutions. It should be noted that increasing
the focusing parameter ǫ can yield solutions which are not focused, especially as shown in
Figure 7.
Table 1: Relative error and final values of the regularization parameter for the inverted models.
Figure
3
6
7
8
9a-b
9c-d
χ2
measure
51.23
47.26
40.03
51.23
51.23
51.23
α(K)
Relative
L-curve GCV L-curve
0.62
7.41
0.4270
0.87
14.15 0.3376
0.50
6.85
0.4014
0.63
8.20
0.7708
27.04 429.38 0.4264
27.23 441.86 0.4404
Error
Fidelity
GCV L-curve
0.4025 19.14
0.3506 26.27
0.4013 21.63
0.6910 17.21
0.4026 19.64
0.4069 17.95
φ(d)
GCV
21.93
24.96
13.61
20.80
16.77
16.10
To assess both the impact of the choice of the bounds on the convergence properties for
the solution and choice of the regularization parameter α, we investigated two additional
situations. First we implemented the same problem as given in figure 2, with noise η1 = 0.03
and η2 = 0.001, but inverted now with upper bounds on the density changed to 2gr/cm3 .
The results are illustrated in figure 8 and detailed as before in Table 1. We see that the solutions are more focused, as anticipated from the previous work of Portniaguine and Zhdanov
(1999) but the relative error overall is increased and the fidelity of the solution is also decreased. On the other hand, it is of greater interest for the purposes of this study to observe
that the performance of the parameter choice techniques is independent of the upper bound,
11
(a)
(b)
(c)
(d)
Figure 5: The L-curve for all iterations with in each case the circle showing point with maximum curvature.
In figures 5(b)-5(d) we show the individual curves at iterations 1, 10 and the final stage.
and the parameters are found stably independent of the constraint bounds. To determine
the necessity of using the MS instead of a smoothness stabilizer, we also considered the
results obtained using smoothness stabilizer, i.e. the We in (9) replaced with the approximation for the second derivative of the model parameters, for the situations in figures 3
and 8. These results are also detailed in Table 1 and illustrated in figures 9(a)-9(b) and
9(c)-9(d), respectively. They demonstrate the relative insensitivity to density limits of the
smoothness-stabilizer obtained solutions. On the other hand, the solutions lack the contrast
that is achieved using the MS regularization. Overall, the solutions obtained with GCV are
apparently more robust than those with the L-curve. It should be noted that using a non-ℓ2
measure of the derivative in geophysical inversion leads to a strongly piecewice constant, or
blocky, reconstruction with sharp jumps, see e.g. Farquharson and Oldenburg (1998).
4. Numerical Results: Practical Data
4.1. Geological Context
The data used for inversion was acquired over the Safo mining camp in Maku-Iran which
is well known for manganese ores. Geologically this area is located in the Khoy ophiolite
12
(a) L-curve
(b) GCV
Figure 6: Density model obtained from inverting the data of Figure 2(b) with MS stabilizer: and noise level
with η1 = 0.01 and η2 = .001. The regularization parameter was found using in (a) the L-curve and in (b)
the GCV.
(a) L-curve
(b) GCV
Figure 7: Density model obtained from inverting the data of Figure 2(b) with MS stabilizer and noise level
with η = 0.05 and η2 = .001. The regularization parameter was found using in (a) the L-curve and in (b)
the GCV.
zone, in the northwest of Iran. Some manganese and iron-manganese deposits are found
within sedimentary pelagic rocks and radiolarian cherts which are accompanied by Khoy
ophiolite, Imamalipour (2005). Most of these deposits have little reserve; the Safo deposit
is the only viable area distinguished so far for mining. In the Safo deposit, depositions of
manganese have been found to occur in different horizons within pelagic rocks. Mineralogically, pyrolusite, bixibite, braunite and hematite are the main minerals present in ore, of
which the pyrolusite is the dominant ore Imamalipour (2005), and Calcite with quartz and
barite present as minor phases. The banded, massive and disseminated textures are seen
in orebodies. Manganese content varies from 7.4% to 69.1% in different regions of the area
Imamalipour (2005).
4.2. Gravity anomaly
The area of the gravity survey extends between UTM coordinates [438276 438609] west
and [4342971 4343187] north, Z38. The gravity survey was performed by the gravity branch
of the institute of Geophysics, University of Tehran. The measurements were corrected for
effects caused by instruments and tidal drift, latitude, free air and the Bouguer correction
to yield the Bouguer gravity anomaly, Figure 10(a). The Bouguer anomaly displays extreme
magnitudes in the central of the area in the north-south direction, related to mineral oc13
(a) L-curve
(b) GCV
Figure 8: Density model obtained from inverting the first data set η1 = 0.03 and η2 = .001 with MS stabilizer
and density limits 0gr/cm3 ≤ mj ≤ 2gr/cm3 . The regularization parameter was found using in 8(a) the
L-curve; and in 8(b) the GCV.
currence which has high density contrast with the host rocks. This geologic structure is
therefore clearly suitable for using a 2-D algorithm. The residual anomaly was obtained
by subtracting the regional anomaly from the Bouguer anomaly using a polynomial fitting
method, Figure 10(b). One of the recommended steps in potential field inversion is upward
continuation of data to a height of half the thickness of the shallowest cell which removes near
surface effects without noticeably degrading the data. Figure 11 shows upward continuation
of the residual data up to 2.5m.
4.3. Inversion result
A profile of the anomaly (SA) consisting of 49 data measurements, sampled every 5m,
is chosen for inversion. The subsurface is divided into 49 × 15 square cells of size 5m,
hence in this case m = 49 and n = 735. Based on geological information Imamalipour
(2005), the background density is set to 2.8gr/cm3 and the density limits for the inversion
are 2.4gr/cm3 ≤ mj ≤ 4.7gr/cm3 . The maximum number of iterations was set to 20.
Each datum is assigned a Gaussian error as in the simulated cases, here with η1 = .05 and
η2 = .001. Figures 12a-b illustrate the reconstructed density model from the inversion of
profile SA using the L-curve and GCV methods for estimating the regularization parameter,
(K)
(K)
yielding αL−curve = 0.65 and αGCV = 4.87 respectively. Figures 13(a)-13(b) illustrate the
profile of the anomaly (SA) which is used for the inversion, indicated by the stars, and
the resulting values obtained from reconstructed models in figures 12(a)-12(b), denoted in
each case by the circles. Both solutions clearly represent the density contrast and geometry
for the occurrence of manganese ore, figures 12(a)-12(b). The horizontal extension of the
obtained model is about 30m and the vertical extension shows a depth interval approximately
between 5m and 35m. These results are close to those obtained by Borehole drilling on the
site; which show extension of manganese ores from 3−4m to 25−30m in the subsurface along
the north-south direction Noorizadeh (2010). The data fidelity, the stabilizer, the parametric
functional and regularization parameter, with iteration k , are shown in Figures 14(a)-14(b).
The convergence histories have properties for the practical data that are similar to those for
the simulated data sets.
14
(a) L-curve
(b) GCV
(c) L-curve
(d) GCV
Figure 9: Density model obtained from inverting the first data set η1 = 0.03 and η2 = .001 with smoothness
stabilizer. In 9(a) and 9(b) density limits are 0gr/cm3 ≤ mj ≤ 1gr/cm3 . In 9(c) and 9(d) density limits are
0gr/cm3 ≤ mj ≤ 2gr/cm3 . The regularization parameter was found using in 9(a) and 9(c) the L-curve; and
in 9(b) and 9(d) the GCV.
5. Conclusions
Tikhonov regularization with the minimum support stabilizer has been demonstrated to
yield non-smooth solutions and is thus an appropriate approach for recovery of geological
structures with sharp boundaries. The presented algorithm is flexible and allows variable
weighting in the stabilizer, including depth weighting, a priori density ranges for the domain
and the inclusion of hard constraints for the a priori information in the inversion process.
The L-curve criterion and GCV method for estimating the regularization parameter were
discussed, and characteristics of each of them for obtaining a solution were introduced.
Numerical tests using synthetic data have demonstrated feasibility of applying both methods
in the iteratively reweighted algorithm. The regularization parameter is seen to converge
as the number of iterations increases. Although GCV leads to inverse solutions which are
slightly smoother than those obtained by the L-curve, both recovered models are close to
the original model. For this small-scale problem, it is shown that the GSVD can be used in
the algorithm, demonstrating the filtering of the solution. Moreover, this use of the GSVD,
which might generally be assumed to be too expensive, is beneficial and worthwhile in the
context of regularization parameter estimation as shown here. For large-scale problems
it is anticipated that a randomized GSVD needs to be developed, along the lines of the
randomized SVD that was introduced in Liberty et al (2007). Finally the method was used
on a profile of gravity data acquired over the Safo manganese mine in the northwest of Iran.
The result shows a density distribution in subsurface from about 5m to 35m in depth and
about 30m horizontally. Future work will consider the inclusion of statistical weighting in
15
(a) L-curve
(b) GCV
Figure 10: Bouguer anomaly over the Safo manganese mine in 10(a) and the residual anomaly over the Safo
manganese mine in 10(b).
Figure 11: Upward continuation of the residual anomaly to height 2.5m.
the solution and the use of regularization parameter estimation using statistical approaches,
Mead and Renaut (2009).
6. Acknowledgements
Rosemary Renaut acknowledges the support of AFOSR grant 025717: “Development
and Analysis of Non-Classical Numerical Approximation Methods”, and NSF grant DMS
1216559: “Novel Numerical Approximation Techniques for Non-Standard Sampling Regimes”.
We would also like to thank the two anonymous referees who raised interesting questions
that lead us to include additional results and clarifications of the method.
16
(a) L-curve
(b) GCV
Figure 12: The density models obtained by inverting field gravity data (profile SA). The regularization
parameter was found using in (a) the L-curve and in (b) the GCV.
(a) L-curve
(b) GCV
Figure 13: The field gravity data (stars) and computed data for the reconstructed density model (circles).
The regularization parameter was found using in 13(a) the L-curve; in 13(b) the GCV.
References
References
Ajo-Franklin J B Minsley B J and Daley T M 2007 Applying compactness constraints to
differential traveltime tomography Geophysics 72 R67-R75
Blakely R J 1996 Potential Theory in Gravity & Magnetic Applications Cambridge University
Press Cambridge United Kingdom
Boulanger O and Chouteau M 2001 Constraints in 3D gravity inversion Geophysical Prospecting 49 2 265-280
Farquharson C G and Oldenburg D W 1998 Non-linear inversion using general measure of
data misfit and model structure Geophysical Journal International 134 213-227
Farquharson C G and Oldenburg D W 2004 A comparison of automatic techniques for estimating the regularization parameter in non-linear inverse problems Geophysical Journal
International 156 411-425
17
(a) L-curve
(b) GCV
Figure 14: The data fidelity φ(d), the stabilizer S(m), the parametric functional P α (m), and the regularization parameter, α ,all plotted against iteration k. The regularization parameter was found using in (a)
the L-curve and in (b) the GCV.
Hansen P C 1992 Analysis of discrete ill-posed problems by means of the L-curve SIAM
Review 34 561-580
Hansen P C 1998 Rank-Deficient and Discrete Ill-Posed Problems: Numerical Aspects of Linear Inversion SIAM Monographs on Mathematical Modeling and Computation 4 Philadelphia
Imamalipour A 2005 Geochemistry mineralogy and origin of Safo manganese deposit Ninth
meeting of Geological Society of Iran (in Persian)
Last B. J. and Kubik K 1983 Compact gravity inversion Geophysics 48 713-721
Li Y and Oldenburg D W 1996 3D inversion of magnetic data Geophysics 61 394-408
Li Y and Oldenburg D W 1998 3D inversion of gravity data Geophysics 63 109-119
Li Y and Oldenburg D W 1999 3D Inversion of DC resistivity data using an L-curve criterion
69th Ann. Internat. Mtg. Soc. Expl. Geophys. Expanded Abstracts 251-254
Liberty E Woolfe F Martinsson PG Rokhlin V and Tygert M 2007 Randomized algorithms for
the low-rank approximation of matrices Proceedings of the National Academy of Sciences
104 20167-20172
Mead J L and Renaut R A 2009 A Newton root-finding algorithm for estimating the regularization parameter for solving ill-conditioned least squares problems Inverse Problems
25 025002 doi: 10.1088/0266-5611/25/2/025002
Noorizadeh A 2010 Geological report, Bore-hole results of Safo manganese mine Spadana
Mining Company
18
Paige C C and Saunders M A 1981 Towards a generalized singular value decomposition
SIAM Journal on Numerical Analysis 18 3 398-405
Pilkington M 1997 3-D magnetic imaging using conjugate gradients Geophysics 62 1132-1142
Portniaguine O and Zhdanov M S 1999 Focusing geophysical inversion images Geophysics
64 874-887
Vatankhah S Ardestani E V and Ashtar J M 2013 A method for 2D inversion of gravity data
Journal of Earth and Space Physics (accepted)
Vogel C R 2002 Computational Methods for Inverse Problems SIAM Frontiers in Applied
Mathematics SIAM Philadelphia U.S.A.
Zhdanov M S 2002 Geophysical Inverse Theory and Regularization Problems Elsevier Amsterdam
Zhdanov M S and Tolstaya E 2004 Minimum support nonlinear parameterization in the
solution of 3−D magnetotelluric inverse problem Inverse Problems 30 937-952
19
| 5cs.CE
|
MATROIDS: A MACAULAY2 PACKAGE
arXiv:1511.04618v1 [math.CO] 14 Nov 2015
JUSTIN CHEN
Abstract. We give an overview of the Macaulay2 package Matroids, which contains functionality to create and
compute with matroids. Examples highlighting the use of all major functions in the package are provided, along
with explanations of some of the algorithms. Interesting connections of matroids to various branches and areas are
also given.
Introduction. A matroid is a combinatorial object which abstracts the notions of (linear algebraic, graphtheoretic) independence. Since their introduction by Whitney [8], matroids have found diverse applications in
combinatorics, graph theory, optimization, and algebraic geometry, in addition to being studied as interesting
objects in their own right.
One striking feature of matroids is the multitude of distinct ways to define them. This variety of equivalent (or
cryptomorphic) ways to characterize matroids is one of the greatest strengths of matroid theory, and one of the
reasons for its ubiquity. The first and foremost definition is via bases:
Definition. Let E be a finite set, and B ⊆ 2E a set of subsets of E. The pair (E, B) is a matroid if B satisfies the
following conditions:
(B1) B 6= ∅, and
(B2) (Basis exchange) For any B1 , B2 ∈ B and b1 ∈ B1 \ B2 , there exists b2 ∈ B2 \ B1 with B1 \ {b1 } ∪ {b2 } ∈ B.
E is called the ground set of the matroid M = (E, B), B is the set of bases, and an element of B is a basis. Any
subset of a basis is an independent set. A subset of E that is not independent is dependent. The minimal (with
respect to inclusion) dependent sets are circuits.
It is easy to see that the sets of bases, independent sets, and circuits all give the same amount of information;
and any one of these determines the other two: the independent and dependent sets partition 2E , and the bases
are precisely the independent sets which are maximal with respect to inclusion.
Important notes: For the reader already familiar with matroids: this package provides capabilities for converting
between various representations of matroids, forming linear and graphic matroids from a matrix or graph, creating
and detecting existence of minors, computing Tutte polynomials, and some additional functions for matroid applications in other areas. Each of these will in turn be illustrated with examples in this article. Virtually all notation
and results mentioned below are from Oxley [5].
In this package, a matroid is, by default, internally represented by its list of bases. Moreover, the ground set is
always internally treated as a list {0, . . . , n − 1}. Although the actual elements of the ground set can be arbitrary
(e.g. integers, column vectors, edges in a graph), all computations involving elements of the ground set are done
by referring to an element by its index in the ground set, which is an integer in {0, . . . , |E| − 1} (in Macaulay2,
enumeration starts at 0). For convenience, methods are provided to easily convert between an actual subset of the
ground set and its list of indices.
Basic properties. For the latest version of the package, see 1. We start with a minimal running example:
i1 : loadPackage "Matroids"
o1 = Matroids
o1 : Package
2010 Mathematics Subject Classification. 05-04, 05B35, 52B40, 05C31.
1https://github.com/jchen419/Matroids-M2
1
2
JUSTIN CHEN
i2 : M = matroid({a,b,c,d},{{0,1},{0,2}})
o2 = M
o2 : Matroid
This creates a matroid on the ground set {a, b, c, d} with 2 bases, whose indices (as subsets of the ground set)
are {0, 1} and {0, 2}. We can view the list of bases:
i3 : bases M
o3 = {set {0, 1}, set {0, 2}}
o3 : List
Notice that the bases are returned as sets of indices. To view the actual subsets of the ground set corresponding
to the indices, use the method groundSet (notice that groundSet converts sets to lists):
i4 : apply(bases M, B -> groundSet_M B)
o4 = {{a, b}, {a, c}}
o4 : List
To convert a subset of the ground set to its list of indices, use the method indicesOf, which is an inverse to
groundSet:
i5 : apply(oo, B -> indicesOf_M B)
o5 = {{0, 1}, {0, 2}}
o5 : List
So far, no attempt has been made to check that M is actually a matroid, i.e. the set of bases satisfies the basis
exchange property. We verify this now, and also give a non-example:
i6 : isValid M
o6 = true
i7 : isValid matroid({a,b,c,d},{{0,1},{2,3}})
o7 = false
We can obtain plenty of matroid-theoretic information for this example:
i8 : rk M -- the rank of M = size of any basis of M
o8 = 2
i9 : rk_M {0,3} -- {0,3} corresponds to the subset {a,d}
o9 = 1
i10 : circuits M
o10 = {set {1, 2}, set {3}}
o10 : List
i11 : isDependent_M {1}
o11 = false
i12 : independents(M, 2) -- independent sets of M of size 2
o12 = {set {0, 1}, set {0, 2}}
o12 : List
We also recall the following definitions:
Definition. A loop in M is a 1-element dependent set, and a coloop in M is an element contained in every basis.
The closure of a subset A ⊆ E is A := {x ∈ E | rk(A) = rk(A ∪ {x})}. A flat of M is a closed subset, i.e. A = A.
A hyperplane of M is a flat of rank equal to rk M − 1.
i13
o13
o13
i14
o14
o14
i15
o15
o15
i16
:
=
:
:
=
:
:
=
:
:
loops M
{3}
List
coloops M
{0}
List
closure_M {2,3}
{1, 2, 3}
List
hyperplanes M
MATROIDS: A MACAULAY2 PACKAGE
3
o16 = {set {0, 3}, set {1, 2, 3}}
o16 : List
i17 : flats M
o17 = {{set {3}}, {set {0, 3}, set {1, 2, 3}}, {set {0, 1, 2, 3}}}
o17 : List
i18 : fvector M
o18 = {1, 2, 1}
o18 : List
(here the f-vector records the number of flats of rank i, for i = 0, . . . , rk M - notice that f0 = frk M = 1 always,
as the set of loops is the unique flat of rank 0 and E is the unique flat of maximal rank). We remark that the
set of flats is closed under intersections, and thus (being finite, with a maximal element) is a lattice, with meet =
intersection, and join given by X ∨ Y := X ∪ Y .
Analogous to the basis exchange property, there are abstract axioms that characterize which sets of subsets of E
can be the set of circuits, flats, or hyperplanes of a matroid on E. Similarly, there are axioms characterizing which
functions 2E → Z≥0 and 2E → 2E are the rank functions and closure operators of a matroid on E, respectively.
Lattices of flats of matroids can also be characterized among all finite lattices. Any one of the following – bases,
circuits, rank function, closure operator, lattice of flats, hyperplanes – determines all others.
Constructing types of matroids. The simplest family of matroids are the uniform matroids, i.e. where the
set of bases equals all subsets of a fixed size:
i19 : U = uniformMatroid(2,4); bases U
o20 = {set {0, 1}, set {0, 2}, set {1, 2}, set {0, 3}, set {1, 3}, set {2, 3}}
o20 : List
Another family of fundamental importance is the class of linear matroids, which arise naturally from a matrix.
The columns of the matrix form the ground set, and a set of column vectors is declared independent if they are
linearly independent in the vector space of columns.
i21 : A = matrix{{0,4,-1,6},{0,2/3,7,1}}; MA = matroid A; groundSet MA
2
4
o21 : Matrix QQ <--- QQ
o23 = {0, | 4
|, | -1 |, | 6 |}
| 2/3 | | 7 | | 1 |
o23 : List
An abstract matroid M is called representable or realizable over a field k if M is isomorphic to a linear matroid
over k, where an isomorphism of matroids is a bijection between ground sets that induces a bijection on bases. We
verify that the matroid M we started with is isomorphic to M A, hence is representable over Q:
i24 : isomorphic(MA, M)
Isomorphism: {3, 1, 0, 2}
o24 = true
Note that the bijection giving the isomorphism is printed: here {3, 1, 0, 2} represents the permutation σ of
{0, 1, 2, 3} with σ(0) = 3, σ(1) = 1, σ(2) = 0, σ(3) = 2. This indicates that the map f : M A → M sending the ith
element of M A to the σ(i)th element of M is an isomorphism.
An important class of representable matroids (over any field) are the graphic matroids, derived from a graph.
If G is a graph, then the graphic matroid M (G) has ground set equal to the edge set of G, and circuits given by
cycles in G (a cycle in G is a connected subgraph where every vertex has degree 2).
This package uses the Graphs package to provide functionality for creating and obtaining information about
graphs. Conversely, the graphic matroid also captures a great deal of information about the graph, much of which
has (at the time of this writing) not yet been exposed by the Graphs package.
The first example of this is the set of cycles. As the graphic matroid is defined in terms of the cycles of the
graph, it was important for this package to have an efficient method of listing all cycles. To this end, two methods
are provided in this package: a local helper function which returns all closed walks of a given length centered at
a given vertex, and a global wrapper function which identifies oppositely oriented closed walks into a single cycle
4
JUSTIN CHEN
(each cycle through a vertex v gives rise to exactly 2 closed walks at v, by reversing orientation). The global
wrapper function also includes significant optimizations to speed up computation time, such as eliminating cases
with few cycles before starting a search, and recursively deleting all previously visited vertices.
In the end though, finding all cycles is an NP-complete problem (e.g. existence of a Hamiltonian cycle can be
simply read off from the list of cycles), so efficiency is – even theoretically – still limited. In real use cases, the
algorithm performs quite well for relatively small graphs; the sparser the better.
i25 : time #getCycles generalizedPetersenGraph(5,2)
-- used 0.0597035 seconds
o25 = 57
i26 : time #getCycles completeGraph 8
-- used 1.64614 seconds
o26 = 8018
i27 : K5 = completeGraph 5
o27 = Graph{0 => {1, 2, 3, 4}}
1 => {0, 2, 3, 4}
2 => {0, 1, 3, 4}
3 => {0, 1, 2, 4}
4 => {0, 1, 2, 3}
o27 : Graph
i28 : M5 = matroid K5; groundSet M5
o29 = {set {0, 1}, set {0, 2}, set {0, 3}, set {0, 4}, set {1, 2}, set {1, 3}, set {1, · · ·
o29 : List
i30 : #bases M5
o30 = 125
The above computation verifies Cayley’s theorem, which states that the number of spanning trees on n vertices
is nn−2 (note that the bases of the graphic matroid are precisely the spanning trees, when the underlying graph is
connected).
As one may suspect from the discussion on cycles, the graphic matroid is created from the list of cycles, i.e. by
specifying circuits. This can be done for an abstract matroid as well, using the optional argument EntryMethod
=> "circuits" in the constructor function. Regardless of the value of EntryMethod, the bases are automatically
computed upon creation. We recreate the matroid M from before, by specifying its circuits. Note that two matroids
are considered equal iff they have the same set of bases and same size ground sets, or equivalently, the identity
permutation is an isomorphism between them.
i31 : M1 = matroid({a,b,c,d},{{1,2},{3}}, EntryMethod => "circuits", TargetRank => 2)
o31 = M1
o31 : Matroid
i32 : M == M1
o32 = true
Often specific matroids of importance are close to uniform, in the sense that relatively few subsets of size rk M
are dependent, so the set of nonbases (i.e. dependent sets of size rk M ) can also be specified, using EntryMethod
=> "nonbases".
i33 : fanoNB = {{0,1,2},{0,4,5},{0,3,6},{1,3,5},{1,4,6},{2,3,4},{2,5,6}};
i34 : F7 = matroid(toList (1..7), fanoNB, EntryMethod => "nonbases"); #bases F7
o35 = 28
We can also construct a new matroid from smaller ones by taking direct sums: if M1 = (E1 , B1 ) and M2 =
(E2 , B2 ) are matroids, then their direct sum is M1 ⊕ M2 := (E1 ⊔ E2 , {B1 ⊔ B2 | B1 ∈ B1 , B2 ∈ B2 }).
i36 : S = directsum(U, matroid completeGraph 3); groundSet S
o37 = {(0, 0), (1, 0), (2, 0), (3, 0), (set {0, 1}, 1), (set {0, 2}, 1), (set {1, 2}, 1)}
o37 : List
i38 : bases S
o38 = {set {0, 1, 4, 6}, set {0, 1, 4, 5}, set {0, 1, 5, 6}, set {0, 2, 4, 6}, set {0, · · ·
MATROIDS: A MACAULAY2 PACKAGE
5
o38 : List
Notice that although elements of the ground set of the direct sum receive a placeholder index to ensure disjointness (and thus become increasingly more complex with repeated direct sums), the bases only change by shifting
the indices of the second argument up. This is one instance where the representation of bases as indices is more
useful than as elements.
A matroid that cannot be written as a direct sum of nonempty matroids is called connected or irreducible. Every
matroid is a direct sum of connected matroids, its connected components. To find the connected components, use
the method componentsOf:
i39 : C = componentsOf S
o39 = {Matroid, Matroid}
o39 : List
i40 : S == directsum(C#0, C#1)
o40 = true
Duality and minors. One of the most important features of matroid theory is the existence of a duality. It is
straightforward to check that if M = (E, B) is a matroid, then {E − B | B ∈ B} also satisfies (B1) and (B2), and
thus is the set of bases of a matroid on E, called the dual matroid of M , denoted by M ∗ .
i41 : D = dualMatroid M; bases D
o42 = {set {2, 3}, set {1, 3}}
o42 : List
i43 : M == dualMatroid D
o43 = true
i44 : loops D
o44 = {0}
o44 : List
i45 : circuits D
o45 = {set {1, 2}, set {0}}
o45 : List
Notice that loops of D are coloops of M , and circuits of D are complements of hyperplanes of M (this is in
fact how the method hyperplanes works). Indeed, virtually any matroid-theoretic property or operation can be
enriched by considering its dual version, which is often distinct, and/or interesting as well. One such operation is
deletion, which dualizes to contraction, which we discuss now:
Definition. Let M = (E, B) be a matroid, and S ⊆ E. The restriction of M to S, denoted M |S , is the matroid
on S with bases {B ∩ S | B ∈ B, |B ∩ S| = rk S}. The deletion of S, denoted M \ S, is the restriction of M to
E − S. The contraction of M by S, denoted M/S, is defined as (M ∗ \ S)∗ .
i46 : N1 = deletion_M {3}; (groundSet N1, bases N1)
o47 = ({a, b, c}, {set {0, 1}, set {0, 2}})
o47 : Sequence
i48 : N2 = contraction_M {1}; (groundSet N2, bases N2)
o49 = ({a, c, d}, {set {0}})
o49 : Sequence
A minor of M is any matroid which can be obtained from M by a sequence of deletions and contractions. It
is a fact that any minor of M is of the form (M/X) \ Y for X, Y disjoint subsets of E (in fact, even more can be
said about the structure of minors, as we shall see shortly).
i50 : minorM5 = minor(M5, {9}, {3,5,8}) -- contracts {9}, then deletes {3,5,8}
o50 = minorM5
o50 : Matroid
i51 : (groundSet minorM5, #bases minorM5)
o51 = ({set {0, 1}, set {0, 2}, set {0, 3}, set {1, 2}, set {1, 4}, set {2, 3}}, 16)
o51 : Sequence
6
JUSTIN CHEN
Minor-theoretic properties can be used to describe many important classes of matroids. For example, a class C
of matroids is said to be minor-closed if every minor of a matroid in C is again in C. The classes of uniform, krepresentable (for arbitrary k), and graphic matroids are all minor-closed. Furthermore, certain classes of matroids
can be characterized by their forbidden or excluded minors, namely the matroids not in the class, but with every
proper minor in the class.
Theorem 1. (Tutte) [6, 7] Let M be a matroid.
i) M is binary (i.e. representable over F2 ) iff M has no U2,4 minor (i.e. no minor of M is isomorphic to U2,4 ).
ii) M is regular (i.e. representable over any field) iff M has no U2,4 , F7 , or F7∗ minor.
iii) M is graphic iff M has no U2,4 , F7 , F7∗ , M (K5 )∗ , or M (K3,3 )∗ minor.
Here U2,4 is the uniform matroid of rank 2 on 4 elements, and F7 is the Fano matroid.
We illustrate this by verifying that M (K5 ) is regular (alternatively, one can simply write down a 0-1 matrix
representing M (K5 )):
i52 : hasMinor(M5, U)
o52 = false
i53 : hasMinor(M5, F7)
o53 = false
i54 : hasMinor(M5, dualMatroid F7)
o54 = false
A word about the algorithm behind the method hasMinor is in order. First, the problem of testing for minors
(even a fixed minor among a specific class of matroids) is NP-complete. Despite this, there are significantly fewer
checks necessary to determine existence of a minor than may initially seem. Indeed, every minor of M is in fact
of the form (M/I) \ I ∗ , where I, I ∗ are disjoint, I is independent, and I ∗ is coindependent (i.e. is independent in
M ∗ ). Such a minor has rank equal to that of M/I, which in turn equals rk M − |I|. Thus checking existence of
a minor N in M can be realized as a two-step process, where the first step contracts independent sets of M of a
fixed size down to the rank of N , and the second step deletes coindependent sets down to the size of N .
In the case of a successful minor test, the sets I and I ∗ are printed (as well as the isomorphism from (M/I) \ I ∗
to N as usual). Note that the indices of I, I ∗ are given with respect to the ground set of M (as opposed to giving
the indices of I ∗ with respect to the ground set of M/I), to agree with the syntax of the method minor.
i55 : M4 = matroid completeGraph 4
o55 = M4
o55 : Matroid
i56 : hasMinor(M5, M4)
Isomorphism: matroids are equal
Contract set {9}, delete set {3, 5, 8}
o56 = true
i57 : minorM5 == M4
o57 = true
Finally, the single most useful piece of information about a matroid that can be obtained from deletions and
contractions is the Tutte polynomial TM (x, y). It is a bivariate polynomial with integer coefficients which satisfies
the deletion-contraction recurrence
TM (x, y) = TM\e (x, y) + TM/e (x, y),
e ∈ E not a loop or coloop
with the initial condition TM (x, y) = xa y b if M consists of a coloops and b loops.
i58 : tuttePolynomial M5
6
5
4
3
4
3
2
2
3
2
2
o58 = y + 4y + x + 5x*y + 10y + 6x + 10x y + 15x*y + 15y + 11x + 20x*y + 15y · · ·
o58 : ZZ[x, y]
The Tutte polynomial is the universal invariant satisfying deletion-contraction, in the sense that any numerical
invariant of matroids which satisfies a (weighted) deletion-contraction recurrence is an evaluation of the Tutte
polynomial, up to a scale factor. For instance, the number of bases is equal to TM (1, 1):
MATROIDS: A MACAULAY2 PACKAGE
7
i59 : tutteEvaluate(M5, 1, 1)
o59 = 125
Even for graphic matroids, the Tutte polynomial contains a wealth of information about the graph; e.g. the
Tutte polynomial specializes to the chromatic polynomial, the flow polynomial, and the reliability polynomial.
For specific points, TM(G) (2, 1) counts the number of forests in G, and TM(G) (2, 0) counts the number of acyclic
orientations of G.
i60 : tutteEvaluate(M5, 2, 1)
o60 = 291
i61 : tutteEvaluate(M5, 2, 0)
o61 = 120
i62 : factor chromaticPolynomial K5
o62 = (x)(x - 4)(x - 3)(x - 2)(x - 1)
o62 : Expression of class Product
Connections. We conclude by presenting some connections of matroids to other areas of mathematics. First,
the matroid polytope PM : let M = ([n], B) be a matroid on the ground set {1, . . . , n}. In the Euclidean space Rn
with standard basis {e1 , . . . , en }, one can form the convex hull of the indicator vectors of the bases of M , namely:
!
X
PM := conv
ei | B ∈ B
i∈B
The matroid polytope can be created as follows:
i63 : needsPackage "Polyhedra"
o63 = Polyhedra
o63 : Package
i64 : convexHull matroidPolytope M4
o64 = {ambient dimension => 6
}
dimension of lineality space => 0
dimension of polyhedron => 5
number of facets => 16
number of rays => 0
number of vertices => 16
o64 : Polyhedron
A theorem of Gelfand, Goresky, MacPherson, and Serganova [3] classifies the subsets B ⊆ 2[n] which are the
bases of a matroid on [n] in terms of the polytope PM .
Next is optimization: let E be a finite set, and I ⊆ 2E a set of subsets that is downward closed: i.e. if X ∈ I
and Y ⊆ X, thenX
Y ∈ I. Let w be a weight function on E, i.e. a function w : E → R, extended to w : 2E → R by
setting w(X) :=
w(x). Consider the optimization problem:
x∈X
(*) Find a maximal member of I of maximum weight, with respect to w.
One attempt to try to solve (*) is to apply the greedy algorithm: namely, after having already selected elements
{x1 , . . . , xi }, choose an element xi+1 ∈ E of maximum weight such that {x1 , . . . , xi , xi+1 } ∈ I, and repeat. It turns
out that the greedy algorithm will always work if I is the independent sets of a matroid, and for nothing else in
general, as the following theorem shows (curiously, this was proved before Whitney introduced matroids):
Theorem 2. [2] Let E be a finite set, and I ⊆ 2E . Then I is the set of independent sets of a matroid on E iff I
is downward closed and for all weight functions w : E → R, the greedy algorithm successfully solves (*).
A solution to (*) provided by the greedy algorithm can be obtained using the method greedyAlgorithm (the
weight function is specified by its list of values on E):
i65 : w = {0, log(2), 4/3, 1, -4, 2, pi_RR};
8
JUSTIN CHEN
i66 : greedyAlgorithm(F7, w)
o66 = {6, 5, 3}
o66 : List
Last but not least is algebraic geometry: analogous to an Orlik-Solomon algebra (the cohomology ring of the
complement of a hyperplane arrangement), one can define a cohomology ring associated to a matroid. For a
matroid M on ground set E with no loops, set
R := Z[xF ]/(I1 + I2 ),
I1 :=
X
i1 ∈F
xF −
X
!
i1 , i2 ∈ E distinct ,
i2 ∈F
′
I2 := (xF xF ′ | F, F incomparable)
′
where F (and F ) runs over all nonempty proper flats of M . When M is realizable over C, R is the cohomology
ring of the complement of a subspace arrangement corresponding to the lattice of flats of M , without the minimal
and maximal elements. It is easy to see that (taking R with coefficients in a field k instead of Z) R is a standard
graded Artinian k-algebra of Castelnuovo-Mumford regularity r = rk M − 1. A recent result of Adiprasito, Katz,
and Huh [1] states that R is a Poincare duality algebra (in particular, is Gorenstein), and also satisfies the Lefschetz
property: for sufficiently general l ∈ R1 , and for j ≤ r/2, multiplication by lr−2j is an isomorphism Rj → Rr−j .
We illustrate the Gorenstein property for R with the Vamos matroid (the smallest matroid not realizable over any
field):
i67 : V = specificMatroids("vamos")
o67 = V
o67 : Matroid
i68 : (#groundSet V, #bases V, rk V, #flatten flats V)
o68 = (8, 65, 4, 79)
o68 : Sequence
i69 : I = idealCohomologyRing V; -- ideal in polynomial ring with 77 = 79 - 2 variables
o69 : Ideal of QQ[x , x , x , x , x , x , x , x , x , x , x , x , x , x , x , x · · ·
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
1 ···
i70 : time toList(0..<rk V)/(i -> hilbertFunction(i, I))
-- used 32.3128 seconds
o70 = {1, 70, 70, 1}
o70 : List
Acknowledgements. This project was partially supported by NSF grant DMS-1001867. The author is very
grateful to June Huh for explaining the connection of the cohomology ring of a matroid. The author would also like
to thank David Eisenbud and Daniel Grayson for advice with Macaulay2, and Joe Kileel for helpful discussions.
References
[1] Adiprasito, K., Katz, E., and Huh, J. Hodge theory for combinatorial geometries. In preparation.
[2] Borůvka, O. O jistém problému minimálnı̀m. Práce Mor. Prı́rodoved. Spol. v Brne 3, 37-58, 1926.
[3] Gelfand, I. M., Goresky, R. M., MacPherson, R. D., and Serganova, V. V. Combinatorial geometries, convex polyhedra, and
Schubert cells. Advances in Mathematics 63 (3): 301-316, 1987.
[4] Grayson, D. R. and Stillman, M. E. Macaulay2, a software system for research in algebraic geometry. Available at
http://www.math.uiuc.edu/Macaulay2/
[5] Oxley, J. G. Matroid Theory. Second Edition, Oxford University Press, New York, 2011.
[6] Tutte, W. T. A homotopy theorem for matroids, I, II. Trans. Amer. Math. Soc. 88, 144-174, 1958.
[7] Tutte, W. T. Matroids and graphs. Trans. Amer. Math. Soc. 90, 527-552, 1959.
[8] Whitney, H. On the abstract properties of linear dependence. Amer. J. Math. 57, 509-533, 1935.
Department of Mathematics, University of California, Berkeley, California, 94720 U.S.A
E-mail address: [email protected]
| 0math.AC
|
HIERARCHICALLY HYPERBOLIC SPACES II:
COMBINATION THEOREMS AND THE DISTANCE FORMULA
arXiv:1509.00632v3 [math.GR] 14 Feb 2017
JASON BEHRSTOCK, MARK F. HAGEN, AND ALESSANDRO SISTO
Abstract. We introduce a number of tools for finding and studying hierarchically hyperbolic spaces (HHS), a rich class of spaces including mapping class groups of surfaces, Teichmüller space with either the Teichmüller or Weil-Petersson metrics, right-angled Artin
groups, and the universal cover of any compact special cube complex. We begin by introducing a streamlined set of axioms defining an HHS. We prove that all HHSs satisfy a
Masur-Minsky-style distance formula, thereby obtaining a new proof of the distance formula
in the mapping class group without relying on the Masur-Minsky hierarchy machinery. We
then study examples of HHSs; for instance, we prove that when M is a closed irreducible
3–manifold then π1 M is an HHS if and only if it is neither N il nor Sol. We establish this
by proving a general combination theorem for trees of HHSs (and graphs of HH groups).
We also introduce a notion of “hierarchical quasiconvexity”, which in the study of HHS is
analogous to the role played by quasiconvexity in the study of Gromov-hyperbolic spaces.
Contents
Introduction
1. The main definition and background on hierarchically hyperbolic spaces
2. Tools for studying HHSs
3. Realization of consistent tuples
4. Hierarchy paths and the distance formula
5. Hierarchical quasiconvexity and gates
6. Hulls
7. The coarse median property
8. Combination theorems for hierarchically hyperbolic spaces
9. Hyperbolicity relative to HHGs
10. Hierarchical hyperbolicity of 3-manifold groups
11. A new proof of the distance formula for mapping class groups
References
1
8
14
17
20
27
31
34
36
47
49
51
53
Introduction
One of the most remarkable aspects of the theory of mapping class groups of surfaces is that
the coarse geometry of the mapping class group, MCGpSq, can be fully reconstructed from
its shadows on a collection of hyperbolic spaces — namely the curve graphs of subsurfaces
of the underlying surface. Each subsurface of the surface S is equipped with a hyperbolic
curve graph and a projection, the subsurface projection, to this graph from MCGpSq; there
Behrstock was supported as a Simons Fellow.
Hagen was supported by NSF Grant Number 1045119 and by EPSRC.
Sisto was supported by the Swiss National Science Foundation project 144373.
1
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
2
are also projections between the various curve graphs. The powerful Masur–Minsky distance
formula [MM00] shows that the distance between points of MCGpSq is coarsely the sum over
all subsurfaces of the distances between the projections of these points to the various curve
graphs. Meanwhile, the consistency/realization theorem [BKMM12] tells us that tuples with
coordinates in the different curve graphs that obey “consistency” conditions characteristic of
images of actual points in MCGpSq are, coarsely, images of points in MCGpSq. Finally, any
two points in MCGpSq are joined by a uniform-quality quasigeodesic projecting to a uniform
unparameterized quasigeodesic in each curve graph — a hierarchy path [MM00].
It is perhaps surprising that analogous behavior should appear in CAT(0) cube complexes,
since the mapping class group cannot act properly on such complexes, c.f., [Bri10, Hag07,
KL96]. However, mapping class groups enjoy several properties reminiscent of nonpositively/negatively curved spaces, including: automaticity (and, thus, quadratic Dehn function)
[Mos95], having many quasimorphisms [BF02], super-linear divergence [Beh06], etc. Mapping class groups also exhibit coarse versions of some features of CAT(0) cube complexes,
including coarse centroids/medians [BM11] and, more generally, a local coarse structure of a
cube complex as made precise in [Bow13], applications to embeddings in trees, [BDS11], etc.
Accordingly, it is natural to seek a common thread joining these important classes of groups
and spaces.
In [Hag14] it was shown that, for an arbitrary CAT(0) cube complex X , the intersectiongraph of the hyperplane carriers — the contact graph — is hyperbolic, and in fact quasiisometric to a tree. This object seems at first glance quite different from the curve graph
(which records, after all, non-intersection), but there are a number of reasons this is quite
natural, two of which we now mention. First, the curve graph can be realized as a coarse
intersection graph of product regions in MCG. Second, the contact graph is closely related
to the intersection graph of the hyperplanes themselves; when X is the universal cover of
the Salvetti complex of a right-angled Artin group, the latter graph records commutation of
conjugates of generators, just as the curve graph records commutation of Dehn twists.
The cube complex X coarsely projects to its contact graph. Moreover, using disc diagram
techniques, it is not hard to show that any two 0–cubes in a CAT(0) cube complex are
joined by a combinatorial geodesic projecting to a geodesic in the contact graph [BHS14].
This observation — that CAT(0) cube complexes have “hierarchy paths” with very strong
properties — motivated a search for an analogue of the theory of curve graphs and subsurface
projections in the world of CAT(0) cube complexes. This was largely achieved in [BHS14],
where a theory completely analogous to the mapping class group theory was constructed for
a wide class of CAT(0) cube complexes, with (a variant of) the contact graph playing the role
of the curve graph. These results motivated us to define a notion of “spaces with distance
formulae”, which we did in [BHS14], by introducing the class of hierarchically hyperbolic spaces
(HHSs) to provide a framework for studying many groups and spaces which arise naturally
in geometric group theory, including mapping class groups and virtually special groups, and
to provide a notion of “coarse nonpositive curvature” which is quasi-isometry invariant while
still yielding some of those properties available via local geometry in the classical setting of
nonpositively-curved spaces.
As mentioned above, the three most salient features of hierarchically hyperbolic spaces
are: the distance formula, the realization theorem, and the existence of hierarchy paths. In
the treatment given in [BHS14], these attributes are part of the definition of a hierarchically
hyperbolic space. This is somewhat unsatisfactory since, in the mapping class group and
cubical settings, proving these theorems requires serious work.
In this paper, we show that although the definition of hierarchically hyperbolic space previously introduced identifies the right class of spaces, there exist a streamlined set of axioms
for that class of spaces which are much easier to verify in practice than those presented
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
3
in [BHS14, Section 13] and which don’t require assuming a distance formula, realization
theorem, or the existence of hierarchy paths. Thus, a significant portion of this paper is
devoted to proving that those results can be derived from the simplified axioms we introduce
here. Along the way, we obtain a new, simplified proof of the actual Masur-Minsky distance
formula for the mapping class group. We then examine various geometric properties of hierarchically hyperbolic spaces and groups, including many reminiscent of the world of CAT(0)
spaces and groups; for example, we show that hierarchically hyperbolic groups have quadratic
Dehn function. Finally, taking advantage of the simpler set of axioms, we prove combination
theorems enabling the construction of new hierarchically hyperbolic spaces/groups from old.
The definition of a hierarchically hyperbolic space still has several parts, the details of
which we postpone to Section 1. However, the idea is straightforward: a hierarchically
hyperbolic space is a pair pX , Sq, where X is a metric space and S indexes a set of δ–
hyperbolic spaces with several features (to each U P S the associated space is denoted CU ).
Most notably, S is endowed with 3 mutually exclusive relations, nesting, orthogonality, and
transversality, respectively generalizing nesting, disjointness, and overlapping of subsurfaces.
For each U P S, we have a coarsely Lipschitz projection πU : X Ñ CU , and there are relative
projections CU Ñ CV when U, V P S are non-orthogonal. These projections are required to
obey “consistency” conditions modeled on the inequality identified by Behrstock in [Beh06],
as well as a version of the bounded geodesic image theorem and large link lemma of [MM00],
among other conditions. A finitely generated group G is hierarchically hyperbolic if it can
be realized as a group of HHS automorphisms (“hieromorphisms”, as defined in Section 1) so
that the induced action on X by uniform quasi-isometries is geometric and the action on S
is cofinite. Hierarchically hyperbolic groups, endowed with word-metrics, are hierarchically
hyperbolic spaces, but the converse does not appear to be true.
Combination theorems. One of the main contributions in this paper is to provide many
new examples of hierarchically hyperbolic groups, thus showing that mapping class groups
and various cubical complexes/groups are just two of many interesting families in this class
of groups and spaces. We provide a number of combination theorems, which we will describe
below. One consequence of these results is the following classification of exactly which 3–
manifold groups are hierarchically hyperbolic:
Theorem 10.1 (3–manifolds are hierarchically hyperbolic). Let M be a closed 3–
manifold. Then π1 pM q is a hierarchically hyperbolic space if and only if M does not have a
Sol or Nil component in its prime decomposition.
This result has a number of applications to the many fundamental groups of 3–manifolds
which are HHS. For instance, in such cases, it follows from results in [BHS14] that: except
for Z3 , the top dimension of a quasi-flat in such a group is 2, and any such quasi-flat is locally
close to a “standard flat” (this generalizes one of the main results of [KL97, Theorem 4.10]);
up to finite index, Z and Z2 are the only finitely generated nilpotent groups which admit
quasi-isometric embeddings into π1 pM q; and, except in the degenerate case where π1 pM q is
virtually abelian, such groups are all acylindrically hyperbolic (as also shown in [MO14]).
Remark (Hierarchically hyperbolic spaces vs. hierarchically hyperbolic groups). There is
an important distinction to be made between a hierarchically hyperbolic space, which is a
metric space X equipped with a collection S of hyperbolic spaces with certain properties,
and a hierarchically hyperbolic group, which is a group acting geometrically on a hierarchically
hyperbolic space in such a way that the induced action on S is cofinite. The latter property
is considerably stronger. For example, Theorem 10.1 shows that π1 M , with any word-metric,
is a hierarchically hyperbolic space, but, as we discuss in Remark 10.2, π1 M probably fails
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
4
to be a hierarchically hyperbolic group in general; for instance we conjecture this is the case
for those graph manifolds which can not be cocompactly cubulated.
In the course of proving Theorem 10.1, we establish several general combination theorems,
including one about relative hyperbolicity and one about graphs of groups. The first is:
Theorem 9.1 (Hyperbolicity relative to HHGs). Let the group G be hyperbolic relative
to a finite collection P of peripheral subgroups. If each P P P is a hierarchically hyperbolic
space, then G is a hierarchically hyperbolic space. Further, if each P P P is a hierarchically
hyperbolic group, then so is G.
Another of our main results is a combination theorem establishing when a tree of hierarchically hyperbolic spaces is again a hierarchically hyperbolic space.
Theorem 8.6 (Combination theorem for HHS). Let T be a tree of hierarchically hyperbolic spaces. Suppose that:
‚ edge-spaces are uniformly hierarchically quasiconvex in incident vertex spaces;
‚ each edge-map is full;
‚ T has bounded supports;
‚ if e is an edge of T and Se is the Ď–maximal element of Se , then for all V P Se˘ ,
the elements V and φ♦
pS q are not orthogonal in Se˘ .
e˘ e
Then XpT q is hierarchically hyperbolic.
Hierarchical quasiconvexity is a natural generalization of both quasiconvexity in the hyperbolic setting and cubical convexity in the cubical setting, which we shall discuss in some
detail shortly. The remaining conditions are technical and explained in Section 8, but are
easily verified in practice.
As a consequence, we obtain a set of sufficient conditions guaranteeing that a graph of
hierarchically hyperbolic groups is a hierarchically hyperbolic group.
Corollary 8.22 (Combination theorem for HHG). Let G “ pΓ, tGv u, tGe u, tφ˘
e uq be
a finite graph of hierarchically hyperbolic groups. Suppose that G equivariantly satisfies the
hypotheses of Theorem 8.6. Then the total group G of G is a hierarchically hyperbolic group.
Finally, we prove that products of hierarchically hyperbolic spaces admit natural hierarchically hyperbolic structures.
As mentioned earlier, we will apply the combination theorems to fundamental groups of 3–
manifolds, but their applicability is broader. For example, they can be applied to fundamental
groups of higher dimensional manifolds such as the ones considered in [FLS11].
The distance formula and realization. As defined in [BHS14], the basic definition of a
hierarchically hyperbolic space is modeled on the essential properties underlying the “hierarchy
machinery” of mapping class groups. In this paper, we revisit the basic definition and provide
a new, refined set of axioms; the main changes are the removal of the “distance formula”
and “hierarchy path” axioms and the replacement of the “realization” axiom by a far simpler
“partial realization”. These new axioms are both more fundamental and more readily verified.
An important result in mapping class groups which provides a starting point for much recent research in the field is the celebrated “distance formula” of Masur–Minsky [MM00] which
provides a way to estimate distances in the mapping class group, up to uniformly bounded
additive and multiplicative distortion, via distances in the curve graphs of subsurfaces. We
give a new, elementary, proof of the distance formula in the mapping class group. The first
step in doing so is verifying that mapping class groups satisfy the new axioms of a hierarchically hyperbolic space. We provide elementary, simple proofs of the axioms for which
elementary proofs do not exist in the literature (most notably, the uniqueness axiom); this
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
5
is done in Section 11. This then combines with our proof of the following result which states
that any hierarchically hyperbolic space satisfies a “distance formula” (which in the case of
the mapping class group provides a new proof of the Masur–Minsky distance formula):
Theorem 4.5 (Distance formula for HHS). Let pX, Sq be hierarchically hyperbolic. Then
there exists s0 such that for all s ě s0 there exist constants K, C such that for all x, y P X ,
ÿ
dX px, yq —pK,Cq
ttdW px, yquus ,
W PS
where dW px, yq denotes the distance in the hyperbolic space CW between the projections of
x, y and ttAuuB “ A if A ě B and 0 otherwise.
Moreover, we show in Theorem 4.4 that any two points in X are joined by a uniform
quasigeodesic γ projecting to a uniform unparameterized quasigeodesic in CU for each U P
S. The existence of such hierarchy paths was hypothesized as part of the definition of a
hierarchically hyperbolic space in [BHS14], but now it is proven as a consequence of the
other axioms.
The Realization Theorem for the mapping class group was established by Behrstock–
Kleiner–Minsky–Mosher in [BKMM12, Theorem 4.3]. This theorem states that given a surface S and, for each subsurface W Ď S, a point in the curve complex of W : this sequence of
points arises as the projection of a point in the mapping class group (up to bounded error),
whenever the curve complex elements satisfy certain pairwise “consistency conditions.” Thus
the Realization Theorem provides another sense in which all of the quasi-isometry invariant
geometry of the mapping class group is encoded by the projections onto the curve graphs of
subsurfaces.1 In Section 3 we show that an arbitrary hierarchically hyperbolic space satisfies
a realization theorem. Given our elementary proof of the new axioms for mapping class
groups in Section 11, we thus obtain a new proof of [BKMM12, Theorem 4.3].
Hulls and the coarse median property. Bowditch introduced a notion of coarse median
space to generalize some results about median spaces to a more general setting, and, in
particular, to the mapping class group [Bow13]. Bowditch observed in [Bow15] that any
hierarchically hyperbolic space is a coarse median space; for completeness we provide a short
proof of this result in Theorem 7.3. Using Bowditch’s results about coarse median spaces, we
obtain a number of applications as corollaries. For instance, Corollary 7.9 is obtained from
[Bow14a, Theorem 9.1] and says that any hierarchically hyperbolic space satisfies the Rapid
Decay Property and Corollary 7.5 is obtained from [Bow13, Corollary 8.3] to show that all
hierarchically hyperbolic groups are finitely presented and have quadratic Dehn functions.
This provides examples of groups that are not hierarchically hyperbolic, for example:
Corollary 7.6 (OutpFn q is not an HHG). For n ě 3, the group OutpFn q is not a hierarchically hyperbolic group.
Indeed, OutpFn q was shown in [BV95, HM13b, BV] to have exponential Dehn function.
This result is interesting as a counter-point to the well-known and fairly robust analogy
between OutpFn q and the mapping class group of a surface; especially in light of the fact
that OutpFn q is known to have a number of properties reminiscent of the axioms for an HHS,
c.f., [BF14b, BF14a, HM13a, SS].
The coarse median property, via work of Bowditch, also implies that asymptotic cones
of hierarchically hyperbolic spaces are contractible. Moreover, in Corollary 6.7, we bound
the homological dimension of any asymptotic cone of a hierarchically hyperbolic space. This
1In [BKMM12], the name Consistency Theorem is used to refer to the necessary and sufficient conditions
for realization; since we find it useful to break up these two aspects, we refer to this half as the Realization
Theorem, since anything that satisfies the consistency conditions is realized.
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
6
latter result relies on the use of hulls of finite sets of points in the HHS X . This construction generalizes the Σ–hull of a finite set, constructed in the mapping class group context
in [BKMM12]. (It also generalizes a special case of the ordinary combinatorial convex hull
in a CAT(0) cube complex.) A key feature of these hulls is that they are coarse retracts of X
(see Proposition 6.3), and this plays an important role in the proof of the distance formula.
Hierarchical spaces. We also introduce the more general notion of a hierarchical space
(HS). This is the same as a hierarchically hyperbolic space, except that we do not require
the various associated spaces CU , onto which we are projecting, to be hyperbolic. Although
we mostly focus on HHS in this paper, a few things are worth noting. First, the realization
theorem 3.1 actually makes no use of hyperbolicity of the CU , and therefore holds in the
more general context of HS; see Section 3.
Second, an important subclass of the class of HS is the class of relatively hierarchically
hyperbolic spaces, which we introduce in Section 6.2. These are hierarchical spaces where the
spaces CU are uniformly hyperbolic except when U is minimal with respect to the nesting
relation. As their name suggests, this class includes all metrically relatively hyperbolic spaces;
see Theorem 9.3. With an eye to future applications, in Section 6.2 we prove a distance
formula analogous to Theorem 4.5 for relatively hierarchically hyperbolic spaces, and also
establish the existence of hierarchy paths. The strategy is to build, for each pair of points
x, y, in the relatively hierarchically hyperbolic space, a “hull” of x, y, which we show is
hierarchically hyperbolic with uniform constants. We then apply Theorems 4.5 and 4.4.
Standard product regions and hierarchical quasiconvexity. In Section 5.1, we introduce the notion of a hierarchically quasiconvex subspace of a hierarchically hyperbolic space
pX , Sq. In the case where X is hyperbolic, this notion coincides with the usual notion of
quasiconvexity. The main technically useful features of hierarchically quasiconvex subspaces
generalize key features of quasiconvexity: they inherit the property of being hierarchically
hyperbolic (Proposition 5.6) and one can coarsely project onto them (Lemma 5.5).
Along with the hulls discussed above, the most important examples of hierarchically quasiconvex subspaces are standard product regions: for each U P S, one can consider the set
PU of points x P X whose projection to each CV is allowed to vary only if V is nested into or
orthogonal to U ; otherwise, x projects to the same place in CV as CU does under the relative
projection. The space PU coarsely decomposes as a product, with factors corresponding to
the nested and orthogonal parts. Product regions play an important role in the study of
boundaries and automorphisms of hierarchically hyperbolic spaces in the forthcoming paper [DHS15], as well as in the study of quasi-boxes and quasiflats in hierarchically hyperbolic
spaces carried out in [BHS14].
Some questions and future directions. Before embarking on the discussion outlined
above, we raise a few questions about hierarchically hyperbolic spaces and groups that we
believe are of significant interest.
The first set of questions concern the scope of the theory, i.e., which groups and spaces
are hierarchically hyperbolic and which operations preserve the class of HHS:
Question A (Cubical groups). Let G act properly and cocompactly on a CAT(0) cube complex. Is G a hierarchically hyperbolic group? Conversely, suppose that pG, Sq is a hierarchically hyperbolic group; are there conditions on the elements of S which imply that G acts
properly and cocompactly on a CAT(0) cube complex?
Substantial evidence for this conjecture was provided in [BHS14] where we established
that a CAT(0) cube complex X containing a factor system is a hierarchically hyperbolic
space, and the associated hyperbolic spaces are all uniform quasi-trees. (Roughly speaking,
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
7
X contains a factor-system if the following collection of subcomplexes has finite multiplicity:
the smallest collection of convex subcomplexes that contains all combinatorial hyperplanes
and is closed under collecting images of closest-point projection maps between its elements.)
The class of cube complexes that are HHS in this way contains all universal covers of special
cube complexes with finitely many immersed hyperplanes, but the cube complexes containing
factor systems have not been completely characterized. In the forthcoming paper [DHS15],
we show that the above question is closely related to a conjecture of the first two authors on
the simplicial boundary of cube complexes [BH, Conjecture 2.8].
More generally, we ask the following:
Question B (Factor systems in median spaces). Is there a theory of factor systems in median
spaces generalizing that in CAT(0) cube complexes, such that median spaces/groups admitting
factor systems are hierarchically hyperbolic?
Presumably, a positive answer to Question B would involve the measured wallspace structure
on median spaces discussed in [CDH10]. One would have to develop an analogue of the
contact graph of a cube complex to serve as the underlying hyperbolic space. One must be
careful since, e.g., the Baumslag-Solitar group BSp1, 2q is median but has exponential Dehn
function [Ger92] and is thus not a hierarchically hyperbolic space, by Corollary 7.5. On the
other hand, if the answer to Question B is positive, one might try to do the same thing for
coarse median spaces.
There are a number of other groups and spaces where it is natural to inquire whether or
not they are hierarchically hyperbolic. For example:
Question C (Handlebody group). Let H be a compact oriented 3–dimensional genus g
handlebody, and let Gg ď MCGpBHq be the group of isotopy classes of diffeomorphisms of
H. Is Gg a hierarchically hyperbolic group?
Question D (Graph products). Let G be a (finite) graph product of hierarchically hyperbolic
groups. Is G hierarchically hyperbolic?
The answer to Question C is presumably no, while the answer to D is most likely yes. The
positive answer to Question D would follow from a strengthened version of Theorem 8.6.
There are other candidate examples of hierarchically hyperbolic spaces. For example, it
is natural to ask whether a right-angled Artin group with the syllable-length metric, introduced in [KK14], which is analogous to Teichmüller space with the Weil-Petersson metric, is
hierarchically hyperbolic.
As far as the difference between hierarchically hyperbolic spaces and groups is concerned,
we conjecture that the following question has a positive answer:
Question E. Is it true that the fundamental group G of a non-geometric graph manifold is
a hierarchically hyperbolic group if and only if G is virtually compact special?
It is known that G as above is virtually compact special if and only if it is chargeless in
the sense of [BS05], see [HP13].
There remain a number of open questions about the geometry of hierarchically hyperbolic
spaces in general. Theorem 7.3 ensures, via work of Bowditch, that every asymptotic cone
of a hierarchically hyperbolic space is a median space [Bow13]; further properties in this
direction are established in Section 6. Motivated by combining the main result of [Sis11] on
3–manifold groups with Theorem 10.1, we ask:
Question F. Are any two asymptotic cones of a given hierarchically hyperbolic space biLipschitz equivalent?
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
8
The notion of hierarchical quasiconvexity of a subgroup of a hierarchically hyperbolic group
pG, Sq generalizes quasiconvexity in word-hyperbolic groups and cubical convex-cocompactness
in groups acting geometrically on CAT(0) cube complexes with factor-systems. Another notion of quasiconvexity is stability, defined by Durham-Taylor in [DT14]. This is a quite different notion of quasiconvexity, since stable subgroups are necessarily hyperbolic. In [DT14],
the authors characterize stable subgroups of the mapping class group; it is reasonable to ask
for a generalization of their results to hierarchically hyperbolic groups.
Many hierarchically hyperbolic spaces admit multiple hierarchically hyperbolic structures.
However, as discussed in [BHS14], a CAT(0) cube complex with a factor-system has a “minimal” factor-system, i.e., one that is contained in all other factor systems. In this direction,
it is natural to ask whether a hierarchically hyperbolic space pX , Sq admits a hierarchically
hyperbolic structure that is canonical in some way.
Organization of the paper. Section 1 contains the full definition of a hierarchically hyperbolic space (and, more generally, a hierarchical space) and some discussion of background.
Section 2 contains various basic consequences of the definition, and some tricks that are
used repeatedly. In Section 3, we prove the realization theorem (Theorem 3.1). In Section 4 we establish the existence of hierarchy paths (Theorem 4.4) and the distance formula
(Theorem 4.5). Section 5 is devoted to hierarchical quasiconvexity and product regions, and
Section 6 to coarse convex hulls and relatively hierarchically hyperbolic spaces. The coarse
median property and its consequences are detailed in Section 7. The combination theorems
for trees of spaces, graphs of groups, and products are proved in Section 8, and groups hyperbolic relative to HHG are studied in Section 9. This is applied to 3-manifolds in Section 10.
Finally, in Section 11, we prove that mapping class groups are hierarchically hyperbolic.
Acknowledgments. MFH and AS thank the organizers of the Ventotene 2015 conference,
and JB and MFH thank CIRM and the organizers of GAGTA 2015, where some of the work
on this paper was completed.
1. The main definition and background on hierarchically hyperbolic spaces
1.1. The axioms. We begin by defining a hierarchically hyperbolic space. We will work in
the context of a quasigeodesic space, X , i.e., a metric space where any two points can be
connected by a uniform-quality quasigeodesic. Obviously, if X is a geodesic space, then it
is a quasigeodesic space. Most of the examples we are interested in are geodesic spaces, but
in order to construct hierarchically hyperbolic structures on naturally-occurring subspaces
of hierarchically hyperbolic spaces, we must work in the slightly more general setting of
quasigeodesic spaces.
Definition 1.1 (Hierarchically hyperbolic space). The q–quasigeodesic space pX , dX q is a
hierarchically hyperbolic space if there exists δ ě 0, an index set S, and a set tCW : W P Su
of δ–hyperbolic spaces pCU, dU q, such that the following conditions are satisfied:
(1) (Projections.) There is a set tπW : X Ñ 2CW | W P Su of projections sending
points in X to sets of diameter bounded by some ξ ě 0 in the various CW P S.
Moreover, there exists K so that each πW is pK, Kq–coarsely Lipschitz.
(2) (Nesting.) S is equipped with a partial order Ď, and either S “ H or S contains
a unique Ď–maximal element; when V Ď W , we say V is nested in W . We require
that W Ď W for all W P S. For each W P S, we denote by SW the set of V P S
such that V Ď W . Moreover, for all V, W P S with V Ĺ W there is a specified
CV .
subset ρVW Ă CW with diamCW pρVW q ď ξ. There is also a projection ρW
V : CW Ñ 2
V
(The similarity in notation is justified by viewing ρW as a coarsely constant map
CV Ñ 2CW .)
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
9
(3) (Orthogonality.) S has a symmetric and anti-reflexive relation called orthogonality:
we write V KW when V, W are orthogonal. Also, whenever V Ď W and W KU , we
require that V KU . Finally, we require that for each T P S and each U P ST for
which tV P ST | V KU u ‰ H, there exists W P ST ´ tT u, so that whenever V KU
and V Ď T , we have V Ď W . Finally, if V KW , then V, W are not Ď–comparable.
(4) (Transversality and consistency.) If V, W P S are not orthogonal and neither is
nested in the other, then we say V, W are transverse, denoted V &W . There exists
κ0 ě 0 such that if V &W , then there are sets ρVW Ď CW and ρW
V Ď CV each of
diameter at most ξ and satisfying:
(
min dW pπW pxq, ρVW q, dV pπV pxq, ρW
V q ď κ0
for all x P X .
For V, W P S satisfying V Ď W and for all x P X , we have:
(
min dW pπW pxq, ρVW q, diamCV pπV pxq Y ρW
V pπW pxqqq ď κ0 .
&
The preceding two inequalities are the consistency inequalities for points in X .
V
Finally, if U Ď V , then dW pρU
W , ρW q ď κ0 whenever W P S satisfies either V Ĺ W
or V &W and W U .
(Finite complexity.) There exists n ě 0, the complexity of X (with respect to S),
so that any set of pairwise–Ď–comparable elements has cardinality at most n.
(Large links.) There exist λ ě 1 and E ě maxtξ, κ0 u such that the following holds.
Let W P S and let x, x1 P X . Let N “ λdW pπW pxq, πW px1 qq ` λ. Then there exists
tTi ui“1,...,tN u Ď SW ´ tW u such that for all T P SW ´ tW u, either T P STi for some
i, or dT pπT pxq, πT px1 qq ă E. Also, dW pπW pxq, ρTWi q ď N for each i.
(Bounded geodesic image.) For all W P S, all V P SW ´ tW u, and all geodesics
V
γ of CW , either diamCV pρW
V pγqq ď E or γ X NE pρW q ‰ H.
(Partial Realization.) There exists a constant α with the following property. Let
tVj u be a family of pairwise orthogonal elements of S, and let pj P πVj pX q Ď CVj .
Then there exists x P X so that:
‚ dVj px, pj q ď α for all j,
(5)
(6)
(7)
(8)
V
‚ for each j and each V P S with Vj Ď V , we have dV px, ρVj q ď α, and
V
‚ if W &Vj for some j, then dW px, ρWj q ď α.
(9) (Uniqueness.) For each κ ě 0, there exists θu “ θu pκq such that if x, y P X and
dpx, yq ě θu , then there exists V P S such that dV px, yq ě κ.
We say that the q–quasigeodesic metric spaces tXi u are uniformly hierarchically hyperbolic
if each Xi satisfies the axioms above and all constants, including the complexities, can be
chosen uniformly. We often refer to S, together with the nesting and orthogonality relations,
the projections, and the hierarchy paths, as a hierarchically hyperbolic structure for the space
X . Observe that X is hierarchically hyperbolic with respect to S “ H, i.e., hierarchically hyperbolic of complexity 0, if and only if X is bounded. Similarly, X is hierarchically hyperbolic
of complexity 1 with respect to S “ tX u, if and only if X is hyperbolic.
Notation 1.2. Where it will not cause confusion, given U P S, we will often suppress the
projection map πU when writing distances in CU , i.e., given x, y P X and p P CU we write
dU px, yq for dU pπU pxq, πU pyqq and dU px, pq for dU pπU pxq, pq. Note that when we measure
distance between a pair of sets (typically both of bounded diameter) we are taking the
minimum distance between the two sets. Given A Ă X and U P S we let πU pAq denote
YaPA πU paq.
Remark 1.3 (Surjectivity of projections). One can always replace each CU with a thickening
of πU pX q, and hence make each πU coarsely surjective. The details can be found in the
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
10
forthcoming paper [DHS15], where this procedure gets used; the resulting spaces are termed
normalized hierarchically hyperbolic spaces.
Remark 1.4 (Large link function). It appears as though there is no actual need to require in
Definition 1.1.(6) that N depend linearly on dW px, x1 q. Instead, we could have hypothesized
that for any C ě 0, there exists N pCq so that the statement of the axiom holds with
N “ N pCq whenever dW px, x1 q ď C. However, one could deduce from this and the rest of
the axioms that N pCq grows linearly in C, so we have elected to simply build linearity into
the definition.
Remark 1.5 (Summary of constants). Each hierarchically hyperbolic space pX , Sq is associated with a collection of constants often, as above, denoted δ, ξ, n, κ0 , E, θu , K, where:
(1) CU is δ–hyperbolic for each U P S,
(2) each πU has image of diameter at most ξ and each πU is pK, Kq–coarsely Lipschitz,
and each ρU
V has (image of) diameter at most ξ,
(3) for each x P X , the tuple pπU pxqqU PS is κ0 –consistent,
(4) E is the constant from the bounded geodesic image axiom.
Whenever working in a fixed hierarchically hyperbolic space, we use the above notation freely.
We can, and shall, assume that E ě q, E ě δ, E ě ξ, E ě κ0 , E ě K, and E ě α.
Remark 1.6. We note that in Definition 1.1.(1), the assumption that the projections are
Lipschitz can be replaced by the weaker assumption that there is a proper function of the
projected distance which is a lower bound for the distance in the space X . From this weaker
assumption, the fact that the projections are actually coarsely Lipschitz then follows from
the fact that we assume X to be quasi-geodesic. Since the Lipschitz hypothesis is cleaner to
state and, in practice, fairly easy to verify, we just remark on this for those that might find
this fact useful in proving that more exotic spaces are hierarchically hyperbolic.
&
1.2. Comparison to the definition in [BHS14]. Definition 1.1 is very similar to the definition of a hierarchically hyperbolic space given in [BHS14], with the following differences:
(1) The existence of hierarchy paths and the distance formula were stated as axioms
in [BHS14]; below, we deduce them from the other axioms. Similarly, the below
realization theorem was formerly an axiom, but has been replaced by the (weaker)
partial realization axiom.
(2) We now require X to be a quasigeodesic space. In [BHS14], this follows from the
existence of hierarchy paths, which was an axiom there.
(3) We now require the projections πU : X Ñ CU to be coarsely Lipschitz; although
this requirement was not imposed explicitly in [BHS14], it follows from the distance
formula, which was an axiom there.
(4) In [BHS14], there were five consistency inequalities; in Definition 1.1.(4), there are
two. The last three inequalities in the definition from [BHS14] follow from Proposition 1.7 below. (Essentially, the partial realization axiom has replaced part of the
old consistency axiom.)
V
(5) In Definition 1.1.(4), we require that, if U Ď V , then dW pρU
W , ρW q ď κ0 whenever
W P S satisfies either V Ĺ W or V &W and W U . In the context of [BHS14], this
follows by considering the standard product regions constructed using realization
(see [BHS14, Section 13.1] and Section 5.2 of the present paper).
Proposition 1.7 (ρ–consistency). There exists κ1 so that the following holds. Suppose that
U, V, W P S satisfy both of the following conditions: U Ĺ V or U &V ; and U Ĺ W or U &W .
Then, if V &W , then
(
V
U
W
min dCW pρU
W , ρW q, dCV pρV , ρV q ď κ1
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
11
and if V Ĺ W , then
(
V
U
W U
min dCW pρU
W , ρW q, diamCV pρV Y ρV pρW qq ď κ1 .
Proof. Suppose that U Ĺ V or U &V and the same holds for U, W . Suppose that V &W
or V Ď W . Choose p P πU pX q. There is a uniform α so that partial realization (DefiniU
tion 1.1.(8)) provides x P X so that dU px, pq ď α and dT px, ρU
T q ď α whenever ρT is defined
U
and coarsely constant. In particular, dV px, ρU
V q ď α and dW px, ρW q ď α. The claim now
follows from Definition 1.1.(4), with κ1 “ κ0 ` α.
l
In view of the discussion above, we have:
Proposition 1.8. The pair pX , Sq satisfies Definition 1.1 if and only if it is hierarchically
hyperbolic in the sense of [BHS14].
In particular, as observed in [BHS14]:
Proposition 1.9. If pX , Sq is a hierarchically hyperbolic space, and X 1 is a quasigeodesic
space quasi-isometric to X , then there is a hierarchically hyperbolic space pX 1 , Sq.
1.3. A variant on the axioms. Here we introduce two slightly simpler versions of the
HHS axioms and show that in the case, as in most situations which arise naturally, that the
projections are coarsely surjective, then it suffices to verify the simpler axioms.
The following is a subset of the nesting axiom; here we remove the definition of the
CV in the case V Ĺ W .
projection map ρW
V : CW Ñ 2
Definition 1.1.(2)1 (Nesting variant). S is equipped with a partial order Ď, and either
S “ H or S contains a unique Ď–maximal element; when V Ď W , we say V is nested in
W . We require that W Ď W for all W P S. For each W P S, we denote by SW the set of
V P S such that V Ď W . Moreover, for all V, W P S with V Ĺ W there is a specified subset
ρVW Ă CW with diamCW pρVW q ď ξ.
The following is a subset of the transversality and consistency axiom.
Definition 1.1.(4)1 (Transversality). If V, W P S are not orthogonal and neither is nested
in the other, then we say V, W are transverse, denoted V &W . There exists κ0 ě 0 such
that if V &W , then there are sets ρVW Ď CW and ρW
V Ď CV each of diameter at most ξ and
satisfying:
(
min dW pπW pxq, ρVW q, dV pπV pxq, ρW
V q ď κ0
&
for all x P X .
V
Finally, if U Ď V , then dW pρU
W , ρW q ď κ0 whenever W P S satisfies either V Ĺ W or
V &W and W U .
The following is a variant of the bounded geodesic image axiom:
Definition 1.1.(7)1 (Bounded geodesic image variant). Suppose that x, y P X and V Ĺ W
have the property that there exists a geodesic from πW pxq to πW pyq which stays pE ` 2δq-far
from ρVW . Then dV px, yq ď E.
Proposition 1.10. Given a quasigeodesic space X and an index set S, then pX , Sq is an
HHS if it satisfies the axioms of Definition 1.1 with the following changes:
‚ Replace Definition 1.1.(2) by Definition 1.1.(2)1 .
‚ Replace Definition 1.1.(4) by Definition 1.1.(4)1 .
‚ Replace Definition 1.1.(7) by Definition 1.1.(7)1 .
‚ Assume that for each CU the map πU pX q is uniformly coarsely surjective.
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
12
Proof. To verify Definition 1.1.(2), for each V, W P S with V Ĺ W , we define a map
CV as follows. If p P CW ´ N pρV q, then let ρW ppq “ π pxq for some x P X
ρW
E W
V
V : CW Ñ 2
V
with πW pxq — p. Since p does not lie E–close to ρVW , this definition is coarsely independent
of x by Definition 1.1.(7)1 . On NE pρVW q, we define ρW
V arbitrarily. By definition, the resulting
map satisfies Definition 1.1.(4). Moreover, coarse surjectivity of πW and Definition 1.1.(7)1
ensure that Definition 1.1.(7) holds. The rest of the axioms hold by hypothesis.
l
Remark 1.11. The definition of an HHS provided by Proposition 1.10 is convenient because
it does not require one to define certain maps between hyperbolic spaces: Definition 1.1.(2)1
is strictly weaker than Definition 1.1.(2). On the other hand, it is often convenient to work
with HHSes in which some of the projections πU are not coarsely surjective; for example,
this simplifies the proof that hierarchically quasiconvex subspaces inherit HHS structures in
Proposition 5.6. Hence we have included both definitions.
In practice, we almost always apply consistency and bounded geodesic image in concert,
which involves applying bounded geodesic image to geodesics of CW joining points in πW pX q.
Accordingly, Definition 1.1.(7)1 is motivated by the following easy observation:
Proposition 1.12. Let pX , Sq be an HHS. Then the conclusion of Definition 1.1.(7)1 holds
for all x, y P X and V, W P S with V Ĺ W .
1.4. Hierarchical spaces. Although most of our focus in this paper is on hierarchically
hyperbolic spaces, there are important contexts in which hyperbolicity of the spaces CU, U P
S is not used; notably, this is the case for the realization theorem (Theorem 3.1). Because of
the utility of a more general definition in later applications, we now define the following more
general notion of a hierarchical space; the reader interested only in the applications to the
mapping class group, 3–manifolds, cube complexes, etc. may safely ignore this subsection.
Definition 1.13 (Hierarchical space). A hierarchical space is a pair pX , Sq as in Definition 1.1, with X a quasigeodesic space and S an index set, where to each U P S we associate
a geodesic metric space CU , which we do not require to be hyperbolic. As before, there are
coarsely Lipschitz projections πU : X Ñ CU and relative projections ρU
V : CU Ñ CV whenever
U, V are non-orthogonal. We require all statements in the Definition 1.1 to hold, except for
hyperbolicity of the CU .
Remark 1.14. Let X be a quasigeodesic space that is hyperbolic relative to a collection P
of subspaces. Then X has a hierarchical space structure: the associated spaces onto which
we project are the various P, together with the space Xp obtained by coning off the elements
of P in X . When the elements of P are themselves hierarchically hyperbolic, we obtain a
hierarchically hyperbolic structure on X (see Section 9). Otherwise, the hierarchical structure
need not be hierarchically hyperbolic since Xp is the only one of the elements of S known to
be hyperbolic.
Remark 1.15. Other than hierarchically hyperbolic spaces, we are mainly interested in
hierarchical spaces pX , Sq where for all U P S, except possibly when U is Ď–minimal, we
have that CU is hyperbolic. This is the case, for example, in relatively hyperbolic spaces.
1.5. Consistency and partial realization points. The following definitions, which abstract the consistency inequalities from Definition 1.1.(4) and the partial realization axiom,
Definition 1.1.(8), play important roles throughout our discussion. We will consider this topic
in depth in Section 3.
ś
Definition 1.16 (Consistent). Fix κ ě 0 and let ~b P U PS 2CU be a tuple such that for each
U P S, the coordinate bU is a subset of CU with diamCU pbU q ď κ. The tuple ~b is κ–consistent
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
13
if, whenever V &W ,
(
min dW pbW , ρVW q, dV pbV , ρW
V q ďκ
and whenever V Ď W ,
(
min dW pbW , ρVW q, diamCV pbV Y ρW
V pbW qq ď κ.
Definition 1.17 (Partial realization point). Given θ ě 0 and a κ–consistent tuple ~b, we say
that x is a θ–partial realization point for tVj u Ď S if
(1) dVj px, bVj q ď θ for all j,
V
(2) for all j, we have dV px, ρVj q ď θ for any V P S with Vj Ď V , and
V
(3) for all W such that W &Vj for some j, we have dW px, ρWj q ď θ.
1.6. Level. The following definition is very useful for proving statements about hierarchically
hyperbolic spaces inductively. Although it is natural, and sometimes useful, to induct on
complexity, it is often better to induct on level:
Definition 1.18 (Level). Let pX, Sq be hierarchically hyperbolic. The level `U of U P S is
defined inductively as follows. If U is Ď-minimal then we say that its level is 1. The element
U has level k ` 1 if k is the maximal integer such that there exists V Ď U with `V “ k and
V ‰ U . Given U P S, for each ` ě 0, let S`U be the set of V Ď U with `U ´ `V ď ` and let
T`U “ S`U ´ S`´1
U .
1.7. Maps between hierarchically hyperbolic spaces.
Definition 1.19 (Hieromorphism). Let pX , Sq and pX 1 , S1 q be hierarchically hyperbolic
structures on the spaces X , X 1 respectively. A hieromorphism, consists of a map f : X Ñ X 1 ,
an injective map f ♦ : S Ñ S1 preserving nesting, transversality, and orthogonality, and, for
each U P S a map f ˚ pU q : CU Ñ Cpf ♦ pU qq which is a quasi-isometric embedding where
the constants are uniform over all elements of S and for which the following two diagrams
coarsely commute (with uniform constants) for all nonorthogonal U, V P S:
f
X
πU
CpU q
f ˚ pU q
/ X1
πf ♦ pU q
/ Cpf ♦ pU qq
and
CU
ρU
V
f ˚ pU q
/ Cpf ♦ pU qq
♦
CV
f ˚ pV
q
f pU q
ρf ♦ pV q
/ Cpf ♦ pV qq
where ρU
V : CU Ñ CV is the projection from Definition 1.1, which, by construction, is coarsely
constant if U &V or U Ď V . As the functions f, f ˚ pU q, and f ♦ all have distinct domains, it
is often clear from the context which is the relevant map; in that case we periodically abuse
notation slightly by dropping the superscripts and just calling all of the maps f .
Definition 1.20 (Automorphism, hierarchically hyperbolic group). An automorphism of the
hierarchically hyperbolic space pX , Sq is a hieromorphism f : pX , Sq Ñ pX , Sq such that f ♦
is bijective and each f ˚ pU q is an isometry; hence f : X Ñ X is a uniform quasi-isometry
by the distance formula (Theorem 4.5). The full automorphism group of pX , Sq is denoted
AutpSq.
The finitely generated group G is hierarchically hyperbolic if there exists a hierarchically
hyperbolic space pX , Sq and an action G Ñ AutpSq so that the uniform quasi-action of G
on X is metrically proper and cobounded and S contains finitely many G–orbits. Note that
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
14
if G is hierarchically hyperbolic by virtue of its action on the hierarchically hyperbolic space
pX , Sq, then pG, Sq is a hierarchically hyperbolic structure with respect to any word-metric
on G; for any U P S the projection is the composition of the projection X Ñ CU with a
G–equivariant quasi-isometry G Ñ X . In this case, pG, Sq (with the implicit hyperbolic
spaces and projections) is a hierarchically hyperbolic group structure.
Definition 1.21 (Equivariant hieromorphism). Let pX , Sq and pX 1 , G1 q be hierarchically
hyperbolic spaces and consider actions G Ñ AutpSq and G1 Ñ AutpS1 q. For each g P G, let
pfg , fg♦ , tfg˚ pU quq denote its image in AutpSq (resp., for g 1 P G1 we obtain pfg1 , fg♦1 , tfg˚1 pU quq in
AutpS1 q). Let φ : G Ñ G1 be a homomorphism. The hieromorphism pf, f ♦ , tf ˚ pU quq : pX , Sq Ñ
♦
pX 1 , S1 q is φ–equivariant if for all g P G and U P S, we have f ♦ pfg♦ pU qq “ fφpgq
pf ♦ pU qq and
the following diagram (uniformly) coarsely commutes:
f ˚ pU q
CU
fg˚ pU q
Cpfg♦ pU qq
f ˚ pfg♦ pU qq
/ Cpf ♦ pU qq
/
˚
pU qq
fφpgq
♦ ♦
Cpf pfg pU qqq
œ
œ
In this case, f : X Ñ X 1 is (uniformly) coarsely φ-equivariant in the usual sense. Also, we
♦
: S1 , and f ♦ : S Ñ S1 .
note for the reader that fg♦ : S , while fφpgq
2. Tools for studying HHSs
We now collect some basic consequences of the axioms that are used repeatedly throughout the paper. However, this section need not all be read in advance. Indeed, the reader
should feel free to skip this section on a first reading and return to it later when necessary.
Throughout this section, we work in a hierarchically hyperbolic space pX , Sq.
2.1. Handy basic consequences of the axioms.
Lemma 2.1 (“Finite dimension”). Let pX , Sq be a hierarchically hyperbolic space of complexity n and let U1 , . . . , Uk P S be pairwise-orthogonal. Then k ď n.
Proof. By Definition 1.1.(3), there exists W1 P S, not Ď–maximal, so that U2 , . . . , Uk Ď W1 .
Applying Definition 1.1 inductively yields a sequence Wk´1 Ď Wk´2 Ď . . . Ď W1 Ď S of
distinct elements, where S is Ď–maximal, so that Ui´1 , . . . , Uk Ď Wi for 1 ď i ď k ´ 1.
Hence k ď n by Definition 1.1.(5).
l
Lemma 2.2. There exists χ so that |S1 | ď χ whenever S1 Ď S does not contain a pair of
transverse elements.
Proof. Let S1 Ď S be a collection of pairwise non-transverse elements, and let n be large
enough that any collection of pairwise orthogonal (resp. pairwise Ď-comparable) elements
of S has cardinality at most n; the complexity provides such an n, by Definition 1.1.(5) and
Lemma 2.1. By Ramsey’s theorem, there exists N so that if |S1 | ą N then S1 either contains
a collection of elements, of cardinality at least n ` 1, that are pairwise orthogonal elements
or pairwise Ď-comparable. Hence, |S1 | ď N .
l
Lemma 2.3 (Consistency for pairs of points). Let x, y P X and V, W P S satisfy V &W and
dV px, yq, dW px, yq ą 10E. Then, up to exchanging V and W , we have dV px, ρW
V q ď E and
V
dW py, ρW q ď E.
V
Proof. Suppose that dV px, ρW
V q ą E, which then, by consistency, implies dW px, ρW q ď
V
E. Then, either dW py, ρW q ď 9E, in which case dW px, yq ď 10E, a contradiction, or
dW py, ρVW q ą E, in which case consistency implies that dV py, ρW
l
V q ď E.
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
15
Corollary 2.4. For x, y, V, W as in Lemma 2.3, and any z P X , there exists U P tV, W u
such that dU pz, tx, yuq ď 10E.
V
Proof. By Lemma 2.3, we may assume that dV px, ρW
V q, dW py, ρW q ď E. Suppose that
dW pz, tx, yuq ą 10E. Then dW pz, ρVW q ą 9E, so that, by consistency, dV pz, ρW
V q ď E,
whence dV pz, xq ď 2E.
l
The following is needed for Theorem 3.1 and in the forthcoming [DHS15].
Lemma 2.5 (Passing large projections up the Ď–lattice). For every C ě 0 there exists N
with the following property. Let V P S, let x, y P X , and let tSi uN
i“1 Ď SV be distinct and
satisfy dCSi px, yq ě E. Then there exists S P SV and i so that Si Ĺ S and dCS px, yq ě C.
Proof. The proof is by induction on the level k of a Ď-minimal S P SV into which each Si is
nested. The base case k “ 1 is empty.
Suppose that the statement holds for a given N “ N pkq when the level of S is at most
k. Suppose further that |tSi u| ě N pk ` 1q (where N pk ` 1q is a constant much larger than
N pkq that will be determined shortly) and there exists a Ď-minimal S P SV of level k ` 1
into which each Si is nested. There are two cases.
If dCS px, yq ě C, then we are done. If not, then the large link axiom (Definition 1.1.(6))
yields K “ KpCq and T1 , . . . , TK , each properly nested into S (and hence of level less than
k ` 1), so that any given Si is nested into some Tj . In particular, if N pk ` 1q ě KN pkq,
there exists a j so that at least N pkq elements of tSi u are nested into Tj . By the induction
hypothesis and the finite complexity axiom (Definition 1.1.(5)), we are done.
l
The next lemma is used in the proof of Proposition 4.12, on which the existence of hierarchy
paths (Theorem 4.4) relies. It is used again in Section 7 to construct coarse media.
Lemma 2.6 (Centers are consistent). There exists κ with the following property. Let x, y, z P
X . Let ~b “ pbW qW PS be so that bW is a point in CW with the property that there exists a
geodesic triangle in CW with vertices in πW pxq, πW pyq, πW pzq each of whose sides contains
a point within distance δ of bW . Then ~b is κ–consistent.
Proof. Recall that for w P tx, y, zu the tuple pπV pwqqV PS is E–consistent. Let U, V P S be
transverse. Then, by E–consistency, up to exchanging U and V and substituting z for one
U
U
of x, y, we have dV px, ρU
V q, dV py, ρV q ď E, so dV px, yq ď 3E (recall that the diameter of ρV
is at most E). Since bV lies at distance δ from the geodesic joining πV pxq, πV pyq, we have
dV pbV , ρU
V q ď 3E ` δ, whence the lemma holds with κ “ 3E ` δ.
Suppose now U Ĺ V . If bV is within distance 10E of ρU
V , then we are done. Otherwise, up to
permuting x, y, z, any geodesic rπV pxq, πV pyqs is 5E–far from ρU
V . By consistency of pπW pxqq,
pπW pyqq and bounded geodesic image, we have dU px, yq ď 10E, diamU pρVU pπV pyqqYπU pyqq ď
E, and diamU pρVU pbV Y πV pyqqq ď 10E. The first inequality and the definition of bU imply
dU pbU , yq ď 20E, and taking into account the other inequalities we get diamU pρVU pbV qYbU q ď
100E.
l
2.2. Partially ordering sets of maximal relevant elements of S. In this subsection, we
describe a construction used several times in this paper, including: in the proof of realization
(Theorem 3.1), in the construction of hierarchy paths (Theorem 4.4), and in the proof of
the distance formula (Theorem 4.5). We expect that this construction will have numerous
other applications, as is the case with the corresponding partial ordering in the case of the
mapping class group, see for example [BKMM12, BM11, CLM12].
ś
Fix x P X and a tuple ~b P U PS 2impπU q , where the U –coordinate bU is a set of diameter
at most some fixed ξ ě 0. For example, ~b could be the tuple pπU pyqq for some y P X .
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
16
In the remainder of this section, we choose κ ě 0 and require that ~b is κ–consistent. (Recall
that if ~b is the tuple of projections of a point in X , then ~b is E–consistent.)
Definition 2.7 (Relevant). First, fix θ ě 100 maxtκ, Eu. Then U P S is relevant (with
respect to x, ~b, θ) if dU px, bU q ą θ. Denote by Relpx, ~b, θq the set of relevant elements.
Let Relmax px, ~b, θq be a subset of Relpx, ~b, θq whose elements are pairwise Ď–incomparable
(for example, they could all be Ď–maximal in Relpx, ~b, θq, or they could all have the same
level). Define a relation ĺ on Relmax px, ~b, θq as follows. Given U, V P Relmax px, ~b, θq, we
have U ĺ V if U “ V or if U &V and dU pρVU , bU q ď κ. Figure 1 illustrates U ă V .
Figure 1. Heuristic picture of U ă V (for ~b the coordinates of y P X , for
concreteness). The idea is that “on the way” from x to y one “first encounters”
U and is forced to change the projection from πU pxq to πU pyq „ ρVU . In doing
so the projection to V is not affected.
Proposition 2.8. The relation ĺ is a partial order. Moreover, either U, V are ĺ–comparable
or U KV .
Proof. Clearly ĺ is reflexive. Antisymmetry follows from Lemma 2.9. Suppose that U, V are
ĺ–incomparable. If U KV , we are done, and we cannot have U Ď V or V Ď U , so suppose
U &V . Then, by ĺ–incomparability of U, V , we have dU pρVU , bU q ą κ and dV pρU
V , bV q ą
~
κ, contradicting κ–consistency of b. This proves the assertion that transverse elements of
Relmax px, ~b, θq are ĺ–comparable. Finally, transitivity follows from Lemma 2.10.
l
Lemma 2.9. The relation ĺ is antisymmetric.
Proof. If U ĺ V and U ‰ V , then dU pbU , ρVU q ď κ, so dU px, ρVU q ą θ ´ κ ě 99κ ą E. Then,
U
dV px, ρU
l
V q ď E, by consistency. Thus dV pbV , ρV q ą κ, and so, by definition V ł U .
Lemma 2.10. The relation ĺ is transitive.
Proof. Suppose that U ĺ V ĺ W . If U “ V or V “ W , then U ĺ W , and by Lemma 2.9,
we cannot have U “ W unless U “ V “ W . Hence suppose U &V and dU pρVU , bU q ď κ, while
~
V &W and dV pρW
V , bV q ď κ. By the definition of Relmax px, b, θq, we have dT px, bT q ą 100κ
for T P tU, V, W u.
W
V
V
We first claim that dV pρU
V , ρV q ą 10E. Indeed, dU pbU , ρU q ď κ, so dU pρU , xq ě 90κ,
U
whence dV pρV , xq ď E ď κ by E–consistency of the tuple pπT pxqqT PS . On the other hand,
U
W
dV pρW
V , bV q ď κ, so dV pρV , ρV q ą 10E as claimed. Hence, by Lemma 2.11, we have U &W .
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
17
Since diampimpπW qq ą 100κ — indeed, dW px, bW q ą 100κ and bW P impπW pX qq — partial
V
realization (Definition 1.1.(8)) provides a P X satisfying dW pa, tρU
W , ρW uq ě 10κ.
We thus have dU pa, ρW
U q ď E by E–consistency of pπT paqqT PS , and the same is true
V
with V replacing U . Hence dV pρU
V , aq ą E, so consistency implies dU pa, ρU q ď E. Thus
W
W
dU pρVU , ρW
U q ď 2E. Thus dU pbU , ρU q ď 2E ` κ ă 10κ, whence dU px, ρU q ą 50κ ą E, so
U
dW px, ρW q ď E by consistency and the fact that U &W . It follows that dW pbW , ρU
Wq ě
100κ ´ E ą κ, so, again by consistency, dU pbU , ρU
q
ď
κ,
i.e.,
U
ĺ
W
.
l
W
Lemma 2.11. Let U, V, W P S satisfy diampimpπU qq, diampimpπV qq, diampimpπW qq ą 10E,
W
and U &V, W &V , and dV pρU
V , ρV q ą 10E. Suppose moreover that U and W are Ď–incomparable.
Then U &W .
Proof. If U KW , then by the partial realization axiom (Definition 1.1.(8)) and the lower bound
on diameters, there exists an E–partial realization point x for tU, W u so that
dU pρVU , xq, dW pρVW , xq ą E.
W
U
This contradicts consistency since dV pρU
V , ρV q ą 10E; indeed, by consistency dV pρV , xq ď
W
U
W
E, dV pρV , xq ď E, i.e., dV pρV , ρV q ď 2E. Hence U &W .
l
2.3. Coloring relevant elements. In this subsection, the key result is Lemma 2.14, which
we will apply in proving the existence of hierarchy paths in Section 4.3.
Fix x, y P X . As above, let Relpx, y, 100Eq consist of those V P S for which dV px, yq ą
100E. Recall that, given U P S, we denote by T`U the set of V P SU such that `U ´`V “ `. In
particular, if V, V 1 P T`U and V Ď V 1 , then V “ V 1 . Let Rel`U px, y, 100Eq “ Relpx, y, 100EqX
T`U , the set of V Ď U so that dV px, yq ą 100E and `U ´ `V “ `.
By Proposition 2.8, the relation ĺ on Rel`U px, y, 100Eq defined as follows is a partial order:
1
V ĺ V 1 if either V “ V 1 or dV py, ρVV q ď E.
Definition 2.12 (Relevant graph). Denote by G the graph with vertex-set Rel`U px, y, 100Eq,
with two vertices adjacent if and only if the corresponding elements of Rel`U px, y, 100Eq are
orthogonal. Let G c denote the complementary graph of G, i.e., the graph with the same
vertices and edges corresponding to ĺ–comparability.
The next lemma is an immediate consequence of Proposition 2.8:
Lemma 2.13. Elements of V, V 1 P Rel`U px, y, 100Eq are adjacent in G if and only if they
are ĺ–incomparable.
Lemma 2.14 (Coloring relevant elements). Let χ be the maximal cardinality of a set of
pairwise orthogonal elements of T`U . Then there exists a χ–coloring of the set of relevant
elements of T`U such that non–transverse elements have different colors.
Proof. Since each clique in G — i.e., each ĺ–antichain in Rel`U px, y, 100Eq — has cardinality
at most χ, Dilworth’s theorem [Dil50, Theorem 1.1] implies that G can be colored with χ
colors in such a way that ĺ–incomparable elements have different colors; hence non-transverse
elements have different colors.
l
Remark 2.15. The constant χ provided by Lemma 2.14 is bounded by the complexity of
pX , Sq, by Lemma 2.2.
3. Realization of consistent tuples
The goal of this section is to prove Theorem 3.1. In this section we will work with a fixed
hierarchical space pX , Sq. We will use the concepts of consistency and partial realization
points; see Definition 1.16 and Definition 1.17.
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
18
Theorem 3.1 (Realization of consistent tuples). For each κ ě 1 there exist θe , θu ě 0 such
ś
that the following holds. Let ~b P W PS 2CW be κ–consistent; for each W , let bW denote the
CW –coordinate of ~b.
Then there exists x P X so that dW pbW , πW pxqq ď θe for all CW P S. Moreover, x is
coarsely unique in the sense that the set of all x which satisfy dW pbW , πW pxqq ď θe in each
CW P S, has diameter at most θu .
Proof. The main task is to prove the following claim about a κ–consistent tuple ~b:
Claim 1. Let tVj u be a family of pairwise orthogonal elements of S, all of level at most `.
Then there exists θe “ θe p`, κq ą 100Eκα and pairwise-orthogonal tUi u so that:
(1) each Ui is nested into some Vj ,
(2) for each Vj there exists some Ui nested into it, and
(3) any E–partial realization point x for tUi u satisfies dW pbW , xq ď θe for each W P S
for which there exists j with W Ď Vj .
Applying Claim 1 when ` “ `S , where S P S is the unique Ď–maximal element, along
with the Partial Realization axiom (Definition 1.1.(8)), completes the existence proof, giving
us a constant θe . If x, y both have the desired property, then dV px, yq ď 2θe ` κ for all
V P S, whence the uniqueness axiom (Definition 1.1.(9)) ensures that dpx, yq ď θu , for an
appropriate θu . Hence to prove the theorem it remains to prove Claim 1, which we do now.
The claim when ` “ 1 follows from the partial realization axiom (Definition 1.1.8), so we
assume that the claim holds for ` ´ 1 ě 1, with θe p` ´ 1, κq “ θe1 , and prove it for level `.
Reduction to the case |tVj u| “ 1: It suffices to prove the claim in the case where tVj u
has a single element, V . To see this, note that once we prove the claim for each Vj separately,
yielding a collection of pairwise-orthogonal sets tUij Ď Vj u with the desired properties, then
we take the union of these sets to obtain the claim for the collection tVj u.
The case tVj u “ tV u: Fix V P S so that `V “ `. If for each x P X that satisfies
dV px, bV q ď E we have dW pbW , xq ď 100Eκα for W P SV , then the claim follows with
tUi u “ tV u. Hence, we can suppose that this is not the case.
We are ready for the main argument, which is contained in Lemma 3.2 below. We will
construct tUi u incrementally, using Lemma 3.2, which essentially says that either we are done
at a certain stage or we can add new elements to tUi u.
We will say that the collection U of elements of SV is totally orthogonal if any pair of
distinct elements of U are orthogonal. Given a totally orthogonal family U we say that
W P SV is U–generic if there exists U P U so that W is not orthogonal to U . Notice that no
W is H–generic.
A totally orthogonal collection U Ď SV is C–good if any E–partial realization point x
for U has the property that for each W P SV we have dW px, bW q ď C. (Notice that our
goal is to find such U.) A totally orthogonal collection U Ď SV is C–generically good if any
E–partial realization point x for U has the property that for each U–generic W P SV we have
dW px, bW q ď C (e.g., for U “ H).
We can now quickly finish the proof of the claim using Lemma 3.2 about extending generically good sets, which we state and prove below. Start with U “ H. If U is C–good for
C “ 100Eκα, then we are done. Otherwise we can apply Lemma 3.2 and get U1 “ U1 as in
the lemma. Inductively, if Un is not 10n C–good, we can apply the lemma and extend Un to a
new totally orthogonal set Un`1 . Since there is a bound on the cardinality of totally orthogonal sets by Lemma 2.1, in finitely many steps we necessarily get a good totally orthogonal
set, and this concludes the proof of the claim, and hence of the theorem.
l
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
19
Lemma 3.2. For every C ě 100Eκα the following holds. Let U Ď SV ´ tV u be totally
orthogonal and C–generically good but not C–good. Then there exists a totally orthogonal,
10C–generically good collection U1 Ď SV with U Ĺ U1 .
Proof. Let x0 be an E–partial realization point for U so that there exists some W Ď V for
which dW pbW , x0 q ą C.
The idea is to try to “move towards” ~b starting from x0 , by looking at all relevant elements
of SV that lie between them and finding out which ones are the “closest” to ~b.
Let Vmax be the set of all W Ď V for which:
(1) dW pbW , x0 q ą C and
(2) W is not properly nested into any element of SV satisfying the above inequality.
We now establish two facts about Vmax .
Applying Proposition 2.8 to partially order Vmax : For U, U 1 P Vmax , write U ĺ U 1
1
if either U “ U 1 or U &U 1 and dU pρU
U , bU q ď 10Eκ; this is a partial order by Proposition 2.8,
which also implies that if U, U 1 P Vmax are transverse then they are ĺ-comparable. Hence
any two ĺ–maximal elements of Vmax are orthogonal, and we denote by V1max the set of
ĺ–maximal (hence pairwise-orthogonal) elements of Vmax .
Finiteness of Vmax : We now show that |Vmax | ă 8. By Lemma 2.2 and Ramsey’s
theorem, if Vmax was infinite then it would contain an infinite subset of pairwise transverse
elements, so, in order to conclude that |Vmax | ă 8, it suffices to bound the cardinality of a
pairwise-transverse subset of Vmax .
Suppose that W1 ă ¨ ¨ ¨ ă Ws P Vmax are pairwise transverse. By partial realization
s
(Definition 1.1.(8)) there exists z P X such that dWs pz, bWs q ď α and dWi pρW
Wi , zq ď α
s
~
for each i ‰ s, and such that dV pz, ρW
V q ď α. By consistency of b and bounded geodesic
s
image, ρW
V has to be within distance 10Eκ of a geodesic in CV from x0 to bV . In particular
dV px0 , zq ď θe1 ` 100Eκα ` 10Eκ. Also, for each i ‰ s,
Ws
s
dWi px0 , zq ě dWi px0 , bWi q ´ dWi pbWi , ρW
Wi q ´ dWi pρWi , zq
ě 100Eκα ´ 10Eκ ´ α ě 50Eκα ě 50E.
Ws
s
Indeed, dWi pbWi , ρW
Wi q ď 10Eκ since Wi ă Ws , while dWi pρWi , zq ď α by our choice of z.
Lemma 2.5 now provides the required bound on s.
Choosing U1 : Since `U ă `V for all U P V1max , by induction there exists a totally
orthogonal set tUi u so that any E–partial realization point x for tUi u satisfies dT pbT , xq ď θe1
for each T P S nested into some U P V1max . Let U1 “ tUi u Y U.
Choose such a partial realization point x and let W Ď V be U1 –generic. Our goal is to
bound dW px, bW q, and we will consider 4 cases.
If there exists U P U that is not orthogonal to W , then we are done by hypothesis, since
any E–partial realization point for U1 is also an E–partial realization point for U.
Hence, from now on, assume that W is orthogonal to each U P U, i.e. W is not U–generic.
If W Ď U for some U P V1max , then we are done by induction.
Suppose that W &U for some U P V1max . For each Ui Ď U — and our induction hypothesis
i
implies that there is at least one such Ui — we have dW px, ρU
W q ď E since x is a partial
realization point for tUi u and either Ui Ď W or Ui &W (since W is U1 –generic but not
U–generic). The triangle inequality therefore yields:
U
U
i
dW px, bW q ď E ` dW pρU
W , ρW q ` dW pbW , ρW q.
U
U
i
By Definition 1.1.(4), dW pρU
W , ρW q ď E, and we will show that dW pbW , ρW q ď 2C, so that
dW px, bW q ď 2E ` 2C.
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
20
W
Suppose, for a contradiction, that dW pbW , ρU
W q ą 2C. If dU pρU , x0 q ď E, then
dU pρW
U , bU q ě C ´ E ą κ,
by consistency, whence dW pρU
W , bW q ď κ, a contradiction.
U
On the other hand, if dU pρW
U , x0 q ą E, then dW px0 , ρW q ď E by consistency. Hence
dW px0 , bW q ě 2C ´ E. Hence there exists a Ď–maximal W 1 ‰ V with the property that
W Ď W 1 Ď V and dW 1 px0 , bW q ą C (possibly W 1 “ W ). Such a W 1 is in Vmax by definition.
Since W &U , and W 1 and U are Ď–incomparable, W 1 &U . Thus U and W 1 are ĺ–
comparable, by Proposition 2.8. Since W 1 ‰ U and U is ĺ–maximal, we have W 1 ĺ U , i.e.,
W1
dW 1 pbW 1 , ρU
W 1 q ď 10Eκ. Since ĺ is antisymmetric, by Lemma 2.9, we have dU pbU , ρU q ą
W1
W
10Eκ. Since dU pρW
U , ρU q ď E (from Definition 1.1.(4)), we have dU pbU , ρU q ą 10Eκ ´ E ą
U
κ, since E ě 1, so, by consistency, dW pbW , ρW q ď κ, a contradiction.
Finally, suppose U Ĺ W for some U P V1max . In this case, by Ď–maximality of U , we
i
have dW px0 , bW q ď C. Also, dW px, ρU
W q ď E for any Ui Ď U since x is a partial realization
Ui
U
point, so that dW px, ρU
W q ď 2E, since dW pρW , ρW q ď E by Definition 1.1.(4). If dW px, bW q ą
2C, then we claim dU px0 , bU q ď 10Eκ, a contradiction. Indeed, any geodesic in CW from
πW px0 q to bW does not enter the E–neighborhood of ρU
W . By bounded geodesic image,
W
W
diamU pρU pπW px0 qq Y ρU pbW qq ď E and by consistency, diamU pρW
U pπW px0 qq Y πU px0 qq ď E
and diamU pρW
pb
qYb
q
ď
κ,
and
we
obtain
the
desired
bound
on
dU px0 , bU q. This completes
W
U
U
the proof of the lemma.
l
4. Hierarchy paths and the distance formula
Throughout this section, fix a hierarchically hyperbolic space pX , Sq.
4.1. Definition of hierarchy paths and statement of main theorems. Our goal is to
deduce the existence of hierarchy paths (Theorem 4.4) from the other axioms and to prove
the distance formula (Theorem 4.5).
Definition 4.1 (Quasigeodesic, unparameterized quasigeodesic). A pD, Dq–quasigeodesic in
the metric space M is a pD, Dq–quasi-isometric embedding f : r0, `s Ñ M ; we allow f to be
a coarse map, i.e., to send points in r0, `s to uniformly bounded sets in M . A (coarse) map
f : r0, `s Ñ M is a pD, Dq–unparameterized quasigeodesic if there exists a strictly increasing
function g : r0, Ls Ñ r0, `s such that f ˝ g : r0, Ls Ñ M is a pD, Dq–quasigeodesic and for each
j P r0, Ls X N, we have diamM pf pgpjqq Y f pgpj ` 1qqq ď D.
Definition 4.2 (Hierarchy path). For D ě 1, a (not necessarily continuous) path γ : r0, `s Ñ
X is a D–hierarchy path if
(1) γ is a pD, Dq-quasigeodesic,
(2) for each W P S, the path πW ˝ γ is an unparameterized pD, Dq–quasigeodesic.
Notation 4.3. Given A, B P R, we denote by ttAuuB the quantity which is A if A ě B and
0 otherwise. Given C, D, we write A —C,D B to mean C ´1 A ´ D ď B ď CA ` D.
Theorem 4.4 (Existence of Hierarchy Paths). Let pX , Sq be hierarchically hyperbolic. Then
there exists D0 so that any x, y P X are joined by a D0 -hierarchy path.
Theorem 4.5 (Distance Formula). Let pX, Sq be hierarchically hyperbolic. Then there exists
s0 such that for all s ě s0 there exist constants K, C such that for all x, y P X ,
ÿ
dX px, yq —pK,Cq
ttdW px, yquus .
W PS
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
21
The proofs of the above two theorems are intertwined, and we give the proof immediately below. This relies on several lemmas, namely Lemma 4.11, proved in Section 4.3, and
Lemmas 4.19 and 4.18, proved in Section 4.4.
Proof of Theorems 4.5 and 4.4. The lower bound demanded by Theorem 4.5 is given by
Lemma 4.19 below. By Lemma 4.11 and Lemma 4.18, there is a monotone path (see Definition 4.8) whose length realizes the upper bound on dX px, yq, and the same holds for any
subpath of this path, which is therefore a hierarchy path, proving Theorem 4.4 and completing the proof of Theorem 4.5.
l
4.2. Good and proper paths: definitions. We now define various types of (non-continuous)
paths in X that will appear on the way to hierarchy paths.
Definition 4.6 (Discrete path). A K–discrete path is a map γ : I Ñ X , where I is an interval
in Z and dX pγpiq, γpi ` 1qq ď K whenever i, i ` 1 P I. The length |α| of a discrete path α is
max I ´ min I.
Definition 4.7 (Efficient path). A discrete path α with endpoints x, y is K–efficient if
|α| ď KdX px, yq.
Definition 4.8 (Monotone path). Given U P S, a K–discrete path α and a constant L, we
say that α is L–monotone in U if whenever i ď j we have dU pαp0q, αpiqq ď dU pαp0q, αpjqq`L.
A path which is L–monotone in U for all U P S is said to be L–monotone.
Definition 4.9 (Good path). A K–discrete path that is L–monotone in U is said to be
pK, Lq–good for U . Given S1 Ď S, a path α that is pK, Lq–good for each V P S1 is pK, Lq–
good for S1 .
Definition 4.10 (Proper path). A discrete path α : t0, . . . , nu Ñ X is pr, Kq–proper if for
0 ď i ă n ´ 1, we have dX pαpiq, αpi ` 1qq P rr, r ` Ks and dX pαpn ´ 1q, αpnqq ď r ` K.
Observe that pr, Kq-properness is preserved by passing to subpaths.
4.3. Good and proper paths: existence. Our goal in this subsection is to join points in
X with proper paths, i.e., to prove Lemma 4.11. This relies on the much more complicated
Proposition 4.12, which produces good paths (which are then easily made proper).
Lemma 4.11. There exists K so that for any r ě 0, any x, y P X are joined by a Kmonotone, pr, Kq–proper discrete path.
Proof. Let α0 : t0, . . . , n0 u Ñ X be a K–monotone, K–discrete path joining x, y, which exists
by Proposition 4.12. We modify α0 to obtain the desired path in the following way. Let j0 “ 0
and, proceeding inductively, let ji be the minimal j ď n such that either dX pα0 pji´1 q, α0 pjqq P
rr, r ` Ks or j “ n. Let m be minimal so that jm “ n and define α : t0, . . . , mu Ñ X
by αpjq “ α0 pij q. The path α is pr, Kq-proper by construction; it is easily checked that
K–monotonicity is not affected by the above modification; the new path is again discrete,
although for a larger discreteness constant.
l
It remains to establish the following proposition, whose proof is postponed until the end
of this section, after several preliminary statements have been obtained.
Proposition 4.12. There exists K so that any x, y P X are joined by path that is pK, Kq–
good for each U P S.
Definition 4.13 (Hull of a pair of points). For each x, y P X , θ ě 0, let Hθ px, yq be the set
of all p P X so that, for each W P S, the set πW ppq lies at distance at most θ from a geodesic
in CW joining πW pxq to πW pyq. Note that x, y P Hθ px, yq.
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
22
Remark 4.14. The notion of a hull is generalized in Section 6 to hulls of arbitrary finite
sets, but we require only the version for pairs of points in this section.
Lemma 4.15 (Retraction onto hulls). There exist θ, K ě 0 such that, for each x, y P X ,
there exists a pK, Kq–coarsely Lipschitz map r : X Ñ Hθ px, yq that restricts to the identity
on Hθ px, yq.
Proof. Let κ be the constant from Lemma 2.6, let θe be chosen as in the realization theorem
ś
(Theorem 3.1), and let p P X ´ Hθe px, yq. Define a tuple ~b “ pbpW q P W PS 2CW so that bpW is
on a geodesic in CW from πW pxq to πW pyq and is within distance δ of the other two sides of
a triangle with vertices in πW pxq, πW pyq, πW ppq. By the realization theorem (Theorem 3.1),
there exists rppq P Hθe px, yq so that dW pπW prppqq, bpW q ď θe . For p P Hθe px, yq, let rppq “ p.
To see that r is coarsely Lipschitz, it suffices to bound dX prppq, rpqqq when p, q P X satisfy
dX pp, qq ď 1. For such p, q we have dW pbpW , bqW q ď 100E, so that Theorem 3.1 implies
dX prppq, rpqqq ď θu p100Eq, as required.
l
Corollary 4.16. There exist θ, K ě 0 such that, for each x, y P X , there exists a K–discrete
and K–efficient path that lies in Hθ px, yq and joins x to y.
Proof. We can assume that dX px, yq ě 1. Since X is a quasigeodesic space, there exists C “
CpX q ě 1 and a pC, Cq–quasi-isometric embedding γ : r0, Ls Ñ X with γp0q “ x, γpLq “ y.
Let ρ be the path obtained by restricting r ˝ γ : r0, Ls Ñ Hθ px, yq to r0, Ls X N, where r is
the retraction obtained in Lemma 4.15. Then dX pρpiq, ρpi ` 1qq ď 10KC since r is pK, Kq–
coarsely Lipschitz and γ is pC, Cq–coarsely Lipschitz, i.e., ρ is 10KC–discrete. Finally, ρ is
efficient because L ď CdX px, yq ` C ď 2CdX px, yq.
l
The efficiency part of the corollary is used in Lemma 4.19.
4.3.1. Producing good paths. We will need the following lemma, which is a special case of
Proposition 6.4.2. We give a proof in the interest of a self-contained exposition.
Lemma 4.17. For any θ0 there exists a constant θ such that for every x, y P X and every
x1 , y 1 P Hθ0 px, yq, we have Hθ0 px1 , y 1 q Ď Hθ px, yq.
Proof. For any z P Hθ0 px1 , y 1 q and W P S the projection πW pzq lies 2pδ ` θ0 q–close to a
geodesic in CW from πW pxq to πW pyq, by a thin quadrilateral argument.
l
We now prove the main proposition of this subsection.
Proof of Proposition 4.12. Recall that, for ` ě 0 and U P S, the set S`U consists of those
V P SU with `U ´ `V ď `, and that T`U consists of those V P SU with `U ´ `V “ `.
We prove by induction on ` that there exist θ, K such that for any ` ě 0, x, y P X and
U P S, there is a path α in Hθ px, yq connecting x to y such that α is pK, Kq–good for S`U .
It then follows that for any x, y P X , there exists a path α in Hθ px, yq connecting x to y such
that α is pK, Kq–good for S; this latter statement directly implies the proposition.
For a, b P X , denote by ra, bsW a geodesic in CW from πW paq to πW pbq.
Fix U P S.
The case ` “ 0: In this case, S0U “ tU u. By Corollary 4.16, there exists θ0 , K and a
K–discrete path α01 : t0, . . . , ku Ñ Hθ0 px, yq joining x to y.
Similarly, for each x1 , y 1 P Hθ0 px, yq there exists a K–discrete path β contained in Hθ0 px1 , y 1 q,
joining x1 to y 1 , and recall that Hθ0 px1 , y 1 q is contained in Hθ px, yq for a suitable θ in view of
Lemma 4.17.
We use the term straight path to refer to a path, such as β, which for each V P S projects
uniformly close to a geodesic of CpV q.
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
23
We now fix U P S, and, using the observation in the last paragraph explain how to modify
α01 to obtain a K–discrete path α0 in Hθ px, yq that is K–monotone in U ; the construction
will rely on replacing problematic subpaths with straight paths.
A point t P t0, . . . , ku is a U –omen if there exists t1 ą t so that dU pα01 p0q, α01 ptqq ą
dU pα01 p0q, α01 pt1 qq ` 5KE. If α01 has no U –omens, then we can take α0 “ α01 , so suppose
that there is a U –omen and let t0 be the minimal U –omen, and let t10 ą t0 be maximal so
that dU pα01 p0q, α01 pt0 qq ą dU pα01 p0q, α01 pt10 qq. Inductively define tj to be the minimal U –omen
with tj ě t1j´1 , if such tj exists; and when tj exists, we define t1j to be maximal in t0, . . . , ku
satisfying dU pα01 p0q, α01 ptj qq ą dU pα01 p0q, α01 pt1j qq. For each j ě 0, let x1j “ α01 ptj q and let
yj1 “ α01 pt1j q. See Figure 2.
α01 pt1j q
Ðx
α01 ptj q
yÑ
Figure 2. The picture above shows part of α01 in X , and that below shows its
projection to U . The point tj is an omen, as witnessed by the point marked
with a square. Inserting the dashed path βj , and deleting the corresponding
subpath of α01 , makes tj cease to be an omen.
For each j, there exists a K–discrete path βj which lies in Hθ0 px1j , yj1 q Ď Hθ px, yq and is a
straight path from x1j to yj1 . Let α0 be obtained from α01 by replacing each α01 prtj , t1j sq with
βj . Clearly, α0 connects x to y, is K–discrete, and is contained in Hθ px, yq. For each j we
have that diamCU pβj q ď dU px1j , yj1 q ` 2θ0 .
Notice that dU px1j , yj1 q ă 2KE ` 10θ0 . In fact, since α01 p0q, α01 ptj q, α01 pt1j q lie θ0 -close to a
common geodesic and dU pα01 p0q, α01 ptj qq ě dU pα01 p0q, α01 pt1j qq, we would otherwise have
dU pα01 p0q, α01 ptj qq ´ dU pα01 p0q, α01 pt1j qq ě dU px1j , yj1 q ´ 5θ0 ě 2KE ` θ0 .
However, dU pα01 ptj q, α01 ptj ` 1qq ď 2KE because of K–discreteness and the projection map
to CU being E-coarsely Lipschitz. Hence, the inequality above implies
dU pα01 p0q, α01 ptj qq ą dU pα01 p0q, α01 pt1j qq ` 2KE ě dU pα01 ptj q, α01 ptj ` 1qq,
which contradicts the maximality of t1j . (Notice that t1j ‰ k, and hence t1j ` 1 P t0, . . . , ku
because dU pα01 p0q, α01 pt1j qq ` θ0 ă dU pα01 p0q, α01 ptj qq ď dU pα01 p0q, α01 pkqq ` θ0 .)
In particular, we get diamCU pβj q ď 2KE ` 12θ0 , and it is then easy to check that α0 is
maxt5KE, 2KE ` 12θ0 u–monotone in U . By replacing K with maxt5KE, 2KE ` 12θ0 u, we
thus have a K–discrete path α0 Ă Hθ px, yq that joins x, y and is K–monotone in U .
We now proceed to the inductive step. Specifically, we fix ` ě 0 and we assume there exist
θind , K such that for any x, y P X and U P S, there is a path α in Hθind px, yq connecting x
to y such that α is pK, Kq–good for S`U .
Let now x, y, U be as above.
The coloring: For short, we will say that V P S is A–relevant if dU px, yq ě A, see
Definition 2.7. Notice that to prove that a path in Hθ px, yq is monotone, it suffices to restrict
our attention to only those W P S which are, say, 10KE–relevant.
By Lemma 2.14, there exists χ ě 0, bounded by the complexity of X , and a χ–coloring c
of the 10KE–relevant elements of T`U such that cpV q “ cpV 1 q only if V &V 1 . In other words,
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
24
Ůχ´1 ´1
the set of 10KE–relevant elements of T`U has the form i“0
c piq, where c´1 piq is a set of
pairwise-transverse relevant elements of T`U .
Induction hypothesis: Given p ă χ ´ 1, assume by induction (on ` and p) that there
exist θp ě θind , Kp ě K, independent of x, y, U , and a path αp : t0, . . . , ku Ñ Hθp px, yq,
Ů
joining x, y, that is pK, Kq–good for pi“0 c´1 piq and good for S`´1
U .
Resolving backtracks in the next color: Let θp`1 be provided by Lemma 4.17 with
input θp . We will modify αp to construct a Kp`1 –discrete path αp`1 in Hθp`1 px, yq, for some
Ůp`1 ´1
Kp`1 ě Kp , that joins x, y and is pKp`1 , Kp`1 q–good in i“0
c piq.
Notice that we can restrict our attention to the set Cp`1 of 100pKp E`θp q–relevant elements
of c´1 pp ` 1q.
A point t P t0, . . . , ku is a pp ` 1q–omen if there exists V P Cp`1 and t1 ą t so that
dV pαp p0q, αp ptqq ą dV pαp p0q, αp pt1 qq ` 5Kp E. If αp has no pp ` 1q–omens, then we can
take αp`1 “ αp , since αp is good in each V with cpV q ă p ` 1. Therefore, suppose that
there is a pp ` 1q–omen, let t0 be the minimal pp ` 1q–omen, witnessed by V0 P Cp`1 . We
can assume that t0 satisfies dV0 ptx, yu, αp pt0 qq ą 10Kp E. Let t10 ą t0 be maximal so that
dV0 pαp p0q, αp pt0 qq ą dV0 pαp p0q, αp pt10 qq. In particular dV0 py, αp pt10 qq ě 10E.
Let x10 “ α0 pt0 q and y01 “ α0 pt10 q. Inductively, define tj as the minimal pp ` 1q–omen,
witnessed by Vj P Cp`1 , with tj ě t1j´1 , if such tj exists and let t1j be maximal so that
dVj pαp p0q, αp ptj qq ą dVj pαp p0q, αp pt1j qq and dVj py, αp pt1j qq ą 100E. We can assume that tj
satisfies dVj ptx, yu, αp ptj qq ą 10Kp E. Also, let x1j “ αp ptj q, yj1 “ αp pt1j q.
Let βj be a path in Hθp px1j , yj1 q joining x1j to yj1 that is pKp , Kp q–good for each relevant V
with cpV q ď p and each relevant V P S`´1
U . Such paths can be constructed by induction.
By Lemma 4.17 βj lies in Hθp`1 px, yq. Let αp`1 be obtained from αp by replacing each
αp pttj , . . . , t1j uq with βj . Clearly, αp`1 connects x to y, is Kp –discrete, and is contained in
Hθp`1 px, yq.
We observe that the same argument as in the case ` “ 0 gives dVj px1j , yj1 q ď 2Kp E ` 10θp .
Verification that αp`1 is good for current colors: We next check that each βj is
Ůp`1 ´1
103 pKp E ` θp q–monotone in each W P i“0
c piq. We have to consider the following cases.
(We can and shall assume below W is 100pKp E ` θp q–relevant.)
‚ If W Ď Vj , then W “ Vj , since `W “ `Vj . Since the projections on CW of the
endpoints of the straight path βj coarsely coincide, βj is p2Kp E ` 12θp q–monotone
in W . (See the case ` “ 0.)
‚ Suppose Vj Ď W , and Vj ‰ W . We claim that the projections of the endpoints of βj
lie at a uniformly bounded distance in CW .
V
We claim that ρWj has to be E–close to either rx, x1j sW or ryj1 , ysW . In fact, if this
was not the case, we would have
dVj px, yq ď dVj px, x1j q ` dVj px1j , yj1 q ` dVj pyj1 , yq ď 2E ` 2Kp E ` 10θp ,
where we applied bounded geodesic image (Definition 1.1.(7)) to the first and last
terms.
This is a contradiction with Vj being 100pKp E ` θp q–relevant.
Suppose by contradiction that dW px1j , yj1 q ě 500pKp E ` θp q. Suppose first that
V
V
ρWj is E–close to rx, x1j sW . Then, by monotonicity, ρWj is E–far from rαp pt1j q, ysW .
V
By bounded geodesic image, this contradicts dVj py, αp pt1j qq ě 10E. If instead ρWj is
E–close to ryj1 , ysW , then by bounded geodesic image we have dVj px, αp ptj qq ď E,
contradicting that tj is an omen witnessed by Vj . See Figure 3.
Hence dW px1j , yj1 q ě 500pKp E ` θp q and βj is 103 pKp E ` θp q–monotone in W .
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
25
y
α0 pt1j ´ 1q
yj1 “ α0 pt1j q
V
ρWj
x1j
x
CW
Figure 3. The situation in CW .
‚ Suppose W &Vj . We again claim that the projections of the endpoints of βj are
V
uniformly close in CW , by showing that they both coarsely coincide with ρWj . Since
W
Vj is relevant, either dVj px, ρW
Vj q ě E or dVj py, ρVj q ě E. Thus, by consistency,
V
dW pρWj , tx, yuq ď E. Suppose for a contradiction, that dW px1j , yj1 q ą 100pKp E ` θp q.
V
V
We consider separately the cases where dW px, ρWj q ď E and dW py, ρWj q ď E
V
V
First, suppose that dW px, ρWj q ď E. Then dW py, ρWj q ě 10Kp E ´ E ą E, so by
1
1
consistency, dVj py, ρW
Vj q ď E. If dVj px, txj , yj uq ą E, then consistency implies that
V
V
dW px1j , ρWj q ď E and dW pyj1 , ρWj q ď E, whence dW px1j , yj1 q ď 2E, a contradiction.
If dVj px, tx1j , yj1 uq ď E, then since dVj px1j , yj1 q ď 2Kp E ` 10θp , we have dVj px, x1j q ď
5Kp E ` 10θp , contradicting that, since tj was a pp ` 1q–omen witnessed by Vj , we
must have dVj px, x1j q ą 5Kp E.
V
Second, suppose dW py, ρWj q ď E. Then by relevance of W and consistency,
1
1
1
dVj px, ρW
Vj q ď E. As above, we have dVj pxj , xq ą 5Kp E ` 10θp , so dVj px, txj , yj uq ą
Kp E ą 3E (since dVj px1j , yj1 q ď 2Kp E ` 10θp and we may assume Kp ą 3), so
1
1
1
1
dVj pρW
Vj , txj , yj uq ą E. Thus, by consistency, πW pxj q, πW pyj q both lie at distance at
V
most E from ρWj , whence dW px1j , yj1 q ď 3E.
‚ Finally, suppose that W KVj . Then either cpW q ă cpVj q and βj is Kp –monotone in
W , or W is irrelevant.
Hence, each βj is 103 pKp E ` θp q–monotone in each W P c´1 pt0, . . . , p ` 1uq.
Verification that αp`1 is monotone: Suppose that there exist t, t1 such that t ă t1 and
dV pαp`1 p0q, αp`1 ptqq ą dV pαp`1 p0q, αp`1 pt1 qq ` 104 pKp E ` θp q for some V P c´1 pt0, . . . , p `
1uq. We can assume t, t1 R Yi pti , t1i q. Indeed, if t P pti , t1i q (respectively, t1 P ptj , t1j q), then since
all βm are 103 pKp E ` θp q–monotone, we can replace t with t1i (respectively, t1 with tj ). After
such a replacement, we still have dV pαp`1 p0q, αp`1 ptqq ą dV pαp`1 p0q, αp`1 pt1 qq ` 5Kp E.
Let i be maximal so that t1i ď t (or let i “ ´1 if no such t1i exists). By definition of ti`1 ,
we have ti`1 ď t, and hence ti`1 “ t. But then t1i`1 ą t1 , which is not the case.
Conclusion: Continuing the above procedure while p ă χ produces the desired path
which is pKχ , Kχ q–good for S`U . In particular, when U “ S is Ď–maximal and ` is the
length of a maximal Ď–chain, the proposition follows.
l
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
26
4.4. Upper and lower distance bounds. We now state and prove the remaining lemmas
needed to complete the proof of Theorem 4.5 and 4.4.
Lemma 4.18 (Upper bound). For every K, s there exists r with the following property. Let
α : t0, . . . , nu Ñ X be a K-monotone, pr, Kq-proper discrete path connecting x to y. Then
ÿ
|α| ´ 1 ď
ttdW px, yquus .
W PS
Proof. Let r “ rpK, E, sq be large enough that, for any a, b P X , if dX pa, bq ě r, then there
exists W P S so that dW pa, bq ě 100KEs. This r is provided by Definition 1.1.(9).
For 0 ď j ď n ´ 1, choose Vj P S so that dVj pαpjq, αpj ` 1qq ě 100KEs. By monotonicity
of α in Vj , for any j 1 ą j we have
dVj pαp0q, αpj 1 qq ě dVj pαp0q, αpjqq ` 50KEs.
ř
It follows by induction on j ď n that W PS ttdW pαp0q, αpjqquus ě mintj, n ´ 1u.
l
Lemma 4.19 (Lower bound). There exists s0 such that for all s ě s0 , there exists C with
the following property.
1 ÿ
ttdW px, yquus .
dX px, yq ě
C W PS
Proof. From Corollary 4.16, we obtain a K–discrete path α : t0, nu Ñ X joining x, y and having the property that the (coarse) path πV ˝ α : t0, . . . , nu Ñ CV lies in the K–neighborhood
of a geodesic from πV pxq to πV pyq. Moreover, α is K–efficient, by the same corollary.
Fix s0 ě 103 KE. A checkpoint for x, y in V P S is a ball Q in CV so that πV ˝ α intersects
Q and dV ptx, yu, Qq ě 10KE ` 1. Note that any ball of radius 10KE centered on a geodesic
from πV pxq to πV pyq is a checkpoint for x, y in V , provided
Q it isUsufficiently far from tx, yu.
px,yq
3
For each V P Relpx, y, 10 KEq, choose a set CV of dV 10
checkpoints for x, y in V ,
subject to the requirement that dV pC1 , C2 q ě 10KE for all distinct C1 , C2 P CV . For each
V P Relpx, y, 103 KEq, we have 10|CV | ě dV px, yq, so
ÿ
|CV | ě
V PS
1 ÿ
ttdW px, yquus0 .
10 W PS
Each j P t0, . . . , nu is a door if there exists V P Relpx, y, 103 KEq and C P CV such that
πV pαpjqq P C but πV pαpj ´ 1qq R C. The multiplicity of a door j is the cardinality of the
set Mpjq of V P Relpx, y, 103 KEq for which there exists C P CV with πV pαpjqq P C and
πV pαpj ´ 1qq R C. Since CV is a set of pairwise-disjoint checkpoints, j is a door for at most
one element of CV , for each V . Hence the multiplicity of j is precisely the total number of
checkpoints in YV PRelpx,y,103 KEq CV for which j is a door.
We claim that the set Mpjq does not contain a pair of transverse elements. Indeed, suppose
that U, V P Mpjq, satisfy U &V . Let QV P CV , QU P CU be the checkpoints containing
πV pαpjqq, πU pαpjqq respectively, so that dU pαpjq, tx, yuq, dV pαpjq, tx, yuq ě 10KE `1 ą 10E,
contradicting Corollary 2.4. Lemma 2.2řthus gives |MV | ď χ. Now, |α| is at least the number
of doors in t0, . . . , nu, whence |α| ě χ1 V PS |CV |. Since α is K–efficient, we obtain
dX px, yq ě
For s ě s0 ,
ř
W PS ttdW px, yquus
ď
ř
ÿ
1
ttdW px, yquus0 .
10χK W PS
W PS ttdW px, yquus0
, so the claim follows.
l
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
27
5. Hierarchical quasiconvexity and gates
We now introduce the notion of hierarchical quasiconvexity, which is essential for the
discussion of product regions, the combination theorem of Section 8, and in the forthcoming [DHS15].
Definition 5.1 (Hierarchical quasiconvexity). Let pX , Sq be a hierarchically hyperbolic
space. Then Y Ď X is k–hierarchically quasiconvex, for some k : r0, 8q Ñ r0, 8q, if the
following hold:
(1) For all U P S, the projection πU pYq is a kp0q–quasiconvex subspace of the δ–
hyperbolic space CU .
ś
(2) For all κ ě 0 and κ-consistent tuples ~b P U PS 2CU with bU Ď πU pYq for all U P S,
each point x P X for which dU pπU pxq, bU q ď θe pκq (where θe pκq is as in Theorem 3.1)
satisfies dpx, Yq ď kpκq.
Remark 5.2. Note that condition (2) in the above definition is equivalent to: For every
κ ą 0 and every point x P X satisfying dU pπU pxq, πU pYqq ď κ for all U P S, has the property
that dpx, Yq ď kpκq.
The following lemma and its corollary allow us to check the second condition of hierarchical
quasiconvexity for one value of κ only, which is convenient.
Lemma 5.3. For each Q there exists κ so that the following holds. Let Y Ď X be such that
πV pYq is Q–quasiconvex for each V P S. Let x P X and, for each V P S, let pV satisfy
dV px, pV q ď dV px, Yq ` 1. Then ppV q is κ–consistent.
Proof. For each V , choose yV P Y so that πV pyV q “ pV .
Suppose that V &W or V Ď W . By Lemma 2.6 and Theorem 3.1, there exists z P X so
that for all U P S, the projection πU pzq lies C–close to each of the geodesics rπU pxq, πU pyV qs,
rπU pxq, πU pyW qs, and rπU pyW q, πU pyV qs, where C depends on X . Hence dV ppV , zq and
dW ppW , zq are uniformly bounded, by quasiconvexity of πV pYq and πW pYq.
Suppose that V &W . Since the tuple pπU pzqq is consistent, either yV lies uniformly close in
CV to ρW
V , or the same holds with V and W interchanged, as required. Suppose that V Ď W .
Suppose that dW ppW , ρVW q is sufficiently large, so that we have to bound diamV pρW
V ppW q Y
pV q. Since dW pz, pW q is uniformly bounded, dW pz, ρVW q is sufficiently large that consistency
ensures that diamV pρW
V pπW pzqq Y πV pzqq is uniformly bounded. Since any geodesic from pW
W
to z lies far from ρVW , the sets ρW
V pπW pzqq and ρV ppV q coarsely coincide. Since πV pzq coarsely
coincides with pV by construction of z, we have the required bound. Hence the tuple with
V –coordinate pV is κ–consistent for uniform κ.
l
Definition 5.4 (Gate). A coarsely Lipschitz map gY : X Ñ Y is called a gate map if for each
x P X it satisfies: gY pxq is a point y P Y such that for all V P S, the set πV pyq (uniformly)
coarsely coincides with the projection of πV pxq to the kp0q–quasiconvex set πV pYq. The point
gpxq is called the gate of x in Y. The uniqueness axiom implies that when such a map exists
it is coarsely well-defined.
We first establish that, as should be the case for a (quasi)convexity property, one can
coarsely project to hierarchically quasiconvex subspaces. The next lemma shows that gates
exist for k–hierarchically quasiconvex subsets.
Lemma 5.5 (Existence of coarse gates). If Y Ď X is k–hierarchically quasiconvex and nonempty, then there exists a gate map for Y, i.e., for each x P X there exists y P Y such that
for all V P S, the set πV pyq (uniformly) coarsely coincides with the projection of πV pxq to
the kp0q–quasiconvex set πV pYq.
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
28
Proof. For each V P S, let pV satisfy dV px, pV q ď dV px, Yq ` 1. Then ppV q is κ–consistent
for some κ independent of x by Lemma 5.3.
Theorem 3.1 and the definition of hierarchical quasiconvexity combine to supply y 1 P
Nkpκq pYq with the desired projections to all V P S; this point lies at distance kpκq from some
y P Y with the desired property.
We now check that this map is coarsely Lipschitz. Let x0 , xn P X be joined by a uniform
quasigeodesic γ. By sampling γ, we obtain a discrete path γ 1 : t0, . . . , nu Ñ X such that
dX pγ 1 piq, γ 1 pi ` 1qq ď K for 0 ď i ď n ´ 1, where K depends only on X , and such that
řn´1
dX pgY pγ 1 piqq, gY pγ 1 pi`1qqq,
γ 1 p0q “ x0 , γ 1 pnq “ xn . Observe that dX pgY px0 q, gY pxn qq ď i“0
1
so it suffices to exhibit C such that dX pgY pxq, gY px qq ď C whenever dX px, x1 q ď K. But if
dX px, x1 q ď K, then each dU px, x1 q ď K 1 for some uniform K 1 , by Definition 1.1.(1), whence
the claim follows from the fact that each CU Ñ πU pYq is coarsely Lipschitz (with constant
depending only on δ and kp0q) along with the uniqueness axiom (Definition 1.1.(9)).
l
5.1. Hierarchically quasiconvex subspaces are hierarchically hyperbolic.
Proposition 5.6. Let Y Ď X be a hierarchically k-quasiconvex subset of the hierarchically
hyperbolic space pX , Sq. Then pY, dq is a hierarchically hyperbolic space, where d is the metric
inherited from X .
Proof. There exists K so that any two points in Y are joined by a uniform quasigeodesic.
Indeed, any two points in Y are joined by a hierarchy path in X , which must lie uniformly
close to Y.
We now define a hierarchically hyperbolic structure. For each U , let rU : CU Ñ πU pYq be
the coarse projection, which exists by quasiconvexity. The index set is S, and the associated
hyperbolic spaces are the various CU . For each U , define a projection πU1 : ÑCU by πU1 “
rU ˝ πU , and for each non-orthogonal pair U, V P S, the corresponding relative projection
CU Ñ CV is given by rV ˝ ρU
V . All of the requirements of Definition 1.1 involving only the
various relations on S are obviously satisfied, since we have only modified the projections.
The consistency inqualities continue to hold since each rU is uniformly coarsely Lipschitz.
The same is true for bounded geodesic image and the large link lemma. Partial realization
holds by applying the map gY to points constructed using partial realization in pX , Sq. l
Remark 5.7 (Alternative hierarchically hyperbolic structures). In the above proof, one can
replace each CU by a thickening CUY of πU pYq (this set is quasiconvex; the thickening is
to make a hyperbolic geodesic space). This yields a hierarchically hyperbolic structure with
coarsely surjective projections.
5.2. Standard product regions. In this section, we describe a class of hierarchically quasiconvex subspaces called standard product regions that will be useful in future applications.
We first recall a construction from [BHS14, Section 13].
Definition 5.8 (Nested partial tuple). ś
Recall SU “ tV P S : V Ď U u. Fix κ ě κ0 and let
FU be the set of κ–consistent tuples in V PSU 2CV .
Definition 5.9 (Orthogonal partial tuple). Let SK
U “ tV P S : V KU u Y tAu, where A is
a Ď–minimal element W
ś such that V Ď W for all V KU . Fix κ ě κ0 , let EU be the set of
κ–consistent tuples in V PSK ´tAu 2CV .
U
Construction 5.10 (Product regions in X ). Given X and U P S, there are coarsely welldefined maps φĎ , φK : FU , EU Ñ X whose images are hierarchically quasiconvex and which
extend to a coarsely well-defined map φU : FU ˆ EU Ñ X with hierarchically quasiconvex
image. Indeed, for each p~a, ~bq P FU ˆ EU , and each V P S, define the co-ordinate pφU p~a, ~bqqV
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
29
as follows. If V Ď U , then pφU p~a, ~bqqV “ aV . If V KU , then pφU p~a, ~bqqV “ bV . If V &U , then
pφU p~a, ~bqqV “ ρU
a, ~bqqV “ ρU
V . Finally, if U Ď V , and U ‰ V , let pφU p~
V.
~
We now verify that the tuple φU p~a, bq is consistent. If W, V P S, and either V or W
is transverse to U , then the consistency inequality involving W and V is satisfied in view
of Proposition 1.7. The same holds if U Ď W or U Ď V . Hence, it remains to consider
the cases where V and W are each either nested into or orthogonal to U : if V, W Ď U
or V, W KU then consistency holds by assumption; otherwise, up to reversing the roles of
V and W we have V Ď U and W KU , in which case V KW and there is nothing to check.
Theorem 3.1 thus supplies the map φU , and the maps φĎ and φK come from restricting to
the appropriate factor. Where it will not introduce confusion, we abuse notation and regard
FU , EU as subspaces of X , i.e., FU “ im φĎ , EU “ im φK .
Proposition 5.11. When EU , FU are endowed with the subspace metric d, the spaces pFU , SU q
and pEU , SK
U q are hierarchically hyperbolic; if U is not Ď–maximal, then their complexity is strictly less than that of pX , Sq. Moreover, φĎ and φK determine hieromorphisms
pFU , SU q, pEU , SK
U q Ñ pX , Sq.
Proof. For each V Ď U or V KU , the associated hyperbolic space CV is exactly the one used
in the hierarchically hyperbolic structure pX , Sq. For A, use an appropriate thickening C ˚ A
of πA pim φK q to a hyperbolic geodesic space. All of the projections FU Ñ CV, V P SU and
K
˚
EU Ñ CV, V P SK
U are as in pX , Sq (for A, compose with a quasi-isometry πA pim φ q Ñ C A).
K
Observe that pFU , SU q and pEU , SU q are hierarchically hyperbolic (this can be seen using a
simple version of the proof of Proposition 5.6). If U is not Ď–maximal in S, then neither is
A, whence the claim about complexity.
K
The hieromorphisms are defined by the inclusions SU , SK
U Ñ S and, for each V P SU YSU ,
˚
K
the identity CV Ñ CV , unless V “ A, in which case we use C A Ñ πA pim φ q ãÑ CA. These
give hieromorphisms by definition.
l
&
Remark 5.12 (Dependence on A). Note that A need not be the unique Ď–minimal element
of S into which each V KU is nested. Observe that EU (as a set and as a subspace of X ) is
defined independently of the choice of A. It is the hierarchically hyperbolic structure from
Proposition 5.11 that a priori depends on A. However, note that A Ď U , since there exists
V Ď A with V KU , and we cannot have V Ď U and V KU simultaneously. Likewise, A U by
definition. Finally, if U Ď A, then the axioms guarantee the existence of B, properly nested
into A, into which each V KU is nested, contradicting Ď–minimality of A. Hence U &A. It
follows that πA pEU q is bounded — it coarsely coincides with ρU
A . Thus the hierarchically
hyperbolic structure on EU , and the hieromorphism structure of φK , is actually essentially
canonical: we can take the hyperbolic space associated to the Ď–maximal element to be a
point, whose image in each of the possible choices of A must coarsely coincide with ρU
A.
Remark 5.13 (Orthogonality and product regions). If U KV , then we have FU Ď EV and
FV Ď EU , so there is a hierarchically quasiconvex map φU ˆ φV : FU ˆ FV Ñ X extending
K
to φU ˆ φK
U and φV ˆ φV .
Remark 5.14. Since FU , EU are hierarchically quasiconvex spaces, Definition 5.4 provides
coarse gates gFU : X Ñ FU and gEU : X Ñ EU . These are coarsely the same as the following
maps: given x P X , let ~x be the tuple defined by xW “ πW pxq when W Ď U and xW “ πW pxq
when W KU and ρU
x is consistent and coarsely equals gFU ˆEU pxq.
W otherwise. Then ~
Definition 5.15 (Standard product region). For each U P S, let PU “ im φU , which is
coarsely FU ˆ EU . We call this the standard product region in X associated to U .
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
30
&
The next proposition follows from the definition of the product regions and the fact that,
V
if U Ď V , then ρU
W , which holds
W , ρW coarsely coincide whenever V Ď W or V &W and U
by Definition 1.1.(4).
Proposition 5.16 (Parallel copies). There exists ν ě 0 such that for all U P S, all V P SU ,
and all u P EU , there exists v P EV so that φV pFV ˆ tvuq Ď Nν pφU pFU ˆ tuuqq.
5.2.1. Hierarchy paths and product regions. Recall that a D–hierarchy path γ in X is a
pD, Dq–quasigeodesic γ : I Ñ X such that πU ˝γ is an unparameterized pD, Dq–quasigeodesic
for each U P S, and that Theorem 4.4 provides D ě 1 so that any two points in X are joined
by a D–hierarchy path. In this section, we describe how hierarchy paths interact with standard product regions.
Given x, y P X , declare V P S to be relevant if dV px, yq ě 200DE.
Proposition 5.17 (“Active” subpaths). There exists ν ě 0 so that for all x, y P X , all V P S
with V relevant for px, yq, and all D–hierarchy paths γ joining x to y, there is a subpath α
of γ with the following properties:
(1) α Ă Nν pPV q;
(2) πU |γ is coarsely constant on γ ´ α for all U P SV Y SK
V.
Proof. We may assume γ : t0, nu Ñ X is a 2D–discrete path. Let xi “ γpiq for 0 ď i ď n.
First consider the case where V is Ď–maximal among relevant elements of S. Lemma 5.18
provides ν 2 ě 0, independent of x, y, and also provides i ď n, such that dS pxi , ρVS q ď ν 2 .
Let i be minimal with this property and let i1 be maximal with this property. Observe that
there exists ν 1 ě ν 2 , depending only on ν 2 and the (uniform) monotonicity of γ in CS, such
that dS pxj , ρVS q ď ν 1 for i ď j ď i1 .
For j P ti, . . . , i1 u, let x1j “ gPV pxj q. Let U P S. By definition, if U Ď V or U KV , then
πU pxj q coarsely coincides with πU px1j q, while πU px1j q coarsely coincides with ρVU if V Ď U or
V &U . We claim that there exist i1 , i11 with i ď i1 ď i11 ď i1 such that for i1 ď i ď i11 and
U P S with V Ď U or U &V , the points πU pxj q and πU px1j q coarsely coincide; this amounts
to claiming that πU pxj q coarsely coincides with ρVU .
If V Ď U and some geodesic σ in CU from πU pxq to πU pyq fails to pass through the Eneighborhood of ρVU , then bounded geodesic image shows that ρU
V pσq has diameter at most
E. On the other hand, consistency shows that the endpoints of ρU
V pσq coarsely coincide
with πV pxq and πV pyq, contradicting that V is relevant. Thus σ passes through the E–
neighborhood of ρVU . Maximality of V implies that U is not relevant, so that πV pxq, πV pyq,
and πV pxj q all coarsely coincide, whence πV pxj q coarsely coincides with ρVU .
If U &V and U is not relevant, then πU pxj q coarsely coincides with both πU pxq and πU pyq,
each of which coarsely coincides with ρVU , for otherwise we would have dV px, yq ď 2E by
consistency and the triangle inequality, contradicting that V is relevant. If U &V and U
is relevant, then, by consistency, we can assume that πU pyq,ρVU coarsely coincide, as do
V
πV pxq,ρU
V . Either πU pxj q coarsely equals ρU , or πV pxj q coarsely equals πV pxq, again by
consistency. If dV px, xj q ď 10E or dV py, xj q ď 10E, discard xj . Our discreteness assumption
and the fact that V is relevant imply that there exist i1 ď i11 between i and i1 so that xj is
not discarded for i1 ď j ď i11 . For such j, the distance formula now implies that dpxj , x1j q is
bounded by a constant ν independent of x, y.
We thus have i1 , i11 such that xj P Nν pPV q for i ď j ď i1 and xj R Nν pPV q for j ă i or j ą i1 ,
provided V is Ď–maximal relevant. If W Ď V and W is relevant, and there is no relevant
W 1 ‰ W with W Ď W 1 Ď V , then we may apply the above argument to γ 1 “ gPV pγ|i,...,i1 q to
produce a subpath of γ 1 lying ν–close to PW Ď PV , and hence a subpath of γ lying 2ν–close to
PW . Finiteness of the complexity (Definition 1.1.(5)) then yields assertion (1). Assertion (2)
is immediate from our choice of i1 , i11 .
l
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
31
Lemma 5.18. There exists ν 1 ě 0 so that for all x, y P X , all relevant V P S, and all
D–hierarchy paths γ joining x to y, there exists t P γ so that dS pt, ρVS q ď ν 1 .
Proof. Let σ be a geodesic in CS joining the endpoints of πS ˝ γ. Since dV px, yq ě 200DE,
the consistency and bounded geodesic image axioms (Definition 1.1.(4),(7)) imply that σ
enters the E–neighborhood of ρVS in CS, whence πS ˝ σ comes uniformly close to ρVS .
l
6. Hulls
In this section we build “convex hulls” in hierarchically hyperbolic spaces. This construction is motivated by, and generalizes, the concept in the mapping class group called Σ–hull,
as defined by Behrstock–Kleiner–Minsky–Mosher [BKMM12]. Recall that given a set A of
points in a δ–hyperbolic space H, its convex hull, denoted hullH pAq, is the union of geodesics
between pairs of points in this set. We will make use of the fact that the convex hull is
2δ–quasiconvex (since, if p P rx, ys, q P rx1 , y 1 s, then rp, qs Ď N2δ prp, xs Y rx, x1 s Y rx1 , qsq Ď
N2δ pry, xs Y rx, x1 s Y rx1 , y 1 sq).
The construction of hulls is based on Proposition 6.3, which generalizes Lemma 4.15;
indeed, the construction of hulls in this section generalizes the hulls of pairs of points used
in Section 4 to prove the distance formula. The second part of Proposition 6.3 (which is not
used in Section 4) relies on the distance formula.
Definition 6.1 (Hull of a set). For each set A Ă X and θ ě 0, let Hθ pAq be the set of all
p P X so that, for each W P S, the set πW ppq lies at distance at most θ from hullCW pAq.
Note that A Ă Hθ pAq.
Lemma 6.2. There exists θ0 so that for each θ ě θ0 there exists k : R` Ñ R` and each
A Ď X , we have that Hθ pAq is k–hierarchically quasiconvex.
Proof. For any θ and U P S, due to δ–hyperbolicity we have that πU pHθ pAqq is 2δ–quasiconvex,
so we only have to check the condition on realization points.
Let A1 be the union of all D0 –hierarchy paths joining points in A, where D0 is the constant
from Theorem 4.4. Then the Hausdorff distance between πU pA1 q and πU pAq is bounded by
C “ Cpδ, D0 q for each U P S. Also, πU pA1 q is Q “ Qpδ, D0 q–quasiconvex. Let κ be the
constant from Lemma 5.3, and let θ0 “ θe pκq be as in Theorem 3.1.
Fix any θ ě θ0 , and any κ ě 0. Let pbU q be a κ1 –consistent tuple with bU Ď Nθ phullCU pAqq
for each U P S. Let x P X project θe pκ1 q–close to each bU . We have to find y P Hθ pAq
uniformly close to x. By Lemma 5.3, ppU q is κ–consistent, where pU P hullCW pAq satisfies
dU px, pU q ď dU px, hullCW pAqq ` 1. It is readily seen from the uniqueness axiom (Definition
1.1.9) that any y P X projecting close to each pU has the required property, and such a y
exists by Theorem 3.1.
l
We denote the Hausdorff distance in the metric space Y by dHaus,Y p¨, ¨q. The next proposition directly generalizes [BKMM12, Proposition 5.2] from mapping class groups to general
hierarchically hyperbolic spaces.
Proposition 6.3 (Retraction onto hulls). For each sufficiently large θ there exists C ě
1 so that for each set A Ă X there is a pK, Kq–coarsely Lipschitz map r : X Ñ Hθ pAq
restricting to the identity on Hθ pAq. Moreover, if A1 Ă X also has cardinality at most N ,
then dX prA pxq, rA1 pxqq is C–coarsely Lipschitz in dHaus,X pA, A1 q.
Proof. By Lemma 6.2, for all sufficiently large θ, Hθ pAq is hierarchically quasiconvex. Thus,
by Lemma 5.5 there exists a map r : X Ñ Hθ pAq, which is coarsely Lipschitz and which is
the identity on Hθ pAq.
We now prove the moreover clause. By Definition 1.1.(1), for each W the projections
πW are each coarsely Lipschitz and thus dHaus,CW pπW pAq, πW pA1 qq is bounded by a coarsely
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
32
Lipschitz function of dHaus,X pA, A1 q. It is then easy to conclude using the distance formula
(Theorem 4.5) and the construction of gates (Definition 5.4) used to produce the map r. l
6.1. Homology of asymptotic cones. In this subsection we make a digression to study
homological properties of asymptotic cones of hierarchically hyperbolic spaces. This subsection is not needed for the proof of distance formula, and in fact we will use the distance
formula in a proof.
Using Proposition 6.3, the identical proof as used in [BKMM12, Lemma 5.4] for mapping
class groups, yields:
Proposition 6.4. There exists θ0 ě 0 depending only on the constants of the hierarchically
hyperbolic space pX , Sq such that for all θ, θ1 ě θ0 there exist K, C, and θ2 such that given
two sets A, A1 Ă X , then:
(1) diampHθ pAqq ď K diampAq ` C
(2) If A1 Ă Hθ pAq then Hθ pA1 q Ă Hθ2 pAq.
(3) dHaus,X pHθ pAq, Hθ pA1 qq ď KdHaus,X pA, A1 q ` C.
(4) dHaus,X pHθ pAq, Hθ1 pAqq ď C
Remark 6.5. Proposition 6.4 is slightly stronger than the corresponding [BKMM12, Lemma
5.4], in which A, A1 are finite sets and the constants depend on their cardinality. The source of
the strengthening is just the observation that hulls in δ–hyperbolic spaces are 2δ–quasiconvex
regardless of the cardinality of the set (cf. [BKMM12, Lemma 5.1]).
It is an easy observation that given a sequence A of sets An Ă X with bounded cardinality,
the retractions to the corresponding hulls Hθ pAn q converge in any asymptotic cone, Xω , to
a Lipschitz retraction from that asymptotic cone to the ultralimit of the hulls, HpAq. A
general argument, see e.g., [BKMM12, Lemma 6.2] implies that the ultralimit of the hulls
is then contractible. The proofs in [BKMM12, Section 6] then apply in the present context
using the above Proposition, with the only change needed that the reference to the rank
theorem for hierarchically hyperbolic spaces as proven in [BHS14, Theorem J] must replace
the application of [BM08]. In particular, this yields the following two results:
Corollary 6.6. Let X be a hierarchically hyperbolic space and Xω one of its asymptotic
cones. Let X Ă Xω be an open subset and suppose that for any sequence, A, of finite subsets
of X we have HpAq Ă X. Then X is acyclic.
Corollary 6.7. If pU, V q is an open pair in Xω , then Hk pU, V q “ t0u for all k greater than
the complexity of X .
6.2. Relatively hierarchically hyperbolic spaces and the distance formula. In this
section, we work in the following context:
Definition 6.8 (Relatively hierarchically hyperbolic spaces). The hierarchical space pX , Sq
is relatively hierarchically hyperbolic if there exists δ such that for all U P S, either U is
Ď–minimal or CU is δ–hyperbolic.
Our goal is to prove the following two theorems, which provide hierarchy paths and a
distance formula in relatively hierarchically hyperbolic spaces. We will not use these theorems
in the remainder of this paper, but they are required for future applications.
Theorem 6.9 (Distance formula for relatively hierarchically hyperbolic spaces). Let pX , Sq
be a relatively hierarchically hyperbolic space. Then there exists s0 such that for all s ě s0 ,
there exist constants C, K such that for all x, y P X ,
ÿ
ttdU px, yquus .
dX px, yq —K,C
U PS
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
33
Proof. By Proposition 6.14 below, for some suitably-chosen θ ě 0 and each x, y P X , there
exists a subspace Mθ px, yq of X (endowed with the induced metric) so that pMθ px, yq, Sq is a
hierarchically hyperbolic space (with the same nesting relations and projections from pX , Sq,
so that for all U P S, we have that πU pMθ px, yqq Ă Nθ pγU q, where γU is an arbitrarilychosen geodesic in CU from πU pxq to πU pyq. We emphasize that all of the constants from
Definition 1.1 (for Mθ px, yq) are independent of x, y. The theorem now follows by applying
the distance formula for hierarchically hyperbolic spaces (Theorem 4.5) to pMθ px, yq, Sq. l
Theorem 6.10 (Hierarchy paths in relatively hierarchically hyperbolic spaces). Let pX , Sq be
a relatively hierarchically hyperbolic space. Then there exists D ě 0 such that for all x, y P X ,
there exists a pD, Dq–quasigeodesic γ in X joining x, y so that πU pγq is an unparameterized
pD, Dq–quasigeodesic.
Proof. Proceed exactly as in Theorem 6.9, but apply Theorem 4.4 instead of Theorem 4.5. l
We now define hulls of pairs of points in the relatively hierarchically hyperbolic space
pX , Sq. Let θ be a constant to be chosen (it will be the output of the realization theorem for
a consistency constant depending on the constants associated to pX , Sq), and let x, y P X .
For each U P S, fix a geodesic γU in CU joining πU pxq to πU pyq. Define maps rU : CU Ñ γU
as follows: if CU is hyperbolic, let rU be the coarse closest-point projection map. Otherwise,
if CU is not hyperbolic (so U is Ď–minimal), define rU as follows: parametrize γU by arc
length with γU p0q “ x, and for each p P CU , let mppq “ mintdU px, pq, dU px, yqu. Then
rU ppq “ γU pmppqq. This rU is easily seen to be an L–coarsely Lipschitz retraction, with L
independent of U and x, y. (When U is minimal, rU is 1–Lipschitz.)
Next, define the hull Mθ px, yq to be the set of points x P X such that dU px, γU q ď θ for all
U P S. In the next proposition, we show that Mθ px, yq is a hierarchically hyperbolic space,
with the following hierarchically hyperbolic structure:
(1) the index set is S;
(2) the nesting, orthogonality, and transversality relations on S are the same as in pX , Sq;
(3) for each U P S, the associated hyperbolic space is γU ;
(4) for each U P S, the projection πU1 : Mθ px, yq Ñ γU is given by πU1 “ rU ˝ πU ;
(5) for each pair U, V P S of distinct non-orthogonal elements, the relative projection
CU Ñ CV is given by rV ˝ ρU
V.
Since there are now two sets of projections (those defined in the original hierarchical space
pX , Sq, denoted π˚ , and the new projections π˚1 ), in the following proofs we will explicitly
write all projections when writing distances in the various CU .
Lemma 6.11 (Gates in hulls). Let Mθ px, yq be as above. Then there exists a uniformly
coarsely Lipschitz retraction r : X Ñ Mθ px, yq such that for each U P S, we have, up to
uniformly (independent of x, y) bounded error, πU ˝ r “ rU ˝ πU .
Remark 6.12. It is crucial in the following proof that CU is δ–hyperbolic for each U P S
that is not Ď–minimal.
Proof of
śLemma 6.11. Let z P X and, for each U , let tU “ rU ˝ πU pzq; this defines a tuple
ptU q P U PS 2CU which we will check is κ–consistent for κ independent of x, y. Realization
(Theorem 3.1) then yields m P X such that dU pπU pmq, tU q ď θ for all U P S. By definition,
tU P γU , so m P Mθ px, yq and we define gx,y pzq “ m. Note that up to perturbing slightly, we
may take gx,y pzq “ z when z P Mθ . Hence it suffices to check consistency of ptU q.
First let U, V P S satisfy U &V . Then dV pπV pxq, πV pyqq ď 2E (up to exchanging U and
V ), and moreover each of πU pxq, πU pyq is E–close to ρU
V . Since tV lies on γV , it follows that
U
dV ptV , ρV q ď 2E.
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
34
Next, let U, V P S satisfy U Ĺ V . Observe that in this case, CV is δ–hyperbolic because V is not Ď–minimal. First suppose that dV pγV , ρU
V q ą 1. Then by consistency
and bounded geodesic image, dU px, yq ď 3E, and diamU pρVU pγV qq ď E. It follows that
diamU ptU Y ρU
V ptV qq ď 10E.
U
Hence, suppose that dV pρU
V , γV q ď 10E but that dV ptV , ρV q ą E. Without loss of gener1
ality, ρU
V lies at distance ď E from the subpath of γV joining tV to πV pyq. Let γV be the
subpath joining x to tV . By consistency, bounded geodesic image, and the fact that CV
is δ–hyperbolic and tV “ rV ˝ πV pzq, the geodesic triangle between πV pxq, πV pzq, and tV
projects under ρVU to a set, of diameter bounded by some uniform ξ, containing πU pxq, πU pzq,
and ρVU ptV q. Hence, since tU “ rU ˝ πU pzq, and πU pxq P γU , the triangle inequality yields
a uniform bound on diamU ptU Y ρVU ptV qq. Hence there exists a uniform κ, independent of
x, y, so that ptU q is κ–consistent. Finally, gx,y is coarsely Lipschitz by the uniqueness axiom
(Definition 1.1.(9)), since each rU is uniformly coarsely Lipschitz.
l
Lemma 6.13. Let m, m1 P Mθ px, yq. Then there exists C ě 0 such that m, m1 are joined by
a pC, Cq–quasigeodesic in Mθ px, yq.
Proof. Since X is a quasigeodesic space, there exists K ě 0 so that m, m1 are joined by a
K–discrete pK, Kq–quasigeodesic σ : r0, `s Ñ X with σp0q “ m, σp`q “ m1 . Note that gx,y ˝ σ
is a K 1 –discrete, efficient path for K 1 indepdent of x, y, since the gate map is uniformly
coarsely Lipschitz. A minimal-length K 1 –discrete efficient path in Mθ px, yq from x to y has
the property that each subpath is K 1 –efficient, and is a uniform quasigeodesic, as needed. l
Proposition 6.14. For all sufficiently large θ, the data (1)–(5) above makes pMθ px, yq, Sq
a hierarchically hyperbolic space, where Mθ px, yq inherits its metric as a subspace of X .
Moreover, the associated constants from Definition 1.1 are independent of x, y.
Proof. By Lemma 6.13, Mθ px, yq is a uniform quasigeodesic space. We now verify that the
enumerated axioms from Definition 1.1 are satisfied. Each part of the definition involving
only S and the Ď, K, & relations is obviously satisfied; this includes finite complexity. The
consistency inequalities hold because they hold in pX , Sq and each rU is L–coarsely Lipschitz.
The same holds for bounded geodesic image and the large link lemma. We now verify the
two remaining claims:
Uniqueness: Let m, m1 P Mθ px, yq, so that dU pπU pmq, γU q, dU pπU pm1 q, γU q ď θ for all U P
S. The definition of rU ensures that dU prU ˝ πU pmq, rU ˝ πU pm1 qq ě dU pπU pmq, πU pm1 qq ´ 2θ,
and uniqueness follows.
Partial realization: Let tUi u be a totally orthogonal subset of S and choose, for each i,
some pi P γUi . By partial realization in pX , Sq, there exists z P X so that dUi pπUi pzq, pi q ď E
1
i
for each i and dV pπV pzq, ρU
V q ď E provided Ui Ĺ V or Ui &V . Let z “ gx,y pzq P Mθ px, yq.
Then, by the definition of the gate map and the fact that each rU is L–coarsely Lipschitz, there
i
exists α, independent of x, y, so that dUi prUi ˝ πUi pz 1 q, pi q ď α, while dV prV ˝ πV pz 1 q, ρU
V qďα
1
whenever Ui &V or Ui Ď V . Hence z is the required partial realization point. This completes
the proof that pMθ px, yq, Sq is an HHS.
l
7. The coarse median property
In this section, we study the relationship between hierarchically hyperbolic spaces and
spaces that are coarse median in the sense defined by Bowditch in [Bow13]. In particular,
this discussion shows that OutpFn q is not a hierarchically hyperbolic space, and hence not a
hierarchically hyperbolic group, for n ě 3.
Definition 7.1 (Median graph). Let Γ be a graph with unit-length edges and path-metric d.
Then Γ is a median graph if there is a map m : Γ3 Ñ Γ such that, for all x, y, z P Γ, we have
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
35
dpx, yq “ dpx, mq ` dpm, yq, and likewise for the pairs x, z and y, z, where m “ mpx, y, zq.
Note that if x “ y, then mpx, y, zq “ x.
Chepoi established in [Che00] that Γ is a median graph precisely when Γ is the 1-skeleton
of a CAT(0) cube complex.
Definition 7.2 (Coarse median space). Let pM, dq be a metric space and let m : M 3 Ñ M
be a ternary operation satisfying the following:
(1) (Triples) There exist constants κ, hp0q such that for all a, a1 , b, b1 , c, c1 P M ,
`
˘
dpmpa, b, c, q, mpa1 , b1 , c1 qq ď κ dpa, a1 q ` dpb, b1 q ` dpc, c1 q ` hp0q.
(2) (Tuples) There is a function h : N Y t0u Ñ r0, 8q such that for any A Ď M with
p0q
1 ď |A| “ p ă 8, there is a CAT(0) cube complex Fp and maps π : A Ñ Fp and
p0q
λ : Fp Ñ M such that dpa, λpπpaqqq ď hppq for all a P A and such that
dpλpmp px, y, zqq, mpλpxq, λpyq, λpzqqq ď hppq
p0q
for all x, y, z P Fp , where mp is the map that sends triples from Fp to their median.
Then pM, d, mq is a coarse median space. The rank of pM, d, mq is at most d if each Fp above
can be chosen to satisfy dim Fp ď d, and the rank of pM, d, mq is exactly d if d is the minimal
integer such that pM, d, mq has rank at most d.
The next fact was observed by Bowditch [Bow15]; we include a proof for completeness.
Theorem 7.3 (Hierarchically hyperbolic implies coarse median). Let pX , Sq be a hierarchically hyperbolic space. Then X is coarse median of rank at most the complexity of pX , Sq.
Proof. Since the spaces CU, U P S are δ-hyperbolic for some δ independent of U , there exists
for each U a ternary operation mU : CU 3 Ñ CU so that pCU, dU , mU q is a coarse median space
of rank 1, and the constant κ and function h : N Y t0u Ñ r0, 8q from Definition 7.2 can be
chosen to depend only on δ (and not on U ).
Definition of the median: Define a map m : X 3 Ñ X as follows. Let x, y, z P X and, for
ś
each U P S, let bU “ mU pπU pxq, πU pyq, πU pzqq. By Lemma 2.6, the tuple ~b P U PS 2CU whose
U -coordinate is bU is κ–consistent for an appropriate choice of κ. Hence, by the realization
theorem (Theorem 3.1), there exists θe and m “ mpx, y, zq P X such that dU pm, bU q ď θu for
all U P S. Moreover, this is coarsely well-defined (up to the constant θe from the realization
theorem).
Application of [Bow13, Proposition 10.1]: Observe that, by Definition 1.1.1, the projections πU : X Ñ CU, U P S are uniformly coarsely Lipschitz. Moreover, for each U P S,
the projection πU : X Ñ CU is a “quasimorphism” in the sense of [Bow13, Section 10],
i.e., dU pmU pπU pxq, πU pyq, πU pzqq, πU pmpx, y, zqqq is uniformly bounded, by construction, as
U varies over S and x, y, z vary in X . Proposition 10.1 of [Bow13] then implies that m is a
coarse median on X , since that the hypothesis pP 1q of that proposition holds in our situation
by the distance formula.
l
The following is a consequence of Theorem 7.3 and work of Bowditch [Bow13, Bow14a]:
Corollary 7.4 (Contractibility of asymptotic cones). Let X be a hierarchically hyperbolic
space. Then all the asymptotic cones of X are contractible, and in fact bi-Lipschitz equivalent
to CAT(0) spaces.
Corollary 7.5 (HHGs have quadratic Dehn function). Let G be a finitely generated group
that is a hierarchically hyperbolic space. Then G is finitely presented and has quadratic Dehn
function. In particular, this conclusion holds when G is a hierarchically hyperbolic group.
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
Proof. This follows from Theorem 7.3 and [Bow13, Corollary 8.3].
36
l
Corollary 7.6. For n ě 3, the group OutpFn q is not a hierarchically hyperbolic space, and
in particular is not a hierarchically hyperbolic group.
Proof. This is an immediate consequence of Corollary 7.5 and the exponential lower bound
on the Dehn function of OutpFn q given by the combined results of [BV95, HM13b, BV]. l
We also recover a special case of Theorem I of [BHS14], using Corollary 7.5 and a theorem
of Gersten-Holt-Riley [GHR03, Theorem A]:
Corollary 7.7. Let N be a finitely generated virtually nilpotent group. Then G is quasiisometric to a hierarchically hyperbolic space if and only if N is virtually abelian.
Corollary 7.8. Let S be a symmetric space of non-compact type, or a thick affine building.
Suppose that the spherical type of S is not Ar1 . Then S is not hierarchically hyperbolic.
Proof. This follows from Theorem 7.3 and Theorem A of [Hae14].
l
Finally, Theorem 9.1 of [Bow14a] combines with Theorem 7.3 to yield:
Corollary 7.9 (Rapid decay). Let G be a group whose Cayley graph is a hierarchically
hyperbolic space. Then G has the rapid decay property.
7.1. Coarse media and hierarchical quasiconvexity. The natural notion of quasiconvexity in the coarse median setting is related to hierarchical quasiconvexity.
Definition 7.10 (Coarsely convex). Let pX , Sq be a hierarchically hyperbolic space and let
m : X 3 Ñ X be the coarse median map constructed in the proof of Theorem 7.3. A closed
subspace Y Ď X is µ–convex if for all y, y 1 P Y and x P X , we have mpy, y 1 , xq P Nµ pYq.
Remark 7.11. We will not use µ–convexity in the remainder of the paper. However, it is
of independent interest since it parallels a characterization of convexity in median spaces: a
subspace Y of a median space is convex exactly when, for all y, y 1 P Y and x in the ambient
median space, the median of x, y, y 1 lies in Y.
Proposition 7.12 (Coarse convexity and hierarchical quasiconvexity). Let pX , Sq be a hierarchically hyperbolic space and let Y Ď X . If Y is hierarchically k-quasiconvex, then there
exists µ ě 0, depending only on k and the constants from Definition 1.1, such that Y is
µ–convex.
Proof. Let Y Ď X be k–hierarchically quasiconvex, let y, y 1 P Y and x P X . Let m “
mpx, y, y 1 q. For any U P S, the projection πU pYq is by definition kp0q-quasiconvex, so that,
for some k 1 “ k 1 pkp0q, δq, we have dU pmU , πU pYqq ď k 1 , where mU is the coarse median of
πU pxq, πU pyq, πU py 1 q coming from hyperbolicity of CU . The tuple pmU qU PS was shown above
to be κ–consistent for appropriately-chosen κ (Lemma 2.6), and dU pmU , mpx, y, y 1 qq ď θe pκq,
so, by hierarchical quasiconvexity dX pmpx, y, y 1 q, Yq is bounded by a constant depending on
kpκq and k 1 .
l
8. Combination theorems for hierarchically hyperbolic spaces
The goal of this section is to prove Theorem 8.6, which enables the construction of new
hierarchically hyperbolic spaces and groups from a tree of given ones. We postpone the
statement of the theorem until after the relevant definitions.
Definition 8.1 (Quasiconvex hieromorphism, full hieromorphism). Let pf, f ♦ , tf ˚ pU quU PS q
be a hieromorphism pX , Sq Ñ pX 1 , S1 q. We say f is k–hierarchically quasiconvex if its image
is k–hierarchically quasiconvex. The hieromorphism is full if:
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
37
(1) there exists ξ ě 0 such that each f ˚ pU q : CU Ñ Cpf ♦ pU qq is a pξ, ξq–quasi-isometry,
and
(2) for each U P S, if V 1 P S1 satisfies V 1 Ď f ♦ pU q, then there exists V P S such that
V Ď U and f ♦ pV q “ V 1 .
Remark 8.2. Observe that Definition 8.1.(2) holds automatically unless V 1 is bounded.
Definition 8.3 (Tree of hierarchically hyperbolic spaces). Let V, E denote the vertex and
edge-sets, respectively, of the simplicial tree T . A tree of hierarchically hyperbolic spaces is a
quadruple
`
˘
T “ T, tXv u, tXe u, tφe˘ : v P V, e P Eu
satisfying:
(1) tXv u and tXe u are uniformly hierarchically hyperbolic: each Xv has index set Sv , and
each Xe has index set Se . In particular, there is a uniform bound on the complexities
of the hierarchically hyperbolic structures on the Xv and Xe .
(2) Fix an orientation on each e P E and let e` , e´ denote the initial and terminal
vertices of e. Then, each φe˘ : Xe Ñ Xe˘ is a hieromorphism with all constants
bounded by some uniform ξ ě 0. (We adopt the hieromorphism notation from Definition 1.19. Hence we actually have maps φe˘ : Xe Ñ Xe˘ , and maps φ♦
e˘ : Se Ñ Se˘
preserving nesting, transversality, and orthogonality, and coarse ξ–Lipschitz maps
φ˚e˘ pU q : CU Ñ Cpφ♦
e˘ pU qq satisfying the conditions of that definition.)
Given a tree TŮ of hierarchically hyperbolic spaces, denote by X pT q the metric space constructed from vPV Xv by adding edges of length 1 as follows: if x P Xe , we declare φe´ pxq to
be joined by an edge to φe` pxq. Given x, x1 P X in the same vertex space Xv , define d1 px, x1 q
to be dXv px, x1 q. Given x, x1 P X joined by an edge, define d1 px, x1 q “ 1. Given a sequence
x0 , x1 , . . . , xk P X , with consecutive points either joined by an edge or in a common vertex
ř
1
1
1
space, define its length to be k´1
i“1 d pxi , xi`1 q. Given x, x P X , let dpx, x q be the infimum
of the lengths of such sequences x “ x0 , . . . , xk “ x1 .
Remark 8.4. Since the vertex spaces are (uniform) quasigeodesic spaces, pX , dq is a quasigeodesic space.
Definition 8.5 (Equivalence, support, bounded support). Let T be a tree of hierarchically
hyperbolic spaces. For each e P E, and each We´ P Se´ , We` P Se` , write We´ „d We`
♦
The transitive closure „ of „d is an
if there exists We P Se so that φ♦
e´ pWe q “ φe` pWe q.
Ť
Ť
equivalence relation on v Sv . The „–class of W P v Sv is denoted rW s.
The support of an equivalence class rW s is the induced subgraph TrW s of T whose vertices
are those v P T so that Sv contains a representative of rW s. Observe that TrW s is connected.
The tree T of hierarchically hyperbolic spaces has bounded supports if there exists n P N such
that each „–class has support of diameter at most n.
We can now state the main theorem of this section:
Theorem 8.6 (Combination theorem for hierarchically hyperbolic spaces). Let T be a tree
of hierarchically hyperbolic spaces. Suppose that:
(1) there exists a function k so that each edge–hieromorphism is k–hierarchically quasiconvex;
(2) each edge–hieromorphism is full;
(3) T has bounded supports of diameter at most n;
(4) if e is an edge of T and Se is the Ď–maximal element of Se , then for all V P Se˘ ,
the elements V and φ♦
pS q are not orthogonal in Se˘ .
e˘ e
Then XpT q is hierarchically hyperbolic.
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
38
We postpone the proof until after the necessary lemmas and definitions. For the remainder of this section, fix a tree of hierarchically hyperbolic spaces T “ pT, tXv u, tXe u, tφe˘ uq
satisfying the hypotheses of Theorem 8.6; let n be the constant
Ť implicit in assumption 3.
Denote by S the union of tT u with the set of all „–classes in v Sv .
Definition 8.7 (Nesting, orthogonality, transversality in S). For all rW s P S, declare
rW s Ď T . If rV s, rW s are „–classes, then rV s Ď rW s if and only if there exists v P T such
that rV s, rW s are respectively represented by Vv , Wv P Sv and Vv Ď Wv ; this relation is
nesting. For convenience, for A P S, we write SA to denote the set of B P S such that
B Ď A.
Likewise, rV sKrW s if and only if there exists a vertex v P T such that rV s, rW s are respectively represented by Vv , Wv P Sv and Vv KWv ; this relation is orthogonality. If rV s, rW s P S
are not orthogonal and neither is nested into the other, then they are transverse, written
rV s&rW s. Equivalently, rV s&rW s if for all v P TrV s X TrW s , the representatives Vv , Wv P Sv
of rV s, rW s satisfy Vv &Wv .
Fullness (Definition 8.1.(2)) was introduced to enable the following two lemmas:
Lemma 8.8. Let T be a tree of hierarchically hyperbolic spaces, let v be a vertex of the
underlying tree T , and let U, U 1 P Sv satisfy U Ď U 1 . Then either U “ U 1 or U U 1 .
Proof. Suppose that U „ U 1 , so that there is a closed path v “ v0 , v1 , . . . , vn “ v in T and a
sequence U “ U0 , U1 , . . . , Un “ U 1 such that Ui P Svi and Ui „d Ui`1 for all i. If U ‰ U 1 ,
then Condition (2) (fullness) from Definition 8.1 and the fact that hieromorphisms preserve
nesting yields U 2 P Sv , different from U 1 , such that U 2 „ U and U 2 Ĺ U 1 Ĺ U (where Ĺ
denotes proper nesting). Repeating this argument contradicts finiteness of complexity.
l
Lemma 8.9. The relation Ď is a partial order on S, and T is the unique Ď–maximal element.
Moreover, if rV sKrW s and rU s Ď rV s, then rU sKrW s and rV s, rW s are not Ď–comparable.
Proof. Reflexivity is clear. Suppose that rVv s Ď rUu s Ď rWw s. Then there are vertices
v1 , v2 P V and representatives Vv1 P rVv s, Uv1 P rUu s, Uv2 P rUu s, Wv2 P rWw s so that Vv1 Ď Uv1
and Uv2 Ď Wv2 . Since edge–hieromorphisms are full, induction on dT pv1 , v2 q yields Vv2 Ď Uv2
so that Vv2 „ Vv1 . Transitivity of the nesting relation in Sv2 implies that Vv2 Ď Wv2 , whence
rVv s Ď rWw s. Suppose that rUu s Ď rVv s and rVv s Ď rUu s, and suppose by contradiction
that rUu s ‰ rVv s. Choose v1 , v2 P V and representatives Uv1 , Uv2 , Vv1 , Vv2 so that Uv1 Ď Vv1
and Vv2 Ď Uv2 . The definition of „ again yields Uv2 „ Uv1 with Uv2 Ď Vv2 ‰ Uv2 . This
contradicts Lemma 8.8 below. Hence Ď is antisymmetric, whence it is a partial order. The
underlying tree T is the unique Ď–maximal element by definition.
Suppose that rV sKrW s and rU s Ď rV s. Then there are vertices v1 , v2 and representatives Vv1 , Wv1 , Uv2 , Vv2 such that Vv1 KWv1 and Uv2 KVv2 . Again by fullness of the edge–
hieromorphisms, there exists Uv1 „ Uv2 with Uv1 Ď Vv1 , whence Uv1 KWv1 . Thus rU sKrW s
as required. Finally, Ď–incomparability of rV s, rW s follows from fullness and the fact that
edge-hieromorphisms preserve orthogonality and nesting.
l
Lemma 8.10. Let rW s P S and let rU s Ď rW s. Suppose moreover that
(
rV s P SrW s : rV sKrU s ‰ H.
Then there exists rAs P SrW s ´ trW su such that rV s Ď rAs for all rV s P SrW s with rV sKrU s.
Proof. Choose some v P V so that there exists Vv P Sv and Uv P Sv with rUv s “ rU s and
Vv KUv . Then by definition, there exists Av P Sv so that Bv Ď Av whenever Bv KUv and so
that rBv s Ď rW s. It follows from the fact that the edge-hieromorphisms are full and preserve
(non)orthogonality that rBs Ď rAv s whenever rBsKrU s.
l
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
39
Lemma 8.11. There exists χ ě 0 such that if tV1 , . . . , Vc u Ă S consists of pairwise orthogonal or pairwise Ď–comparable elements, then c ď χ.
Proof. For each v P T , let χv be the complexity of pXv , Sv q and let χ “ maxv χv ` 1. Let
rV1 s, . . . , rVc s P S´ tT u be „–classes that are pairwise orthogonal or pairwise Ď–comparable.
The Helly property for trees yields a vertex v lying in the support of each rVi s; let Viv P Sv
represent rVi s. Since edge–hieromorphisms preserve nesting, orthogonality, and transversality, c ď χv . However, each Vi Ď T , whence χ is the correct upper bound.
l
Definition 8.12 (Favorite representative, hyperbolic spaces associated to equivalence classes).
Let CT “ T . For each „–class rW s, choose a favorite vertex v of TrW s and let Wv P SWv be
the favorite representative of rW s. Let CrW s “ CWv . Note that each CrW s is δ–hyperbolic,
where δ is the uniform hyperbolicity constant for T .
Definition 8.13 (Gates in vertex spaces). For each vertex v of T , define a gate map gv : X Ñ
Xv as follows. Let x P Xu for some vertex u of T . We define gv pxq inductively on dT pu, vq.
If u “ v, then set gv pxq “ x. Otherwise, u “ e´ for some edge e of T so that dT pe` , vq “
dT pu, vq´1. Then set gv pxq “ gv pφe` pφ´1
pg
pxqqqq. We also have a map βVv : X Ñ CVv ,
e´ φe´ pXe q
defined by βVv pxq “ πVv pgv pxqq. (Here, gφe´ pXe q : Xe´ “ Xu Ñ φe´ pXe q is the usual gate map
to a hierarchically quasiconvex subspace, described in Definition 5.4, and φ´1
is a quasie˘
inverse for the edge-hieromorphism.)
Lemma 8.14. There exists K, depending only on E and ξ, such that the following holds.
Let e, f be edges of T and v a vertex so that e´ “ f ´ “ v. Suppose for some V P Sv
that there exist x, y P φe´ pXe q Ď Xv with dV pgφf ´ pXf q pxq, gφf ´ pXf q pyqq ą 10K. Then V P
φ♦
pSe q X φ♦
pSf q.
e´
f´
Proof. Let Ye “ φe´ pXe q and let Yf “ φf ´ pXf q; these spaces are uniformly hierarchically
quasiconvex in Xv . Moreover, by fullness of the edge-hieromorphisms, we can choose K ě
100E so that the map πV : Ye Ñ CV is K–coarsely surjective for each V P φ♦
pSe q, and
e´
♦
♦
likewise for φf ´ pSf q and Yv . If V P Sv ´ φf ´ pSf q, then πV is K–coarsely constant on Yf ,
by the distance formula, since Xf is quasi-isometrically embedded. Likewise, πV is coarsely
constant on Ye if V R φ♦
pSe q. (This also follows from consistency when V is transverse
e´
to some unbounded element of φ♦
pSe q and from consistency and bounded geodesic image
e´
otherwise.)
Suppose that there exists V P Sv such that dV pgφf ´ pXf q pxq, gφf ´ pXf q pyqq ą 10K. Since
gφf ´ pXf q pxq, gφf ´ pXf q pyq P Xf , we therefore have that V P φ♦
pSf q. On the other hand, the
f´
definition of gates implies that dV px, yq ą 8K, so V P φ♦
pSe q.
e´
l
Lemma 8.15. There exists a constant K 1 such that the following holds. Let e, f be edges of
T and suppose that there do not exist Ve P Se , Vf P Sf for which φ♦
pV q „ φ♦
pV q. Then
e´ e
f´ f
1
ge´ pXf q has diameter at most K . In particular, the conclusion holds if dT pe, f q ą n, where
n bounds the diameter of the supports.
Proof. The second assertion follows immediately from the first in light of how n was chosen.
We now prove the first assertion by induction on the number k of vertices on the geodesic
in T from e to f . The base case, k “ 1, follows from Lemma 8.14.
For k ě 1, let v0 , v1 , . . . , vk be the vertices on a geodesic from e to f , in the obvious order.
Let b be the edge joining vk´1 to vk , with b´ “ vk .
It follows from the definition of gates that ge´ pXf q has diameter (coarsely) bounded above
by that of gφb´ pXb q pXf q and that of ge´ pXb q. Hence suppose that diampgφb´ pXb q pXf qq ą 10K
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
40
Xe
Xb
Xb´
Xf
Figure 4. Schematic of the subset of X near Xb´ .
and diampge´ pXb qq ą 10K. Then, by induction and Lemma 8.14, we see that there exists
Ve P Se , Vf P Sf for which φ♦
pV q „ φ♦
pV q, a contradiction.
l
e´ e
f´ f
Lemma 8.16. The map gv : X Ñ Xv is coarse Lipschitz, with constants independent of v.
Proof. Let x, y P X . If the projections of x, y to T lie in the ball of radius 2n ` 1 about v,
then this follows since gv is the composition of a bounded number of maps, each of which is
uniformly coarsely Lipschitz by Lemma 5.5. Otherwise, by Remark 8.4, it suffices to consider
x, y with dX px, yq ď C, where C depends only on the metric d. In this case, let vx , vy be the
vertices in T to which x, y project. Let v 1 be the median in T of v, vx , vy . Observe that there is
a uniform bound on dT pvx , v 1 q and dT pvy , v 1 q, so it suffices to bound dv pgv pgv1 pxqq, gv pgv1 pyqqq.
Either dT pv, v 1 q ď 2n ` 1, and we are done, or Lemma 8.15 gives the desired bound, since
equivalence classes have support of diameter at most n.
l
Definition 8.17 (Projections). For each rW s P S, define the projection πrW s : X Ñ CrW s
by πrW s pxq “ βWv pxq, where Wv is the favorite representative of rW s. Note that these
projections take points to uniformly bounded sets, since the collection of vertex spaces is
uniformly hierarchically hyperbolic. Define πT : X Ñ T to be the usual projection to T .
Lemma 8.18 (Comparison maps). There exists a uniform constant ξ ě 1 such that for all
Wv P Sv , Ww P Sw with Wv „ Ww , there exists a pξ, ξq–quasi-isometry c : Wv Ñ Ww such
that c ˝ βv “ βw up to uniformly bounded error.
Definition 8.19. A map c as given by Lemma 8.18 is called a comparison map.
Proof of Lemma 8.18. We first clarify the situation by stating some consequences of the
definitions. Let e` , e´ be vertices of T joined by an edge e. Suppose that there exists W ` P
Se` , W ´ P Se´ such that W ` „ W ´ , so that there exists W P Se with πpφe˘ qpW q “ W ˘ .
Then the following diagram coarsely commutes (with uniform constants):
X
ge´
ge`
Xe´
πW ´
X`
? e
_
CW ´_
Xe
πW ´
`
CW
?
πW
CW
where Xe Ñ Xe˘ is the uniform quasi-isometry φe˘ , while Xe˘ Ñ Xe is the composition of
a quasi-inverse for φe˘ with the gate map Xe˘ Ñ φe˘ pXe q, and the maps CW Ø CW ˘ are
the quasi-isometries implicit in the edge hieromorphism or their quasi-inverses. The proof
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
41
essentially amounts to chaining together a sequence of these diagrams as e varies among the
edges of a geodesic from v to w; an important ingredient is played by the fact that such a
geodesic has length at most n.
Let v “ v0 , v1 , . . . , vm , vm`1 “ w be the geodesic sequence in T from v to w and let ei be the
edge joining vi to vi`1 . For each i, choose Wi P Sei and Wi˘ P Sei so that (say) W0´ “ Wv
˘
` “ W and so that φ♦ pW q “ W ˘ for all i. For each i, let q ˘ : CW Ñ CW ˘ be
and Wm
w
i
i
i
i
i
ei
˘
qi “ φ˚ei pWi q be the pξ 1 , ξ 1 q–quasi-isometry packaged in the edge–hieromorphism, and let q̄i˘
˘
be a quasi-inverse; the constant ξ 1 is uniform by hypothesis, and m ď n since T has bounded
supports. The hypotheses on the edge–hieromorphisms ensure that the Wi˘ are uniquely
determined by Wv , Ww , and we define c by
1
1
m m
c “ qm
q̄m ¨ ¨ ¨ q11 q̄11 ,
where i , 1i P t˘u depend on the orientation of ei , and 1i “ ` if and only if i “ ´. This is
a pξ, ξq–quasi-isometry, where ξ “ ξpξn1 q.
If v “ w, then c is the identity and c ˝ βv “ βv . Let d ě 1 “ dT pv, wq and let w1 be the
penultimate vertex on the geodesic of T from v to w. Let c1 : CWv Ñ CWw1 be a comparison
map, so that, by induction, there exists λ1 ě 0 so that dCWw1 pc1 ˝ βv pxq, βw1 pxqq ď λ1 for
all x P X . Let c2 “ q̄k` qk´ : CWw1 Ñ CWw be the pξ 1 , ξ 1 q–quasi-isometry packaged in the
edge–hieromorphism, so that the following diagram coarsely commutes:
X
gv
Xv
πWv
z
CWv
gw
gw1
$
/ Xw1
/
Xw
gw
πWw
gw 1
πW
c1
w1
/ CWw1
/ CWw
2
c
c2 ˝c1
Since c “
and the constants implicit in the coarse commutativity of the diagram depend
only on the constants of the hieromorphism and on d ď n, the claim follows.
l
Lemma 8.20. There exists K such that each πrW s is pK, Kq–coarsely Lipschitz.
Proof. For each vertex v of T and each V P Sv , the projection πV : Xv Ñ CV is uniformly
coarsely Lipschitz, by definition. By Lemma 8.16, each gate map gv : X Ñ Xv is uniformly
coarsely Lipschitz. The lemma follows since πrW s “ πWv ˝ gv , where v is the favorite vertex
carrying the favorite representative Wv of rW s.
l
Definition 8.21 (Projections between hyperbolic spaces). If rV s Ď rW s, then choose vertices
v, v 1 , w P V so that Vv , Ww are respectively the favorite representatives of rV s, rW s, while
Vv1 , Wv1 are respectively representatives of rV s, rW s with Vv1 , Wv1 P Sv1 and Vv1 Ď Wv1 . Let
cV : CVv1 Ñ CVv and cW : CWv1 Ñ CWw be comparison maps. Then define
´
¯
V 1
rV s
ρrW s “ cW ρWv 1 ,
v
rW s
which is a uniformly bounded set, and define ρrV s : CrW s Ñ CrV s by
W
rW s
1
ρrV s “ cV ˝ ρV v1 ˝ c̄W ,
v
W
1
where c̄W is a quasi–inverse of cW and ρV v1 : CWv1 Ñ CVv1 is the map provided by Definiv
tion 1.1.(2). Similarly, if rV s&rW s, and there exists w P T so that Sw contains representatives
Vw , Ww of rV s, rW s, then let
´
¯
rV s
ρrW s “ cW ρVWww .
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
42
Otherwise, choose a closest pair v, w so that Sv (respectively Sw ) contains a representative
of rV s (respectively rW s). Let e be the first edge of the geodesic in T joining v to w, so
v “ e´ (say). Let S be the Ď–maximal element of Se , and let
¯
´
πpφ ´ qpSq
rW s
.
ρrV s “ cV ρVv e
This is well-defined by hypothesis (4).
rW s
For each „–class rW s, let ρT be the support of rW s (a uniformly bounded set since T
has bounded supports). Define ρTrW s : T Ñ CW as follows: given v P T not in the support
of rW s, let e be the unique edge with e´ (say) separating v from the support of rW s. Let
rφ ´ pSqs
. If v is in the support of rW s, then let
S P Se be Ď–maximal. Then ρTrW s pvq “ ρrWe s
ρTrW s pvq be chosen arbitrarily.
We are now ready to complete the proof of the combination theorem.
Proof of Theorem 8.6. We claim that pX pT q, Sq is hierarchically hyperbolic. We take the
nesting, orthogonality, and transversality relations for a tree of spaces given by Definition 8.7. In Lemmas 8.9 and 8.10, it is shown that these relations satisfy all of the conditions (2) and (3) of Definition 1.1 not involving the projections. Moreover, the complexity
of pX pT q, Sq is finite, by Lemma 8.11, verifying Definition 1.1.(5). The set of δ–hyperbolic
spaces tCA : A P Su is provided by Definition 8.12, while the projections πrW s : X Ñ CrW s
required by Definition 1.1.(1) are defined in Definition 8.17 and are uniformly coarsely LipsrV s
chitz by Lemma 8.20. The projections ρrW s when rV s, rW s are non-orthogonal are described
in Definition 8.21. To complete the proof, it thus suffices to verify the consistency inequalities (Definition 1.1.(4)), the bounded geodesic image axiom (Definition 1.1.(7)), the large
link axiom (Definition 1.1.(6)), partial realization (Definition 1.1.(8)), and uniqueness (Definition 1.1.(9)).
Consistency: Suppose that rU s&rV s or rU s Ď rV s and let x P X . Choose representatives
Uu P Su , Vu P Sv of rU s, rV s so that dT pu, vq realizes the distance between the supports of
rU s, rV s. By composing the relevant maps in the remainder of the argument with comparison
maps, we can assume that Uu , Vv are favorite representatives. Without loss of generality, there
exists a vertex w P T so that x P Xw . If u “ v, then consistency follows since it holds in each
vertex space, so assume that u, v have disjoint supports and in particular rU s&rV s.
If w R ru, vs, then (say) u separates w from v. Then πrV s pxq “ πVv pgv pxqq. Let e be the
rV s
edge of the geodesic ru, vs emanating from u, so that ρrU s “ ρSUu , where S is the image in
&
Su of the Ď–maximal element of Se . If dCUu pgu pxq, ρSUu q ď E, then we are done. Otherwise,
u
by consistency in Su , we have dCS pgu pxq, ρU
S q ď E, from which consistency follows. Hence
suppose that w P ru, vs. Then without loss of generality, there is an edge e containing v and
separating w from v. As before, projections to V factor through the Ď-maximal element of
Se , from which consistency follows.
It remains to verify consistency for T, rW s for each „–class rW s. Choose x P Xv . If
dT pv, TrW s q ě n ` 1, then let e be the edge incident to TrW s separating it from v, so that (up
to a comparison map) ρTrW s pvq “ ρSW , where W P Se` represents W , and e` P TrW s , and S
is the image in Se` of the Ď–maximal element of Se . (Note that W &S by hypothesis (4) of
the theorem and the choice of e). On the other hand (up to a comparison map) πrW s pxq “
πW pge` pxqq — πW pge` pXv qq — πW pFS q “ ρSW , as desired.
Finally, suppose that rU s Ď rV s and that either rV s Ĺ rW s or rV s&rW s and rU s rW s. We
rU s
rV s
claim that drW s pρrW s , ρrW s q is uniformly bounded. Choose representatives Uu P Su , Vu P Su
of rU s, rV s with Uu Ď Vu . Choose representatives Vw , Ww P Sw so that Vw Ď Ww or Vw &Uw
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
43
&
according to whether rV s Ď rW s or rV s&rW s, and choose representatives Uv , Wv P Sv of
rU s, rW s so that Uv Ď Wv or Uv &Wv according to whether rU s Ď rW s or rU s&rW s. Let
m P T be the median of u, v, w. Since u, w lie in the support of rU s, rW s, so does m, since
supports are connected. Likewise, m lies in the support of rV s. Let Um , Vm , Wm be the
representatives of rU s, rV s, rW s in m. Since edge-maps are full hieromorphisms, we have
Um Ď Vm and Um Wm and either Vm Ď Wm or Vm &Wm . Hence Definition 1.1.(4) implies
Vm
m
that dWm pρU
Wm , ρWm q is uniformly bounded. Since the comparison maps are uniform quasirU s
rV s
isometries, it follows that drW s pρrW s , ρrW s q is uniformly bounded, as desired.
Bounded geodesic image and large link axiom in T : Let γ be a geodesic in T and
rW s
let rW s be a „–class so that dT pγ, ρT q ą 1, which is to say that γ does not contain vertices
in the support of rW s. Let e be the terminal edge of the geodesic joining γ to the support of
rSs
rW s. Then for all u P γ, we have by definition that ρTrW s puq “ ρrW s for some fixed „–class
rSs. This verifies the bounded geodesic image axiom for T, rW s.
By Lemma 8.15, there exists a constant K 2 such that if x, x1 P X respectively project
to vertices v, v 1 , then any rW s P S ´ tT u with drW s px, x1 q ě K 2 is supported on a vertex
vrW s on the geodesic rv, v 1 s and is hence nested into rSvW s, where SvW is maximal in SvrW s .
Indeed, choose w in the support of rW s. Then either drW s px, x1 q is smaller than some specified
constant, or dXw pgw pxq, gw px1 q ą K 1 . Thus gXw pXm q has diameter at least K 1 , where m is the
median of v, v 1 , w. Hence m lies in the support of rW s, and m P rv, v 1 s, and rW s Ď rSs, where
rSv
s
S is Ď–maximal in Sm . Finally, for each such SvW , it is clear that dT px, ρT W q ď dT px, x1 q,
verifying the conclusion of the large link axiom for T .
Bounded geodesic image and large link axiom in W Ă T : Let rW s be non-Ď–
maximal, let rV s Ď rW s, and let γ be a geodesic in CrW s. Then γ is a geodesic in CWw , by
definition, where w is the favorite vertex of rW s with corresponding representative Ww . Let
rV s
Vw be the representative of rV s supported on w, so that ρrW s “ ρVWww , so that γ avoids the
rV s
E–neighborhood of ρrW s exactly when it avoids the E–neighborhood of ρVWww . The bounded
geodesic image axiom now follows from bounded geodesic image in Sw , although the constant
E has been changed according to the quasi-isometry constant of comparison maps.
Now suppose x, x1 P Xv , Xv1 and choose w to be the favorite vertex in the support of
rW s. Suppose for some rV s Ď rW s that drV s px, x1 q ě E 1 , where E 1 depends on E and the
quasi-isometry constants of the edge-hieromorphisms. Then dVw pgw pxq, gw px1 qq ě E, for
some representative Vw P Sw of rV s, by our choice of E 1 . Hence, by the large link axiom
in Sw , we have that Vw Ď Ti , where tTi u is a specified set of N “ tdWw pgw pxq, gw px1 qqu “
tdrW s px, x1 qu elements of Sw , with each Ti Ă Ww . Moreover, the large link axiom in Sw
rT s
implies that drW s px, ρrWi s q “ dWw pgw pxq, ρTWi w q ď N for all i. This verifies the large link
axiom for pX pT q, Sq.
Partial realization: Let rV1 s, . . . , rVk s P S be pairwise–orthogonal, and, for each i ď k,
let pi P CrVi s. For each i, let Ti Ď T be the induced subgraph spanned by the vertices w
such that rVi s has a representative in Sw . The definition of the „–relation implies that each
Ti is connected, so by the Helly property of trees, there exists a vertex v P T such that for
each i, there exists Vvi P Sv representing rVi s. Moreover, we have Vvi KVvj for i ‰ j, since the
edge–hieromorphisms preserve the orthogonality relation. Applying the partial realization
axiom (Definition 1.1.(8)) to tp1i P CVvi u, where p1i is the image of pi under the appropriate
comparison map, yields a point x P Xv such that πVvi pxq is coarsely equal to p1i for all i, whence
drVi s px, pi q is uniformly bounded. If rVi s Ď rW s, then W has a representative Wv P Sv such
rV s
that Vvi Ď W , whence drW s px, ρWi q is uniformly bounded since x is a partial realization
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
44
point for tVvi u in Sv . Finally, if rW s&rVi s, then either the subtrees of T supporting rW s and
rV s
rVi s are disjoint, in which case drW s px, ρrWi s q is bounded, or rW s has a representative in Sv
transverse to Vvi , in which case the same inequality holds by our choice of x.
Uniqueness of realization points: Suppose x, y P X satisfy drV s px, yq ď K for all rV s P
S. Then for each vertex v P T , applying the uniqueness axiom in Xv to gv pxq, gv pyq shows
that, since dXv pgv pxq, gv pyqq ď ζ “ ζpKq. Indeed, otherwise we would have dV pgv pxq, gv pyqq ą
ξK`ξ for some V P Sv , whence drV s px, yq ą K. There exists k ď K and a sequence v0 , . . . , vk
of vertices in T so that x P Xv0 , y P Xvk . For each j, let xj “ gvj pxq and yj “ gvj pyq. Then
x “ x0 , y0 , x1 , y1 , . . . , yj´1 , xj , yj , . . . , xk , yk “ y is a path of uniformly bounded length joining x to y. Indeed, dXvj pxj , yj q ď ζ and k ď K by the preceding discussion, while xj coarsely
coincides with a point on the opposite side of an edge-space from yj´1 by the definition of
the gate of an edge-space in a vertex-space and the fact that xj´1 and yj´1 coarsely coincide.
This completes the proof.
l
8.1. Equivariant definition of pX pT q, Sq. Let T denote the tree of hierarchically hyperbolic spaces pT , tXv u, tXe u, tπe˘ uq, and let pX pT q, Sq be the hierarchically hyperbolic structure defined in the proof of Theorem 8.6. Various arbitrary choices were made in defining the
constituent hyperbolic spaces and projections in this hierarchically hyperbolic structure, and
we now insist on a specific way of making these choices in order to describe automorphisms
of pX pT q, Sq.
Recall that an automorphism of pX pT q, Sq is determined by a bijection g : S Ñ S and a
set of isometries g : CrV s Ñ CgrV s, for rV s P S. Via the distance formula, this determines a
uniform quasi-isometry
X pT q Ů
Ñ X pT q.
Ů
A bijection g : vPV Sv Ñ vPV Sv is T –coherent
if there is an induced isometry g of the
Ů
underlying tree, T , so that f g “ gf , where f : vPV Sv Ñ T sends each V P Sv to v, for all
v P V. The T –coherent bijection g is said to be T –coherent if it also preserves the relation „.
Noting that the composition of T –coherent bijections is T –coherent, denote by PT the group
of T –coherent bijections. For each g P PT , there is an induced bijection g : S Ñ S.
Recall that the hierarchically hyperbolic structure pX pT q, Sq was completely determined
except for the following three types of choices which were made arbitrarily.
(1) For each rV s P S, we chose an arbitrary favorite vertex v in the support of rV s; and,
(2) we chose an arbitrary favorite representative Vv P Sv with rV s “ rVv s. (Note that
if, as is often the case in practice, edge-hieromorphisms Se Ñ Sv are injective, then
Vv is the unique representative of its „–class that lies in Sv , and hence our choice is
uniquely determined.)
(3) For each rW s P S, the point ρTrW s is chosen arbitrarily in CW , where W is the favorite
representative of rW s.
We now constrain these choices so that they are equivariant. For each PT –orbit in S,
choose a representative rV s of that orbit, choose a favorite vertex v arbitrarily in its support,
and choose a favorite representative Vv P Sv of rV s. Then declare gVv P Sgv to be the
favorite representative, and gv the favorite vertex, associated to grV s, for each g P PT .
Recall that, for each rW s P S, we defined CrW s to be CW , where W is the favorite
representative of rW s. Suppose that we have specified a subgroup G ď PT and, for each
rW s P S and g P PT , an isometry g : CrW s Ñ CgrW s. Then we choose ρTrW s in such a way
that ρTgrW s “ gρTrW s for each rW s P S and g P G.
8.2. Graphs of hierarchically hyperbolic groups. Recall that the finitely generated
group G is hierarchically hyperbolic if there is a hierarchically hyperbolic space pX , Sq such
that G ď AutpSq and the action of G on X is metrically proper and cobounded and the
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
45
action of G on S is co-finite (this, together with the definition of an automorphism, implies
that only finitely many isometry types of hyperbolic space are involved in the HHS structure).
Endowing G with a word-metric, we see that pG, Sq is a hierarchically hyperbolic space.
If pG, Sq and pG1 , S1 q are hierarchically hyperbolic groups, then a homomorphism of hierarchically hyperbolic groups φ : pG, Sq Ñ pG1 , S1 q is a homomorphism φ : G Ñ G1 that is
also a φ–equivariant hieromorphism as in Definition 1.21.
Recall that a graph of groups G is a graph Γ “ pV, Eq together with a set tGv : v P V u
of vertex groups, a set tGe : e P Eu of edge groups, and monomorphisms φ˘
e : Ge Ñ Ge˘ ,
where e˘ are the vertices incident to e. As usual, the total group G of G is the quotient
of p˚vPV Gv q ˚ FE , where FE is the free group generated by E, obtained by imposing the
following relations:
‚ e “ 1 for all e P E belonging to some fixed spanning tree T of Γ;
´
´1 for e P E and g P G .
‚ φ`
e
e pgq “ eφe pgqe
We are interested in the case where Γ is a finite graph and, for each v P V, e P E, we have
sets Sv , Se so that pGv , Sv q and pGe , Se q are hierarchically hyperbolic group structures for
which φ˘
e : Ge Ñ Ge˘ is a homomorphism of hierarchically hyperbolic groups. In this case, G
is a finite graph of hierarchically hyperbolic groups. If in addition each φ˘
e has hierarchically
quasiconvex image, then G has quasiconvex edge groups.
r denote the Bass-Serre tree, observe that T “ Gr “ pΓ,
r tGṽ u, tGẽ u, tφẽ ˘uq is a
Letting Γ
r and each
tree of hierarchically hyperbolic spaces, where ṽ ranges over the vertex set of Γ,
r
Gṽ is a conjugate in the total group G to Gv , where ṽ ÞÑ v under Γ Ñ Γ, and an analogous
statement holds for edge-groups. Each φ˘
ẽ is conjugate to an edge-map in G in the obvious
way. We say G has bounded supports if T does.
Corollary 8.22 (Combination theorem for HHGs). Let G “ pΓ, tGv u, tGe u, tφ˘
e uq be a finite
graph of hierarchically hyperbolic groups. Suppose that:
(1) G has quasiconvex edge groups;
(2) each φ˘
e , as a hieromorphism, is full;
(3) G has bounded supports;
(4) if e is an edge of Γ and Se the Ď–maximal element of Se , then for all V P Se˘ , the
elements V and πpφe˘ qpSe q are not orthogonal in Se˘ .
Then the total group G of G is a hierarchically hyperbolic group.
Proof. By Theorem 8.6, pG, Sq is a hierarchically hyperbolic space. Observe that G ď PG ,
r and this action is induced by an action on Ť
since G acts on the Bass-Serre tree Γ,
vPV Sv
preserving the „–relation. Hence the hierarchically hyperbolic structure pG, Sq can be chosen according to the constraints in Section 8.1, whence it is easily checked that G acts by
automorphisms on pG, Sq. The action is obviously proper and cocompact; the action on S
is co-finite since each Gv is a hierarchically hyperbolic group.
l
If we apply the Corollary to hyperbolic groups, we immediately recover the following
special case of the celebrated Bestvina–Feighn Combination theorem [BF92]. The bounded
supports condition appears to be an obstruction to obtaining any substantially larger portion
of their theorem.
Remark 8.23 (Combination theorem for some graphs of hyperbolic groups). Corollary 8.22
implies that, if G is a finite graph of hyperbolic groups such that:
(1) for all edges e of Γ, Ge is quasiconvex in Ge˘ and
(2) for all vertices v of Γ, the collection tGe : v P te˘ uu is almost malnormal in Gv
then the total group G admits a hierarchically hyperbolic structure with trivial orthogonality
relation, from which it follows that G is again hyperbolic, by a result in [DHS15].
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
46
Remark 8.24 (Examples where the combination theorem does not apply). Examples where
one cannot apply Theorem 8.6 or Corollary 8.22 are likely to yield examples of groups that
are not hierarchically hyperbolic groups, or even hierarchically hyperbolic spaces.
(1) Let G be a finite graph of groups with Z2 vertex groups and Z edge groups, i.e., a
tubular group. In [Wis14], Wise completely characterized the tubular groups that act
freely on CAT(0) cube complexes, and also characterized those that admit cocompact
such actions; Woodhouse recently gave a necessary and sufficient condition for the
particular cube complex constructed in [Wis14] to be finite-dimensional [Woo15].
These results suggest that there is little hope of producing hierarchically hyperbolic
structures for tubular groups via cubulation, except in particularly simple cases.
This is because the obstruction to cocompact cubulation is very similar to the obstruction to building a hierarchically hyperbolic structure using Theorem 8.6. Indeed,
if some vertex-group Gv – Z2 has more than 2 independent incident edge-groups,
then, if G satisfied the hypotheses of Theorem 8.6, the hierarchically hyperbolic structure on Gv would include 3 pairwise-orthogonal unbounded elements, contradicting
partial realization. This shows that such a tubular group does not admit a hierarchically hyperbolic structure by virtue of the obvious splitting, and in fact shows
that there is no hierarchically hyperbolic structure in which Gv and the incident
edge-groups are hierarchically quasiconvex.
(2) Let G “ F ¸φ Z, where F is a finite-rank free group and φ : F Ñ F an automorphism. When F is atoroidal, G is a hierarchically hyperbolic group simply by
virtue of being hyperbolic [BF92, Bri00]. There is also a more refined hierarchically
hyperbolic structure in this case, in which all of the hyperbolic spaces involved are
quasi-trees. Indeed, by combining results in [HW15] and [Ago13], one finds that G
acts freely, cocompactly, and hence virtually co-specially on a CAT(0) cube complex,
which therefore contains a G-invariant factor system in the sense of [BHS14] and is
hence a hierarchically hyperbolic group; the construction in [BHS14] ensures that the
hierarchically hyperbolic structure for such cube complexes always uses a collection
of hyperbolic spaces uniformly quasi-isometric to trees. However, the situation is
presumably quite different when G is not hyperbolic. In this case, it seems that G is
rarely hierarchically hyperbolic.
8.3. Products. In this short section, we briefly describe a hierarchically hyperbolic structure
on products of hierarchically hyperbolic spaces.
Proposition 8.25 (Product HHS). Let pX0 , S0 q and pX1 , S1 q be hierarchically hyperbolic
spaces. Then X “ X0 ˆ X1 admits a hierarchically hyperbolic structure pX , Sq such that for
each of i P t0, 1u the inclusion map Xi Ñ X induces a quasiconvex hieromorphism.
Proof. Let pXi , Si q be hierarchically hyperbolic spaces for i P t0, 1u. Let S be a hierarchically hyperbolic structure consisting of the disjoint union of S0 and S1 (together with
their intrinsic hyperbolic spaces, projections, and nesting, orthogonality, and transversality
relations), along with the following domains whose associated hyperbolic spaces are points:
S, into which everything will be nested; Ui , for i P t0, 1u, into which everything in Si is
nested; for each U P Si a domain VU , with |CVU | “ 1, into which is nested everything in
Si`1 and everything in Si orthogonal to U . The elements VU are pairwise–transverse, and
are all transverse to U0 and U1 . Projections πU : X0 ˆ X1 Ñ U P S are defined in the obvious
way when U R S0 Y S1 ; otherwise, they are the compositions of the existing projections
with projection to the relevant factor. Projections of the form ρU
V are either defined already,
uniquely determined, or are chosen to coincide with the projection of some fixed basepoint
(when V P S0 YS1 and U is not). It is easy to check that this gives a hierarchically hyperbolic
structure on X1 ˆ X2 .
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
47
The hieromorphisms pXi , Si q Ñ pX , Sq are inclusions on Xi and S; for each U P Si ,
the map Si Q CU Ñ CU P S is the identity. It follows immediately from the definitions
that the diagrams from Definition 1.19 coarsely commute, so that these maps are indeed
hieromorphisms. Hierarchical quasiconvexity likewise follows from the definition.
l
Product HHS will be used in defining hierarchically hyperbolic structures on graph manifolds in Section 10. The next result follows directly from the proof of the previous proposition.
Corollary 8.26. Let G0 and G1 be hierarchically hyperbolic groups. Then G0 ˆ G1 is a
hierarchically hyperbolic group.
9. Hyperbolicity relative to HHGs
Relatively hyperbolic groups possess natural hierarchically hyperbolic structures:
Theorem 9.1 (Hyperbolicity relative to HHGs). Let the group G be hyperbolic relative to
a finite collection P of peripheral subgroups. If each P P P is a hierarchically hyperbolic
space, then G is a hierarchically hyperbolic space. Further, if each P P P is a hierarchically
hyperbolic group, then so is G.
Proof. We prove the statement about hierarchically hyperbolic groups; the statement about
spaces follows a fortiori.
For each P P P, let pP, SP q be a hierarchically hyperbolic group structure. For convenience, assume that the P P P are pairwise non-conjugate (this will avoid conflicting
hierarchically hyperbolic structures). For each P and each left coset gP , let SgP be a copy
of SP (with associated hyperbolic spaces and projections), so that there is a hieromorphism
pP, SP q Ñ pgP, SgP q, equivariant with respect to the conjugation isomorphism P Ñ P g .
p be the usual hyperbolic space formed from G by coning off each left coset of each
Let G
p YŮ
P P P. Let S “ tGu
gP PGP SgP . The nesting, orthogonality, and transversality relations
on each SgP are as defined above; if U, V P SgP , Sg1 P 1 and gP ‰ g 1 P 1 , then declare U &V .
p The hyperbolic space C G
p is G,
p while the hyperbolic space
Finally, for all U P S, let U Ď G.
CU associated to each U P SgP was defined above.
p is the inclusion. For each U P SgP , let
The projections are defined as follows: πGp : G Ñ G
ggP : G Ñ gP be the closest-point projection onto gP and let πU “ πU ˝ ggP , to extend the
domain of πU from gP to G. Since each πU was coarsely Lipschitz on U , and the closest-point
projection is uniformly coarsely Lipschitz, the projection πU is uniformly coarsely Lipschitz.
For each U, V P SgP , the coarse maps ρVU and ρU
V were already defined. If U P SgP and
U
1
1
1
1
V P Sg P , then ρV “ πV pgg P pgP qq, which is a uniformly bounded set (here we use relative
hyperbolicity, not just the weak relative hyperbolicity that is all we needed so far). Finally,
p we define ρU to be the cone-point over the unique gP with U P SgP , and
for U ‰ G,
p
G
p p
p
G
p is a cone-point
ρ : G Ñ CU is defined as follows: for x P G, let ρG pxq “ πU pxq. If x P G
U
S
U
1
1
g P
over g 1 P 1 ‰ gP , let ρG
, where Sg1 P 1 P Sg1 P 1 is Ď–maximal. The cone-point over
U pxq “ ρU
gP may be sent anywhere in U .
By construction, to verify that pG, Sq is a hierarchically hyperbolic group structure, it
suffices to verify that it satisfies the remaining axioms for a hierarchically hyperbolic space
given in Definition 1.1, since the additional G–equivariance conditions hold by construction.
Specifically, it remains to verify consistency, bounded geodesic image and large links, partial
realization, and uniqueness.
Consistency: The nested consistency inequality holds automatically within each SgP , so
p but this follows directly from the definition:
it remains to verify it only for U P SgP versus G,
p from the cone-point over gP , then ρGp pxq “ πU pxq, by definition. To
if x P G is far in G
U
p
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
48
verify the transverse inequality, it suffices to consider U P SgP , V P Sg1 P 1 with gP ‰ g 1 P 1 .
Let x P G and let z “ gg1 P 1 pxq. Then, if dU px, zq is sufficiently large, then dgP px, zq is
correspondingly large, so that by Lemma 1.15 of [Sis13], gg1 P 1 pxq and gg1 P 1 pgP q coarsely
coincide, as desired.
The last part of the consistency axiom, Definition 1.1.(4), holds as follows. Indeed, if
U Ď V , then either U “ V , and there is nothing to prove. Otherwise, if U Ď V and either
p The latter
V Ĺ W or W &V , then either U, V P SgP for some g, P , or U P SgP and V “ G.
situation precludes the existence of W , so we must be in the former situation. If W P SgP , we
p then U, V both project to the cone-point
are done since the axiom holds in SgP . If W “ G,
U
V
over gP , so ρW “ ρW . In the remaining case, W P Sg1 P 1 for some g 1 P 1 ‰ gP , in which case
V
ρU
W , ρW both coincide with πW pgg 1 P 1 pgP qq.
Bounded geodesic image: Bounded geodesic image holds within each SgP by construcp Let γ̂ be a geodesic in
tion, so it suffices to consider the case of U P SgP nested into G.
p
G avoiding gP and the cone on gP . Lemma 1.15 of [Sis13] ensures that any lift of γ̂ has
p
uniformly bounded projection on gP , so ρG
U ˝ γ̂ is uniformly bounded.
Large links: The large link axiom (Definition 1.1.(6)) can be seen to hold in pG, Sq by
combining the large link axiom in each gP with malnormality of P and Lemma 1.15 of [Sis13].
Partial realization: This follows immediately from partial realization within each SgP
and the fact that no new orthogonality was introduced in defining pG, Sq, together with the
p and the definition of projection between elements of SgP and Sg1 P 1 when
definition of G
1
1
gP ‰ g P . More precisely, if U P SgP and p P CU , then by partial realization within gP ,
there exists x P gP so that dU px, pq ď α for some fixed constant α and dV px, ρU
Vq ď α
U
for all V P SgP with U Ď V or U &V . Observe that dGp px, ρ p q “ 1, since x P gP and
G
ρUp is the cone-point over gP . Finally, if g 1 P 1 ‰ gP and V P Sg1 P 1 , then dV px, ρU
Vq “
G
dV pπV pgg1 P 1 pxqq, πV pgg1 P 1 pgP qqq “ 0 since x P gP .
p then either they are uniformly close in G,
Uniqueness: If x, y are uniformly close in G,
or they are uniformly close to a common cone-point, over some gP , whence the claim follows
from the uniqueness axiom in SgP .
l
Remark 9.2. The third author established a characterization of relative hyperbolicity in
terms of projections in [Sis13]. Further, there it was proven that like mapping class groups,
there was a natural way to compute distances in relatively hyperbolic groups from certain
related spaces, namely: if pG, Pq is relatively hyperbolic, then distances in G are coarsely
p together
obtained by summing the corresponding distance in the coned-off Cayley graph G
with the distances between projections in the various P P P and their cosets. We recover a
new proof of Sisto’s formula as a consequence of Theorem 9.1 and Theorem 4.5.
Theorem 9.1 will be used in our analysis of 3–manifold groups in Section 10. However, there
is a more general statement in the context of metrically relatively hyperbolic spaces (e.g.,
what Drutu–Sapir call asymptotically tree-graded [DS05], or spaces that satisfy the equivalent
condition on projections formulated in [Sis12]). For instance, arguing exactly as in the proof
of Theorem 9.1 shows that if the space X is hyperbolic relative to a collection of uniformly
hierarchically hyperbolic spaces, then X admits a hierarchically hyperbolic structure (in
which each peripheral subspace embeds hieromorphically).
More generally, let the geodesic metric space X be hyperbolic relative to a collection P
of subspaces, and let Xp be the hyperbolic space obtained from X by coning off each P P P.
Then we can endow X with a hierarchical space structure as follows:
‚ the index-set S consists of P together with an additional index S;
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
49
‚ for all P, Q P P, we have P &Q, while P Ĺ S for all P P P (the orthogonality relation
is empty and there is no other nesting);
‚ for each P P P, we let CP “ P ;
‚ we declare CS “ Xp;
‚ the projection πS : X Ñ Xp is the inclusion;
‚ for each P P P, let πP : X Ñ P be the closest-point projection onto P ;
‚ for each P P P, let ρPS be the cone-point in Xp associated to P ;
‚ for each P P P, let ρSP : Xp Ñ P be defined by ρSP pxq “ πP pxq for x P X , while
ρSP pxq “ πP pQq whenever x lies in the cone on Q P P.
‚ for distinct P, Q P P, let ρPQ “ πQ pP q (which is uniformly bounded since X is
hyperbolic relative to P).
The above definition yields:
Theorem 9.3. Let the geodesic metric space X be hyperbolic relative to the collection P
of subspaces. Then, with S as above, we have that pX , Sq is a hierarchical space, and is
moreover relatively hierarchically hyperbolic.
Proof. By definition, for each U P S, we have that either U “ S and CS “ Xp is hyperbolic,
or U is Ď–minimal. The rest of the conditions of Definition 1.1 are verified as in the proof
of Theorem 9.1.
l
10. Hierarchical hyperbolicity of 3-manifold groups
In this section we show that fundamental groups of most 3–manifolds admit hierarchical
hyperbolic structures. More precisely, we prove:
Theorem 10.1 (3–manifolds are hierarchically hyperbolic). Let M be a closed 3–manifold.
Then π1 pM q is a hierarchically hyperbolic space if and only if M does not have a Sol or Nil
component in its prime decomposition.
Proof. It is well known that for a closed irreducible 3–manifold N the Dehn function of π1 pN q
is linear if N is hyperbolic, cubic if N is Nil, exponential if N is Sol, and quadratic in all
other cases. Hence by Corollary 7.5, if π1 pM q is a hierarchically hyperbolic space, then M
does not contain Nil or Sol manifolds in its prime decomposition. It remains to prove the
converse.
Since the fundamental group of any reducible 3–manifold is the free product of irreducible
ones, the reducible case immediately follows from the irreducible case by Theorem 9.1.
When M is geometric and not Nil or Sol, then π1 pM q is quasi-isometric to one of the
following:
‚ R3 is hierarchically hyperbolic by Proposition 8.25;
‚ H3 , S3 , S2 ˆ R are (hierarchically) hyperbolic;
‚ H2 ˆ R and P SL2 pRq: the first is hierarchically hyperbolic by Proposition 8.25,
whence the second is also since it is quasi-isometric to the first by [Rie01].
We may now assume M is not geometric. Our main step is to show that any irreducible
non-geometric graph manifold group is a hierarchically hyperbolic space.
Let M be an irreducible non-geometric graph manifold. By [KL98, Theorem 2.3], by
replacing M by a manifold whose fundamental group is quasi-isometric to that of M , we may
assume that our manifold is a flip graph manifold, i.e., each Seifert fibered space component
is a trivial circle bundles over a surfaces of genus at least 2 and each pair of adjacent Seifert
fibered spaces are glued by flipping the base and fiber directions.
Let X be the universal cover of M . The decomposition of M into geometric components
induces a decomposition of X into subspaces tSv u, one for each vertex v of the Bass-Serre
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
50
tree T of M . Each such subspace Sv is bi-Lipschitz homeomorphic to the product of a copy
Rv of the real line with the universal cover Σv of a hyperbolic surface with totally geodesic
boundary, and there are maps φv : Sv Ñ Σv and ψv : Sv Ñ Rv . Notice that Σv is hyperbolic,
and in particular hierarchically hyperbolic. However, for later purposes, we endow Σv with
the hierarchically hyperbolic structure originating from the fact that Σv is hyperbolic relative
to its boundary components, see Theorem 9.1.
By Proposition 8.25 each Sv is a hierarchically hyperbolic space and thus we have a tree
of hierarchically hyperbolic spaces. Each edge space is a product B0 Σv ˆ Rv , where B0 Σv is
a particular boundary component of Σv determined by the adjacent vertex. Further, since
the graph manifold is flip, we also have that for each vertices v, w of the tree, the edgehieromorphism between Sv and Sw sends B0 Σv to Rw and Rv to B0 Σw .
We now verify the hypotheses of Corollary 8.22. The first hypothesis is that there exists
k so that each edge-hieromorphism is k–hierarchically quasiconvex. This is easily seen since
the edge-hieromorphisms have the simple form described above. The second hypothesis of
Theorem 8.6, fullness of edge-hieromorphisms, also follows immediately from the explicit
description of the edges here and the simple hierarchically hyperbolic structure of the edge
spaces.
The third hypothesis of Theorem 8.6 requires that the tree has bounded supports. We can
assume that the product regions Sv are maximal in the sense that each edge-hieromorphism
sends the fiber direction Rv to B0 Σw in each adjacent Sw . It follows that the support of each
„–class (in the language of Theorem 8.6) consists of at most 2 vertices. The last hypothesis
of Theorem 8.6 is about non-orthogonality of maximal elements and again follows directly
from the explicit hierarchically hyperbolic structure.
All the hypotheses of Theorem 8.6 are satisfied, so π1 M (with any word metric) is a
hierarchically hyperbolic space for all irreducible non-geometric graph manifolds M .
The general case that the fundamental group of any non-geometric 3–manifold is an hierarchically hyperbolic space now follows immediately by Theorem 9.1, since any 3–manifold
group is hyperbolic relative to its maximal graph manifold subgroups.
l
Remark 10.2 ((Non)existence of HHG structures for 3–manifold groups). The proof of Theorem 10.1 shows that for many 3–manifolds M , the group π1 M is not merely a hierarchically
hyperbolic space (when endowed with the word metric arising from a finite generating set),
but is actually a hierarchically hyperbolic group. Specifically, if M is virtually compact special, then π1 M acts freely and cocompactly on a CAT(0) cube complex X that is the universal
cover of a compact special cube complex. Hence X contains a π1 M –invariant factor system
(see [BHS14, Section 8]) consisting of a π1 M –finite set of convex subcomplexes. This yields
a hierarchically hyperbolic structure pX , Sq where π1 M ď AutpSq acts cofinitely on S and
geometrically on X , i.e., π1 M is a hierarchically hyperbolic group.
The situation is quite different when π1 M is not virtually compact special. Indeed, when
M is a nonpositively-curved graph manifold, π1 M virtually acts freely, but not necessarily cocompactly, on a CAT(0) cube complex X , and the quotient is virtually special; this is a result
of Liu [Liu11] which was also shown to hold in the case where M has nonempty boundary by
Przytycki and Wise [PW11]. Moreover, π1 M acts with finitely many orbits of hyperplanes.
Hence the π1 M –invariant factor system on X from [BHS14] yields a π1 M –equivariant HHS
structure pX , Sq with S π1 M –finite. However, this yields an HHG structure on π1 M only
if the action on X is cocompact. In [HP13], the second author and Przytycki showed that
π1 M virtually acts freely and cocompactly on a CAT(0) cube complex, with special quotient,
only in the very particular situation where M is chargeless. This essentially asks that the
Ă from the proof of Theorem 10.1
construction of the hierarchically hyperbolic structure on M
Ă
can be done π1 M –equivariantly. In general, this is impossible: recall that we passed from M
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
51
to the universal cover of a flip manifold using a (nonequivariant) quasi-isometry. Motivated
by this observation and the fact that the range of possible HHS structures on the universal
cover of a JSJ torus is very limited, we conjecture that π1 M is a hierarchically hyperbolic
group if and only if π1 M acts freely and cocompactly on a CAT(0) cube complex.
11. A new proof of the distance formula for mapping class groups
We now describe the hierarchically hyperbolic structure of mapping class groups. In
[BHS14] we gave a proof of this result using several of the main results of [Beh06, BKMM12,
MM99, MM00]. Here we give an elementary proof which is independent of the Masur–Minsky
“hierarchy machinery.” One consequence of this is a new and concise proof of the celebrated
Masur–Minsky distance formula [MM00, Theorem 6.12], which we obtain by combining Theorem 11.1 and Theorem 4.5.
(1) Let S be closed connected oriented surface of finite type and let MpSq be its marking
complex.
(2) Let S be the collection of isotopy classes of essential subsurfaces of S, and for each
U P S let CU be its curve complex.
(3) The relation Ď is nesting, K is disjointness and & is overlapping.
(4) For each U P S, let πU : MpSq Ñ CU be the (usual) subsurface projection. For
U, V P S satisfying either U Ď V or U &V , denote ρU
V “ πV pBU q P CV , while for
CV be the subsurface projection.
:
CU
Ñ
2
V Ď U let ρU
V
Theorem 11.1. Let S be closed connected oriented surface of finite type. Then, pMpSq, Sq
is a hierarchically hyperbolic space, for S as above. In particular the mapping class group
MCGpSq is a hierarchically hyperbolic group.
Proof. Hyperbolicity of curve graphs is the main result of [MM99]; more recently, elementary
proofs of this were found in [Aou13, Bow14b, CRS15, HPW13, PS15].
Axioms 1, 2, 3 and 5 are clear.
Both parts of axiom 4 can be found in [Beh06]. The nesting part is elementary, and a
short elementary proof in the overlapping case was obtained by Leininger and can be found
in [Man10].
Axiom 7 was proven in [MM00], and an elementary proof is available in [Web13]. In
fact, in the aforementioned papers it is proven that there exists a constant C so that for
any subsurface W , markings x, y and geodesic from πW pxq to πW pyq the following holds. If
V Ď W and V ‰ W satisfies dV px, yq ě C then some curve along the given geodesic does
not intersect BV . This implies Axiom 6, since we can take the Ti to be the complements of
curves appearing along the aforementioned geodesic.
Axiom 8 follows easily from the following statement, which clearly holds: For any given
collection of disjoint subsurfaces and curves on the given subsurfaces, there exists a marking
on S that contains the given curves as base curves (or, up to bounded error, transversals in
the case that the corresponding subsurface is an annulus).
Axiom 9 is hence the only delicate one. We are finished modulo this last axiom which we
verify below in Proposition 11.2.
l
Proposition 11.2. pMpSq, Sq satisfies the uniqueness axiom, i.e., for each κ ě 0, there
exists θu “ θu pκq such that if x, y P MpSq satisfy dU px, yq ď κ for each U P S then
dMpSq px, yq ď θu .
Proof. Note that when the complexity (as measured by the quantity 3g ` p ´ 3 where g is the
genus and p the number of punctures) is less than 2 then MpSq is hyperbolic and thus the
axiom holds. We thus will proceed by inducting on complexity: thus we will fix S to have
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
52
complexity at least 2 and assume that all the axioms for a hierarchically hyperbolic space,
including the uniqueness axiom, hold for each proper subsurface of S.
Now, fixed our surface S, the proof is by induction on dCS pbasepxq, basepyqq.
If dCS pbasepxq, basepyqq “ 0, then x and y share some non-empty multicurve σ “ c1 Y
¨ ¨ ¨ Y ck . For x1 , y 1 the restrictions of x, y to S ´ σ we have that, by induction, dMpS´σq px1 , y 1 q
is bounded in terms of κ. We then take the markings in a geodesic in MpS ´ σq from x1 to
y 1 and extend these all in the same way to obtain markings in MpSq which yield a path in
MpSq from x to ŷ whose length is bounded in terms of κ, where ŷ is the marking for which:
‚ ŷ has the same base curves as y,
‚ the transversal for each ci is the same as the corresponding transversal for x, and
‚ the transversal for each curve in basepyq ´ tci u is the same as the corresponding
transversal for y.
Finally, it is readily seen that dMpSq pŷ, yq is bounded in terms of κ because the transversals
of each ci in the markings x and y are within distance κ of each other. This completes the
proof of the base case of the Proposition.
Suppose now that that the statement holds whenever dCS pbasepxq, basepyqq ď n, and let
us prove it in the case dCS pbasepxq, basepyqq “ n ` 1. Let cx P basepxq and cy P basepyq
satisfy dCS pcx , cy q “ n ` 1. Let cx “ σ0 , . . . , σn`1 “ cy be a tight geodesic (hence, each σi
is a multicurve). Let σ be the union of σ0 and σ1 . Using the realization theorem in the
subsurface S ´ σ we can find a marking x1 in S ´ σ whose projections onto each CU for
U Ď S ´ σ coarsely coincide with πU pyq. Let x̂ be the marking for which:
‚ basepx̂q is the union of basepx1 q and σ,
‚ the transversal in x̂ of curves in basepx̂q X basepx1 q are the same as those in x1 ,
‚ the transversal of cx in x̂ is the same as the one in x,
‚ the transversal in x̂ of a curve c in σ1 is πAc pyq, where Ac is an annulus around y.
Note that dCS pbasepx̂q, basepyqq “ n. Hence, the following two claims conclude the proof.
Claim 1. dMpSq px, x̂q is bounded in terms of κ.
Proof. It suffices to show that we have a bound on dCU px, x̂q in terms of κ for each U Ď
S ´ cx . In fact, once we do that, by induction on complexity we know that we can bound
dMpS´cx q pz, ẑq, where z, ẑ are the restrictions of x, x̂ to S ´ cx , whence the conclusion easily
follows.
If U is contained in S ´ σ, then the required bound follows since πU px̂q coarsely coincides
with πU px1 q in this case.
If instead BU intersects σ1 , then πU px̂q coarsely coincides with πU pσ1 q.
At this point, we only have to show that πU pσ1 q coarsely coincides with πU pyq, and in
order to do so we observe that we can apply the bounded geodesic image theorem to the
geodesic σ1 , . . . , σn`1 . In fact, σ1 intersects BU by hypothesis and σi intersects BU for i ě 3
because of the following estimate that holds for any given boundary component c of BU
dCpSq pσi , cq ě dCpSq pσi , σ0 q ´ dCpSq pσ0 , cq ě i ´ 1 ą 1.
Finally, σ2 intersects BU because of the definition of tightness: BU intersects σ1 , hence it
must intersect σ0 Y σ2 . However, it does not intersect σ0 , whence it intersects σ2 .
Claim 2. There exists κ1 , depending on κ, so that for each subsurface U of S we have
dCU px̂, yq ď κ1 .
Proof. If σ0 intersects BU , then πU px̂q coarsely coincides with πU pσ0 q. In turn, πU pσ0 q
coarsely coincides with πU pxq, which is κ–close to πU pyq.
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
53
On the other hand, if U does not intersect σ, then we are done by the definition of x1 .
Hence, we can assume that U is contained in S´σ0 and that σ1 intersects BU . In particular,
πU px̂q coarsely coincides with πU pσ1 q. But we showed in the last paragraph of the proof of
Claim 1 that πU pσ1 q coarsely coincides with πU pyq, so we are done.
As explained above, the proofs of the above two claims complete the proof.
l
References
[Ago13]
Ian Agol. The virtual Haken conjecture (with an appendix by Ian Agol, Daniel Groves and Jason
Manning). Doc. Math., J. DMV, 18:1045–1087, 2013.
[Aou13]
Tarik Aougab. Uniform hyperbolicity of the graphs of curves. Geometry & Topology, 17(5):2855–
2875, 2013.
[BDS11]
J. Behrstock, C. Druţu, and M. Sapir. Median structures on asymptotic cones and homomorphisms into mapping class groups. Proc. London Math. Soc., 102(3):503–554, 2011.
[Beh06]
J. Behrstock. Asymptotic geometry of the mapping class group and Teichmüller space. Geometry
& Topology, 10:2001–2056, 2006.
[BF92]
M. Bestvina and M. Feighn. A combination theorem for negatively curved groups. J. Differential
Geometry, 35:85–101, 1992.
[BF02]
M. Bestvina and K. Fujiwara. Bounded cohomology of subgroups of mapping class groups. Geometry & Topology, 6:69–89 (electronic), 2002.
[BF14a]
Mladen Bestvina and Mark Feighn. Hyperbolicity of the complex of free factors. Adv. Math.,
256:104–155, 2014.
[BF14b]
Mladen Bestvina and Mark Feighn. Subfactor projections. J. Topol., 7(3):771–804, 2014.
[BH]
Jason Behrstock and Mark F. Hagen. Cubulated groups: thickness, relative hyperbolicity, and
simplicial boundaries. Arxiv:1212.0182. To appear in Groups, Geometry, and Dynamics.
[BHS14]
Jason Behrstock, Mark F. Hagen, and Alessandro Sisto. Hierarchically hyperbolic spaces I: curve
complexes for cubical groups. arXiv:1412.2171, 2014.
[BKMM12] Jason Behrstock, Bruce Kleiner, Yair Minsky, and Lee Mosher. Geometry and rigidity of mapping
class groups. Geom. Topol, 16(2):781–888, 2012.
[BM08]
Jason A Behrstock and Yair N Minsky. Dimension and rank for mapping class groups. Annals of
Mathematics, 167:1055–1077, 2008.
[BM11]
Jason A. Behrstock and Yair N. Minsky. Centroids and the rapid decay property in mapping
class groups. J. Lond. Math. Soc. (2), 84(3):765–784, 2011.
[Bow13]
Brian H. Bowditch. Coarse median spaces and groups. Pacific J. Math, 261(1):53–93, 2013.
[Bow14a] Brian H Bowditch. Embedding median algebras in products of trees. Geometriae Dedicata,
170(1):157–176, 2014.
[Bow14b] Brian H. Bowditch. Uniform hyperbolicity of the curve graphs. Pacific J. Math, 269(2):269–280,
2014.
[Bow15]
Brian H. Bowditch. Large–scale rigidity properties of the mapping class groups. Preprint, 2015.
[Bri00]
Peter Brinkmann. Hyperbolic automorphisms of free groups. Geometric & Functional Analysis,
10(5):1071–1089, 2000.
[Bri10]
Martin R Bridson. Semisimple actions of mapping class groups on CAT(0) spaces. Geometry of
Riemann surfaces, 368:1–14, 2010.
[BS05]
S Buyalo and P Svetlov. Topological and geometric properties of graph-manifolds. St. Petersburg
Mathematical Journal, 16(2):297–340, 2005.
[BV]
M R Bridson and K Vogtmann. The Dehn functions of OutpFn q and AutpFn q. arXiv:1011.1506.
[BV95]
M R Bridson and K Vogtmann. On the geometry of the automorphism group of a free group.
Bulletin of the London Mathematical Society, 27(6):544–552, 1995.
[CDH10]
Indira Chatterji, Cornelia Druţu, and Frédéric Haglund. Kazhdan and Haagerup properties from
the median viewpoint. Advances in Mathematics, 225(2):882–921, 2010.
[Che00]
Victor Chepoi. Graphs of some CATp0q complexes. Adv. in Appl. Math., 24(2):125–179, 2000.
[CLM12]
Matt T. Clay, Christopher J. Leininger, and Johanna Mangahas. The geometry of right-angled
Artin subgroups of mapping class groups. Groups Geom. Dyn., 6(2):249–278, 2012.
[CRS15]
Matt Clay, Kasra Rafi, and Saul Schleimer. Uniform hyperbolicity of the curve graph via surgery
sequences. Algebraic and Geometric Topology, 14(6):3325–3344, 2015.
[DHS15]
Matthew Gentry Durham, Mark F. Hagen, and Alessandro Sisto. Boundaries of hierarchically
hyperbolic spaces. In preparation, 2015.
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
[Dil50]
[DS05]
[DT14]
[FLS11]
[Ger92]
[GHR03]
[Hae14]
[Hag07]
[Hag14]
[HM13a]
[HM13b]
[HP13]
[HPW13]
[HW15]
[KK14]
[KL96]
[KL97]
[KL98]
[Liu11]
[Man10]
[MM99]
[MM00]
[MO14]
[Mos95]
[PS15]
[PW11]
[Rie01]
[Sis11]
[Sis12]
[Sis13]
[SS]
[Web13]
[Wis14]
[Woo15]
54
R. P. Dilworth. A decomposition theorem for partially ordered sets. Ann. of Math, 51:161–166,
1950.
C. Druţu and M. Sapir. Tree-graded spaces and asymptotic cones of groups. Topology, 44(5):959–
1058, 2005. With an appendix by Denis Osin and Mark Sapir.
Matthew Gentry Durham and Samuel J Taylor. Convex cocompactness and stability in mapping
class groups. arXiv:1404.4803, 2014.
Roberto Frigerio, Jean-François Lafont, and Alessandro Sisto. Rigidity of high dimensional graph
manifolds. arXiv:1107.2019, 2011. To appear in Asterisque.
Stephen M Gersten. Dehn functions and `1 -norms of finite presentations. Springer, 1992.
Steve M Gersten, Derek F Holt, and Tim R Riley. Isoperimetric inequalities for nilpotent groups.
Geometric & Functional Analysis, 13(4):795–814, 2003.
Thomas Haettel. Higher rank lattices are not coarse median. arXiv:1412.1714, 2014.
Frédéric Haglund. Isometries of CAT(0) cube complexes are semi-simple. arXiv:0705.3386, 2007.
Mark F. Hagen. Weak hyperbolicity of cube complexes and quasi-arboreal groups. J. Topol.,
7(2):385–418, 2014.
Michael Handel and Lee Mosher. The free splitting complex of a free group, I: hyperbolicity.
Geom. Topol., 17(3):1581–1672, 2013.
Michael Handel and Lee Mosher. Lipschitz retraction and distortion for subgroups of OutpFn q.
Geometry & Topology, 17(3):1535–1579, 2013.
Mark F. Hagen and Piotr Przytycki. Cocompactly cubulated graph manifolds. To appear in Isr.
J. Math. arXiv:1310.1309, 2013.
Sebastian Hensel, Piotr Przytycki, and Richard CH Webb. Slim unicorns and uniform hyperbolicity for arc graphs and curve graphs. arXiv:1301.5577, 2013.
Mark F. Hagen and Daniel T. Wise. Cubulating hyperbolic free-by-cyclic groups: the general
case. Geometric and Functional Analysis, 25(1):134–179, 2015.
Sang-Hyun Kim and Thomas Koberda. The geometry of the curve graph of a right-angled Artin
group. International Journal of Algebra and Computation, 24(02):121–169, 2014.
Michael Kapovich and Bernhard Leeb. Actions of discrete groups on nonpositively curved spaces.
Math. Ann., 306(2):341–352, 1996.
M. Kapovich and B. Leeb. Quasi-isometries preserve the geometric decomposition of Haken
manifolds. Invent. Math., 128(2):393–416, 1997.
M. Kapovich and B. Leeb. 3-manifold groups and nonpositive curvature. Geom. Funct. Anal.,
8(5):841–852, 1998.
Yi Liu. Virtual cubulation of nonpositively curved graph manifolds. ArXiv:1110.1940, 2011.
Johanna Mangahas. Uniform uniform exponential growth of subgroups of the mapping class
group. Geometric and Functional Analysis, 19(5):1468–1480, 2010.
Howard A Masur and Yair N Minsky. Geometry of the complex of curves I: Hyperbolicity. Inventiones mathematicae, 138(1):103–149, 1999.
Howard A Masur and Yair N Minsky. Geometry of the complex of curves II: Hierarchical structure. Geometric and Functional Analysis, 10(4):902–974, 2000.
Ashot Minasyan and Denis Osin. Acylindrical hyperbolicity of groups acting on trees. Mathematische Annalen, pages 1–51, 2014.
L. Mosher. Mapping class groups are automatic. Ann. of Math., 142:303–384, 1995.
Piotr Przytycki and Alessandro Sisto. A note on acylindrical hyperbolicity of mapping class
groups. arXiv:1502.02176, 2015.
Piotr Przytycki and Daniel T. Wise. Graph manifolds with boundary are virtually special.
ArXiv:1110.3513, 2011.
E. Rieffel. Groups quasi-isometric to H2 ˆ R. J. London Math. Soc. (2), 64(1):44–60, 2001.
Alessandro Sisto. 3-manifold groups have unique asymptotic cones. arXiv:1109.4674, 2011.
Alessandro Sisto. On metric relative hyperbolicity. arXiv:1210.8081, 2012.
Alessandro Sisto. Projections and relative hyperbolicity. Enseignment Math., (59):165–181, 2013.
L. Sabalka and D. Savchuk. Submanifold projection. arXiv:1211.3111.
Richard CH Webb. A short proof of the bounded geodesic image theorem. arXiv:1301.6187, 2013.
Daniel Wise. Cubular tubular groups. Transactions of the American Mathematical Society,
366(10):5503–5521, 2014.
Daniel J Woodhouse. Classifying finite dimensional cubulations of tubular groups.
arXiv:1502.02619, 2015.
HHS II: COMBINATION THEOREMS AND THE DISTANCE FORMULA
Lehman College and The Graduate Center, CUNY, New York, New York, USA
Current address: Barnard College, Columbia University, New York, New York, USA
E-mail address: [email protected]
Dept. of Pure Maths and Math. Stat., University of Cambridge, Cambridge, UK
E-mail address: [email protected]
ETH, Zürich, Switzerland
E-mail address: [email protected]
55
| 4math.GR
|
A Self-Taught Artificial Agent for Multi-Physics
Computational Model Personalization
Dominik Neumanna,c,∗, Tommaso Mansib , Lucian Itud,e ,
Bogdan Georgescub , Elham Kayvanpourf , Farbod Sedaghat-Hamedanif ,
Ali Amrf , Jan Haasf , Hugo Katusf , Benjamin Mederf , Stefan Steidlc ,
Joachim Horneggerc , Dorin Comaniciub
arXiv:1605.00303v1 [cs.CE] 1 May 2016
a
Medical Imaging Technologies, Siemens Healthcare GmbH, Erlangen, Germany
b
Medical Imaging Technologies, Siemens Healthcare, Princeton, USA
c
Pattern Recognition Lab, FAU Erlangen-Nürnberg, Erlangen, Germany
d
Siemens Corporate Technology, Siemens SRL, Brasov, Romania
e
Transilvania University of Brasov, Brasov, Romania
f
Department of Internal Medicine III, University Hospital Heidelberg, Germany
Abstract
Personalization is the process of fitting a model to patient data, a critical
step towards application of multi-physics computational models in clinical
practice. Designing robust personalization algorithms is often a tedious,
time-consuming, model- and data-specific process. We propose to use artificial intelligence concepts to learn this task, inspired by how human experts
manually perform it. The problem is reformulated in terms of reinforcement learning. In an off-line phase, Vito, our self-taught artificial agent,
learns a representative decision process model through exploration of the
computational model: it learns how the model behaves under change of
parameters. The agent then automatically learns an optimal strategy for
on-line personalization. The algorithm is model-independent; applying it
to a new model requires only adjusting few hyper-parameters of the agent
and defining the observations to match. The full knowledge of the model
itself is not required. Vito was tested in a synthetic scenario, showing that
it could learn how to optimize cost functions generically. Then Vito was
applied to the inverse problem of cardiac electrophysiology and the personalization of a whole-body circulation model. The obtained results suggested
that Vito could achieve equivalent, if not better goodness of fit than standard methods, while being more robust (up to 11% higher success rates)
and with faster (up to seven times) convergence rate. Our artificial intelligence approach could thus make personalization algorithms generalizable
and self-adaptable to any patient and any model.
∗
Corresponding author
Email address: [email protected] (Dominik Neumann)
Preprint submitted to Medical Image Analysis
May 3, 2016
Keywords: Computational Modeling, Model Personalization,
Reinforcement Learning, Artificial Intelligence.
1. Introduction
Computational modeling attracted significant attention in cardiac research over the last decades (Clayton et al., 2011; Frangi et al., 2001; Hunter
and Borg, 2003; Kerckhoffs et al., 2008; Krishnamurthy et al., 2013; Kuijpers et al., 2012; Noble, 2002). It is believed that computational models
can improve patient stratification and therapy planning. They could become the enabling tool for predicting disease course and therapy outcome,
ultimately leading to improved clinical management of patients suffering
from cardiomyopathies (Kayvanpour et al., 2015). A crucial prerequisite for
achieving these goals is precise model personalization: the computational
model under consideration needs to be fitted to each patient. However, the
high complexity of cardiac models and the often noisy and sparse clinical
data still hinder this task.
A wide variety of manual and (semi-)automatic model parameter estimation approaches have been explored, including Aguado-Sierra et al. (2010,
2011); Augenstein et al. (2005); Chabiniok et al. (2012); Delingette et al.
(2012); Itu et al. (2014); Konukoglu et al. (2011); Le Folgoc et al. (2013);
Marchesseau et al. (2013); Neumann et al. (2014a,b); Prakosa et al. (2013);
Schmid et al. (2006); Seegerer et al. (2015); Sermesant et al. (2009); Wallman et al. (2014); Wang et al. (2009); Wong et al. (2015); Xi et al. (2013);
Zettinig et al. (2014). Most methods aim to iteratively reduce the misfit
between model output and measurements using optimization algorithms,
for instance variational (Delingette et al., 2012) or filtering (Marchesseau
et al., 2013) approaches. Applied blindly, those techniques could easily fail
on unseen data, if not supervised, due to parameter ambiguity, data noise
and local minima (Konukoglu et al., 2011; Neumann et al., 2014a; Wallman
et al., 2014). Therefore, complex algorithms have been designed combining
cascades of optimizers in a very specific way to achieve high levels of robustness, even on larger populations, i.e. 10 or more patients (Kayvanpour et al.,
2015; Neumann et al., 2014b; Seegerer et al., 2015). However, those methods
are often designed from tedious, trial-and-error-driven manual tuning, they
are model-specific rather than generic, and their generalization to varying
data quality cannot be guaranteed. On the contrary, if the personalization
task is assigned to an experienced human, given enough time, he almost always succeeds in manually personalizing a model for any subject (although
solution uniqueness is not guaranteed, but this is inherent to the problem).
There are several potential reasons why a human expert is often superior to standard automatic methods in terms of personalization accuracy
and success rates. First, an expert is likely to have an intuition of the
2
model’s behavior from his prior knowledge of the physiology of the modeled
organ. Second, knowledge about model design and assumptions, and model
limitations and implementation details certainly provide useful hints on the
“mechanics” of the model. Third, past personalization of other datasets
allows the expert to build up experience. The combination of prior knowledge, intuition and experience enables to solve the personalization task more
effectively, even on unseen data.
Inspired by humans and contrary to previous works, we propose to address the personalization problem from an artificial intelligence (AI) perspective. In particular, we apply reinforcement learning (RL) methods (Sutton
and Barto, 1998) developed in the AI community to solve the parameter
estimation task for computational physiological models. With its roots in
control theory on the one hand, and neuroscience theories of learning on the
other hand, RL encompasses a set of approaches to make an artificial agent
learn from experience generated by interacting with its environment. Contrary to standard (supervised) machine learning (Bishop, 2006), where the
objective is to compute a direct mapping from input features to a classification label or regression output, RL aims to learn how to perform tasks. The
goal of RL is to compute an optimal problem-solving strategy (agent behavior), e.g. a strategy to play the game “tic-tac-toe” successfully. In the AI
field, such a behavior is often represented as a policy, a mapping from states,
describing the current “situation” the agent finds itself in (e.g. the current
locations of all “X” and “O” on the tic-tac-toe grid), to actions, which allow
the agent to interact with the environment (e.g. place “X” on an empty cell)
and thus influence that situation. The key underlying principle of RL is that
of reward (Kaelbling et al., 1996), which provides an objective means for the
agent to judge the outcome of its actions. In tic-tac-toe, the agent receives
a high, positive reward if the latest action led to a horizontal, vertical or
diagonal row full of “X” marks (winning), and a negative reward (punishment) if the latest action would allow the opponent to win in his next move.
Based on such rewards, the artificial agent learns an optimal winning policy
through trial-and-error interactions with the environment.
RL was first applied to game (e.g. Tesauro, 1994) or simple control
tasks. However, the past few years saw tremendous breakthroughs in RL for
more complex, real-world problems (e.g. Barreto et al., 2014; Kveton and
Theocharous, 2012; Nguyen-Tuong and Peters, 2011). Some noteworthy examples include Mülling et al. (2013), where the control entity of a robot
arm learned to select appropriate motor primitives to play table tennis, and
Mnih et al. (2015), where the authors combine RL with deep learning to
train an agent to play 49 Atari games, yielding better performance than an
expert in the majority of them.
Motivated by these recent successes and building on our previous work
(Neumann et al., 2015), we propose an RL-based personalization approach,
henceforth called Vito, with the goal of designing a framework that can, for
3
Off-line phase
Assimilate model
behavior
On-line phase
Learn optimal
personalization
strategy
Personalize model
for new patients
Figure 1: Overview of Vito: a self-taught artificial agent for computational model personalization, inspired by how human operators approach the personalization problem.
the first time to our knowledge, learn by itself how to estimate model parameters from clinical data while being model-independent. As illustrated in
Fig. 1, first, like a human expert, Vito assimilates the behavior of the physiological model under consideration in an off-line, one-time only, data-driven
exploration phase. From this knowledge, Vito learns the optimal strategy
using RL (Sutton and Barto, 1998). The goal of Vito during the on-line
personalization phase is then to sequentially choose actions that maximize
future rewards, and therefore bring Vito to the state representing the solution of the personalization problem. To setup the algorithm, the user needs
to define what observations need to be matched, the allowed actions, and a
single hyper-parameter related to the desired granularity of the state-space.
Then everything is learned automatically. The algorithm does not depend
on the underlying model.
Vito was evaluated on three different tasks. First, in a synthetic experiment, convergence properties of the algorithm were analyzed. Then,
two tasks involving real clinical data were evaluated: the inverse problem
of cardiac electrophysiology and the personalization of a lumped-parameter
model of whole-body circulation. The obtained results suggested that Vito
can achieve equivalent (or better) goodness of fit as standard optimization
methods, increased robustness and faster convergence rates.
A number of novelties and improvements over Neumann et al. (2015)
are featured in this manuscript. First, an automatic, data-driven statespace quantization method is introduced that replaces the previous manual
technique. Second, the need to provide user-defined initial parameter values
is eliminated by employing a new data-driven technique to initialize personalization of unseen data. Third, a stochastic personalization policy is
introduced, for which the previously used standard deterministic policy is
a special case. Fourth, the convergence properties are evaluated in parameter space using a synthetic personalization scenario. In addition, thorough
evaluation of Vito’s performance with increasing amount of training samples was conducted and personalization of the whole-body circulation model
was extended to several variants involving two to six parameters. Finally,
the patient database used for experimentation was extended from 28 to 83
patients for the cardiac electrophysiology experiments, and from 27 to 56
for the whole-body circulation experiments.
4
The remainder of this manuscript is organized as follows. Sec. 2 presents
the method. In Sec. 3, the experiments are described and the results are
presented. Sec. 4 concludes the manuscript with a summary and discussions
about potential limitations and extensions of the method.
2. Method
This section presents the reinforcement-learning (RL) framework for
computational model personalization. Sec. 2.1 introduces Markov decision
process (MDP). Sec. 2.2 defines the personalization problem and how it can
be reformulated in terms of an MDP. Sec. 2.3 describes how the artificial
agent, Vito, learns how the model behaves. Next, Sec. 2.4 provides details
about state-space quantization, and Sec. 2.5 describes how the model knowledge is encoded in the form of transition probabilities. All steps mentioned
so far are performed in an off-line training phase. Finally, Sec. 2.6 explains
how the learned knowledge is applied on-line to personalize unseen data.
2.1. Model-based Reinforcement Learning
2.1.1. MDP Definition
A crucial prerequisite for applying RL is that the problem of interest, here personalization, can be modeled as a Markov decision process
(MDP). An MDP is a mathematical framework for modeling decision making when the decision outcome is partly random and partly controlled by
a decision maker (Sutton and Barto, 1998). Formally, an MDP is a tuple
M = (S, A, T , R, γ), where:
• S is the finite set of states that describe the agent’s environment, nS
is the number of states, and st ∈ S is the state at time t.
• A is the finite set of actions, which allow the agent to interact with
the environment, nA is the number of actions, and at ∈ A denotes the
action performed at time t.
• T : S × A × S → [0; 1] is the stochastic transition function, where
T (st , at , st+1 ) describes the probability of arriving in state st+1 after
the agent performed action at in state st .
• R : S × A × S → R is the scalar reward function, where rt+1 =
R(st , at , st+1 ) is the immediate reward the agent receives at time t + 1
after performing action at in state st resulting in state st+1 .
• γ ∈ [0; 1] is the discount factor that controls the importance of future
versus immediate rewards.
5
2.1.2. Value Iteration
The value of a state, V ∗ (s), is the expected discounted reward the agent
accumulates when it starts in state s and acts optimally in each step:
(∞
)
X
∗
k
V (s) = E
γ rt+k+1 st = s ,
(1)
k=0
where E{} denotes the expected value given the agent always selects the
optimal action, and t is any time step. Note that the discount factor γ is
a constant and the superscript k its exponent. V ∗ can be computed using
value iteration (Sutton and Barto, 1998), an iterative algorithm based on
dynamic programming. In the first iteration i = 0, let Vi : S → R denote
an initial guess for the value function that maps states to arbitrary values.
Further, let Qi : S × A → R denote the ith “state-action value function”guess, which is computed as:
X
Qi (s, a) =
T (s, a, s0 ) R(s, a, s0 ) + γVi (s0 ) .
(2)
s0 ∈S
Value iteration iteratively updates Vi+1 from the previous Qi :
∀s ∈ S :
Vi+1 (s) = max Qi (s, a) ,
a∈A
(3)
until the left- and right-hand side of Eq. 3 are equal for all s ∈ S; then
V ∗ ← Vi+1 and Q∗ ← Qi+1 . From this equality relation, also known as
the Bellman equation (Bellman, 1957), one can obtain an optimal problemsolving strategy for the problem described by the MDP (assuming that all
components of the MDP are known precisely). It is encoded in terms of a
deterministic optimal policy π ∗ : S → A:
π ∗ (s) = arg max Q∗ (s, a) ,
(4)
a∈A
i.e. a mapping that tells the agent in each state the optimal action to take.
2.1.3. Stochastic Policy
In this work not all components of the MDP are known precisely, instead some are approximated from training data. Value iteration, however,
assumes an exact MDP to guarantee optimality of the computed policy.
Therefore, instead of relying on the deterministic policy π ∗ (Eq. 4), a generalization to stochastic policies π̃ ∗ is proposed here to mitigate potential
issues due to approximations. Contrary to Eq. 4, where for each state only
the one action with maximum Q∗ -value is considered, a stochastic policy
stores several candidate actions with similar high Q∗ -value and returns one
6
Model
parameters
Model
Model
state
Objectives
Measured
data
Patient
Figure 2: A computational model f is a dynamic system that maps model input parameters x to model state (output) variables y. The goal of personalization is to tune x such
that the objectives c, defined as the misfit between y and the corresponding measured
data z of a given patient, are optimized (the misfit is minimized).
of them through a random process each time it is queried. To this end, the
Q∗ (s, ·)-values for a given state s are first normalized:
Q̃∗s (a) =
Q∗ (s, a) − mina0 ∈A [Q∗ (s, a0 )]
.
maxa0 ∈A [Q∗ (s, a0 )] − mina0 ∈A [Q∗ (s, a0 )]
(5)
All actions whose normalized Q̃∗s -value is below a threshold of = 54 (set
empirically and used throughout the entire manuscript) are discarded, while
actions with large values are stored as potential candidates. Each time
the stochastic policy is queried, a = π̃∗ (s), it returns one of the candidate
actions P
a selected randomly with probability proportional to its Q̃∗s -value:
∗
Q̃s (a)/ a0 Q̃∗s (a0 ); the sum is over all candidate actions a0 .
2.2. Reformulation of the Model Personalization Problem into an MDP
2.2.1. Problem Definition
As illustrated in Fig. 2, any computational model f is governed by a set of
parameters x = (x1 , . . . , xnx )> , where nx denotes the number of parameters.
x is bounded within a physiologically plausible domain Ω, and characterized
by a number of ny (observable) state variables y = (y1 , . . . , yny )> . The
state variables can be used to estimate x. Note that some parameters may
be pre-estimated or assigned fixed values. The goal of personalization is to
optimize a set of nc objectives c = (c1 , . . . , cnc )> . The objectives are scalars
defined as ci = d(yi , zi ), where d is a measure of misfit, and zi denotes
the patient’s measured data (z) corresponding to yi . In this work d(yi , zi ) =
yi −zi . Personalization is considered successful if all user-defined convergence
criteria ψ = (ψ1 , . . . , ψnc )> are met. The criteria are defined in terms of
maximum acceptable misfit per objective: ∀i ∈ {1, . . . , nc } : |ci | < ψi .
2.2.2. Problem Reformulation
Personalization is mapped to a Markov decision process as follows:
States: An MDP state encodes the misfit between the computed model
state (outcome of forward model run) and the patient’s measurements.
Thus, MDP states carry the same type of information as objective vectors c,
yet the number of MDP states has to be finite (Sec. 2.1), while there are an
infinite number of different objective vectors due to their continuous nature.
Therefore the space of objective vectors in Rnc is reduced to a finite set of
7
representative states: the MDP states S, each s ∈ S covering a small region
in that space. One of those states, ŝ ∈ S, encodes personalization success
as it is designed such that it covers exactly the region where all convergence
criteria are satisfied. The goal of Vito is to learn how to reach that state.
Actions: Vito’s actions modify the parameters x to fulfill the objectives c.
An action a ∈ A consists in either in- or decrementing one parameter xi by
1×, 10× or 100× a user-specified reference value δi with δ = (δ1 , . . . , δnx )> .
This empirically defined quantization of the intrinsically continuous action
space yielded good results for the problems considered in this work.
Transition function: T encodes the agents knowledge about the computational model f and is learned automatically as described in Sec. 2.5.
Rewards: Inspired by the “mountain car” benchmark (Sutton and Barto,
1998), the rewards are defined as always being equal to R(s, a, s0 ) = −1
(punishment), except when the agent performs an action resulting in personalization success, i.e. when s0 = ŝ. In that case, R(·, ·, ŝ) = 0 (no punishment).
Discount factor: The large discount factor γ = 0.99 encourages policies
that favor future over immediate rewards, as Vito should always prefer the
long-term goal of successful personalization to short-term appealing actions
in order to reduce the risk of local minima.
2.3. Learning Model Behavior through Model Exploration
Like a human operator, Vito first learns how the model “behaves” by
experimenting with it. This is done through a “self-guided sensitivity analysis”. A batch of sample transitions is collected through model exploration
episodes E p = {ep1 , ep2 , . . . }. An episode epi is a sequence of ne-steps consecutive transitions generated from the model f and the patient p for whom
the target measurements zp are known. An episode is initiated at time
t = 0 by generating random initial model parameters xt within the physiologically plausible domain Ω. From the outputs of a forward model run
yt = f (xt ), the misfits to the patient’s corresponding measurements are
computed, yielding the objectives vector ct = d(yt , zp ). Next, a random exploration policy πrand that selects an action according to a discrete uniform
probability distribution over the set of actions is employed. The obtained
at ∈ A is then applied to the current parameter vector, yielding modified
parameter values xt+1 = at (xt ). From the output of the forward model run
yt+1 = f (xt+1 ) the next objectives ct+1 are computed. The next action at+1
is then selected according to πrand , and this process is repeated ne-steps − 1
times. Hence, each episode can be seen as a set of consecutive tuples:
e = {(xt , yt , ct , at , xt+1 , yt+1 , ct+1 ),
t = 0, . . . , ne-steps − 1} .
(6)
In this work, ne-steps = 100 transitions are created in each episode as a tradeoff between sufficient length of an episode to cover a real personalization
scenario and sufficient exploration of the parameter space.
8
The model is explored with many different training patients and the
S resulting episodes are combined into one large training episode set E = p E p .
The underlying hypothesis (verified in experiments) is that the combined
E allows to cancel out peculiarities of individual patients, i.e. to abstract
from patient-specific to model-specific knowledge.
2.4. From Computed Objectives to Representative MDP State
As mentioned above, the continuous space of objective vectors is quantized into a finite set of representative MDP states S. A data-driven approach is proposed. First, all objective vectors observed during training are
clustered according to their distance to each other. Because the ranges of
possible values for the individual objectives can vary significantly depending on the selected measurements (due to different types of measurements,
different units, etc.), the objectives should be normalized during clustering to avoid bias towards objectives with relatively large typical values. In
this work the distance measure performs implicit normalization to account
for these differences: the distance between two objective vectors (c1 , c2 ) is
defined relative to the inverse of the convergence criteria ψ:
q
kc1 − c2 kψ = (c1 − c2 )> diag(ψ)−1 (c1 − c2 ) ,
(7)
where diag(ψ)−1 denotes a diagonal matrix with ( ψ11 , ψ12 , . . . ) along its diagonal. The centroid of a cluster is the centroid of a representative state.
In addition, a special “success state” ŝ representing personalization success
is created, which covers the region in state-space where all objectives are
met: ∀i : |ci | < ψi . The full algorithm is described in Appendix A. Finally, an operator φ : Rnc → S that maps continuous objective vectors c to
representative MDP states is introduced:
φ(c) = arg min kc − ξs kψ
(8)
s∈S
where ξs denotes the centroid corresponding to state s. For an example
state-space quantization see Fig. 3.
2.5. Transition Function as Probabilistic Model Representation
In this work, the stochastic MDP transition function T encodes the
agent’s knowledge about the computational model f . It is learned from the
training data E. First, the individual samples (xt , yt , ct , at , xt+1 , yt+1 , ct+1 )
are converted to state-action-state transition tuples Ê = {(s, a, s0 )}, where
s = φ(ct ), a = at and s0 = φ(ct+1 ). Then, T is approximated from statistics
over the observed transition samples:
T (s, a, s0 ) = P
{(s, a, s0 ) ∈ Ê}
,
s00 ∈S
9
{(s, a, s00 ) ∈ Ê}
(9)
Data-driven quantization
Manual quantization
150
150
100
100
50
50
0
0
-50
-50
-100
-100
-150
-150
-150
-100
-50
0
50
100
150
-150
-100
-50
0
50
100
150
Figure 3: State-space quantization. Left: Example data-driven quantization of a twodimensional state-space into nS = 120 representative states. The states are distributed
according to the observed objective vectors c in one of the experiments in Sec. 3.2. The
objectives were QRS duration [ms] (c1 ) and electrical axis [deg] (c2 ). The center rectangle
(green region) denotes the success state ŝ where all objectives are met (∀i : |ci | < ψi ); see
text for details. Right: Manual quantization as used in Neumann et al. (2015).
where |{·}| denotes the cardinality of the set {·}. If nS and nA are large
compared to the total number of samples it may occur that some stateaction combinations are not observed: |{(s, a, ·) ∈ Ê}| = 0. In that case
uniformity is assumed: ∀s00 ∈ S : T (s, a, s00 ) = 1/nS .
M is now fully defined. Value iteration (Sec. 2.1) is applied and the
stochastic policy π̃∗ is computed, which completes the off-line phase.
2.6. On-line Model Personalization
On-line personalization, as illustrated in Fig. 4, can be seen as a two-step
procedure. First, Vito initializes the personalization of unseen patients from
training data. Second, Vito relies on the computed policy π̃∗ to guide the
personalization process.
2.6.1. Data-driven Initialization
Good initialization can be decisive for a successful personalization. Vito’s
strategy is to search for forward model runs in the training database E for
which the model state f (x) = y ≈ zp is similar to the patient’s measurements. To this end, Vito examines all parameters Ξ = {x ∈ E | f (x) ≈ zp }
that yielded model states similar to the patient’s measurements. Due to
ambiguities induced by the different training patients, data noise and model
assumptions, Ξ could contain significantly dissimilar parameters. Hence,
picking a single x ∈ Ξ might not yield the best initialization. Analyzing Ξ
probabilistically instead helps Vito to find likely initialization candidates.
The details of the initialization procedure are described in Appendix B.
10
Unseen
patient
Personalized
parameters
Initialization
Run forward model
Update parameters
Check convergence
Observe state
Select action
and detect oscillations
Figure 4: Vito’s probabilistic on-line personalization phase. See text for details.
Given the patient’s measurements zp , the procedure outputs a list of initialization candidates X0 = (x00 , x000 , . . . ). The list is sorted by likelihood with
the first element, x00 , being the most likely one.
2.6.2. Probabilistic Personalization
The first personalization step initializes the model parameter vector x0
with the most likely among all initialization candidates, x0 ∈ X0 (see previous section for details). Then, as illustrated in Fig. 4, Vito computes the
forward model y0 = f (x0 ) and the misfit between the model output and the
patient’s measurements c0 = d(y0 , zp ) to derive the first state s0 = φ(c0 ).
Given s0 , Vito decides from its policy the first action to take a0 = π̃∗ (s0 ),
and walks through state-action-state sequences to personalize the computational model f by iteratively updating the model parameters through MDP
actions. Bad initialization could lead to oscillations between states as observed in previous RL works (Kveton and Theocharous, 2012; Neumann
et al., 2015). Therefore, upon detection of an oscillation, which is done by
monitoring the parameter traces to detect recurring sets of parameter values,
the personalization is re-initialized at the second-most-likely x0 ∈ X0 , etc. If
all |X0 | initialization candidates have been tested, a potential re-initialization
defaults to fully random within the physiologically plausible parameter domain Ω. The process terminates once Vito reaches state ŝ (success), or when
a pre-defined maximum number of iterations is reached (failure).
3. Experiments
Vito was applied to a synthetic parameter estimation problem and to two
challenging problems involving real clinical data: personalization of a cardiac electrophysiology (EP), and a whole-body-circulation (WBC) model.
All experiments were conducted using leave-one-out cross-validation. The
numbers of datasets and transition samples used for the different experiments are denoted ndatasets and nsamples , respectively.
3.1. Synthetic Experiment: the Rosenbrock Function
First, Vito was employed in a synthetic scenario, where the ground-truth
model parameters were known. The goals were to test the ability of Vito to
11
5
≥5
0.25
4.5
4
2
3.5
0.25
3
1
3
0
2.5
2
0.25
-1
-2
1.5
1
-4
0.5
-5
-5 -4 -3 -2 -1
0.25
-3
0
1
2
3
4
5 -5 -4 -3 -2 -1
0
1
2
3
4
5 -5 -4 -3 -2 -1
0
1
2
3
4
5
Maximum observerd error
4
0
Figure 5: Synthetic experiment. Left: Contour plot of the Rosenbrock function f α=1 with
global minimum at x = (1, 1)> (red dot). The color scale is logarithmic for visualization
purposes: the darker, the lower the function value. Mid: Maximum L2 -error in parameter
space after personalization over all functions for varying initial parameter values. See text
for details. Yellow represents errors ≥ 5 (maximum observed error ≈ 110). Right: Same
as mid panel, except the extended action set was used. The red dots are the 100 groundtruth parameters x = (α, α2 )> generated for random α.
optimize cost functions generically, and to directly evaluate the performance
in the parameter space.
3.1.1. Forward Model Description
The Rosenbrock function (Rosenbrock, 1960), see Fig. 5, left panel, is
a non-convex function that is often used to benchmark optimization algorithms. It was treated as the forward model in this experiment:
f α (x1 , x2 ) = (α − x1 )2 + 100 · (x2 − x21 )2 ,
(10)
where x = (x1 , x2 )> were the model parameters to estimate for any α, and
f α : Ω → R. As described in Sec. 2.2.2, each of Vito’s actions a ∈ A in- or
decrements a parameter value by multiples (1×, 10×, 100×) of parameterspecific reference values. The reference values were set to δ = (0.01, 0.01)> ,
determined as 0.1% of the defined admissible parameter space per dimension,
Ω = [−5; 5]2 . The parameter α ∈ R defines a family of functions {f α }. The
goal was to find generically arg minx1 ,x2 f α (x1 , x2 ).
The Rosenbrock function has a unique global minimum at x = (α, α2 )> ,
where both terms T1 = (α − x1 ) and T2 = (x2 − x21 ) evaluate to 0. The
personalization objectives were therefore defined as c = (|T1 − 0|, |T2 − 0|)> ,
with the measured data z = (0, 0)> were zero for both objectives and the
computed data y = (T1 , T2 )> . The convergence criteria were set empirically
to ψ = (0.05, 0.05)> .
3.1.2. Evaluation
Vito was evaluated on ndatasets = 100 functions f α with randomly generated α ∈ [−2, 2]. In the off-line phase, for each function, nsamples =
12
10 · ne-steps = 1000 samples, i.e. ten training episodes, each consisting in
ne-steps = 100 transitions (Sec. 2.3), were generated to learn the policy. The
number of representative states was set to nS = 100. To focus on Vito’s
on-line personalization capabilities, both the data-driven initialization and
the re-initialization on oscillation (Sec. 2.6) were disabled. In total, 441 experiments with different initializations (sampled on a 21 × 21 uniform grid
spanned in Ω) were conducted. For each experiment all 100 functions were
personalized using leave-one-family-function-out cross validation, and the
error value from the function exhibiting the maximum L2 -error (worst-case
scenario) between ground-truth (α, α2 ) and estimated parameters was plotted. As one can see from the large blue region in Fig. 5, mid panel, for the
majority of initial parameter values Vito always converged to the solution
(maximum L2 -error < 0.25; the maximum achievable accuracy depended on
the specified convergence criteria ψ and on the reference values δ, which
“discretized” the parameter space). However, especially for initializations
far from the ground-truth (near border regions of Ω), Vito was unable to personalize some functions properly, which was likely due to the high similarity
of the Rosenbrock function shape in these regions.
To investigate this issue, the experiment was repeated after additional
larger parameter steps were added to the set of available actions: A0 =
A ∪ {±500δ1 ; ±500δ2 }. As shown in Fig. 5, right panel, Vito could now
personalize successfully starting from any point in Ω. The single spot with
larger maximum error (bright spot at approximately x = (−1, 2)> ) can be
explained by Vito’s stochastic behavior: Vito may have become unlucky if
it selected many unfavorable actions in sequence due to the randomness
introduced by the stochastic policy. Enabling re-initialization on oscillation
solved this issue entirely. In conclusion, this experiment showed that Vito
can learn how to minimize a cost function generically.
3.2. Personalization of Cardiac Electrophysiology Model
Vito was then tested in a scenario involving a complex model of cardiac
electrophysiology coupled with 12-lead ECG. Personalization was performed
for real patients from actual clinical data. A total of ndatasets = 83 patients were available for experimentation. For each patient, the end-diastolic
bi-ventricular anatomy was segmented from short-axis cine magnetic resonance imaging (MRI) stacks as described in Zheng et al. (2008). A tetrahedral anatomical model including myofibers was estimated and a torso atlas
affinely registered to the patient based on MRI scout images. See Zettinig
et al. (2014) for more details.
3.2.1. Forward Model Description
The depolarization time at each node of the tetrahedral anatomical
model was computed using a shortest-path graph-based algorithm, similar
13
to the one proposed in Wallman et al. (2012). Tissue anisotropy was modeled by modifying the edge costs to take into account fiber orientation. A
time-varying voltage map was then derived according to the depolarization
time: at a given time t, mesh nodes whose depolarization time was higher
than t were assigned a trans-membrane potential of −70 mV, 30 mV otherwise. The time-varying potentials were then propagated to a torso model
where 12-lead ECG acquisition was simulated, and QRS duration (QRSd)
and electrical axis (EA) were derived (Zettinig et al., 2014). The model
was controlled by the conduction velocities (in m/s) of myocardial tissue
and left and right Purkinje network: x = (vMyo , vLV , vRV )> . The latter two
domains were modeled as fast endocardial conducting tissue. The admissible parameter space Ω was set to [200; 1000] for vMyo and [500; 5000] for
both vLV and vRV . Reference increment values to build the action set A
were set to δ = (5, 5, 5)> m/s for the three model parameters. The goal of
EP personalization was to estimate x from the measured QRSd and EA.
Accounting for uncertainty in the measurements and errors in the model,
a patient was considered personalized if QRSd and EA misfits were below
ψ = (5 ms, 10◦ )> , respectively.
3.2.2. Number of Representative States
In contrast to Neumann et al. (2015), where state-space quantization
required manual tuning of various threshold values, the proposed approach
relies on a single hyper-parameter only: nS , the number of representative
states (Sec. 2.4). To specify nS , eight patients were selected for scouting.
Exhaustive search was performed for nS ∈ {10, 20, . . . , 490, 500} representative states. The goodness of a given configuration was evaluated based
on the success rate (relative number of successfully personalized cases according to convergence criteria ψ) over five independent, consecutive, leaveone-patient-out cross-validated personalization runs of the eight patients.
Furthermore, the average number of required forward model runs was considered. To this end, 100 training episodes (100·ne-steps = 104 transition samples) per patient were generated for each personalization run as described in
Sec. 2.3. As one can see from Fig. 6, good performance was achieved from 50
to 300 representative states. The large range of well performing nS indicates
a certain level of robustness with respect to that hyper-parameter. A slight
performance peak at 120 representative states was observed. Therefore,
nS = 120 was selected for further experimentation as compromise between
maintaining a low number of states and sufficient state granularity. An example quantization with nS = 120 is visualized in Fig. 3. The eight scouting
datasets were discarded for the following experiments to avoid bias in the
analysis.
14
0.8
80
0.6
60
0.4
40
0.2
20
0
10
50
100
150
200
250
300
350
Number of representative states
400
450
0
500
Average # forward model runs
100
120
Success rate
1
Figure 6: Hyper-parameter scouting. Vito’s performance for varying number of representative states nS on eight scouting datasets. The solid and dashed curves represent success
rate and average number of forward runs until convergence, respectively, aggregated over
five personalization runs with varying training data.
3.2.3. Reference Methods
Vito’s results were compared to two standard personalization methods
based on BOBYQA (Powell, 2009), a widely-used gradient-free optimizer
known for its robust performance and fast convergence. The first approach,
“BOBYQA simple”, mimicked the most basic estimation setup, where only
the minimum level of model and problem knowledge were assumed.
Pnc The
objective function was the sum of absolute QRSd and EA errors: i=1 |ci |.
It was minimized in a single optimizer run where all three parameters in
x were tuned simultaneously. The algorithm terminated once all convergence criteria ψ were satisfied (success) or if the number of forward model
evaluations exceeded 100 (failure). The second approach, “BOBYQA cascade”, implemented an advanced estimator with strong focus on robustness,
which computed the optimum parameters in a multi-step iterative fashion.
It is based on Seegerer et al. (2015), where tedious manual algorithm and
cost function tuning was performed on a subset of the data used in this
manuscript. In a first step, the myocardial conduction velocity was tuned to
yield good match between computed and measured QRS duration. Second,
left and right endocardial conduction velocities were optimized to minimize
electrical axis error. Both steps were repeated until all parameter estimates
were stable.
To remove bias towards the choice of initial parameter values, for each
of the two methods all datasets were personalized 100 times with different
random initializations within the range of physiologically plausible values
Ω. The differences in performance were striking: only by changing the
initialization, the number of successfully personalized cases varied from 13 to
37 for BOBYQA simple, and from 31 to 51 for BOBYQA cascade (variability
of more than 25% of the total number of patients). These results highlight
the non-convexity of the cost function to minimize.
15
0
135
Data-driven
initialization
90
Full personalization
Absolute EA error [deg]
50
Full personalization
Fixed initialization
Absolute QRSd error [ms]
100
Data-driven
initialization
Fixed initialization
180
150
45
0
Training samples per dataset
Training samples per dataset
Figure 7: Absolute errors over all patients after initialization with fixed parameter values
(blue), after data-driven initialization for increasing amount of training data (white), and
after full personalization with Vito (green). Data-driven initialization yielded significantly
reduced errors if sufficient training data were available (> 102 ) compared to initialization
with fixed values. Full personalization further reduced the errors by a significant margin.
The red bar and the box edges indicate the median absolute error, and the 25 and 75
percentiles, respectively. Left: QRS duration errors. Right: Electrical axis errors.
3.2.4. Full Personalization Performance
First, Vito’s overall performance was evaluated. The full personalization
pipeline consisting in off-line learning, initialization, and on-line personalization was run on all patients with leave-one-patient-out cross-validation
using 1000 training episodes (nsamples = 1000 · ne-steps = 105 transition samples) per patient. The maximum number of iterations was set to 100. The
green box plots in the two panels of Fig. 7 summarize the results. The
mean absolute errors were 4.1 ± 5.6 ms and 12.4 ± 13.3◦ in terms of QRSd
and EA, respectively, a significant improvement over the residual error after
initialization. In comparison to the reference methods, the best BOBYQA
simple run yielded absolute errors of 4.4 ± 10.8 ms QRSd and 15.5 ± 18.6◦
EA on average, and the best BOBYQA cascade run 0.1 ± 0.2 ms QRSd and
11.2 ± 15.8◦ EA, respectively. Thus, in terms of EA error all three methods
yielded comparable performance, and while BOBYQA simple and Vito performed similarly in terms of QRSd, BOBYQA cascade outperformed both
in this regard. However, considering success rates, i.e. successfully personalized patients according to the defined convergence criteria (ψ) divided by
total number of patients, both the performance of Vito (67%) and BOBYQA
cascade (68%) were equivalent, while BOBYQA simple reached only 49% or
less. In terms of run-time, i.e. average number of forward model runs until convergence, Vito (31.8) almost reached the high efficiency of BOBYQA
simple (best: 20.1 iterations) and clearly outperformed BOBYQA cascade
(best: 86.6 iterations), which means Vito was ≈ 2.5× faster.
16
3.2.5. Residual Error after Initialization
A major advantage over standard methods such as the two BOBYQA approaches is Vito’s automated, data-driven initialization method (Sec. 2.6.1),
which eliminates the need for user-provided initial parameter values. To
evaluate the utility of this step, personalization using Vito was stopped directly after initialization (the most likely x0 was used) and the errors in terms
of QRSd and EA resulting from a forward model run f with the computed
initial parameter values were quantified. This experiment was repeated for
increasing number of transition samples per dataset: nsamples = 100 . . . 105 ,
and the results were compared to the error after initialization when fixed
initial values were used (the initialization of the best performing BOBYQA
experiment was used). As one can see from Fig. 7, with increasing amount
of training data both errors decreased notably. As few as 102 transitions per
dataset already provided more accurate initialization than the best tested
fixed initial values. Thus, not only does this procedure simplify the setup
of Vito for new problems (no user-defined initialization needed), this experiment showed that it can reduce initial errors by a large margin, even
when only few training transitions were available. It should be noted that
Vito further improves the model fit in its normal operating mode (continue
personalization after initialization), as shown in the previous experiment.
3.2.6. Convergence Analysis
An important question in any RL application relates to the amount of
training needed until convergence of the artificial agent’s behavior. For Vito
in particular, this translates to the amount of transition samples required
to accurately estimate the MDP transition function T to compute a solid
policy on the one hand, and to have enough training data for reliable parameter initialization on the other hand. To this end, Vito’s overall performance
(off-line learning, initialization, personalization) was evaluated for varying
number of training transition samples per dataset. As one can see from the
results in Fig. 8, with increasing amount of training data the performance
increased, suggesting that the learning process was working properly. Even
with relatively limited training data of only nsamples = 102 samples per patient, Vito outperformed the best version of BOBYQA simple (49% success
rate). Starting from nsamples ≈ 3000, a plateau at ≈66% success rate was
reached, which remained approximately constant until the maximum tested
number of samples. This was almost on par with the top BOBYQA cascade
performance (68% success rate). Also the run-time performance increased
with more training data. For instance, Vito’s average number of iterations
was 36.2 at 103 samples, 31.5 at 104 samples, or 31.8 at 105 samples.
These results suggested that not only Vito can achieve similar performance as an advanced, manually engineered method, but also the number
of required training samples was not excessive. In fact, a rather limited
and thus well manageable amount of data, which can be computed in a
17
60
0.4
40
# Forward model runs
0.2
0
100
101
102
103
Training samples per dataset
104
20
0
0.6
0.4
0.2
0
BOBYQA
simple
BOBYQA
cascade
100
80
60
40
20
0
Average # forward model runs
0.6
# Forward model runs
0.8
Success rate
80
Reference Methods
# Forward model runs
Success rate
1
Success rate
Success rate
0.8
100
Average # forward model runs
Success rate
Vito
1
Figure 8: EP personalization results. Personalization success rate in blue and average
number of iterations in red. Left: Vito’s performance for increasing number of training
transition samples per dataset. Each dot represents results from one experiment (crossvalidated personalization of all 75 datasets), solid/dashed line is low-pass filtered mean,
shaded areas represent 0.5× and 1× standard deviation. Right: Performance of both
reference methods. Each shade represents 10% of the results, sorted by performance.
reasonable time-frame, sufficed.
3.3. Personalization of Whole-Body Circulation Model
Next, Vito was asked to personalize a lumped-parameter whole-body
circulation (WBC) model from pressure catheterization and volume data.
A subset of ndatasets = 56 patients from the EP experiments were used for
experimentation. The discrepancy was due to missing catheterization data
for some patients, which was required for WBC personalization only. For
each patient, the bi-ventricular anatomy was segmented and tracked from
short-axis cine MRI stacks throughout one full heart cycle using shapeconstraints, learned motion models and diffeomorphic registration (Wang
et al., 2013). From the time-varying endocardial meshes, ventricular volume
curves were derived. Manual editing was performed whenever necessary.
3.3.1. Forward Model Description
The WBC model to personalize was based on Itu et al. (2014). It contained a heart model (left ventricle (LV) and atrium, right ventricle and
atrium, valves), the systemic circulation (arteries, capillaries, veins) and the
pulmonary circulation (arteries, capillaries, veins). Time-varying elastance
models were used for all four chambers of the heart. The valves were modeled
through a resistance and an inertance. A three-element Windkessel model
was used for the systemic and pulmonary arterial circulation, while a twoelement Windkessel model was used for the systemic and pulmonary venous
circulation. We refer the reader to Itu et al. (2014); Neumann et al. (2015);
Westerhof et al. (1971) for more details. Personalization was performed with
respect to the patient’s heart rate as measured during catheterization.
The goal of this experiment was to compare Vito’s personalization performance for the systemic part of the model in setups with increasing number of parameters to tune and objectives to match. To this end, Vito was
18
x
Initial volume
LV max. elastance
Aortic resistance
Aortic compliance
Dead volume
Time to Emax
Default value
400 mL
2.4 mmHg/mL
1100 g/(cm4 s)
1.4 ·109 cm4 s2 /g
10 mL
300 ms
Ω
[200; 1000] mL
[0.2; 5] mmHg/mL
[500; 2500] g/(cm4 s)
[0.5; 6] ·109 cm4 s2 /g
[−50; 500] mL
[100; 600] ms
Setups
6, 5, 3, 2
6, 5, 3, 2
6, 5, 3
6, 5
6, 5
6
Table 1: WBC parameters x, their default values and domain Ω. The last column denotes the experiment setups in which a parameter was personalized (e.g. “5”: parameter
was among the estimated parameters in 5p experiment). Default values were used in
experiments where the respective parameters were not personalized.
c
End-diastolic LV volume
End-systolic LV volume
Mean aortic pressure
Peak-systolic aortic pressure
End-diastolic aortic pressure
Ejection time
ψ
20 mL
20 mL
10 mmHg
10 mmHg
10 mmHg
50 ms
Measured range
[129; 647] mL
[63; 529] mL
[68; 121] mmHg
[83; 182] mmHg
[48; 99] mmHg
[115; 514] ms
Setups
6, 5, 3, 2
6, 5, 3, 2
6, 5, 3
6, 5
6, 5
6
Table 2: WBC objectives c, their convergence criteria ψ and range of measured values in
the patient population used for experimentation.
employed on setups with two to six parameters (2p, 3p, 5p, 6p): initial
blood volume, LV maximum elastance, time until maximum elastance is
reached, total aortic resistance and compliance, and LV dead volume. The
reference values δ to define Vito’s allowed actions A were set to .5% of the
admissible parameter range Ω for each individual parameter, see Table 1 for
details. The personalization objectives were MRI-derived end-diastolic and
end-systolic LV volume, ejection time (time duration during which the aortic valve is open and blood is ejected), and peak-systolic, end-diastolic, and
mean aortic blood pressures as measured during cardiac catheterization, see
Fig. 9. To account for measurement noise, personalization was considered
successful if the misfits per objective were below acceptable threshold values
ψ as listed in Table 2.
3.3.2. Number of Representative States
Along the same lines as Sec. 3.2.2, the hyper-parameter for state-space
quantization was tuned based on the eight scouting patients. The larger the
dimensionality of the state-space, the more representative states were needed
to yield good performance. In particular, for the different WBC setups, the
numbers of representative states (nS ) yielding the best scouting performance
were 70, 150, 400 and 600 for the 2p, 3p, 5p and 6p setup, respectively. The
scouting datasets were discarded for the following experiments.
19
LV volume [mL]
Aortic pressure [mmHg]
160
140
WBC 2p
Measured
Computed
160
WBC 3p
160
140
140
120
120
120
100
100
100
80
80
60
60
60
40
40
40
20
200
20
200
20
200
185
End-diastolic
volume
185
Mean
pressure
80
End-diastolic
volume
185
WBC 5p
Peak-systolic
pressure
160
140
100
Mean
pressure
80
End-diastolic
volume
40
20
200
End-diastolic
volume
Ejection
time
170
155
155
155
155
140
140
140
110
0 0.2 0.4 0.6 0.8 1
Time [s]
125
110
125
110
0 0.2 0.4 0.6 0.8 1
Time [s]
Endsystolic
volume
0 0.2 0.4 0.6 0.8 1
Time [s]
End-diastolic
aortic pressure
185
170
125
Mean
pressure
60
End-diastolic
aortic pressure
170
Endsystolic
volume
Peak-systolic
pressure
120
170
Endsystolic
volume
WBC 6p
140
125
110
Endsystolic
volume
0 0.2 0.4 0.6 0.8 1
Time [s]
Figure 9: Goodness of fit in terms of time-varying LV volume and aortic pressure for Vito
personalizing an example patient based on the different WBC setups. The added objectives
per setup are highlighted in the respective column. With increasing number of parameters
and objectives Vito manages to improve the fit between model and measurements.
3.3.3. Reference Method
A gradient-free optimizer (Lagarias et al., 1998) based on the simplex
method was used to benchmark Vito. The objective function was the sum
of squared differences between computed and measured values, weighted by
the inverse of the convergence criteria to counter the different ranges of
objective values (e.g. due to different types of measurements and different
units): kckψ (Eq. 7). Compared to non-normalized optimization, the algorithm converged up to 20% faster and success rates increased by up to
8% under otherwise identical conditions. Personalization was terminated
once all convergence criteria were satisfied (success), or when the maximum
number of iterations was reached (failure). To account for the increasing
complexity of optimization with increasing number of parameters nx , the
maximum number of iterations was set to 50 · nx for the different setups.
As one can see from Fig. 10, right panels, with increasing number of
parameters to be estimated, the performance in terms of success rate and
number of forward model runs decreased slightly. This is expected as the
problem becomes harder. To suppress bias originating from (potentially
poor) initialization, the reference method was run 100 times per setup (as
in EP experiments), each time with a different, randomly generated set of
initial parameter values. The individual performances varied significantly
for all setups.
20
Vito
Success rate
1
1
0.8
0.8
0.6
0.6
0.4
0.4
2p
3p
5p
6p
Average # forward model runs
0.2
0
0
10
300
1
10
2
10
3
0.2
0
4
10
10
2p
3p
5p
6p
300
2p
3p
5p
6p
250
200
250
200
150
150
100
100
50
50
0
100
Reference Method
101
102
103
Training samples per dataset
104
0
2p
3p
5p
#Parameters
6p
Figure 10: WBC model personalization results (top: success rate, bottom: average number
of forward model runs until convergence) for various estimation setups (different colors),
see text for details. Left: Vito’s performance for increasing number of training transition
samples per dataset. Each dot represents results from one experiment (cross-validated
personalization of all 48 datasets), solid/dashed lines are low-pass filtered mean, shaded
areas represent 0.5× and 1× standard deviation. Right: Performance of reference method.
Each shade represents 10% of the results, sorted by performance; darkest shade: best 10%.
21
3.3.4. Convergence Analysis
For each WBC setup the full Vito personalization pipeline was evaluated
for increasing training data (nsamples = 100 . . . 105 ) using leave-one-patientout cross-validation. The same iteration limits as for the reference method
were used. The results are presented in Fig. 10, left panels. With increasing
data, Vito’s performance, both in terms of success rate and run-time (iterations until convergence), increased steadily until reaching a plateau. As
one would expect, the more complex the problem, i.e. the more parameters
and objectives involved in the personalization, the more training data was
needed to reach the same level of performance. For instance, Vito reached
80% success rate with less than nsamples = 50 training samples per dataset in
the 2p setup, whereas almost 90× as many samples were required to achieve
the same performance in the 6p setup.
Compared to the reference method, given enough training data, Vito
reached equivalent or better success rates (e.g. up to 11% higher success
rate for 6p) while significantly outperforming the reference method in terms
of run-time performance. In the most basic setup (2p), if nsamples ≥ 103 , Vito
converged after 3.0 iterations on average, while the best reference method
run required 22.6 iterations on average, i.e. Vito was seven times faster. For
the more complex setups (3p, 5p, 6p), the speed-up was not as drastic. Yet,
in all cases Vito outperformed even the best run of the reference method by
a factor of 1.8 or larger.
4. Conclusion
4.1. Summary and Discussion
In this manuscript, a novel personalization approach called Vito has been
presented. To our knowledge, it is the first time that biophysical model
personalization is addressed using artificial intelligence concepts. Inspired
by how humans approach the personalization problem, Vito first learns the
characteristics of the computational model under consideration using a datadriven approach. This knowledge is then utilized to learn how to personalize
the model using reinforcement learning. Vito is generic in the sense that it
requires only minimal and intuitive user input (parameter ranges, authorized
actions, number of representative states) to learn by itself how to personalize
a model.
Vito was applied to a synthetic scenario and to two challenging personalization tasks in cardiac computational modeling. The problem setups
and hyper-parameter configurations are listed in Table 3. In most setups
the majority of hyper-parameters were identical and only few (nS ) required
manual tuning, suggesting good generalization properties of Vito. Another
key result was that Vito was up to 11% more robust (higher success rates)
compared to standard personalization methods. Vito’s ability to generalize
22
Application
Rosenbrock
Rosenbrock ext.
EP
WBC 2p
WBC 3p
WBC 5p
WBC 6p
nx
2
2
3
2
3
5
6
nc
2
2
2
2
3
5
6
ndatasets
100
100
83 (75)
56 (48)
56 (48)
56 (48)
56 (48)
nS
100
100
120
70
150
400
600
nA /nx
6
8
6
6
6
6
6
nplateau
n/a
n/a
3 000
450
2 000
3 500
20 000
Table 3: Applications considered in this manuscript described in terms of the number
of parameters (nx ), objectives (nc ) and datasets (ndatasets ) used for experimentation (in
brackets: excluding scouting patients, if applicable); and Vito’s hyper-parameters: the
number of representative MDP states (nS ) and the number of actions per parameter
(nA /nx ). The last column (nplateau ) denotes the approximate number of samples needed
to reach the performance “plateau” (see convergence analyses in Sec. 3.2.6 and Sec. 3.3.4).
the knowledge obtained from a set of training patients to personalize unseen patients was shown as all experiments reported in this manuscript were
based on cross-validation. Furthermore, Vito’s robustness against training
patients for whom we could not find a solution was tested. In particular, for
about 20% of the patients, in none of the electrophysiology experiments in
Sec. 3.2 any personalization (neither Vito nor the reference methods) could
produce a result that satisfied all convergence criteria. Hence, for some
patients no solution may exist under the given electrophysiology model configuration1 . Still, all patients were used to train Vito, and surprisingly Vito
was able to achieve equivalent success rate as the manually engineered personalization approach for cardiac EP.
Generating training data could be considered Vito’s computational bottleneck. However, training is i) performed off-line and one-time only, and
ii) it is independent for each training episode and each patient. Therefore,
large computing clusters could be employed to perform rapid training by
parallelizing this phase. On-line personalization, on the contrary, is not parallelizable in its current form: the parameters for each forward model run
depend on the outcome of the previous iteration. Since the forward computations are the same for every “standard” personalization method (not
including surrogate-based approaches), the number of forward model runs
until convergence was used for benchmarking: Vito was up to seven times
faster compared to the reference methods. The on-line overhead introduced
by Vito (convert data into an MDP state, then query policy) is negligible.
As such, Vito could become a unified framework for personalization of
any computational physiological model, potentially eliminating the need for
1
Potential solution non-existence may be due to possibly invalid assumptions of the
employed EP model for patients with complex pathologies.
23
an expert operator with in-depth knowledge to design and engineer complex
optimization procedures.
4.2. Challenges and Outlook
Important challenges still remain, such as the incorporation of continuous
actions, the definition of states and their quantization. In this work we propose a data-driven state-space quantization strategy. Contrary to Neumann
et al. (2015), where a threshold-based state-quantization involving several
manually tuned threshold values (Fig. 3) was employed, the new method
is based on a single hyper-parameter only: the number of representative
states. Although it simplifies the setup of Vito, this quantization strategy
may still not be optimal, especially if only little training data is available.
Therefore, advanced approaches for continuous reinforcement learning with
value function approximation (Mnih et al., 2015; Sutton and Barto, 1998)
could be integrated to fully circumvent quantization issues.
At the same time, such methods could improve Vito’s scalability towards
high-dimensional estimation tasks. In this work we showed that Vito can be
applied to typical problems emerging in cardiac modeling, which could be
described as medium-scale problems with moderate number of parameters to
personalize and objectives to match. In unreported experiments involving
>10 parameters, however, Vito could no longer reach satisfactory performance, which is likely due to the steeply increasing number of transition
samples needed to sample the continuous state-space of increasing dimensionality sufficiently during training. The trends in Sec. 3.3 confirm the need
for more data. In the future, experience replay (Adam et al., 2012; Lin, 1993)
or similar techniques could be employed to increase training data efficiency.
Furthermore, massively parallel approaches (Nair et al., 2015) are starting
to emerge, opening up new avenues for large-scale reinforcement learning.
Although the employed reinforcement learning techniques guarantee convergence to an optimal policy, the computed personalization strategy may
not be optimal for the model under consideration as the environment is only
partially observable and the personalization problem ill-posed: there is no
guarantee for solution existence or uniqueness. Yet, we showed that Vito
can solve personalization more robustly and more effectively than standard
methods under the same conditions. However, a theoretical analysis in terms
of convergence guarantees and general stability of the method would be desirable, in particular with regards to the proposed re-initialization strategy.
As a first step towards this goal, in preliminary (unreported) experiments on
the EP and the WBC model we observed that the number of patients which
do not require re-initialization (due to oscillation) to converge to a successful
personalization consistently increased with increasing training data.
The data-driven initialization proposed in this work simplifies Vito’s
setup by eliminating the need for user-provided initialization. However,
currently there is no guarantee that the first initialization candidate is the
24
one that will yield the “best” personalization outcome. Therefore, one could
investigate the benefits of a fuzzy personalization scheme: many personalization processes could be run in parallel starting from the different initialization candidates. Parameter uncertainty quantification techniques (Neumann
et al., 2014a) could then be applied to compute a probability density function over the space of model parameters. Such approaches aim to gather
complete information about the solution-space, which can be used to study
solution uniqueness and other interesting properties.
An important characteristic of any personalization algorithm is its stability against small variations of the measured data. A preliminary experiment indicated good stability of Vito: the computed parameters from
several personalization runs, each involving small random perturbations of
the measurements, were consistent. Yet in a small group of patients some
parameter variability was observed, however, it was below the variability of
the reference method under the same conditions. To what extent certain
degrees of variability will impact other properties of the personalized model
such as its predictive power will be subject of future research. We will also
investigate strategies to improve Vito’s stability further. For instance, the
granularity of the state-space could provide some flexibility to tune the stability: less representative states means a larger region in state space per
state, thus small variations in the measured data might have less impact
on personalization outcome. However, this could in turn have undesirable
effects on other properties of Vito such as success rate or convergence speed
(see Sec. 3.2.2).
Beyond these challenges, Vito showed promising performance and versatility, making it a first step towards an automated, self-taught model personalization agent. The next step will be to investigate the predictive power
of the personalized models, for instance for predicting acute or long-term
response in cardiac resynchronization therapy (Kayvanpour et al., 2015; Sermesant et al., 2009).
Appendix A. Data-driven State-Space Quantization
This section describes the details of the proposed data-driven quantization approach to define the set of representative MDP states S (see Sec. 2.4).
It is based on clustering, in particular on the weighted k -means algorithm
described in Arthur and Vassilvitskii (2007). To this end, all objective vectors C = {c ∈ E} are extracted from the training data (Sec. 2.3). C ⊂ Rnc
represents all observed “continuous states”. The goal is to convert C into
the finite set of representative MDP states S while taking into account that
Vito relies on a special “success state” ŝ encoding personalization success.
The success state ŝ does not depend on the data, but on the maximum
acceptable misfit ψ. In particular, since personalization success implies that
all objectives are met, ŝ should approximate a hyperrectangle centered at
25
Figure A.11: Preprocessing of k -means input data to enforce the success state ŝ. Left:
Continuous state-space with observed objective vectors c (blue points). The points with
dashed outline will be canceled out. Right: Delineation of ŝ in green, enforced by inserted
vectors (green / red points) with large weights. See text for details.
0 and bounded at ±ψ, i.e. a small region in Rnc where ∀i : |ci | < ψi . To
enforce ŝ, the input to weighted k -means is preprocessed as follows.
First, the 0-vector is inserted into C, along with two vectors per dimension i, where all components are zero, except the ith component, which is set
to ±2ψi . These 2nc + 1 inserted vectors are later converted into centroids
of representative states to delineate the desired hyperrectangle for ŝ as illustrated in Fig. A.11. Furthermore, to avoid malformation of ŝ, no other
representative state should emerge within that region. Therefore, all vectors
c ∈ C, where ∀i : |ci | < 2ψi (except for the inserted vectors) are canceled
out by assigning zero weight, while the inserted vectors are assigned large
weights → ∞ and all remaining vectors weights of 1.
Next, k -means is initialized by placing a subset of the initial centroids at
the locations of the inserted states, and the remaining nS − 2nc − 1 centroids
at random vectors in C. Both the large weight and the custom initialization
enforce the algorithm to converge to a solution where one cluster centroid
is located at each inserted vector, while the other centroids are distributed
according to the training data. To ensure equal contribution of all objectives
(cancel out different units, etc.), similarity is defined relative to the inverse
of the user-defined convergence criteria (Eq. 7).
Finally, after k -means converged, the resulting centroids, denoted ξs , are
used to delineate the region in Rnc assigned to a representative state s.
Appendix B. Data-driven Initialization
This section describes the details of the proposed data-driven initialization approach to compute a list of candidate initialization parameter vectors
X0 = (x00 , x000 , . . . ) for a new patient p based on the patient’s measurements
zp and the training database E (see Sec. 2.6.1).
26
First, all model states are extracted from the training database: Υ =
{y ∈ E}. Next, Υ is fed to a clustering algorithm (e.g. k -means). As in
Appendix A, the distance measure is defined relative to the inverse of the
convergence criteria (Eq. 7). The output is a set of centroids (for simplicity,
in this work the number of centroids was set to nS ), and each vector is
assigned to one cluster based on its closest centroid. Let Υp ⊆ Υ denote the
members of the cluster whose centroid is closest to zp and Ξp = {x ∈ E |
f (x) ∈ Υp } the set of corresponding model parameters. For each cluster, an
approximation of the likelihood over the generating parameters is computed
in terms of a probability density function. In this work a Gaussian mixture
model is assumed:
GMMp (x) =
M
X
νm N (x; µm , Σm ) .
(B.1)
m=1
The parameter vectors in Ξp are treated as random samples drawn from
GMMp . Its properties, namely the number of mixture components M , their
weights νm , and their means µm and covariance matrices Σm , are estimated
from these samples using a multivariate kernel density estimator with automated kernel bandwidth estimation, see Kristan et al. (2011) for more
details. Finally, the M estimated means are selected as initialization candidates and stored in a list X0 = (µm0 , µm00 , . . . ). The elements of X0 are
sorted in descending order according to their corresponding νm -values to prioritize more likely initializations: µm0 is the mean with m0 = arg maxm νm .
References
References
Adam, S., Busoniu, L., Babuska, R., 2012. Experience replay for real-time
reinforcement learning control. IEEE Sys. Man. Cybern. 42 (2), 201–212.
Aguado-Sierra, J., Kerckhoffs, R. C. P., Lionetti, F., Hunt, D., Villongco,
C., Gonzales, M., Campbell, S. G., McCulloch, A. D., 2010. A computational framework for patient-specific multi-scale cardiac modeling. In:
Kerckhoffs, R. C. (Ed.), Patient-Specific Modeling of the Cardiovascular
System. Springer, pp. 203–223.
Aguado-Sierra, J., Krishnamurthy, A., Villongco, C., Chuang, J., Howard,
E., Gonzales, M. J., Omens, J., Krummen, D. E., Narayan, S., Kerckhoffs,
R. C. P., McCulloch, A. D., 2011. Patient-specific modeling of dyssynchronous heart failure: a case study. Prog. Biophys. Mol. Bio. 107 (1),
147–155.
Arthur, D., Vassilvitskii, S., 2007. k-means++: The advantages of careful
seeding. In: ACM-SIAM Symp. Discr. Algorithm. pp. 1027–1035.
27
Augenstein, K. F., Cowan, B. R., LeGrice, I. J., Nielsen, P. M., Young, A. A.,
2005. Method and apparatus for soft tissue material parameter estimation
using tissue tagged magnetic resonance imaging. J. Biomech. Eng. 127 (1),
148–157.
Barreto, A., Precup, D., Pineau, J., 2014. Practical kernel-based reinforcement learning. arXiv preprint arXiv:1407.5358.
Bellman, R., 1957. Dynamic Programming. Princeton University Press.
Bishop, C. M., 2006. Pattern recognition and machine learning. Vol. 4.
Springer New York.
Chabiniok, R., Moireau, P., Lesault, P.-F., Rahmouni, A., Deux, J.-F.,
Chapelle, D., 2012. Estimation of tissue contractility from cardiac cinemri using a biomechanical heart model. Biomech. Model. Mechan. 11 (5),
609–630.
Clayton, R., Bernus, O., Cherry, E., Dierckx, H., Fenton, F., Mirabella,
L., Panfilov, A., Sachse, F., Seemann, G., Zhang, H., 2011. Models of
cardiac tissue electrophysiology: progress, challenges and open questions.
Prog. Biophys. Mol. Bio. 104 (1), 22–48.
Delingette, H., Billet, F., Wong, K. C. L., Sermesant, M., Rhode, K., Ginks,
M., Rinaldi, C. A., Razavi, R., Ayache, N., 2012. Personalization of cardiac motion and contractility from images using variational data assimilation. IEEE T. Biomed. Eng. 59 (1), 20–24.
Frangi, A. F., Niessen, W. J., Viergever, M., 2001. Three-dimensional modeling for functional analysis of cardiac images, a review. IEEE T. Med. Imaging 20 (1), 2–5.
Hunter, P. J., Borg, T. K., 2003. Integration from proteins to organs: the
physiome project. Nat. Rev. Mol. Cell Bio. 4 (3), 237–243.
Itu, L., Sharma, P., Georgescu, B., Kamen, A., Suciu, C., Comaniciu, D.,
2014. Model based non-invasive estimation of PV loop from echocardiography. IEEE Eng. Med. Biol. Soc., 6774–6777.
Kaelbling, L. P., Littman, M. L., Moore, A. W., 1996. Reinforcement learning: A survey. J. Artif. Intell. Res., 237–285.
Kayvanpour, E., Mansi, T., Sedaghat-Hamedani, F., Amr, A., Neumann, D.,
Georgescu, B., Seegerer, P., Kamen, A., Haas, J., Frese, K. S., Irawati, M.,
Wirsz, E., King, V., Buss, S., Mereles, D., Zitron, E., Keller, A., Katus,
H. A., Comaniciu, D., Meder, B., 2015. Towards personalized cardiology:
Multi-scale modeling of the failing heart. PLoS ONE 10 (7), e0134869.
28
Kerckhoffs, R. C. P., Lumens, J., Vernooy, K., Omens, J., Mulligan, L., Delhaas, T., Arts, T., McCulloch, A., Prinzen, F., 2008. Cardiac resynchronization: insight from experimental and computational models. Prog. Biophys. Mol. Bio. 97 (2), 543–561.
Konukoglu, E., Relan, J., Cilingir, U., Menze, B. H., Chinchapatnam, P.,
Jadidi, A., Cochet, H., Hocini, M., Delingette, H., Jaı̈s, P., Haı̈ssaguerre,
M., Ayache, N., Sermesant, M., 2011. Efficient probabilistic model personalization integrating uncertainty on data and parameters: Application to eikonal-diffusion models in cardiac electrophysiology. Prog. Biophys. Mol. Bio. 107 (1), 134–146.
Krishnamurthy, A., Villongco, C. T., Chuang, J., Frank, L. R., Nigam, V.,
Belezzuoli, E., Stark, P., Krummen, D. E., Narayan, S., Omens, J. H.,
McCulloch, A. D., Kerckhoffs, R. C. P., 2013. Patient-specific models of
cardiac biomechanics. J. Comput. Phys. 244, 4–21.
Kristan, M., Leonardis, A., Skočaj, D., 2011. Multivariate online kernel
density estimation with gaussian kernels. Pattern Recogn. 44 (10), 2630–
2642.
Kuijpers, N. H., Hermeling, E., Bovendeerd, P. H., Delhaas, T., Prinzen,
F. W., 2012. Modeling cardiac electromechanics and mechanoelectrical
coupling in dyssynchronous and failing hearts. J. Cardiovasc. Transl. Res.
5 (2), 159–169.
Kveton, B., Theocharous, G., 2012. Kernel-based reinforcement learning on
representative states. In: Association for the Advancement of Artificial
Intelligence. pp. 977–983.
Lagarias, J. C., Reeds, J. A., Wright, M. H., Wright, P. E., 1998. Convergence properties of the Nelder-Mead simplex method in low dimensions.
SIAM J. Optimiz. 9 (1), 112–147.
Le Folgoc, L., Delingette, H., Criminisi, A., Ayache, N., 2013. Current-based
4D shape analysis for the mechanical personalization of heart models. In:
Medical Computer Vision. Recognition Techniques and Applications in
Medical Imaging. Vol. 7766 of LNCS. Springer, pp. 283–292.
Lin, L.-J., 1993. Reinforcement learning for robots using neural networks.
Tech. rep., DTIC Document.
Marchesseau, S., Delingette, H., Sermesant, M., Cabrera-Lozoya, R., TobonGomez, C., Moireau, P., Figueras i Ventura, R. M., Lekadir, K., Hernandez, A., Garreau, M., Donal, E., Leclercq, C., Duckett, S. G., Rhode,
K., Rinaldi, C. A., Frangi, A. F., Razavi, R., Chapelle, D., Ayache, N.,
2013. Personalization of a cardiac electromechanical model using reduced
29
order unscented kalman filtering from regional volumes. Med. Image Anal.
17 (7), 816–829.
Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare,
M. G., Graves, A., Riedmiller, M., Fidjeland, A. K., Ostrovski, G., Petersen, S., Beattie, C., Sadik, A., Antonoglou, I., King, H., Kumara, D.,
Wierstra, D., Legg, S., Hassabis, D., 2015. Human-level control through
deep reinforcement learning. Nature 518 (7540), 529–533.
Mülling, K., Kober, J., Kroemer, O., Peters, J., 2013. Learning to select and
generalize striking movements in robot table tennis. Int. J. Robot. Res.
32 (3), 263–279.
Nair, A., Srinivasan, P., Blackwell, S., Alcicek, C., Fearon, R., De Maria,
A., Panneershelvam, V., Suleyman, M., Beattie, C., Petersen, S., Legg, S.,
Mnih, V., Kavukcuoglu, K., Silver, D., 2015. Massively parallel methods
for deep reinforcement learning. arXiv:1507.04296.
Neumann, D., Mansi, T., Georgescu, B., Kamen, A., Kayvanpour, E., Amr,
A., Sedaghat-Hamedani, F., Haas, J., Katus, H., Meder, B., Hornegger,
J., Comaniciu, D., 2014a. Robust image-based estimation of cardiac tissue
parameters and their uncertainty from noisy data. In: MICCAI. Vol. 8674
of LNCS. Springer, pp. 9–16.
Neumann, D., Mansi, T., Grbic, S., Voigt, I., Georgescu, B., Kayvanpour,
E., Amr, A., Sedaghat-Hamedani, F., Haas, J., Katus, H., Meder, B.,
Hornegger, J., Kamen, A., Comaniciu, D., 2014b. Automatic image-tomodel framework for patient-specific electromechanical modeling of the
heart. In: IEEE Int. Symp. Biomed. Imaging. pp. 935–938.
Neumann, D., Mansi, T., Itu, L., Georgescu, B., Kayvanpour, E., SedaghatHamedani, F., Haas, J., Katus, H., Meder, B., Steidl, S., Hornegger,
J., Comaniciu, D., 2015. Vito – a generic agent for multi-physics model
personalization: Application to heart modeling. In: MICCAI. Vol. 9350
of LNCS. Springer, pp. 442–449.
Nguyen-Tuong, D., Peters, J., 2011. Model learning for robot control: a
survey. Cogn. Process. 12 (4), 319–340.
Noble, D., 2002. Modeling the heart – from genes to cells to the whole organ.
Science 295 (5560), 1678–1682.
Powell, M. J., 2009. The BOBYQA algorithm for bound constrained optimization without derivatives. Cambridge NA Report NA2009/06.
Prakosa, A., Sermesant, M., Allain, P., Villain, N., Rinaldi, C., Rhode,
K., Razavi, R., Delingette, H., Ayache, N., 2013. Cardiac electrophysio30
logical activation pattern estimation from images using a patient-specific
database of synthetic image sequences. IEEE T. Biomed. Eng.
Rosenbrock, H., 1960. An automatic method for finding the greatest or least
value of a function. Comput. J. 3 (3), 175–184.
Schmid, H., Nash, M., Young, A., Hunter, P., 2006. Myocardial material parameter estimation – a comparative study for simple shear.
J. Biomech. Eng. 128 (5), 742–750.
Seegerer, P., Mansi, T., Jolly, M.-P., Neumann, D., Georgescu, B., Kamen,
A., Kayvanpour, E., Amr, A., Sedaghat-Hamedani, F., Haas, J., Katus, H., Meder, B., Comaniciu, D., 2015. Estimation of regional electrical
properties of the heart from 12-lead ECG and images. In: Statistical Atlases and Computational Models of the Heart – Imaging and Modelling
Challenges. Vol. 8896 of LNCS. Springer, pp. 204–212.
Sermesant, M., Billet, F., Chabiniok, R., Mansi, T., Chinchapatnam, P.,
Moireau, P., Peyrat, J.-M., Rhode, K., Ginks, M., Lambiase, P., Arridge,
S., Delingette, H., Sorine, M., Rinaldi, C. A., Chapelle, D., Razavi, R.,
Ayache, N., 2009. Personalised electromechanical model of the heart for
the prediction of the acute effects of cardiac resynchronisation therapy.
In: Functional Imaging and Modeling of the Heart. Vol. 5528 of LNCS.
Springer, pp. 239–248.
Sutton, R. S., Barto, A. G., 1998. Reinforcement learning: An introduction.
Vol. 1. MIT press Cambridge.
Tesauro, G., 1994. Td-gammon, a self-teaching Backgammon program,
achieves master-level play. Neural Comput. 6 (2), 215–219.
Wallman, M., Smith, N. P., Rodriguez, B., 2012. A comparative study of
graph-based, eikonal, and monodomain simulations for the estimation of
cardiac activation times. IEEE T. Biomed. Eng. 59 (6), 1739–1748.
Wallman, M., Smith, N. P., Rodriguez, B., 2014. Computational methods
to reduce uncertainty in the estimation of cardiac conduction properties
from electroanatomical recordings. Med. Image Anal. 18 (1), 228–240.
Wang, V. Y., Lam, H., Ennis, D. B., Cowan, B. R., Young, A. A., Nash,
M. P., 2009. Modelling passive diastolic mechanics with quantitative mri
of cardiac structure and function. Med. Image. Anal. 13 (5), 773–784.
Wang, Y., Georgescu, B., Chen, T., Wu, W., Wang, P., Lu, X., Ionasec, R.,
Zheng, Y., Comaniciu, D., 2013. Learning-based detection and tracking
in medical imaging: a probabilistic approach. In: Deformation Models.
Vol. 7 of LNCVB. Springer, pp. 209–235.
31
Westerhof, N., Elzinga, G., Sipkema, P., 1971. An artificial arterial system
for pumping hearts. J. Appl. Physiol. 31 (5), 776–781.
Wong, K. C., Sermesant, M., Rhode, K., Ginks, M., Rinaldi, C. A.,
Razavi, R., Delingette, H., Ayache, N., 2015. Velocity-based cardiac contractility personalization from images using derivative-free optimization.
J. Mech. Behav. Biomed. 43, 35–52.
Xi, J., Lamata, P., Niederer, S., Land, S., Shi, W., Zhuang, X., Ourselin,
S., Duckett, S. G., Shetty, A. K., Rinaldi, C. A., Rueckert, D., Razavi,
R., Smith, N. P., 2013. The estimation of patient-specific cardiac diastolic
functions from clinical measurements. Med. Image. Anal. 17 (2), 133–146.
Zettinig, O., Mansi, T., Neumann, D., Georgescu, B., Rapaka, S., Seegerer,
P., Kayvanpour, E., Sedaghat-Hamedani, F., Amr, A., Haas, J., Steen,
H., Katus, H., Meder, B., Navab, N., Kamen, A., Comaniciu, D., 2014.
Data-driven estimation of cardiac electrical diffusivity from 12-lead ECG
signals. Med. Image Anal., 1361–1376.
Zheng, Y., Barbu, A., Georgescu, B., Scheuering, M., Comaniciu, D., 2008.
Four-chamber heart modeling and automatic segmentation for 3-D cardiac
CT volumes using marginal space learning and steerable features. IEEE
T. Med. Imaging 27 (11), 1668–1681.
32
| 5cs.CE
|
1
Model-Free MLE Estimation for Online Rotor
Angle Stability Assessment with PMU Data
arXiv:1702.03629v1 [cs.SY] 13 Feb 2017
Shaopan Wei, Student Member, IEEE, Ming Yang, Member, IEEE, Junjian Qi, Member, IEEE,
Jianhui Wang, Senior Member, IEEE, Shiying Ma, and Xueshan Han
Abstract—Recent research has demonstrated that the rotor angle stability can be assessed by identifying the sign of the system’s
maximal Lyapunov exponent (MLE). A positive (negative) MLE
implies unstable (stable) rotor angle dynamics. However, because
the MLE may fluctuate between positive and negative values for
a long time after a severe disturbance, it is difficult to determine
the system stability when observing a positive or negative MLE
without knowing its further fluctuation trend. In this paper, a new
approach for online rotor angle stability assessment is proposed
to address this problem. The MLE is estimated by a recursive
least square (RLS) based method based on real-time rotor angle
measurements, and two critical parameters, the Theiler window
and the MLE estimation initial time step, are carefully chosen to
make sure the calculated MLE curves present distinct features
for different stability conditions. By using the proposed stability
assessment criteria, the developed approach can provide timely
and reliable assessment of the rotor angle stability. Extensive
tests on the New-England 39-bus system and the Northeast Power
Coordinating Council 140-bus system verify the effectiveness of
the proposed approach.
Index Terms—Lyapunov exponent, model-free, online stability assessment, phasor measurement unit, rotor angle stability,
Theiler window.
I. I NTRODUCTION
T
RANSIENT rotor angle stability refers to the ability
of synchronous generators of an interconnected power
system to remain in synchronism after a severe disturbance [1].
With the development of synchrophasor technologies, utilities
are now able to track rotor angle deviations and take actions
to respond to emergency events. However, since the dynamics
of power systems are complex, online rotor angle stability
assessment is still very challenging [2], [3].
In [4], an adaptive out-of-step relay is proposed for the
Florida-Georgia system. The equal area criterion is applied
to change the settings of the protection system based on
This work was supported by the National Basic Research Program of China
(973 Program) under Grant 2013CB228205, State Grid Corporation of China
under Grant XT71-15-056, and the National Science Foundation of China
under Grant 51007047 and 51477091.
M. Yang (Corresponding Author) is with Key Laboratory of Power System
Intelligent Dispatch and Control, Shandong University, Jinan, Shandong
250061 China. He was a visiting scholar with Argonne National Laboratory,
Argonne, IL 60439 USA (e-mail: [email protected]).
S. Wei and X. Han are with Key Laboratory of Power System Intelligent
Dispatch and Control, Shandong University, Jinan, Shandong 250061 China
(e-mail: spw [email protected]; [email protected]).
J. Qi and J. Wang are with the Energy Systems Division, Argonne
National Laboratory, Argonne, IL 60439 USA (e-mail: [email protected]; [email protected]).
S. Ma is with Institute of Electric Power System, China Electric
Power Research Institute, Haidian District, Beijing 100192 China (e-mail:
[email protected]).
phasor measurement unit (PMU) measurements. In [5], the
dynamics of the power transfer paths are monitored based on
the energy functions of the two-machine equivalent system,
and the PMU data are used to identify the parameters of
the energy functions. In [6], PMU measurements are used
as inputs for estimating the differential/algebraic equation
model to predict the post-fault dynamics. In [7], an online
dynamic security assessment scheme is proposed based on
self-adaptive decision trees, where the PMU data are used for
online identification of the system critical attributes. In [8],
the rotor angle stability is estimated by using artificial neural
networks and the measured voltage and current phasors are
used as inputs of the offline trained estimation model. In [9],
a systematic scheme for building fuzzy rule-based classifiers
for fast stability assessment is proposed. By testing on a large
and highly diversified database, it is demonstrated that the
analysis of post-fault short-term PMU data can extract useful
features satisfying the requirements of stability assessment.
Lyapunov exponents (LEs) those characterize the separation
rate of infinitesimally close trajectories are important indices
for quantifying the stability of dynamical systems. If the
system’s maximal Lyapunov exponent (MLE) is positive, the
system is unstable, and vice versa. LEs are first applied to
power system stability analysis in [10], in which it is verified
that LEs can predict the out-of-step conditions of power
systems. In [2], a model-based MLE method is proposed
for online prediction of the rotor angle stability with PMU
measurements. The work builds solid analytical foundations
for the LE-based rotor angle stability assessment. In [11],
the LEs are calculated with dynamic component and network
models to identify the coherent groups of generators.
Although the model-based MLE estimation approaches have
made significant progresses on online rotor angle stability
assessment, they are usually computationally expensive especially when applied to large power systems. Therefore, two
model-free MLE estimation approaches have been proposed
for transient voltage stability assessment [12] and rotor angle
stability assessment [3], for which the MLEs can be estimated
by only using PMU measurements.
The model-free LE-based stability assessment approaches
are attractive, because they can eliminate model errors and
simplify the calculation. However, when applying these approaches, a time window has to be pre-specified for the MLE
observation. The window size is crucial for obtaining reliable
and timely assessment results, i.e., too small window size will
lead to unreliable assessment results while too large window
size will lead to accurate but untimely results. The window
2
size is difficult to be determined in advance, because the
estimated MLEs may fluctuate between positive and negative
values for quite a long time after disturbances and the window
size should change with different fault scenarios.
In this paper, a LE-based model-free rotor angle stability
assessment approach is proposed. The MLEs are estimated
by a recursive least square (RLS) based method based on
real-time rotor angle measurements. By properly choosing
two critical parameters according to the characteristics of the
relative rotor angles of the selected generator pairs, the calculated MLE curves will present distinct features for different
stability conditions, based on which the stability criteria are
correspondingly designed to capture the MLE features and
perform online rotor angle stability assessment. Compared
with the existing approaches, the proposed approach does not
need a pre-specified time window to identify the sign of MLEs.
Instead, the proposed approach can always make a reliable and
timely assessment as soon as the crucial features are observed
from the estimated MLE curves.
The remainder of this paper is organized as follows. Section
II introduces the theoretical basis of model-free MLE estimation. Section III proposes a rotor angle stability assessment
approach, and discusses the parameter selection principles
and stability criteria. In Section IV, simulation results on
the New-England 39-bus system and the Northeast Power
Coordinating Council (NPCC) 140-bus system are presented
to validate the effectiveness of the proposed approach. Finally,
the conclusions are drawn in Section V.
II. MLE E STIMATION F ROM T IME S ERIES
LEs can reflect the exponential divergence or convergence of
neighboring trajectories in the state space of a dynamic system
[13]. An N -dimensional dynamic system has N LEs and the
largest one is defined as the MLE of the system. MLE is a
useful indicator of system stability: A positive MLE indicates
unstable system dynamics while a negative MLE indicates
asymptotically stable dynamics [2]. The MLE can be estimated
by using Jacobian matrix based (model-based) methods [2],
[13] or direct model-free methods [3], [12], [13]. Compared
with the Jacobin matrix based methods, direct methods are
more suitable for online stability assessment mainly because
they do not need the repeated computing of the Jacobian
matrix or even the dynamic model of the system.
According to Oseledec’s multiplicative ergodic theorem
[13]–[15], for a reference point X 0 and its neighboring point
X m(0) chosen from the state space of a nonlinear dynamic
system, the distance between the trajectories emerged from
X 0 and X m(0) , i.e., the original trajectory and the neighboring trajectory, will have three different growth phases as
illustrated in Fig. 1. In Phase I, the difference vector between
the states of the trajectories gradually converges towards
the most expanding direction, and the distance between the
trajectories will exhibit fluctuations. In Phase II, the distance
experiences an exponential growth characterized by the MLE,
which corresponds to a linear segment in the semi-logarithmic
plot. Finally, in Phase III the separation of the trajectories is
saturated and the distance converges to a constant value.
Fig. 1. Logarithmic distance of neighboring states on different trajectories.
As proved by [13] and [15]–[17], the MLE can be estimated
from the mean logarithmic separation rate of the trajectories
in Phase II as
d (m(n), n, k)
1
log
λk ≈
k∆t
d (m(n), n, 0)
||X m(n)+k − X n+k ||
1
log
,
=
k∆t
||X m(n) − X n ||
X n , X m(n) , X n+k , and X m(n)+k ∈ Phase II,
(1)
where λk is the estimated MLE, k is the lagged time steps
for the MLE estimation, ∆t is the time duration for each
time step, X n and X m(n) are the initial points for the
MLE estimation on the original and neighboring trajectories,
respectively, X n+k is the kth point behind X n on the original
trajectory, X m(n)+k is the kth point behind X m(n) on the
neighboring trajectory, d (m(n), n, 0) is the Euclidean distance
between the MLE estimation initial points, d (m(n), n, k) is
the Euclidean distance between the kth points behind the MLE
estimation initial points, and ||A − B|| denotes the Euclidean
distance between points A and B.
It should be noted that the original and the neighboring
trajectories are usually from the same observed time series
with different initial points, as illustrated in Fig. 2. In order
to make sure the trajectories are temporally separated and
thus can be seen as different trajectories, the trajectory initial
points should satisfy |m(0) − 0| > w, where w is called the
Fig. 2. Trajectories and the effect of the Theiler window.
3
Theiler window [13] and should be determined according to
the characteristics of the system.
III. LE-BASED ROTOR A NGLE S TABILITY A SSESSMENT
Here we propose a rotor angle stability assessment approach
based on MLE estimation. It is data-driven and can perform
online stability assessment only using the rotor angle and rotor
speed of the generators [2], [3]. Although the rotor angle
and rotor speed may not be directly available from PMU
measurements, they can be estimated by various dynamic state
estimation methods [18]–[22].
A. State Variable Selection
Even a moderate-size power system may still have hundreds
of state variables. Due to both the calculation intractability and
the insufficiency of measurements, it is impractical to use all
these variables to form the state space for MLE estimation.
It is more feasible to reconstruct the power system dynamics
only with a small number of state variables.
According to Takens’ theorem [23], [24], the dynamics of a
nonlinear system can be reconstructed by the observations of
a single state variable, and the reconstructed state vector can
be expressed as
>
Θt = [θt , θt−τ , θt−2τ , · · · , θt−2N τ ] ,
of the severely disturbed generators, if |ωtc ,g | /ωt∗c > σ,
where σ is a predetermined threshold. In this paper σ is
chosen as 0.7, as in [27].
3) Find the least disturbed generator with the minimal
absolute value of the rotor speed.
4) Form a SDGP by combing one of the severely disturbed
generators and the least disturbed generator. Iterate over
the severely disturbed generators and form all SDGPs.
B. RLS-Based MLE Estimation
The MLE can be estimated by calculating the slope of
the logarithmic distance curve in Phase II. Considering the
influences of measurement errors and nonlinear fluctuations,
we adopt the least square algorithm to estimate the MLE.
From (1) and Fig. 2 it is seen that the MLE is estimated
starting from time step m(n). From this time step, k + 1
sequential logarithmic distances can be obtained as
L(m(n) + i) = log (d (m(n), n, i))
= log |θm(n)+i − θn+i | , i = 0, 1, . . . , k. (5)
Then, the MLE estimation model can be expressed as
L(m(n) + i) = λk · (m(n) + i)∆t + Ck + ξk ,
i = 0, 1, . . . , k,
(2)
(6)
where N is the dimension of the original system, τ is the lag
time, and θt , θt−τ , θt−2τ , · · · , θt−2N τ are the observations of
the observed variable at the corresponding time steps.
In the reconstructed state space, the Euclidean distance in
(1) becomes
where λk is the MLE to be estimated, Ck is the constant term,
and ξk is the residual term.
The solution of the least square estimation is
λk
−1
Ê k =
(7)
X>
= (X >
k Y k,
k Xk)
Ck
d(m(n), n, i) = ||Θm(n)+i − Θn+i ||, i ∈ {0, k},
where X k is the coefficient matrix and Y k is the observation
vector. According to the aforementioned definitions, X k and
Y k can be expressed as
(3)
where Θn+i and Θm(n)+i are the observations on the reconstructed trajectory.
Moreover, another simplification can be made based on
[13] and [17], in which it is shown that it is sufficient to
only consider the first component of the reconstructed state
vector to estimate MLE, because all components will grow
exponentially at the rate of MLE. Therefore, the Θn+i and
Θm(n)+i in (3) can be replaced by their first components, and
the distance in (1) can be calculated by
d(m(n), n, i) = |θm(n)+i − θn+i |, i ∈ {0, k}.
(4)
In this paper, the relative rotor angle of the severely disturbed generator pair (SDGP) is selected as the observed state
variable, because these SDGPs are, in general, responsible for
the system dynamics after considerable disturbances [25]. An
SDGP should be composed of a severely disturbed generator
and the least disturbed generator in order to reflect the dynamics of the severely disturbed generators [26]. In particular, the
SDGPs can be identified as follows.
1) Obtain the rotor speed of all generators at the fault
clearing moment, ωtc ,n , n = 1, 2, . . . , NG , from PMU
measurements, where NG is the number of generators.
2) Obtain the maximal absolute value of the rotor speed,
i.e., ωt∗c =
max
|ωtc ,n |. Define generator g as one
n=1,2,...,NG
m(n)∆t
(m(n) + 1)∆t
Xk =
···
(m(n) + k)∆t
1
1
,
···
1
L(m(n))
L(m(n) + 1)
Y k =
.
···
L(m(n) + k)
(8)
(9)
Moreover, to avoid the repeated calculation of the inverse
matrix in (7), a recursive estimation algorithm is applied [28],
which can be formulated as
h
i
Ê k+1 = Ê k + Gk+1 yk+1 − x>
k+1 Ê k ,
Gk+1 =
P k xk+1
,
1 + x>
k+1 P k xk+1
P k+1 = P k − Gk+1 x>
k+1 P k ,
>
(10)
where xk+1 is [(m(n) + k + 1)∆t 1] , yk+1 is the new
observation L(m(n)+k +1), Ê k and Ê k+1 are the estimation
results before and after obtaining the new observation, P k and
P k+1 are the covariance matrices, and Gk+1 is the gain vector.
The algorithm includes the following three steps:
4
1) With the first 2 groups of data (k = 1 in (5)), set the
−1
initial values of Ê 1 and P 1 to be (X >
X>
1 X 1)
1Y1
>
−1
and (X 1 X 1 ) , respectively.
2) Obtain a new observation of the logarithmic distance
between the trajectories, and then sequentially calculate
Gk+1 , Ê k+1 , and P k+1 according to (10).
3) Set k = k + 1 and return to step 2.
•
Pattern I: In this pattern, the relative rotor speed increases after fault clearing. No decelerating area exists
for the SDGP and the system will lose stability during
the first swing. Since the original and the neighboring
trajectories separate rapidly, w is set to 1 (the smallest
positive integer) to minimize the estimation waiting time.
•
Pattern II: In this pattern, the relative rotor speed decreases after fault clearing. However, since the decelerating area is relatively small, the relative rotor speed
increases again after a short time period, and the increasing trend continues until the system loses stability. The
key feature of this pattern is that the relative rotor speed
at the fault clearing moment, denoted by v0 in Fig. 3,
appears again after the initial decrease. In order to achieve
obvious separation between the trajectories, w is set to
be the time step lags of the reappearance of v0 .
•
Pattern III: In this pattern, the relative rotor speed first
decreases to −v0 , and then exhibits periodic oscillations.
The decelerating area is large enough to reduce the
relative rotor speed to zero, and the system stability
depends on the damping characteristics of the post-fault
equilibrium point. In this case, w is set to be the time
step lags of the first appearance of −v0 .
•
Pattern IV: In this pattern, the relative rotor speed first
decreases to some value greater than −v0 , and then
oscillates periodically. The key feature of this pattern
is that the relative rotor speed v0 and −v0 cannot be
observed after fault clearing. This pattern is a special case
of Pattern III, and w is set to be the time step lags of the
first appearance of the local minimal relative rotor speed
after fault clearing.
•
Pattern V: This pattern is similar to Pattern III, and
usually appears after very quick fault clearing. The key
feature of this pattern is that the relative rotor speed shows
decelerated growth immediately after fault clearing, and
the relative rotor speed −v0 can be observed after that.
The w is set in the same way as in Pattern III.
•
Pattern VI: This pattern is a special case of Pattern V.
The key feature of this pattern is that the relative rotor
speed shows decelerated growth immediately after the
fault clearing, and the relative rotor speed −v0 cannot
be observed during the oscillations. The w is set in the
same way as in Pattern IV.
C. Parameter Setting
Because the Theiler window w and the MLE estimation
initial time step m(n) determine the shape of the estimated
MLE curve, they are crucial for a quick and reliable rotor
angle stability assessment. Here, we discuss how to choose
these parameters according to the rotor angle swing features.
(1) Theiler window selection
The Theiler window w determines the temporal separation
between the initial points θ0 and θm(0) . It should be large
enough to ensure that θ0 and θm(0) are the initial points
of different trajectories. However, too large w will cause
unnecessary waiting time and delay the stability assessment.
The post-fault rotor angles of the SDGPs have significant
swing patterns [29]–[31]. According to the features of the
relative rotor speed of the SDGPs, six distinct swing patterns
can be identified as shown in Fig. 3. Different w are chosen
for different patterns as follows.
(2) MLE estimation initial time step selection
Fig. 3. Relative rotor speed curves of different patterns.
The MLE estimation initial time step m(n) should ensure
that the slope estimation is performed for Phase II of the
logarithmic distance growth. In this phase, the original and
neighboring trajectories have been sufficiently separated, and
thus the logarithmic distance curve has shown clear development trend [13].
In fact, the distance between the rotor angle trajectories
has a close relationship with the relative rotor speed of the
SDGP. According to the definition of the Theiler window,
the distance between the corresponding points of the original
5
and the neighboring trajectories at any time step j can be
reformulated as
dj = d(m(0), 0, j) = θm(0)+j − θ0+j
j+w
j+w
X
X
vt ·∆t ,
vt ·∆t−θ0+j =
≈ θ0+j +
t=j+1
(11)
t=j+1
where vt is the relative rotor speed at the relevant time step.
It is revealed in (11) that the distance dj is equal to the area
enclosed by the time axis and the relative rotor speed curve
within the Theiler window, as indicated by the shaded regions
in Fig. 3. Therefore, m(n) can be determined as follows.
•
Patterns I–II: With the selected w, it is seen in Fig. 3 that
the distance dj of these patterns monotonically increases
after fault clearing, which indicates that the logarithmic
distance growth is in Phase II as soon as the fault is
cleared. Therefore, in these two patterns, m(n) is set to
be w to minimize the estimation waiting time.
•
Patterns III–VI: For these patterns the distance will
exhibit periodic fluctuations after fault clearing. To ensure
the trajectories have been sufficiently separated and to
catch the main trend of the fluctuations, m(n) is set to
be w + j ∗ , where j ∗ is the time step when dj reaches its
first local maximum after fault clearing.
Criteria II and III correspond to Patterns III–VI. In these
patterns, the relative rotor angle will oscillate after fault
clearing. However, by using the selected parameters, the logarithmic distance will exhibit significant development trend as
illustrated in Fig. 5. If the relative rotor speed has a undamped
oscillation, the SDGP is unstable and the logarithmic distance
curve will fluctuate periodically as shown in Fig. 5(a). Since
the MLE is the average slope of the logarithmic distance curve
from the MLE estimation initial time step, it will first decrease
and then increase to a positive peak value, as in Fig. 4. By
contrast, if the relative rotor speed has a damped oscillation,
the SDGP is stable and the logarithmic distance curve will look
like Fig. 5(b). The MLE will first decrease and then increase
to a negative peak value, as in Fig. 4.
Fig. 5. The logarithmic distance curves for Pattern III–VI.
D. Rotor Angle Stability Assessment Criteria
When the MLE curve of a SDGP is estimated, the following
criteria can be used to determine the stability of the SDGP.
Criterion I: If the MLE of the SDGP increases at the
beginning, the SDGP is unstable.
• Criterion II: If the MLE decreases at the beginning,
it will have oscillations. If the first peak point of the
oscillation is positive, the SDGP is unstable.
• Criterion III: If the MLE decreases at the beginning
and the first peak point of the oscillation is negative, the
SDGP is stable.
Typical MLE curves corresponding to these criteria are
shown in Fig. 4. If the condition in Criterion I is satisfied,
Pattern I or II in Fig. 3 will happen, for which the distance
will increase immediately after the MLE estimation initial time
step and the trend will last until the SDGP loses stability.
•
Fig. 4. Typical MLE curves for Criterion I–III.
Because SDGPs are responsible for the system dynamics
after disturbances [25], we finally have the following criterion
on the angle stability of the system.
•
Criterion IV: If all SDGPs are stable, the system is
stable; otherwise, the system is unstable.
E. Rotor Angle Stability Assessment Procedure
The proposed online rotor angle stability assessment procedure is shown in Fig. 6, which includes the measurement
data preparation module, the parameter setting module, the
MLE estimation module, and the stability assessment module.
Specifically,
1) When a fault is detected, the measurement data preparation module will be immediately activated to identify
the SDGPs and collect the corresponding relative rotor
angle and rotor speed measurements in real time.
2) The relative rotor speed variation patterns are identified
online according to the rules described in Section III-C,
and the parameters are assigned accordingly.
3) The MLE sequences can be calculated by using the RLSbased algorithm in Section III-B.
4) Finally the stability condition can be assessed by using
the criteria provided in Section III-D according to the
features of the estimated MLE curves.
IV. C ASE S TUDIES
The proposed approach is tested on the New-England 39bus system and the NPCC 140-bus system. Simulations are
performed with Power System Analysis Toolbox (PSAT) [32]
in MATLAB.
6
Fig. 6. Flow chart of the assessment procedure.
A. New-England 39-Bus System
The New-England 39-bus system has 10 generators and
46 branches. The parameters can be found in [33]. Unless
otherwise specified, all generators in the tests are described by
the fourth-order transient model with Type I turbine governor
(TG), Type II automatic voltage regulator (AVR), and Type
II power system stabilizer (PSS) (see PSAT documentation).
All loads are described by the ZIP model and the ratios of
the constant impedance, constant current, and constant power
loads are 0.4, 0.5, and 0.1, respectively. The sampling rate
of the PMU measurements used for MLE estimation is 120
samples/s [3].
To verify the effectiveness of the proposed approach, a
three-phase-to-ground fault is applied at bus 2 at t = 1 s, and
the fault is cleared by opening line 2–3 at tc = 1.243 s and
tc = 1.244 s for Scenarios I and II, respectively. According
to the time-domain simulation, the system is stable under
Scenario I and unstable under Scenario II.
By using the method in Section III-A, the generator pairs
38–39 and 37–39 are identified as the SDGPs for both scenarios. Figs. 7–8 show the relative rotor angles and the estimated
MLEs of the SDGPs.
From Fig. 7(a), it is seen that the relative rotor angles of the
SDGPs tend to be stable after a long period of oscillations. In
Fig. 7(b), the MLEs of the SDGPs decrease immediately after
fault clearing, and their first peak points of the oscillations are
both less than 0. Therefore, according to Criteria III and IV,
the system is stable under this scenario.
By contrast, under Scenario II the MLE of the generator pair
38–39 increases immediately after fault clearing, as shown in
Fig. 8(b). Therefore, according to Criteria I and IV, the system
is unstable under this scenario.
For the two scenarios, the system stability can be assessed,
respectively, within 2.82 s and 1.40 s after fault clearing,
which demonstrates that the proposed approach can provide
early detection of instability. In fact, according to the proposed approach, the first-swing instability can be identified
very quickly, because the instability can be detected at the
beginning of the MLE curve. On the other hand, for the multiswing stability or instability, the assessment time is a little
longer since the approach needs to check the first peak point
of the MLE curve after fault clearing.
In industrial applications, a predetermined relative rotor
angle value, i.e., π rad, is usually set as the threshold for
determining rotor angle stability [34]. Although this pragmatic
criterion is easy to execute, the relative rotor angle corresponding to rotor angle instability may change significantly with the
change of topologies, parameters, and operating conditions.
For instance, under Scenario I, the relative rotor angle between
generators 38 and 39 can reach up to 3.207 rad while the
system is still stable.
It should also be noted that there are foundational differences between the proposed approach and that in [3]. In
the proposed approach, the system stability can be explicitly
determined at latest when the first peak point of the MLE curve
is observed. By contrast, for the approach in [3], the MLE
curve must be observed for a long period to ensure the sign
of MLE. Unfortunately, it is actually difficult for the approach
in [3] to predetermine the observing window size to provide
reliable and timely assessment results (for instance, see Figs.
Fig. 7. Simulation results of Scenario I.
Fig. 8. Simulation results of Scenario II.
7
1–2 in [3]). From this perspective, the proposed approach is
more time efficient and more reliable.
In order to further verify the accuracy of the proposed
approach, extensive tests are performed on the New-England
39-bus system. Specifically, a three-phase-to-ground fault is
created for each bus (except the generator buses) at t = 1 s,
and is cleared at 1.08 s, 1.16 s, 1.24 s, and 1.32 s, respectively. According to the test results, the proposed approach
successfully determines the rotor angle stability in all 224
tests. The occurrence frequencies of the fault patterns and the
success rate of the proposed stability assessment approach are
summarized in Table I.
TABLE I
S UMMARY OF THE T ESTS ON THE N EW-E NGLAND 39-B US S YSTEM
tc /s
1.08
1.16
1.24
1.32
Success Rate
I
N/A
N/A
4
38
100%
time is 2.36 s, which is much shorter than the time required
for directly looking at the relative rotor angle curve.
B. NPCC 140-Bus System
The proposed approach is also tested on the NPCC 140bus system [35]. The settings are the same as those for the
New-England 39-bus system.
A three-phase-to-ground fault is first applied at bus 35 at
t = 0.1 s, and is cleared by opening line 34–35 at tc = 0.307 s
(Scenario III, stable) and tc = 0.308 s (Scenario IV, unstable),
respectively. Generator pairs 1–48 and 2–48 are identified as
the SDGPs for both cases. The relative rotor angle curves and
MLE curves under these two scenarios are shown in Figs.
10–11, respectively.
Occurrence Times of the Swing Patterns
II
III
IV
V
VI
N/A
4
43
6
3
6
10
36
3
1
13
9
27
2
1
12
2
4
N/A
N/A
100%
100%
100%
100%
100%
It is seen in Table I that Patterns I and II those correspond to
first-swing instability usually occur with longer fault clearing
time. In contrast, Patterns V and IV usually occur with the
shorter fault clearing time, and the system will have a good
chance to maintain stability in these patterns. Among the tests,
the first-swing instability identification time ranges from 1.2 s
to 1.5 s, and the multi-swing stability assessment time ranges
from 2.2 s to 2.5 s, which further confirms the efficient of the
proposed approach.
Because the system has sufficient damping to ensure the
stability of the post-fault equilibrium points, there are no multiswing instability cases. In order to generate a multi-swing
instability case, all of the PSSs are removed and the parameters
of the AVRs are tuned. In the modified system, a three-phaseto-ground fault is applied at bus 28 at t = 1 s, and is cleared by
opening line 27–28 at tc = 1.12 s. In this test, only generator
pair 38–39 is identified as the SDGP, whose relative rotor angle
curve and MLE curve are shown in Fig. 9.
Fig. 9. Simulation results of the multi-swing unstable case.
In Fig. 9(b), the MLE decreases immediately after fault
clearing, and the following peak MLE is positive. According
to Criteria II and IV, the MLE variation pattern indicates that
the system is multi-swing unstable, which is verified by the
relative rotor angle curve. In this case, the stability assessment
Fig. 10. Simulation results of Scenario III.
Fig. 11. Simulation results of Scenario IV.
In Fig. 10, it is seen that the curves of Scenario III are
similar to those of Scenario I. In this case, the system is
determined as stable because of the same reason as Scenario
I. As shown in Fig. 11, under Scenario IV the MLE curves
of both SDGPs exhibit unstable features, which indicates that
both generator pairs will lose stability and thus the system is
unstable.
Extensive tests are also executed on the NPCC 140-bus
system. A three-phase-to-ground fault is applied at each bus
(except the generator buses) at t = 0.1 s, and is cleared at
0.18 s, 0.26 s, 0.32 s and 0.40 s, respectively. Table II lists
the occurrence frequencies of the fault patterns and the success
rate of the proposed stability assessment approach during the
tests.
According to the test results, the proposed approach can
accurately determine system stability in all 716 tests. The firstswing instability identification time ranges from 1.1 s to 1.7
8
TABLE II
S UMMARY OF THE T ESTS ON THE NPCC 140-B US S YSTEM
tc /s
0.18
0.26
0.32
0.40
Success Rate
I
N/A
N/A
8
60
100%
Occurrence Times of
II
III
N/A
6
10
26
57
38
12
49
100%
100%
the Swing Patterns
IV
V
VI
166
2
5
143
N/A
N/A
76
N/A
N/A
58
N/A
N/A
100%
100%
100%
s, and the multi-swing stability assessment time ranges from
1.8 s to 2.4 s.
V. C ONCLUSION
In this paper, a model-free approach for online rotor angle
stability assessment is proposed based on MLE. By using the
proposed MLE estimation algorithm, parameter setting rules
and the stability criteria, the approach can online identify
the system stability condition with PMU measurements. The
approach does not need a predetermined observing window
to identify the sign of the MLE, and can provide reliable
and timely assessment results by analyzing the features of
the estimated MLE curve. To verify the performance of the
proposed approach, extensive tests are performed on the NewEngland 39-bus system and the NPCC 140-bus system. The
proposed approach can successfully determine the system
stability conditions in all 945 tests. Moreover, among all the
tests, the first-swing stability can be assessed within 1.7 s and
the multi-swing stability can be assessed within 2.5 s.
R EFERENCES
[1] P. Kundur, J. Paserba, V. Ajjarapu, G. Andersson, A. Bose et al.,
“Definition and classification of power system stability,” IEEE Trans.
Power Syst., vol. 19, no. 3, pp. 1387–1401, Aug. 2004.
[2] J. Yan, C. C. Liu, and U. Vaidya, “PMU-based monitoring of rotor angle
dynamics,” IEEE Trans. Power Syst., vol. 26, no. 4, pp. 2125–2133, Mar.
2011.
[3] S. Dasgupta, M. Paramasivam, U. Vaidya, and V. Ajjarapu, “PMU-based
model-free approach for real-time rotor angle monitoring,” IEEE Trans.
Power Syst., vol. 30, no. 5, pp. 2818–2819, Sep. 2015.
[4] V. Centeno, A. Phadke, A. Edris, J. Benton, M. Gaudi, and G. Michel,
“An adaptive out-of-step relay,” IEEE Trans. Power Delivery, vol. 12,
no. 1, pp. 61–71, Jan. 1997.
[5] J. H. Chow, A. Chakrabortty, M. Arcak, B. Bhargava, and A. Salazar,
“Synchronized phasor data based energy function analysis of dominant
power transfer paths in large power systems,” IEEE Trans. Power Syst.,
vol. 22, no. 2, pp. 727–734, May 2007.
[6] C. W. Liu and J. S. Thorp, “New methods for computing power system
dynamic response for real-time transient stability prediction,” IEEE
Trans. Circuits Syst. I, vol. 47, no. 3, pp. 324–337, Mar. 2000.
[7] K. Sun, S. Likhate, V. Vittal et al., “An online dynamic security
assessment scheme using phasor measurements and decision trees,”
IEEE Trans. Power Syst., vol. 22, no. 4, pp. 1935–1943, Nov. 2007.
[8] A. Del Angel, P. Geurts, D. Ernst, M. Glavic, and L. Wehenkel,
“Estimation of rotor angles of synchronous machines using artificial
neural networks and local PMU-based quantities,” Neurocomputing,
vol. 70, no. 16, pp. 2668–2678, Oct. 2007.
[9] I. Kamwa, S. Samantaray, and G. Joos, “Development of rule-based
classifiers for rapid stability assessment of wide-area post-disturbance
records,” IEEE Trans. Power Syst., vol. 24, no. 1, pp. 258–270, Feb.
2009.
[10] C. W. Liu, J. S. Thorp, J. Lu et al., “Detection of transiently chaotic
swings in power systems using real-time phasor measurements,” IEEE
Trans. Power Syst., vol. 9, no. 3, pp. 1285–1292, Aug. 1994.
[11] S. K. Khaitan and J. D. McCalley, “Vantage: A lyapunov exponents
based technique for identification of coherent groups of generators in
power systems,” Electr. Pow. Syst. Res., vol. 105, pp. 33–38, Dec. 2013.
[12] S. Dasgupta, M. Paramasivam, U. Vaidya, and V. Ajjarapu, “Real-time
monitoring of short-term voltage stability using PMU data,” IEEE Trans.
Power Syst., vol. 28, no. 4, pp. 3702–3711, Nov. 2013.
[13] C. Skokos, G. A. Gottwald, and J. Laskar, Chaos Detection and
Predictability. Heidelberg, Germany: Springer, 2016.
[14] V. I. Oseledec, “A multiplicative ergodic theorem. lyapunov characteristic numbers for dynamical systems,” Trans. Moscow Math. Soc, vol. 19,
no. 2, pp. 197–231, Dec. 1968.
[15] M. T. Rosenstein, J. J. Collins, and C. J. De Luca, “A practical method
for calculating largest lyapunov exponents from small data sets,” Physica
D, vol. 65, no. 1-2, pp. 117–134, May 1993.
[16] J. Gao and Z. Zheng, “Local exponential divergence plot and optimal
embedding of a chaotic time series,” Phys. Lett. A, vol. 181, no. 2, pp.
153–158, Oct. 1993.
[17] H. Kantz, “A robust method to estimate the maximal lyapunov exponent
of a time series,” Phys. Lett. A, vol. 185, no. 1, pp. 77–87, Jan. 1994.
[18] Z. Huang, K. Schneider, and J. Nieplocha, “Feasibility studies of applying kalman filter techniques to power system dynamic state estimation,”
in 2007 International Power Engineering Conference (IPEC 2007).
IEEE, Dec. 2007, pp. 376–382.
[19] E. Ghahremani and I. Kamwa, “Dynamic state estimation in power
system by applying the extended kalman filter with unknown inputs
to phasor measurements,” IEEE Trans. Power Syst., vol. 26, no. 4, pp.
2556–2566, Nov. 2011.
[20] K. Sun, J. Qi, and W. Kang, “Power system observability and dynamic
state estimation for stability monitoring using synchrophasor measurements,” Control Eng. Pract., vol. 53, pp. 160–172, Aug. 2016.
[21] E. Ghahremani and I. Kamwa, “Local and wide-area pmu-based decentralized dynamic state estimation in multi-machine power systems,”
IEEE Trans. Power Syst., vol. 31, no. 1, pp. 547–562, Jan. 2016.
[22] J. Qi, K. Sun, J. Wang, and H. Liu, “Dynamic state estimation for
multi-machine power system by unscented Kalman filter with enhanced
numerical stability,” IEEE Trans. Smart Grid, to be published.
[23] F. Takens, “Detecting strange attractors in turbulence,” in Dynamical
systems and turbulence, Warwick 1980. Heidelberg, Germany: Springer,
1981, pp. 366–381.
[24] N. H. Packard, J. P. Crutchfield, J. D. Farmer, and R. S. Shaw, “Geometry
from a time series,” Phys. Rev. Lett., vol. 45, no. 9, pp. 712–716, Nov.
1980.
[25] M. Haque and A. Rahim, “Determination of first swing stability limit
of multimachine power systems through taylor series expansions,” IEE
Proc.-C, vol. 136, no. 6, pp. 373–380, Nov. 1989.
[26] M. Yin, C. Chung, K. Wong, Y. Xue, and Y. Zou, “An improved iterative
method for assessment of multi-swing transient stability limit,” IEEE
Trans. Power Syst., vol. 26, no. 4, pp. 2023–2030, Nov. 2011.
[27] M. Haque, “Further developments of the equal-area criterion for multimachine power systems,” Electr. Pow. Syst. Res., vol. 33, no. 3, pp.
175–183, Jun. 1995.
[28] P. C. Young, Recursive estimation and time-series analysis: An introduction for the student and practitioner. Heidelberg, Germany: Springer,
2011.
[29] B. Spalding, H. Yee, and D. Goudie, “Coherency recognition for
transient stability studies using singular points,” IEEE Trans. Power App.
Syst., vol. 96, no. 4, pp. 1368–1375, Jul. 1977.
[30] M. Takahashi, K. Matsuzawa, M. Sato, K. Omata, R. Tsukui, T. Nakamura, and S. Mizuguchi, “Fast generation shedding equipment based
on the observation of swings of generators,” IEEE Trans. Power Syst.,
vol. 3, no. 2, pp. 439–446, May 1988.
[31] G. G. Karady, A. A. Daoud, and M. A. Mohamed, “On-line transient
stability enhancement using multi-agent technique,” in Proc. IEEE
Power Eng. Soc. Winter Meeting, 2002, vol. 2, pp. 893–899.
[32] F. Milano, “An open source power system analysis toolbox,” IEEE Trans.
Power Syst., vol. 20, no. 3, pp. 1199–1206, Aug. 2005.
[33] A. Pai, Energy function analysis for power system stability. Boston,
MA, USA: Kluwer Academic Publishers, 1989.
[34] Y. Xu, Z. Y. Dong, J. H. Zhao, P. Zhang, and K. P. Wong, “A reliable
intelligent system for real-time dynamic security assessment of power
systems,” IEEE Trans. Power Syst., vol. 27, no. 3, pp. 1253–1263, Aug.
2012.
[35] J. H. Chow, R. Galarza, P. Accari, and W. Price, “Inertial and slow
coherency aggregation algorithms for power system dynamic model
reduction,” IEEE Trans. Power Syst., vol. 10, no. 2, pp. 680–685, May
1995.
| 3cs.SY
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.